Sentiment Analysis For Social Media

Sentiment Analysis
for Social Media

Edited by
Carlos A. Iglesias and Antonio Moreno
Printed Edition of the Special Issue Published in Applied Sciences
www.mdpi.com/journal/applsci
Sentiment Analysis for Social Media
Special Issue Editors

Carlos A. Iglesias
Antonio Moreno
MDPI • Basel • Beijing • Wuhan • Barcelona • Belgrade • Manchester • Tokyo • Cluj • Tianjin
Special Issue Editors
Carlos A. Iglesias Antonio Moreno
Departamento de Ingenierı́a de Departament d’Enginyeria
Sistemas Telemáticos, Informàtica i Matemàtiques,
ETSI Telecomunicación Escola Tècnica Superior
Spain d’Enginyeria, Universitat Rovira
i Virgili (URV)
Spain
Editorial Office
MDPI
St. Alban-Anlage 66
4052 Basel, Switzerland
This is a reprint of articles from the Special Issue published online in the open access journal
Applied Sciences (ISSN 2076-3417) (available at: https://www.mdpi.com/journal/applsci/special
issues/Sentiment Social Media).
For citation purposes, cite each article independently as indicated on the article page online and as
indicated below:
LastName, A.A.; LastName, B.B.; LastName, C.C. Article Title. Journal Name Year, Article Number,
Page Range.
ISBN 978-3-03928-572-3 (Pbk)

ISBN 978-3-03928-573-0 (PDF)

c 2020 by the authors. Articles in this book are Open Access and distributed under the Creative
Commons Attribution (CC BY) license, which allows users to download, copy and build upon
published articles, as long as the author and publisher are properly credited, which ensures maximum
dissemination and a wider impact of our publications.
The book as a whole is distributed by MDPI under the terms and conditions of the Creative Commons
license CC BY-NC-ND.
Contents
About the Special Issue Editors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
Carlos A. Iglesias and Antonio Moreno

Reprinted from: Appl. Sci. 2019, 9, 5037, doi:10.3390/app9235037 . . . . . . . . . . . . . . . . . . . 1
Hyoji Ha, Hyunwoo Han, Seongmin Mun, Sungyun Bae, Jihye Lee and Kyungwon Lee
An Improved Study of Multilevel Semantic Network Visualization for Analyzing Sentiment
Word of Movie Review Data
Hannah Kim and Young-Seob Jeong

Sentiment Classification Using Convolutional Neural Networks
Xingliang Mao, Shuai Chang, Jinjing Shi, Fangfang Li and Ronghua Shi
Sentiment-Aware Word Embedding for Emotion Classification
Mohammed Jabreel and Antonio Moreno

A Deep Learning-Based Approach for Multi-Label Emotion Classification in Tweets
Eline M. van den Broek-Altenburg and Adam J. Atherly

Using Social Media to Identify Consumers’ Sentiments towards Attributes of Health Insurance
during Enrollment Season
Sunghee Park and Jiyoung Woo

Gender Classification Using Sentiment Analysis and Deep Learning in a Health Web Forum
Hui Liu, Yinghui Huang, Zichao Wang, Kai Liu, Xiangen Hu and Weijun Wang
Personality or Value: A Comparative Study of Psychographic Segmentation Based on an Online
Review Enhanced Recommender System
Guadalupe Obdulia Gutiérrez-Esparza, Maite Vallejo-Allende and José Hernández-Torruco

Classification of Cyber-Aggression Cases Applying Machine Learning
v
About the Special Issue Editors
Carlos A. Iglesias (Professor). Prof. Carlos Iglesias Associate Professor at the Universidad
Politécnica de Madrid. He holds a Ph.D. in Telecommunication Engineering. He was previously
Deputy Director at Grupo Gesfor and Innovation Director at Germinus XXI. He has been actively
involved in research projects funded by private companies as well as national and European
programs. His research interests are focused on intelligent systems (knowledge engineering,
multi-agent systems, machine learning, and natural language processing).
Antonio Moreno (Professor). Dr. Moreno is a Full Professor of Artificial Intelligence at

University Rovira i Virgili (URV) in Tarragona, Spain. He was the founder and director of the
ITAKA (Intelligent Technologies for Advanced Knowledge Acquisition) research group until 2019.
Since 2018, he has been the Deputy Director of URV Engineering School. He has been the author
of more than 60 journal papers and over 125 conference papers. He has supervised 8 Ph.D. theses
on different topics, including ontology learning, agents applied in health care, intelligent data
analysis applied on healthcare data, recommender systems, and multi-criteria decision making. His
current research interests are focused on sentiment analysis, recommender systems, and multi-criteria
decision support systems.
vii
applied
sciences
Editorial
Carlos A. Iglesias 1, *,† and Antonio Moreno 2,†
1 Intelligent Systems Group, ETSI Telecomunicación, Avda. Complutense 30, 28040 Madrid, Spain
2 Intelligent Technologies for Advance Knowledge Acquisition (ITAKA) Group,
Escola Tècnica Superior d’Enginyeria, Departament d’Enginyeria Informàtica i Matemàtiques,
Universitat Rovira i Virgili, 43007 Tarragona, Spain; [email protected]
* Correspondence: [email protected]; Tel.: +34-910671900
† These authors contributed equally to this work.

Received: 31 October 2019; Accepted: 19 November 2019; Published: 22 November 2019
Abstract: Sentiment analysis has become a key technology to gain insight from social networks.
The field has reached a level of maturity that paves the way for its exploitation in many different
fields such as marketing, health, banking or politics. The latest technological advancements, such as
deep learning techniques, have solved some of the traditional challenges in the area caused by the
scarcity of lexical resources. In this Special Issue, different approaches that advance this discipline
are presented. The contributed articles belong to two broad groups: technological contributions
and applications.
Keywords: sentiment analysis; emotion analysis; social media; affect computing
1. Introduction
Sentiment analysis technologies enable the automatic analysis of the information distributed
through social media to identify the polarity of posted opinions [1]. These technologies have been
extended in the last years to analyze other aspects, such as the stance of a user towards a topic [2]
or the users’ emotions [3], even combining text analytics with other inputs, including multimedia
analysis [4] or social network analysis [5].
This Special Issue “Sentiment Analysis for Social Media" aims to reflect recent developments in
sentiment analysis and to present new advances in sentiment analysis that enable the development
of future sentiment analysis and social media monitoring methods. The following sections detail the
selected works in the development of new techniques as well as new applications.
2. New Paths in Sentiment Analysis on Social Media

Traditionally, sentiment analysis has focused on text analysis using Natural Language Processing
and feature-based Machine Learning techniques. The advances in disciplines such as Big Data and
Deep Learning technologies have impacted and benefited the evolution of the field. This Special Issue
includes four works that propose novel techniques.
In the first work, titled “An Improved Study of Multilevel Semantic Network Visualization for Analyzing
Sentiment Word of Movie Review Data” [6], Ha et al. propose a method for sentiment visualization in
massive social media. For this purpose, they design a multi-level sentiment network visualization
mechanism based on emotional words in the movie review domain. They propose three visualization
methods: a heatmap visualization of the semantic words of every node, a two-dimensional scaling
map of semantic word data, and a constellation visualization using asterism images for each cluster
of the network. The proposed visualizations have been used as a recommender system that suggest
movies with similar emotions to the previously watched ones. This novel idea of recommending
contents based on similar emotional patterns can be applied to other social networks.
Appl. Sci. 2019, 9, 5037; doi:10.3390/app9235037 1 www.mdpi.com/journal/applsci

Appl. Sci. 2019, 9, 5037
In the second contribution, titled “Sentiment Classification Using Convolutional Neural Networks” [7],
Kim and Jeong deal with the problem of textual sentiment classification. They propose a Convolutional
Neural Network (CNN) model consisting of an embedding layer, two convolutional layers, a pooling
layer, and a fully-connected layer. The model is evaluated in three datasets (movie review data,
customer review data and Stanford Sentiment Treebank data) and compared with traditional Machine
Learning models and state of the art Deep Learning models. Their main conclusion is that the use of
consecutive convolutional layers is effective for relatively long texts.
In the third work, titled “Sentiment-Aware Word Embedding for Emotion Classification” [8], Mao et al.
suggest the use of a sentiment-aware word embedding for improving emotional analysis. The proposed
method builds a hybrid representation that combines emotional word embeddings based on
an emotional lexicon with semantic word embeddings based on Word2Vec [9]. They use the emotional
lexicon DUTIR, which is a Chinese ontology resource collated and labeled by the Dalian University
of Technology Information Retrieval Laboratory [10]. This resource annotates lexicon entries with
a model of seven emotions (happiness, trust, anger, sadness, fear, disgust and surprise). The evaluation
is done with data from Weibo, a popular Chinese social networking site. The paper evaluates two
methods (direct combination and addition) for building the hybrid representation in several datasets.
They conclude that the experiments prove that the use of hybrid word vectors is effective for supervised
emotion classification, improving significantly the classification accuracy.
Finally, in the fourth theoretical contribution, titled “A Deep Learning-Based Approach for Multi-Label
Emotion Classification in Tweets” [11], Jabreel and Moreno address the problem of multi-class emotion
classification based on Deep Learning techniques. The most popular approach for this problem is
to transform it into multiple binary classification problems, one for each emotion class. This paper
proposes a new transformation approach, so-called xy-pair-set, that transforms the original problem
into just one binary classification problem. The transformation problem is solved with a Deep
Learning-based system, so-called BNet. This system consists of three modules: an embedding module
that uses three embedding models and an attention function, an encoding module based on Recurrent
Neural Networks (RNNs), and a classification module that uses two feed-forward layers with the
ReLU activation function followed by a sigmoid unit. The system is evaluated using the dataset “Affect
in Tweets” of SemEval-2019 Task 1 [2], and it outperformed the state of the art systems.
3. Applications of Sentiment Analysis in Social Media

The wide range of applications of sentiment analysis has fostered its evolution. Sentiment analysis
techniques have enabled to make sense of big social media data to make more informed decisions and
understand social events, product marketings or political events. Four works selected in this Special
Issue deal with the application of sentiment analysis for improving health insurances, understanding
AIDS patients, e-commerce user profiling and cyberagression detection.
In the first work, titled “Using Social Media to Identify Consumers’ Sentiments towards Attributes
of Health Insurance during Enrollment Season” [12], van den Broek-Altenburg and Atherly aim at
understanding the consumers’ sentiments towards health insurances. For this purpose, they mined
Twitter discussions and analyzed them using a dictionary-based approach using the NRC Emotion
Lexicon [13], which provides for each word its polarity as well as its related emotion (anger, anticipation,
disgust, fear, joy, sadness, surprise and trust). The main finding of this study is that consumers are
worried about providers networks, prescription drug benefits and political preferences. In addition,
consumers trust medical providers but fear unexpected events. These results suggest that more
research is needed to understand the origin of the sentiments that drive consumers so that insurers can
provide better insurance plans.
In the second contribution, titled “Gender Classification Using Sentiment Analysis and Deep Learning
in a Health Web Forum” [14], Park and Woo deal also with the application of sentiment analysis
techniques to health-related topics. In particular, they apply sentiment analysis for identifying
gender in health forums based on Deep Learning techniques. The authors analyze messages from
2
Appl. Sci. 2019, 9, 5037
an AIDS-related bulletin board from HealthBoard.com and evaluate both traditional and Deep Learning
techniques for gender classification.
In the third approach [15], titled “Personality or Value: A Comparative Study of Psychographic
Segmentation Based on an Online Review Enhanced Recommender System”, Liu et al. analyze the predictive
and explanatory capability of psychographic characteristics in e-commerce user preferences. For this
purpose, they construct a pychographic lexicon based on seed words provided by psycholinguistics that
are expanded using synonyms from WordNet [16], resulting in positive and negative lexicons for two
psychographic models, Schwartz Value Survey (SVS) [17] and Big Five Factor (BFF) [18]. Then they
construct word embeddings using Word2Vec [9] and extend the corpus with word embeddings
from an Amazon corpus [19]. Finally, they incorporate the lexicons in a deep neural network-based
recommender system to predict the users’ online purchasing behaviour. They also evaluate customer
segmentation based on BDSCAN clustering [20], but this does not provide a significant improvement.
The main insight of this research is that psychographic variables improve the explanatory power of
e-consumer preferences, but their prediction capability is not significant.
Finally, in the fourth work [21], titled “Classification of Cyber-Aggression Cases Applying Machine
Learning”, Gutiérrez-Esparza et al. deal with the detection of cyberagression. They build and label
a corpus of cyberagression news from Facebook in Latinamerica and develop a classification model
based on Machine Learning techniques. The developed corpus can foster research in this field, given
the scarcity of lexical resources in languages different from English.
4. Conclusions
The diversity of approaches of the articles included in this Special Issue shows the great interest
and dynamism of this field. Moreover, this Special Issue of Applied Sciences contributes to provide
a good overview of some of the main areas of research in this field.
Funding: This research received no external funding.

Acknowledgments: The Guest Editors would like to thank all the authors that have participated in this Special
Issue and also the reference contact in MDPI, Nyssa Yuan, for all the support and work dedicated to the success of
this Special Issue.
Conflicts of Interest: The authors declare no conflict of interest.
References
1. Liu, B. Sentiment analysis and opinion mining. Synth. Lect. Hum. Lang. Technol. 2012, 5, 1–167.
2. Mohammad, S.; Bravo-Marquez, F.; Salameh, M.; Kiritchenko, S. Semeval-2018 task 1: Affect in tweets.
In Proceedings of the 12th International Workshop on Semantic Evaluation, New Orleans, LA, USA,
5–6 June 2018; pp. 1–17.
3. Cambria, E.; Poria, S.; Hussain, A.; Liu, B. Computational Intelligence for Affective Computing and
Sentiment Analysis [Guest Editorial]. IEEE Comput. Intell. Mag. 2019, 14, 16–17.
4. Li, Z.; Fan, Y.; Jiang, B.; Lei, T.; Liu, W. A survey on sentiment analysis and opinion mining for social
multimedia. Multimed. Tools Appl. 2019, 78, 6939–6967.
5. Sánchez-Rada, J.F.; Iglesias, C.A. Social context in sentiment analysis: Formal definition, overview of current
trends and framework for comparison. Inf. Fusion 2019, 52, 344–356.
6. Ha, H.; Han, H.; Mun, S.; Bae, S.; Lee, J.; Lee, K. An Improved Study of Multilevel Semantic Network
Visualization for Analyzing Sentiment Word of Movie Review Data. Appl. Sci. 2019, 9, 2419. [CrossRef]
7. Kim, H.; Jeong, Y.S. Sentiment Classification Using Convolutional Neural Networks. Appl. Sci. 2019, 9, 2347.
[CrossRef]
8. Mao, X.; Chang, S.; Shi, J.; Li, F.; Shi, R. Sentiment-Aware Word Embedding for Emotion Classification.
Appl. Sci. 2019, 9, 1334. [CrossRef]
9. Mikolov, T.; Chen, K.; Corrado, G.; Dean, J. Efficient estimation of word representations in vector space.
arXiv 2013, arXiv:1301.3781.
3
Appl. Sci. 2019, 9, 5037
10. Chen, J. The Construction and Application of Chinese Emotion Word Ontology. Master’s Thesis, Dailian
University of Technology, Dalian, China, 2008.
11. Jabreel, M.; Moreno, A. A Deep Learning-Based Approach for Multi-Label Emotion Classification in Tweets.
12. Van den Broek-Altenburg, E.M.; Atherly, A.J. Using Social Media to Identify Consumers’ Sentiments towards
Attributes of Health Insurance during Enrollment Season. Appl. Sci. 2019, 9, 2035. [CrossRef]
13. Mohammad, S.M.; Kiritchenko, S.; Zhu, X. NRC-Canada: Building the state-of-the-art in sentiment analysis
of tweets. arXiv 2013, arXiv:1308.6242.
14. Park, S.; Woo, J. Gender Classification Using Sentiment Analysis and Deep Learning in a Health Web Forum.
15. Liu, H.; Huang, Y.; Wang, Z.; Liu, K.; Hu, X.; Wang, W. Personality or Value: A Comparative Study of
Psychographic Segmentation Based on an Online Review Enhanced Recommender System. Appl. Sci. 2019,
9, 1992. [CrossRef]
16. Miller, G.A. WordNet: A lexical database for English. Commun. ACM 1995, 38, 39–41. [CrossRef]
17. Sagiv, L.; Schwartz, S.H. Cultural values in organisations: Insights for Europe. Eur. J. Int. Manag. 2007,
1, 176–190. [CrossRef]
18. McCrae, R.R.; Costa, P.T., Jr. The five-factor theory of personality. In Handbook of Personality: Theory and
Research; The Guilford Press: New York, NY, USA, 2008; pp. 159–181.
19. McAuley, J.; Targett, C.; Shi, Q.; Van Den Hengel, A. Image-based recommendations on styles and
substitutes. In Proceedings of the 38th International ACM SIGIR Conference on Research and Development
in Information Retrieval, Santiago, Chile, 9–13 August 2015; pp. 43–52.
20. Ester, M.; Kriegel, H.P.; Sander, J.; Xu, X.. A density-based algorithm for discovering clusters in large spatial
databases with noise. In KDD-96 Proceddings; AAAI Press: Portland, OR, USA, 1996; pp. 226–231.
21. Gutiérrez-Esparza, G.O.; Vallejo-Allende, M.; Hernández-Torruco, J. Classification of Cyber-Aggression
Cases Applying Machine Learning. Appl. Sci. 2019, 9, 1828. [CrossRef]
c 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).
4
applied
sciences
Article
An Improved Study of Multilevel Semantic Network
Visualization for Analyzing Sentiment Word of
Movie Review Data
Hyoji Ha 1 , Hyunwoo Han 1 , Seongmin Mun 1,2 , Sungyun Bae 1 , Jihye Lee 1 and
Kyungwon Lee 3, *
1 Lifemedia Interdisciplinary Program, Ajou University, Suwon 16499, Korea; [email protected] (H.H.);
[email protected] (H.H.); [email protected] (S.M.); [email protected] (S.B.); [email protected] (J.L.)
2 Science of Language, MoDyCo UMR 7114 CNRS, University Paris Nanterre, 92000 Nanterre, France
3 Department of Digital Media, Ajou University, Suwon 16499, Korea
* Correspondence: [email protected]

Received: 30 April 2019; Accepted: 31 May 2019; Published: 13 June 2019
Abstract: This paper suggests a method for refining a massive amount of collective intelligence data
and visualizing it with a multilevel sentiment network in order to understand the relevant information
in an intuitive and semantic way. This semantic interpretation method minimizes network learning
in the system as a fixed network topology only exists as a guideline to help users understand.
Furthermore, it does not need to discover every single node to understand the characteristics of each
clustering within the network. After extracting and analyzing the sentiment words from the movie
review data, we designed a movie network based on the similarities between the words. The network
formed in this way will appear as a multilevel sentiment network visualization after the following
three steps: (1) design a heatmap visualization to effectively discover the main emotions on each
movie review; (2) create a two-dimensional multidimensional scaling (MDS) map of semantic word
data to facilitate semantic understanding of network and then fix the movie network topology on
the map; (3) create an asterism graphic with emotions to allow users to easily interpret node groups
with similar sentiment words. The research also presents a virtual scenario about how our network
visualization can be used as a movie recommendation system. We next evaluated our progress to
determine whether it would improve user cognition for multilevel analysis experience compared to
the existing network system. Results showed that our method provided improved user experience in
terms of cognition. Thus, it is appropriate as an alternative method for semantic understanding.
Keywords: collaborative schemes of sentiment analysis and sentiment systems; review data mining;
semantic networks; sentiment word analysis
1. Introduction
1.1. Background and Purpose of Research

At present, we are faced with an enormous amount of information every day due to the vast
growth of information and communications technology. Thus, there is increased interest in effective
data processing and analysis. In particular, big data is playing an increasingly important role since
it is suitable for refined and semantic processing, even if the amount of data is considerable or if its
structure is complex [1]. Big data has also attracted great attention in the field of data visualization
primarily for the design of efficient processing and semantic analysis. Data visualization is a redesigned
concept of data analysis with better readability, offering distinct insights that cannot be grasped from
a table or graph [2]. Network visualization is a visual tool to semantically analyze data if there is a

Appl. Sci. 2019, 9, 2419
massive amount or if the structure is complex [3]. Therefore, this study aims to demonstrate massive
collective intelligence data through network visualization. This study also proposes an intuitive
and semantic analysis method by designing multilevel sentiment network visualization based upon
emotion words [4]. Social network analysis plays a significant role in understanding and finding
solutions for society-functional problems by examining the original structure and relationships of the
network. Therefore, network visualization is applied in a wide variety of fields, including network
analysis based on data similarity, network analysis about social-scientific situations, graph theory, and
recommendation systems.
Force-directed graph drawing algorithm is a standard layout algorithm for designing a network
graph. It is considered to be highly useful since it allows related nodes to form a cluster [5]. However,
the location of the node varies each time the graph is formed if a force-directed graph drawing
algorithm is used because the entry value of the node’s location is random and the eventual position is
determined by relative connections between nodes. Therefore, users must repeat the learning of the
system if in a force-directed network, since the absolute location information is not fixed (Figure 1).
This is a notable drawback. Such difficulties can become great obstacles when interpreting the network
if it consists of a considerable amount of data. Furthermore, the collective intelligence can deliver the
wrong data if it is visualized based on a force-directed layout since locations of nodes may vary.
Figure 1. Force-directed layout network (left: 300 nodes; right: 400 nodes). Locations of nodes continue
to change whenever data is added or modified. Reproduced with permission from [6]; published by
[IEEE], 2015.
Following a preliminary study of ours [6–8] which introduced the sentiment network visualization
approaches that are the basis of this work, we designed a multilevel sentiment network visualization
so as to facilitate intuitive understanding of the complex collective intelligence data. Also, we present
an approach to find solutions for those difficulties of a force-directed layout. Specifically, the primary
contributions of our work are:
• More complete description of the sentiment movie network developing process: we present
the explanation of flock algorithm method. Even if nodes increase in large scale, with this
method, nodes won’t be overlapped making bigger sentiment network configuration possible
with improved delivery.
• Considering more diversity visual metaphorical graphic model, and conducting a survey about
selection of graphic to better support interaction with the sentiment movie network.
• An evaluation for comparing the awareness level on network location.
• An evaluation for comparing between its final visualization map (Network + MDS+ Constellation
Metaphor) and our previous visualization map [8].
6
Appl. Sci. 2019, 9, 2419
1.2. Research Process

After targeting “movie review data” among examples of collective intelligence, we selected 36
sentiment words that generally appeared in movie reviews. These were classified into seven clustering
characteristics through principal component analysis. These keywords were also analyzed through
multidimensional scaling (MDS) to discover any correlations, such as similarities and differences.
A two-dimensional distribution map was designed in the next step according to these correlations.
We then designed a sentiment-based movie similarity network based on this two-dimensional
distribution map through three steps. First, assuming that each node contains one piece of information
about a movie, edges were created to match with each other if the nodes shared the most similar
emotion characteristics. This eventually led to a network of 678 movie nodes. Heatmap visualization
was also applied to allow us to more easily grasp the frequency of sentiment words of each movie [9].
A two-dimensional distribution map with sentiment words related to the movie review was created
accordingly to distribute nodes in semantic positions. As each node was influenced by the semantic
point depending on its attribute value, absolute positions of nodes were designed to reflect attributes
of nodes [10]. Thus, nodes representing each movie formed a network layout which attracted nodes
to where sentiment words were spread on the two-dimensional distribution map according to its
frequency. The whole structure was referred to as the sentiment-movie network. Third, we applied a
constellation visualization to label the characteristics of each cluster when nodes in a network structure
showed clustering on a two-dimensional keyword distribution map. An asterism graphic consisted
of objects representing traits of each clustering formed by nodes and edges of a network clustering
structure. We next suggested a virtual scenario in the general public’s view in order to determine how
this network visualization could be applied in the actual movie recommendation process. Assuming
that a random user enjoyed a certain movie, we demonstrated the process which the user went through
in order to find a movie with similar emotions. Finally, three kinds of evaluations were conducted to
verify whether the visualization method that we proposed could be linked to cognition improvement
for users’ multilevel analysis experience. The first test was designed to verify whether users would
show a satisfactory understanding of the meaning of the location structure of nodes in a movie network
visualization. The next test compared the two groups provided with or without heatmap visualization
to see how well they could adopt sentiment words data when discovering the network. The last test
was designed to determine which visualization case worked the most effectively for subject groups to
conduct a semantic analysis among the following three cases: (1) the first visualization with network
nodes; (2) the second visualization that involved fusing the first visualization with the two-dimensional
sentiment word map indicating locations of the sentiments; and (3) the third visualization which
superimposed the constellation metaphor over the second visualization. Figure 2 illustrates three
processes for the sentiment analysis, data visualizing, and the evaluation workflow.
Figure 2. Research framework.
7
Appl. Sci. 2019, 9, 2419
2. Materials and Methods
2.1. Sentiment Analysis and Visualization

Sentiment word analysis is a fundamental process in numerous applications of opinion and
emotion mining, such as review crawling, data refining, and SNS classification [11]. In order to
contribute to sentiment analysis and data visualization research fields, there is an abundance of studies
that applied sentiment words and sentiment visualization system.
MyungKyu et al. [12] was used to deal with sentiment words shown in online postings and social
media blogs. JoungYeon et al. [13] was used to illustrate adjectives used to describe the texture of
haptic and indicated relations between adjectives on multi-dimensional scaling.
Ali et al. [14] provided a clear and logical taxonomy of sentiment analysis work. This taxonomy
was made up of two main subcategories in this field: opinion mining and emotion mining. Also,
they present a set of significant resources, including lexicons and datasets that researchers need for a
polarity classification task. This would allow them to study the sentiment of any topic and determine
the polarity of said topic, be it either positive or negative.
Kostiantyn et al. [15] suggest the state of the art sentiment visualization techniques and trends by
providing the 132 cases of “sentimentvis” interactive survey browser. In this research, all the cases are
classified based on the categorization standards. This categorization consists of the data domain, data
source, data properties, analytic tasks, visualization tasks, visual variable, and visual representation.
Also, the collected data indicates the growing multidisciplinary insight for visualization of sentiment
with regard to multiple data domains and tasks.
2.2. Movie Recommendation

Studies on movie recommendation methods have mainly focused on “content-based
recommendation systems utilizing information filtering technology” and “corporative recommendation
system.” According to Oard et al. [16], a content-based recommendation system was used to extract
the characteristics from individual information and its following preference.
Movie recommendation systems based on corporative filtering have been analyzed by
Sarwar et al. [17] and Li et al. [18]. They were used to recommend options selected by the group that
shared similar information with individual users.
While these previous studies made recommendations based on individual information of users, we
further managed user’s experience data as in “emotional review data felt during the movie watching,”
thus enriching emotional attributes that fit the purpose of movie recommendation.
Recently, Ziani et al. [19] suggest a recommendation algorithm based on sentiment analysis to
help users decide on products, restaurants, movies, and other services using online product reviews.
The main goal of this study is to combine both the recommendation system and sentiment analysis in
order to generate the most accurate recommendations for users.
However, this work has a limitation that did not present a user-centered recommendation system
and focused on developing the automatic recommendation algorithm based on semi-supervised
support vector machines (S3VM). Meanwhile, our system provides user-centered recommendation
experience through using the network visualization and metaphorical sentiment graphic, which is
easy to analyze.
2.3. Network Visualization and Layouts

A number of studies have been conducted on network visualization methods, including several
recent studies on user’s perception. For example, Cody Dunne et al. [20] have introduced a technique
called motif simplification, in which common patterns of nodes and links are replaced with compact
and meaningful glyphs, allowing users to analyze network visualization easily.
While this method identifies the maximal motif more accurately, even enabling the estimation of
size through glyph and interaction, it has several difficulties for ordinary users. For example, users
8
Appl. Sci. 2019, 9, 2419
must put considerable effort toward learning concepts of motif and interpreting glyph. In addition,
they have difficulty in discovering the optimal set of motifs.
Another study by Giorgio et al. [21] presented a tool called “Knot” with a focus on the analysis of
multi-dimensional and heterogeneous data to achieve interface design and information visualization
in a multidisciplinary research context.
Furthermore, Nathalie et al. [22] suggested methods to solve clustering ambiguity and increase
readability in network visualization. That paper states that major challenges facing social network
visualization and analysis include the lack of readability of resulting large graphs and the ambiguous
assignment of actors shared among multiple communities to a single community.
They proposed using actor duplication in social networks in order to assign actors to multiple
communities without substantially affecting the readability. Duplications significantly improve
community-related tasks but can interfere with other graph readability tasks.
Their research provided meaningful insights as to how central actors could bridge the community.
However, it also left confusions when distinguishing duplicated nodes and analyzing visualizations
that exceeded a certain size since node duplications could artificially distort visualization.
Gloria et al. [23] present a novel method that uses semantic network analysis as an efficient way
to analyze vaccine sentiment. This study enhanced understanding of the scope and variability of
attitudes and beliefs toward vaccination by using the Gephi network layout.
These four studies mentioned above all aimed to improve network visualization from the
perspective of users, with a particular focus on settling challenges of visualization distortion and
existing network through users’ learning of new technologies. However, network layouts used in
previous studies could not handle overlapping of nodes when faced with increasing amounts of data.
In addition, the user may have to repeat the system since the location may vary. Furthermore,
these network layouts were inefficient in that users had to check each node one by one in order to
identify characteristics of a cluster.
Our research is consistent with previous studies in that it will fix the network data based upon
sentiment words. It was designed to minimize users’ learning and prevent distorted interpretation by
applying a metaphor based on the characteristics of nodes in the network.
While previous studies made recommendations based on user’s individual information, we
further managed user’s experience data as in “emotional review data felt during the movie watching,”
thus enriching emotional attributes that fit the purpose of movie recommendation.
3. Sentiment Analysis Data Processing

Sentiment analysis is the field of study that analyzes people’s opinions, sentiment, attitudes,
evaluations, survey, and emotions towards entities such as issues, events, topics, and their attributes.
For sentiment analysis, we performed three data processes as follows.
3.1. Sentiment Words Collection

We selected 100 sentiment words filtered from 834 sentiment words based on the research
conducted by DougWoong et al. [24] in order to create a sentiment word distribution map. A further
survey of 30 subjects aged from 20 to 29 years determined the most frequently used words among
these 100 sentiment words. Following basic instruction on the concept of sentiment words during
movies, we investigated to what degree the emotion represented in each sentiment word could be
drawn from watching some movies.
The survey began with the question “Describe how much you feel as in each sentiment words after
watching the movies with the following genres, based on your previous experience.” The questionnaire
used a 7-point Likert Scale to evaluate responses ranging from “Strongly irrelevant” to “Strongly
relevant.” After eliminating 32 sentiment words that scored below the average, 68 sentiment words
were finally selected [9].
9
Appl. Sci. 2019, 9, 2419
3.2. Sentiment Words Refinement

In order to select the final sentiment words used for a two-dimensional distribution map among
68 sentiment words from the user survey, we collected and compared sentiment word data in existing
movie reviews, eliminating words that were rarely used. This procedure consisted of three phases as
shown below.
Crawling: Movie review data used in this research were collected from movie information
service NAVER [25], a web portal site with the largest number of users in Korea. We designed a web
crawler to automate sentiment word collection from movie reviews. This crawler covered three stages:
(1) collecting unrefined movie reviews and tags in the NAVER movie web page, (2) refining collected
data suitable for this research, and (3) extracting sentiment words based on analysis of refined data.
As a result, we obtained 4,107,605 reviews of 2289 movies from 2004 to 2015.
Establishing sentiment word dictionary: We divided text data into morphemes collected through
the crawling process using a mecab-ko-lucene-analyzer [26] and further extracted sentiment morphemes.
A total of 133 morpheme clusters were selected through several text mining processes. A morpheme is
the smallest grammatical unit in a language. In other words, it is the smallest meaningful unit of a
language [27].
Each emotion morpheme selected was classified according to 68 kinds of detailed sentiment word
categories. A sentiment word dictionary classified by the chosen sentiment word was then established.
Extracting emotion morphemes and classifying them by category were conducted in consultation with
Korean linguists.
To produce more accurate results, we eliminated less influential sentiment word clusters after
matching them with actual movie review data. We calculated Term (w) Frequency (tf : Term Frequency)
of each sentiment word cluster (t) suggested by the following formula.
j
t f (t, d) = Σ i=0 f (wi , d) (1)
j = number of words in sentimental group t

The number of times that term t occurs in document d
Then, inverse document frequency (idf ) was also drawn from this formula to lower the weight of
the general sentiment word group. The inverse document frequency (idf ) is “the logarithmically scaled
inverse fraction of the documents that contain the word” [28]. It reduces the number of terms with a
high incidence across documents.

N
id f (t, D) = log (2)
|{d ∈ D : t ∈ d}|
N = total number of documents in the corpus N = |D|

D = document set
TF-IDF score of sentiment word clusters on each movie was calculated using the following formula:
TFIDF(t, d, D) = t f (t, d) ∗ id f (t, D) (3)
We next considered the maximum TF-IDF score that might appear from each sentiment word in
order to decrease the number of sentiment words. For example, the word “Aghast” showed a TF-IDF
score of no more than 0.8% in every movie, whereas “Sweet” scored 42% for at least one movie. We
eliminated sentiment words with TF-IDF scores under 10%. Eventually, we selected 36 sentiment
words. These sentiment word clusters were broadly divided into “Happy,” “Surprise,” “Boring,”
“Sad,” “Anger,” “Disgust,” and “Fear” as shown in Table 1.
10
Appl. Sci. 2019, 9, 2419
Table 1. Final sentiment words. Reproduced with permission from [6]; published by [IEEE], 2015.
Clustering Characteristics Sentiment Words

Happy Happy, Sweet, Funny, Excited, Pleasant, Fantastic, Gratified, Enjoyable, Energetic
Surprise Surprised, Ecstatic, Awesome, Wonderful, Great, Touched, Impressed
Boring Calm, Drowsy, Bored
Sad Pitiful, Lonely, Mournful, Sad, Heartbroken, Unfortunate
Anger Outraged, Furious
Disgust Ominous, Cruel, Disgusted
Fear Scared, Chilly, Horrified, Terrified, Creepy, Fearsome
3.3. Movie Data Collection

Movie samples used in network visualization were also collected from NAVER movie service in
the same way as movie review data [25]. Based on 2289 movie samples from 2004 to 2015 registered
in the NAVER movie service, movies with more than 1000 emotion morphemes were used to filter
the emotion level. As a result, 678 movie samples were ultimately selected and used as network
sample data.
4. Visualization Proposal
We proposed three methods to solve problems of the existing social network visualization layouts
as follows.
4.1. Heatmap Visualization

Heatmap visualization is a method consisting of rectangular tiling shaded in scaled color [29,30].
It is used to search for anomalous patterns of data metrics as outlined in previous research [31]. To
comprehend convergences, the method described by Young-Sik et al. [32] was used. A recent study
proposed 100 cases of IT fractures, visualized and analyzed using the 2D parametric fracture probability
heatmap [33]. The proposed map projection technique can project the high dimensional proximal
femur fracture information onto a single 2D plane. Also, they applied heatmap in order to present the
level of risk.
In this research, we showed TF-IDF size frequency of sentiment words in a heatmap, utilizing
coordinate space in 2-dimensional distribution map of each sentiment word in order to visualize a
sentimental distribution graphic for each movie node. The detailed research process is shown below.
First, we measured space among the selected 36 sentiment words and analyzed correlations in order to
design a two-dimensional sentiment word distribution map. We then conducted multi-dimensional
scaling (MDS). We conducted a survey on the semantic distance among 36 sentiment words by enrolling
20 college students majoring in Digital Media Technology. These 36 sentiment words were placed
on both axes (36 × 36) and the distance between words was scored in a range of plus/minus 3 points
by considering their emotional distance. We used UNICET to facilitate a variety of network analysis
methods based on data obtained from the 20 survey participants [10]. We also created Metric MDS as
shown in Figure 3 based on the semantic distance among movie review semantic words. As a result,
positive emotions such as “Happy” and “Surprise” were distributed on the right side of the X-axis while
negative feelings such as “Anger” and “Disgust” were distributed on the left side. Active emotions
that were generally exaggerated gestures such as “Fear” and “Surprise” were distributed on the top of
the Y-axis while static emotions such as “Sad” and “Boring” were on the bottom. Furthermore, each
type of sentiment words was clustered concretely based on particular characteristics such as “Happy,”
“Surprise,” “Boring,” “Sad,” “Anger,” “Disgust,” and “Fear” on the two-dimensional distribution map.
Results showed that cluster characteristic “Surprise” could be divided into “Happy” and “Fear” clusters.
This implies that both overflowing joy and sudden fright are dominant in particular movies [10].
11
Appl. Sci. 2019, 9, 2419
Figure 3. 36 sentiment words multidimensional scaling (MDS) map. Reproduced with permission
from [6]; published by [IEEE], 2015.
Heatmap visualization was then designed based on a two-dimensional sentiment word distribution
map as well as frequencies of these 36 sentiment words, which consisted of a two-dimensional
distribution map. One optional movie was needed to design the heatmap. We measured frequencies
of sentiment words on each movie by contrasting sentiment words in the movie review data obtained
through the data construction process and sentiment words in the morphological dictionary. In addition,
we measured numerical values by calculating the TF-IDF score to lower the weight of particular
sentiment words that frequently emerged regardless of the typical characteristics of the movie.
Therefore, TF-IDF score on each sentiment word could be interpreted as a numerical value reflected
on the heat map visualization graph for target movies. The final heatmap graph consisted of a
two-dimensional distribution map with sentiment words and tiny triangular cells. Every cell was
initialized at numerical value, 0. The value then increased depending on the TF-IDF score of the
sentiment word located in the pertinent cell. As the numerical value of the cell increased, the color
of the cell changed, making it easier to discover the value of TF-IDF score of a pertinent sentiment
word. Furthermore, as high-valued cells influenced the values of surrounding cells, the heatmap graph
became gradually significant.
Figure 4b is a heatmap graph representing the distribution map of sentiment words from movie
reviews written by viewers about the movie “Snowpiercer.” This graph shows high frequencies of
emotions such as “Pitiful and boring” as well as “Funny and great.” Some reviews noted “Not so
impressive, below expectation. It was pitiful and boring.” “It was crueler than I thought, and it lasts
pretty long in my head.” and “The movie was pretty pitiful, probably because my expectations were
too high.” As shown in these reviews, there were various spectators with different emotions about this
movie, including disappointment.
12
Appl. Sci. 2019, 9, 2419
(a)
(b)
Figure 4. (a) Heatmap of “Don’t Cry Mommy” showing single emotions (Furious, Outraged).
(b) Heatmap of “Snowpiercer” that shows various emotions (Cruel, Pitiful, Lonely, Bored, Funny, Great,
and Energetic). Reproduced with permission from [6]; published by [IEEE], 2015.
Therefore, a heatmap showing movie sentiment words can be divided into two cases. Thus,
there are two types of heatmaps [7,8]. The first one is a type including movies with a single frequent
sentiment word. The table shown above presented a ranking of sentiment words in a movie called
“Don’t cry mommy” (Figure 4a), while the table shown below is its actual movie reviews. Since the
movie describes a sexual crime, emotions related to “Furious” and “Outraged” are prominent. Such
parts are shown in red in the heatmap. The second is a type with various frequent sentiment words.
For instance, the word frequency table and reviews on the movie “Snowpiercer” (Figure 4b) revealed
different sentiment words such as “Boring, Cruel, Funny, and Pitiful.” Thus, its heatmap suggests
patterns based on such information. Furthermore, heatmap visualization helps us easily understand
the types of emotions that viewers have experienced in a movie by reflecting the frequency of each
sentiment word in the movie.
4.2. Sentiment Movie Network

In this session, we aim to describe the basic structures of suggested graphs and examples. Locations
of nodes can be altered depending on the main sentiment word from the movie review. The suggested
graph is similar to the artifact actor network, a type of multi-layered social network. the artifact actor
network connects between the artifact network and social network using the semantic connection.
Thus, it expresses the semantic relation between two networks [34]. In our proposed graph, we
connected sentiment words on the 2-dimensional scaling map with movie network. In this paper, we
13
Appl. Sci. 2019, 9, 2419
referred to this network as the sentiment-movie network. Figure 5 shows the basic structure of the
sentiment movie network [6].
Figure 5. The multilevel structure of the sentiment movie network. Reproduced with permission
from [6]; published by [IEEE], 2015.
As shown in Figure 5, the suggested graph is comprised of two layers. The first layer is called the
semantic layer. It consists of semantic points based on the 36 sentiment words. The semantic point of
the sentiment word is located at an initially set value. It is immovable. The second layer is called the
network layer, which includes nodes comprising the movie network. Each movie node forms the edge
of other movie nodes based on similarities. It also forms imaginary edges with the sentiment word in
the two-dimensional distribution map based on the sentiment word connoted by the pertinent node.
Nodes connected by edges have both attractive force and repulsive forces based on a force-directed
algorithm. By contrast, semantic points of sentiment words are immovable, leaving only attractive
force. For edge composition between nodes, we calculated cosine similarity between movies based on
TF-IDF scores of the 36 sentiment words. The similarity between movie A and movie B or SIM (A, B) is
shown as follows. n A ∗B
i=0 i i
SIM(A, B) = (4)
n (A )2 ∗ n (B )2
i=0 i i=0 i
The edge between each node and the semantic point sets up a fixed threshold value and generates
an edge by designating a sentiment word with a value that is greater than a threshold value as semantic
feature. Figure 6a,b show an example that the location of a node on the graph can be altered depending
on the frequency of sentiment word indicated in the Heatmap Visualization [7,8].
Figure 6a shows that the node is located in the space of the sentiment word with overwhelmingly
high frequency. If a movie has more than two frequent sentiment words, its nodes will be located in the
way, as shown in Figure 6b. If shown in the heatmap visualization of Figure 6b, nodes in “Snowpiercer”
are influenced by keywords such as Pitiful, Bored, Energetic, Great, and Cruel. In addition, nodes
are particularly influenced by locations of “pitiful” and “bored” since those two are the strongest that
they affect nodes to be placed around the negative y-axis. As such, frequent sentiment words will
place nodes.
To show users a host of nodes in a meaningful position without overlapping, flock algorithm has
been applied. Flock algorithm consists of separation, alignment, and cohesion [35]. We composed
nodes applying separation method because nodes of developed visualization do not move and have
no direction.
Solution: One node compares all nodes to each other’s positions and moves in opposite directions
if overlaid.
14
Appl. Sci. 2019, 9, 2419
Algorithm 1. Separate nodes–avoid crowding neighbors

Input: nodes→Node objects to be displayed on the screen.
Output: Node objects whose locations have been recalculated so that they are distributed
without overlap.
Method
1: Initialize SEPARATION_DISTANCE
2: WHILE true
3: change = false
4: FOR src in nodes
5: FOR dst in nodes
6: IF src == dst
7: continue
8: ELSE
9: dist = src.position − dst.poisiton
10: IF dist.size < SEPARATION_DISTANCE
11: change = true
12: src.position += 1/dist * dist.direction
13: dst.position −= 1/dist * dist.direction
14: END IF
15: END IF
16: END FOR
17: END FOR
18: IF change == false
19: break
20: END WHILE
Even if nodes increase in large scale or similar nodes are added, with this method, nodes will not
be overlapped making bigger network configuration possible with improved delivery.
As every node connected by the network made of suggested methods is located in the graph,
clustering is formed by combining similar movies in the space of sentiment word with high frequency
considering connections between movies as well as connections between related sentiment word.
Figure 7 shows an extreme position of a node and a cluster [8]. This network allows users to easily
understand the network structure even if the number of nodes is changed by fixating the topology of
movie networks based on the sentiment word distribution map of movie reviews. Finally, k-means
clustering operation using cosine similarity value for classifying cluster characteristics of each node
was conducted. The number of clusters considered ranged from 9 to 12. The final cluster number was
chosen to be 11 as the node number of each cluster was evenly distributed and various characteristics
were well clustered. Furthermore, each node was colored to classify each node group based on these
11 clusters.
(a)
Figure 6. Cont.
15
Appl. Sci. 2019, 9, 2419
(b)
Figure 6. (a) Heatmap visualization and positioning on the sentiment-movie network. (one point
position) in the case of “Don’t Cry Mommy.” (b) Heatmap visualization and positioning on
the sentiment-movie network. (more than two point positions) in the case of “Snowpiercer.”
Reproduced with permission from [8]; published by [Journal of The Korea Society of Computer
and Information], 2016.
Figure 7. Sentiment movie network (678 Movie nodes). Reproduced with permission from [6];
published by [IEEE], 2015.
4.3. Constellation Visualization

This session facilitates a cognitive understanding of the process used to design constellation image
visualization based on specific nodes and edges with significant sentiment word frequency in order to
clarify semantic parts of each clustering. Metaphorical visualizations can map characteristics of certain
well-understood visual images/patterns to a more poorly understood data source so that aspects of the
target can be more understandable.
Recently, Information graphics and the field of information visualization apply a variety of
metaphorical devices to make large size, complex, abstract, or otherwise difficult-to-comprehend
information understandable in graphical terms [36,37].
We created an asterism graphic of each cluster network by considering the significant sentiment
words, information on movies, and synopses in each cluster. Our research makes it easy to understand
the sentiment data system to the general public by using the visual metaphor graphics.
In order to realize asterism images, we referred to the labeling data of the 11 different clusters
yielded from k-means clustering, the most dominant categories of sentiment words in each cluster,
and their following information on movies and synopsis. In order to select a pool of constellation
graphics, experts with academic backgrounds in graphic design and movie scenario professionals were
16
Appl. Sci. 2019, 9, 2419
consulted, ultimately narrowing down options to 2~3 graphics per sentiment cluster. Table 2 below
contains the lists of image options corresponding to each of these 11 clusters.
For the next phase, 30 students majoring in design and visual media were surveyed to choose
the final constellation graphics for representing sentiment words around each cluster. The survey
incorporated statements that described relations between sentiment words shown in clusters as well
as the choice of images in a 5-point Likert scale. For instance, when a statement of “An image of a
dynamite bears relation with the feeling of surprise” is presented, the test subject was supposed to
mark an answer from a range of “not at all (1 point) ~ very much (5 points)” and evaluate the strength
of the link between the sentiment word and the graphic. Table 2 presents the average drawn from
these 30 students’ data. The graphics with the highest average was chosen as the final constellation to
represent the relevant cluster.
Table 2. List of candidate graphics linked to a cluster of sentiment words.
Cluster Name List of Candidate Graphics

Cruel and dreadful Red pyramid (4.7), Piranha (4.1), Jigsaw (3.3)
Dramatic Emotional Comet (2.1), Ballerinas (3.5), Whale (4.1)
Dynamic mood change Chameleon (3.3), Persona mask (4.3)
Thrilling and horrifying Alien (3.8), Jaws (3.0)
Surprising A surprised face mask (1.7), Dynamite (3.4), Jack in the box (4.2)
Pleasing and exciting Firecracker (3.3), Gambit (3.7), Party hat (3.2)
Authentic fear Reaper (4.5), Scream Mask (4.1), Dracula (2.3)
Generally Monotonous Sloth (3.0), Snail (2.1), Yawner (2.3)
Fun and cheerful Wine Glass (4.3), Heart (3.2), A diamond ring (3.5)
Cute Gingerbread Cookie (3.5), Kitty (3.0)
Sad and touching Mermaid (4.2), Teardrop (2.8)
Table 3 below shows the main emotions and movie examples contained in each cluster as well
as motives for choosing each asterism name [6]. This constellation visualization helped us naturally
perceive characteristics of a node without having to individually review parts of the network.
Table 3. Definition of constellation visualization. Reproduced with permission from [6]; published by
[IEEE], 2015.
Cluster Name Movie Examples Asterism Name Motives for Each Name
Final Destination 3, Symbolized the cruelly murdering character in
Cruel and dreadful Red pyramid
Piranha 3D a movie <Silent Hill>
Inspired from the scene when grampus appears
Pride & Prejudice, The
Dramatic Emotional Whale in a movie <Life of Pi>, which aroused
Notebook
dramatic and emotional image simultaneously
Persona masks are supposed to express various
Snowpiercer,
Dynamic mood change Persona mask emotions, which is similar to movies with
Transformers
dynamic mood changes
Resident Evil, War Of Aliens arouse fear and suspense in
Thrilling and horrifying Alien
The Worlds unrealistic situations
Symbolized an object popping out of the box to
Surprising Saw, A Perfect Getaway Jack in the Box
express surprising moments
Relevant to the magician character of a movie
Pleasing and exciting Iron Man, Avatar Gambit
<X-men>, who is fun and surprising
Paranormal Activity, The Symbolized as a reaper to show the authentic
Authentic fear Reaper
Conjuring and intrinsic fear
Originated from the idea that sloths are boring
Generally Monotonous 127 Hours, Changeling Sloth
and mundane
Hairspray, The Spy: Wine glass is a symbol of romantic and
Fun and cheerful Wine Glass
Undercover Operation festive atmosphere
Despicable Me, Gingerbread men cookies represent cute and
Cute Gingerbread Cookie
Puss In Boots sweet sensations
Million Dollar Baby, The story of little mermaid shows touching,
Sad and touching Mermaid
Man on fire magical and sad ambience at the same time
17
Appl. Sci. 2019, 9, 2419
Constellation visualization helped us naturally perceive characteristics of a node without having

to review parts of the network individually. A comprehensive network map based on information
presented in this table is shown in Figure 8. We call this movie data network based on three proposals
“CosMovis.” Figure 8 also indicates that it is substantially easier to semantically analyze network
visualization using overlapping asterism images on each sentiment word and symbolic nodes with
connection structure of edges.
Figure 8. Comprehensive “CosMovis” constellation map of sentiment word-based movies. (Demo Site
URL: https://hyunwoo.io/cosmovis/). Reproduced with permission from [7]; published by [IEEE], 2014.
5. CosMovis Network Analysis Scenario

Our research fixed the network data based upon sentiment words MDS map and was designed
to minimize users’ learning and prevent distorted interpretation, by applying a metaphor based on
the sentiment keyword characteristics of the nodes in the network. In order to show that the user’s
target can be applied widely, this section presents an ideal scenario about how our sentiment network
visualization is used as a movie recommendation and how it affects a general public’s decision-making
process [8]. The scenario is shown as follows.
Scenario: Getting a movie recommendation from the cluster with similar emotions based on
previous movies that the general public has watched previously.
This scenario describes a situation of obtaining a movie recommendation from clusters that share
similar emotions based on sentiment words implied by users’ previous movie preferences. Assume
that the user watched “Star Trek” and was satisfied with the emotions provided by the movie. The
user will then search for a new movie from the network. He/she will first find nodes of “Star Trek” in
the network and then look for movies closely located to “Star Trek.” As a result, the user will be able to
find three other movie nodes based on “Star Trek” as shown in red in Figure 9.
Next, the user will focus on the heatmap to understand what kinds of emotions those three movies
contain. Figure 10 on the right shows heatmaps of movie nodes A, B, and C as well as their sentiment
word distribution table. Node A is for the movie “Avatar” and Node B is for the movie “Ironman 3.”
One can see that heatmaps of these two nodes are very similar to “Star Trek.” Thus, if the user wants to
have a similar emotional experience, one could watch Avatar or Ironman 3. Node C is for the movie
“Masquerade.” According to its heatmap, this movie also contains “touching” aspects in addition to
other emotions present in “Star Trek.” Therefore, if one wants a movie more emotional than “Star Trek,”
the user could pick this movie.
18
Appl. Sci. 2019, 9, 2419
We can also note that node C is for a historical genre movie, “Masquerade,” unlike “Star Trek,”
“Avatar,” or “Ironman 3”. This implies that sentiment-based movie similarity network can also
recommend movies in different genres as long as they contain similar sentiments. Likewise, users can
make more effective decisions if they want to obtain movie recommendations based on movies they
have watched by understanding the network structure, selecting candidate movies among ones similar
to his/her previously enjoyed movies, and analyzing sentiment word frequency of each candidate node.
Figure 9. (upper) Movie “Star Trek” Poster and TF-IDF size frequency of sentiment words. (middle)
“Star Trek” heatmap visualization. (lower) After selecting the movie “Star Trek (blue node),” three
movies A, B, and C (red nodes) that are located closely within the same cluster are discovered.
Reproduced with permission from [8]; published by [Journal of The Korea Society of Computer and
Information], 2016.
19
Appl. Sci. 2019, 9, 2419
Figure 10. (upper) Movie “Avatar” TF-IDF size frequency of sentiment words and heatmap visualization.
(middle) Movie “Ironman3,” (lower) movie “Masquerade.”
6. Evaluation
The scenario is as follows. Using an experimental method, we will explore whether the network
visualization put forth by this research can improve user cognition. For the purpose of verification,
three evaluations were conducted as follows. All evaluations were designed based on social science
and statistics in accordance with the International Review Board regulations.
6.1. Evaluation 1: User Awareness Level Test on Network Location

Purpose and Method of Cognition Evaluation: To gauge the users’ awareness level of the
network structure created through this research, an evaluation involving 40 participants was designed.
The participants were divided into two groups of 20, with one group given a simple explanation
regarding the visualization for two minutes while the other group was provided with a deeper
20
Appl. Sci. 2019, 9, 2419
explanation for five minutes. After a single movie node was selected from the network and the
corresponding heatmap shown, participants were asked to perform a series of tasks such as “Choose
the heatmap located adjacent to (or across from) the selected node.” The movie network was verified
through three axes, as illustrated below (Figure 11). The survey contained a total of 20 questions,
including five questions per axis along with an additional five questions pertaining to the node in the
center of the network.
Figure 11. Three axis and sentiment word directional guide for verifying the network structure.
Comparison of group average: For the purpose of average comparison, an independent t-test
was conducted. An independent t-test was also used to verify whether a difference in population
means existed between two groups that followed normal distribution N (μ1 , σ21 ) and N (μ2 , σ22 ) The two
groups were assumed to be independent of each other [38]. For this research, in order to discover the
difference in awareness levels between the two groups given either a simple or a detailed explanation
on the network structure, a survey asked participants to find the heatmap that represented a particular
node on the visualization map. We recruited 40 participants (20 in each group) who were majoring
in related fields with an understanding of visualization. We also assumed a normal distribution in
accordance with the central limit theorem. Three results on these two groups regarding network
structure are presented in Table 4.
Table 4. Results of independent t-test analysis.
Equal Variance P-Value/Alternative

Question P-Value T-Value P-Value
Assumption Hypothesis Adoption
b_A~a_A 0.01079 ** Heteroscedasticity ** −2.1963 0.03612 ** Adopt **
b_B~a_B 0.8118 Equal variance −2.5591 0.01461 ** Adopt **
b_C~a_C 0.3576 Equal variance −1.9321 0.06117 * Adopt *
b_M~b_M 0.6122 Equal variance −1.3635 0.1809 Dismissal
* 90% Confidence level, ** 95% Confidence level (A: A-axis/B: B-axis/C: C-axis/M: Middle area)/ (a: 2 min—Group/
b: 5 min—Group)/ (Left~Right: Compare Left and Right).
Independent t-test results between the more-informed group and the less-informed group showed
that P-values for questions in the three axes (excluding the middle area) were smaller than the
alpha-value (0.05), which was the critical value of a 95% confidence level, or (0.1), which was the
critical value of a 90% confidence level. This indicated a significant difference between the two groups
21
Appl. Sci. 2019, 9, 2419
that were more-instructed and less-instructed about the network. Table 5 and Figure 12 present details
regarding the two groups compared through independent t-tests.
Table 5. Details regarding the two test groups.
Question 95% Confidence Before After

−1.351534< μ
b_A~a_A 3.9 4.6
<−0.048465
−1.612026< μ
b_B~a_B 3.4 4.3
<−0.187974
−1.741933< μ
b_C~a_C 2.5 3.35
<−0.041933
b_M~a_M −1.242694< μ <0.242694 2.15 2.65
Figure 12. Comparison of correct answer rate between test groups.
Details on the two test groups indicated that, in terms of the rate of providing correct answers
for questions on all axes, the well-informed group outperformed the less-informed group in terms
of network structure. For the A-axis, after being provided a further explanation about the network
structure, the percentage of getting correct answers increased by 0.7. Similarly, for B-Axis, C-Axis,
and center node, percentages of getting correct answers increased by 0.9, 0.85, and 0.5, respectively.
Accuracy improvement was most pronounced for the A-Axis, whereas questions regarding the center
node produced the lowest enhancement in performance. Thus, we can infer that explaining the
network structure in advance of relevant visualization use can help raise the level of user awareness
on visualization. To further boost awareness regarding the center node, along with an explanation
on the network structure, a supplementary method may be required. Nevertheless, with the correct
answer rate for the less-informed group hovering around 50% on average, we can conclude that user
awareness of visualization is considered high in general.
6.2. Evaluation 2: Test on Usability Impact of the Heatmap Function

Purpose and Method of Usability Measurement Experiment: The first part of this research
concluded that explaining the network structure prior to visualization usage could enhance user
awareness [8]. To further improve the awareness level on center nodes, an additional prop up measure
was required. As one possible method for boosting user awareness on center nodes, this study
introduced a new function, through which dragging the mouse over to a particular center node would
offer heatmap visualization. The second test was thus designed to confirm whether a difference in
22
Appl. Sci. 2019, 9, 2419
visualization usability existed between group 1 (provided with the heatmap function) and group
2 (not provided such a function). The usability was measured through a survey comprised of five
questionnaires regarding learnability, efficiency, understandability, feature functionality, and accuracy
measured with a Likert scale of 7. Forty college students (two groups of 20) who were currently
studying data visualization with knowledge of the visualization field were selected as samples.
Reliability Analysis: Before analyzing the data gathered, a reliability analysis of the above survey
was carried out in order to verify its reliability. Reliability analysis measures internal consistency
among questions/statements employed in the survey based on Cronbach’s α. Cronbach’s α has a value
of between 0 and 1. The closer the value is to 1, the higher the question’s reliability. Generally, a
Cronbach’s α of 0.6 or higher is considered to be reliable. This single index can be used for all questions
to yield a comprehensive analysis. Results of reliability analysis performed on the final data set are
presented in Table 6.
Table 6. Results of reliability analysis. Reproduced with permission from [8]; published by [Journal of
The Korea Society of Computer and Information], 2016.
Cronbach’s α Cronbach’s α Total

Categories Statements
(Provide) (Non-Provide) Cronbach’s α
1. It is easy to select a movie based on
Learnability 0.698 0.827
the sentiment words.
2. It is efficient to select the node
Efficiency 0.698 0.826 0.666
based on the sentiment of the movie.
3. It is easy to understand the
Understandability sentiment distribution depending on 0.661 0.747
varying node locations.
4. It provides an adequate function to
Feature
help user choose a movie based on 0.663 0.749
Functionality
certain sentiment words.
5. The selected movie and the
Accuracy sentiment distribution predicted from 0.742 0.725
the movie’s map coordinate matches.
Comparing maximal values of Cronbach’s α after eliminating a certain item indicated that the
number was the highest (0.742) after eliminating Accuracy if the heatmap was provided while it had
the maximal value of 0.827 after eliminating Learnability if the heatmap was not provided. In addition,
the Cronbach’s α value exceeded 0.6, suggesting the absence of unreliability per each statement as well
as a high level of internal consistency across the survey. Thus, it can be concluded that the survey is
highly reliable.
Average Comparison per Group: In order to compare averages, an independent t-test was
conducted. An independent t-test was also used to verify whether there was a difference in population
means between two groups that followed a normal distribution N (μ1 , σ21 ) and N (μ2 , σ22 ). The two
groups are assumed to be independent of each other [39]. This research, through further data distillation,
examined 30 sets of survey data in accordance with the central limit theorem. The two groups were
assumed to follow a normal distribution and be independent of each other in comparing averages.
Results of average comparison analysis of the two groups are shown in Table 7.
23
Appl. Sci. 2019, 9, 2419
Table 7. Results of the independent t-test analysis. Reproduced with permission from [8]; published by
[Journal of The Korea Society of Computer and Information], 2016.
Equal Variance P-Value/Alternative

Question P-Value T-Value P-Value
Assumption Hypothesis Adoption
1_1~1_2 0.08203 Heteroscedasticity 4.8295 0.00003 ** Adopt **
5_1~5_2 0.0026 Equal variance 4.9205 0.000036 ** Adopt *
* 90% Confidence level, ** 95% Confidence level (a: Heatmap group/b: No heatmap group)/ (Left~Right: Compare
Left and Right).
After conducting an independent t-test on the two groups to which heatmap function was either
available or not available, it revealed significant differences between the two. Table 8 below presents
details of statements along with the alternative hypothesis.
Table 8. Details of statements with alternative hypothesis. Reproduced with permission from [8];
published by [Journal of The Korea Society of Computer and Information], 2016.
Question 95% Confidence Provide Non-Provide

1_1~1_2 1.128555< μ <2.771445 5.45 3.5
2_1~2_2 1.86879< μ <3.33121 5.95 3.35
3_1~3_2 1.118809< μ <2.781191 5.8 3.85
4_1~4_2 0.9908482< μ <2.7091518 5.65 3.8
5_1~5_2 1.254113< μ <3.045887 5.75 3.6
(Left~Right: Compare Left and Right).
For all statements that rejected the null hypothesis (providing heatmap function does not produce
a significant difference between the two groups) and adopted the alternative hypothesis (providing
heatmap function produces a significant difference between the two groups), it was confirmed that
the average value of the group provided with the heatmap function was higher than that of the
control group. Thus, we could conclude that a valid difference existed between the two groups based
on whether or not heatmap function was provided. We could further conclude that providing the
heatmap function reinforced the usability of the visualization map. It was also effective in enhancing
the awareness level of middle area nodes, which required an additional measure to be understood
better based on the first part of this research.
6.3. Evaluation 3: Reaction Time Test on Adding Constellation Visualization Function

Purpose and Method of Measuring Reaction Time: This research proceeds with a comparison
between its final visualization map and previous visualization maps [6–8]. The final visualization in
this study was formed by combining clustered (colored) nodes based on sentiment words, the sentiment
word MDS map indicating locations of sentiments, and the constellation metaphor applied to sentiment
words. This comparison test was designed to determine whether there was any difference in reaction
time spent when selecting the node for each visualization. Visualizations applied in this test consisted
of three different types: visualizations 1 to 3. Table 9, Table 10 and Figure 13 present further details of
these types. Test subjects were divided into three groups, one for each version of the visualization. In
this study, visualization scenes were divided into four quadrants. Subjects were given 30 s to observe
each quadrant of the visualization. Thus, the total observation time was 120 s. Meanwhile, subjects
were allowed to discover the content and structure of “CosMovis” freely. They were asked questions
(e.g., on which part of the map is the movie cluster that stirs touching and sad feelings located?)
on complex sentiments included in six polygons. Reaction time for recognizing the location of the
complex sentiment was then measured. College students currently studying data visualization with
24
Appl. Sci. 2019, 9, 2419
knowledge of the visualization field were selected as samples. A total of 45 participants were tested
(15 participants for each visualization group) [40].
Table 9. The basic structure of visualization (1 to 3).
Visualization Type Basic Structure

Visualization 1 (Network) Visualization on clustered (colored) nodes from sentiment words
Combined visualization on clustered nodes and MDS map
Visualization 2 (Network + MDS)
which shows the location of sentiments
Combined visualization on clustered nodes, MDS map which
Visualization 3 (Network + MDS +
shows the location of sentiments, and constellation metaphors
Constellation Metaphor)
based on sentiment words
Table 10. Polygon-related questions.
Polygon Name
Polygon Equal Variance Assumption
(Sentiment Words)
1. Where are the movies that stir up touching

Mermaid (Sad, Touching)
and sad feelings clustered around?
2. Where are the movies that arouse fun,

Wine Glass (Funny, Sweet,
sweet and pleasing emotions
Cheerful)
clustered around?
3. Where are the movies that trigger feelings

Jack in the box (Surprising)
of surprise clustered around?
4. Where are the movies that induce cruel and

Red-Pyramid (Cruel, Dreadful)
dreadful emotions clustered around?
5. Where are the movies that set off genuine

Reaper (Authentic fear)
fear clustered around?
6. Where are the movies that elicit dramatic

Whale (Dramatic Emotional)
sentiments clustered around?
Figure 13. Cont.
25
Appl. Sci. 2019, 9, 2419
Figure 13. (upper) First visualization: network; (middle) second visualization: network + MDS (lower)
third visualization: network + MDS + constellation metaphor.
ANOVA per group: To analyze the variance of data pertaining to multiple groups, this study
conducted two-way ANOVA. Two-way ANOVA can be used when a data set is comprised of two
independent variables and one dependent variable in order to verify the existence of a significant
difference in variance among different groups caused by changes in an independent variable [41]. This
study collected data through a closed testing method in which a series of 1:1 individual tests were
undertaken. We used questionnaires and measured reaction times of the three groups that were given
different versions of visualization for accurately locating the complex sentiment suggested by the six
polygons. Results of two-way ANOVA are presented in Table 11.
Table 11. Results of the two-way ANOVA.
DF Sum Sq Mean Sq F-Value Pr(>F)

Group 1 618.5 618.5 278.387 2 × 10−16 **
Question 1 83.8 83.8 37.74 2.16 × 10−9 **
Question: Group 1 0.1 0.1 0.056 0.814
Residuals 356 790.9 2.2
* 90% Confidence level, ** 95% Confidence level.
We reviewed the difference in reaction time among groups given different versions of visualization
maps and six kinds of polygon questions. The visualization map-based test result revealed an F-Value
of 278.387 and a P-Value of 2 × 10−16 lower than alpha value of 0.05, leading us to discard the null
hypothesis (no difference exists in reaction time to different polygon questions among groups) and
adopt the alternative hypothesis (a difference exists in reaction time to different polygon questions
among groups). In the polygon-based test, two-way ANOVA yielded an F-Value of 37.74 and a
P-Value of 2.16 × 10−9 , lower than the threshold alpha value of 0.05. Therefore, we replaced the null
hypothesis (no difference exists in reaction time among groups given different versions of polygon-based
visualization) with the alternative hypothesis (a difference exists in reaction time among groups given
different versions of polygon-based visualization). Finally, we analyzed the difference in reaction
time among nested groups with different polygon-visualization and obtained an F-Value of 0.056 and
a P-Value of 0.814, both of which were larger than the alpha value of 0.05. Thus, we accepted the
null hypothesis (a difference exists in reaction time among nested groups given different versions of
polygon-visualization), forgoing the alternative hypothesis (no difference exists in reaction time among
nested groups given different versions of polygon-visualization).
Based on ANOVA test results, we concluded that a significant difference satisfying the 95%
confidence level existed among the groups provided with different versions of the visualization and
among different questions of the test. In order to further delve into the makeup of such differences, a
post-hoc analysis was conducted using a Boxplot. Results are shown in Figure 14.
The post-hoc analysis using boxplot revealed that the first group’s (given the first version of the
visualization) reaction time in solving questions to find complex sentiment clusters was the longest.
26
Appl. Sci. 2019, 9, 2419
They spent an average of 12.01 s to complete the task. The second group (given the second version of
the visualization) took 10.41 s on average to locate the sentiment cluster, showing a reduction of 1.6 s
compared to the first group. The third group (given the third version of the visualization) had the
shortest reaction time in problem-solving recording, with an average of 8.8 s. Based on these results, it
can be concluded that using the final visualization form which combines clustered (colored) nodes,
the sentiment word MDS map, and the constellation metaphor can help users better understand the
visualization relative to the first two versions while minimizing the reaction time, thereby facilitating
the use of visualization.
Figure 14. Post-hoc analysis for the question*group.
7. Conclusions
This study proposed three different methods for intuitive and semantic analysis based on the
multilevel sentiment network visualization that we created from movie review data to serve as collective
intelligence. These three methods were: (1) heatmap visualization that indicated semantic word
characteristics of each node; (2) a method that described network nodes based upon a two-dimensional
sentiment word map; and (3) a method using an asterism graphic for semantic interpretation of
clustering followed by an evaluation to verify the suggestions.
Our system presents the insight which makes it easy to understand for users by using the three
methods mentioned above.
The first evaluation revealed that participants understood the relations between locations of the
nodes and the heatmap relatively well. However, their levels of awareness dropped significantly
when considering nodes in the middle area of the network, which was the lowest even when subjects
had been informed of relevant content. From the second evaluation, we concluded that providing
heatmap visualization reinforced the understanding of emotions in each movie since it delivered
emotion characteristics. The final evaluation led to the conclusion that the emotion word guideline of
27
Appl. Sci. 2019, 9, 2419
nodes, as well as the asterism metaphor, allowed users to understand the semantic network better than
being exposed to the network solely with nodes.
Results of these three methods signify that a heatmap visualization is a useful tool for identifying
subtle differences in sentiment word distribution among nodes placed in adjacent locations. This
sentiment-movie network allows users to promptly understand the characteristics of the nodes since it
assigns nodes based on sentiment words of each movie. Moreover, a two-dimensional distribution
map showing sentiment words facilitates the understanding of the main emotion of movie nodes.
Likewise, it is expected that these three methods can help users understand semantic and effective
multilevel sentiment analysis in different ways.
Our results also imply that the general public could efficiently select a movie by proposing a
virtual scenario for obtaining a movie recommendation using our network visualization, heatmap,
and asterism graphic. This indicates that “CosMovis” could be adopted to create a novel movie
recommendation algorithm that provides new contents with emotional patterns. However, this would
only be possible after further improvements regarding nodes in the middle area since its awareness
level needs to be enhanced according to our second evaluation.
Our research can also be applied to any numerical sentiment data (or other factors that might
work as collective intelligence) as long as it can be formed within a network structure in addition to
movie review data which has been covered so far. For instance, analyzing a trending issue in Twitter
can help us deduce a direct and intuitive outcome if a certain issue is placed as a node in the network
such as “CosMovis” by examining relations between the emotion pattern of a certain topic and those
of other topics. Future research is needed to analyze data in ontology structure as well as sentiment
analysis so that multilevel semantic visualization can be adopted in order to better clarify the criteria
or meanings of the ontology structure data. In addition, we will conduct additional evaluation to
select visual metaphorical graphics for the general public. Through this evaluation, we will select
appropriate graphics which can make the general public empathize with sentiment information. Also,
we will perform the three evaluations that dealt with at session 6 again in order to prove the user
effectiveness considering the general public.
Author Contributions: Conceptualization, H.H.; Methodology, H.H., J.L. and K.L.; Software, H.H.; Validation,
S.M.; Formal Analysis, H.H. and J.L.; Investigation, S.M.; Data Curation, H.H.; Writing—Original Draft Preparation,
H.H.; Writing—Review & Editing, H.H., S.B. and K.L.; Visualization, H.H., H.H. and S.B.; Supervision, K.L.
Funding: This research was funded by [the Ministry of Education of the Republic of Korea and the National
Research Foundation of Korea] grant number [NRF-2018S1A5B6075104] And [Brain Korea 21 Plus Digital Therapy
Research Team] grant number [NRF31Z20130012946].
References
1. Hwansoo, L.; Dongwon, L.; Hangjung, Z. Personal Information Overload and User Resistance in the Big
Data Age. J. Intell. Inf. Syst. 2013, 19, 125–139.
2. Fayyad, U.M.; Wierse, A.; Grinstein, G.G. Information Visualization in Data Mining and Knowledge Discovery;
Morgan Kaufmann: Burlington, MA, USA, 2002.
3. Shneiderman, B.; Aris, A. Network Visualization by Semantic Substrates. IEEE Trans. Vis. Comput. Graph.
2006, 12, 733–740. [CrossRef] [PubMed]
4. Hao, M.; Rohrdantz, C.; Janetzko, H.; Dayal, U.; Keim, D.; Haug, L.E.; Hsu, M.C. Visual sentiment analysis
on twitter data streams. In Proceedings of the Visual Analytics Science and Technology (VAST), Providence,
RI, USA, 23–28 October 2011.
5. Thomas, M.J.; Edward, M. Graph Drawing by Force-direct Placement. Softw. Pract. Exp. 1991, 21, 1129–1164.
6. Hyoji, H.; Wonjoo, H.; Sungyun, B.; Hanmin, C.; Hyunwoo, H.; Gi-nam, K.; Kyungwon, L. CosMovis:
Semantic Network Visualization by Using Sentiment Words of Movie Review Data. In Proceedings of the
19th International Conference on Information Visualisation (IV 2015), Barcelona, Spain, 21 July–24 July 2015.
28
Appl. Sci. 2019, 9, 2419
7. Hyoji, H.; Gi-nam, K.; Wonjoo, H.; Hanmin, C.; Kyungwon, L. CosMovis: Analyzing semantic network of
sentiment words in movie reviews. In Proceedings of the IEEE 4th Symposium on Large Data Analysis and
Visualization (LDAV 2014), Paris, France, 9–10 November 2014.
8. Hyoji, H.; Hyunwoo, H.; Seongmin, M.S.; Sungyun, B.; Jihye, L.; Kyungwon, L. Visualization of movie
recommendation system using the sentimental vocabulary distribution map. J. Korea Soc. Comput. Inf. 2016,
21, 19–29.
9. Doi, K.; Park, H.; Junekyu, S.; Sunyeong, P.; Kyungwon, L. Visualization of Movie Recommender System
using Distribution Maps. In Proceedings of the IEEE Pacific Visualization Symposium (PacificVis 2012),
Songdo, Korea, 28 February–2 March 2012.
10. Hyoji, H.; Gi-nam, K.; Kyungwon, L. A Study on Analysis of Sentiment Words in Movie Reviews and the
Situation of Watching Movies. Soc. Des. Converg. 2013, 43, 17–32.
11. Deng, Z.H.; Yu, H.; Yang, Y. Identifying sentiment words using an optimization model with l1 regularization.
In Thirtieth AAAI Conference on Artificial Intelligence; AAAI Press: Menlo Park, CA, USA, 2016.
12. MyungKyu, K.; JungHo, K.; MyungHoon, C.; Soo-Hoan, C. An Emotion Scanning System on Text Documents.
Korean J. Sci. Emot. 2009, 12, 433–442.
13. JoungYeon, S.; KwangSu, C. The Perceived Lexical Space for Haptic Adjective based on Visual Texture
aroused form Need for Touch. Soc. Des. Converg. 2013, 38, 117–128.
14. Yadollahi, A.; Shahraki, A.G.; Zaiane, O.R. Current state of text sentiment analysis from opinion to emotion
mining. ACM Comput. Surv. CSUR 2017, 50, 25. [CrossRef]
15. Kucher, K.; Paradis, C.; Kerren, A. The state of the art in sentiment visualization. Comput. Graph. Forum 2018,
37, 71–96. [CrossRef]
16. Oard, D.W.; Marchionini, G. A Conceptual Framework for Text Filtering; DRUM: College Park, MD, USA, 1996;
pp. 1–6.
17. Sarwar, B.; Karypis, G.; Konstan, J.; Riedl, J. Item-based Collaborative Filtering Recommendation Algorithms.
In Proceedings of the 10th International World Wide Web Conference, HongKong, China, 1–5 May 2001;
pp. 285–295.
18. Li, P.; Yamada, S. A Movie Recommender System Based on Inductive Learning. In Proceedings of the 2004
IEEE Conference on Cybernetics and Intelligent Systems, Singapore, 1–3 December 2004; pp. 318–323.
19. Ziani, A.; Azizi, N.; Schwab, D.; Aldwairi, M.; Chekkai, N.; Zenakhra, D.; Cheriguene, S. Recommender
System Through Sentiment Analysis. In Proceedings of the 2nd International Conference on Automatic
Control, Telecommunications and Signals, Annaba, Algeria, 11–12 December 2017.
20. Cody, D.; Ben, S. Motif simplification: Improving network visualization readability with fan, connector, and
clique glyphs. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI ’13); ACM:
New York, NY, USA, 2013; pp. 3247–3256.
21. Uboldi, G.; Caviglia, G.; Coleman, N.; Heymann, S.; Mantegari, G.; Ciuccarelli, P. Knot: an interface for the
study of social networks in the humanities. In Proceedings of the Biannual Conference of the Italian Chapter of
SIGCHI (CHItaly ’13); ACM: New York, NY, USA, 2013; Volume 15, pp. 1–9.
22. Henry, N.; Bezerianos, A.; Fekete, J. Improving the Readability of Clustered Social Networks using Node
Duplication. IEEE Vis. Comput. Graph. 2008, 14, 1317–1324. [CrossRef] [PubMed]
23. Kang, G.J.; Ewing-Nelson, S.R.; Mackey, L.; Schlitt, J.T.; Marathe, A.; Abbas, K.M.; Swarup, S. Semantic
network analysis of vaccine sentiment in online social media. Vaccine 2017, 35, 3621–3638. [CrossRef]
[PubMed]
24. DougWoong, H.; HyeJa, K. Appropriateness and Frequency of Emotion Terms in Korea. Korean J. Psychol. Gen.
2000, 19, 78–98.
25. NAVER Movie. Available online: https://movie.naver.com/ (accessed on 1 February 2019).
26. Mecab-ko-lucene-analyzer. Available online: http://eunjeon.blogspot.kr (accessed on 15 February 2018).
27. Morpheme. Available online: https://en.wikipedia.org/wiki/Morpheme (accessed on 15 February 2018).
28. Srinivasa-Desikan, B. Natural Language Processing and Computational Linguistics: A Practical Guide to Text
Analysis with Python, Gensim, SpaCy, and Keras; Packt Publishing Ltd.: Birmingham, UK, 2018.
29. Wilkinson, L.; Frendly, M. The History of the Cluster Heat Map. Am. Stat. 2009, 63, 179–184. [CrossRef]
30. Van Eck, N.J.; Waltman, L.; Den Berg, J.; Kaymak, U. Visualizing the computational intelligence field.
IEEE Comput. Intell. Mag. 2006, 1, 6–10. [CrossRef]
29
Appl. Sci. 2019, 9, 2419
31. Robert, G.; Nick, G.; Rose, K.; Emre, S.; Awalin, S.; Cody, D.; Ben, S. Meirav Taieb-Maimon NetVisia: Heat
Map, Matrix Visualization of Dynamic Social Network Statistics & Content. In Proceedings of the Third IEEE
International Conference on Social Computing (the SocialCom 2011), Boston, MA, USA, 9–11 October 2011.
32. Young-Sik, J.; Chung, Y.J.; Jae Hyo, P. Visualisation of efficiency coverage and energy consumption of sensors
in wireless sensor networks using heat map. IET Commun. 2011, 5, 1129–1137.
33. Fu, Y.; Liu, R.; Liu, Y.; Lu, J. Intertrochanteric fracture visualization and analysis using a map projection
technique. Med. Biol. Eng. Comput. 2019, 57, 633–642. [CrossRef]
34. Reinhardt, W.; Moi, M.; Varlem, T. Artefact-Actor-Networks as tie between social networks and artefact
networks. In Proceedings of the 5th International Conference on Collaborative Computing: Networking,
Applications and Worksharing, Washington, DC, USA, 11–14 November 2009.
35. Flocking Algorithms. Available online: https://en.wikipedia.org/wiki/Flocking_(behavior) (accessed on
15 February 2018).
36. Risch, J.S. On the role of metaphor in information visualization. arXiv 2008, arXiv:0809.0884.
37. Hiniker, A.; Hong, S.; Kim, Y.S.; Chen, N.C.; West, J.D.; Aragon, C. Toward the operationalization of visual
metaphor. J. Assoc. Inf. Sci. Technol. 2017, 68, 2338–2349. [CrossRef]
38. Wiebe, K.L.; Bortolotti, G.R. Variation in carotenoid-based color in northern flickers in a hybrid zone. Wilson
J. Ornithol. 2002, 114, 393–401. [CrossRef]
39. Edgell, S.E.; Stephen, E.; Noon, S.M.; Sheila, M. Effect of violation of normality on the t test of the correlation
coefficient. Psychol. Bull. 1984, 95, 579. [CrossRef]
40. Clinch, J.J.; Keselman, H.J. Parametric alternatives to the analysis of variance. J. Educ. Stat. 1982, 7, 207–214.
[CrossRef]
41. Fujikoshi, Y. Two-way ANOVA models with unbalanced data. Discret. Math. 1993, 116, 315–334. [CrossRef]
© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
30
applied
sciences
Article
Sentiment Classification Using Convolutional
Neural Networks
Hannah Kim and Young-Seob Jeong *
Department of Future Convergence Technology, Soonchunhyang University, Asan-si 31538, Korea;
[email protected]

Received: 29 April 2019; Accepted: 4 June 2019; Published: 7 June 2019
Abstract: As the number of textual data is exponentially increasing, it becomes more important to
develop models to analyze the text data automatically. The texts may contain various labels such
as gender, age, country, sentiment, and so forth. Using such labels may bring benefits to some
industrial fields, so many studies of text classification have appeared. Recently, the Convolutional
Neural Network (CNN) has been adopted for the task of text classification and has shown quite
successful results. In this paper, we propose convolutional neural networks for the task of sentiment
classification. Through experiments with three well-known datasets, we show that employing
consecutive convolutional layers is effective for relatively longer texts, and our networks are better
than other state-of-the-art deep learning models.
Keywords: deep learning; convolutional neural network; sentiment classification
1. Introduction
In the Big Data era, the amount of various data, such as image, video, sound, and text, is
increasing exponentially. As text is the largest among them, studies related to text analysis have been
actively conducted from the past to the present. In particular, text classification has drawn much
attention because the text may have categorical labels such as sentiment (e.g., positive or negative),
author gender, language category, or various types (e.g., spam or ham). For example, the users of
Social Network Services (SNS) mostly represent their sentimental feeling, and they often share some
opinions about daily news with the public or friends. Emotional analysis involves mood categories
(e.g., happiness, joy, satisfaction, angry), while sentiment analysis involves categories such as positive,
neutral, and negative. In this paper, we target the sentiment analysis that classifies the given text
into one of sentiment categories. On websites about movies, people are likely to post their comments
that probably contain sentiment or opinions. If such a sentiment is accurately predicted, then it
will be applicable to various industrial fields (e.g., movie recommendation, personalized news-feed).
Indeed, the international market of movies is growing much faster than before, so many companies
(e.g., Netflix) provide movie recommendation services that essentially predict the sentiment or rating
scores of customers.
There have been many studies that have adopted machine learning techniques for text
classification. Although the machine learning techniques have been widely used and have shown
quite successful performance, they strongly depend on manually-defined features, where the feature
definition requires much effort of domain experts. For this reason, deep learning techniques have
been drawing attention recently, as they may reduce the effort for the feature definition and achieve
relatively high performance (e.g., accuracy). In this paper, we aim at sentiment classification for text
data and propose an architecture of the Convolutional Neural Network (CNN) [1–3], which is a type of

Appl. Sci. 2019, 9, 2347
deep learning model. We demonstrate the effectiveness of our proposed network through experimental
comparison with other machine learning models.
The contributions of this paper can be summarized as follows: (1) we design an architecture of two
consecutive convolutional layers to improve performance for long and complex texts; (2) we provide
a discussion about the architectural comparison between our models and two other state-of-the-art
models; (3) we apply the CNN model to binary sentiment classification and achieved 80.96%, 81.4%,
and 70.2% weighted F1 scores for Movie Review (MR) data [4–6], Customer Review (CR) data [7],
and Stanford Sentiment Treebank (SST) data [8], respectively; and (4) we also show that our model
achieved 68.31% for ternary sentiment classification with the MR data.
The remainder of this paper is organized as follows. Section 2 reviews related studies about
sentiment classification, machine learning, and deep learning models. Section 3 presents a detailed
description of our proposed network. Section 4 describes the experimental settings and datasets
and compares the proposed model with some other models. Thereafter, Section 5 discusses the
experimental results. Finally, Section 6 concludes this paper.
2. Background
2.1. Machine Learning for Sentiment Classification

The sentiment can be defined as a view of or an attitude toward a situation or event. It often
involves several types of self-indulgent feelings: happiness, tenderness, sadness, or nostalgia. One may
define the sentiment labels as a polarity or valence (e.g., positive, neutral, and negative) or several
types of emotional feeling (e.g., angry, happy, sad, proud). The definition of sentiment labels will
affect the outcome of the sentiment analysis, so we need to define the sentiment labels carefully. There
have been studies that defined three or more sentiment labels (e.g., opinion rating scores, emotional
feelings) [9–13], and some studies adopted two-dimensional labels (e.g., positive and negative) [14–26].
Although there has been much performance improvement in the field of sentiment analysis, the binary
classification of sentiment still remains challenging; for example, the performance (e.g., accuracy)
of recent studies varied from 70–90% in terms of the characteristics of data. In this paper, we aim
at the binary classification (positive or negative) and the ternary classification (positive, neutral,
and negative).
There have been many studies on classifying sentiments using machine learning models, such
as Support Vector Machine (SVM), Naive Bayes (NB), Maximum Entropy (ME), Stochastic Gradient
Descent (SGD), and ensemble. The most widely-used features for such machine learning models
have been n-grams. Read [14] used unigram features for sentiment binary classification and obtained
88.94% accuracy using the SVM. Kennedy and Inkpen [15] adopted unigram and bigram features
for sentiment binary classification and achieved 84.4% accuracy for movie review data [27] using
the SVM. Wan [20] tackled binary classification with unigram and bigram features and achieved
86.1% accuracy for translated Amazon product review data. In [21], Akaichi utilized a combination
of unigram, bigram, and trigram features and obtained 72.78% accuracy using the SVM. Valakunde
and Patwardhan [28] aimed at five-class (e.g., strong positive, positive, neutral, negative, and strong
negative) sentiment classification and obtained 81% accuracy using the SVM with bigram features. In
the study of Gautam and Yadav [18], they utilized the SVM along with the semantic analysis model
for the sentiment binary classification of Twitter texts and achieved 89.9% accuracy using unigram
features. Tripathy et al. [19] also used the SVM with n-gram features and obtained 88.94% accuracy
for sentiment binary classification. Hasan et al. [25] adopted unigram features for sentiment binary
classification and achieved 79% using the NB for translated Urdu tweets data. All of these studies
that utilized n-gram features generally achieved 70–90% accuracies, and the most effective model was
the SVM.
There also have been studies that have defined hand-crafted features for sentiment classification.
Yassine and Hajj [29] used affective lexicon, misspelling, and emoticons as features and obtained
32
Appl. Sci. 2019, 9, 2347
87% accuracy for ternary classification using the SVM. Denecke [22] defined three scores (e.g.,
positivity, negativity, and objectivity) as features and achieved 67% precision and 66% recall for
binary classification using the Logistic Regression (LR) classifier. Jiang et al. [16] tackled binary
classification using the SVM and achieved 67.8% accuracy for Twitter texts, where they adopted two
kinds of target-independent features (e.g., twitter content features and sentiment lexicon features).
Bahrainian and Dengel [23] used the number of positive words and the number of negative words
as features and achieved 86.7% accuracy for binary classification with Twitter texts. Neethu and
Rajasree [17] combined unigram features and their own Twitter-specific features and obtained 90%
accuracy for binary classification using the SVM. Karamibekr and Ghorbani [30] defined the number
opinion nouns as a feature and combined the feature with unigrams. They achieved 65.46% accuracy
for ternary classification using the SVM. Antai [24] used the normalized frequencies of words as
features and obtained 84% accuracy for binary classification using the SVM. Ghiassi and Lee [31]
aimed at five-class sentiment classification, and they defined a domain-independent feature set for
Twitter data. They achieved a 92.7% F1 score using the SVM. Mensikova and Mattmann [26] utilized
the results of Named Entity (NE) extraction as features and obtained a 0.9 False Positive Rate (FPR).
These studies elaborated on feature definition rather than using only n-gram features, and they
accomplished better performance (e.g., 92.7% F1 score). It is not fair, of course, to compare the
performance between the previous studies because they used different datasets. However, as shown
in [17,30], it is obvious that the combination of hand-crafted features and the n-grams must be better
than using only n-gram features.
Although there has been success using machine learning models with the hand-crafted features
and the n-gram features, these studies have a common limitation that their performance varied
depending how well the features were defined; for different data, much effort of domain experts will
be required to achieve better performance. This limitation also exists in information fusion approaches
for sentiment analysis [32] that combine other resources (e.g., ontology, lexicon), as it will cost much
time and effort of domain experts. Deep learning model is one of the solutions for such a limitation,
as it is known to capture arbitrary patterns (i.e., features) automatically. Furthermore, as stated in [33],
using the deep learning model for sentiment analysis will provide a meta-level feature representation
that generalizes well on new domains.
In this paper, we take the deep learning technique to tackle the sentiment classification. In the
next subsection, we review previous studies that adopted the deep learning techniques for the
sentiment classification.
2.2. Deep Learning for Sentiment Classification

The deep learning technique, one of the machine learning techniques, has been recently widely
used to classify sentiments. Dong et al. [34] designed a new model, namely the Adaptive Recursive
Neural Network (AdaRNN), that classifies the Twitter texts into three sentiment labels: positive,
neutral, and negative. By experimental results, the AdaRNN achieved 66.3% accuracy. Huang et al. [35]
proposed Hierarchical Long Short-Term Memory (HLSTM) and obtained 64.1% accuracy on Weibo
tweet texts. Tang et al. [36] introduced a new variant of the RNN model, the Gated Recurrent Neural
Network (GRNN), which achieved 66.6% (Yelp 2013–2015 data) and 45.3% (IMDB data) accuracies. All
of these studies commonly assumed that there were three or more sentiment labels.
Meanwhile, Qian et al. [37] utilized Long Short-Term Memory (LSTM) [38–41] for binary
classification of sentiment and obtained 82.1% accuracy on the movie review data [27]. Kim [1]
had a result of a maximum of 89.6% accuracy with seven different types of data through their CNN
model with one convolutional layer. Zhang et al. [42] proposed three-way classification and obtained
a maximum 94.2% accuracy with four datasets, where their best three-way model was NB-SVM.
Severyn and Moschitti [43] employed a pretrained Word2Vec for their CNN model and achieved
84.79% (phrase-level) and 64.59% (message-level) accuracies with SemEval-2015 data. The CNN model
used in [43] was essentially the same as the model of [1]. Deriu et al. [44] trained the CNN model that
33
Appl. Sci. 2019, 9, 2347
had a combination of two convolutional layers and two pooling layers to classify tweet data of four
languages sentimentally and obtained a 67.79% F1 score. In the study of Ouyang et al. [45], the CNN
model, which had three convolution/pooling layer pairs, was proposed, and the model outperformed
other previous models including the Matrix-Vector Recursive Neural Network (MV-RNN) [46]. We
conducted experiments with several of our CNN models that had different structures, and two of our
models (e.g., the seventh and eighth models) were similar to the model of [44,45].
As summarized above, for the sentiment classification, there are two dominant types of deep
learning technique: RNN and CNN. In this work, we propose a CNN model, the structure of which
was carefully designed for effective sentiment classification. In the next subsection, we explain the
advantages of the CNN in analyzing text.
2.3. Convolutional Neural Network for Text Classification

Among the existing studies using deep learning to classify texts, the CNN takes advantage of the
so-called convolutional filters that automatically learn features suitable for the given task. For example,
if we use the CNN for the sentiment classification, the convolutional filters may capture inherent
syntactic and semantic features of sentimental expressions, as shown in [47]. It has been shown
that a single convolutional layer, a combination of convolutional filters, might achieve comparable
performance even without any special hyperparameter adjustment [1]. Furthermore, the CNN does
not require expert knowledge about the linguistic structure of a target language [48]. Thanks to these
advantages, the CNN has been successfully applied to various text analyses: semantic parsing [49],
search by query [50], sentence modeling [2].
One may argue that the Recurrent Neural Network (RNN) [51] might be better for the text
classification than for the CNN, as it preserves the order of the word sequence. However, the CNN
is also capable of capturing sequential patterns, as concerns the local patterns by the convolutional
filters; for example, the convolutional filters along with the attention technique have been successfully
applied to machine translation [52]. Moreover, compared to the RNN, the CNN mostly has a smaller
number of parameters, so that the CNN is trainable with a small amount of data [43]. The CNN is also
known to explore the richness of pretrained word embeddings [53].
In this paper, we design a CNN model for the sentiment classification and show that our network
is better than other deep learning models through experimental results.
3. The Proposed Method

CNN, which has been widely used on image datasets, extracts the significant features of the
image, as the “convolutional” filter (i.e., kernel) moves through the image. If the input data are given
as one-dimensional, the same function of CNN could be used in the text as well. In the text area, while
the filter moves, local information of texts is stored, and important features are extracted. Therefore,
using CNN for text classification is effective.
Figure 1 shows a graphical representation of the proposed network. The network consisted of an
embedding layer, two convolutional layers, a pooling layer, and a fully-connected layer. We padded
the sentence vectors to make a fixed size. That is, too long sentences were cut to a certain length, and
too short sentences were appended with the [PAD] token. We set the fixed length S to be the maximum
length of the sentences. An embedding layer that maps each word of a sentence to an E-dimensional
feature vector outputs an S × E matrix, where E denotes the embedding size. For example, suppose
that 10 is king, 11 is shoes, and 20 is queen in the embedding space. 10 and 20 are close in this space
due to the semantic similarity of king and queen, but 10 and 11 are quite far because of the semantic
dissimilarity of king and shoes. In this example, 10, 11, and 20 are not numeric values, they are just
the simple position in this space. In other words, the embedding layer is a process of placing words
received as input into a semantically well-designed space, where words with similar meanings are
located close and words with opposite meanings are located far apart, digitizing them into a vector.
The embedding is the process of projecting a two-dimensional matrix into a low-dimensional vector
34
Appl. Sci. 2019, 9, 2347
space (E-dimension) to obtain a word vector. The embedding vectors can be obtained from other
resources (e.g., Word2Vec) or from the training process. In this paper, our embedding layer was
obtained through the training process, and all word tokens including the [UNK] token for unseen
words would be converted to numeric values using the embedding layer.
Figure 1. The graphical representation of the network, where the output dimensions of each layer are
represented at the bottom of the corresponding layers.
The S × E matrix, the output of the embedding layer, is then laid down as the first convolutional
layer. The first convolutional layer is the C1 × E matrix, which stores the local information needed
to classify the sentiment class in a S × E matrix and convey information to the next convolutional
layer. The C1 × E matrix slides (i.e., convolves) all the values of the S × E matrix with an arbitrary
stride, calculates the dot product, and passes the dot product result to the next layer. The second
convolutional layer uses the C2 × 1 matrix to extract features from the contextual information of the
main word based on the local information stored in the first convolutional layer. C1 and C2 denote the
filter size of each convolutional layer, and the two convolutional layers have K1 and K2 distinct filters,
respectively, to capture unique contextual information. In other words, the first convolutional layer is
utilized to look at simple contextual information while looking over the S × E matrix, and the second
convolutional layer is utilized to capture key features and then extract them (e.g., worst, great) that
contain sentiments affecting classification.
The matrix that passed through the consecutive convolutional layer is used as the input to the
pooling layer. While average-pooling and L2-norm pooling have been used as the pooling layer
position, in this paper, we used the max-pooling, which is a technique for selecting the largest value as
a representative of the peripheral values. Since the sentiment is often determined by a combination of
several words rather than expressing the sentiment in every word in the sentence, we adopted the
max-pooling technique. The pooling layer slides all the values of the matrix, which is the output of the
second convolutional layer, with an arbitrary stride, resulting in output vectors. Since max-pooling
is the layer that passes to the next layer the largest value among several values, it results in output
vectors of a much smaller size. In other words, the convolutional layer looks at the context and extracts
the main features, and the pooling layer plays a role in selecting the most prominent features.
After passing through the pooling layer, a flattening process is performed to convert the
two-dimensional feature map from the output into a one-dimensional format and to delivery it
to an F-dimensional Fully-Connected (FC) layer. Since the FC layer takes a one-dimensional vector
as the input, the two-dimensional vector delivered from the pooling layer needs to be flattened. The
FC layer connects all input and output neurons. A vector that passes through the FC layer forms an
output that is classified as positive or negative. The activation function softmax functions to classify
multiple classes in the FC layer. The softmax function outputs, the value, which is the probability
value, is generated for each class.
Most people might think that with many convolutional layer stacks, it may be better to store local
information and to extract contextual information; however, deep networks do not always have higher
performance than shallow networks. Because a network’s variation (e.g., deep or shallow) might rely
35
Appl. Sci. 2019, 9, 2347
on the length of the data and the number of data and features, in this paper, we argue that data passing
through two consecutive convolutional layers and then passing through the pooling layer is successful
at storing context information and extracting prominent features. This assertion is demonstrated by
experiments and discussed in Section 5.
4. Experiment
4.1. Data
We used three Movie Review (MR) data from kaggle [4–6]. The dataset in [4] consisted of two
labels, positive and negative, while [5] was composed of three labels of positive, neutral, and negative.
Furthermore, the dataset in [6] was composed of five labels of positive, somewhat positive, neutral,
somewhat negative, and negative. The positive class of [4], the positive class of [5], and the positive
and somewhat positive classes of [6] were merged and labeled as the positive class. The neutral class
of [5] and the neutral class of [6] were merged and labeled as the neutral class. Then, the negative
class of [4], the negative class of [5], and the somewhat negative and negative classes of [6] were
merged and labeled as the negative class. Positive, neutral, and negative were assigned as 1, 0, and 2,
respectively. Some 11,800 positive data, 5937 neutral data, and 9698 negative data were used, totaling
27,435. Amazon Customer Review data (CR) [7] were labeled as positive and negative, and the amount
of positive data was more than that of negative data. Therefore, we used only 3671 out of 34,062 data
to control the positive:negative ratio. Stanford Sentiment Treebank data (SST) [8] were labeled as a
value between 0 and 1, so a value of 0.5 or more was relabeled as positive, and a value of less than 0.5
was relabeled as negative.
Table 1 shows the detailed statistics of the data. N is the number of text data, Dist (+, −) is the
ratio of positive:negative classes, and aveL/maxL is the average length/maximum length of the text.
We divide all the data into three sets (i.e., train, validation, and test) with a ratio of 55:20:25, and |V |
was the dictionary size of each dataset.
Table 1. Detailed statistics of the data.
Data N Dist (+,−) aveL/maxL Train:Test:Val |V |

MR 21,498 55:45 31/290 12,095:5375:4031 9396
CR 3671 62:38 19/227 2064:918:689 1417
SST 11,286 52:48 12/41 6348:2822:2116 3550
4.2. Preprocessing
Preprocessing was carried out to modify the text data appropriately in the experiment. We used
decapitalization and did not mark the start and end of the sentences. We deleted #, two or more spaces,
tabs, Retweets (RT), and stop words. We also changed the text that represented the url that began
with “http” to [URL] and the text that represented the account ID that began with “@” to [NAME].
In addition, we changed digits to [NUM], and special characters to [SPE]. We changed “can’t” and
“isn’t” to “can not” and “is not”, respectively, since “not” is important in sentiment analysis.
We split the text data by space and constructed a new dictionary using only the words appearing
more than six times in the whole text. In the MR data, 9394 words out of a total of 38,195 words were
made into a dictionary. In CR data, 1415 out of a total of 5356 words, and in SST data, 3548 out of a
total of 16,870 words were made. The new dictionary also included [PAD] for padding and [UNK] to
cover missing words. Padding is a method of cropping all text to a fixed length and filling it with the
[PAD] token for text shorter than that length. Here, the fixed length was the maximum length of text
in each dataset.
36
Appl. Sci. 2019, 9, 2347
4.3. Performance Comparison

Decision Tree (DT), Naive Bayes (NB), Support Vector Machine (SVM) [54], and Random Forest
(RF) [55] were used for comparing with our CNN models. Table 2 summarizes the description of the
traditional machine learning models and the parameter settings. The parameters were optimized via a
grid searching. We also compared our CNN models with the state-of-the-art models such as Kim’s [1]
and Zhang et al.’s [42].
Table 2. The description of the traditional machine learning models and the parameter settings.
Model Description
- Probabilistic classifier based on the Bayes’ theorem
Naive Bayes (NB)
- Assumes the independence between features
- C4.5 classifier using the J48 algorithm
- Confidence factor for pruning: 0.25
- Minimum number of instances: 1
Decision Tree (DT)
- The number of decimal places: 1
- Determines the amount of data used for pruning: 5
- Reduced error pruning: True
- Non-probabilistic binary classification model that finds a decision
Support Vector Machine (SVM) boundary with a maximum distance between two classes
- Kernel: RBF
- Exponent = 1.0, Complexity (c) = 10.0
- Kind of ensemble model that generates final result by incorporating the
results of multiple decision trees
Random Forest (RF) - Maximum tree depth: 30
- Number of iterations: 150
- Each tree has no depth limitation
In order to optimize the CNN model proposed in this paper, we experimented with various
settings, and the optimal settings were as follows: (1) the embedding dimension E = 25; (2) C1 and C2
were set to 3; (3) K1 and K2 were 16 and 8; and (4) the pooling layer dimension P = 2. We set the strides
for the kernels to be 1. Every layer took the Rectified Linear Unit (ReLU) as an activation function
except for the output layer that took a softmax function. The number of the trainable parameter was
476,331, and we adopted Adam’s optimizer with an initial learning rate of 0.001. We took the L2
regularization for the fully-connected layer. The number of epochs was changed according to the
complexity of the network (i.e., the number of layers and parameters). For example, the first network
was trained for 5 epochs, and the model of Kim [1] was trained for 10 epochs.
In the study of Zhang et al. [42], NB-SVM was the best. The effectiveness of the combination
with different models (e.g., NB, SVM, NB-SVM) may vary for different tasks. We tested different
combinations with our datasets and discovered that the SVM with the RBF kernel was the best
enhanced model for our task. Thus, the CNN enhanced by SVM was used for comparison with
other models.
5. Result and Discussion
5.1. Result
The three datasets (movie review data, customer review data, and Stanford Sentiment Treebank
data) were applied to the proposed CNN models, traditional machine-learning models, and other
state-of-the-art models. Note that, we conducted two experiments: binary classification and ternary
classification. Tables 3 and 4 show the experimental results of binary classification, where the two
values for each cell correspond to the “positive” and “negative” classes, respectively. Table 5 shows the
experimental results of ternary classification in MR data. Table 3 shows the experimental results with
37
Appl. Sci. 2019, 9, 2347
the MR data, while Table 4 describes the experimental results with the CR and SST data. These tables
include the accuracy, precision, recall, F1 score, and weighted-F1 score. For example, the F1 scores of
the decision tree are 67.2% for positive and 31.1% for negative in Table 3.
Table 3. The results with the MR data, where Emb, Conv, Pool, globalpool, FC stand for embedding
layer, convolutional layer, pooling layer, global pooling layer, and fully-connected layer, respectively.
Model Accuracy Precision Recall F1 Weighted-F1

Decision Tree 59.64 58.0/64.0 76.2/39.6 67.4/47.1 57.2
Naive Bayes 56.40 57.5/53.3 77.7/30.7 66.1/38.9 52.0
Support Vector Machine 54.95 57.7/55.2 93.0/9.1 69.3/15.5 44.9
Random Forest 58.73 56.4/59.8 74.7/39.4 66.4/46.4 58.1
Kim [1] 80.85 80.7/80.9 76.2/84.7 78.3/82.8 80.75
Zhang et al. [42] 77.28 72.4/69.1 56.1/82.1 63.2/75.0 69.62
Emb+Conv+Conv+Pool+FC 81.06 81.5/80.7 75.6/85.6 78.4/83.1 80.96
Emb+Conv+Pool+FC 79.70 77.4/81.63 78.2/80.9 77.8/81.3 79.71
Emb+Conv+Conv+Conv+Pool+FC 80.30 80.2/80.4 75.3/84.5 77.7/82.4 80.26
Emb+Conv+Pool+Conv+FC 78.17 74.4/81.8 79.5/77.1 76.8/79.4 78.22
Emb+Conv+globalpool+FC 77.54 77.3/77.7 71.8/82.4 74.4/79.9 77.39
Emb+Conv+Conv+globalpool+FC 79.06 79.1/79.0 73.5/83.8 76.2/81.3 78.98
Emb+Conv+Pool+Conv+Pool+FC 79.11 78.6/79.5 74.3/83.1 76.4/81.2 79.0
Emb+Conv+Pool+Conv+Pool+Conv+Pool+FC 74.61 84.1/72.8 59.5/90.6 69.7/80.7 75.7
Table 4. The results with the CR and SST data, where Emb, Conv, Pool, globalpool, FC stand
for embedding layer, convolutional layer, pooling layer, global pooling layer, and fully-connected
layer, respectively.
CR SST
Model
Weighted-F1 F1 Weighted-F1 F1
Decision Tree 63.7 47.0/74.5 51.5 62.4/41.7
Naive Bayes 61.0 77.7/31.8 35.7 10.6/63.4
Support Vector Machine 59.7 26.5/78.5 37.8 68.6/4.0
Random Forest 64.4 41.8/76.9 51.2 59.0/47.3
Kim [1] 74.8 78.6/65.6 56.1 47.2/66.3
Zhang et al. [42] 54.8 64.7/37.7 52.1 45.6/59.5
Emb+Conv+Conv+Pool+FC 78.3 84.8/67.1 68.3 70.5/65.7
Emb+Conv+Pool+FC 78.3 82.3/71.3 66.5 67.7/65.1
Emb+Conv+Conv+Conv+Pool+FC 70.4 82.0/50.3 68.2 68.4/68.0
Emb+Conv+Pool+Conv+FC 75.3 84.0/60.3 68.6 70.5/66.5
Emb+Conv+globalpool+FC 81.4 86.1/73.4 70.2 72.8/67.2
Emb+Conv+Conv+globalpool+FC 79.4 84.5/70.5 70.0 71.2/68.7
Emb+Conv+Pool+Conv+Pool+FC 73.18 82.8/56.5 66.62 67.7/65.4
Emb+Conv+Pool+Conv+Pool+Conv+Pool+FC 51.57 77.6/6.7 65.21 70.2/59.5
In Table 3, the Random Forest (RF) achieved the best F1 scores, while the Decision Tree (DT)
exhibited comparable results. Our first network was the best among all models including the
state-of-the-art models of [1,42]. Similarly, in Table 4, the RF and the DT achieved the best F1 scores
among the traditional models, but our fifth network outperformed all other models. Note that the best
networks in Tables 3 and 4 were different. That is, the first network was the best with the MR data,
which had relatively longer sentences, as shown in Table 3, whereas the fifth network was the best for
the other two datasets (e.g., CR, SST) that had relatively shorter sentences.
One may argue that stacking more convolutional layers might be better, as the deeper network
is known to capture higher level patterns. This might be true, but it should be noted that the deeper
network is not always better than the shallow networks. The depth of the network needs to be
determined according to the data characteristics; too deep networks will probably overfit, whereas too
38
Appl. Sci. 2019, 9, 2347
shallow networks will underfit. For example, although the third network was deeper than the first
network, the first network outperformed the third network, as shown in Table 3. To clarify this, we
conducted experiments varying the number of convolutional layers from 1–5, and its result is depicted
in Figure 2. Meanwhile, the first network can be compared with the fourth network to show that
stacking two consecutive convolutional layers is better. The fifth and sixth networks may be compared
with the first and second networks, if one can see if the max-pooling and the global max-pooling
contribute to the models. The seventh and eighth networks had similar structures to that in [44,45].
Figure 2. Classification performance according to the depth of the convolutional layer, where the
horizontal axis means the number of stacked convolutional layers.
We also conducted ternary classification. The binary classification assumed that there were only
two labels (e.g., positive and negative), whereas the ternary classification assumed that there were
three labels (e.g., positive, negative, and neutral). As shown in Table 5, our first network was the best
for the ternary classification, but the performance fell to a 68.31% weighted-F1 score.
Table 5. Performance of ternary classification with the MR data, where Emb, Conv, Pool, globalpool, FC
stand for embedding layer, convolutional layer, pooling layer, global pooling layer, and fully-connected
layer, respectively.
Model Weighted-F1
Decision Tree 46.8
Naive Bayes 27.2
Support Vector Machine 34.9
Random Forest 46.2
Kim [1] 68.2
Zhang et al. [42] 67.5
Emb+Conv+Conv+Pool+FC 68.3
Emb+Conv+Pool+FC 55.0
Emb+Conv+Conv+Conv+Pool+FC 64.6
Emb+Conv+Pool+Conv+FC 66.9
Emb+Conv+globalpool+FC 65.3
Emb+Conv+Conv+globalpool+FC 67.5
5.2. Discussion
5.2.1. Comparison with Other Models

In terms of the F1 scores, our network was about 10% greater than the traditional machine models.
We believe that the biggest reason is the inherent ability of the CNN to capture local patterns. The
CNN has the ability to capture local patterns and higher level patterns through its convolutional and
pooling layers, as explained in [48]. We believe that such an ability of the CNN model produces the
performance gap. The proposed network was also better than the state-of-the-art deep learning models
39
Appl. Sci. 2019, 9, 2347
such as those of Kim [1] and Zhang et al. [42]. Kim [1] and Zhang et al. [42] commonly utilized three
filters (sizes of 3, 4, and 5) in parallel, in the hope to capture the filters’ arbitrary features with different
sizes. On the other hand, our network had consecutive filters that had two benefits: (1) hierarchical
(higher) features were extracted, and (2) the high-level filters would have a broader range to see
local patterns, where such advantages of consecutive filters were already reported in many previous
studies [56–58]. The three-way approach, employed in [42], gave worse performance than the model
of [1], although it was designed to combine of a CNN model and a traditional machine learning model.
The reason can be explained by the fact that the CNN model itself was already sufficient to achieve
the high performance, as shown in Table 3, so the combination with other traditional models (e.g.,
Naive Bayes) degraded the overall performance. In Zhang et al. [42], among the sentences classified by
CNN [1], the sentences with weak confidence were re-classified by the traditional machine learning
models, and it improved the overall performance. On the other hand, for the three datasets in this
paper, we found that the ratio of weakly-classified instances was biased, as shown in Figure 3; this
means that the entropy of the weakly-classified instances was small, so it was difficult to expect that
the performance would improve by the combination with the traditional models.
Figure 3. The ratio of positive/negative sentences with weak confidence.
5.2.2. Network Structure

As shown in Tables 3 and 4, the max-pooling had better performance in MR data, whereas the
global max-pooling seemed helpful in the CR data and the SST data. The max-pooling layer gave the
largest value in a certain subarea as an output, while the global max-pooling did this in the whole area.
Figure 4 shows the difference.
Figure 4. The difference of max-pooling and global max-pooling.
If MR data adopted global max-pooling, the large number of missing values would be discovered.
Because the length of the total sentence of the MR data was longer than that of CR data and SST data,
the performance of the MR data was lower. On the other hand, CR data and SST data with relatively
a short length of sentences were not degraded even when using global max-pooling and, in some
cases, performed better than using max-pooling. As a result, CR data and SST data showed the highest
performance when one convolutional layer and global max-pooling layer were combined. The reason
is also that they were shorter and simpler data than the MR data, so they showed higher performance
in relatively monotonous models. However, The network with two convolutional layers never lagged
40
Appl. Sci. 2019, 9, 2347
behind the traditional machine learning models and state-of-the-art models in terms of performance in
this experimental result.
6. Conclusions
In this paper, we designed a convolutional neural network for the sentiment classification.
By experimental results, we showed that the consecutive convolutional layers contributed to better
performance on relatively long text. The proposed CNN models achieved about 81% and 68%
accuracies for the binary classification and ternary classification, respectively. As future work, we
will apply our work to other classification tasks (e.g., gender classification). We will also continue
finding better structures for the sentiment classification; for example, residual connection for stacking
more layers.
Author Contributions: Conceptualization, H.K.; validation, H.K., Y.-S.J.; investigation, H.K.; resources, Y.-S.J.;
data curation, H.K.; writing–original draft preparation, H.K.; writing–review and editing, Y.-S.J.; visualization,
H.K.; supervision, Y.-S.J.; project administration, Y.-S.J.
Acknowledgments: This work was supported by the National Research Foundation of Korea(NRF) grant funded
by the Korea government(MSIP; Ministry of Science, ICT & Future Planning) (No. 2019021348). This work was
supported by the Soonchunhyang University Research Fund.
References
1. Kim, Y. Convolutional neural networks for sentence classification. arXiv 2014, arXiv:1408.5882.
2. Nal, K.; Grefenstette, E.; Blunsom, P. A convolutional neural network for modelling sentences. arXiv 2014,
arXiv:1404.2188.
3. Lei, T.; Barzilay, R.; Jaakkola, T. Molding cnns for text: Non-linear, non-consecutive convolutions. arXiv
2015, arXiv:1508.04112.
4. Amazon Movie Review Dataset. Available online: https://www.kaggle.com/ranjan6806/corpus2#corpus/
(accessed on 11 November 2012).
5. Movie Review Dataset. Available online: https://www.kaggle.com/ayanmaity/movie-review#train.tsv/
6. Rotten Tomatoes Movie Review Dataset. Available online: https://www.kaggle.com/c/movie-review-
sentiment-analysis-kernels-only/ (accessed on 11 November 2012).
7. Consumer Reviews Of Amazon Products Dataset. Available online: https://www.kaggle.com/datafiniti/
consumer-reviews-of-amazon-products/ (accessed on 11 November 2012).
8. Stanford Sentiment Treebank Dataset. Available online: https://nlp.stanford.edu/sentiment/code.html/
9. Pak, A.; Paroubek, P. Twitter as a corpus for sentiment analysis and opinion mining. LREc 2010, 10,
1320–1326.
10. Alm, C.O.; Roth, D.; Sproat, R. Emotions from text: Machine learning for text-based emotion prediction.
In Proceedings of the Conference on Human Language Technology and Empirical Methods in Natural
Language Processing, Vancouver, BC, Canada, 6–8 October 2005; pp. 579–586.
11. Bartlett, M.S.; Littlewort, G.; Frank, M.; Lainscsek, C.; Fasel, I.; Movellan, J. Recognizing facial expression:
Machine learning and application to spontaneous behavior. In Proceedings of the 2005 IEEE Computer
Society Conference on Computer Vision and Pattern Recognition (CVPR’05), San Diego, CA, USA,
20–25 June 2005; Volume 2, pp. 568–573.
12. Heraz, A.; Razaki, R.; Frasson, C. Using machine learning to predict learner emotional state from
brainwaves. In Proceedings of the Seventh IEEE International Conference on Advanced Learning
Technologies (ICALT 2007), Niigata, Japan, 18–20 July 2007; pp. 853–857.
13. Yujiao, L.; Fleyeh, H. Twitter Sentiment Analysis of New IKEA Stores Using Machine Learning.
In Proceedings of the International Conference on Computer and Applications, Beirut, Lebanon,
25–26 July 2018.
41
Appl. Sci. 2019, 9, 2347
14. Read, J. Using emoticons to reduce dependency in machine learning techniques for sentiment classification.
In Proceedings of the ACL Student Research Workshop, Ann Arbor, Michigan, 27 June 2005; pp. 43–48.
15. Kennedy, A.; Inkpen, D. Sentiment classification of movie reviews using contextual valence shifters.
Comput. Intell. 2006, 22, 110–125. [CrossRef]
16. Jiang, L.; Yu, M.; Zhou, M.; Liu, X.; Zhao, T. Target-dependent twitter sentiment classification.
In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human
Language Technologies, Portland, Oregon, 19–24 June 2011; Volume 1, pp. 151–160.
17. Neethu, M.; Rajasree, R. Sentiment analysis in twitter using machine learning techniques. In Proceedings of
the 2013 Fourth International Conference on Computing, Communications and Networking Technologies
(ICCCNT), Tiruchengode, India, 4–6 July 2013; pp. 1–5.
18. Gautam, G.; Yadav, D. Sentiment analysis of twitter data using machine learning approaches and semantic
analysis. In Proceedings of the 2014 Seventh International Conference on Contemporary Computing (IC3),
Noida, India, 7–9 August 2014; pp. 437–442.
19. Tripathy, A.; Agrawal, A.; Rath, S.K. Classification of sentiment reviews using n-gram machine learning
approach. Expert Syst. Appl. 2016, 57, 117–126. [CrossRef]
20. Wan, X. A comparative study of cross-lingual sentiment classification. In Proceedings of the 2012
IEEE/WIC/ACM International Joint Conferences on Web Intelligence and Intelligent Agent Technology,
Macau, China, 4–7 December 2012; Volume 1, pp. 24–31.
21. Akaichi, J. Social networks’ Facebook’statutes updates mining for sentiment classification. In Proceedings
of the 2013 International Conference on Social Computing, Alexandria, VA, USA, 8–14 September 2013;
pp. 886–891.
22. Denecke, K. Using sentiwordnet for multilingual sentiment analysis. In Proceedings of the 2008 IEEE 24th
International Conference on Data Engineering Workshop, Cancun, Mexico, 7–12 April 2008; pp. 507–512.
23. Bahrainian, S.A.; Dengel, A. Sentiment analysis using sentiment features. In Proceedings of the 2013
IEEE/WIC/ACM International Joint Conferences on Web Intelligence (WI) and Intelligent Agent Technologies
(IAT), Atlanta, GA, USA, 17–20 November 2013; Volume 3; pp. 26–29.
24. Antai, R. Sentiment classification using summaries: A comparative investigation of lexical and statistical
approaches. In Proceedings of the 2014 6th Computer Science and Electronic Engineering Conference
(CEEC), Colchester, UK, 25–26 September 2014; pp. 154–159.
25. Hasan, A.; Moin, S.; Karim, A.; Shamshirband, S. Machine learning-based sentiment analysis for twitter
accounts. Math. Comput. Appl. 2018, 23, 11. [CrossRef]
26. Mensikova, A.; Mattmann, C.A. Ensemble sentiment analysis to identify human trafficking in web data.
Workshop on Graph Techniques for Adversarial Activity Analytics (GTA 2018), Marina Del Rey, CA, USA,
5–9 February 2018
27. Pang, B.; Lee, L. A sentimental education: Sentiment analysis using subjectivity summarization based on
minimum cuts. In Proceedings of the 42nd Annual Meeting on Association for Computational Linguistics,
Barcelona, Spain, 21–26 July 2004; pp. 271–278.
28. Valakunde, N.; Patwardhan, M. Multi-aspect and multi-class based document sentiment analysis of
educational data catering accreditation process. In Proceedings of the 2013 International Conference on
Cloud & Ubiquitous Computing & Emerging Technologies, Pune, India, 15–16 November 2013; pp. 188–192.
29. Yassine, M.; Hajj, H. A framework for emotion mining from text in online social networks. In Proceedings of
the 2010 IEEE International Conference on Data Mining Workshops, Sydney, Australia, 13 December 2010;
pp. 1136–1142.
30. Karamibekr, M.; Ghorbani, A.A. A structure for opinion in social domains. In Proceedings of the 2013
International Conference on Social Computing, Alexandria, VA, USA, 8–14 September 2013; pp. 264–271.
31. Ghiassi, M.; Lee, S. A domain transferable lexicon set for Twitter sentiment analysis using a supervised
machine learning approach. Expert Syst. Appl. 2018, 106, 197–216. [CrossRef]
32. Balazs, J.A.; Velásquez, J.D. Opinion mining and information fusion: A survey. Inf. Fusion 2016, 27, 95–110.
[CrossRef]
33. Chaturvedi, I.; Cambria, E.; Welsch, R.E.; Herrera, F. Distinguishing between facts and opinions for sentiment
analysis: Survey and challenges. Inf. Fusion 2018, 44, 65–77. [CrossRef]
42
Appl. Sci. 2019, 9, 2347
34. Dong, L.; Wei, F.; Tan, C.; Tang, D.; Zhou, M.; Xu, K. Adaptive recursive neural network for target-dependent
twitter sentiment classification. In Proceedings of the 52nd Annual Meeting of the Association for
Computational Linguistics (Volume 2: Short Papers), Baltimore, Maryland, USA, 23–25 June 2014; Volume 2,
pp. 49–54.
35. Huang, M.; Cao, Y.; Dong, C. Modeling rich contexts for sentiment classification with lstm. arXiv 2016,
arXiv:1605.01478.
36. Tang, D.; Qin, B.; Liu, T. Document modeling with gated recurrent neural network for sentiment classification.
In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, Lisbon,
Portugal, 17–21 September 2015; pp. 1422–1432.
37. Qian, Q.; Huang, M.; Lei, J.; Zhu, X. Linguistically regularized lstms for sentiment classification. arXiv 2016,
arXiv:1611.03949.
38. Mikolov, T. Statistical language models based on neural networks. Ph.D. Thesis, Brno University of
Technology, Brno, Czechia, 2012.
39. Chung, J.; Gulcehre, C.; Cho, K.; Bengio, Y. Empirical evaluation of gated recurrent neural networks on
sequence modeling. arXiv 2014, arXiv:1412.3555.
40. Tai, K.S.; Socher, R.; Manning, C.D. Improved semantic representations from tree-structured long short-term
memory networks. arXiv 2015, arXiv:1503.00075.
41. Wang, F.; Zhang, Z.; Lan, M. Ecnu at semeval-2016 task 7: An enhanced supervised learning method
for lexicon sentiment intensity ranking. In Proceedings of the 10th International Workshop on Semantic
Evaluation (SemEval-2016), San Diego, CA, USA, 16–17 June 2016; pp. 491–496.
42. Zhang, Y.; Zhang, Z.; Miao, D.; Wang, J. Three-way enhanced convolutional neural networks for
sentence-level sentiment classification. Inf. Sci. 2019, 477, 55–64. [CrossRef]
43. Severyn, A.; Moschitti, A. Unitn: Training deep convolutional neural network for twitter sentiment
classification. In Proceedings of the 9th International Workshop On Semantic Evaluation (SemEval 2015),
Denver, CO, USA, 4–5 June 2015; pp. 464–469.
44. Deriu, J.; Lucchi, A.; De Luca, V.; Severyn, A.; Müller, S.; Cieliebak, M.; Hofmann, T.; Jaggi, M. Leveraging
large amounts of weakly supervised data for multi-language sentiment classification. In Proceedings of the
26th International Conference on World Wide Web, Perth, Australia, 3–7 April 2017; pp. 1045–1052.
45. Ouyang, X.; Zhou, P.; Li, C.H.; Liu, L. Sentiment analysis using convolutional neural network. In Proceedings
of the 2015 IEEE International Conference on Computer and Information Technology; Ubiquitous Computing
and Communications; Dependable, Autonomic and Secure Computing; Pervasive Intelligence and
Computing, Liverpool, UK, 26–28 October 2015; pp. 2359–2364.
46. Socher, R.; Huval, B.; Manning, C.D.; Ng, A.Y. Semantic compositionality through recursive matrix-vector
spaces. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing
and Computational Natural Language Learning, Jeju Island, Korea, 12–14 July 2012; pp. 1201–1211.
47. Rios, A.; Kavuluru, R. Convolutional neural networks for biomedical text classification: Application in
indexing biomedical articles. In Proceedings of the 6th ACM Conference on Bioinformatics, Computational
Biology and Health Informatics, Atlanta, GA, USA, 9–12 September 2015; pp. 258–267.
48. Zhang, X.; Zhao, J.; LeCun, Y. Character-level convolutional networks for text classification. In Advances in
Neural Information Processing Systems 28; Neural Information Processing Systems Foundation, Inc.: Montreal,
QC, Canada, 2015; pp. 649–657.
49. Yih, W.T.; He, X.; Meek, C. Semantic Parsing for Single-Relation Question Answering. In Proceedings
of the 52nd Annual Meeting of the Association for Computational Linguistics, Baltimore, MD, USA,
23–25 June 2014; Volume 2, pp. 643–648.
50. Shen, Y.; Xiaodong, H.; Gao, J.; Deng, L.; Mensnil, G. Learning Semantic Representations Using Convolutional
Neural Networks for Web Search. In Proceedings of the 23rd International Conference on World Wide Web,
Seoul, Korea, 7–11 April 2014; pp. 373–374.
51. Lai, S.; Xu, L.; Liu, K.; Zhao, J. Recurrent convolutional neural networks for text classification. In Proceedings
of the Twenty-Ninth AAAI Conference on Artificial Intelligence, Austin, TX, USA, 25–30 January 2015.
52. Gehring, J.; Auli, M.; Grangier, D.; Dauphin, Y.N. A convolutional encoder model for neural machine
translation. arXiv 2016, arXiv:1611.02344.
43
Appl. Sci. 2019, 9, 2347
53. Dos Santos, C.; Gatti, M. Deep convolutional neural networks for sentiment analysis of short texts.
In Proceedings of the COLING 2014, the 25th International Conference on Computational Linguistics:
Technical Papers, Dublin, Ireland, 23–29 August 2014; pp. 69–78.
54. Boser, B.E.; Guyon, I.M.; Vapnik, V.N. A training algorithm for optimal margin classifiers. In Proceedings
of the Fifth Annual Workshop on Computational Learning Theory, Pittsburgh, PA, USA, 27–29 July 1992;
pp. 144–152.
55. Breiman, L. Random forests. Mach. Learn. 2001, 45, 5–32. [CrossRef]
56. Szegedy, C.; Liu, W.; Jia, Y.; Sermanet, P.; Reed, S.; Anguelov, D.; Erhan, D.; Vanhoucke, V.; Rabinovich, A.
Going deeper with convolutions. In Proceedings of the 2015 IEEE Conference on Computer Vision and
Pattern Recognition, Boston, MA, USA, 7–12 June 2015; pp. 1–9.
57. Szegedy, C.; Vanhoucke, V.; Ioffe, S.; Shlens, J.; Wojna, Z. Rethinking the inception architecture for computer
vision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, NV,
USA, 27–30 June 2016; pp. 2818–2826.
58. Szegedy, C.; Ioffe, S.; Vanhoucke, V.; Alemi, A.A. Inception-v4, inception-resnet and the impact of residual
connections on learning. In Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence,
San Francisco, CA, USA, 4–9 February 2017.
44
applied
sciences
Article
Sentiment-Aware Word Embedding for Emotion
Classification
Xingliang Mao 1,†,‡ , Shuai Chang 2,‡ , Jinjing Shi 2, *, Fangfang Li 2, * and Ronghua Shi 2
1 Science and Technology on Information Systems Engineering Laboratory, National University of Defense
Technology, Changsha 410073, China; [email protected]
2 School of Computer Science and Engineering, Central South University, Changsha 410073, China;
[email protected] (S.C.); [email protected] (R.S.)
* Correspondence: [email protected] (J.S.); [email protected] (F.L.)
† Current address: National University of Defense Technology, Changsha 410073, China.
‡ These authors contributed equally to this work.

Received: 12 February 2019; Accepted: 22 March 2019; Published: 29 March 2019
Abstract: Word embeddings are effective intermediate representations for capturing semantic
regularities between words in natural language processing (NLP) tasks. We propose sentiment-aware
word embedding for emotional classification, which consists of integrating sentiment evidence within
the emotional embedding component of a term vector. We take advantage of the multiple types
of emotional knowledge, just as the existing emotional lexicon, to build emotional word vectors to
represent emotional information. Then the emotional word vector is combined with the traditional
word embedding to construct the hybrid representation, which contains semantic and emotional
information as the inputs of the emotion classification experiments. Our method maintains the
interpretability of word embeddings, and leverages external emotional information in addition to
input text sequences. Extensive results on several machine learning models show that the proposed
methods can improve the accuracy of emotion classification tasks.
Keywords: emotion classification; sentiment lexicon; text feature representation; hybrid vectorization;
sentiment-aware word embedding
1. Introduction
With the rapid increase in the popularity of social media applications, such as Twitter, a larger
amount of sentiment data is being generated. Emotional analysis has attracted much attention.
At the same time, sentiment analysis for Chinese social network data has been gradually developed.
In 2013, the second CCF (China Computer Federation) International Conference on Natural Language
Processing and Chinese Computing (NLPCC) established the task of evaluating the emotions of Weibo,
which attracted many researchers and institutions. The conference drove the development of emotional
analysis in China. Weibo sites have been the main communication tool. They provide information that
is more up-to-date than conventional news sources, and this has encouraged researchers to analyze
emotional information from this data source. There are many differences between Weibo text and
traditional long text, such as movie reviews in sentiment analysis. Firstly, they are short with no
more than 140 characters. Secondly, words used in Weibo are more casual than those in official texts,
and they contain a lot of noise, such as informal text snippets. For example, there are web-popular
words, like “LanShouXiangGu” (a network language/buzzword, means feel awful and want to cry).
Web-popular words might be seen as traditional words but represent different meanings or emotions.
Finally, Chinese is largely different from English; it has more complex syntaxes and sentence structures.
This increases the difficulty of emotional analysis in Chinese.

Appl. Sci. 2019, 9, 1334
Text consists of many ordered words in the emotional analysis task. In order to process a text
document mathematically, the text document is projected into the vector space. A text document is
represented by a vector in the same vector space so that the document can be classified by a model.
The BoW (bag-of-words) model is widely used in text processing applications. It is an approach to
modeling texts numerically [1]. It processes texts regardless of word order and semantic structure and
disregards context which means that it is unable to sufficiently capture complex linguistic features.
At the same time, one of the drawbacks of BoW is its high number of dimensions and excessive sparsity.
The appearance of word embedding overcomes this shortcoming.
Recently, word embedding based approaches [2,3] have learned from low-dimensional,
continuously-valued vector representations using unsupervised methods over the large corpus.
State-of-the-art word embedding algorithms include the C and W model [4], the continuous
bag-of-words (CBOW) model, the Skip-Gram Word2Vec model [3], and the GloVe (Global Vectors for
Word Representation) model [5]. Word embedding techniques have been shown to facilitate a variety
of NLP (natural language processing) tasks, including machine translation [6], word analogy [7], POS
(part of speech) tagging [8], sequence labeling [4], named entity recognition [9], text classification [10],
speech processing [11], and so on. The principle behind these word embedding approaches is the
distributional hypothesis that “You shall know a word by the company it keeps” [12]. By leveraging
statistical information, such as word co-occurrence frequencies, the method could explicitly encode
many linguistic regularities and patterns into vectors. It produces a vector space in which each unique
word in the corpus is assigned a corresponding vector in the space, and words with similar contexts in
the training corpus are located in close proximity. The feasibility of distribution assumptions has been
confirmed in many experiments [13,14].
However, most existing word embedding algorithms only consider statistical information from
documents [3,5,7]. The representations learned from these algorithms are not the most effective for
emotional analysis tasks. In order to improve the performance of word embedding in emotional
analysis tasks, a method that combines the traditional word embedding and provides prior knowledge
from external sources is proposed. The knowledge of the polarity and intensity of emotional words
can be obtained via public sentiment lexicons, and this sentiment information is not directly obtained
in word co-occurrence frequencies, which can greatly enhance the performance of word embedding
for emotional analysis. For example, “happy” and “sad” might appear in the same or a similar
emotional context but represent different emotions, so it is not enough to learn the emotional
information of these two words by counting the word co-occurrence. The method proposed in
this paper builds sentiment-aware word embedding by incorporating prior sentiment knowledge into
the embedding process.
Our primary contribution is therefore to propose such a solution by making use of the external
emotional information, and propose the sentiment-aware word embedding to improve emotional
analysis. While there is an abundant literature in the NLP community on word embedding for text
representations, much less work has been devoted in comparison to hybrid representation (combining
the diverse vector representations into a single representation).The proposed sentiment-aware word
embedding is implemented by jointly embedding the word and prior emotional knowledge in
the same latent space. The method tests on the NLPCC dataset label Weibo data with seven
emotions. First, our method encodes the semantic relationships between words by traditional
word embedding. Second, the method incorporates sentiment information of words into emotional
embedding. Various combinations of word representations are used in this experiment. It is the hybrid
sentiment-aware word embedding that can encode both semantics and sentiments of words. In the
experiments, the results show that the two kinds of semantic evidence can complement each other to
improve the accuracy of identifying the correct emotion.
The paper is organized as follows. Section 2 briefly introduces the emotion analysis method.
Section 3 presents the details of the proposed methodology. Section 4 discusses the experimental
46
Appl. Sci. 2019, 9, 1334
arrangement. Section 5 summarizes the contents of the full text and discusses the future direction
of development.
2. Related Work
The study of emotional analysis is roughly summed up into three categories: Rule-based analysis,
unsupervised classification, and supervised classification.
Rule-based analysis is mainly performed together with the emotion lexicon. In English text
analysis, Kamps proposed a distance measurement method to determine the semantic polarity of
adjectives based on the synonym graph theory model. In the analysis of Chinese text, Zhu et al.
introduced a simple method based on HowNet Chinese lexicon to determine the semantic direction
of Chinese words [15]. Pan et al. identified six kinds of emotion expressed by Weibo with the
lexicon-based method [16]. However, the lexicon-based method is low in accuracy, and the classifying
quality is easily limited by the lexicon [17]. In particular, the lexicon-based method ignores contextual
information.
Unsupervised classification analysis does not use tagged documents but relies on a
documentation’s statistical properties, NLP processes, and existing vocabulary, which has an emotional
or polarizing tendency. Turney presented a simple, unsupervised learning algorithm for classifying
reviews [18]. Lin and He proposed a novel, probabilistic modeling framework based on latent Dirichlet
allocation (LDA), called the joint sentiment/topic model (JST), which detects sentiments and topics
simultaneously from the text [19]. They also explored various ways to obtain prior information to
improve the accuracy of emotional detection. Yili Wang and Hee Yong Youn [20] proposed a novel
feature weighting approach for the sentiment analysis of Twitter data and a fine-grained feature
clustering strategy to maximize the accuracy of the analysis.
The analysis based on supervised classification generates an emotion classification model
with labeled training data. The effectiveness of supervised technologies depends on the features
used in the classification task. The bag-of-words features and their weighting scheme are widely
used in natural language processing, which provides a simplified representation of documents
through various features. However, these methods have limitations in the task of emotional analysis.
Word embedding drives the development of many NLP tasks through the low-dimensional continuous
vector representations of words.
In the framework of the word embedding model, the word vector is generated according to
the distribution hypothesis [12]. It has been found that learning vectors can clearly encode many
linguistic regularities and patterns [21]. However, it is still not enough to rely solely on word-level
distribution information collected from the text corpus to learn high-quality representations [22,23].
Auxiliary information has been shown to help learn task-specific word embedding to improve the
performance in the tasks [24,25].
In order to improve the representation of word embedding, some research work has been proposed
to incorporate various additional resources into the learning framework of word representation.
Some knowledge-enhancing word embedding models incorporate lexical knowledge resources
into the training process of word embedding models [22,23,26,27]. In the study of Levy and
Goldberg [28], the grammatical context of the automatically generated dependency analysis tree
was used for word representation training. Meanwhile, some people learn cross-language word
embedding with a multilingual parallel corpus [29–31]. Fuji Ren and Jiawen Deng pointed out that
background knowledge is composed of keywords and co-occurring words that are extracted from
the external corpus and proposed a background knowledge-based multi-stream neural network [32].
Many effective analysis algorithms take advantage of existing knowledge to improve classification
performance [33–35]. Wang et al. [33] studied the problem of understanding human sentiments from
the large-scale collection of internet images based on both image features and contextual social network
information and proved that both visual feature-based and text-based sentiment analysis approaches
can learn high-quality models. Tang et al. proposed a learning sentiment-specific word embedding
47
Appl. Sci. 2019, 9, 1334
approach dubbed sentiment embedding, which retains the effectiveness of word contexts and exploits
the sentiment of text to learn more powerful, continuous word representations [36].
3. Methodology
Constructing an effective features vector to represent text is a basic component in the NLP tasks.
In view of the specific emotion classification task, we propose the sentiment-aware word embedding
based on the construction of a hybrid word vector method containing emotional information.
This study introduces all of the details about constructing the hybrid feature representations in
the following sections (as shown in Figure 1).

"

!
Figure 1. Framework for the proposed method.
The method comprises three main components: (1) the construction of semantic word vectors
based on Word2Vec; (2) the construction of emotional word vectors based on the emotional lexicon;
and (3) the construction of hybrid sentiment-aware word representations. The third step is based on
the results of the previous two steps.
3.1. Constructing a Semantic Word Vector Based on Word2Vec

Word2Vec is an open source tool based on deep learning. It has become more and more popular
recently because of its high accuracy in analyzing semantic similarities between two words and
the relatively low computational cost. It has two modes: Continuous bag of words (CBOW) and
Skip-Gram, which can be applied to quickly learn word embedding from the original text and capture
word relationships with the built vector representation model (neural network model). In this paper,
the Skip-Gram model was selected for word embedding training. After the text preprocessing step,
the word vector representations of all words in the document are learned by the Skip-Gram model,
and the vector representation for each word encountered in the input is extracted from the model.
Then, the representation of each sentence can be received by averaging over the vectors of all its
comprising words. The vector representation of the new document is derived in the same way during
the sentiment prediction phase.
3.2. Constructed Emotional Word Vector Based on the Emotional Lexicon

After the preprocessing step, every word in each sentence is converted into a vector. The obtained
vectors are then compared based on their cosine similarity degree with the vectors of the emotional
words. The emotional word vectors with high similarity scores are selected to be combined.
The proposal helps to increase the “semantic power” of traditional semantic space models for emotional
analysis by combining different sources of semantic evidence. The hybrid sentiment-aware word
48
Appl. Sci. 2019, 9, 1334
embedding is inputted as the experimental model, showing that two kinds of semantic evidence can
complement each other, and the mix of them can identify the correct emotion.
The lexicon-based feature extraction method is promising due to the existence of the emotional
lexicon. Typically, an emotional lexicon consists of a set of language-specific words, which include
information about the emotional category to which it belongs, the polarity intensity, and so on.
A fine-grained multi-emotional polarity lexicon can improve the classification accuracy in emotional
analysis tasks compared to simple emotional dictionaries [37].
A set of documents and emotional lexicons is given to build the model and generate the emotional
word vectors. A corpus, D, consists of a set of texts, D = {d1 , d2 · · · · , dn }, and the vocabulary,
T = {t1 , t2 · · · · , tm }, which are unique terms extracted from D. The word representation of the terms
ti are mapped from the Word2Vec model, and then a set of word representations of all words in the
vocabulary is derived, i.e., V = {v1 , v2 · · · · , vm } ∈ Rm×d , where m is the size of the vocabulary, and d
represents the dimensions of the vector space.
The core of this method is to construct emotional word vectors representing the emotional
information for each word in the vocabulary, but most of them express some emotions that are not
typical opinion words, so the paper introduces the lexicon-based method. In order to construct the
emotional word vectors, the method utilizes all emotional words in the lexicon as the emotional

vocabulary, E = [ E1 , E2 , ....Ek ], and gets their word representations, VE = VE1 , VE2 , · · · , VEk . Due to
the scale problem of word embedding the model by training the corpus, there is a low degree of
the coverage problem, that is, there are some words that are in the emotional vocabulary but not
in the vector spaces learned from Word2Vec. Thus, these words can be ignored and deleted in
subsequent processing.
The word embedding model captures the semantic relationships in the corpus text, and the
contexts of words with the same emotional polarity are identical. Therefore, the similarity between
two items is estimated by calculating the cosine similarity between the vectors represented by V and E,
that is, sim (V, E) = consine (V, E), defined as

∑id=1 Vi ∗ VEi
sim (V, E) = (1)
∑id=1 Vi2 ∗ ∑id=1 VE2i
where V and E are the vectors of length d.

For each item in T, the similarity to all items in E is calculated and then the similarity results
are collected into the matrix Y ∈ Rm×k , where m is the length of the text glossary, k is the length of
the emotional vocabulary, and Yi,j indicates the cosine similarity between the lexical item i and the
emotion j. Based on matrix Y, the top n emotional words, Ew = {e1 , e2 , · · · en }, are selected by setting
the threshold. For words in Ew , they are the nearest neighbors to the item in T as determined by
distinguishing their sentimental differences.
Similarity in Y means that an item in E and an item in T have the same context; meanwhile,
the emotional intensity provided in the dictionary represents the emotional information of the item in
E and constructs the emotion vector for all items in T by combining the two kinds of information.
The emotional word vector for each word is calculated as a weighted sum rather than a simple
average operation:
1 n
EmotionVeci = ∑ weighti ∗ Vei . (2)
n i =1
Then, Formula (3) is used to compute the weight of every word; based on this formula, the higher
ranked nearest neighbors will receive higher weights:
weighti = Yij ∗ score j . (3)
49
Appl. Sci. 2019, 9, 1334
The Dalian University of Technology Information Retrieval (DUTIR) is used to represent sentiment
lexicon in this study, which is a Chinese ontology resource labeled by the Dalian University of
Technology Information Retrieval Laboratory [38]. It contains 27,466 words, and each word is
associated with a real-valued score of 1, 3, 5, 7, or 9. The score represents the degree of sentiment,
where 9 indicates the maximum strength, and 1 is the minimum strength. Then, we rank the emotional
words in Ew by strength scores. The score j is defined as the reciprocal rank of e j in the Ew , that is
1
score j = (4)
rank j
where rank j denotes the rank of e j generated by the intensity-based ranking process.
The weights have to be normalized to sum to one. This can be obtained with the following relation:
weighti
weighti = . (5)
∑nj=1 weight j
In the process of constructing the emotional word vector of words, the shortcomings of the
context-based word embedding result in words with opposite sentiment polarity having a fairly high
cosine similarity in the vector space. There are great errors generated in the construction of emotional
word vectors using word embedding mapping and cosine similarity, so sentimental lexicons are used
to correct the word representations mapping from the word embedding space. It is a good way
to optimize existing word vectors by using real-valued sentiment intensity scores provided by the
emotional lexicons and word vector refinement model. In this way, the words are closer to semantically
and emotionally similar words in the dictionary (that is, those with similar intensity scores) and stay
away from words that are not emotionally similar but are similar in semantics.
3.3. Constructing Hybrid Sentiment-Aware Word Embedding

In this work, sentiment-aware word embedding is constructed to represent each word by
connecting the emotional word vectors based on the lexicon and the semantic word vectors based
on the Word2Vec model. Sentiment-aware word embedding can capture the emotional orientation
of words, which is the word representation method strictly based on word embedding. In addition,
it also makes the most of the contextual and semantic expression ability of the word embedding model.
The sentiment-aware word embedding combines the emotional word vectors with semantic word
embedding to simplify combinatorial functions, which indicates that the advantages of the two models
can be combined in a single mixed representation. This paper explores different methods of vector
combination and experiments with the proposed vector combination method in Section 4. In particular,
in order to compare the advantages and limitations of various methods, a comparative study of the
two combination methods is conducted.
The first method combines the emotional word vectors with semantic words of a given word
directly, which allows two vector representations with different dimension representations:
xnew = xe xs (6)
where xe (xs ) represents the emotional word vector, (semantic word embedding).
Two vectors, xc and ye , form by linking the corresponding vectors from the original space x1 with
x2 , y1 , and y2 as well. Cosine similarity is used to estimate the similarity between two items, and the
key factor for cosine similarity is the dot product, i.e.,

y1
x c · y c = ( x1 x2 ) · ( y1 y2 ) = x1 x2 · = x1 · y1 + x2 · y2 . (7)
y2
50
Appl. Sci. 2019, 9, 1334
Thus, the cosine similarity in cascade space is determined by the linear combination of the dot
products of the vector component. Therefore, the semantic relations and emotional relations between
two words are distinguished as features.
The second method is to combine these representational spaces by addition, which requires the
two spaces to have the same dimensions, and it can be realized using simple vector addition. The value
for each word in the new space is the normalized sum of its two component spaces, i.e.,
xnew = xe + xs . (8)
From the dot product result of the vector, the direct superposition of the two vectors
which combine the characteristic components of them increases the distinction between different
emotional features.
No matter whether vector connection or vector overlay is used, the experimental results show
that the advantages of the two models can be combined in a single mixed representation by combining
word vector components generated in different ways with simple combinatorial functions.
3.4. Algorithm Implementation

According to data pre-processing, word embedding construction, emotional lexicon processing,
and word similarity computation, Algorithm 1 presents the sentiment-aware word embedding
approach to emotional analysis. From the algorithm description, it can be seen that with the first
iteration, we can get the vocabulary. With the second and third iterations, we can get the vector of
each item in the vocabulary and emotional lexicon. Based on the previous step, we can get the hybrid
sentiment-aware word embedding through the last iteration.
Algorithm 1 sentiment-aware word embedding for emotion classification

Input: a set of unclassified text documents
Output: a set of classified text documents
for each doc ∈ corpus do
for each term ∈ doc do
terms ← pre-processing all term with NLP methods
end
vocabulary ← terms
end
for each term ∈ vocabulary do
SemanticVectorterm ← w2vecterm
end
for each emotional_word ∈ emotional lexicon do
Vectoremotional_word ← w2vecemotional_word
end
for each term ∈ vocabulary do
for each emotional_word ∈ emotional lexicon do
sim(term, emotional_word) = consine(term, emotional_word)
if sim(term, emotional_word) > Threadhold then
weight = sim(term, emotional_word) ∗ score(emotional_word)
end
end
EmotionVectorterm = ∑ weigh
emotional_word
HybridVector1term =SemanticVectorterm + EmotionVectorterm
HybridVector2term = SemanticVectorterm EmotionVectorterm
end
51
Appl. Sci. 2019, 9, 1334
4. Experiments
In this section, the proposed hybrid word vector representation method is evaluated with a series
of experiments on Weibo text datasets. Based on the above steps, the hybrid word vector is constructed
and used as an input to train classification models. The experiments select different approaches to
evaluate the proposed hybrid feature vectors. Then, a concrete discussion is given based on different
classification results.
The Word2Vec model utilizes the genism software package in Python to create the word vectors,
and the Skip-Gram, one method of Word2Vec, was selected for this work. Word2Vec works in a
way that is comparable to the machine learning approach of deep neural networks, which allows the
representation of semantically similar words in the same vector space with adjacent points.
4.1. Description of Datasets

The data in this paper comes from Weibo (a popular Chinese social networking site). The training
corpus for word embedding learning includes 10 million unlabeled blogs of COAE 2014 (The Sixth
Chinese Opinion Analysis Evaluation) and 40 thousand blogs labeled using the emotion category in
the task of NLPCC2018. The text of Microblog is labeled “none” if it does not convey any emotion.
If the text conveys emotion, it is labeled with emotion categories from happiness, trust, anger, sadness,
fear, disgust, or surprise. The number of sentences for each emotion category in training data and test
data are described in Table 1. We can see from the tables that the distribution of different emotion
classes is not balanced.
Table 1. Examples of the emotional vocabulary ontology format.
Training Data Test Data

Emotion Type
Number Percentage Number Percentage
HAPPINESS 5200 13% 3000 15%
TRUST 5200 13% 2200 11%
ANGER 3600 9% 2000 10%
SADNESS 2800 7% 1600 8%
FEAR 2400 6% 1000 5%
DISGUST 4800 12% 2000 10%
SURPRISE 4000 10% 2200 11%
none 12,000 30% 6000 30%
This paper selected DUTIR, which is a Chinese ontology resource collated and labeled by the
Dalian University of Technology Information Retrieval Laboratory, as emotional lexicons. There are
seven kinds of emotion in the lexical ontology, happiness, trust, anger, sadness, fear, disgust and
surprise, a total of 27,466 emotional words, and five emotional intensities, 1, 3, 5, 7, and 9, where 9
indicates the maximum strength and 1 is the least strength. This resource describes a Chinese word
or phrase from different perspectives, including lexical categories, emotional categories, emotional
intensity, and polarity of words.
In the Chinese Emotion Word Ontology, the general format is shown in Table 2:
Table 2. Examples of the emotional vocabulary ontology format. Part of speech (POS).
Word POS Sense Sense Number Emotional Subcategory Strength Polarity

Happy adj 2 1 happy 5 1
sad adj 1 1 sad 5 0
The emotional categories are shown in Table 3.
52
Appl. Sci. 2019, 9, 1334
Table 3. Emotional category.
Emotional Categories Subcategories Example Words

HAPPINESS happy joy, hilarity
relieved relieved, innocent
TRUST respect polite, reverent
praise handsome, excellent
believe trust, reliability
favorite admiration, love
wish longing for, blessing
ANGER anger annoyed, furious
SADNESS sad sorrow, heart
disappointed regret, despair
guilty remorseful, unwilling,
miss lovesickness, hangs
FEAR nervous flustered, overwhelmed
dread timid, scared
shame red-faced, self-confident
DISGUST bored irritated, upset
abomination shameful, hateful
blame vanity, chaotic
jealous envy
doubt suspicious, multi-minded
SURPRISE surprise strange, miraculous
4.2. Experimental Results

In this section, we explain the use of the support vector machine (SVM), logistic regression model,
decision tree model and gradient boost model classifier to evaluate the effectiveness of the emotional
word vector.
The main content is to evaluate the possible advantages of the hybrid word vectors method
relative to other word representation methods. The proposed method tries different mixing methods
on the emotional word vector and the semantic word vector to compare them according to their
classification accuracy on all datasets.
In the emotional classification task, the current widely accepted evaluation indicators are the
precision and recall rate. The F-score, a comprehensive metric, was also selected to measure the
accuracy of the assessment analysis. The concrete calculation process can be expressed as follows:
#system_correc (emotion = i )
Precison = (9)
#system_proposed (emotion = i )
#system_correct (emotion = i )
Recall = (10)
#gold (emotion = i )
where #system_correc(emotion = i ) represents the number of texts correctly categorized into a
class, and #system_proposed(emotion = i ) represents the total number of texts classified into a class.
The calculation formula for the F-score is described as follows:
2 × Precison × Recall
F-measure = . (11)
Precison + Recall
Experiments use hybrid vectors as input into different classifiers. Here, the SVM, logistic
regression classification model, decision tree model and gradient boost model are selected, and the
experimental results can be described as follows.
From Table 4, it can be seen that the classification accuracy and reliability of positive emotions
such as happiness and trust are higher than the negative emotions such as fear and disgust. The main
53
Appl. Sci. 2019, 9, 1334
reason is that in the training corpus, the text scale of positive emotion is larger than the text scale of
negative emotion, for example, the number of texts belonging to happiness and trust accounted for a
total of 26% of the training data, while the number of texts belonging to sadness and fear accounted for
a total of 11%, which indicates that the effect of classification is related to the distribution of training
corpus to a certain extent.
Chinese Weibo text content is short with heavy colloquialism and many novel words on the
internet, which limits the process of constructing emotional word vectors based on the emotion
dictionary. Despite this, the experiment still enhanced the precision, which further proves the validity
of the experimental method. The results of this experiment are related to the quality of emotional
dictionaries to a certain extent, and emotional words with strong emotions should be chosen in the
course of the experiment.
In order to highlight the effectiveness of the emotional word vector, this study carried out
comparative experiments on different word representation methods. Specifically, the performance of
the method based on the Skip-Gram and that based on the hybrid word vectors proposed in this paper
were compared.
Table 4. Performance evaluation sentiment-aware word embedding. Hybrid method 1 is as shown in

Equation (6), hybrid method 2 is as shown in Equation (8), Prec is the abbreviation of Precision which
is as shown in Equation (9), Rec is the abbreviation of Recall which is as shown in Equation (10), F1 is
the abbreviation of F-measure which is as shown in Equation (11).
Model Method Evaluation Indicators HAPPINESS TRUST ANGER SADNESS FEAR DISGUST SURPRISE
Prec 0.7900 0.7976 0.7901 0.6364 0.6579 0.7273 0.6490
Hybrid method 1 Rec 0.8315 0.7974 0.7807 0.549 0.4902 0.7141 0.9402
F1 0.8102 0.7973 0.7793 0.5895 0.5161 0.7093 0.7679
Support vector machine (SVM)
Prec 0.7733 0.7968 0.7873 0.5714 0.5488 0.6933 0.6393
F1 0.7795 0.7964 0.7674 0.5161 0.5118 0.6685 0.7583
Prec 0.7600 0.8030 0.7742 0.6314 0.6571 0.7146 0.6377
F1 0.7600 0.8028 0.7605 0.5652 0.5349 0.7050 0.7500
Logistic Regression
Prec 0.7451 0.8030 0.7742 0.5713 0.5908 0.7190 0.6492
F1 0.7524 0.8028 0.7605 0.5162 0.5475 0.7104 0.7579
Prec 0.6935 0.6966 0.6157 0.6977 0.5116 0.6543 0.6479
F1 0.6917 0.7052 0.6421 0.5911 0.4681 0.6521 0.6682
decision_tree
Prec 0.6624 0.6522 0.6356 0.6989 0.5080 0.6552 0.6531
F1 0.6547 0.6546 0.6383 0.6000 0.4512 0.6487 0.6531
Prec 0.8164 0.7975 0.7900 0.6392 0.6453 0.6344 0.6486
F1 0.8087 0.7937 0.8009 0.7345 0.6376 0.7329 0.7245
gradient_boost
Prec 0.7906 0.7900 0.7800 0.6486 0.6420 0.6255 0.6012
F1 0.7886 0.7900 0.7792 0.7245 0.6395 0.6373 0.7172
As is shown in the Table 5, the experimental results of hybrid word vectors as classifier input are
better than the experimental results of the initial word embedding as classifier input. Among them,
hybrid method 1 is slightly better than hybrid method 2. The main reason is that mixing the
methods increases the difference between the word vectors of different emotions more significantly.
Mixing method 2 also brings about the improvement of precision, but the increase is not significant.
54
Appl. Sci. 2019, 9, 1334
Table 5. Accuracy comparison of emotion classification experiments based on various text representations.
Model Word Vector Prec Rec F1

Skip-gram 0.6685 0.6811 0.6542
svm Hybrid word vector1 0.7211 0.7290 0.7099
Hybrid word vector2 0.6871 0.7027 0.6865
Skip-gram 0.6612 0.6507 0.6674
LR Hybrid word vector1 0.7124 0.7015 0.6969
Skip-gram 0.6366 0.6100 0.6230
decision_tree Hybrid word vector1 0.6453 0.6170 0.6312
Skip-gram 0.6812 0.7511 0.7145
gradient_boost Hybrid word vector1 0.7102 0.7981 0.7475
4.3. Relation to Other Method

Multi-entity sentiment analysis using entity-level feature extraction and word embeddings
approach [39] enhance the word embeddings approach with the deployment of a sentiment
lexicon-based technique. The paper proposes associating a given entity with the adjectives, adverbs,
and verbs describing it and extracting the associated sentiment to try and infer if the text is positive
or negative in relation to the entity or entities. We discuss the major differences between Sweeney’s
model and our method: (i) Lexicon is used in different ways, Sweeney’s model uses the lexicon
against the parsed text to identify the polarity of the descriptor words (ii) Sweeney’s model uses
a Twitter-specific parser to identify the descriptor words that relate to a specific entity for text
that contains multiple entities. The descriptor words of the multi-entity tweets are scored using
SentiWordNet sentiment lexicon. The overall scoring per entity is printed out as output for the
multi-entity tweets. The remaining tweets are classified using a random forest classifier. At the end of
the article, the author points out the research aims to highlight how a hybrid word embeddings and
lexicon-based approach can be used to tackle the problem of sentiment analysis on multiple entities.
But we can see that it isolates dictionary information from word embedding techniques and does not
use so-called hybrid word embedding techniques. And our approach is the perfect way to achieve this.
We have reproduced the experiment of the paper and compared it with our method:
As is shown in the Table 6, sentiment-aware word embedding is superior to Sweeney’s model in
the use of external information, and performs better on emotion classification than Sweeney’s model.
To sum up, the hybrid word vector that combines semantic information with external emotional
information not only provides more word feature information but also involves the emotion labeling
of the word in the model prediction. The experiments showed that this method is effective, and it is
superior to the original model in terms of accuracy, recall rate, and F-score. It can be concluded that in
the emotional analysis task, the quality of the word vector can be improved by incorporating external
emotional information.
Table 6. Accuracy comparison of Sweeney’s model and sentiment-aware word embedding.
Method Prec Rec F1

SentiWordNet + POS 0.66 0.66 0.66
Word2Vec + SentiWordNet + POS 0.6913 0.6754 0.695
55
Appl. Sci. 2019, 9, 1334
5. Conclusions
Sentiment-aware word embedding was proposed for emotion classification tasks. The method
uses mixed emotional word vectors for emotional analysis, through context-sensitive word embedding
provided by Word2Vec combinations of the emotional information provided by the dictionary, which
is used as the input of the classifier model. The experiment proved that the use of hybrid word
vectors is effective for supervised emotion classification, as it greatly improves the accuracy of emotion
classification tasks.
In the future, our work will be aimed at doing some experiments in the following directions to
demonstrate the flexibility and effectiveness of the method and to further improve its performance:
(a) The method will be applied to the other language corpuses to prove its versatility, and (b) novel
ways of combining different word vector components will be explored to increase the differentiation
of features between two words.
Author Contributions: F.L. conceived the idea, designed and performed the experiments, and analyzed the
results, S.C. drafted the initial manuscript, and S.C., J.S., and X.M. revised the final manuscript. R.S. provided
experimental environment and academic guidance.
Funding: This work was supported by the National Natural Science Foundation of China (Grant Nos. 61401519,
61872390), the Natural Science Foundation of Hunan Province (Grant Nos. 2016JJ4119, 2017JJ3415), and the
Postdoctoral Science Foundation of China (Grant No. 2016M592450).
References
1. Nikhil, R.; Tikoo, N.; Kurle, S.; Pisupati, H.S.; Prasad, G. A survey on text mining and sentiment analysis for
unstructured web data. J. Emerg. Technol. Innov. Res. 2015, 2, 1292–1296.
2. Huang, E.H.; Socher, R.; Manning, C.D.; Ng, A.Y. Improving word representations via global context and
multiple word prototypes. In Proceedings of the Meeting of the Association for Computational Linguistics:
Long Papers, Jeju Island, Korea, 8–14 July 2012; pp. 873–882.
3. Mikolov, T.; Chen, K.; Corrado, G.; Dean, J. Efficient Estimation of Word Representations in Vector Space.
arXiv 2013, arXiv:1301.3781.
4. Collobert, R.; Weston, J.; Karlen, M.; Kavukcuoglu, K.; Kuksa, P. Natural Language Processing (Almost) from
Scratch. J. Mach. Learn. Res. 2011, 12, 2493–2537.
5. Pennington, J.; Socher, R.; Manning, C. Glove: Global vectors for word representation. In Proceedings
of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), Doha, Qatar,
25–29 October 2014; pp. 1532–1543.
6. Devlin, J.; Zbib, R.; Huang, Z.; Lamar, T.; Schwartz, R.; Makhoul, J. Fast and Robust Neural Network Joint
Models for Statistical Machine Translation. In Proceedings of the 52nd Annual Meeting of the Association
for Computational Linguistics, Baltimore, MD, USA, 23–25 June 2014; pp. 1370–1380.
7. Mikolov, T.; Sutskever, I.; Chen, K.; Corrado, G.; Dean, J. Distributed Representations of Words and Phrases
and their Compositionality. In Proceedings of the Advances in Neural Information Processing Systems,
Lake Tahoe, NV, USA, 5–10 December 2013; pp. 3111–3119.
8. Lin, C.-C.; Ammar, W.; Dyer, C.; Levin, L. Unsupervised pos induction with word embeddings. arXiv 2015,
arXiv:1503.06760.
9. Turian, J.; Ratinov, L.; Bengio, Y. Word representations: A simple and general method for semi-supervised
learning. In Proceedings of the Meeting of the Association for Computational Linguistics, ACL 2010,
Uppsala, Sweden, 11–16 July 2010; pp. 384–394.
10. Joulin, A.; Grave, E.; Bojanowski, P.; Mikolov, T. Bag of Tricks for Efficient Text Classification. arXiv 2016,
arXiv:1607.01759.
11. Mesnil, G.; Dauphin, Y.; Yao, K.; Bengio, Y.; Deng, L.; Hakkanitur, D.; He, X.; Heck, L.; Tur, G.; Yu, D.
Using Recurrent Neural Networks for Slot Filling in Spoken Language Understanding. IEEE/ACM Trans.
Audio Speech Lang. Process. 2015, 23, 530–539. [CrossRef]
12. Harris, Z.S. Distributional Structure; Springer: Dordrecht, The Netherlands, 1981; pp. 146–162.
13. Charles, W.G. Contextual correlates of meaning. Appl. Psycholinguist. 2000, 21, 505–524. [CrossRef]
56
Appl. Sci. 2019, 9, 1334
14. Rubenstein, H.; Goodenough, J.B. Contextual correlates of synonymy. Commun. ACM 1965, 8, 627–633.
[CrossRef]
15. Zhu, Y.; Min, J.; Zhou, Y. Semantic orientation computing based on HowNet. J. Chin. Inf. Process. 2006, 20, 14–20.
16. Pan, M.H.; Niu, Y. Emotion Recognition of Micro-blogs Based on a Hybrid Lexicon. Comput. Technol. Dev.
2014, 9, 6.
17. Cambria, E.; Schuller, B.; Xia, Y.; Havasi, C. New Avenues in Opinion Mining and Sentiment Analysis.
IEEE Intell. Syst. 2013, 28, 15–21. [CrossRef]
18. Turney, P.D. Thumbs up or thumbs down?: Semantic orientation applied to unsupervised classification of
reviews. In Proceedings of the 40th annual meeting on association for computational linguistics, Philadelphia,
PA, USA, 7–12 July 2002; pp. 417–424.
19. Lin, C.; He, Y. Joint sentiment/topic model for sentiment analysis. In Proceedings of the 18th ACM
Conference on Information and Knowledge Management, Hong Kong, China, 2–6 November 2009;
pp. 375–384.
20. Wang, Y.; Youn, H. Feature Weighting Based on Inter-Category and Intra-Category Strength for Twitter
Sentiment Analysis. Appl. Sci. 2019, 9, 92. [CrossRef]
21. Mikolov, T.; Yih, W.-t.; Zweig, G. Linguistic regularities in continuous space word representations.
In Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational
Linguistics: Human Language Technologies, Atlanta, GA, USA, 9–14 June 2013; pp. 746–751.
22. Liu, Q.; Jiang, H.; Wei, S.; Ling, Z.-H.; Hu, Y. Learning semantic word embeddings based on ordinal
knowledge constraints. In Proceedings of the 53rd Annual Meeting of the Association for Computational
Linguistics and the 7th International Joint Conference on Natural Language Processing, Beijing, China,
26–31 July 2015; pp. 1501–1511.
23. Faruqui, M.; Dodge, J.; Jauhar, S.K.; Dyer, C.; Hovy, E.; Smith, N.A. Retrofitting Word Vectors to Semantic
Lexicons. arXiv 2014, arXiv:1411.4166.
24. Liu, Y.; Liu, Z.; Chua, T.S.; Sun, M. Topical word embeddings. In Proceedings of the Twenty-Ninth AAAI
Conference on Artificial Intelligence, Austin, TX, USA, 25–30 January 2015; pp. 2418–2424.
25. Zhou, C.; Sun, C.; Liu, Z.; Lau, F.C.M. Category Enhanced Word Embedding. arXiv 2015, arXiv:1511.08629.
26. Yu, M.; Dredze, M. Improving Lexical Embeddings with Semantic Knowledge. In Proceedings of the 52nd
Annual Meeting of the Association for Computational Linguistics, Baltimore, MD, USA, 23–25 June 2014;
pp. 545–550.
27. Xu, C.; Bai, Y.; Bian, J.; Gao, B.; Wang, G.; Liu, X.; Liu, T.Y. RC-NET: A General Framework for Incorporating
Knowledge into Word Representations. In Proceedings of the ACM International Conference on Conference
on Information and Knowledge Management, Shanghai, China, 3–7 November 2014; pp. 1219–1228.
28. Levy, O.; Goldberg, Y. Dependency-Based Word Embeddings. In Proceedings of the 52nd Annual Meeting of
the Association for Computational Linguistics, Baltimore, MD, USA, 23–25 June 2014; pp. 302–308.
29. Lu, A.; Wang, W.; Bansal, M.; Gimpel, K.; Livescu, K. Deep Multilingual Correlation for Improved Word
Embeddings. In Proceedings of the Conference of the North American Chapter of the Association for
Computational Linguistics: Human Language Technologies, Denver, CO, USA, 31 May–5 June 2015;
pp. 250–256.
30. Hermann, K.M.; Blunsom, P. Multilingual Models for Compositional Distributed Semantics. arXiv 2014,
arXiv:1404.4641.
31. Zhang, J.; Liu, S.; Li, M.; Zhou, M.; Zong, C. Bilingually-constrained phrase embeddings for machine
translation. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics,
Baltimore, MD, USA, 23–25 June 2014; pp. 111–121.
32. Ren, F.; Deng, J. Background Knowledge Based Multi-Stream Neural Network for Text Classification.
33. Wang, Y.; Wang, S.; Tang, J.; Liu, H.; Li, B. Unsupervised sentiment analysis for social media images.
In Proceedings of the IEEE International Conference on Data Mining Workshop, Washington, DC, USA,
14–17 November 2015; pp. 1584–1591.
34. Hogenboom, A.; Bal, D.; Frasincar, F.; Bal, M.; Jong, F.D.; Kaymak, U. Exploiting emoticons in sentiment
analysis. In Proceedings of the 28th Annual ACM Symposium on Applied Computing, Coimbra, Portugal,
18–22 March 2013; pp. 703–710.
57
Appl. Sci. 2019, 9, 1334
35. Hu, X.; Tang, J.; Gao, H.; Liu, H. Unsupervised sentiment analysis with emotional signals. In Proceedings of
the 22nd International Conference on World Wide Web, Rio de Janeiro, Brazil, 13–17 May 2013; pp. 607–618.
36. Tang, D.; Wei, F.; Qin, B.; Yang, N.; Liu, T.; Zhou, M. Sentiment Embeddings with Applications to Sentiment
Analysis. IEEE Trans. Knowl. Data Eng. 2016, 28, 496–509. [CrossRef]
37. Carrillo-De-Albornoz, J.; Plaza, L. An emotion-based model of negation, intensifiers, and modality for
polarity and intensity classification. J. Assoc. Inf. Sci. Technol. 2013, 64, 1618–1633. [CrossRef]
38. Chen, J. The Construction and Application of Chinese Emotion Word Ontology. Master’s Thesis,
Dailian University of Technology, Dalian, China, 2008.
39. Sweeney, C.; Padmanabhan, D. Multi-entity sentiment analysis using entity-level feature extraction and
word embeddings approach. In Proceedings of the Recent Advances in Natural Language Processing,
Varna, Bulgaria, 4–6 September 2017.
58
applied
sciences
Article
A Deep Learning-Based Approach for Multi-Label
Emotion Classification in Tweets
Mohammed Jabreel 1,2, * and Antonio Moreno 1
1 ITAKA Research Group, Universitat Rovira i Virgili, 43007 Tarragona, Spain; [email protected]
2 Department of Computer Science, Hodeidah University, 1821 Hodeidah, Yemen

Abstract: Currently, people use online social media such as Twitter or Facebook to share their
emotions and thoughts. Detecting and analyzing the emotions expressed in social media content
benefits many applications in commerce, public health, social welfare, etc. Most previous work
on sentiment and emotion analysis has only focused on single-label classification and ignored the
co-existence of multiple emotion labels in one instance. This paper describes the development of
a novel deep learning-based system that addresses the multiple emotion classification problem
in Twitter. We propose a novel method to transform it to a binary classification problem and
exploit a deep learning approach to solve the transformed problem. Our system outperforms the
state-of-the-art systems, achieving an accuracy score of 0.59 on the challenging SemEval2018 Task
1:E-cmulti-label emotion classification problem.
Keywords: opinion mining; sentiment analysis; emotion classification; deep learning; Twitter
1. Introduction
Emotions are the key to people’s feelings and thoughts. Online social media, such as Twitter and
Facebook, have changed the language of communication. Currently, people can communicate facts,
opinions, emotions, and emotion intensities on different kinds of topics in short texts. Analyzing the
emotions expressed in social media content has attracted researchers in the natural language processing
research field. It has a wide range of applications in commerce, public health, social welfare, etc. For
instance, it can be used in public health [1,2], public opinion detection about political tendencies [3,4],
brand management [5], and stock market monitoring [6]. Emotion analysis is the task of determining
the attitude towards a target or topic. The attitude can be the polarity (positive or negative) or an
emotional state such as joy, anger, or sadness [7].
Recently, the multi-label classification problem has attracted considerable interest due to its
applicability to a wide range of domains, including text classification, scene and video classification,
and bioinformatics [8]. Unlike the traditional single-label classification problem (i.e., multi-class or
binary), where an instance is associated with only one label from a finite set of labels, in the multi-label
classification problem, an instance is associated with a subset of labels.
Most previous work on sentiment and emotion analysis has only focused on single-label
classification. Hence, in this article, we focus on the multi-label emotion classification task, which
aims to develop an automatic system to determine the existence in a text of none, one, or more out of
eleven emotions: the eight Plutchik [9] categories (joy, sadness, anger, fear, trust, disgust, surprise, and
anticipation) that are shown in Figure 1, plus optimism, pessimism, and love.

Appl. Sci. 2019, 9, 1123
Figure 1. The set of the eight basic emotions proposed by Plutchik [9].
One of the most common approaches to addressing the problem of multi-label classification is
the problem transformation. With this approach, a multi-label problem is transformed into one or more
single-label (i.e., binary or multi-class) problems. Specifically, single-label classifiers are learned and
employed; after that, the classifiers’ predictions are transformed into multi-label predictions.
Different transformation methods have been proposed in the multi-label literature. The most
common method is called binary relevance [10,11]. The idea of the binary relevance method is simple
and intuitive. A multi-label problem is transformed into multiple binary problems, one problem for
each label. Then, an independent binary classifier is trained to predict the relevance of one of the labels.
Although binary relevance is popular in the literature, due to its simplicity, it suffers from directly
modeling correlations that may exist between labels. However, it is highly resistant to overfitting
label combinations, since it does not expect examples to be associated with previously-observed
combinations of labels.
In this article, we propose a novel transformation method, called xy-pair-set, for the multi-label
classification problem. Unlike binary relevance methods, our method transforms the problem into
only one binary classification problem as described in Section 3. Additionally, we exploit the successes
of deep learning models, especially the word2vecmethods’ family [12] and the recurrent neural
networks [13,14] and attention models [15,16], to develop a system that solves the transformed binary
classification problem. The critical component of our system is the embedding module, which uses
three embedding models and an attention function to model the relationship between the input and
the label.
60
Appl. Sci. 2019, 9, 1123
To summarize, the contribution of this work is four-fold.
• We propose a novel transformation mechanism for the multi-label classification problem.

• We propose a novel, attentive deep learning system, which we call Binary Neural Network
(BNet), which works on the new transformation method. Our system is a data-driven, end-to-end
neural-based model, and it does not rely on external resources such as parts of speech taggers and
sentiment or emotion lexicons.
• We evaluate the proposed system on the challenging multi-label emotion classification dataset of
SemEval-2018 Task1: Affect in Tweets.
• The experimental results show that our system outperforms the state-of-the-art systems.
The rest of the article is structured as follows. In Section 2, we overview the related work on
multi-label problem transformation methods and Twitter sentiment and emotion analysis. In Section 3,
we explain in detail the methodology. In Section 4, we report the experimental results. In Section 5,
the conclusions and the future work are presented.
2. Related Works
In this section, we overview the most popular research studies related to this work. In Section 2.1,
we summarize the most common multi-label problem transformation methods. Section 2.2 gives an
overview of the state-of-the-art works on the problem of multi-label emotion classification on Twitter.
2.1. Problem Transformation Methods

Let X = { x1 , x2 , . . . xn } be the set of all instances and Y = {y1 , y2 , . . . , ym } be the set of all labels.
We can define the set of data:
D = {( xi , Ŷi )| xi ∈ X and Ŷi ⊆ Y is the set of labels associated with xi } (1)
In this expression, D is called a supervised multi-label dataset.

The task of multi-label classification is challenging because the number of label sets grows
exponentially as the number of class labels increases. One common strategy to address this issue
is to transform the problem into a traditional classification problem. The idea is to simplify the
learning process by exploiting label correlations. Based on the order of the correlations, we can group
the existing transformation methods into three approaches [17,18], namely first-order approaches,
second-order approaches, and high-order approaches.
First-order approaches decompose the problem into some independent binary classification
problems. In this case, one binary classifier is learned for each possible class, ignoring the co-existence
of other labels. Thus, the number of independent binary classifiers needed is equal to the number of
labels. For each multi-label training example ( xi , Ŷi ) ∈ D, yk ∈ Y, we construct a binary classification
training set, Dk as the following: xi will be regarded as one positive example if yk ∈ Ŷi and one
negative example otherwise. In the first case, we will get a training example in the form ( xi , 1) ∈ Dk ,
which will be ( xi , 0) ∈ Dk in the second case. Thus, for all labels {y1 , y2 , . . . , ym } ∈ Y, m training
sets { D1 , D2 , . . . , Dm } are constructed. Based on that, for each training set Dk , one binary classifier
can be learned with popular learning techniques such as AdaBoost [19], k-nearest neighbor [20],
decision trees, random forests [21,22], etc. The main advantage of first-order approaches is their
conceptual simplicity and high efficiency. However, these approaches can be less effective due to their
ignorance of label correlations.
Second-order approaches try to address the lack of modeling label correlations by exploiting
pairwise relationships between the labels. One way to consider pairwise relationships is to train
one binary classifier for each pair of labels [23]. Although second-order approaches perform well in
several domains, they are more complicated than the first-order approaches in terms of the number
61
Appl. Sci. 2019, 9, 1123
of classifiers. Their complexity is quadratic, as the number of classifiers needed is (m2 ). Moreover,
in real-world applications, label correlations could be more complex and go beyond second-order.
High-order approaches tackle the multi-label learning problem by exploring high-order
relationships among the labels. This can be fulfilled by assuming linear combinations [24], a nonlinear
mapping [25,26], or a shared subspace over the whole label space [27]. Although high-order approaches
have stronger correlation-modeling capabilities than their first-order and second-order counterparts,
these approaches are computationally demanding and less scalable.
Our transformation mechanism, shown in Section 3.1, is a simple as the first-order approaches
and can model, implicitly, high-order relationships among the labels if some requirements, detailed in
Section 3.1, are fulfilled. It requires only one binary classifier, and the number of training examples
grows polynomially in terms of the number of instances and the number of labels. If the number of
training examples in the multi-label training dataset is n and the number of the labels is m, then the
number of the training examples in the transformed binary training set is n × m.
2.2. Emotion Classification in Tweets

Various machine learning approaches have been proposed for traditional emotion classification
and multi-label emotion classification. Most of the existing systems solve the problem as a text
classification problem. Supervised classifiers are trained on a set of annotated corpora using a different
set of hand-engineered features. The success of such models is based on two main factors: a large
amount of labeled data and the intelligent design of a set of features that can distinguish between the
samples. With this approach, most studies have focused on engineering a set of efficient features to
obtain a good classification performance [28–30]. The idea is to find a set of informative features to
reflect the sentiments or the emotions expressed in the text. Bag-of-Words (BoW) and its variation,
n-grams, is the representation method used in most text classification problems and emotion analysis.
Different studies have combined the BoW features with other features such as the parts of speech tags,
the sentiment and the emotion information extracted from lexicons, statistical information, and word
shapes to enrich the text representation.
Although BoW is a popular method in most text classification systems, it has some drawbacks.
Firstly, it ignores the word order. That means that two documents may have the same or a very
close representation as far as they have the same words, even though they carry a different meaning.
The n-gram method resolves this disadvantage of BoW by considering the word order in a context of
length n. However, it suffers from sparsity and high dimensionality. Secondly, BoW is scarcely able
to model the semantics of words. For example, the words beautiful, wonderful and view have an equal
distance in BoW, where the word beautiful is closer to the word wonderful than the word view in the
semantic space.
Sentiment and emotion lexicons play an essential role in developing efficient sentiment and
emotion analysis systems. However, it is difficult to create such lexicons. Moreover, finding the best
combination of lexicons in addition to the best set of statistical features is a time-consuming task.
Recently, deep learning models have been utilized to develop end-to-end systems in many tasks
including speech recognition, text classification, and image classification. It has been shown that such
systems automatically extract high-level features from raw data [31,32].
Baziotis et al. [33], the winner of the multi-label emotion classification task of SemEval-2018 Task1:
Affect in Tweets, developed a bidirectional Long Short-Term Memory (LSTM) with a deep attention
mechanism. They trained a word2vec model with 800,000 words derived from a dataset of 550 million
tweets. The second place winner of the SemEval leaderboard trained a word-level bidirectional LSTM
with attention, and it also included non-deep learning features in its ensemble [34]. Ji Ho Park et al. [35]
trained two models to solve this problem: regularized linear regression and logistic regression classifier
chain [11]. They tried to exploit labels’ correlation to perform multi-label classification. With the first
model, the authors formulated the multi-label classification problem as a linear regression with label
distance as the regularization term. In their work, the logistic regression classifier chain method was
62
Appl. Sci. 2019, 9, 1123
used to capture the correlation of emotion labels. The idea is to treat the multi-label problem as a
sequence of binary classification problems by taking the prediction of the previous classifier as an extra
input to the next classifier.
In this work, we exploited the deep learning-based approach to develop a system that can extract
a high-level representation of the tweets and model an implicit high-order relationship among the
labels. We used the proposed system alongside the proposed transformation method to train a function
that can solve the problem of multi-label emotion classification in tweets. The next section explains the
details of our proposed system.
3. Methodology
This section shows the methodology of this work. First, we explain in Section 3.1 the proposed
transformation method, xy-pair-set. Afterwards, we describe the proposed system in Section 3.2.
3.1. xy-Pair-Set: Problem Transformation

The proposed transformation method xy-pair-set transforms a multi-label classification dataset D
into a supervised binary dataset D̂ as follows:
∀ xi ∈ X, y ∈ Y and ( xi , Ŷi ) ∈ D, ∃!(( xi , y), φ) ∈ D̂

1 if y ∈ Ŷi (2)
where φ =
0 otherwise
Algorithm 1 explains the implementation of the proposed transformation method. It takes as

inputs a multi-label dataset D (Equation (1)) and a set of labels Y, and it returns a transformed binary
dataset. We show next an illustrative example.
Algorithm 1: xy-pair-set algorithm.

Input: Input: a multi-label classification dataset D and a set of labels Y
Output: Output: a binary classification dataset D̂
1 D̂ = {};
2 foreach ( xi , Ŷi ) ∈ D do
3 foreach y ∈ Y do
4 x̂i = ( xi , y) ; a tuple of xi and y.
5 if y ∈ Ŷi then

6 D̂ = D̂ ( x̂i , 1) ;
7 else

8 D̂ = D̂ ( x̂i , 0) ;
9 end
10 end
11 end
12 return D̂
Let X = {x1, x2}, Y = {a, b, c} and D = {(x1, {a, c}), (x2, {b})}. The output of the
binary relevance transformation method is a set of three independent binary datasets, one for
each label. That is, Da = {(x1, 1), (x2, 0)}, Db = {(x1, 0), (x2, 1)}, and Dc = {(x1, 1), (x2, 0)}.
In contrast, the output of our transformation method is a single binary dataset D̂ =
{((x1, a), 1), ((x1, b), 0), ((x1, c), 1), ((x2, a), 0), ((x2, b), 1), ((x2, c), 0)}.
The task in this case, unlike the traditional supervised binary classification algorithms, is to develop
a learning algorithm to learn a function g : X × Y → {0, 1}. The success of such an algorithm is based
on three requirements: (1) an encoding method to represent an instance x ∈ X as a high-dimensional
63
Appl. Sci. 2019, 9, 1123
vector Vx , (2) a method to encode a label y ∈ Y as a vector Vy , and (3) a method to represent the relation
between the instance x and the label y. These three conditions make g able to capture the relationships
inputs-to-labels and labels-to-labels. In this work, we take advantage of the successes of deep learning
models to fulfill the three requirements listed above. We empirically show the success of our system with
respect to these conditions as reported in Sections 4.6 and 4.7.
3.2. BNet: System Description

This subsection explains the proposed system to solve the transformed binary problem mentioned
above. Figure 2 shows the graphical depiction of the system’s architecture. It is composed of three
parts: the embedding module, the encoding module, and the classification module. We explain in
detail each of them below.
Figure 2. An illustration of the proposed system. Shaded parts are trainable. Fw and Bw refer to the
Forward and Backward cells respectively, and FF means Feed-Forward layer.
3.2.1. Embedding Module

Let (W, y) be the pair of inputs to our system, where W = {w1 , w2 , . . . , wl } is the set of the words
in a tweet and y is the label corresponding to an emotion. The goal of the embedding module is to
represent each word wi by a vector vwi and the label by a vector vy .
Our embedding module can be seen as a function that maps a query and a set of key-value pairs
to an output, where the query, keys, values, and output are all vectors. The output is computed as a
weighted sum of the values, where the weight assigned to each value is computed by a compatibility
function of the query with the corresponding key. The query is the trainable label embedding, EQ (y);
the keys are the pretrained words embeddings, EK (wi ) ∀wi ∈ W; and the values are the trainable
words embeddings, EV (wi ) ∀wi ∈ W.
As shown in Figure 2, we used the output of EQ and EK as inputs to the attention model to find
the alignments, i.e., the weights α, between the label y and the words W of the input tweet. This step
models the relation between the input and the label. As soon as the weights were obtained, we then
multiplied each word’s vector that came from the embedding EV by its corresponding weight. Given
that, the final representation of a word wi ∈ W is given as the following:
v wi = E V ( wi ) · α ( wi , y ) (3)
The function α is an attention-based model, which finds the strength of the relationship between
the word wi and the label y based on their semantic similarity. That is, α(wi , y) is a value based on the
distance Δ(wi , y) between wi and y as:
64
Appl. Sci. 2019, 9, 1123
eΔ(wi ,y)
α ( wi , y ) = Δ(w j ,y)
(4)
∑ w j ∈W e
Here, Δ(wi , y) is a scalar score that represents the similarity between the word wi and the label y:
Δ ( wi , y ) = E k ( wi ) · v y T (5)
vy = E Q (y) (6)
It is worth noting that α(wi , y) ∈ [0, 1] and:
∑ α ( wi , y ) = 1 (7)
w i ∈W
3.2.2. Encoding Module

The goal of the encoding module is to map the sequence of word representations
{vw1 , vw2 , . . . , vwl } that is obtained from the embedding module to a single real-valued dense vector.
In this work, we used a Recurrent Neural Network (RNN) to design our encoder. RNN reads the input
sequence of vectors in a forward direction (left-to-right) starting from the first symbol vw1 to the last one
vwl . Thus, it processes sequences in temporal order, ignoring the future context. However, for many
tasks on sequences, it is beneficial to have access to future, as well as to past information. For example,
in text processing, decisions are usually made after the whole sentence is known. The Bidirectional
Recurrent Neural Network (BiRNN) variant [13] proposed a solution for making predictions based on
both past and future information.
−
→ ←−
A BiRNN consists of forward φ and backward φ RNNs. The first one reads the input sequence
−
→ −
→
in a forward direction and produces a sequence of forward hidden states ( h1 , . . . , hl ), whereas the
former reads the sequence in the reverse order (vwl , . . . , vw1 ), resulting in a sequence of backward
←
− ←−
hidden states ( hl , . . . , h1 ).
We obtained a representation for each word vwt by concatenating the corresponding forward
−
→ ←−
hidden state ht and the backward one ht . The following equations illustrate the main ideas:
−
→ − → −−→
h t = φ ( v w t , h t −1 ) (8)
←
− ← − ←−−
h t = φ ( v w t , h t +1 ) (9)
−→ ← −
ht = [ ht ; ht ] (10)
The final input representation of the sequence is:
c = F ({h1 , h2 , . . . , hl }) (11)
We simply chose F to be the last hidden state (i.e., F ({h1 , h2 , . . . , hn }) = hn ).

−
→ ←−
In this work, we used two Gated Recurrent Units (GRUs) [14], one as φ and the other as φ .
This kind of RNN was designed to have more persistent memory, making them very useful to capture
long-term dependencies between the elements of a sequence. Figure 3 shows a graphical depiction of
a gated recurrent unit.
A GRU has reset (rt ) and update (zt ) gates. The former can completely reduce the past hidden state
ht−1 if it finds that it is irrelevant to the computation of the new state, whereas the later is responsible
for determining how much of ht−1 should be carried forward to the next state ht .
65
Appl. Sci. 2019, 9, 1123
Figure 3. Gated Recurrent Unit (GRU).
The output ht of a GRU depends on the input xt and the previous state ht−1 , and it is computed
as follows:
rt = σ (Wr · [ ht−1 ; xt ] + br ) (12)

zt = σ (Wz · [ ht−1 ; xt ] + bz ) (13)

ht = tanh (Wh · [(rt
ht−1 ); xt ] + bh ) (14)
h t = (1 − z t )
h t −1 + z t

ht (15)
In these expressions, rt and zt denote the reset and update gates, ht is the candidate output state,
and ht is the actual output state at time t. The symbol
stands for element-wise multiplication; σ is a
sigmoid function; and; stands for the vector-concatenation operation. Wr , Wz , Wh ∈ Rdh ×(d+dh ) and
br , bz , bh ∈ Rdh are the parameters of the reset and update gates, where dh is the dimension of the hidden
state and d is the dimension of the input vector.
3.2.3. Classification Module

Our classifier was composed of two feed-forward layers with the ReLU activation function
followed by a Sigmoid unit.
4. Experiments and Results

In this section, we first describe the experimental details, and then, we describe the dataset and
the pre-processing we used. Afterwards, we introduce the state-of-the-art systems we compared our
system with, and finally, we report the empirical validation proving the effectiveness of our system.
4.1. Experimental Details

Table 1 shows the hyperparameters of our system, which was trained using Adam [36], with
a learning rate of 0.0001, β 1 = 0.5, and a mini-batch size of 32 to minimize the binary cross-entropy
loss function:
L(θ, D̂ ) = −E(( x ,y ),φ )∼ D̂ [φi · g( xi , yi ) + (1 − φi ) · (1 − g( xi , yi ))]
i i i
(16)
where, g( xi , yi ) is the predicted value, φi is the real value, and θ is the model’s parameters.
The hyperparameters of our system were obtained by applying Bayesian optimization [37].
We used the development set as a validation set to fine-tune those parameters.
66
Appl. Sci. 2019, 9, 1123
Table 1. Hyperparameters of our system.
Parameter Value
EQ :
Dimensions: 11 × 310
Initialization: Uniform (−0.02, 0.02)
Trainable: Yes
EK :
Embedding Module Dimensions: 13,249 × 310
Initialization: Pretrained model (we used the pretrained embeddings provided in [33]).
Trainable: No
EV :
Dimensions: 13,249 × 310
Initialization: Uniform (−0.02, 0.02)
Trainable: Yes
RNN Cell: GRU
Hidden size: 200
Encoding Module Layers: 2
Encoding: last hidden state
RNN dropout: 0.3
FF1: 1024 units
FF2: 512 units
Classification Module Sigmoid: 1 unit
Activation: ReLU
Dropout: 0.3
4.2. Dataset
In our experiments, we used the multi-label emotion classification dataset of SemEval-2018 Task1:
Affect in Tweets [30]. It contains 10,983 samples divided into three splits: training set (6838 samples),
validation set (886 samples), and testing set (3259 samples). For more details about the dataset, we refer
the reader to [38]. We trained our system on the training set and used the validation set to fine-tune
the parameters of the proposed system. We pre-processed each tweet in the dataset as follows:
• Tokenization: We used an extensive list of regular expressions to recognize the following meta
information included in tweets: Twitter markup, emoticons, emojis, dates, times, currencies,
acronyms, hashtags, user mentions, URLs, and words with emphasis.
• As soon as the tokenization was done, we lowercased words and normalized the recognized
tokens. For example, URLs were replaced by the token “<URL>”, and user mentions were
replaced by the token “<USER>”. This step helped to reduce the size of the vocabulary without
losing information.
4.3. Comparison with Other Systems

We compared the proposed system with the state-of-the-art systems used in the task of multi-label
emotion classification, including:
• SVM-unigrams: a baseline support vector machine system trained using just word unigrams as
features [30].
• NTUA-SLP: the system submitted by the winner team of the SemEval-2018
Task1:E-cchallenge [33].
• TCS: the system submitted by the second place winner [34].
• PlusEmo2Vec: the system submitted by the third place winner [35].
• Transformer: a deep learning system based on large pre-trained language models developed by
the NVIDIA AI lab [39].
67
Appl. Sci. 2019, 9, 1123
4.4. Evaluation Metrics

We used multi-label accuracy (or Jaccard index), the official competition metric used by the
organizers of SemEval-2018 Task 1: Affect in Tweets, for the E-c sub task, which can be defined as the
size of the intersection of the predicted and gold label sets divided by the size of their union.
1 Gt ∩ Pt
| T | t∑
Jaccard = (17)
∈ T Gt ∪ Pt
In this expression, Gt is the set of the gold labels for tweet t, Pt is the set of the predicted labels
for tweet t, and T is the set of tweets. Additionally, we also used the micro-averaged F-score and the
macro-averaged F-score.
Let #c (l ) denote the number of samples correctly assigned to the label l, # p (l ) the number of
samples assigned to l, and #(l ) the number of actual samples in l. The micro-averaged F1-score is
calculated as follows:
∑ #c ( l )
Pmicro = l ∈ L (18)
∑l ∈ L # p (l )
∑ l ∈ L #c ( l )
Rmicro = (19)
∑ l ∈ L #( l )
2 × Pmicro × Rmicro
F1micro = (20)
Pmicro + Rmicro
Thus, Pmicro is the micro-averaged precision score, and Rmicro is the micro-averaged recall score.
Let Pl , Rl , and Fl denote the precision score, recall score, and the F1-score of the label l.
The macro-averaged F1-score is calculated as follows:
#c ( l )
Pl = (21)
# p (l )
#c ( l )
Rl = (22)
#( l )
2 × Pl × Rl
Fl = (23)
Pl + Rl
1
| L| l∑
F1macro = Fl (24)
∈L
4.5. Results
We submitted our system’s predictions to the SemEval Task1:E-C challenge. The results were
computed by the organizers on a golden test set, for which we did not have access to the golden labels.
Table 2 shows the results of our system and the results of the compared models (obtained from
their associated papers). As can be observed from the reported results, our system achieved the top
Jaccard index accuracy and macro-averaged F1 scores among all the state-of-the-art systems, with a
competitive, but slightly lower score for the micro-average F1.
To get more insight about the performance of our system, we calculated the precision score, the
recall score, and the F1 score of each label. The results of this analysis are shown in Figure 4. We found
that our system gave the best performance on the “joy” label followed by the “anger”, “fear”, “disgust”,
and “optimism” labels. The obtained F1-score of these labels was above 70%. The worst performance
was obtained on the “trust”, “surprise”, “anticipation”, and “pessimism” labels. In most cases, our
system gave a recall score higher than the precision score. It seems that the system was aggressive
against the emotions “trust”, “surprise”, “anticipation”, and “pessimism” (i.e., the system associated a
68
Appl. Sci. 2019, 9, 1123
low number of samples to these labels). This can be attributed to the low number of training examples
for these emotions and to the Out-Of-Vocabulary (OOV) problem.
Table 2. Results of our system and state-of-the-art systems. The best values are in bold.
Model Accuracy (Jaccard) Micro F1 Macro F1

BNet(Our System) 0.590 0.692 0.564
SVM-Unigrams 0.442 0.57 0.443
Transformer 0.577 0.690 0.561
NTUA-SLP 0.588 0.701 0.528
TCS 0.582 0.693 0.530
PlusEmo2Vec 0.576 0.692 0.497
1
0.9
0.8
0.7
0.6
Scores
0.5
0.4
0.3
0.2
0.1
0
Emotions
Precision Recall F1-Score
Figure 4. Performance analysis.
4.6. Attention Visualizations

We visualized the attention weights to get a better understanding of the performance of our
system. The results are described in Figures 5–8, which show heat-maps of the attention weights on
the top four example tweets from the validation set. The color intensity refers to the weight given to
each word by the attention model. It represents the strength of the relationship between the word and
the emotion, which reflects the importance of this word in the final prediction. We can see that the
attention model gave the important weights to the common words, such as the stop words, in case the
tweet was not assigned to the emotion; for example, the word “for” in Figure 5 and the word “this”
in Figure 7 and the token “<user>” in Figure 8. Moreover, it also gives a high weight for the words
and the emojis related to emotions (e.g., “cheering” and “awesome” for joy, “birthday” for love, etc.).
An interesting observation is that when emojis were present, they were almost always selected as
important if they were related to the emotion. For instance, we can see in Figure 7 that the sadness
emotion relied heavily on the emoji. We also found that considering only one word to model the
relation between the tweet and the emotions was not enough. In some cases, the emotion of a word
may be flipped based on the context. For instance, consider the following tweet as an example: “When
being #productive (doing the things that NEED to be done), #anxiety level decreases and #love level
increases. #personalsexuality”, the word “anxiety” is highly related to the emotion fear, but in this
context, it shows optimism and trust emotions. However, our system misassociated this example with
the fear emotion.
69
Appl. Sci. 2019, 9, 1123
im clapping and cheering for both teams . ..1 ..2
Anger 0.43 0.10 0.18 0.01 0.00 0.01 0.02 0.08 0.08 0.08
Anticipation 0.03 0.02 0.05 0.04 0.28 0.07 0.19 0.11 0.11 0.11
Disgust 0.15 0.08 0.26 0.01 0.01 0.02 0.04 0.14 0.14 0.14
Fear 0.01 0.01 0.02 0.00 0.23 0.01 0.00 0.24 0.24 0.24
Joy 0.01 0.07 0.00 0.74 0.13 0.04 0.01 0.00 0.00 0.00
Love 0.00 0.00 0.00 0.01 0.97 0.00 0.00 0.01 0.01 0.01
Optimism 0.01 0.01 0.02 0.13 0.50 0.18 0.09 0.02 0.02 0.02
Pessimism 0.01 0.01 0.02 0.01 0.30 0.01 0.03 0.20 0.20 0.20
Sadness 0.03 0.02 0.03 0.02 0.21 0.02 0.02 0.21 0.21 0.21
Surprise 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00
Trust 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00
Figure 5. Attention visualization example. Golden labels are {joy, optimism} and predicted labels are
{joy (0.91), optimism (0.51)}.
i got a free dr . pepper from the vending machine awesome
Anger 0.02 0.03 0.28 0.04 0.01 0.12 0.07 0.04 0.33 0.02 0.04 0.01
Anticipation 0.28 0.03 0.04 0.06 0.27 0.07 0.04 0.01 0.09 0.04 0.05 0.02
Disgust 0.02 0.03 0.22 0.04 0.02 0.14 0.04 0.09 0.31 0.04 0.04 0.01
Fear 0.11 0.00 0.16 0.00 0.00 0.12 0.00 0.00 0.58 0.00 0.01 0.00
Joy 0.02 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.95
Love 0.76 0.01 0.01 0.01 0.00 0.01 0.00 0.00 0.04 0.00 0.00 0.16
Optimism 0.33 0.10 0.01 0.06 0.13 0.02 0.03 0.01 0.02 0.03 0.02 0.23
Pessimism 0.08 0.01 0.20 0.01 0.01 0.18 0.01 0.01 0.46 0.01 0.03 0.00
Sadness 0.07 0.02 0.19 0.01 0.04 0.19 0.04 0.04 0.34 0.04 0.02 0.01
Surprise 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Trust 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Figure 6. Attention visualization example. Golden labels are {joy, surprise} and predicted labels are
{joy (0.97), optimism (0.87)}.
can not believe zain starting secondary this year ᰁ
Anger 0.25 0.13 0.02 0.02 0.02 0.06 0.49 0.00 0.02
Anticipation 0.09 0.03 0.06 0.17 0.34 0.17 0.04 0.08 0.02
Disgust 0.14 0.17 0.04 0.03 0.04 0.10 0.45 0.01 0.03
Fear 0.19 0.02 0.02 0.01 0.02 0.03 0.69 0.01 0.01
Joy 0.04 0.00 0.10 0.12 0.12 0.02 0.01 0.57 0.03
Love 0.13 0.01 0.08 0.14 0.01 0.00 0.21 0.08 0.34
Optimism 0.03 0.01 0.19 0.26 0.12 0.07 0.02 0.27 0.02
Pessimism 0.34 0.06 0.03 0.03 0.02 0.04 0.35 0.01 0.13
Sadness 0.01 0.02 0.01 0.01 0.00 0.01 0.03 0.00 0.91
Surprise 0.14 0.00 0.00 0.00 0.00 0.00 0.85 0.00 0.00
Trust 0.34 0.00 0.00 0.00 0.00 0.00 0.65 0.00 0.00
Figure 7. Attention visualization example. Golden labels are {sadness, surprise} and predicted labels are
{love (0.74), sadness (0.98)}.
70
Appl. Sci. 2019, 9, 1123
<user> happy birthday gorg , have a good one ᯷ x
Anger 0.05 0.00 0.00 0.01 0.27 0.33 0.21 0.00 0.09 0.00 0.03
Anticipation 0.49 0.01 0.04 0.01 0.05 0.14 0.06 0.03 0.07 0.04 0.07
Disgust 0.04 0.00 0.00 0.01 0.34 0.32 0.16 0.00 0.10 0.00 0.03
Fear 0.60 0.00 0.00 0.00 0.07 0.03 0.28 0.00 0.01 0.00 0.01
Joy 0.01 0.67 0.08 0.09 0.00 0.00 0.00 0.09 0.00 0.06 0.00
Love 0.02 0.01 0.73 0.04 0.00 0.00 0.00 0.00 0.00 0.20 0.00
Optimism 0.10 0.29 0.09 0.04 0.00 0.00 0.01 0.33 0.01 0.11 0.01
Pessimism 0.40 0.00 0.00 0.00 0.34 0.05 0.17 0.01 0.01 0.01 0.02
Sadness 0.16 0.01 0.03 0.02 0.42 0.16 0.13 0.01 0.02 0.01 0.02
Surprise 0.98 0.00 0.00 0.00 0.00 0.00 0.02 0.00 0.00 0.00 0.00
Trust 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Figure 8. Attention visualization example. Golden labels are {joy, love, optimism} and predicted labels
are {joy (0.98), love (0.91) optimism (0.95)}.
4.7. Correlation Analysis

Figure 9 shows the correlation analysis of emotion labels in the validation set. Each cell in the
figure represents the correlation score of each pair of emotion labels. The reported values show exciting
findings. Our system captured the relations among the emotion labels. The correlation scores of the
predicted labels were almost identical to the ground-truth. There was an exception in the surprise
and trust emotions. Our system was unsuccessful in capturing the relationships between these two
emotions and the inputs or the other emotions. We attribute this apparent lack of correlation to the
low number of training examples of these two emotions.
Moreover, there was always a positive correlation between related emotions such as “joy” and
“optimism” (the score from the ground truth labels and from the predicted labels was 0.74). On the
other side, we can see that there was a negative correlation between unlinked emotions like “anger”
and “love”. The scores were −0.27 and −0.3, respectively.
This result further strengthened our hypothesis that the proposed system was able to, implicitly,
model the relationships between the emotion labels.
Anger Anticipation Disgust Fear Joy Love Optimism Pessimism Sadness Surprise Trust
Anger ͳ ǦͲǤͳͷ ͲǤ͹ ǦͲǤͲʹͳ ǦͲǤͷʹͳ ǦͲǤʹ͹ ǦͲǤͶͷ ͲǤͲͷͷ ͲǤͳͷ ǦͲǤͲͷͶ ǦͲǤͳ͸
Anticipation ǦͲǤͳͷ ͳ ǦͲǤͳͷ ǦͲǤͲ͸͸ ͲǤͲͻʹ ǦͲǤͲͷ ͲǤͳͶ ǦͲǤͲͷͳ ǦͲǤͳʹ ͲǤͳͷ ͲǤͳͳ
Disgust ͲǤ͹ ǦͲǤͳͷ ͳ ͲǤͲʹͶ ǦͲǤͷ͵ ǦͲǤʹͻ ǦͲǤͶͻ ͲǤͲͳͷ ͲǤʹ ǦͲǤͲ͸ͺ ǦͲǤͳ͹
Fear ǦͲǤͲʹͳ ǦͲǤͲ͸͸ ͲǤͲʹͶ ͳ ǦͲǤʹͶ ǦͲǤͳͷ ǦͲǤͳͻ ͲǤͲ͹͸ ͲǤͲͲͷͺ ǦͲǤͲ͸Ͷ ǦͲǤͲʹͻ
Joy ǦͲǤͷʹͳ ͲǤͲͻʹ ǦͲǤͷ͵ ǦͲǤʹͶ ͳ ͲǤͶ ͲǤͷͺ ǦͲǤʹʹ ǦͲǤ͵ͷ ͲǤͲͶͻ ͲǤͳ͵
Love ǦͲǤʹ͹ ǦͲǤͲͷ ǦͲǤʹͻ ǦͲǤͳͷ ͲǤͶ ͳ ͲǤ͵ͳ ǦͲǤͳͶ ǦͲǤʹ ǦͲǤͲ͸ͻ ͲǤͳͳ
Optimism ǦͲǤͶͷ ͲǤͳͶ ǦͲǤͶͻ ǦͲǤͳͻ ͲǤͷͺ ͲǤ͵ͳ ͳ ǦͲǤʹ ǦͲǤʹͺ ͲǤͲͳͳ ͲǤʹͶ
Pessimism ͲǤͲͷͷ ǦͲǤͲͷͳ ͲǤͲͳͷ ͲǤͲ͹͸ ǦͲǤʹʹ ǦͲǤͳͶ ǦͲǤʹ ͳ ͲǤ͵Ͷ ǦͲǤͲ͵͸ ǦͲǤͲͺͳ
Sadness ͲǤͳͷ ǦͲǤͳʹ ͲǤʹ ͲǤͲͲͷͺ ǦͲǤ͵ͷ ǦͲǤʹ ǦͲǤʹͺ ͲǤ͵Ͷ ͳ ǦͲǤͲ͸ͻ ǦͲǤͳʹ
Surprise ǦͲǤͲͷͶ ͲǤͳͷ ǦͲǤͲ͸ͺ ǦͲǤͲ͸Ͷ ͲǤͲͶͻ ǦͲǤͲ͸ͻ ͲǤͲͳͳ ǦͲǤͲ͵͸ ǦͲǤͲ͸ͻ ͳ ͲǤͲ͸ʹ
Trust ǦͲǤͳ͸ ͲǤͳͳ ǦͲǤͳ͹ ǦͲǤͲʹͻ ͲǤͳ͵ ͲǤͳͳ ͲǤʹͶ ǦͲǤͲͺͳ ǦͲǤͳʹ ͲǤͲ͸ʹ ͳ
(a) The ground-truth labels.

Anger Anticipation Disgust Fear Joy Love Optimism Pessimism Sadness Surprise Trust
Anger ͳ ǦͲǤͳͶ ͲǤ͹ͺ ͲǤͲͷͶ ǦͲǤͷ͵ ǦͲǤ͵ ǦͲǤͷ͹ ǦͲǤͲ͵ͺ ͲǤʹͶ ͲǤͲͶ ǦͲǤͲ͹ͷ
Anticipation ǦͲǤͳͶ ͳ ǦͲǤͳͷ ǦͲǤͲ͸Ͷ ͲǤʹͳ ǦͲǤͲ͵ ͲǤʹ͵ ǦͲǤͲͶͶ ǦͲǤͳʹ ͲǤͳͷ ͲǤͲͺͺ
Disgust ͲǤ͹ͺ ǦͲǤͳͷ ͳ ͲǤͲͺ͸ ǦͲǤͷͶ ǦͲǤ͵ʹ ǦͲǤͷͻ ͲǤͲͳͷ ͲǤ͵ͳ ͲǤͲͶ ǦͲǤͲͺͳ
Fear ͲǤͲͷͶ ǦͲǤͲ͸Ͷ ͲǤͲͺ͸ ͳ ǦͲǤʹʹ ǦͲǤͳʹ ǦͲǤͳ͸ ͲǤͲͶͶ ͲǤͲͶͻ ͲǤͲͺʹ ǦͲǤͲ͵ͻ
Joy ǦͲǤͷ͵ ͲǤʹͳ ǦͲǤͷͶ ǦͲǤʹʹ ͳ ͲǤͶͳ ͲǤ͹Ͷ ǦͲǤͳͺ ǦͲǤ͵ͺ ͲǤͲ͵ͷ ͲǤͲͷͳ
Love ǦͲǤ͵ ǦͲǤͲ͵ ǦͲǤ͵ʹ ǦͲǤͳʹ ͲǤͶͳ ͳ ͲǤͶ͸ ǦͲǤͲͺͺ ǦͲǤʹͳ ͲǤͲ͹͹ ͲǤͲʹ͵
Optimism ǦͲǤͷ͹ ͲǤʹ͵ ǦͲǤͷͻ ǦͲǤͳ͸ ͲǤ͹Ͷ ͲǤͶ͸ ͳ ǦͲǤͳ͹ ǦͲǤͶ ͲǤͲͶ ͲǤͲ͸ͷ
Pessimism ǦͲǤͲ͵ͺ ǦͲǤͲͶͶ ͲǤͲͳͷ ͲǤͲͶͶ ǦͲǤͳͺ ǦͲǤͲͺͺ ǦͲǤͳ͹ ͳ ͲǤ͵ͷ ͲǤͳʹ ǦͲǤͲʹ͹
Sadness ͲǤʹͶ ǦͲǤͳʹ ͲǤ͵ͳ ͲǤͲͶͻ ǦͲǤ͵ͺ ǦͲǤʹͳ ǦͲǤͶ ͲǤ͵ͷ ͳ ͲǤͲͶͶ ǦͲǤͲͶͺ
Surprise ͲǤͲͶ ͲǤͳͷ ͲǤͲͶ ͲǤͲͺʹ ͲǤͲ͵ͷ ͲǤͲ͹͹ ͲǤͲͶ ͲǤͳʹ ͲǤͲͶͶ ͳ ǦͲǤͲͲ͵ʹ
Trust ǦͲǤͲ͹ͷ ͲǤͲͺͺ ǦͲǤͲͺͳ ǦͲǤͲ͵ͻ ͲǤͲͷͳ ͲǤͲʹ͵ ͲǤͲ͸ͷ ǦͲǤͲʹ͹ ǦͲǤͲͶͺ ǦͲǤͲͲ͵ʹ ͳ
(b) The predicted labels.
Figure 9. Correlation matrices of emotion labels of the development set.
71
Appl. Sci. 2019, 9, 1123
5. Conclusions
In this work, we presented a new approach to the multi-label emotion classification task. First,
we proposed a transformation method to transform the problem into a single binary classification
problem. Afterwards, we developed a deep learning-based system to solve the transformed problem.
The key component of our system was the embedding module, which used three embedding models
and an attention function. Our system outperformed the state-of-the-art systems, achieving a Jaccard
(i.e., multi-label accuracy) score of 0.59 on the challenging SemEval2018 Task 1:E-c multi-label emotion
classification problem.
We found that the attention function can model the relationships between the input words and
the labels, which helps to improve the system’s performance. Moreover, we showed that our system is
interpretable by visualizing the attention weights and analyzing them. However, some limitations
have been identified. Our system does not model the relationships between the phrases and the labels.
Phrases play a key role in determining the most appropriate set of emotions that must be assigned to
a tweet. For instance, an emotion word that reflects “sadness” can be flipped in a negated phrase or
context. Thus, in our future work, we plan to work on solving this drawback. One possible solution is
to adapt the attention function to model the relationships between different n-gram tokens and labels.
Structured attention networks [40] can also be adapted and used to address this issue.
Moreover, we plan to work on developing a non-aggressive system that performs robustly and
equally on all the emotion labels by experimenting with different ideas like using data augmentation
to enrich the training data or using transfer learning.
Author Contributions: Conceptualization, M.J. and A.M.; methodology, M.J.; software, M.J.; validation, M.J.
and A.M.; formal analysis, M.J. and A.M.; investigation, M.J.; resources, M.J.; data curation, M.J.; writing,
original draft preparation, M.J.; writing, review and editing, A.M.; visualization, M.J.; supervision, A.M.; funding
acquisition, A.M.
Acknowledgments: The authors acknowledge the research support of URV(2018PFR-URV-B2 and Martí i
Franqués PhD grant).
References
1. Chen, Y.; Zhou, Y.; Zhu, S.; Xu, H. Detecting Offensive Language in Social Media to Protect Adolescent Online
Safety. In Proceedings of the 2012 International Conference on Privacy, Security, Risk and Trust (PASSAT),
and 2012 International Conference on Social Computing (SocialCom), Amsterdam, The Netherlands, 3–5
September 2012; pp. 71–80.
2. Cherry, C.; Mohammad, S.M.; Bruijn, B. Binary Classifiers and Latent Sequence Models for Emotion Detection
in Suicide Notes. Biomed. Inform. Insights 2012, 5, BII-S8933. [CrossRef] [PubMed]
3. Mohammad, S.M.; Zhu, X.; Kiritchenko, S.; Martin, J. Sentiment, Emotion, Purpose, and Style in Electoral
Tweets. Inf. Process. Manag. 2015, 51, 480–499. [CrossRef]
4. Cambria, E. Affective computing and sentiment analysis. IEEE Intell. Syst. 2016 31, 102–107. [CrossRef]
5. Jabreel, M.; Moreno, A.; Huertas, A. Do Local Residents and Visitors Express the Same Sentiments on
Destinations Through Social Media? In Information and Communication Technologies in Tourism; Springer: New
York, NY, USA, 2017; pp. 655–668.
6. Yun, H.Y.; Lin, P.H.; Lin, R. Emotional Product Design and Perceived Brand Emotion. Int. J. Adv. Psychol. IJAP
2014, 3, 59–66. [CrossRef]
7. Mohammad, S.M. Sentiment Analysis: Detecting Valence, Emotions, and Other Affectual States from Text.
In Emotion Measurement; Meiselman, H.L., Ed.; Woodhead Publishing: Cambridge, UK, 2016; pp. 201–237,
ISBN 9780081005088.
8. Read, J.; Pfahringer, B.; Holmes, G.; Frank, E. Classifier Chains for Multi-label Classification. Mach. Learn.
2011, 85, 333. [CrossRef]
9. Robert, P. Emotions: A general Psychoevolutionary Theory. In Approaches to Emotion; Scherer, K.R., Ekman, P.,
Eds.; Psychology Press: London, UK, 2014; pp. 197–219.
72
Appl. Sci. 2019, 9, 1123
10. Tsoumakas, G.; Katakis, I. Multi-label Classification: An Overview. Int. J. Data Warehous. Min. IJDWM 2007,
3, 1–13. [CrossRef]
11. Read, J. Scalable Multi-label Classification. Ph.D. Thesis, University of Waikato, Hamilton,
New Zealand, 2010.
12. Mikolov, T.; Chen, K.; Corrado, G.; Dean, J. Efficient Estimation of Word Representations in Vector Space.
arXiv 2013, arXiv:1301.3781.
13. Schuster, M.; Paliwal, K.K. Bidirectional Recurrent Neural Networks. IEEE Trans. Signal Process. 1997, 45,
2673–2681. [CrossRef]
14. Chung, J.; Gulcehre, C.; Cho, K.; Bengio, Y. Empirical Evaluation of Gated Recurrent Neural Networks on
Sequence Modeling. arXiv 2014, arXiv:1412.3555.
15. Al-Molegi, A.; Jabreel, M.; Martínez-Ballesté, A. Move, Attend and Predict: An attention-based neural model
for people’s movement prediction. Pattern Recognit. Lett. 2018, 112, 34–40. [CrossRef]
16. Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A.N.; Kaiser, Ł.; Polosukhin, I.
Attention is All You Need. In Proceedings of the 31st Conference on Neural Information Processing Systems
(NIPS 2017), Long Beach, CA, USA, 4–9 December 2017; pp. 5998–6008.
17. Zhang, M.L.; Zhou, Z.H. A review on Multi-label Learning Algorithms. IEEE Trans. Knowl. Data Eng. 2014,
26, 1819–1837. [CrossRef]
18. Zhang, M.L.; Zhang, K. Multi-label Learning by Exploiting Label Dependency. In Proceedings of the 16th
ACM International Conference on Knowledge Discovery and Data Mining (SIGKDD), Washington, DC,
USA, 25–28 July 2010; pp. 999–1008.
19. Schapire, R.E.; Singer, Y. BoosTexter: A boosting-Based System for Text Categorization. Mach. Learn. 2000 39,
135–168. [CrossRef]
20. Zhang, M.L.; Zhou, Z.H. ML-KNN: A Lazy Learning Approach to Multi-label Learning. Pattern Recognit.
2007, 40, 2038–2048. [CrossRef]
21. Clare, A.; King, R.D. Knowledge Discovery in Multi-label Phenotype Data. In Proceedings of the European
Conference on Principles of Data Mining and Knowledge Discovery, Freiburg, Germany, 3–5 September
2001; pp. 42–53.
22. De Comite, F., Gilleron, R.; Tommasi, M. Learning Multi-label Alternating Decision Trees from Texts and
Data. In Proceedings of the 3rd International Conference on Machine Learning and Data Mining in Pattern
Recognition, Leipzig, Germany, 5–7 July 2003; Volume 2734, pp. 35–49.
23. Mencia, E.L.; Fürnkranz, J. Efficient Pairwise Multilabel Classification for Large-Scale Problems in the Legal
Domain. In Proceedings of the Joint European Conference on Machine Learning and Knowledge Discovery
in Databases, Antwerp, Belgium, 15–19 September 2008; Springer: Berlin/Heidelberg, Germany, 2008;
pp. 50–65.
24. Cheng, W.; Hüllermeier, E. Combining Instance-Based Learning and Logistic Regression for Multilabel
Classification. Mach. Learn. 2009, 76, 211–225. [CrossRef]
25. Godbole, S.; Sarawagi, S. Discriminative Methods for Multi-labeled Classification. In Proceedings of the
Pacific-Asia Conference on Knowledge Discovery and Data Mining, Osaka, Japan, 20–23 May 2008; Springer:
Berlin/Heidelberg, Germany, 2008; pp. 22–30.
26. Younes, Z.; Abdallah, F.; Denoeux, T.; Snoussi, H. A Dependent Multilabel Classification Method Derived
From the k-Nearest Neighbor Rule. J. Adv. Signal Process. 2011, 1, 645964. [CrossRef]
27. Yan, R.; Tesic, J.; Smith, J.R. Model-Shared Subspace Boosting for Multi-label Classification. In Proceedings
of the 13th ACM SIGKDD, San Jose, CA, USA, 12–15 August 2007; pp. 834–843.
28. Jabreel, M.; Moreno, A. SentiRich: Sentiment Analysis of Tweets Based on a Rich Set of Features. In Artificial
Intelligence Research and Development; Nebot, Á., Binefa, X., López de Mántaras, R., Eds.; IOS Press:
Amsterdam, The Netherlands, 2016; Volume 288, pp. 137–146, ISBN 978-1-61499-695-8.
29. Jabreel, M.; Moreno, A. SiTAKA at SemEval-2017 Task 4: Sentiment Analysis in Twitter Based on a Rich Set
of Features. In Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017),
Vancouver, BC, Canada, 3–4 August 2017; pp. 694–699.
30. Mohammed, S., M.; Bravo-Marquez, F.; Salameh, M.; Kiritchenko,S. Semeval-2018 task 1: Affect in Tweets.
In Proceedings of the 12th International Workshop on Semantic Evaluation, New Orleans, LA, USA, 5–6 June
2018; pp. 1–17.
31. LeCun, Y.; Bengio, Y.; Hinton, G. Deep Learning. Nature 2015 521, 436–444. [CrossRef]
73
Appl. Sci. 2019, 9, 1123
32. Tang, D.; Qin, B.; Liu, T. Deep Learning for Sentiment Analysis: Successful Approaches and Future
Challenges. Wiley Interdiscip. Rev. Data Min. Knowl. Discov. 2015 5, 292–303. [CrossRef]
33. Baziotis, C.; Athanasiou, N.; Chronopoulou, A.; Kolovou, A.; Paraskevopoulos, G.; Ellinas, N.; Narayanan, S.;
Potamianos, A. NTUA-SLP at SemEval-2018 Task 1: Predicting Affective Content in Tweets with Deep
Attentive RNNs and Transfer Learning. In Proceedings of the 12th International Workshop on Semantic
Evaluation, New Orleans, LA, USA, 5–6 June 2018; pp. 245–255.
34. Meisheri, H.; Dey, L. TCS Research at Semeval2018 Task 1: Learning Robust Representations using
Multi-Attention Architecture. In Proceedings of the 12th International Workshop on Semantic Evaluation,
New Orleans, LA, USA, 5–6 June 2018; pp. 291–299.
35. Park, J.H.; Xu, P.; Fung, P. PlusEmo2Vec at SemEval-2018 Task 1: Exploiting Emotion Knowledge from Emoji
and #hashtags. In Proceedings of the 12th International Workshop on Semantic Evaluation, New Orleans,
LA, USA, 5–6 June 2018; pp. 264–272.
36. Kingma, D.P.; Ba, J. Adam: A Method for Stochastic Optimization. arXiv 2014, arXiv:1412.6980.
37. James, B.; Yamins, D.; Cox, D.D. Hyperopt: A python library for optimizing the hyperparameters of
machine learning algorithms. In Proceedings of the 12th Python in Science Conference, Austin, TX, USA,
24–29 June 2013.
38. Mohammad, S.; Kiritchenko, S. Understanding emotions: A dataset of tweets to study interactions between
affect categories. In Proceedings of the Eleventh International Conference on Language Resources and
Evaluation, Miyazaki, Japan, 7–12 May 2018.
39. Kant, N.; Puri, R.; Yakovenko, N.; Catanzaro, B. Practical Text Classification With Large Pre-Trained Language
Models. arXiv 2018, arXiv:1812.01207.
40. Kim, Y.; Denton, C.; Hoang, L.; Rush, A.M. Structured Attention Networks. arXiv 2017, arXiv:1702.00887.
74
applied
sciences
Article
Using Social Media to Identify Consumers’
Sentiments towards Attributes of Health Insurance
during Enrollment Season
Eline M. van den Broek-Altenburg * and Adam J. Atherly
Center for Health Services Research, The Larner College of Medicine, University of Vermont,
Burlington, VT 05405, USA; [email protected]
* Correspondence: [email protected]; Tel.: +1-(802)-656-2722; Fax: +1-(802)-656-8577

Received: 27 April 2019; Accepted: 13 May 2019; Published: 17 May 2019
Featured Application: The health insurance choice literature has found that financial
considerations, such as premiums, deductible, and maximum out-of-pocket spending caps,
are important to consumers. But these financial factors are just part of the cost-benefit trade-off
consumers make. Publicly available datasets often do not include these other factors. Researchers
in other fields have increasingly used web data from social media platforms, such as Twitter and
search engines to analyze consumer behavior using Natural Language Processing. NLP combines
machine learning, computational linguistics, and computer science, to understand natural
language including consumer’s sentiments, attitudes, and emotions from social media. This study
is among the first to use natural language from an online platform to analyze sentiments
when consumers are discussing health insurance. By clarifying what the expressed attitudes
or sentiments are, we get an idea of what variables we may want to include in future studies of
health insurance choice.
Abstract: This study aims to identify sentiments that consumers have about health insurance by
analyzing what they discuss on Twitter. The objective was to use sentiment analysis to identify
attitudes consumers express towards health insurance and health care providers. We used an
Application Programming Interface to gather tweets from Twitter with the words “health insurance”
or “health plan” during health insurance enrollment season in the United States in 2016–2017.
Word association was used to find words associated with “premium,” “access,” “network,” and
“switch.” Sentiment analysis established which specific emotions were associated with insurance and
medical providers, using the NRC Emotion Lexicon, identifying emotions. We identified that provider
networks, prescription drug benefits, political preferences, and norms of other consumers matter.
Consumers trust medical providers but they fear unexpected health events. The results suggest
that there is a need for different algorithms to help consumers find the plans they want and need.
Consumers buying health insurance in the Affordable Care Act marketplaces in the United States
choose lower-cost plans with limited benefits, but at the same time express fear about unexpected
health events and unanticipated costs. If we better understand the origin of the sentiments that drive
consumers, we may be able to help them better navigate insurance plan options and insurers can
better respond to their needs.
Keywords: social media; Twitter; text mining; sentiment analysis; word association; health insurance;
provider networks

Appl. Sci. 2019, 9, 2035
1. Introduction
In the Affordable Care Act health insurance marketplaces in the United States (USA), consumers
are mandated to choose a health insurance plan. Plans may differ by premiums, benefits, and other
plan attributes, such as the network of providers or how tightly managed the plan is. Consumers
ideally pick the best combination of plan attributes, switching plans if necessary.
The health insurance choice literature has found that financial considerations, such as premiums,
deductibles, and maximum out-of-pocket spending caps, are indeed important to consumers [1–5].
However, these considerations are just part of the cost-benefit trade-off consumers make. Surveys and
discrete choice experiments suggest that other plan attributes, such as choice of personal doctors [6,7],
continuity of care [8–11], or how “tightly managed” the plan is [4], also have an effect on consumers’
choices. Information about quality of service or other aspects of care delivery may also play a role [12].
The more we know about the trade-offs consumers make and what factors play a role in insurance
choice, the better we can predict or anticipate future choices.
This study identifies sentiments that consumers have when discussing health insurance in the
USA by using an alternative data source: Twitter. Twitter has grown exponentially in recent years
and computer and data scientists have learned how to extract information from the 328 million
monthly active Twitter users, 70 million of whom live in the USA [13], Every second, on average,
around 6000 tweets are sent via Twitter, which corresponds to 500 million tweets per day and around
200 billion tweets per year [14].
Twitter’s “tweets,” which were at the time of our study limited to 140 characters, have been
shown to have surprising predictive power. Numerous studies across different academic fields
have used Twitter as a tool for forecasting or prediction. Researchers in industrial organization and
marketing have used Twitter data to analyze what consumers want and need. In fields like finance
and macroeconomics, text from social media has been used to make predictions about the stock
market [15–17], oil [18], sales [19], and unemployment rates [20,21], or as a surveillance tool to track
messages related to security breaches [22]. In the political arena, Twitter has been used to predict the
outcome of elections or to poll political sentiment [23–25]. It has been suggested that analysis of social
media data more accurately predicted Trump’s win than election polls [26].
More recently, text mining of web content has been used in the context of public health. Twitter data
have been used to evaluate health care quality, poll reactions to health policy reforms and in various
other public health contexts. Additionally, researchers have used text from Twitter for influenza
surveillance [27–31]. For example, an analysis of three million tweets between May and December
2009 showed that the 2009 H1N1 flu outbreak could have been identified on Twitter one week before
it emerged in official records from general practitioner reports. Researchers at the Department of
Computer Science at Johns Hopkins University created a model for Twitter that groups symptoms and
treatments into latent ailments.
Other examples include using tweets to compute the average happiness of cancer patients for
each cancer diagnosis [32], to measure patient-perceived quality of care in hospitals [33], and to predict
asthma prevalence by combining Twitter data with other data sources [34]. The latter study provides
evidence that monitoring asthma-related tweets may provide real-time information that can be used to
predict outcomes from traditional surveys.
Some recent studies have used web data from search engines such as Google to analyze consumer
behavior in health insurance. One study examined factors associated with health insurance-related
Google searches during the first open enrollment period [35]. The authors found that search volumes
were associated with local uninsured rates. Another study used text data from Twitter to identify
consumers’ sentiments to predict insurance enrollment [36].
A number of studies have used Twitter data in a similar context to this study. Beyond the health
insurance studies mentioned above, Twitter has also been used to assess public opinion about the
Affordable Care Act over time: a study found substantial spikes in the volume of Affordable Care
Act-related tweets in response to key events in the law’s implementation [37].
76
Appl. Sci. 2019, 9, 2035
The aim of this study is to identify sentiments that consumers express on Twitter when they
discuss health insurance. The first objective of this paper is to identify words that are associated with
the word “switch” in the tweets. In the context of tweets gathered on the search “health insurance,”
we assume that “switch” is related to health insurance at least some of the time. The second objective
of this paper is to identify what attitudes or sentiments consumers have when communicating about
health insurance in online social networks. The study is hypothesis-generating: gaining insights into
the words consumers use when they communicate about health insurance on an online social network
may lead to better-informed theory regarding health plan choices. By clarifying what the expressed
attitudes or sentiments are, we may find variables we can include in future studies and we may be able
to generate testable hypotheses.
2. Materials and Methods
2.1. Data
Using an Application Programming Interface (API), we gathered tweets from the Twitter server
with the words “health insurance,” “health plan,” “health provider” or “doctor” in them during open
enrollment period from 1 November 2016 until 31 January 2017. This is the yearly period when U.S.
citizens can enroll in or switch a health insurance plan. Beyond this timeframe, they have to stay with
the plan they have. API is code that allows two software programs, in our case Twitter and Python 3.6,
to communicate with each other. With the API, Python authenticated, requested, and received the data
from the Twitter server. The words “health insurance” and “health plan” generated approximately one
tweet every 3 s, adding up to 28,800 per day; 892,800 per month; and 2,678,400 total tweets during the
ACA open enrollment season for 2017.
We used the API to create a body of text, called “VCorpus,” in R 3.4. At each index of the
“VCorpus object,” there is a PlainTextDocument object, which is essentially a list that contains the
actual text data of the tweet, as well as some corresponding metadata such as the location from
which the tweet was sent, the date, and other elements. In other words, the tweets were gathered
in one text document and pre-processed for analysis. This pre-processing gets rid of punctuation,
hashtags, and retweets, strips white space, and removes stop words and custom terms so that they
are now represented as lemmatized plain words. To illustrate, the tweet text “Obama care is a joke fr.
My health care plan is just not affordable no more. Cheaper to pay the penalty I guess” was changed to:
“obama care joke fr health care plan just affordable cheaper pay penalty guess” after pre-processing.
The most important way that text differs from more typical data sources is that text is naturally
high-dimensional, which makes analysis difficult, often referred to as the “curse of dimensionality.”
For example, suppose that a sample of tweets, each of which is 20 words long, and that each word is
drawn from a vocabulary of 2000 possible words. It follows that the unique representation of these
tweets has a very high dimensionality, with 40,000 data columns.
To reduce dimensionality, we use the “bag of words” (BoW) model. The BoW model, also known
as a vector space model, reduces dimensionality by simplifying the representation of the words used
in natural language processing and information retrieval. In this model, a text document (such as a
tweet) is represented as the bag (multiset) of its words, disregarding grammar and word order [38].
Subsequently, our “bag of words” model learned a vocabulary from the millions of tweets and
then modeled each tweet by counting the number of times each word appears in the tweet [38].
Through automatic text categorization, we extracted features or “token sets” from the text by
representing the tweets by the words that occur in it.
To explain how we converted text to numeric data, here is an example. Sentence 1: “The health
insurance plan is too expensive to cover my health needs”; Sentence 2: “The health insurance company
offers an expensive health plan.” We can see that, from these two sentences, our vocabulary is: {The,
health, insurance, plan, is, too, expensive, to, cover, my, needs, company, offers, an}. To get the bags of
words, the number of times each word occurs was counted in each sentence. In Sentence 1,” health”
77
Appl. Sci. 2019, 9, 2035
appears twice, and the other words each appear once, so the feature vector for Sentence 1 is: {1, 2, 1, 1,
1, 1, 1, 1, 1, 1, 1, 0, 0, 0} and Sentence 2: {1, 2, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1} We created a Term-Document
matrix (TDM) where each row is a 1/0 representation of whether a single word is contained within the
tweet and every column is a tweet. We then removed terms with a sparse factor of less than 0.001.
These are the terms that occur less than 0.01% of times in a tweet. The resulting matrix contained
1354 words.
2.2. Analytic Approach

To find which words are associated with switching, we used Word Association: a function that
calculates the association of a word with another word in the TDM. We used the findAssocs() in R to
calculate the association of a word with every other word in the TDM. The output scores range from 0
to 1, where a score of 1 means that two words always appear together, and a score of 0 means that
they never appear together. To find associations, we set a minimum of 0.05, meaning that the program
would look for all words that were associated in one tweet (that has a maximum of 140 characters) with
“premium” at least 5% of the tweets. Since we were interested in attitudes to plan attributes, we tested
three attributes of health insurance plans: “premium,” “access,” and “network.” We also looked for all
the words that were associated with the word “switch” at least 5% of the time. We chose “premium”
because we know from the literature that the premium matters when consumers buy health insurance,
as well as “access” to doctors. Since we were particularly interested in whether provider networks,
which refers to insurance coverage of doctors in-network, matters when consumers discuss health
insurance, we also looked at “network.”
To identify sentiments in the tweets, we use methods of classification. We used sentiment lexicons,
a dictionary-based approach, which depends on finding opinion seed words, and then searches
the dictionary for their synonyms. While various sentiment lexicons all have their advantages and
disadvantages in the context of topic-specific subjectivity scores, interpretation, and completeness,
the choice for a specific sentiment lexicon is context-specific. We used the NRC Emotion Lexicon
(NRC Emolex) [39], which classifies words in a binary yes/no for classes of attitude “positive” and
“negative”; and for classes of emotion: anger, anticipation, disgust, fear, joy, sadness, surprise, and trust.
We wanted to not only find out the overall sentiment of tweets, but also what specific emotion they
embodied, and identify which words represented those emotions.
3. Results
3.1. Word Association

Figure 1 shows that the most common word consumers used in combination with the word
“switch” was the word “bait” (0.31), meaning that in 31% of tweets with the word “switch,” the word
“bait” was also used. In 8% of the tweets that use the word “switch,” the word “premium” is also used.
This suggests that insurance was often described as a bait and switch, such as in this example
tweet: “The healthcare bait-and-switch. In network hospital, out of network doctor.” This was followed
by “lockedin” and “rxrights.” “Rxrights” refers RxRights.org, which serves as a forum for individuals
to share experiences and voice opinions regarding the need for affordable prescription drugs. It is an
example of how “switch” could be used in a different context than insurance, such as in this tweet:
“In the USA, this is how we are forced to switch insulins without any regard to our health or to doctors’
orders.” The next most common word associated with switch was “network.” Networks were used in
tweets about switching, such as in this example: “Gods blessings is like health insurance, you have in
network benefits and out of network benefits, but in network is way better of course.” There were
tweets discussing the role of provider networks in insurance such as this one: “$252 for a doctor visit.
I wasn’t even there for 20 minutes. Thanks insurance for not letting me know the doc was no longer in
my network.”
78
Appl. Sci. 2019, 9, 2035
EDLW
U[ULJKWV ORFNHGLQ
6ZLWFK
SUHPLXP QHWZRUN

Figure 1. Word associations for plan attributes and switching among health insurance tweets.
“Network” was associated with the word switch as often as “premium,” which was 0.08,
meaning that 8% of tweets that had the word “switch” in them also contained the work “premium.”
Consumers expressed concerns about premiums, deductibles, and co-pays, such as in this example:
“Dude, as expensive as my insurance is. Copays, premiums, etc., I might as well not even have it.
Costs 2 much to use.”
3.2. Sentiment Analysis

The results of the sentiment analysis showed that two emotions prevailed in tweets during
enrollment season: “trust” and “fear” (Table 1). The emotion “trust” was the primary driver of positive
sentiment expressed in tweets, while the emotion “fear” was the primary driver of negative sentiments
and accounted for the slightly negative overall sentiment. Trust was expressed in the context of doctors,
nurses, and other medical providers. Here is an example of a tweet discussing this trustworthy role:
“Patients value their family doctor at the bedside when gravely ill. Healing presence is so powerful.”
Table 1. Positive and negative emotions expressed on Twitter, by the words used in the tweets.
Positive Negative
Doctor (trust) Pressure (negative)
Physician (trust) Die (fear, sadness)
Hospital (trust) Emergency (fear, anger, disgust, sadness)
Nurse (trust) Disease (fear, anger, disgust, sadness)
Plan (anticipation) Pain, surgery (fear, sadness)
Save money (joy) Miscarriage (fear, sadness)
Choice (positive) Cancer (fear, anger, disgust, sadness)
Note: In brackets are the emotions that the NRC Emolex associated with the word used in the tweets.
In a tweet like this, the NRC Emolex classified the word “value” as positive and associated
with “trust,” while the word “healing” is classified as positive and associated with the emotions
of anticipation, joy, and trust. Another tweet referred to the importance of continuity of care:
“Seeing anxiety in culture from lack of relationship with continuity of fam Doctor.” The word “anxiety”
is classified by the NRC Emolex as negative and associated with the emotions of fear, anticipation,
and sadness. In this way, the words used in tweets are given a subjectivity score and common ones are
reported in Table 1.
Fear was conveyed in Tweets about medical events such as “lose,” “disease,” “emergency,”
“surgery,” “cancer,” and unanticipated costs. Consumers expressed both negative and positive
79
Appl. Sci. 2019, 9, 2035
sentiments about choice, but the NRC Emolex could not specify the exact emotion. For example,
one tweet stated: “I hate choice so much that I’ve essentially disregarded my doctor because I’ve
had such a low care of my own wellbeing.” It also identified “pressure” as being negative but could
not specify what kind of emotion consumers expressed. Overall, the sentiment of consumers toward
health insurance was slightly negative, although most sentiments were not at either extreme (Figure 2).
Figure 2. Sentiments expressed in health insurance-related tweets in 2016-2017 Affordable Care Act
enrollment season.
The figure shows how the words used in the tweets were classified: either positive or negative.
It follows from the histogram that most words were classified as “slightly negative” (-1) and few words
were classified as either extremely negative or extremely positive.
To understand what attitudes consumers expressed regarding specific attributes of health plans,
we examined “premium,” “access,” and “network.” Table 2 illustrates that consumers used the
insurance attribute “premium” most often in combination with words like “increase” or “relief.”
Table 2. Words associated with premium, access, and network.
Term in Tweets Words Used Association

Premium Increase 0.17
relief 0.16
Obamacare 0.08
Access No-copay 0.74
ppact (planned parenthood), birth, women 0.66, 0.51, 0.40
Affordable Care Act 0.15
Network narrow 0.23
providers 0.16
covering 0.08
Switch Bait 0.31
lockedin 0.14
rxrights 0.13
premium 0.08
network 0.08
Access was associated most of the time with “nocopay,” suggesting that consumers who care
about access also care about copays. The attribute “network” was associated with “narrow” 23% of
the time, and with “providers” 16% of the time, suggesting that many consumers talk about narrow
network plans when discussing health insurance.
80
Appl. Sci. 2019, 9, 2035
4. Discussion
In this study, we used text data from Twitter to identify attitudes that consumers express towards
health insurance, plan attributes, and health care providers. The health insurance choice literature
focuses primarily on well-defined features of plans that are easily observable from administrative
data such as benefit design, co-insurance rates, and deductibles. Previous studies found that
financial considerations, such as premiums, deductibles, and maximum out-of-pocket spending caps,
are important to consumers. This study reinforces some results from previous research. The sensitivity
of consumers to higher premiums that our study finds is well documented in other literature. The role
of provider networks has been debated recently—our study reinforces the importance of the networks
to consumers.
There are limitations associated with the bag of words approach that we used. The main
disadvantage is that it severely limits the context of the tweet and loses the order of specific information.
Also, it requires supervised machine learning, which entails modeling linguistic knowledge through the
use of dictionaries containing words that are tagged with their semantic orientation [39]. This means
that we used an existing data dictionary, but we accept the classification of the English words to
identify emotions.
It is a challenge to capture the essential meaning of a tweet in a machine-understandable
format [40] as issues like short length, informal words, misspellings, and unusual grammar make
it difficult to obtain a good representation to capture these text aspects. More recently, there has
been a paradigm shift in machine learning towards using distributed representations for words [41]
and sentences [42,43]. Most studies analyzing tweets have not been able to use a fully unsupervised
approach for message-level and phrase-level sentiment analysis of tweets. The advantage to such an
approach would have been, for example, that we would have been able to convey emotions in the
same manner in the tweets as in newspaper articles or blogs, reviews, or other types of user-generated
content. In everyday life, we rely on context to interpret a piece of text or comment, so with bag of
words it is harder to capture context as it merely focuses on isolated words or term frequencies.
We do not know how the demographics of people tweeting about health insurance compares
to the Affordable Care Act marketplace population, the Medicaid expansion population, and the
uninsured. Tweets contain macro data about the user, but it is limited to whatever information the user
decides to give. In practice, a small percentage of users provide personal information such as gender
or age. We do have some information about location, but we lack this information for a substantial
part of the sample and the level of information (city, state) differs by user.
Another limitation is that tweets have social network-oriented properties, and therefore we believe
that a good representation of our tweets should also capture social aspects. A social network analysis is
beyond the scope of this study, as well as conversational science approach looking at how a comment
is influenced by the previous one.
5. Conclusions
This study suggests that other, non-financial factors, might be important in the choice of health
insurance plan, such as the sentiments that consumers have. The discussion of “fear” in relation
to health insurance plan choice may seem unusual; however, the basic economic model for health
insurance posits risk aversion as a key motivator for individuals to buy coverage. In some sense, “fear”
is simply “risk-averse” in the vernacular. In another sense, however, this study provides specificity
about the nature of the risk aversion and suggests that consumers lack confidence in their choices and
express fear towards adverse health events and unanticipated costs.
If we better understand the origin of the fear and other sentiments that drive consumers, we may
be able to help them to better navigate insurance plan options and insurers can make sure to better
respond to their needs. Additionally, plan finders often provide consumers with actuarial spending
estimates for “average” consumers; our study suggests that the average outcome is not the outcome of
interest to consumers. Even though some plan finders include individual-specific calculations [44],
81
Appl. Sci. 2019, 9, 2035
insurance companies may want to react to consumers’ sentiments in addition to financial considerations.
Consumers are concerned about an unusual event—cancer, accident, surgery, or disease—and whether
they can afford care when death is a real possibility. Plan finders could be reconfigured to give coverage
data for consumers experiencing these extreme health events.
Text mining is an important addition to research in this area because of the sheer volume of
information and the possibility of looking quantitatively at formative data. In social science in
general, and public health research in particular, a common practice is to rely on small convenience
samples to generate formative data that are qualitative to generate hypotheses. Potential testable
hypotheses that were generated from the analysis may include “Provider networks are associated
with health plan switching” or “Sentiments expressed on Twitter predict health insurance choice.”
Where qualitative research usually involves very small samples, text data can yield similar insights
with substantially larger sample sizes. This study illustrates that we can use another, perhaps equally
effective, advanced method and data to generate testable hypotheses.
Author Contributions: Both authors made substantial contributions to the conception and design of the work.
E.M.v.d.B.-A. was responsible for the acquisition, analysis, and interpretation of data; and for drafting the
work. A.J.A. played an important role in the interpretation of the data and results and substantively revised the
manuscript. Both authors have approved the submitted version and agree to be personally accountable for their
own contributions and for ensuring that questions related to the accuracy or integrity of any part of the work are
appropriately investigated, resolved, and documented in the literature.
References
1. Atherly, A.; Dowd, B.E.; Feldman, R. The effect of benefits, premiums, and health risk on health plan choice
in the Medicare program. Health Serv. Res. 2004, 39, 847–864. [CrossRef] [PubMed]
2. Buchmueller, T.C. Consumer Demand for Health Insurance; NBER Reporter, National Bureau of Economic
Research, Inc.: Cambridge, MA, USA, 2006; pp. 10–13.
3. Short, P.F.; Taylor, A.K. Premiums, benefits, and employee choice of health insurance options. J. Health Econ.
1989, 8, 293–311. [CrossRef]
4. Trujillo, A.J.; Ruiz, F.; Bridges, J.F.P.; Amaya, J.L.; Buttorff, C.; Quiroga, A.M. Understanding consumer
preferences in the context of managed competition. Appl. Health Econ. Health Policy 2012, 10, 99–111.
[CrossRef]
5. Blumberg, L.J.; Long, S.K.; Kenney, G.M.; Goin, D. Factors Influencing Health Plan Choice among the Marketplace
Target Population on the Eve of Health Reform; Urban Institute: Washington, DC, USA, 2013.
6. Altman, D. What new data tells us about doctors choice. The Wall Street Journal. 4 February 2016.
Available online: https://blogs.wsj.com/washwire/2016/02/04/what-new-data-tell-us-about-doctor-choice/
(accessed on 28 September 2017).
7. Stokes, T.; Tarrant, C.; Mainous, A.G.; Schers, H.; Freeman, G.; Baker, R. Continuity of care: Is the personal
doctor still important? A survey of general practitioners and family physicians in England and Wales,
the United States, and The Netherlands. Ann. Fam. Med. 2005, 3, 353–359. [CrossRef] [PubMed]
8. Guthrie, B.; Saultz, J.W.; Freeman, G.K.; Haggerty, J.L. Continuity of care matters. BMJ Br. Med. J. 2008, 337,
a867. [CrossRef]
9. Turner, D.; Tarrant, C.; Windridge, K.; Bryan, S.; Boulton, M.; Freeman, G.; Baker, R. Do patients value
continuity of care in general practice? An investigation using stated preference discrete choice experiments.
J. Health Serv. Res. Policy 2007, 12, 132–137. [CrossRef]
10. Higuera, L.; Carlin, C.S.; Dowd, B. Narrow provider networks and willingness to pay for continuity of care
and network breadth. J. Health Econ. 2018, 60, 90–97. [CrossRef]
11. Mainous, A.G.; Goodwin, M.A.; Stange, K.C. Patient-physician shared experiences and value patients place
on continuity of care. Ann. Fam. Med. 2004, 2, 452–454. [CrossRef]
12. Enthoven, A.; Kronick, R. Competition 101: Managing demand to get quality care. Bus. Health 1988, 5, 38–40.
13. Statistica. Number of Monthly Active Twitter Users in the United States from 1st Quarter 2010 to 1st Quarter
2017 (in Millions). 2017. Available online: https://www.statista.com/statistics/274564/monthly-active-twitter-
users-in-the-united-states/ (accessed on 22 July 2017).
82
Appl. Sci. 2019, 9, 2035
14. Stats IL. Twitter Usage Statistics. 2017. Available online: http://www.internetlivestats.com/twitter-statistics/
(accessed on 22 July 2017).
15. Ben-Ami, Z.; Feldman, R.; Rosenfeld, B. Using multi-view learning to improve detection of investor sentiments
on twitter. Computación y Sistemas 2014, 18, 477–490. [CrossRef]
16. Bing, L.; Chan, K.C.; Ou, C. Public sentiment analysis in Twitter data for prediction of a company’s stock
price movements. In Proceedings of the 2014 IEEE 11th International Conference on e-Business Engineering
(ICEBE), Guangzhou, China, 5–7 November 2014.
17. Chen, R.; Lazer, M. Sentiment analysis of twitter feeds for the prediction of stock market movement. Stanf. Edu.
Retrieved 2013, 25, 2013.
18. Rao, T.; Srivastava, S. Using Twitter Sentiments and Search Volumes Index to Predict Oil, Gold, Forex and Markets
Indices; Delhi Institutional Repository: Delhi, India, 2012.
19. Dijkman, R.; Ipeirotis, P.; Aertsen, F.; van Helden, R. Using twitter to predict sales: A case study. arXiv 2015,
arXiv:150304599.
20. Antenucci, D.; Cafarella, M.; Levenstein, M.; Ré, C.; Shapiro, M.D. Using Social Media to Measure Labor Market
Flows; National Bureau of Economic Research: Cambridge, MA, USA, 2014.
21. Llorente, A.; Garcia-Herranz, M.; Cebrian, M.; Moro, E. Social media fingerprints of unemployment.
PLoS ONE 2015, 10, e0128692. [CrossRef]
22. Hao, J.; Hao, J.; Dai, H.; Dai, H. Social media content and sentiment analysis on consumer security breaches.
J. Financ. Crime 2016, 23, 855–869. [CrossRef]
23. Bermingham, A.; Smeaton, A.F. On using Twitter to monitor political sentiment and predict election results.
In Proceedings of the Workshop at the International Joint Conference for Natural Language Processing
(IJCNLP), Chiang Mai, Thailand, 13 November 2011.
24. Tumasjan, A.; Sprenger, T.O.; Sandner, P.G.; Welpe, I.M. Predicting elections with twitter: What 140 characters
reveal about political sentiment. ICWSM 2010, 10, 178–185.
25. Tumasjan, A.; Sprenger, T.O.; Sandner, P.G.; Welpe, I.M. Election forecasts with Twitter: How 140 characters
reflect the political landscape. Soc. Sci. Comput. Rev. 2011, 29, 402–418. [CrossRef]
26. Perez, S. Analysis of Social Media Did a Better Job at Predicting Trump’s win than the Polls. Tech Crunch. 2016.
Available online: https://techcrunch.com/2016/11/10/social-media-did-a-better-job-at-predicting-trumps-
win-than-the-polls/ (accessed on 24 July 2017).
27. Broniatowski, D.A.; Paul, M.J.; Dredze, M. National and local influenza surveillance through Twitter:
An analysis of the 2012–2013 influenza epidemic. PLoS ONE 2013, 8, e83672. [CrossRef] [PubMed]
28. Culotta, A. Towards detecting influenza epidemics by analyzing Twitter messages. In Proceedings of the
First Workshop on Social Media Analytics, Washington, DC, USA, 25–28 July 2010.
29. Lamb, A.; Paul, M.J.; Dredze, M. Separating fact from fear: tracking flu infections on Twitter. In Proceedings
of the North American Chapter of the Association for Computational Linguistics—Human Language
Technologies (NAACL HLT) 2013 Conference, Atlanta, GA, USA, 9–14 June 2013.
30. Lampos, V.; De Bie, T.; Cristianini, N. Flu detector-tracking epidemics on Twitter. In Proceedings of the Joint
European Conference on Machine Learning and Knowledge Discovery in Databases, Barcelona, Spain, 20–24
September 2010; pp. 599–602.
31. Signorini, A.; Segre, A.M.; Polgreen, P.M. The use of Twitter to track levels of disease activity and public
concern in the US during the influenza A H1N1 pandemic. PLoS ONE 2011, 6, e19467. [CrossRef] [PubMed]
32. Crannell, W.C.; Clark, E.; Jones, C.; James, T.A.; Moore, J. A pattern-matched Twitter analysis of US
cancer-patient sentiments. J. Surg. Res. 2016, 206, 536–542. [CrossRef]
33. Hawkins, J.B.; Brownstein, J.S.; Tuli, G.; Runels, T.; Broecker, K.; Nsoesie, E.O.; McIver, D.J.; Rozenblum, R.;
Wright, A.; Bourgeois, F.T.; et al. Measuring patient-perceived quality of care in US hospitals using Twitter.
BMJ Qual. Saf. 2016, 25, 404–413. [CrossRef]
34. Dai, H.; Lee, B.R.; Hao, J. Predicting asthma prevalence by linking social media data and traditional surveys.
ANNALS Am. Acad. Polit. Soc. Sci. 2017, 669, 75–92. [CrossRef]
35. Gollust, S.E.; Qin, X.; Wilcock, A.D.; Baum, L.M.; Barry, C.L.; Niederdeppe, J.; Fowler, E.F.; Karaca-Mandic, P.
Search and you shall find: Geographic characteristics associated with google searches during the affordable
care act’s first enrollment period. Med. Care Res. Rev. 2016. [CrossRef] [PubMed]
36. Wong, C.A.; Sap, M.; Schwartz, A.; Town, R.; Baker, T.; Ungar, L.; Merchant, R.M. Twitter sentiment predicts
affordable care act marketplace enrollment. J. Med. Internet Res. 2015, 17, e51. [CrossRef] [PubMed]
83
Appl. Sci. 2019, 9, 2035
37. Davis, M.A.; Zheng, K.; Liu, Y.; Levy, H. Public response to Obamacare on Twitter. J. Med. Internet Res. 2017,
19, e167. [CrossRef] [PubMed]
38. Deepu, S.; Raj, P.; Rajaraajeswari, S. A Framework for Text Analytics using the Bag of Words (BoW) Model for
Prediction. In Proceedings of the 1st International Conference on Innovations in Computing & Networking
(ICICN16), Bangalore, India, 12–13 May 2016.
39. Mohammad, S.M.; Kiritchenko, S.; Zhu, X. NRC-Canada: Building the state-of-the-art in sentiment analysis
of tweets. arXiv 2013, arXiv:13086242.
40. Ganesh, J.; Gupta, M.; Varma, V. Interpretation of semantic tweet representations. arXiv 2017, arXiv:170400898.
41. Mikolov, T.; Sutskever, I.; Chen, K.; Corrado, G.S.; Dean, J. Distributed representations of words and
phrases and their compositionality. In Proceedings of the Advances in Neural Information Processing
Systems 26, 27th Annual Conference on Neural Information Processing Systems 2013, Lake Tahoe, NV, USA,
5–10 December 2013.
42. Hill, F.; Cho, K.; Korhonen, A. Learning distributed representations of sentences from unlabelled data. arXiv
2016, arXiv:160203483.
43. Le, Q.; Mikolov, T. Distributed representations of sentences and documents. In Proceedings of the 31st
International Conference on Machine Learning (ICML-14), Bejing, China, 22–24 June 2014.
44. Wong, C.A.; Polsky, D.E.; Jones, A.T.; Weiner, J.; Town, R.J.; Baker, T. For third enrollment period, marketplaces
expand decision support tools to assist consumers. Health Aff. 2016, 35, 680–687. [CrossRef]
84
applied
sciences
Article
Gender Classification Using Sentiment Analysis and
Deep Learning in a Health Web Forum
Sunghee Park and Jiyoung Woo *
Department of Future Convergence Technology, Soonchunhyang University, Asan-si 31538, Korea;
[email protected]

Featured Application: This work can be applied to detect the gender of online users who do not
disclose this information.
Abstract: Sentiment analysis is the most common text classification tool that analyzes incoming
messages and tells whether the underlying sentiment is positive, negative, or neutral. We can use
this technique to understand people by gender, especially people who are suffering from a sensitive
disease. People use health-related web forums to easily access health information written by and
for non-experts and also to get comfort from people who are in a similar situation. The government
operates medical web forums to provide medical information, manage patients’ needs and feelings,
and boost information-sharing among patients. If we can classify people’s emotional or information
needs by gender, age, or location, it is possible to establish a detailed health policy specialized
into patient segments. However, people with sensitive illness such as AIDS tend to hide their
information. Especially, in the case of sexually transmitted AIDS, we can detect problems and needs
according to gender. In this work, we present a gender detection model using sentiment analysis and
machine learning including deep learning. Through the experiment, we found that sentiment features
generate low accuracy. However, senti-words give better results with SVM. Overall, traditional
machine learning algorithms have a high misclassification rate for the female category. The deep
learning algorithm overcomes this drawback with over 90% accuracy.
Keywords: sentiment analysis; gender classification; machine learning; deep learning; medical
web forum
1. Introduction
Sentiment analysis is contextual mining of text that identifies and extracts subjective information
in source material, helping a business to understand the social sentiment of their brand, product,
or service while monitoring the online conversation. With recent advances in machine learning,
text mining techniques have improved considerably. Creative use of advanced artificial intelligence
techniques can be an effective tool for doing in-depth research.
Sentiment analysis of web content is becoming increasingly important due to augmented
communication through Internet sources such as e-mail, websites, forums, and chat rooms.
By collecting these articles and analyzing people’s emotions expressed in them, we can figure out
people’s feelings and opinions about polices, products, brands, and so on. Compared to traditional
surveys and other research techniques, this rich information can be obtained with less cost and
effort. People can search for particular information based on their individual needs. Patients,
patients’ significant others, or caregivers also use health-related web forums to get health and medical
information and to get comfort from people who are similar to themselves. They also ask questions

Appl. Sci. 2019, 9, 1249
about the disease and find information that is easy to understand [1]. The medical forum also reflects
their feelings [2]. The government relies on web forums to act as a helpdesk to promote people’s
well-being. Some governments directly operate medical web forums to provide medical information,
get to know patients’ needs, manage their emotions, and help people share information [3]. Users
divulge some pieces of personal information when they sign up, and this information can be used to
establish national health care policies. However, for sensitive disease such as AIDS, patients tend not
to expose their personal information. In terms of health policy, it is important to know who is looking
for more information. If we can guess the information from the non-disclosed data, we can understand
who is suffering from a specific disease and what types of worries they might have. Especially, in
sensitive diseases, by understanding patients we can help them deal with their situation properly and
adapt to society. The analysis of AIDS patients’ communication could be useful from a government
perspective, too.
Categorizing and analyzing according to demographic information such as gender, age, and
region is essential to obtain information such as consumers’ emotions, values, and attitudes in all areas
of marketing. In particular, it is important to distinguish gender for detailed policy establishment
because the situation and necessity can be different by gender. It can also help users to see what topics
are most talked about by males and females, and what services are liked or disliked by men and
women [4]. Knowing this information is crucial for market intelligence because the information can be
used in targeted advertising and service improvement.
Our study is based on a real-life application of a web forum. In this study, we collect data from
the AIDS-related bulletin board at Healthboards.com [5], which is one of the top 20 health websites
according to Consumer Reports Health Web Watch. Under the premise that “Emotions expressed by
men and women will be different,” we propose a model that extracts words and emotions from text
messages and distinguishes the gender. By establishing a learning model using gender information,
we figure out unclassified gender information and identify the gender awareness of AIDS patients
by gender.
2. Related Works
In most previous studies on gender classification, the various features fed to machine learning
algorithms. At the Islamic Women’s political forum, Zhang et al. built a machine learning model that
classifies genders by properly combining the characteristics of vocabulary, syntax, structure, uni-gram,
and bi-grams [6]. In the study of Ryu et al. [7], they used logistic regression and SVM (a support
vector machine) to estimate the gender, age, and location of Twitter users. Wang et al. [8] suggested
research based on the Latent Dirichlet Allocation (LDA) model using text data from Facebook or
Twitter, in which women have been shown to use a lot of personal themes while men tend to post a lot
of philosophical or informative text. In the study of Na and Cho [9], they surveyed the emotions of
male and female college students and showed that gender can be distinguished by analyzing emotions
using Fastcluster.
In the study of Yan et al. [10], they presented a naïve Bayes classification approach to identify the
gender of weblog authors. They used weblog-specific features such as webpage background colors
and emoticons. They report an F-measure of around 64% using their features. Mukherjee and Liu. [11]
proposed two new techniques. They used their own POS Sequence mining algorithm and an Ensemble
feature selection technique, and achieved an accuracy of 88%.
Pennacchiotti and Popescu [12] proposed a user classification model applying machine learning
algorithm to the feature set including user profile, user tweeting behavior, linguistic content of user
messages, and user social network features. They explored the feature importance and found that
linguistic features are consistently effective in user classification.
Dwivedi et al. [13] present two systems, a manual feature extraction system and a deep learning
method, to automatically classify the gender of blog authors. For the deep-learning-based model,
they apply a Bidirectional Long Short-Term Memory Network. Barlte and Zheng [14] report an
86
Appl. Sci. 2019, 9, 1249
accuracy of 86% in gender classification on blog datasets by applying deep learning models based on
the Windowed Recurrent Convolutional Neural Network (WRCNN).
Filho et al. [15] proposed textual meta-attributes, taking into account the characters, syntax, words,
structure, and morphology of short, multi-genre, content-free text posted to Twitter to classify an
author’s gender via three different machine learning algorithms. The novel contribution of this work is
to employ a word-based meta-attribute such as Ratio between hapax dislegomena (a word that appears
only twice in a whole text) and the total number of words. Furthermore, they developed a textual
morphology based on the meta-attributes from textual structures such as the ratio between the number
of pronouns and the total number of words. This work achieved 81% accuracy, but the performance
by each class is not presented. Garibo-Orts [16] proposed a statistical approach to the task of gender
classification in tweets. The statistical features include skewness, kurtosis, and central moments;
these statistical features make the learning algorithm context-free and language-independent.
Recently, Convolutional Neural Networks (CNNs) have also been successful in various text
classification tasks. Kim [17] showed that a simple CNN with little hyper-parameter tuning and static
vectors achieves excellent results. Severyn et al. [18] proposed a deep learning model for Twitter sentiment
classification. They advanced a CNN model to adjust word embedding using unsupervised learning.
Based on previous works, females and males are found to have differences in terms of the
vocabulary used and the emotional expression. During a literature review, we found some limitations
in the current literature. First, most recent works focus on Twitter or weblogs. Even though the
gender information disclosed in the medical web forum, especially on sexually transmitted diseases,
is important, we hardly found any work dealing with medical web forums. Secondly, studies using
sentiment features derived from text are lacking, while word features are explored a lot. Thirdly, deep
learning is applied a lot to text classification, but its application to gender classification from textual
information is rare.
In this work, we aim to figure out how the sentiment features derived from emotions expressed
in articles work in gender classification. We will do this by using both machine learning and deep
learning methods.
3. Experiment
We retrieved the users’ gender information from an AIDS-related medical web forum and
presented a gender detection model that classifies gender based on the emotions expressed in posts and
comments. We developed the sentiment feature set that expresses how often the posts contain emotions,
and assessed the emotional complexity, which is the number of emotion categories shown in a post.
Using vocabulary characteristics and emotions from the disclosed data, we built machine learning
models and measured the accuracy, varying the feature sets to select the best model. The proposed
framework is presented in Figure 1.
Figure 1. Gender classification model for medical web forum.
87
Appl. Sci. 2019, 9, 1249
3.1. Dataset
From 14 November 2000 to 25 February 2010, we collected 3451 posts and 18,944 comments—in
total, 22,395 messages—from an AIDS-related bulletin board from HealthBoard.com. It is one
of the first independent health community websites and has more than 280 bulletin boards for
communication about diseases. As shown in Table 1, this dataset has 8954 male-written messages
and 6078 female-written messages. A total of 3282 users participated in the AIDS bulletin board,
with 849 female users and 1054 male users. This implies that 1379 users have not disclosed their
gender. The average post length is 113 words for men and 106 words for women. The percentage of
users who disclosed gender information is 59% for males and 41% for females; 33% of the total has no
gender information. As for the Alzheimer’s-related board, only 1708 (6.2%) out of 27,874 (posts: 2486,
comments: 25,134) have not disclosed their gender. The number of users in the Alzheimer’s bulletin
board is 1498. The number of female users is 924 and male users 185. In total, 1109 users have gender
information. Compared to the users in the Alzheimer’s-related board, users in the AIDS-related board
have a low ratio of disclosed gender information.
Table 1. The gender information disclosed in an AIDS-related web forum.
Gender Numbers of Messages Ratio (%)

Male 8954 40
Female 6078 27
N/A 7363 33
Total 22,395 100
3.2. Feature Extractions

Before feature extraction, we first perform data preprocessing. We first converted everything
to lowercase, removing punctuations, numbers, and stopwords. Stopwords are words that have no
special meaning but are frequently used. For example, in English, “a,” “the,” and “is” are examples of
stopwords. In this work, we excluded these words to reduce the feature set in text mining, with the
assumption that the stopwords are no different in distribution between women and men.
Then, we build the tokenizer only on the training data. We could add the testing data, but results
may go up or down. Testing is needed to determine whether or not adding the testing data will help.
In most text classification studies, words are the most important features. Here are the features that we
considered in this work.
3.2.1. Unigram
We extracted the word features from the text to construct a learning model. The first step
was to derive words after preprocessing. We used the Bag-of-Words technique to tokenize each
word into a DocumentTermMatrix that indicates how many times the word appears in each article.
The Bag-of-Words technique transforms the sentence into a set of numeric vectors as follows. First,
the word set is built by collecting words from all sentences in all documents. Then, a sentence is
expressed as the word counts in word order. For example, in “symptoms appear after at least two”
and “it was two years of unnecessary stress,” these two sentences generate the word set, {symptoms,
appear, after, at, least, two, it, was, years, of, unnecessary, stress}. Two sentences are expressed as the
counts of each word as {1,1,1,1,1,1,0,0,0,0,0,0} and {0,0,0,0,0,1,1,1,1,1,1,1}, respectively. The numeric
vector is derived for a document by aggregating the word count for all sentences in a document. The
document term matrix (DTM) is represented in Figure 2.
88
Appl. Sci. 2019, 9, 1249
Figure 2. Representation of DocumentTermMatrix.
We extracted the document term matrix in two ways, the word occurrence and the occurrence
frequency, to test which feature extraction method is efficient.
To reduce the feature set, we set the threshold value to the number of appearances for extracting
word features. We incrementally increased the threshold value from 5 to 15 by 5s and checked the
classification performance.
3.2.2. Sentiment Features

To derive the sentiment features, we used tidytext, which is a dictionary package built in R,
to measure the rate of emotions. There are several dictionaries, but we used NRC [19] and BING.
The NRC dictionary has 10 categories: trust, fear, negative, sadness, anger, surprise, positive, disgust,
joy, and anticipation. Based on this, we developed sentiment-related features: the number of emotion
types expressed in a message, objectivity (1-number of used emotion types/total emotion (10)), and
emotional complexity (the number of emotion types/10).
The Bing dictionary classifies words only into positive or negative, and has 6788 words. Based on
the Bing dictionary, we calculated the positive rate (the number of positive words/the total number of
words), negative rate (the number of negative words/the total number of words that contain emotions),
and the total number of words.
As shown in Tables 2 and 3, women use the words ‘shine,’ ‘thank,’ ‘bless,’ and ‘glad’ (positive
words) and ‘problem,’ ‘scary,’ and ‘illness’ (negative words) about twice as often as men. On the other
hand, men use the words ‘accurate,’ ‘important,’ ‘receptive’ (positive words) and ‘issue,’ ‘fever,’ and
‘aches’ (negative words) more than twice as often as woman.
Table 2. Differences in positive words between men and women.
Word Men Women

Shine 1 419
Thank 655 714
Glad 154 233
Bless 168 267
Accurate 350 166
Important 196 76
Correct 149 54
Table 3. Differences in negative words between men and women.
Word Men Women

Problem 246 337
Scary 80 147
Illness 144 211
Issue 565 179
Fever 358 140
Aches 184 39
89
Appl. Sci. 2019, 9, 1249
3.3. Classification Algorithms

We employed a representative machine learning algorithm for text classification.
3.3.1. Naïve Bayes

Naïve Bayes is the first approach we tried. Let C = (c1 , c2 ) be the gender class, and F = ( f 1 , f 2 , . . .
f n ) are features, according to Bayes’s theorem:
P(c) P( F |c)
P(c| F ) = . (1)
P( F )
The naïve Bayes assumption is that:

n
P̂( F |c) = ∏ P̂( fi |c). (2)
i =1
Based on the naïve Bayes assumption, the probability of belonging to each class is as follows:
n
P̂( F |c) = argmaxP(C = c|F) = argmaxP(C|c) ∏ P( f = f i |C = c). (3)
c c
i =1
The basic assumption is that all features are independent. This assumption is too strong, but it
works well in reality.
3.3.2. Support Vector Machine

The support vector machine is known as a powerful algorithm until deep learning algorithm
replaced it [20]. Various lines can be drawn between classes A and B. If new data comes in and
overlaps with a line that is on one side, it is hard to judge whether those data belong to class A or
class B. Therefore, the best way to increase the classification accuracy is to deploy a line in a center
between groups with the maximum distance from the line to groups. The maximum distance from
the line to groups is called the margin. At the maximum margin, the middle boundary is called the
optical hyper-plane.
3.3.3. Random Forest

Random forest is an algorithm that generates multiple decision trees by randomly sampled data
for learning and assembles the results of the decision trees by majority voting. It is a type of ensemble
leaning method that implements classification, regression, and clustering based on group learning.
Random forest trains the model using the subset of features and the subset of data in a repetitive
way. This operation reduces the risk of overfitting. This algorithm has been widely used in various
applications in recent years and has been proved to outperform other algorithms.
3.4. Deep Learning Approach

The deep learning algorithm is a kind of machine learning. Machine learning focuses on solving
real-world problems with neural networks created by mimicking the decision-making processes that
take place in the human brain by adopting some of the main ideas of artificial intelligence. Deep
learning focuses more on the tools and techniques of machine learning. The difference between
machine learning and deep learning is that, in the course of machine learning, people are still involved
in the process of extracting features from the given data to train the machine, but deep learning uses
the given data as input data. It does not feature an engineering process to extract the features of
training by human intervention, but automatically performs learning of important features in the data
itself. So deep learning is called end-to-end machine learning. This is because the machine learns from
beginning to end, by itself, without human intervention.
90
Appl. Sci. 2019, 9, 1249
3.4.1. Convolution Neural Network

CNNs are widely used on image datasets [21,22], but researchers have found that they work
great with text, too. Previous works showed that CNN outperforms the state-of-art algorithm in
many cases [21,22] That is because text can be considered a 1D image. Now, the explanation for why
CNNs work for images is quite complicated, but they (partly) revolve around something called ‘spatial
locality’. What this means is that CNNs can find patterns not just from each feature (e.g., a pixel) but
from locations of pixels (e.g., neighborhoods of pixels correlated with another). This makes sense
because images are not a random collection of pixels, but pixels are connected across the image.
3.4.2. Convolution Neural Network (CNN) for Text Classification

To make the sentences in text into an image, we need to encode words with numeric values. Word
embedding is employed to express words as numeric values. The method that assigns a unique integer
value to each word would cause the dimension explosion. Word embedding set the dimensionality,
k, and expresses the word using k-dimensional values. For example, when k is set to 5, “hospital”
is expressed as {0.1,0.2,0.3,0.8,0}. This embedding method dramatically reduces the dimensions.
However, this method does not consider the relationship between words when placing a word in a
matrix. The advanced technique is to consider the relationship between words when transforming
the input data into an input matrix like an image. Word2Vec considers that words that appear in a
similar context have similarities in semantic meaning. We adopted Google’s word2Vec [23], which is a
pretrained model from the GoogleNews dataset.
We performed the experiment using word-embedding method and Word2Vec method as well.
The architecture of the text classification is shown in Figure 3. We transformed words into a
k-dimensional vector. We set the number of words to make an input matrix, denoted as n. A sentence is
tokenized and n words organize the input data. At the end of the sentence, there could not be enough
words to fill the input. In this case, we pad the rows with zeros until the number n is filled.
Figure 3. Model architecture with an example sentence.
To derive abstract features from input data and make the algorithm derive useful features
automatically, we employ a hidden layer and a convolutional layer to the network.
In a convolution layer, we feed the input matrix to the filter of k*h size, then the input of h*k
size is mapped to a cell in the convolutional layer. In Figure 3, the rectangle with a red line is a 3*k
filter. The filter works as an aggregator of h words. The filter needs to keep the word meaning when
91
Appl. Sci. 2019, 9, 1249
aggregating words, so the horizontal size of the filter is set to k. As a result, the convolutional layer is
one-dimensional. The number of nodes in a convolution layer is calculated as n-h + 1.
The vertical size of the filter can be varied. In this work, we test two sizes of filter, 3 (in red) and 8
(in yellow). The number of filters depends on the user’s decision. More filters are needed to enrich
features. The max pooling layer is also employed to avoid overfitting a result of too much features.
The max pooling layer adopts a filter with the size of (n-h + 1)*1 and finds the maximum value for all
values in a precedent convolutional layer. The max pooling layer generates one-dimensional m rows.
Finally, m rows are fully connected to the output. The output layer has two nodes to express bi-class,
gender in our case. Two nodes are filled with 1 or 0, it indicates male or female respectively.
4. Results
4.1. Traditional Machine Learning Results

Initially, 70% of data, including gender information, were used as training data and 30% as test
data. We experimented with naïve Bayes (NB), support vector machine (SVM), and random forest (RF),
which are representative algorithms proven to outperform other algorithms. The precision measure
and the recall measure are calculated for each class. For the female class, the precision is calculated as
TP/(TP + FP) and recall is calculated as TP/(TP + FN). For the male class, the precision and recall are
calculated as FN/(FN + TN) and TN/(FP + TN), respectively. The positive class is defined as the class
of interest. In this work, the female class and male class are treated equally, but the positive class is
set as the female class. The accuracy, the true positive rate, and the true negative rate are calculated
considering two classes. Accuracy indicates how well a binary classification test correctly identifies
each class. False positive rate is calculated as FP/(FP + TN) and indicates the ratio of the wrongly
classified females among the males. False negative rate is calculated as FN/(FN + TP). Using only
sentiment features, overall, the female class has poor performance. Three algorithms did not recognize
the female classes well and classified most of the users into the male class. This indicates that sentiment
features such as emotion types and their density are not effective for classifying authors’ gender on
their own.
To test the effects of sentiment features, we performed three experiments as follows. At first,
we only built a sentiment feature set including the ratio of 10 types of emotional occurrences, objectivity,
and emotional complexity. The accuracy of them are 58.33% (NB), 60.86% (SVM), and 58.66% (RF),
respectively, as shown in Table 4. Compared to the male class, the female class has a poor performance.
Over 50% of female users were classified into the male class.
Table 4. The accuracy of gender detection using sentiment feature set.
Algorithm NB SVM RF
Real
Female Male Female Male Female Male
Female 174 (True positive: TP) 200 (False negative: FN) 81 42 402 413
Male 1679 (False positive: FP) 2456 (True negative: TN) 1723 2663 1451 2243
Precision 9.39% 92.47% 4.49% 98.45% 21.69% 84.45%
Recall 46.52% 59.40% 65.85% 60.72% 49.33% 60.72%
Accuracy 58.33% 60.86% 58.66%
False positive rate 40.60% 39.28% 39.28%
False negative rate 53.48% 34.15% 50.67%
The second experiment was performed using senti-word features. Instead of calculating sentiment
value from senti-word, we checked the senti-word occurrence. Among the 7954 words that appeared
more than five times each, we selected only 1594 senti-words as features that are in the NRC
dictionary. The accuracy increased a little bit with the SVM and RF algorithms, as shown in Table 5.
The precision for the female class improved. The senti-words used to express the author’s emotion
were comparatively effective for gender classification. Among the three algorithms, SVM had the
highest performance for the sentiment feature set and the senti-word feature set.
92
Appl. Sci. 2019, 9, 1249
Table 5. The accuracy of gender detection using senti-word feature set.
NB SVM RF
Female 1706 2434 701 236 514 210
Male 117 252 1122 2450 1309 2470
Precision 93.58% 9.38% 38.45% 91.21% 28.20% 92.16%
Recall 41.21% 68.29% 74.81% 68.59% 70.99% 65.36%
Accuracy 43.42% 69.88% 66.27%
The final experiment was performed using the entire feature set, combining word features and
sentiment features. When we assessed all 7954 words that occurred more than five times along with
the sentiment features, the accuracy improved. However, the performance of SVM rather decreased.
In this case, RF has the highest performance and the precision for the female class somewhat improved.
However, the female class still has a poor performance as shown in Table 6.
Table 6. The accuracy of gender detection using entire feature set (word features + sentiment features).
NB SVM RF
Female 1838 2623 1 0 754 135
Male 14 34 1822 2686 1069 2551
Precision 99.24% 1.28% 0.05% 100.00% 41.36% 94.97%
Recall 41.20% 70.83% 100.00% 59.58% 84.81% 70.47%
Accuracy 41.52% 59.55% 73.33%
Based on the three experiments varying the feature set, we obtained two findings, one is
the algorithm perspective and the other is the feature perspective. We can conclude that SVM
underperforms other algorithms in the case of a few features. On the other hand, RF works well with
more features. In addition, we can conclude that NB is not appropriate for text classification task.
For feature perspective, words are good features for gender classification in the web forum. Types of
emotion and their density are not enough, although these sentiment features reduce the computational
complexity with a small number of features. However, when using with SVM, the senti-word feature
set generates the comparable performance to the word feature set.
Even though words are good features, we cannot employ all words because we need to consider
the computational complexity and cost to make the algorithm work. Machine learning algorithms
support a limited number of features, unlike the deep learning algorithm, and we need to reduce the
number of words included in the feature set. For feature reduction, we explored the threshold value to
extract words from 5 to 100, as shown in Table 7. The number of words involved in learning is displayed
at the end of Table 7. We also tested which feature extraction method among the occurrence with
boolean and the term frequency is useful. From the experiments results, the occurrence with boolean
outperforms the term frequency. In naïve Bayes, which assumes the independence of the variables,
the number of words decreases when setting a higher threshold value, and the performance improves.
SVM has the opposite behavior with NB. The best performance is achieved when applying the RF
model to the term frequency features with a threshold value of 10, resulting in 4822 words. However,
the performance of RF begins to decline beyond a certain number of features from the first two rows of
Table 7. From all the experiments, we could conclude that traditional machine learning algorithms are
93
Appl. Sci. 2019, 9, 1249
not good at gender classification, resulting in a large performance gap between two classes. In the next
section, we will see how much the deep learning algorithm improved the performance.
Table 7. The accuracy of gender detection using different feature set.
Feature|Accuracy NB (%) SVM (%) RF (%)

TF/Boolean (5) 1 + senti 41.52/41.32 59.55/59.57 73.3/72.79
TF/Boolean (10) 2 + senti 44.95/45.13 70.68/71.66 74.41/73.59
TF/Boolean (15) 3 + senti 49.48/51.08 71.72/72.79 73.16/73.05
TF/Boolean (100) 4 + senti 61.5/66.16 71.99/72.52 72.9/73.52
1 including 7588 words, 2 including 4822 words, 3 including 4732 words, 4 including 1108 words.
4.2. Deep Learning Results

We tested two methods of embedding words into vectors, random word-embedding and the
word2Vec method. We also varied the CNN structure with two filter sizes of length 3 or 8, which
implies that three words or eight words are aggregated. The number of filters is set to 10. For parameter
optimization, we used ADAM [24]. This uses squared gradients to scale the learning rate and takes
advantage of momentum by using the moving average of the gradient instead of the gradient
itself. The experiment results show that the random word-embedding method outperforms the
word2Vec method.
We guess that word2Vec is built with formal text from GoogleNews, but our text dataset is written
in spoken language. Thus, word2Vec does not work well in our case. The accuracy of the random
word embedding method ranges between 88% and 91% depending on the CNN structure. On the
other hand, the accuracy with the word2Vec ranges between 67% and 71%.
Secondly, we varied the feature set from word features to senti-word features. The sentiment
feature set is low-dimensional with a small number of features, so we judged that deep learning is not
necessary for this feature set. Thus, we tested CNN for the reduced feature consisting of senti-words
and the full feature set consisting of word features. We achieved 88.7% with senti-word features and
90.6% with word features. The number of senti-word features is at 20% of the number of word features,
but it generates a comparable performance. Compared to traditional machine learning algorithms,
CNN dramatically increases the performance. For the female class, the precision and recall are much
improved and the performance for the two classes is equally good.
We changed the CNN structure to find the best structure for our context. Starting from a simple
structure, we designed more complicated structures. The combination of two convolution layers
proceeded by the pooling layer, the dropout layer after that, and the fully-connected layer at last,
outperformed other structures. A more complicated model with two convolutional layers outperforms
the simplest model with one convolutional layer. However, the accuracy increases by at most 0.004.
Slight modifications to the baseline model, in the second row, degrade the accuracy. From the other
models, we found one fully connected layer is enough for comparing the second through fourth
models. The max pooling layer employed to reduce feature dimension is best deployed right after
each convolutional layer. In future work, we will test more complicated structures and more diversity
in terms of the number of filters and sizes. Table 8 indicates the result of CNN performance according
to each structure.
Table 8. Comparison of CNN performance according to the structures.
Female Male
CNN Structure Accuracy
Precision Recall Precision Recall
Conv + Pool + Dropout + FC 0.906 0.93 0.91 0.87 0.90
Conv + Pool + Conv + Pool + Dropout + FC 0.910 0.93 0.92 0.88 0.89
Conv + Pool + Conv + FC + Dropout + FC 0.905 0.94 0.89 0.86 0.92
Conv + Conv + Pool + FC + Dropout +FC 0.887 0.91 0.90 0.85 0.87
94
Appl. Sci. 2019, 9, 1249
5. Discussion and Conclusions

In this study, we constructed a model to detect gender information by using machine learning
algorithms based on word and sentiment feature sets. We developed various sentiment feature sets
including the number of emotion types expressed in a message, objectivity, and emotional complexity.
To check the effectiveness of sentiment features in gender classification, we varied the feature sets,
including sentiment, senti-word, sentiment+word feature sets and using machine learning algorithms
or deep learning algorithms. From the experiments, we found that sentiment features with a small
number of features and domain-independent are not enough for classifying gender in the medical web
forum. Words should be incorporated to classify the author’s gender. We found that it is possible to
construct gender classification models by using a sentiment+word feature set. In cases of a lightweight
system with a limited number of features, SVM performed moderately with senti-word features.
However, overall, the traditional machine learning algorithm failed, having low precision in the
female class. Many male authors were misidentified as female authors. When we applied the deep
learning algorithm, both gender classes were identified well. The senti-word feature set and word
feature set both worked well, with a 2% accuracy gap. The number of senti-word features is 20% of the
number of word features, but generates a comparable performance.
As a design issue, to embed text into a matrix form, we tested word-embedding methods, one
random and the other pre-built word2Vec. To find the better model, we varied the CNN structure by
adding and deleting hidden layers, the convolutional layer, and the max pooling layer.
This work poses remaining challenges as follows. We tested the word2Vec built by Google and
found that it does not work well for our dataset. In future work, we will build our own word2Vec
suitable for medical web forums by adopting Google’s training method. A consideration of word
similarity would improve the detection performance. Secondly, in the current work, we eliminated
stopwords, numbers, and special characters to reduce the feature set. However, previous work has
suggested that these words could be helpful in the text classification task. In the deep learning model,
feature reduction is not necessary, so we will check their effectiveness. In addition, we intend to
add features in the context of social networks by analyzing the relationship between the posts and
comments in order to improve the accuracy of the proposed model. The social network features are
also generic across disease types and forums.
Author Contributions: S.P. performed the experiments and writing. J.W. collected the dataset and participated
in writing.
Funding: This work was supported by the research fund of Soonchunhyang University (project number 20171094)
and a National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIP; Ministry of
Science, ICT & Future Planning (NRF-2017R1D1A3B03036050)).
Acknowledgments: The authors gratefully acknowledge the support by the research fund of Soonchunhyang
University (project number 20171094) and a National Research Foundation of Korea (NRF) grant funded by the
Korean government (MSIP; Ministry of Science, ICT & Future Planning (NRF-2017R1D1A3B03036050)).
Conflicts of Interest: The authors have no conflicts of interests.
References
1. Weaver, J.B., III; Mays, D.; Weaver, S.S.; Hopkins, G.L.; Eroğlu, D.; Bernhardt, J.M. Health information–seeking
behaviors, health indicators, and health risks. Am. J. Public health 2010, 100, 1520–1525. [CrossRef] [PubMed]
2. Woo, J.; Lee, M.J.; Ku, Y.; Chen, H. Modeling the dynamics of medical information through web forums in
medical industry. Technol. Forecast. Soc. Chang. 2015, 97, 77–90. [CrossRef]
3. Denecke, K.; Nejdl, W. How valuable is medical social media data? Content analysis of the medical web.
Inf. Sci. 2009, 179, 1870–1880. [CrossRef]
4. Sullivan, C.F. Gendered cybersupport: A thematic analysis of two online cancer support groups. J. Health
Psychol. 2003, 8, 83–104. [CrossRef] [PubMed]
5. Healthboard. Available online: https://www.healthboards.com/ (accessed on 25 March 2019).
95
Appl. Sci. 2019, 9, 1249
6. Zhang, Y.; Dang, Y.; Chen, H. Gender classification for web forums. IEEE Trans. Syst. Man Cybern. Part A
Syst. Hum. 2011, 41, 668–677. [CrossRef]
7. Ryu, K.; Jeong, J.; Moon, S. Inferring Sex, Age, Location of Twitter Users. J. KIISE 2014, 32, 46–53.
8. Wang, Y.-C.; Burke, M.; Kraut, R.E. Gender, topic, and audience response: An analysis of user-generated
content on facebook. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems,
Paris, France, 27 April–2 May 2013; pp. 31–34.
9. Na, Y.; Cho, G. Grouping preferred sensations of college students using sementic differential methods of
sensation words. Korean J. Sci. Emot. Sensib. 2002, 5, 9–16.
10. Yan, X.; Yan, L. Gender Classification of Weblog Authors. In Proceedings of the AAAI Spring Symposium:
Computational Approaches to Analyzing Weblogs, Palo Alto, CA, USA, 27–29 March 2006; pp. 228–230.
11. Mukherjee, A.; Liu, B. Improving gender classification of blog authors. In Proceedings of the 2010
Conference on Empirical Methods in Natural Language Processing, Cambridge, MA, USA, 9–11 October
2010; pp. 207–217.
12. Pennacchiotti, M.; Popescu, A.-M. A machine learning approach to twitter user classification. In Proceedings
of the Fifth International AAAI Conference on Weblogs and Social Media, Barcelona, Spain, 17–21 July 2011.
13. Dwivedi, V.P.; Singh, D.K.; Jha, S. Gender Classification of Blog Authors: With Feature Engineering and Deep
Learning using LSTM Networks. In Proceedings of the 2017 Ninth International Conference on Advanced
Computing (ICoAC), Chennai, India, 14–16 December 2017; pp. 142–148.
14. Bartle, A.; Zheng, J. Gender Classification with Deep Learning; Stanford cs224d Course Project Report;
The Stanford NLP Group: Stanford, CA, USA, 2015.
15. Lopes Filho, J.A.B.; Pasti, R.; de Castro, L.N. Gender classification of twitter data based on textual meta-attributes
extraction. In New Advances in Information Systems and Technologies; Springer: Berlin/Heidelberg, Germany,
2016; pp. 1025–1034.
16. Garibo-Orts, O. A big data approach to gender classification in twitter. In Proceedings of the Ninth
International Conference of the CLEF Association (CLEF 2018), Avignon, France, 10–14 September 2018.
17. Kim, Y. Convolutional neural networks for sentence classification. arXiv 2014, arXiv:1408.5882.
18. Severyn, A.; Moschitti, A. Unitn: Training deep convolutional neural network for twitter sentiment
classification. In Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015),
Denver, CO, USA, 4–5 June 2015; pp. 464–469.
19. Mohammad, S.M. Challenges in sentiment analysis. In A Practical Guide to Sentiment Analysis; Springer:
Berlin/Heidelberg, Germany, 2017; pp. 61–83.
20. Nayak, J.; Naik, B.; Behera, H. A comprehensive survey on support vector machine in data mining tasks:
Applications & challenges. Int. J. Database Theory Appl. 2015, 8, 169–186.
21. Zhang, Y.D.; Dong, Z.; Chen, X.; Jia, W.; Du, S.; Muhammad, K.; Wang, S.H. Image based fruit category
classification by 13-layer deep convolutional neural network and data augmentation. Multimedia Tools Appl.
2019, 78, 3613. [CrossRef]
22. Wang, S.H.; Sun, J.; Phillips, P.; Zhao, G.; Zhang, Y.D. Polarimetric synthetic aperture radar image
segmentation by convolutional neural network using graphical processing units. J. Real-Time Image Process.
2018, 15, 631. [CrossRef]
23. word2Vec. Available online: https://code.google.com/archive/p/word2vec/ (accessed on 25 March 2019).
24. Kingma, D.P.; Ba, J.L. Adam: A method for stochastic optimization. arXiv 2014, arXiv:1412.6980v9.
96
applied
sciences
Article
Personality or Value: A Comparative Study of
Psychographic Segmentation Based on an Online
Review Enhanced Recommender System
Hui Liu 1,† , Yinghui Huang 1,2,† , Zichao Wang 3 , Kai Liu 1 , Xiangen Hu 1,4 and Weijun Wang 1, *
1 Key Laboratory of Adolescent Cyberpsychology and Behavior, Ministry of Education,
Central China Normal University, Wuhan 430079, China; [email protected] (H.L.);
[email protected] (Y.H.); [email protected] (K.L.); [email protected] (X.H.)
2 School of Information Management, Central China Normal University, Wuhan 430079, China
3 Department of Electrical and Computer Engineering, Rice University, Houston, TX 77005, USA;
[email protected]
4 Department of Psychology, The University of Memphis, Memphis, TN 38152, USA
* Correspondence: [email protected]; Tel.: +86-153-0715-0076
† Hui Liu and Yinghui Huang are co-first authors.

Received: 1 March 2019; Accepted: 8 May 2019; Published: 15 May 2019
Abstract: Big consumer data promises to be a game changer in applied and empirical marketing
research. However, investigations of how big data helps inform consumers’ psychological aspects
have, thus far, only received scant attention. Psychographics has been shown to be a valuable
market segmentation path in understanding consumer preferences. Although in the context of
e-commerce, as a component of psychographic segmentation, personality has been proven to be
effective for prediction of e-commerce user preferences, it still remains unclear whether psychographic
segmentation is practically influential in understanding user preferences across different product
categories. To the best of our knowledge, we provide the first quantitative demonstration of the
promising effect and relative importance of psychographic segmentation in predicting users’ online
purchasing preferences across different product categories in e-commerce by using a data-driven
approach. We first construct two online psychographic lexicons that include the Big Five Factor (BFF)
personality traits and Schwartz Value Survey (SVS) using natural language processing (NLP) methods
that are based on behavior measurements of users’ word use. We then incorporate the lexicons
in a deep neural network (DNN)-based recommender system to predict users’ online purchasing
preferences considering the new progress in segmentation-based user preference prediction methods.
Overall, segmenting consumers into heterogeneous groups surprisingly does not demonstrate
a significant improvement in understanding consumer preferences. Psychographic variables (both BFF
and SVS) significantly improve the explanatory power of e-consumer preferences, whereas the
improvement in prediction power is not significant. The SVS tends to outperform BFF segmentation,
except for some product categories. Additionally, the DNN significantly outperforms previous
methods. An e-commerce-oriented SVS measurement and segmentation approach that integrates both
BFF and the SVS is recommended. The strong empirical evidence provides both practical guidance for
e-commerce product development, marketing and recommendations, and a methodological reference
for big data-driven marketing research.
Keywords: psychographic segmentation; user preference prediction; lexicon construction; online review;
recommender system; big data-driven marketing

Appl. Sci. 2019, 9, 1992
1. Introduction
In big data era, the research paradigms of marketing service have been greatly changed by the
enormous marketing data accumulated from the internet, such as on demographics, user behavior
(we will henceforth use the word “user”, “customer”, and “consumer” interchangeably), and social
relationships. These fine-grained marketing data are informative, thus providing marketers with extra
opportunities to evaluate users’ preferences [1], predicting the next product users will buy [2,3], delivering
targeted advertisements [4,5], uncovering consumer perceptions of brand [6], and acquiring competitive
intelligence [7]. In particular, psychological aspects of user behavior, critical for understanding rather than
merely predicting consumer preference, ultimately contribute to intelligent decision-making in marketing.
Unfortunately, investigations of how big data helps inform has, thus far, only received scant attention [8].
Studies have demonstrated that psychological variables, such as value and personality, are the
main theories and tools for user psychographic segmentation, which is an important determinant of user
purchase behaviors and preferences [9,10]. In particular, personality traits, one of the psychographic
segmentation components, has been the main consumer psychological and marketing characteristic
used in e-commerce. Unfortunately, the predictive and explanatory power of personality traits for
online user behavior remains controversial [11]. This controversy motivates the investigation of other
types of psychographic segmentation, such as value, in understanding user preferences, which have
not been investigated. Furthermore, it is unclear whether the predictive and explanatory power
varies between product categories and between different segment-wise preference prediction methods.
The main reason is that collecting psychographic segmentation data from e-commerce users is difficult
on a large scale, typically requiring consumers to complete lengthy questionnaires.
Psychological variables, such as personality and value, are deeply embedded in the language
that people use today [12]. With massive user data in the form of natural language, natural language
data provides a clearer picture of people’s cognitive and behavioral processes than data collected from
traditional and widely used self-report surveys [13–15]. Practically, natural language processing (NLP)
techniques can be applied to identifying psychographic variables, such as e-commerce users’ online
word use, to understand and predict users’ purchase behaviors and preferences on a large scale.
E-commerce websites, in particular, have accumulated a large amount of user-generated
content (UGC), which provides the basis for observing users’ psychographics and predicting user
preferences directly. With the rapid development of techniques such as big data, artificial intelligence,
and computational linguistics, UGC provides a reliable path for automatically identifying consumer
psychographics, including personality and values, based on unstructured data. Inspired by recent
advances in big data-driven psycholinguistic research, which indicate the behavioral evidence of online
word use related to psychographics [16,17], we base our research on real-world Amazon consumer
review and rating data. We propose psychographics-related word use evidence, extract consumers’
psychographic characteristics using sentiment analysis methods, and introduce a deep neural network
(DNN) to predict and explain user preferences in e-commerce.
We found that, overall, psychographic variables significantly improved the explanatory power
of e-consumer preferences across most product categories we studied, whereas the improvement
in predictive power was not significant. Specifically, the Schwartz Value Survey (SVS) tended to
outperform Big Five Factor (BFF) segmentation in predicting and explaining user preferences, with the
exception of a few product categories. However, somewhat surprisingly, dividing e-consumers into
heterogeneous groups using a clustering method did not significantly improve the predictive and
explanatory power of online consumer preferences compared to handling consumers as a whole.
Furthermore, the DNN method that we proposed demonstrated the best performance in understanding
e-consumer preferences, and regarding product category, there were more significant differences for
psychographic segmentation in predicting, than explaining, e-consumer preferences. Additionally,
we recommend an e-commerce-oriented SVS measurement and segmentation approach that integrates
both BFF and the SVS.
98
Appl. Sci. 2019, 9, 1992
Our work extends the depth and breadth of psychographic theories through user preference
prediction in e-commerce scenarios. Specifically, we found that subdimensions of psychographic
variables and product types provide practical references for psychographic measurement development
and applications for the specific e-commerce product categories that we studied. By introducing
psychographic-related word use behavioral evidence, followed by big data approaches, we have
attempted to overcome the difficulties of obtaining e-consumer psychographics on a large scale,
and have provided a promising psychographic-based consumer preference prediction method for
subsequent research.
2. Literature Review and Research Design

In this section, we review the foundational work on consumer psychographics and trace its
relationship with user preferences, particularly in online scenarios. Building on the current development
in online behavioral measures of psychographics, we then introduce and propose related methods and
techniques in computer science and psycholinguistics. Finally, we organize and outline our research
questions and framework.
2.1. Psychographic Segmentation

Market segmentation has had a longstanding history since Smith (1956) [18] first suggested it
as a product differentiation strategy to increase competitiveness. It has been widely acknowledged
as a fundamental tool in understanding customer behaviors [19,20]. Wedel and Kamakura (2012)
defined segmentation as “a set of variables or characteristics used to assign potential customers to
homogeneous groups” [21]. Segmentation is critical because a company has limited resources and
must focus on how to best identify and serve its customers.
Initially, market segmentation was mainly rooted in personality profiles. The most frequently used
scale for measuring general aspects of personality as a way to define homogeneous submarkets is the
Edwards Personal Preference Schedule. Generally, however, early studies based on Edwards Personal
Preference Schedule were plagued with low and even inconsistent correlations with customer behavior
and preferences, and hence failed to satisfy marketers’ needs [22]. One of the main reasons was that
the studies used standardized personality tests that were originally developed in clinical or academic
instead of business contexts [23]. Since then, the Big Five Factor personality traits—that frequently
and consistently appeared in most attempts to define the basic human factors, that is, neuroticism,
extraversion, agreeableness, conscientiousness and openness—have replaced the Edwards Personal
Preference Schedule.
Lifestyle is defined as a set of behaviors that mirrors individual psychological considerations
and sociological consequences [24], and is more valid in shaping individual behavior compared
to personality traits. The lifestyle construct used in market segmentation is based on research on
motivation [23] and personal values [22]. Mitchell (1994) proposed that lifestyle is a mixture of
a personal life and perceived value, whereas value is a synthesis of individual attitudes, beliefs, hopes,
prejudices, and demands [25]. Activities, interests, and opinions (AIO) is one of the most widely
used lifestyle measurement tools [23]. It encompasses variables from rational, specific, and behavioral
psychology fields as well as from sociodemographic attributes, and provides good insight into the
lifestyles of individuals [26]. Indeed, individuals’ value priorities are part of their basic world views,
and are commonly defined as desirable, trans-situational goals that vary in importance and serve as
guiding principles in individuals’ lives [27]. Moreover, values are broader in scope than attitudes or
the types of variables contained in AIO measures, and can be an important basis for segmentation.
Generally, value inventories often only contain a handful of values instead of 200 or 300 AIO items.
Therefore, scholars have begun to use simple, intuitive, and elegant values as another lifestyle-related
psychographic segmentation method to replace the very extensive and burdensome AIO approach [22].
Early value-related works widely used three lifestyle measurements that are simpler alternatives
to AIO: value, attitude, and lifestyle (VALS); list of values (LOV); and the Rokeach Value Survey
99
Appl. Sci. 2019, 9, 1992
(RVS). One of the most widely used value segmentation theories is the Stanford Research Institute’s
consulting values and lifestyles (VALS2) typology, which contains 35 psychographic questions and four
demographic questions that link demographics and purchase patterns with psychological attitudes [28].
The RVS proposed by Rokeach (1973) [29], LOV [30] developed by the Survey Research Center at
the University of Michigan, and SVS [31] developed by Schwartz et al. approximately two decades ago,
are three widely used instruments to assess individual values and lifestyles. RVS requires respondents
to rank 18 terminal values and 18 instrumental values. The terminal values are considered to be either
self-centered or society-centered, and intrapersonal or interpersonal in focus, whereas the instrumental
values are moral values and competence. LOV, proposed by Kahle, Beatty, and Homer (1986) [30],
is a shorter and more easily implemented instrument which includes only nine values. SVS is a more
acceptable and widely used value instrument nowadays, and suggests that there are 10 primary values
organized into a circumplex [27], as shown in Figure 1. This circumplex serves as the umbrella under
which the majority of individual value judgments fall. Schwartz (2012) noted the shared motivational
emphases of adjacent values: power (Pow), achievement (Achiev), hedonism (Hed), stimulation (Stim),
self-direction (S-D), universalism (Univ), benevolence (Benev), tradition (Trad), conformity (Conf),
and security (Sec) [27]. Although the theory underlying SVS discriminates 10 values, it postulates
that, at a more basic level, values form a continuum of related motivations. The Schwartz value
implies that the entire set of 10 values relates to any other variable (e.g., behavior, attitude, age, etc.) in
an integrated manner.
Self-Direction Universalism
Stimulation Benevolence
Hedonism Conformity
Tradition
Achievement
Power Security
Figure 1. Schwartz Value Survey circumplex.
2.2. Psychographic Segmentation and Consumer Behavior

Market segmentation consists of viewing a heterogeneous market as a number of smaller
homogeneous markets, with differing product preferences among important market segments for the
purpose of developing better firm strategies. Consumer behavior and predictive preference models
can be much more effective when the target audience is split into different customer segments and
individually tailored predictive models are developed for each segment [32], with some of them based
on psychographics.
Researchers have based the psychographic segmentation of markets on various factors, including
personality, lifestyle, and values. Early psychographics mainly considered personality, but research
based on one of the personality profiles, that is, Edwards Personal Preference Schedule, failed in
predicting consumer purchase decisions. Other scholars used the BFF personality model. Hirsh,
Kang, and Bodenhausen (2012) surveyed people about various marketing messages and found that
people respond more positively to messages tailored to their personality [33]. In the music domain,
I Fernández-Tobías et al. (2016) showed that online music recommendations are more successful
100
Appl. Sci. 2019, 9, 1992
when they leverage the correlations between people’s personality and their music preferences [34].
Similarly, Karumur, Nguyen, and Konstan (2016) discovered correlations between personality and
movie preferences among Netflix users [35]. Additionally, studies in marketing have found that
personality explains only a surprisingly small amount of the overall variance in consumer behavior [11].
Various studies have been conducted on the effect of user lifestyles on consumers’ products
purchases [36] and preferences [37,38]. The comprehensive results have confirmed the effect of
consumer lifestyle in predicting users’ online preferences. Lee et al. (2009) assessed consumer lifestyles
regarding the adoption of electronic products, and the results based on structural equation modeling
demonstrated that consumer lifestyle factors (fashion consciousness, leisure orientation, internet
involvement, and e-shopping preferences) are direct and indirect antecedents of consumers’ intentions
to adopt high-tech products [36]. Based on the AIO survey, Piazza et al. (2017) investigated the
relationship between lifestyle and online user preferences, and found that there are weak monotonic
correlations between Facebook likes and the lifestyles of individuals [38]. Based on a web-usage
lifestyle scale developed by AIO, Pan et al. (2014) applied a tagging method to identify online lifestyles
and to make recommendations according to the similarity of a user lifestyle tag to a telecommunications
company product lifestyle tag [37].
In addition to lifestyle and personality, the relationship between value and consumer preferences
has also been asserted. Scholars have used the LOV construct in relation to consumer fashion
leadership [39], involvement [40], and brand preferences [41]. Based on questionnaire data,
Weng and Run (2013) investigated the effects of Malaysian consumers’ personal values using mainly
LOV measurements with regard to their overall behavioral intention and purchase satisfaction [42].
Fraj and Martinez (2006) adopted VALS and identified which values and lifestyles best explain
environmentally friendly behaviors [43]. Based on RVS, Padgett and Mulvey (2007) demonstrated that
technology must be leveraged via links to customer values [44]. Wiedmann, Hennigs, and Siebels (2009)
constructed a luxury-related value for identifying different types of luxury consumers, and found that
hedonic value aspects as components of the SVS of self-directed pleasure and life enrichment were
most important for their perception of luxury value [45].
The combination of segmentation and prediction, where segmentation is used to help build
segment-wise prediction models, has been a common segmentation approach and has been applied
to different business scenarios. Antipov and Pokryshevskaya (2010) conducted user segmentation
research based on the decision tree method and then built a separate logistic regression scoring model
for each segment using the churn dataset [46]. Reutterer et al. (2006) proposed a two-stage approach for
deriving behaviorally persistent segments and predicting a customer’s tendency to symptomatically
(re)purchase using retail purchase history data [47]. Ge et al. (2011) exploited multifocal learning to
divide consumers into different focal groups and automatically predict a problem category based on
real-world customer problem log data [48].
2.3. Behavioral Measures of Psychographic Segmentation

Although it plays a promising role in predicting online user preferences, psychographic
segmentation has not been widely studied in online shopping scenarios. There are two main
reasons. First, data collection for such research typically requires online customers to complete lengthy
surveys and, therefore, cannot be easily applied on a large scale. Second, acquiring the abstract
psychographic characteristics of online customers through the mandatory selection of questionnaires
is plagued by issues of reliability and validity for the questionnaire [16]. At the present time, with the
in-depth development of big data techniques, researchers can access a large number of open-ended
reports of user psychological characteristics embedded in user-generated content (UGC). In recent
years, an increasing number of studies have demonstrated that such reports are ecologically valid and
driven entirely by what people say they are doing and thinking, in their own words [16].
Psychology and marketing research indicate that human language reflects psychological
characteristics. The frequency with which we use certain categories of words provides clues to
101
Appl. Sci. 2019, 9, 1992
these characteristics. Several researchers have found that variations in word use embedded in
writing such as blogs, essays, and tweets can predict aspects of personality, thinking style, social
connections, and purchase decisions [49–51]. These works, that explore the feasibility of deriving
psychological characteristics from UGC text, have demonstrated that computational models based
on derived psychological characteristics have competitive performance compared with models using
a self-reported questionnaire [52,53]. Research indicates cases in which natural language data have
provided a clearer picture of people’s cognitive and behavioral processes than data collected from
a traditional and widely used self-report survey [16].
As one of the psychographic segmentations, human values are thought to become generalized
across broad swaths of time and culture, and are deeply embedded in the language that people use
every day [12]. Therefore, values could be extracted from a user’s online words using behavior.
Boyd et al. (2015) investigated the link between user word use behavioral measurements and Schwartz
value scores, and proposed the theme words associated with each SVS value dimension [16]. Regarding
personality, Yarkoni (2010) reported individual words used in online blogs (N = 694) that positively
and negatively correlated with personality [17]. These works constitute a proof-of-concept study that
demonstrates the utility of relying on natural language markers of abstract psychological phenomena,
including values and personality, and present significant opportunities to better predict and understand
their connection to consumers’ behaviors and thoughts in a broader sense.
2.4. Research Questions and Design
2.4.1. Q1: What Is the Predictive and Explanatory Power of Psychographic Segmentation in
E-Commerce User Preferences?
Marketers and scholars have highlighted the need to account for customer perceptions and
expectations in a specific market segment. In the offline context, relationships between psychographic
segmentation and user behaviors and preferences have been asserted in various scenarios, and results
have demonstrated that psychographic variables are effective in understanding and predicting user
behaviors and preferences [26,54]. However, in the e-commerce context, the potential predictive and
explanatory power between segmentations—particularly psychographic variables and the clustering
method—in predicting e-commerce consumer preferences has not been comprehensively studied.
2.4.2. Q2: What Are the Differences between SVS and BFF Segmentation in Predicting and Explaining
Researchers argue that values are a useful basis for segmentation or psychographics because,
compared with AIO, values are less numerous, more central and, compared with personality, more
immediately related to motivations and consumer behaviors [21]. The comparison between personality
and values mostly shares the context of brand preference. For example, Mulyanegara (2009; 2011)
compared the predictive power of personality and values on brand preferences within a fashion context
and found that values are indeed better predictors of brand preferences [55,56]. Although scholars
have studied the predictive and explanatory power of psychological segmentation on consumer
behavior, few studies have examined the relative importance of different psychographic variables,
that is, personality and values, in understanding e-commerce consumer behavior.
2.4.3. Q3: What Is the Adaptability of Psychographic Measurements in Predicting and Explaining
As two of the psychographic measures, the SVS and BFF models have been proven to improve
the understanding of consumer purchase behaviors and preferences. However, Schwartz (2012)
highlighted that not all value dimensions can effectively predict individual behaviors [27]. In an online
context, it is reasonable to partition the value items into more or less fine-tuned distinct values according
to the needs and objectives of the analysis.
102
Appl. Sci. 2019, 9, 1992
2.4.4. Overview
Psychographic segmentation has been criticized by the well-known public opinion analyst
and social scientist Daniel Yankelovich, who said that psychographics is “very weak” at predicting
people’s purchases, making it a “very poor” tool for corporate decision-makers [57]. Studies have
demonstrated that there is a significant difference between customer behaviors among different product
types [58]. Different segment groups create different purchasing patterns. However, the relationship
between segmentation and user buying behavior and preferences is strongly affected by the product
category [42,59]. In recent years, deep learning has made great progress and has demonstrated
significantly better performance compared to traditional algorithms in online big data environments.
Considering the scale of online user preferences, the use of deep neural networks (DNNs) will
potentially make better psychological segmentation-based preference regression models.
In this research, we introduce empirical evidence in psycholinguistic research and the recent
progress of DNN algorithms in identifying and integrating consumer psychographic segments from
online shopping behaviors, and investigate its performance in an e-commerce consumer preference
regression model. Moreover, accounting for the predictive and explanatory power of psychographic
segmentation in e-commerce consumer preferences, we use both RMSE and R2 to conduct evaluations
across five product categories. Figure 2 shows the overview of our research.
In the next section, we present the results related to the three main study questions in this research.
Additionally, we investigate the promising influence of the product category and preference regression
method that supports our research.
Research Questions
Psychographic Segmentation on E-commerce

Consumer Preference
Q1 Q2 Q3
Experimental Process
Five Amazon datasets

Findings
Psycholinguistic
Psychographic Recognition
Lexicon
Construction
Clustering Segmentation Generation
DNN User Preference Regression
Evaluation
RMSE R2
Figure 2. Research framework.
3. Methodologies
We now describe the proposed approach, beginning with intuition and then going into detail.
Given that Amazon.com is the most studied online consumer domain in the context of e-commerce,
103
Appl. Sci. 2019, 9, 1992
we used Amazon consumer behavior data to present our methods. In this section, we introduce our
methods, and apply these methods to identify Amazon customers’ SVS and BFF scores to predict their
online preference.
3.1. Word Use-Based Psychographic Inference

Lexicon is the basis for utilizing customer word use behaviors to automatically extract psychological
traits. The combination of semantic knowledge and online corpora not only utilizes the semantic
relationships as prior knowledge to provide accurate seed words but, also, help to foster the word
association information, such as term position and word co-concurrency in the corpus [60]. Based on
semantic knowledge and Amazon corpus, our research applies the NLP method to automatically
construct SVS and BFF consumer psychographic lexicons for consumer value and personality
identification. First, we obtained two seed words according to the SVS- and BFF-related word
use behavior proposed by prior psycholinguistic works. Second, we applied the semantic knowledge
embedded in WordNet, which is an English lexical database, to extend the seed words. Third, we further
extended these seed words on the large-scale Amazon consumer review corpus, to construct the
psychographic lexicons. Finally, we calculated the multidimensional psychographic scores for
consumers and products based on the lexicons and consumer online reviews.
3.1.1. Construction of Psychographic Seed Words by Psycholinguistics

The empirical evidence between psychographics and consumer word use behavior is the solid
basis for seed word set construction. Boyd et al. (2015) provided three value-related word use
behavior sources: self-report “Value Essay”, Self-Report Behavior Essay, and Facebook updates.
They then presented the theme words as positive or negative using the Schwartz values [16]. Basically,
we combine these three sources to build two SVS lexicon seed word sets. However, there are some
contradictions, such as that Religion-related words in the Value Essay are negatively related with
Self-Direction (R2 ≥ 0.01), but in the Self-Report Behavior Essay, these words are positively related with
user Self-Direction (R2 ≥ 0.01)]. Considering that behavioral measures tend to outperform self-report
as addressed by Boyd et al. 2015, we give priority to the user behavior measurement or the use of
words in Facebook updates. Similarly, the work of Yarkoni (2010) provides the basis to construct BFF
seed word sets [17]. Tables 1 and 2 summarize themes related to SVS and seed words related to BFF;
see Boyd et al. 2015 for more details.
Table 1. Schwartz Value Survey (SVS)-related themes.
SVS Direction SVS Dimensions Positively Related Themes Negatively Related Themes
Self-Enhancement Achievement Swearing Empathy, Lord,
Faith (Positive), Faith
(Negative), Empathy,
Self-Transcendence Benevolence
Honesty, Family Care,
School, Meetings
(Negative), Family Growth,
Faith (Negative),
Conservation Conformity Family Care, Religiosity,
Faith (Positive)
Celebration,
Group Success, Proselytizing
Family Care, Meetings,
Self-Enhancement Hedonism Relaxation, Swearing Achievement,
Religiosity, Religiosity
Self-Enhancement Power Student Faith (Negative)
104
Appl. Sci. 2019, 9, 1992
Table 1. Cont.
SVS Direction SVS Dimensions Positively Related Themes Negatively Related Themes
Strive, Action, Nature,
Family Growth, Personal, Free, Band, Guitar,
Conservation Security
Proselytizing Rehearsal, Perform, Money,
Enjoy, Spend, Free, Change
Faith (Negative), Faith
Faith (Negative), Faith
(Positive), Social, Knowledge
(Positive), Social, Growth,
Openness to change Self-Direction Gain, Family Care, Social,
Indulgence,
Growth, Indulgence,
Caring/Knowledge
Caring/Knowledge
Indulgence, Knowledge
Openness to change Stimulation (Negative), Family Care,
Gain, Exhaustion
Religiosity, Celebration
Religiosity, Celebration,
Proselytizing, Faith
Conservation Tradition (Positive), Faith (Negative), Enjoyment
Family Growth, Social,
Family Care
Empathy, Faith (Positive),
Daily Routine, Family Care,
Faith (Negative), Empathy,
Student, Faith (Positive),
Self-Transcendence Universalism Family Growth, Social,
Faith (Negative), Social,
Growth, Caring/
Knowledge Gain
Knowledge, Openness
Table 2. Big Five Factor (BFF)-related seed words.
Big Five Factors Positively Related Words Negatively Related Words

Awful, though, lazy, worse, depressing, irony,
Road, Southern, visited, ground,
Neuroticism terrible, stressful, horrible, sort, annoying,
oldest, invited, completed
ashamed, ban
Bar, drinks, restaurant, dancing, restaurants,
Extraversion grandfather, Miami, countless, drinking, shots, Other, cats, computer, minor
girls, glorious, pool, crowd, sang, grilled
Folk, humans, of, poet, art, by, universe, poetry,
Openness narrative, culture, century, sexual, films, novel, giveaway
decades, ink, passage, literature, blues
Wonderful, together, visiting, morning, spring,
Agreeableness walked, beautiful, staying, felt, share, gray, joy, Porn, cock, fuck
afternoon, day, moments, hug, glad
Completed, adventure, adventures, enjoying, Stupid, boring, desperate, saying,
Conscientiousness
Hawaii, deck utter, it’s, extreme
WordNet is a large English lexical database that groups nouns, verbs, adjectives, and adverbs
into sets of cognitive synonyms, each expressing a distinct concept. Synonym sets are interlinked by
means of conceptual semantic and lexical relations. WordNet’s structure makes it a useful tool for NLP.
Based on WordNet and the psychographic seed words, we extracted the related words by following
the Synonym Rule: for each positive seed word, its synonym is considered to have the same SVS or
BFF dimension. Therefore, we obtain the consumer psychographic (SVS and BFF) candidate thesaurus
composed of seed words and their synonyms from WordNet.
105
Appl. Sci. 2019, 9, 1992
3.1.2. Psychographic Candidate Thesaurus Extension by Amazon Corpus

Consumer word use behaviors in UGC are very diverse. For one thing, the WordNet-based word
extension ignores contexts where words often have different meanings [49,61]. The word “depressed”,
for example, can variously refer to sadness, an economic condition, or even the physical state of
an object. Thus, in specific online shopping context, it is difficult to identify a potential word use
characteristic using the candidate thesaurus alone.
In recent years, many scholars have adopted an NLP method called word embedding to discover
new context-wise words in which a specific meaning is embedded. Based on the large-scale internet
corpus, these methods have achieved remarkable progress [62]. Word embeddings map words to
high-dimensional numerical vectors that contain the semantic information in these words. The basic
idea of our method is to map each word into a K-dimensional vector space through word embeddings
and to calculate the semantic similarity between words in corpus and words in a thesaurus using the
cosine distance between vectors. Based on the psychographic candidate thesaurus, we use Word2Vec
to construct word vector models which can identify new words in large-scale internet corpora by
word similarity.
We apply cosine distance to measure word semantic similarity between w w word vectors in
w
the psychographic candidate thesaurus and the corpus. Let word1 = v1 1 , v2 1 , . . . vm1 , word2 =
w w
v1 2 , v2 2 , . . . vw
m , where m is the word vector dimension; we use m = 200 in our experiments. We can
2
then write cosine distance as follows:

m w
vk 1 × vw 2
k
sim(word1 , word2 ) = cosθ = , (1)
m w1 2 m w2 2
k =1
k = 1 vk × k = 1 vk
where the numerator represents the dot product of the two vectors and the denominator represents the
modular product of the word vectors.
Based on the cosine distance, we utilize the Top10 algorithm in the Gensim [63] software
implementation of Word2vec to calculate the 10 most similar words to the seed words in the thesaurus.
We set 0.45 as the threshold for the similarity value according to our empirical results, and traverse
the word embedding trained on Amazon corpus [64] to obtain the top 10 most similar words for the
psychographic (SVS and BFF) candidate thesaurus. We then extend the candidate thesaurus by adding
the top 10 most similar words. We obtain the final candidate thesaurus by carrying out the above
procedure repeatedly, until there are no new words to be extracted.
Considering that consumer online word use is sometime ambiguous and that the adjacent
values of psychographic are continuous, we calculate multidimensional values corresponding to the
psychographics for every candidate word. According to the online psychographic score equation
below, we calculate the semantic distances between each pair of words in the extended psychographic
seed word sets to obtain the score of the word in all SVS or BFF dimensions. The online psychographic
scores equation is

wordpsy_scores wext , wpsydim

(2)
= Max sim(wext , wsed1 ), sim(wext , wsed2 ), . . . , sim wext , wsedp ,
where wext is a word from the final online psychographic candidate thesaurus, wpsy_dim is a specific
psychographic dimension, wsed1 and wsed2 , . . . , wsedp are seed words attached to wpsy_dim in the candidate
thesaurus constructed in Section 3.1.1. In addition, three experts are hired to manually verify the
SVS and BFF lexicon and to remove words marked as irrelevant by at least two experts. We then get
positive or negative SVS and BFF lexicons with similar meanings to the seed words included.
106
Appl. Sci. 2019, 9, 1992
3.1.3. Lexicon-Based Psychographic Inference

We define the p-dimensional
positive/negative psychographic scores attached to an consumer/
product as L = L1 , L2 , . . . , Lp , user/product reviews set as {r1 , r2 , . . . , rm }, and the total number of
a user’s reviews as m and one of the review ri as {wi1 , wi2 , . . . , win }, where n is the total number of words
in the review. According to the psychographic lexicon, we calculate consumer psychographic scores

using Equation (2), where p-dimensional psychographic scores are defined as Lu ∈ Lp u , where wij is
the psychographic score attached to a word in the consumer reviews.

m
n
Lu ∈ Lup = wij (3)
i=1 j=1
The values are normalized as

Lu
Lu10 = . (4)
Max Lu10
Overall, we calculate the positive and negative psychographic (BFF and SVS) scores for each
consumer and product in the Amazon review dataset.
3.2. User Preference Prediction Based on Psychographic Segmentation and Neural Network
Based on the psychographic inference method, we obtained the psychographic scores for each
consumer and product in the Amazon dataset. In this section, we further introduce density-based
spatial clustering of applications with noise (DBSCAN) and a few regression methods, including DNN,
to predict user preference.
3.2.1. Psychographic Segmentation Based on DBSCAN

Clustering-based segmentation is often used to categorize consumers on the basis of the relative
importance they place on various product attributes, benefits, personality, and value [45,65]. DBSCAN
cluster analysis does not require one to specify the number of clusters in the data a priori, as opposed
to k-means. DBSCAN can find arbitrarily shaped clusters. It can even find a cluster completely
surrounded by, but not connected to, a different cluster. We conduct DBSCAN on the consumers’
psychographic scores, namely, the SVS and BFF scores (see Section 3.1). In the subsequent e-commerce
consumer preference-predicting process, we predict the preferences based on the positive and negative
consumer/product psychographic scores and the segmented groups they belong to.
3.2.2. User Preference Prediction Based on Psychographic Segmentation and Deep Neural Network
Let U and I be the number of consumers and products, R the training interaction matrix, and R̂
the predicted interaction matrix. Let Y be a set of yu,i , Ŷ be a set of ŷui , yui be the preference of user u to
product i, and ŷui denote the corresponding predicted score. We use a partially observed vector (rows of
the features set X and (u)
X̂) as a combination of consumer representation
x and product representation
psy psy psy psy psy psy psy psy psy psy psy psy
x(i) , where x(u) = uP , uP . . . , uP , uN , uN , . . . , uN , x(i) = iP , iP , · · · , iP , iN , iN , . . . , iN .
1 2 D 1 2 D 1 2 D 1 2 D
psy psy
uP and uNrepresent positive and negative psychographic scores, respectively, “psy” represents
D D
psychographic methods (SVS or BFF), and PD and ND represent the number of dimensions of
psychographic lexicons (positive or negative), respectively. We define R as {X, Y} and R̂ as {X̂, Ŷ}. Then,
we conduct three main segment-wise regression algorithms to build the preference-predicting models.
Random Forest (RF): RF has been shown to be effective in a wide range of segment-wise user

preference predicting problems [66]. Random forest is an ensemble of binary trees T1 (X), · · · , TB (X) ,
where X = x(u) ∩ x(i) = {x1 , · · · , xP } is a p-dimensional vector of molecular descriptors or properties
associated with a molecule. Each of these trees is stochastically trained

on random subsets of the
data. The ensemble produces B outputs Ŷ1 = T1 (X), · · · , ŶB = TB (X) , where Ŷb , b = 1, · · · , B is the
107
Appl. Sci. 2019, 9, 1992
prediction for a molecule by the bth tree. Outputs of all trees are aggregated to produce one final
prediction. Ŷ in regression is the average of the individual tree predictions.
Support Vector Machine (SVM): The ability of the support vector machine has been recognized
in automated user/product modeling to predict retailer outcomes [67,68]. An SVM discriminates
2
between data by constructing a hyperplane wT ϕ(X) + b = 0 by minimizing w 2 +C i , subject to

yi w ϕ(xi ) + b ≥ 1 − i , i ≥ 0 ∀i, where ϕ(xi ) is either xi or a higher dimensional representation of
T
xi , C is the cost parameter, w is the weight vector, i is margin of tolerance where no penalty is given
to errors [69].
Deep Neural Network (DNN): Thanks to deep learning’s (DL) capability in solving many complex
tasks while providing state-of-the-art results, academia and industry alike have been in a race to apply
deep learning to a wider range of applications [70]. Although neural networks are widely used in
market segmentation [71,72], few researchers have introduced deep neural networks (DNNs) into
market segmentation to predict user preferences. DNN enables effective capture of non-linear and
non-trivial user-product relationships. A typical DNN consists of an input layer, a few hidden layers,
and an output layer with number of nodes equal to the cardinality of categories. We use the typical mean

square error (MSE), i.e., Bi=1 ( yi − ŷi )2 as a loss function of our DNN method, where ŷi = h( g(xi )), g is
some linear combination of node values in the hidden layers of the network, and h is an activation
function (typically sigmoid or hyperbolic tangent function). Training a neural network involves
minimizing the loss function defined above using gradient descent and backpropagation of the
gradient across multiple hidden layers to update the weights between nodes. Here, we initialize the
weight using Gaussian distribution with expectation 0 and variance 1, then form a neural network
with two hidden layers for consumer preference regression tasks.
3.3. Evaluation
RMSE and R2 are two of the most widely used consumer preference evaluation methods for rating
prediction and can reflect the performance of psychographic segmentations in user preference and
rating prediction. We apply both in an evaluation method for the prediction. Let ŷu,i be the predicted
ratings, yu, i the true ratings, yu the mean rating for consumer u, and n the size of test dataset. Then,
RMSE and R2 are defined as follows:

1 n
RMSE = ( yu, i − ŷu,i )2 , (5)
n i=1
n
i=1 ( ŷu,i − yu )2 − ni=1 ( yu, i − ŷu,i )2
R2 = n 2
. (6)
i=1 ( yu, i − yu )
4. Experiment
In the previous section, we have constructed the positive and negative e-commerce psychographic
lexicons, namely SVS-pos, SVS-neg and BFF-pos, BFF-neg lexicons (see S1.csv, S2.csv, S3.csv and
S4.csv respectively in supplementary materials for more details), and conducted e-commerce consumer
segmentation based on the identified SVS and BFF scores and DBSCAN. In this section, we further
proposed a DNN method to build the segment-wise consumer rating regression model. Then,
we proceeded to utilize the online shopping data in Amazon to conduct our experiments.
4.1. Dataset Description

Amazon is one of the largest e-commerce platforms in the world and has accumulated a large
amount of user buying behavior data. The Amazon review dataset, published by McAuley et al. [64],
contains product reviews and metadata from Amazon.com, including 142.8 million reviews spanning
from May 1996 to July 2014. We selected 5 review datasets from 5 product categories according
to “K-core” values of “10”, whereby each of the remaining users or items have at least 10 reviews.
108
Appl. Sci. 2019, 9, 1992
According to the work by Arnoux et al. (2017) [73], we consider that 10 reviews (whereby the
average length of review is 189 words) is capable for consumer/product psychographic inference and
comparable to 25 tweets. Table 3 shows the detailed dataset description.
Table 3. Experimental dataset description.
Total Number Total Number

Item Category K-Core
of Users of Items
Beauty 1397 548 10
Office Products 724 395 10
Baby 946 231 10
Grocery and Gourmet Food 1556 494 10
Toys and Games 1037 398 10
The sample review shows details about our data:

{
“reviewerID ‘: ’A2SUAM1J3GNN3B”,
“asin‘: ’0000013714”,
“reviewerName‘: ’J. McDonald”,
“helpful‘: [2,3],
“reviewText‘: ’I bought this for my husband who plays the piano. He is having a wonderful time
playing these old hymns. The music is at times hard to read because we think the book was published
for singing from more than playing from. Great purchase though!”,
“overall”: 5.0,
“summary‘: ’Heavenly Highway Hymns”,
“unixReviewTime”: 1252800000,
“reviewTime‘: ’09 13, 2009”
}
4.2. Experimental Procedure

The data processing process is divided into the following steps. Figure 3 shows the whole picture
of our experiment process across Sections 3 and 4.
First of all, we keep the “reviewerID”, “asin”, “overall”, “reviewText”, and “summary” tags in
the seven datasets above and combined “reviewText” and “summary” as the word use behaviors in
recognized online psychographics.
Second, we conduct text preprocessing, including normalization, tokenization, removing stop words,
and stemming on the textual contents using Amazon reviews and lexicons by Python, machine learning
tool Scikit-learn (https://scikit-learn.org/stable/modules/generated/sklearn.cluster.DBSCAN.html) and the
Natural Language Toolkit (NLTK) [74]. Normalization is a process that converts a list of words to a more
uniform sequence. Given a character sequence and a defined document unit, tokenization is the task of
chopping it up into pieces, called tokens. Some extremely common words in reviews, which would be of
little value in helping select texts matching our need, are excluded from the vocabulary entirely. These words
are called stop words. The goal of stemming is to reduce inflectional forms, and sometimes derivationally
related forms of a word, to a common base form. We perform stemming by Lancaster Stemmer in NLTK
on both words in lexicon and reviews. We removed stop words using the English stop word list in
NLTK, obtained the of words in lowercase using the text lower method in Python, and conducted Z-score
normalization using scale method in Scikit-learn. Based on the SVS-pos, SVS-neg, BFF-pos, and BFF-neg
lexicons in Section 3.1, and all the data preprocessing steps above, we calculate the psychographic scores by
matching the used words and the words in these lexicons. Thus, we get the SVS and BFF scores for each
Amazon consumer and product.
109
Appl. Sci. 2019, 9, 1992
Psychographic Segments
Bulider
Semantic Similarity
Online Lifestyles Scale
Manual
Semantic Knowledge Verification
Online Lifestyles Lexicon
üSynonyms
Internet Corpus Word Co-occurrence

üTextual Contents of Amazon Review
Online Lifestyles Dimensions

NLP Preprocessing
Online Lifestyles Vectors Combination Lexcions
üNormalization
üUsers
üTokenization
üProducts
üĂĂĂ Amazon Review Dataset
Cosine
DBSCAN
Similarity
Regression 5 Product
Online Lifestyles Scores LR Categories Recommendation Q1
SVM Model Evaluation
RF ˄RMSE˅ Q2
Predicting
Ratings DNN
Matrix
Dimension
RFE-SVM Q3
Evaluation
Online Lifestyles Segments
User Preference Prediction
Figure 3. Experiment process.
Third, for each of the product categories, we perform the DBSCAN algorithm in the consumers’
SVS or BFF scores, using Scikit-learn, to get the consumers’ positive and negative psychographic scores
and attached psychographic segment tags. We then build the rating predicting dataset which combines
psychographic scores (as independent variables) with the rating given by the consumer to the product
(as a dependent variable). We also construct feature sets for each product category which contains
random values between 0 and 1 as the control group. For each of these datasets, we optimize the
parameters for DNN by gradient descent. The optimal number of epochs (a single pass through
the full training set) used for the neural network algorithm is decided by the performance on the
validation set. For SVM, we use the validation set for optimizing the cost parameter C. We use
Scikit-learn (https://scikit-learn.org/stable/modules/generated/sklearn.linear_model.LinearRegression.html)
implementation of linear regression (LR), SVM with a radial basis function kernel, random forest, and Keras
implementation of Google TensorFlow software (https://www.tensorflow.org) for developing DNN and
baseline. A 5-fold cross-validation is applied to select training and test datasets for each fold and avoid
overfitting for linear regression, SVM, RF, and DNN. An example of RMSE evolution with epochs for DNN
is shown in Figure 4. In Figure 4, we can see that the best epoch is around 15.
110
Appl. Sci. 2019, 9, 1992
Figure 4. Deep neural network training and testing with different epochs.
Fourth, we apply feature ranking with a recursive feature elimination method to investigate
the feature importance of different psychographic subdimensions in understanding consumer online
preference. The support vector machine-based recursive feature elimination (RFE-SVM) approach
is a popular technique for feature selection and subsequent regression task, especially in predicting
consumer preference [75]. At each iteration, a linear SVM is trained, followed by removing one or
more “bad” features from further consideration. The quality of the features is determined by the
absolute value of the corresponding weights used in the SVM. The features remaining after a number
of iterations are deemed to be the most useful for discrimination and can be used to provide insights
into the given data [76]. By introducing the RFE-SVM in segment-wise consumer rating prediction,
we can dive into the subdimensions of SVS and BFF to explore whether these dimensions are effective
in predicting and explaining preferences.
Finally, we perform an online consumer preference-predicting experiment that contains 5 product
categories * 4 predictive algorithms (LR, SVM, RM, NN) * 3 psychographic variables tools (Random,
SVS, and BFF) * 2 clustering method (clustering consumers based on DBSCAN or not).
5. Result Analysis
5.1. Analysis of the Significance of Different Experimental Settings in Predicting and Explaining E-Commerce
User Preferences
We evaluate the predictive and explanatory power of user preference prediction based on LR, support
vector machine (SVM), random forest (RF), and DNN, integrating consumer psychographic segmentation
across different Amazon product categories. Table 4 shows part of the results of RMSE and R2 under different
experimental settings for the user preference prediction process. The nine columns are “ProductCatrgory”,
“ClusterId”, “SampleNum”, “Psychographic Variable”, “Clustering or Not”, “RegressionMethods”, “RMSE”,
“R2” and “ReviewLen”. The “ProductCatrgory” column contains five Amazon product categories from
Table 3; “ClusterId” is the tag for each segment; “SampleNum ” is the number of Amazon consumer ratings
and reviews; “Psychographic Variables ”; “ Psychographic Variables” is the psychographic variable that we
use to perform consumer segmentation; “Clustering or Not” is the experiment setting of applied DBSCAN
clustering that determines whether consumer segmentation is performed; “RegressionMethods” is the
regression model used in preference prediction; and “RMSE” and “R2” are the RMSE and R2, respectively,
for regression model evaluation; “ReviewLen” is the average amount of words in review for each dataset.
We perform analysis of variance (ANOVA) on the results to test whether different experimental settings
(product category, predictive algorithms, psychographic variables, and clustering or not) have an influence
on the performance (R2 and RMSE) of preference prediction.
111
Table 4. Evaluation results of user preference prediction under different experimental settings (partial part of).
Product Category ClusterId SampleNum Psychographic Variables Clustering or Not RegressionMethods RMSE R2 ReviewLen
Toys and Games Random 9577 Random Random RF 0.9544 −0.0462 105.3555
Toys and Games Random 9577 Random Random RF 0.9532 −0.0434 105.3555
Toys and Games Random 9577 Random Random LR 0.9362 −0.0065 105.3555
Appl. Sci. 2019, 9, 1992
Toys and Games Random 9577 Random Random LR 0.9351 −0.0042 105.3555
Toys and Games All 9577 BFF Not Clustered by DBSCAN SVR 0.9091 0.0509 105.3555
Toys and Games All 9577 BFF Not Clustered by DBSCAN LR 0.9084 0.0524 105.3555
Toys and Games All 9577 SVS Not Clustered by DBSCAN SVR 0.9067 0.0558 105.3555
Toys and Games All 9577 SVS Not Clustered by DBSCAN LR 0.9007 0.0684 105.3555
Toys and Games Cluster_2 2044 SVS Clustered by DBSCAN NN 1.1217 0.0311 105.3555
Toys and Games Cluster_2 2044 SVS Clustered by DBSCAN RF 0.9350 0.0746 105.3555
Digital Music Cluster_1 3920 SVS Clustered by DBSCAN SVR 1.1081 −0.0632 205.5029
Digital Music Cluster_1 3920 SVS Clustered by DBSCAN RF 1.0575 0.0317 205.5029
112
Appl. Sci. 2019, 9, 1992
Table 5 shows the ANOVA results which demonstrate that the product category, clustering
method, and regression methods have a significant effect on both R2 and RMSE (p-value < 0.000).
For the segmentation variables that we use (SVS, BFF, and random segmentation), there are significant
differences between the three groups (p-value < 0.05) for R2 , whereas the difference is not significant
for RMSE.
Table 5. ANOVA results for experimental settings in predicting consumer rating.
Index Evaluation F PR (>F) df Sum_sq

Psychographic R2 36.39 0.0000 2 0.4120
Regression Methods R2 32.55 0.0000 3 0.5526
Product Category R2 7.44 0.0000 4 0.1684
Residual R2 147 0.8320
Clustering or Not R2 35.20 0.0000 2 0.4028
Psychographic RMSE 9.93 0.0001 2 0.0516
Clustering or Not RMSE 13.17 0.0000 2 0.0659
Regression Methods RMSE 102.35 0.0000 3 0.7677
Product Category RMSE 76.31 0.0000 4 0.7632
Residual RMSE 147 0.3676
5.2. Analysis of the Effect of Psychographic Segmentation on Predicting and Explaining E-Commerce
User Preferences
To investigate our three main questions, we further compare the differences between the SVS,
BFF, and random segmentation, and their subdimensions, in predicting and explaining consumer
preferences, while excluding the influence of other experimental settings.
5.2.1. Q1: Analysis of the Effect of the Clustering-Based Segmentation in Predicting and Explaining
User Preferences
We conduct Tukey’s range test, which is a single-step multiple comparison procedure and
statistical test, on Table 4. Table 6 shows the Tukey’s range test results for the preferences based
on the psychographic variables and DBSCAN segmentation method. Surprisingly, we can see that
there is no significant difference in both RMSE and R2 , regardless of whether we use DBSCAN to
obtain psychographic segments. Additionally, the psychographic variables significantly improve
the performance of user preference explanation (R2 ), whereas the improvement of efficiency is not
significant in user preference prediction (RMSE).
Table 6. Tukey’s range test for psychographic segmentation as a whole in understanding consumer preference.
Tukey’s Test for Segmentation

Evaluations G1 G2 Lower Meandiff (G2−G1) Upper Reject (p < 0.05)
R2 Clustered by DBSCAN Not Clustered by DBSCAN −0.05 −0.0027 0.0446 False
R2 Clustered by DBSCAN Random −0.167 −0.1201 −0.0732 True
R2 Not Clustered by DBSCAN Random −0.1721 −0.1174 −0.0626 True
RMSE Clustered by DBSCAN Not Clustered by DBSCAN −0.0623 −0.011 0.0403 False
RMSE Clustered by DBSCAN Random −0.0103 0.0406 0.0914 False
RMSE Not Clustered by DBSCAN Random −0.0078 0.0516 0.1109 False
5.2.2. Q2: Analysis of the Significance and Differences between Psychographic Tools in Predicting and
Explaining User Preferences
We applied Tukey’s range test to further compare the differences between the SVS, BFF, and random
segmentation within psychographic variables. Table 7 shows the results. For all product categories that
we study, there are no significant differences between the three groups in terms of RMSE. By contrast,
there are significant differences between both the SVS and BFF with the random segmentations in R2 .
113
Appl. Sci. 2019, 9, 1992
Table 7. Tukey’s test for different psychographic variables.
Evaluations G1 G2 Lower Meandiff (G2-G1) Upper Reject (p < 0.05)

RMSE BFF Random −0.0145 0.0393 0.0932 False
RMSE BFF SVS −0.052 −0.0037 0.0445 False
RMSE Random SVS −0.0964 −0.0431 0.0102 False
R2 BFF Random −0.158 −0.1091 −0.0602 True
R2 BFF SVS −0.0196 0.0242 0.0681 False
R2 Random SVS 0.0849 0.1333 0.1817 True
We plot the mean values of RMSE and R2 for the three segmentation variables in the different
product categories; Figure 5 shows the results. For differences between the SVS and BFF in user rating
explanation, the SVS is superior to BFF in explaining consumer preferences (R2 ) across all product
categories, except for Toys and Games. For user rating prediction, the SVS and BFF tend to perform
better than random segmentation in predicting user preferences (RMSE) in Beauty, Digital Music,
and Office Products. Compared with BFF, the SVS is slightly more powerful in predicting user ratings
(RMSE) for Beauty and Digital Music, whereas there is no difference between them for Grocery and
Gourmet Food, and Office Products. However, the SVS was useless in predicting user ratings (RMSE)
for Toys and Games, whereas BFF improved the predictive power of user ratings.
(a) (b)
Figure 5. Mean values of RMSE and R2
about three segmentation variables across different product
Categories. (a) Mean values of RMSE; (b) Mean values of R2 .
5.2.3. Q3: Analysis of the Adaptability of Psychographic Measurement in Predicting and Explaining
E-Commerce User Preferences
In Section 5.2.2, we studied the predictive and explanatory power of psychographic segmentation
across different product categories. However, it remains unclear whether all subdivision dimensions
of psychographic variables help understand user preferences across different e-commerce product
categories. Table 8 shows part of the results of different subdimensions’ psychographic variables
in predicting and explaining user preferences using the RFE-SVM approach. In Table 8, feature
ranking corresponds to the ranking position of the ith psychographic subdimension, where the
selected, i.e., estimated best dimensions are assigned rank 1; “importance support” represents these
selected dimensions (with value TRUE); “feature source” represents whether a dimension is positively
or negatively related to the SVS or BFF; “product category” contains five products from Table 8;
“psychographic” contains three psychographic segmentations including the SVS, BFF, and random
segmentation; and “subdimension” represents the subdimensions of positive and negative SVS and
BFF. There are 810 rows in Table 8, and 150 rows with psychographic subdimensions that support
our rating prediction models. We group the importance ranking in the 150 rows by product category
and plot the total rankings for different subdimensions of the SVS and BFF. Only approximately
114
Appl. Sci. 2019, 9, 1992
18.5% of the subdimensions of the psychographic measurements were effective in understanding

specific e-commerce preferences. We can see that within Table 8, with the exception of Sec and Univ,
all subdimensions of the SVS and BFF are useful. However, in different product categories, there are
more subdimensions that are not effective in understanding consumer preferences, including Benev,
Hed, Trad, Openness, and Extraversion in Beauty; Benev, Conf, and Openness in Office Products;
Benev, Hed, Trad, Openness, and Extraversion in Toys and Games; Achiev, Hed, and Stim in Digital
Music; and Achiev, Benev, Conf, and Openness in Grocery and Gourmet Food.
Table 8. Importance of different subdimensions of psychographic variable in predicting and explaining

user preference.
Importance Importance Feature Product Consumer

Psychographic Subdimension
Ranking Support Source Category or Product
8 FALSE pos Digital Music BFF Openness U
1 TRUE pos Digital Music BFF Agreeableness U
1 TRUE pos Digital Music BFF Conscientiousness U
6 FALSE pos Digital Music BFF Neuroticism I
5 FALSE pos Digital Music BFF Extraversion I
3 FALSE pos Digital Music BFF Openness I
... ... ... Digital Music ... ... ...
7 FALSE pos Digital Music SVS Stim U
26 FALSE pos Digital Music SVS Conf U
25 FALSE pos Digital Music SVS Hed U
5 FALSE pos Digital Music SVS Trad U
14 FALSE pos Digital Music SVS Achiev U
17 FALSE pos Digital Music SVS Benev U
4 FALSE pos Digital Music SVS Pow U
13 FALSE pos Digital Music SVS Univ U
16 FALSE pos Toys and Games SVS S-D I
29 FALSE pos Toys and Games BFF Sec I
1 TRUE pos Toys and Games BFF Stim I
9 FALSE pos Toys and Games BFF Conf I
24 FALSE pos Toys and Games BFF Hed I
19 FALSE pos Toys and Games SVS Trad I
22 FALSE pos Toys and Games SVS Achiev I
30 FALSE pos Toys and Games SVS Benev I
We conduct Tukey’s range test based on the 150 rows with psychographic subdimensions
supporting our regression models; Table 9 shows the results. There are 37 significant comparisons that
indicate the relationship between subdimensions of the SVS and BFF. From Table 9, we can see that,
in terms of the effectiveness of psychographic subdimensions in understanding consumer preferences,
the SVS subdimensions tend to follow the descending order of Pow, Trad, Conf, SD, Benev, and Stim.
For subdimensions of BFF, the descending order is Conscientiousness, Neuroticism, and Openness.
Figures 6 and 7 show two intuitive expressions of these orders, respectively.
115
Appl. Sci. 2019, 9, 1992
Figure 6. Difference between subdimensions of SVS in understanding consumer rating.
Figure 7. Difference between subdimensions of BFF in understanding consumer rating.
Table 9. Difference between subdimensions of psychographic in understanding user preference.
Ps1–Ps2 (Importance Ranking) Diff Lwr Upr P.adj

Subdimensions of SVS
Pow–Achiev −9.5667 −14.4328 −4.7006 0.0000
Pow–Hed −10.4833 −15.3494 −5.6172 0.0000
Sec–Pow 14.4780 9.2220 19.7340 0.0000
Stim–Pow 9.5667 4.7006 14.4328 0.0000
Univ–Pow 14.5667 9.7006 19.4328 0.0000
Univ–Trad 8.1167 3.2506 12.9828 0.0000
Univ–Conf 7.9833 3.1172 12.8494 0.0000
Trad–Sec −8.0280 −13.2840 −2.7720 0.0000
Sec–Conf 7.8946 2.6386 13.1506 0.0000
Univ–S-D 7.7776 2.5216 13.0336 0.0001
Sec–S-D 7.6889 2.0700 13.3078 0.0004
116
Appl. Sci. 2019, 9, 1992
Table 9. Cont.
Ps1–Ps2 (Importance Ranking) Diff Lwr Upr P.adj

Pow–Conf −6.5833 −11.4494 −1.7172 0.0005
Trad–Pow 6.4500 1.5839 11.3161 0.0007
S-D–Pow 6.7891 1.5331 12.0451 0.0012
Pow–Benev −7.4661 −13.4258 −1.5064 0.0021
Univ–Benev 7.1006 1.1408 13.0603 0.0049
Sec–Benev 7.0118 0.7297 13.2940 0.0131
Univ–Stim 5.0000 0.1339 9.8661 0.0371
Univ–Achiev 5.0000 0.1339 9.8661 0.0371
Subdimensions of BFF
Openness–Conscientiousness 6.7000 1.8339 11.5661 0.0003
Openness–Neuroticism 6.0498 0.7938 11.3058 0.0084
5.3. Analysis of Product Category and Regression Methods in Predicting and Explaining E-Commerce
User Preferences
In this section, we analyze the internal differences in the new segment-wise preference regression
methods and product categories that support the main research questions. Based on Table 4, we perform
the Tukey’s range test on the preference prediction method and product categories; the details are
shown in Table 10.
Regarding RMSE, there were more significant differences in RMSE between different product
categories compared to those in R2 , which is shown in Figure 8. The rating prediction performance
of psychographics tends to follow a descending order from Office Products, which obtains the
best predictive power, to Toys and Games, Beauty, Grocery and Gourmet Food, and Digital Music.
Regarding R2 , as shown in Figure 9, the psychographic variables in Beauty, and Toys and Games
obtain a significantly greater R2 than that of Office Products (p-value < 0.05). In terms of the regression
method, as shown in Figures 10 and 11, DNN is significantly better than the other three methods in
terms of both RMSE and R2 . The performance of methods tends to follow a descending order from
DNN to RF, LR, and SVR in the performance of preference prediction.
Figure 8. Multiple comparisons between all pairs (Tukey) in product categories for psychographic
variable segmentation-based rating prediction (RMSE).
117
Appl. Sci. 2019, 9, 1992
Figure 9. Multiple comparisons between all pairs (Tukey) in product categories for psychographic
segmentation-based rating explaining (R2 ).
Figure 10. Multiple comparisons between all pairs (Tukey) in regression methods for psychographic
segmentation-based rating prediction (RMSE).
Figure 11. Multiple comparisons between all pairs (Tukey) in regression methods for psychographic
segmentation-based rating explanation (R2 ).
118
Table 10. Tukey’s range test for product category and regression methods.
Tukey’s Test for Product Category

Evaluations g1 g2 Lower Meandiff (g2−g1) Upper Reject (p Value < 0.05)
RMSE Beauty Digital Music 0.057 0.119 0.181 True
RMSE Beauty Grocery and Gourmet Food 0.017 0.079 0.142 True
Appl. Sci. 2019, 9, 1992
RMSE Beauty Office Products −0.130 −0.067 −0.004 True

RMSE Digital Music Office Products −0.248 −0.186 −0.124 True
RMSE Digital Music Toys and Games −0.202 −0.141 −0.079 True
RMSE Grocery and Gourmet Food Office Products −0.208 −0.146 −0.084 True
RMSE Grocery and Gourmet Food Toys and Games −0.162 −0.101 −0.040 True
R2 Beauty Office Products −0.172 −0.095 −0.017 True
R2 Office Products Toys and Games 0.004 0.081 0.157 True
Tukey’s Test for Regression Methods
RMSE LR NN −0.199 −0.146 −0.094 True
RMSE NN RF 0.088 0.140 0.192 True
RMSE NN SVR 0.130 0.182 0.235 True
R2 LR NN 0.036 0.094 0.151 True
119
R2 LR SVR −0.135 −0.078 −0.022 True
R2 NN RF −0.140 −0.082 −0.025 True
R2 NN SVR −0.229 −0.172 −0.115 True
R2 RF SVR −0.146 −0.090 −0.034 True
Appl. Sci. 2019, 9, 1992
6. Discussion
In this study, we have focused on the promising role that different psychographic segmentations
play in the understanding of e-commerce consumer preferences. Based on real-world user behavior data
from Amazon, we have introduced psychographic-related behavioral evidence from psycholinguistics
and applied NLP, clustering, and DNN methods to identify users’ psychographic segments and to
further predict their preferences.
We summarize our results as follows. First, we have found that dividing e-consumers into
heterogeneous groups using a clustering method did not significantly improve the predictive and
explanatory power of e-commerce consumer preferences. By contrast, psychographic variables
significantly improved the explanatory power of e-consumer preferences, whereas the improvement in
predictive power was not significant. This finding is consistent with past studies [11,42] which showed
that individual segments based on their psychographic measures do not seem to provide a great deal
of predictive power in the context of buying behavior.
Second, we have found that both value and personality segmentations significantly improve user
preference explanation under different e-commerce scenarios, whereas no significant improvement was
shown in user preference prediction. These findings have verified previous research that psychographic
variables do not seem to provide substantial predictive power in the context of offline buying behavior [11].
However, these findings somehow contradict works suggesting that customer segmentation based on these
variables may be easy to understand, but may not provide the best possible explanatory power [42,58].
These findings show that psychographic variables may play a more important role in understanding
shopping behaviors in online, rather than offline, shopping scenarios. Additionally, although there is
no significant difference between the SVS and BFF across all product categories, both the SVS and BFF
tend to predict e-consumer preferences better across most of the product categories that we have studied,
and the SVS seems to outperform BFF in all the product categories that we have studied, except Toys and
Games. Values that characterize human motivation may be a better psychographic variable compared
with personality, which emphasizes individual differences.
Third, we have found that only 18.5% of the subdimension combinations within a psychographic
tool have stable effectiveness in understanding e-commerce preferences related to a specific product
category. However, all BFF subdimensions are important in understanding e-consumer preferences,
whereas the security and benevolence of the SVS do not demonstrate their effectiveness. We have found
that using more subdimensions does not significantly improve the predictive and explanatory power
within different product categories. Although the SVS is capable of explaining e-consumer preferences,
our work indicates that there may be some subdimensions of the SVS that are able to perform better in
understanding e-consumer preferences. This finding is consistent with Schwartz’s argument that it is
reasonable to partition the value items into more or less fine-tuned distinct values according to the
needs and objectives of the analysis [27].
Finally, the DNN method that we have proposed obtained the best predictive and explanatory
power in understanding e-consumer preferences; it is significantly better than RF and SVM that were
applied in previous research. Regarding product categories, there are more significant differences for
psychographic variables in predicting than explaining e-consumer preferences. Moreover, the role
of psychographic variables in predicting and explaining e-consumer preferences typically does not
demonstrate consistency in different product categories: psychographic variables demonstrates the
best predictive power in Office Products and the least explanatory power in Office Products. In most
cases, prediction models are likely to possess some level of both explanatory and predictive power [77].
We can, therefore, visualize performances of different models in terms of a trade-off between explanatory
and predictive power on a two-dimensional plot, where the two axes indicate prediction accuracy and
explanatory power, respectively. We leave such visualization for future work. As the influence of
psychographic variables on user preferences is moderated by product category, our findings indicate
a more complicated relationship between them.
120
Appl. Sci. 2019, 9, 1992
This study has extended the depth and breadth of psychographic-related studies through user preference
prediction in real-world e-commerce scenarios. Our findings regarding psychographic segmentation and
segment-wise preference prediction have provided theoretical guidance for psychographic variable adoption,
selection, and use in electronic marketing researches like online advertising, retail, and recommendation.
Practically, our findings regarding subdimensions of psychographic variables have provided a practical
reference for psychographic measurement development in each e-commerce product category that we have
studied. Additionally, the influence of product category on psychographic-based preference prediction and
explanation indicates promising e-commerce product research directions and applications. By introducing
psychographic-related word use behavioral evidence, followed by natural language processing and DNN
techniques, we have attempted to overcome the difficulties of observing e-consumer psychographics on
a large scale, and have provided a promising psychographic-based consumer preference prediction method
for subsequent research and applications.
However, our study has some limitations. First, we have only applied the review dataset with “K-core”
values of 10, whereas there are a huge number of consumers who have either a limited number of reviews
or words in their reviews, potentially causing bias in the psychographic inference. Second, the original
dataset does not provide demographic information and we have not evaluated the difference between
the psychographic scores and the scores assessed by self-report, which may have caused biases in our
research results. Third, in addition to psychographic-related single words, there may be other linguistic clues
embedded in phrases, sentences, and paragraphs that we have not taken into consideration. Fourth, although
our research has demonstrated the significant explanatory power of psychographic tools in understanding
e-consumer preferences, no significant predictive power or difference was found. Psychographic variables
such as AIO, VALS, and LOV, or their combinations, should be taken into consideration in fully assessing
the influence of psychographics on the understanding of e-consumer preferences.
In future studies, we will improve the consumer diversity of Amazon data to verify our findings.
Promising future research directions include: evaluating the current psychographic inference method;
developing a new psychographic variable identification method by introducing more advanced DNN
models and textual features like SVS-related themes [16] and word embeddings [78]; verifying and
comparing word use behaviors (K-core); combining different psychographic variables, together with
other well-known segmentation approaches (e.g., demographic segmentation, behavioral segmentation,
geographic segmentation, etc.) to further understand their influence in predicting and understanding
e-consumer preferences.
Supplementary Materials: The following are available online at http://www.mdpi.com/2076-3417/9/10/1992/s1,

Table S1–S4.
Author Contributions: The study was carried out in collaboration between all authors. H.L., Y.H., X.H. and W.W.
designed the research topic. Y.H. and H.L. conducted the experiment and wrote the paper. Z.W. and K.L. examined
the experimental data and checked the experimental results. All authors agreed to submission of the manuscript.
Funding: This research is supported by Program of National Natural Science Foundation of China (No. 71571084,
No. 71271099) and China Scholarship Council.
References
1. Jacobs, B.J.D.; Donkers, B.; Fok, D. Model-based purchase predictions for large assortments. Marketing Sci.
2016, 35, 389–404. [CrossRef]
2. Lu, S.; Xiao, L.; Ding, M. A video-based automated recommender (VAR) system for garments. Marketing Sci.
2016, 35, 484–510. [CrossRef]
3. Xie, K.L.; Zhang, Z.; Zhang, Z. The business value of online consumer reviews and management response to
hotel performance. Int. J. Hospit. Manag. 2014, 43, 1–12. [CrossRef]
4. Chen, J.; Haber, E.; Kang, R.; Hsies, G.; Mahmud, J. Making use of derived personality: The case of social
media ad targeting. In Proceedings of the Ninth International AAAI Conference on Web and Social Media,
Oxford, UK, 26–29 May 2015.
121
Appl. Sci. 2019, 9, 1992
5. Trusov, M.; Ma, L.; Jamal, Z. Crumbs of the cookie: User profiling in customer-base analysis and behavioral
targeting. Marketing Sci. 2016, 35, 405–426. [CrossRef]
6. Culotta, A.; Cutler, J. Mining brand perceptions from twitter social networks. Marketing Sci. 2016, 35, 343–362.
[CrossRef]
7. Jin, J.; Ji, P.; Gu, R. Identifying comparative customer requirements from product online reviews for competitor
analysis. Eng. Appl. Artif. Intell. 2016, 49, 61–73. [CrossRef]
8. Matz, S.C.; Netzer, O. Using big data as a window into consumers’ psychology. Curr. Opin. Behave. Sci. 2017,
18, 7–12. [CrossRef]
9. Jih, W.J.K.; Lee, S.F. An exploratory analysis of relationships between cellular phone uses’ shopping motivators
and lifestyle indicators. J. Comp. Inf. Syst. 2004, 44, 65–73.
10. Ko, E.; Kim, E.; Taylor, C.R.; Kim, K.H.; Kang, J.I. Cross-national market segmentation in the fashion industry:
A study of European, Korean, and US consumers. Int. Marketing Rev. 2007, 24, 629–651. [CrossRef]
11. Sandy, C.J.; Gosling, S.D.; Durant, J. Predicting consumer behavior and media preferences: The comparative
validity of personality traits and demographic variables. Psychol. Market. 2013, 30, 937–949. [CrossRef]
12. Pennebaker, J.W.; Chung, C.K.; Frazee, J.; Lavergne, G.M.; Beaver, D.I. When small words foretell academic
success: The case of college admissions essays. PLoS ONE 2014, 9, e115844. [CrossRef] [PubMed]
13. Buettner, R. Predicting user behavior in electronic markets based on personality-mining in large online social
networks. Electr. Marketing 2017, 27, 247–265. [CrossRef]
14. Lee, A.J.T.; Yang, F.-C.; Chen, C.-H.; Wang, C.-S.; Sun, C.-Y. Mining perceptual maps from consumer reviews.
Decis. Support Syst. 2016, 82, 12–25. [CrossRef]
15. Wang, Y.; Lu, X.; Tan, Y. Impact of product attributes on customer satisfaction: An analysis of online reviews
for washing machines. Electr. Commerce Res. Appl. 2018, 29, 1–11. [CrossRef]
16. Boyd, R.L.; Wilson, S.R.; Pennebaker, J.W.; Kosinski, M.; Stillwell, J.D.; Michaela, R. Values in words: Using
language to evaluate and understand personal values. In Proceedings of the ICWSM, Oxford, UK, 26–29
May 2015; pp. 31–40.
17. Yarkoni, T. Personality in 100,000 words: A large-scale analysis of personality and word use among bloggers.
J. Res. Personal. 2010, 44, 363–373. [CrossRef]
18. Smith, W.R. Product differentiation and market segmentation as alternative marketing strategies. J. Marketing
1956, 21, 3–8. [CrossRef]
19. Karlsson, L.; Dolnicar, S. Someone’s been sleeping in my bed. Ann. Tour. Res. 2016, 58, 159–162. [CrossRef]
20. Shukla, P.; Babin, B.J. Effects of consumer psychographics and store characteristics in influencing shopping
value and store switching. J. Consum. Behav. 2013, 12, 194–203. [CrossRef]
21. Wedel, M.; Kamakura, W.A. Market Segmentation: Conceptual and Methodological Foundation; Springer Science
& Business Media: Berlin, Germany, 2012.
22. Vyncke, P. Lifestyle segmentation: From attitudes, interests and opinions, to values, aesthetic styles, life
visions and media preferences. Eur. J. Commun. 2002, 17, 445–463. [CrossRef]
23. Gunter, B.; Furnham, A. Consumer Profiles (Rle Consumer Behaviour): An Introduction to Psychographics;
Routledge: London, UK, 2014.
24. Walters, G.D. Lifestyle Theory: Past, Present, and Future; Nova Science Publishers: Commack, NY, USA, 2006.
25. Mitchell, V.W. How to identify psychographic segments: Part 1. Marketing Intell. Plann. 1994, 12, 4–10. [CrossRef]
26. Bruwer, J.; Li, E. Wine-related lifestyle (WRL) market segmentation: Demographic and behavioural factors.
J. Wine Res. 2007, 18, 19–34. [CrossRef]
27. Schwartz, S.H. An overview of the Schwartz theory of basic values. Online Read. Psychol. Cult. 2012, 2, 11. [CrossRef]
28. Lin, C.F. Segmenting customer brand preference: Demographic or psychographic. J. Product. Brand Manag.
2002, 11, 249–268. [CrossRef]
29. Rokeach, M. The Nature of Human Values; Free Press: Detroit, MI, USA, 1973.
30. Kahle, L.R.; Beatty, S.E.; Homer, P. Alternative measurement approaches to consumer values: The list of
values (LOV) and values and life style (VALS). J. Consum. Res. 1986, 13, 405–409. [CrossRef]
31. Sagiv, L.; Schwartz, S.H. Cultural values in organisations: Insights for Europe. Eur. J. Int. Manag. 2007, 1,
32. Yang, J.; Liu, C.; Teng, M.; Liao, M.; Xiong, H. Buyer targeting optimization: A unified customer segmentation
perspective, Big Data. In Proceedings of the 2016 IEEE International Conference on IEEE, Washington, DC, USA,
5–8 December 2016; pp. 1262–1271.
122
Appl. Sci. 2019, 9, 1992
33. Hirsh, J.B.; Kang, S.K.; Bodenhausen, G.V. Personalized persuasion: Tailoring persuasive appeals to recipients’
personality traits. Psychol. Sci. 2012, 23, 578–581. [CrossRef]
34. Fernández-Tobías, I.; Braunhofer, M.; Elahi, M.; Ricci, F.; Cantador, I. Alleviating the new user problem in collaborative
filtering by exploiting personality information. User Model. User Adapt. Interact. 2016, 26, 221–255. [CrossRef]
35. Karumur, R.P.; Nguyen, T.T.; Konstan, J.A. Exploring the value of personality in predicting rating behaviors:
A study of category preferences on movielens. In Proceedings of the 10th ACM Conference on Recommender
Systems, Boston, MA, USA, 15–19 September 2016; pp. 139–142.
36. Lee, H.J.; Lim, H.; Jolly, L.D.; Lee, J. Consumer lifestyles and adoption of high-technology products: A case
of South Korea. J. Int. Consum. Marketing 2009, 21, 153–167. [CrossRef]
37. Pan, Y.; Luo, L.; Liu, D.; Xu, X.; Shen, W.; Gao, J. How to recommend by online lifestyle tagging (olt).
Int. J. Inf. Technol. Decis. Making 2014, 13, 1183–1209. [CrossRef]
38. Piazza, A.; Zagel, C.; Haeske, J.; Bodendorf, F. Do you like according to your lifestyle? a quantitative analysis
of the relation between individual facebook likes and the users’ lifestyle. In Proceedings of the International
Conference on Applied Human Factors and Ergonomics; Springer: Cham, Switzerland, 2017; pp. 128–139.
39. Goldsmith, R.E.; Freiden, J.B.; Kilsheimer, J.C. Social values and female fashion leadership: A cross-cultural
study. Psychol. Marketing 1993, 10, 399–412. [CrossRef]
40. Kim, H.S. Consumer profiles of apparel product involvement and values. J. Fashion Marketing Manag. Int. J.
2005, 9, 207–220. [CrossRef]
41. Heine, K.; Trommsdorff, V. Practicable value-cascade positioning of luxury fashion brands. In Proceedings
of the 9th International Marketing Trends Conference, Venice, Italy, 20–23 January 2010; pp. 21–23.
42. Teck, W.J.; Cyril de Run, E. Consumers’ personal values and sales promotion preferences effect on behavioural
intention and purchase satisfaction for consumer product. Asia Pac. J. Marketing Logist. 2013, 25, 70–101. [CrossRef]
43. Fraj, E.; Martinez, E. Environmental values and lifestyles as determining factors of ecological consumer
behaviour: An empirical analysis. J. Consum. Marketing 2006, 23, 133–144. [CrossRef]
44. Padgett, D.; Mulvey, M.S. Differentiation via technology: Strategic positioning of services following the
introduction of disruptive technology. J. Retail. 2007, 83, 375–391. [CrossRef]
45. Wiedmann, K.P.; Hennigs, N.; Siebels, A. Value-based segmentation of luxury consumption behavior.
Psychol. Marketing 2009, 26, 625–651. [CrossRef]
46. Antipov, E.; Pokryshevskaya, E. Applying CHAID for logistic regression diagnostics and classification
accuracy improvement. J. Target. Meas. Anal. Marketing 2010, 18, 109–117. [CrossRef]
47. Reutterer, T.; Mild, A.; Natter, M.; Taudes, A. A dynamic segmentation approach for targeting and customizing
direct marketing campaigns. J. Interact. Marketing 2006, 20, 43–57. [CrossRef]
48. Ge, Y.; Xiong, H.; Zhou, W.; Li, S.; Sahoo, R. Multifocal learning for customer problem analysis.
ACM Trans. Intell. Syst. Technol. 2011, 2, 24. [CrossRef]
49. Schwartz, H.A.; Eichstaedt, J.C.; Kern, M.L.; Dziruizynski, L.; Ramones, S.M.; Agrawal, M.; Shah, A.;
Kosinski, M.; Stillwell, D.; Seligman, M.E.P.; et al. Personality, gender, and age in the language of social
media: The open-vocabulary approach. PLoS ONE 2013, 8, e73791. [CrossRef] [PubMed]
50. Gosling, S.D.; Mason, W. Internet research in psychology. Ann. Rev. Psychol. 2015, 66, 877–902. [CrossRef]
51. Adamopoulos, P.; Ghose, A.; Todri, V. The impact of user personality traits on word of mouth: Text-mining
social media platforms. Inf. Syst. Res. 2018, 29, 612–640. [CrossRef]
52. Bleidorn, W.; Hopwood, C.J. Using machine learning to advance personality assessment and Theory.
Personal. Soc. Psychol. Rev. 2018, 23, 190–203. [CrossRef] [PubMed]
53. Park, G.; Schwartz, H.A.; Eichstaedt, J.C.; Kern, C.J.; Kosinski, M.L.; Stillwell, M.; Ungar, J.D.; Seligman, H.L.;
Martin, E.P. Automatic personality assessment through social media language. J. Personal. Soc. Psychol. 2015,
108, 934. [CrossRef] [PubMed]
54. Grunert, K.G.; Perrea, T.; Zhou, Y.; Huang, G.; Sorensen, B.T.; Krystallis, A. Is food-related lifestyle (FRL) able
to reveal food consumption patterns in non-Western cultural environments? Its adaptation and application
in urban China. Appetite 2011, 56, 357–367. [CrossRef]
55. Casidy, M.R.; Tsarenko, Y. Predicting brand preferences: An examination of the predictive power of consumer
personality and values in the Australian fashion market. J. Fashion Marketing Manag. Int. J. 2009, 13, 358–371.
[CrossRef]
56. Mulyanegara, R.C. The relationship between market orientation, brand orientation and perceived benefits in
the non-profit sector: A customer-perceived paradigm. J. Strateg. Marketing 2011, 19, 429–441. [CrossRef]
123
Appl. Sci. 2019, 9, 1992
57. Yankelovich, D.; Meer, D. Rediscovering market segmentation. Harvard Bus. Rev. 2006, 84, 122.
58. Sinha, P.K.; Uniyal, D.P. Using observational research for behavioural segmentation of shoppers.
J. Retail. Consum. Serv. 2005, 12, 35–48. [CrossRef]
59. Oly Ndubisi, N.; Tung, M.C. Awareness and usage of promotional tools by Malaysian consumers: The case
of low involvement products. Manag. Res. News 2006, 29, 28–40. [CrossRef]
60. Wang, W.; Li, Y.; Huang, Y.; Liu, H.; Zhang, T. A method for identifying the mood states of social network
users based on cyber psychometrics. Future Internet 2017, 9, 22. [CrossRef]
61. Pero, S.; Huettner, A. Affect analysis of text using fuzzy semantic typing. In IEEE Transactionsons on Fuzzy
Systems; IEEE: New York, NY, USA, 2001; pp. 483–496.
62. Turian, J.; Ratinov, L.; Bengio, Y. Word representations: A simple and general method for semi-supervised
learning. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics,
Uppsala, Sweden, 11–16 July 2010; pp. 384–394.
63. “Gensim_ models.word2vec – Word2vec embeddings”. (2019, May 13). Available online: https://radimrehurek.
com/gensim/models/word2vec.html (accessed on 18 February 2019).
64. McAuley, J.; Targett, C.; Shi, Q.; van den Hengel, A. Image-based recommendations on styles and substitutes.
In Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information
Retrieval, Santiago, Chile, 9–13 August 2015; pp. 43–52.
65. Müllensiefen, D.; Hennig, C.; Howells, H. Using clustering of rankings to explain brand preferences with
personality and socio-demographic variables. J. Appl. Statistics 2018, 45, 1009–1029. [CrossRef]
66. Levin, N.; Zahavi, J. Predictive modeling using segmentation. J. Interact. Marketing 2001, 15, 2–22. [CrossRef]
67. Cui, D.; Curry, D. Prediction in marketing using the support vector machine. Marketing Sci. 2005, 24, 595–615.
[CrossRef]
68. Cortez, P.; Cerdeira, A.; Almeida, F.; Matos, T.; Reis, J. Modeling wine preferences by data mining from
physicochemical properties. Decis. Support Syst. 2009, 47, 547–553. [CrossRef]
69. Sharang, A.; Rao, C. Using machine learning for medium frequency derivative portfolio trading. arXiv 2015,
arXiv:1512.06228.
70. Esteva, A.; Robicquet, A.; Ramsundar, B.; Kuleshov, V.; DePristo, M.; Chou, K.; Cui, C.; Corrado, G.; Thurn, S.;
Dean, J. A guide to deep learning in healthcare. Nat. Med. 2019, 25, 24. [CrossRef] [PubMed]
71. Vellido, A.; Lisboa, P.J.G.; Meehan, K. Segmentation of the on-line shopping market using neural networks.
Exp. Syst. Appl. 1999, 17, 303–314. [CrossRef]
72. Boone, D.S.; Roehm, M. Retail segmentation using artificial neural networks. Int. J. Res. Marketing 2002, 19,
73. Arnoux, P.H.; Xu, A.; Boyette, N.; Mahmud, J.; Akkiraju, R.; Sinha, V. 25 Tweets to Know You: A New Model
to Predict Personality with Social Media. In Proceedings of the Eleventh International AAAI Conference on
Web and Social Media, Montréal, QC, Canada, 15–18 May 2017.
74. Perkins, J. Python 3 Text Processing with NLTK 3 Cookbook; Packt Publishing Ltd.: Birmingham, UK, 2014.
75. Guyon, I.; Weston, J.; Barnhill, S.; Vapnik, V. Gene selection for cancer classification using support vector
machines. Mach. Learn. 2002, 46, 389–422. [CrossRef]
76. Bedo, J.; Sanderson, C.; Kowalczyk, A. An efficient alternative to svm based recursive feature elimination
with applications in natural language processing and bioinformatics. In Australasian Joint Conference on
Artificial Intelligence; Springer: Berlin, Germany, 2006; pp. 170–180.
77. Shmueli, G. To explain or to predict? Statistics Sci. 2010, 25, 289–310. [CrossRef]
78. Zheng, L.; Noroozi, V.; Yu, P.S. Joint deep modeling of users and items using reviews for recommendation.
In Proceedings of the Tenth ACM International Conference on Web Search and Data Mining, ACM,
Cambridge, UK, 6–10 February 2017; pp. 425–434.
124
applied
sciences
Article
Classification of Cyber-Aggression Cases Applying
Machine Learning
Guadalupe Obdulia Gutiérrez-Esparza 1,2 , Maite Vallejo-Allende 2 and
José Hernández-Torruco 3, *
1 Cátedras CONACYT Consejo Nacional de Ciencia y Tecnología, Ciudad de México 08400, Mexico;
[email protected]
2 Instituto Nacional de Cardiología Ignacio Chávez, Ciudad de México 14080, Mexico;
[email protected]
3 División Académica de Informática y Sistemas, Universidad Juárez Autónoma de Tabasco, Cunduacán,
Tabasco 86690, Mexico
* Correspondence: [email protected]; Tel.: +52-993-2077-904

Received: 26 March 2019; Accepted: 26 April 2019; Published: 2 May 2019
Abstract: The adoption of electronic social networks as an essential way of communication has
become one of the most dangerous methods to hurt people’s feelings. The Internet and the
proliferation of this kind of virtual community have caused severe negative consequences to the
welfare of society, creating a social problem identified as cyber-aggression, or in some cases called
cyber-bullying. This paper presents research to classify situations of cyber-aggression on social
networks, specifically for Spanish-language users of Mexico. We applied Random Forest, Variable
Importance Measures (VIMs), and OneR to support the classification of offensive comments in three
particular cases of cyber-aggression: racism, violence based on sexual orientation, and violence against
women. Experimental results with OneR improve the comment classification process of the three
cyber-aggression cases, with more than 90% accuracy. The accurate classification of cyber-aggression
comments can help to take measures to diminish this phenomenon.
Keywords: cyber-aggression; sentiment analysis; random forest; racism; violence based on sexual
orientation; violence against women; social networks
1. Introduction
The growing access to combined telecommunication along with the increase of electronic social
network adoption has granted users a convenient method of sharing posts and comments on the
Internet. However, even if this is an improvement in human communication, this environment also
has provided proper conditions resulting in serious negative consequences to the welfare of society,
due to a type of user who posts offensive comments and does not care about the psychological impact
of his/her words, harming other users feelings. This phenomenon is called cyber-aggression [1].
Cyber-aggression is a frequently used keyword in the literature to describe a wide range of offensive
behaviors other than cyber-bullying [2–5].
Unfortunately, this problem has spread into a wide variety of mass media; Bauman [6] says
that some of the most-used digital media for cyber-aggression are: social networks (e.g., Facebook
and Twitter), short message services, forums, trash-polling sites, blogs, video sharing websites,
and chat rooms, among others. Their accessibility and fast adoption are a double-edged sword
because it is impossible to have moderators to keep an eye on every post made and filter its content.
Therefore, cyber-aggression [7] has become a threat to society’s welfare, generating electronic violence.
When cyber-aggression is constant, then it becomes cyber-bullying, mainly characterized by the
invasion of privacy, harassment, and use of obscene language against one user [8], in most of the cases

Appl. Sci. 2019, 9, 1828
against a minor [9]. Unlike bullying, cyber-bullying can happen 24/7, and the consequences for the
victim can be more dramatic, because this not only creates insecurity, trust issues, and depression, but
can also create suicidal thoughts [10] with fatal consequences. Both cyber-bullying and cyber-aggression
can harm a user by mocking and ridiculing them an indefinite number of times [11,12]; therefore it is
crucial to support the increase of detection of these social problems. According to Ditch The Label and
their Cyber-Bullying Survey [13] some of the social networks with the highest number of users that
have reported cases of cyber-aggression are Facebook, Twitter, Instagram, YouTube, and Ask.fm.
Currently, there is a wide variety of research aiming to mitigate cyber-bullying. Those from
psychology or social sciences aim to detect criminal behavior and provide prevention strategies [14],
and those from the field of computer science seek to develop effective techniques or tools to support the
prevention of cyber-aggression and cyber-bullying [15]. Nevertheless, it is still complicated to identify
a unique behavior pattern of aggression. It is also necessary to highlight that most of the research is
carried out by English-speaking countries, so it is difficult to obtain free resources on the Internet for
the development of tools that allow the analysis of comments written in Spanish. The Spanish language
has an extensive vocabulary with multiple expressions, words that vary in different Spanish-speaking
countries, synonyms and colloquialisms that are different in different countries and regions. For this
reason, we consider it prudent to focus on the Spanish language of Mexico. However, we do not rule
out the possibility of extending this research in other Spanish-speaking countries.
There are different types of cyber-aggression. Bauman [6] identifies and describes some forms
of cyber-bullying that are used to attack victims by digital means, such as flaming, harassment,
denigration, masquerading, outing, trickery, social exclusion, and cyber-stalking. Peter [16] highlights
death threats, homophobia, sexual acts, the threat of physical violence, and damage to existing
relationships, among others. Ringrose and Walker [17,18] indicate that women and children are the
most vulnerable groups in case of cyber-aggression.
In Mexico, there has been a wave of hate crimes; in 2014, the World Health Organization reported
that Mexico occupies second place in the world of hate crimes for factors such as “fear”. In Mexico,
the case of violence based on sexual orientation is one of the most vulnerable groups, according to
the National Survey on Discrimination [19]. As found in an investigation by a group of civil society
organizations [20], in Mexico and Latin America, 84% of lesbian, gay, bisexual, and transgender (LGBT)
population have been verbally harassed. On the other hand, violence against women is an exponential
problem that in Mexico has affected young women between 18 and 30 [21]. Mexico has a public
organization known as Instituto Nacional de Estadística y Geografía (INEGI) [22,23], responsible for
regulating and coordinating the National System of Statistical and Geographic Information, as well
as conducting a national census. INEGI [22] has reported that nine million Mexicans have suffered
at least one incident of digital violence in one of its different forms. In 2016, INEGI reported that of
the approximately 61.5 million women in Mexico, at least 63% aged 15 and older had experienced
acts of violence. Racism is another problem in Mexico that has been growing through offensive
messages related to “the wall” with the United States (US), the mobilization of immigrant caravans
and discrimination by skin color. According to an INEGI study, skin color can affect a person’s job
growth possibilities, as well as their socioeconomic status [23].
The American Psychological Association [24] says that different combinations of contextual factors
such as gender, race, social class, and other sources of identity can result in different coping styles,
and urges psychologists to learn guidelines for psychological practice with people with different
problems (e.g., lesbian, gay, and bisexual patients).
For many years, problems from various research areas have been addressed through Artificial
Intelligence (AI) techniques. AI is a discipline that emerged in the 1950s and is basically defined as
the construction of automated models that can solve real problems by emulating human intelligence.
Some recent applications include the following: in [25], two chemical biodegradability prediction
models are evaluated against another commonly used biodegradability model; Li et al. [26] applied
Multiscale Sample Entropy (MSE), Multiscale Permutation Entropy (MPE), and Multiscale Fuzzy
126
Appl. Sci. 2019, 9, 1828
Entropy (MFE) feature extraction methods along with Support Vector Machines (SVMs) classifier to
analyze Motor Imagery EEG (MI-EEG) data; Li and coworkers [27] proposed the new Temperature
Sensor Clustering Method for thermal error modeling of machine tools, then the weight coefficient in
the distance matrix and the number of the clusters (groups) were optimized by a genetic algorithm
(GA); in [28], fuzzy theory and a genetic algorithm are combined to design a Motor Diagnosis System
for rotor failures; the aim in [29] is to design a new method to predict click-through rate in Internet
advertising based on a Deep Neural Network; Ocaña and coworkers [30] proposed the evolutionary
algorithm TS-MBFOA (Two-Swim Modified Bacterial Foraging Optimization Algorithm) and proved a
real problem that seeks to optimize the synthesis of a four-bar mechanism in a mechatronic systems.
In this research, we used IA algorithms to classify comments in three cases of cyber-aggression:
racism, violence based on sexual orientation, and violence against women. The accurate identification
of cybernetic aggression cases is the first step of a process to reduce the incidence of this phenomenon.
We applied IA techniques, specifically Random Forest, Variable Importance Measures (VIMs),
and OneR. The comments used to create the data set were collected from Facebook, considering
specific news related to the cyber-aggression cases included in this study.
In line with the aims of this study, we propose to answer the following research questions:
1. Is it possible to create a model of automatic detection of cyber-aggression cases with high

precision? To answer this question, we will experiment with two classifiers of different approaches
and compare their performance.
2. What are the terms that allow detection of cyber-aggression cases included in this work effectively?
We will seek the answer to this question using methods to identify the relevant features for each
cyber-aggression case.
The present work is organized as follows: Section 2 presents a review of related research works,
Section 3 describes the materials and methods used in this research, Section 4 shows the architecture of
the proposed computational model, Section 5 describes the experiments and results obtained, and the
last section concludes the article.
2. Related Research
Due to the numerous victims of cyber-bullying and the fatal consequences it causes, today there is a
need to study this phenomenon in terms of its detection, prevention, and mitigation. The consequences
of cyber-bullying are worrisome when victims cannot cope with the emotional stress of abusive,
threatening, humiliating, and aggressive messages. Presently there are several types of research to
avoid or reduce online violence; even so, it is necessary to develop more precise techniques or online
tools to support the victims.
Raisi [31] proposes a model to detect offensive comments on social networks, in order to intervene
by filtering or advising those involved. To train this model, they used comments with offensive
words from Twitter and Ask.fm. Other authors [8,32] have developed conversation systems, based on
intelligent agents that allow supportive emotional feedback to victims who suffer from cyber-bullying.
Reynolds [33] proposed a system to detect cyber-bullying in the Formspring social network, based on
the recognition of violence patterns in the user posts, through the analysis of offensive words;
in addition, it uses a ranking level of the detected threat. Likewise, it obtained an accuracy of 81.7%
with J48 decision trees.
Ptaszynski [34] describes in his research the development of an online application in Japan,
for school staff and parents with the duty of detecting inappropriate content on unofficial secondary
websites. The goal is to report cases of cyber-bullying to federal authorities; in this work they used
SVMs and got 79.9% accuracy. Rybnicek [12] proposes an application for Facebook to protect minor
users from cyber-bullying and sex-teasing. This application aims to analyze the content of images and
videos, as well as the activity of the user to record changes in behavior. In another study, a collection
127
Appl. Sci. 2019, 9, 1828
of bad words was made using 3915 published messages tracked from the website Formspring.me.
The accuracy obtained in this study was only 58.5% [35].
Different studies fight against cyber-bullying by supporting the classification of situations,
topics or types of it. For example, at the Massachusetts Institute of Technology, [35] developed a
system to detect cyber-bullying in YouTube video comments. The system can identify the topic of the
message, such as sexuality, race, and intelligence. The overall success of this experiment was 66.7%
accuracy using SVMs. Similarly, the study carried out by Nandhini [36] proposes a system to detect
cyber-bullying activities and classify them as flaming, harassment, racism, and terrorism. The author
uses a fuzzy classification rule; however, the accuracy of the results are very low (around 40%), but he
increased the efficiency of classifier up to 90% using a series of rules. In the same way, Chen [37]
proposes an architecture to detect offensive content and identify potential offensive users in social
media. The system achieves an accuracy of 98.24% in sentence offensive detection and an accuracy
of 77.9% in user offensiveness detection. Similar work to this is presented by Sood [38], in which
comments were tagged from a news site using Amazon’s Mechanical Turk to create a profanity-labeled
data set. In this study, they use SVMs with a profanity list-based and Levenshtein edit distance tool,
getting an accuracy of 90%.
In addition to studies that aim to detect messages with offensive content or classify messages into
types of cyber-bullying, other studies are trying to prove that a system dedicated to taking care of user
behavior can reduce situations of cyber-bullying. Bosse [8] performed an experiment consisting of a
normative multi-agent game with children 6 to 12 years old. In the experiment, the author highlights
a particular case: a girl who, regardless of the agent’s warnings, continued to violate rules and was
removed from the game. However, she changed her attitude and began to follow the rules of the
game. Through this research, the author shows that in the long term, the system manages to reduce the
number of rule violations. Therefore it is possible to affirm that research with technological proposals
using sentiment analysis, text mining, multi-agent, or other AI techniques can support the reduction of
violence online.
3. Materials, Methods and Metrics
3.1. Materials
Study data. Since we did not find a Spanish data set for the study, we had to collect data through
the Facebook API from relevant news in Latin America related to three cases of cyber-aggression:
racism, violence based on sexual orientation, and violence against women. We collected 5000 comments;
however, we used only 2000, those free of spam (spam characterizes in this study as texts with rare
characters, images of expression or humorous thought such as memes, empty spaces, or comments
unrelated to the problem). We then grouped the comments (instances) as follows: 700 comments
about violence based on sexual orientation, 700 comments about violence against women and
600 racist comments.
Labeling process. In the literature, we found that some researchers [33,38–40] used the web
service of Amazon’s Mechanical Turk and paid for anonymous online workers to manually label
comments, reviews, words, images, among others. However, in [41], tagging by workers from
Amazon’s Mechanical Turk showed that at least 2.5% of reviews, classified as “no cyber-bullying”
should be tagged as “cyber-bullying”, for this reason, they sought support from graduate students and
some undergraduates of psychology.
Due to the above, we decided to use a group of three teachers with experience in machine-learning
algorithms supported by psychologists with experience in evaluation and intervention in cases of
bullying in high schools to manually tag the comments. The psychologists explained to professors
when a comment is considered offensive in the three cases of cyber-aggression and how to label the
comment according to the predefined range of values. So, the purpose of the labeling process was to
add an offensive value to each comment considering the case of cyber-aggression and a predefined
128
Appl. Sci. 2019, 9, 1828
numerical scale. In the cases of comments with violence against women we used a scale from zero
to two, with the lowest value of those comments the least offensive. For comments about violence
based on sexual orientation, a scale of values of four to six was used, in the same way with the lowest
value for the least offensive comments. Finally, a scale of eight to ten was used for racist comments,
with the lowest value for the least offensive comments and the highest value for the most. As a result,
we obtained the data set of offensive comments and then used them in the feature-selection procedure
and training process. This data set consisted of two columns; the first column contained the instance
or comment and second the offensive value according to each comment. We describe, in the following
section, the algorithms and methods used in this research for the feature-selection procedure and
training process.
3.2. Methods
Random Forest. The Random Forest classification algorithm, developed by Breiman [42], is a set
of decision trees that outputs a predictive value and is robust against over-fitting. This algorithm has
two parameters, mtry and ntree, which may vary to improve its performance. The first represents the
number of input variables chosen at random in each division and the second the number of trees.
Random Forest obtains a class vote of each tree and proceeds to classify according to the vote of
the majority. We show the functioning of Random Forest in Algorithm 1.
Algorithm 1: Random Forest for classification

1 Let N be the number of training cases, and M the number of variables.
2 Select the value of m, in order to determine the decision in a given node, which must be less
than the value of M.
3 Vary and choose N for the tree (completing all training cases) and use the rest of the test cases
to estimate the error.
4 For each node of the tree, randomly choose m. Obtain the vote of each tree and classify
according to the vote of the majority.
√
In classification processes, the default value for m is calculated by p or [log2 p], where p is the
number of features. Random Forest provides the VIMs method, through which it is possible to rank
the importance of variables in regression or classification problems.
Variable Importance Measures (VIMs). VIMs, based on CART classification trees [43], allows
calculation of the weight to identify key attributes or features. In VIMs there are two embedded
methods for measuring variable importance proposed by Breiman: Mean Decrease Impurity
importance (MDI) and Mean Decrease Accuracy (MDA). MDI ranks each feature importance as
the sum over the number of splits that include the feature, proportionally to the number of samples
it splits; this method is used in classification cases and applied the Gini index to measure the node
impurity [44,45]. Random Forest also uses the Gini index for determining a final class in each tree.
MDA is also called the permutation importance and is based on out-of-bag (OOB) [46] samples to
measure accuracy; this method is suitable for regression problems [43]. In this work we used MDI
with Gini index to classify comments in cases of cyber-aggression, which is defined by:
Q
G (t) = 1 − ∑ p2 ( k | t ) (1)
k =1
where Q is the number of classes, p(k |t) is the estimated class probability for feature t or node t in a
decision tree and k is an output class.
At the end of the process, VIMs calculate a score for each variable, normalized by the standard
deviation. The higher the score, the more important the feature. In this work, we used R [47] and
applied the randomForest [48] library to develop the classification model.
129
Appl. Sci. 2019, 9, 1828
OneR. OneR stands for One Rule. OneR is a rule-based classification algorithm. A classification
rule has the form: if attribute1 <relational operator> value1 <logical operator> attribute2 <relational
operator> value2 <...> then decision-value. OneR generates one rule for each predictor in the data
set, and among them all, it selects the rule with the smallest total error as the only one rule [49].
One advantage of rules is that they are simple for humans to interpret.
3.3. Evaluation Metrics

Confusion Matrix. It shows the performance of the prediction model, comparing the results of the
predictive model against real values.
Accuracy (ACC). This is the evaluation of the predictive model performance. In binary
classification, the accuracy is calculated by dividing the number of correctly identified cases among
the total cases. It is computed by:
TP + TN
ACC = (2)
T+N
Sensitivity (TRP). This metric measures the proportion of true positives which were correctly
classified and is calculated by:
TP
TRP = (3)
TP + FN
Specificity (SPC). This metric measures the proportion of true negatives, which were correctly
classified. It is given by:
TN
SPC = (4)
FP + TN
where TP = True Positive, TN = True Negative, FP = False Positive and FN = False Negative,
in Equations (2)–(4).
Kappa statistic. This method introduced by Cohen [50] and Ben-David [51] is used to measure the
accuracy of machine-learning algorithms. Kappa measures the agreement between the classifier itself
and the ground truth corrected by the effect of the agreement between them by chance. This method is
defined by:
Po + Pc
K= (5)
1 − Pc
where Po is the proportion of agreement between the classifier and the ground truth, Pc is the proportion
of agreement expected between the classifier and the ground truth by chance.
4. Experimental Procedure
4.1. Random Forest

This section describes the steps followed to carry out the experimental procedure with Random
Forest (see Figure 1).
Phase 1. Comment Extraction. We detail this step in the Materials section (study data).
Phase 2. Feature Selection. In this phase, we applied the VIMs method to identify the key
attributes for each cyber-aggression case [43]. The feature-selection process is shown in Figure 2.
We applied this process for each cyber-aggression case. The first step, after we cleaned the comments,
was to separate them according to a numerical range, as follows: 0–2 for comments related to violence
based on sexual orientation, 4–6 for those about violence against women, and 8–10 for those related
to racism. Then, we identified the frequent terms to eliminate those repeated and not important.
Later, we partitioned the set of comments (corpus) using 2/3 for training and 1/3 for testing. At the
same time, we applied the VIMs method along with the Gini index to obtain the terms with the higher
130
Appl. Sci. 2019, 9, 1828
weight to create the feature corpus. After this process, we configured a vector of 30 seeds, and then we
applied the random. f orest.importance instruction. We executed this process 30 times, considering a
different seed to train with different partitions of the training corpus, in order to evaluate the results of
Random Forest. We considered high-performance values greater than 70%, and then it was possible to
use the corpus of features in the parameter optimization process, which we applied to find the most
appropriate value of mtry. In this process, we used the Fselector library in the R system for statistical
computing [47].
Figure 1. Architecture of the proposed computational model.
Figure 2. Feature-selection process.
Phase 3. Classification. This step is the training process considering the feature corpus, identified
by VIMs (applied in phase 2 of the Model). We used 10-fold cross-validation and Random Forest
algorithm to adjust the mtry and ntree parameters. After the training process, we carried out the
classification process using Random Forest, which we executed another 30 times, implementing in
131
Appl. Sci. 2019, 9, 1828
each execution a different seed, generated by the Mersenne-Twister method [52]. To test the result of
the classification in different executions with different sections of training-sets, a variety of seeds is
important. Figure 3 represents the steps followed in this phase.
Figure 3. Classification process.
4.2. OneR
This section describes the experimental procedure for OneR classifier.
For each aggression case, we first calculated the frequency of appearance of each term and the
average frequency of appearance of all the terms. Then, we selected the terms above this average to
use the most frequent terms as predictors. Afterwards, we made a subsetting of the original data set
containing only the selected terms and created a new data set. From this new data set, we created three
data sets to perform the classification experiments. In the first data set, we kept the LGBT class and
marked the other two classes as ALL; in the second we kept the Machismo class and marked the other
two as ALL; and also with the third. Finally, we performed 30 independent runs, where we calculated
the average values of the metrics used across all the study.
5. Results
5.1. Feature Selection Using VIMs

We used the VIMs method to identify the key features of each cyber-aggression case, as shown
in Figure 2 and described in Phase 2 (feature selection). The result of this process (see Table 1) was
made in two executions—initial and final. In the initial execution, we showed the results without
a cleaning process; in the final execution, the results are shown considering a cleaning process and
the application of VIMs method. Column 1, “Case”, of Table 1, represents the cyber-aggression case,
where VS is violence based on sexual orientation, VM is violence against women and R is racism.
The second column, “execution”, indicates the initial and final executions from each cyber-aggression
case. The third column, “Potential features”, presents the arithmetic mean of potential features in
each cyber-aggression case. Column 4 shows the arithmetic mean of weights obtained from all terms,
where we can see that values from the final execution increase in comparison with the result from the
initial execution. In column five, “Maximum weight” represents the maximum weights obtained in
each cyber-aggression case and each type of execution; this term appears in Table 2, in the column of
Importance from each type of cyber-aggression. The final column, “Minimum weight”, is the minimum
weights obtained in each case of cyber-aggression and each case of execution.
132
Appl. Sci. 2019, 9, 1828
Table 1. Results of the feature-selection procedure.
Case Execution Potential Features Average Weights Maximum Weight Minimum Weight
VS Initial 33.33% 5.43 29.44 1.00
VS Final 89.33% 6.44 33.45 0.13
VM Initial 29.16% 4.33 21.09 0.02
VM Final 88.49% 5.24 27.04 0.075
R Initial 48.06% 4.65 23.66 1.00
R Final 89.10% 6.87 28.04 0.20
The processing and cleaning of comments influenced the results of the final execution, because in
the first execution some irrelevant words such as “creo” (“I think” in English) obtained a high value of
11.70, “comentarios” (“comments” in English) obtained a value of 5.36 and “cualquiera” (“anyone” in
English) obtained a value of 2.34.
We measured the importance of the terms with VIMs using the Gini index. Table 2 shows the
most important terms that were identified by this method, which are those with the highest value.
In the case of violence based on sexual orientation, the most important term was “asco” (“disgust”)
with a value of 33.4539, in the case of violence against women, the most important term was “estúpida”
(“stupid”) with a value of 27.0473, and in racist comments, the most important term was “basura”
(“garbage”) with a value of 28.0465. We show in Table 2 an extract of important term results (feature
selection) for each cyber-aggression case after the cleaning of comments; some terms include an “*”,
which is not part of the original term, but was used to censor the offense.
Table 2. Results of the importance of the terms in the feature-selection procedure.
VS Term Importance VM Term Importance R Term Importance

asc* 33.4539 est*pida 27.0473 b*sura 28.0465
circo 28.0842 c*chetadas 19.5111 escl*vos 26.3688
dios 25.7459 c*ger 17.7479 color 22.8754
d*generados 24.1965 c*lera 16.3509 ch*nos 20.9468
ab*rración 23.9967 cualquiera 14.4007 as*sinos 19.5273
an*rmal 18.0855 estorban 13.0006 r*za 18.9392
asqu*roso 17.8648 p*rras 12.7073 n*gro 12.4821
en*ermedad 17.6001 cocina 12.6314 en*no 12.2413
depr*vados 14.4772 drama 12.3626 an*mal 10.6022
an*rmales 13.5536 br*ta 11.6915 b*gotonas 10.0998
mar*cas 12.6927 c*lebra 11.0332 cr*stianos 9.2143
est*pidez 11.0387 p*ta 10.3944 pr*etos 7.3209
biblia 10.9030 ab*sivas 10.3768 t*rroristas 6.6306
cabr*nes 10.8184 b*rra 9.5017 c*zar 6.2582
enf*rmos 10.6553 débil 8.9978 m*tar 6.0199
cerrada 10.2412 descuidada 8.7899 enemigos 5.5840
an*males 9.5388 arg*enderas 8.4999 ch*cha 4.6080
arrepientanse 9.1431 asqu*rosas 8.1761 ind*o 4.4520
d*sviados 8.9325 c*gan 7.8184 d*lincuente 4.3617
deberían 8.4167 m*ntenidas 7.6039 gr*ngo 4.3598
Once we identified the features for each problem, we carried out the training process (phase 3),
where we evaluated the results using the metrics described in Section 3.3 (evaluation metrics).
Finally, we applied the classifier model, if the accuracy exceeded 70%.
5.2. Training Process and Parameter Tuning

To allow the algorithm to be prepared to perform a better classification, we carried out a training
process with Random Forest. We selected the 50 features with the highest weight of each problem;
also, we performed 10-fold cross-validation and a variation of the value of the mtry parameter using
133
Appl. Sci. 2019, 9, 1828
random numbers between zero and 50. We used these random numbers in the same way for each
problem. We measured the training performance of the algorithm using the accuracy and Kappa
metrics. In Table 3, we show the overall results obtained in this process, where R is racism, VS is
violence based on sexual orientation and VM is violence against women. In column two (Mtry),
we show the optimal values of the mtry parameter obtained for each case of cyber-aggression, column
three presents the accuracy obtained with respect of the value of the mtry parameter, and column four
shows the value of Kappa with respect of the value of the mtry parameter. The values used in the mtry
parameter were in a range of 23 to 44. Figure 4 shows the complete results for the adjustment of the
mtry parameter.
Table 3. Results of the adjustment of the mtry parameter.
Case Mtry Accuracy Kappa

VS 50 0.8211 0.5789
VM 23 0.8327 0.6217
R 44 0.9169 0.7737
Figure 4. Graph of adjustment of the results of the mtry parameter.
In the case of the ntree parameter, we considered two values: 100 and 200, based on the
investigations reported in [53,54]. Oshiro [53] concludes that the variation of the mtry with values that
indicate a higher number of trees only increases the computational cost and does not have a significant
improvement ratio.
Once we obtained the value of mtry for each case of cyber-aggression, as well as the value of
ntree, we carried out the classification process with Random Forest. It is essential to highlight that we
made this process with 30 executions varying the training and test set, in order to test the performance
of Random Forest with different sections of the data set (explained in phase three of the architecture of
the model).
Based on the results obtained with accuracy and confusion matrices, in the 30 executions, metrics
such as balanced accuracy, sensitivity, and specificity were obtained. These results are presented in
Table 4, as well as the standard deviation (sd) of each metric depending on cyber-aggression case.
134
Appl. Sci. 2019, 9, 1828
As can be seen, the model obtains a better performance in the presented metrics, even though in
the three problems it highlights the sensitivity with low values and the specificity with high values.
The performance of the model indicates that negative comments were correctly classified; this proves
that the results were very close to the expected performance because 95% of the comments in this
data set were offensive. Since this work focuses on detecting potential features to classify comments
on three problems of cyber-aggression (racism, sexual violence, and violence against women), it was
justified in implementing a data set composed primarily of negative comments.
Table 4. Results of comment classification. Standard deviation (sd) is also shown.
Case Accuracy Balanced Accuracy Sensitivity Specificity

R 0.9030 0.9237 0.6783 0.9883
sd 0.0129 0.0129 0.0466 0.0060
VS 0.8036 0.8313 0.5320 0.9591
sd 0.0162 0.0203 0.0355 0.0115
VM 0.8204 0.8233 0.6325 0.9261
sd 0.0148 0.0187 0.0311 0.0165
5.3. Results with OneR

As Table 5 shows, we found an average frequency of terms for VS aggression of 13.41, 21.61
for VM, and 18.60 for R. This means that for example in VS, terms appear 13.41 times as an average,
where the most frequent term is “mar*c*nes”, with 29 appearances. In VM, terms appear 21.61 times
as an average, the most frequent term being “vieja” with 76 appearances. Finally, in R, terms appear
18.60 times as an average, with “negros” as the most frequent with 63 appearances. We show the
selected terms in the same table, i.e., those with a frequency of appearance greater than the average
frequency of appearance of all the terms. Figure 5, known as word clouds, show the importance
of the terms for each type of aggression. The bigger the term in the figure, the more relevant it is.
These pictures match the terms shown in Table 5. As expected, the most common offensive terms in
Spanish are present. The most common offensive words are “maric*nes”, “asco”, “put*s”, “gays”,
“gay”, “mi*rda” for VS; “vieja”, “pinch*”, “mujeres”, “put*”, “viejas”, “pinch*s” for VM, and ”negros”,
“mi*rda”, “indio”, “negro”, “judíos”, “musulmanes” for R. We used the terms from Table 5 as input to
OneR classifier to build the classification models for each cyber-aggression case.
Table 5. Selected terms by their average frequency for each type of aggression.
Case Average Frequency of Terms Selected Terms (Quantitiy)

VS 13.41 maric*nes, asco, put*s, gays, gay, mi*rda,
hombres, pinch*s, jot*s, maric*s, dios, mundo,
gente, homosexuales, ser, respeto, put* (17)
VM 21.61 vieja, pinch*, mujeres, put*, viejas, pinch*s,
bien, mujer, p*nd*ja, put*s, solo, así (12)
negros, mi*rda, indio, negro, judíos, musulmanes,
R 18.6 malditos, p*nd*ejo, mexicanos, pinch* (10)
135
Appl. Sci. 2019, 9, 1828
(a) (b)
(c)
Figure 5. Word clouds for each cyber-aggression case. (a) Violence based on sexual orientation (VS).
(b) Violence against women (VM). (c) Racism (R).
Table 6 shows the classification results with OneR using only the most relevant terms found,
as described in the Experimental Procedure section. We show the results averaged over 30 independent
runs. We also show the standard deviation (sd). We found an accuracy above 0.90 for all types of
aggression. Moreover, all metrics reached 0.90 in most cases. Results prove the efficiency of the
methodology applied.
Table 6. Classification results using only the most relevant terms for each type of aggression. The results
are shown averaged over 30 runs. Standard deviation (sd) is also shown.
Case Accuracy Balanced Accuracy Sensitivity Specificity Kappa

R 0.9623 0.9569 0.9712 0.9426 0.9119
sd 0.0080 0.0103 0.0071 0.0184 0.0188
VS 0.9525 0.9192 0.8383 1.0000 0.8795
sd 0.0102 0.0174 0.0348 0.0000 0.0274
VM 0.9545 0.9615 0.9274 0.9957 0.9066
sd 0.0099 0.0112 0.0138 0.0237 0.0206
136
Appl. Sci. 2019, 9, 1828
In Table 7, we show the best classification rule generated by OneR for each type of aggression.
Also, the number of correctly classified instances and the balanced accuracy in the test set represents
the effectiveness of each rule. We can see that OneR selected the same most relevant terms for each
type of aggression. Some irrelevant terms in Spanish were included, such as “así”, “bien”, “solo”,
among others. They represent adverbs and prepositions instead of nouns or adjectives.
Table 7. Best rule for each aggression case.
Case VS VM R
Classification termino: termino: termino:
Rule asco -> lgbt asco -> ALL asco -> ALL
dios -> lgbt dios -> ALL dios -> ALL
g*y -> lgbt g*y -> ALL g*y -> ALL
g*ys -> lgbt g*ys -> ALL g*ys -> ALL
gente -> lgbt gente -> ALL gente -> ALL
hombres -> lgbt hombres -> ALL hombres -> ALL
homosexuales -> lgbt homosexuales -> ALL homosexuales -> ALL
j*tos -> lgbt j*tos -> ALL j*tos -> ALL
m*ric*s -> lgbt m*ric*s -> ALL m*ric*s -> ALL
m*ric*nes -> lgbt m*ric*nes -> ALL m*ric*nes -> ALL
mi*rda -> ALL mi*rda -> ALL mi*rda -> racismo
mundo -> lgbt mundo -> ALL mundo -> ALL
pinch*s -> ALL pinch*s -> machismo pinch*s -> ALL
p*ta -> ALL p*ta -> machismo p*ta -> ALL
p*tos -> lgbt p*tos -> ALL p*tos -> ALL
respeto -> lgbt respeto -> ALL respeto -> ALL
ser -> lgbt ser -> ALL ser -> ALL
así -> ALL así -> machismo así -> ALL
bien -> ALL bien -> machismo bien -> ALL
mujer -> ALL mujer -> machismo mujer -> ALL
mujeres -> ALL mujeres -> machismo mujeres -> ALL
p*nd*ja -> ALL p*nd*ja -> machismo p*nd*ja -> ALL
p*nche -> ALL p*nche -> machismo p*nche -> ALL
p*tas -> ALL p*tas -> machismo p*tas -> ALL
solo -> ALL solo -> machismo solo -> ALL
vieja -> ALL vieja -> machismo vieja -> ALL
viejas -> ALL viejas -> machismo viejas -> ALL
indio -> ALL indio -> ALL indio -> racismo
judíos -> ALL judíos -> ALL judíos -> racismo
malditos -> ALL malditos -> ALL malditos -> racismo
mexicanos -> ALL mexicanos -> ALL mexicanos -> racismo
musulmanes -> ALL musulmanes -> ALL musulmanes -> racismo
negro -> ALL negro -> ALL negro -> racismo
negros -> ALL negros -> ALL negros -> racismo
p*nd*jo -> ALL p*nd*jo -> ALL p*nd*jo -> racismo
Correct instances 772/818 775/818 782/818
Balanced Accuracy 0.9458 0.9776 0.9774
6. Discussion and Conclusions

Cyber-aggression has increased negatively as the use of social networks increases, which is why in
this work we have sought to develop computational tools analyzing offensive comments and classified
into three categories.
There are already a variety of software tools or applications that operate under AI techniques to
detect offensive comments, filter them or send messages of support to the victim. However, it is still
necessary to improve the performance of these tools to get more effective predictions. The development
of tools that work in the Spanish language is also required, since most of the research targets
English-speaking countries, which is why it is difficult to obtain resources for algorithm training,
such as data sets, lexicons, corpora, among others. Moreover, it is crucial to consider certain idioms or
137
Appl. Sci. 2019, 9, 1828
colloquialisms of the region where the model applies, so the translation of the available resources of
the Web is not always convenient. Therefore it is necessary to create resources in the Spanish language.
According to the need for a data set of offensive comments in Spanish, and considering the example of
other related research, the authors have created their own data sets of comments using social networks
such as Twitter [55], Formspring.me [35], Facebook [56], Ask.fm [31] and others where cyber-bullying
has been increasing [57].
We decided to create a data set of offensive comments using Facebook. At first, Twitter was
used to extract comments on these three cases of cyber-aggression, but most of the comments were
irrelevant. The Twitter API allows download of comments using a hashtag, but there are few users
who make offensive comments and use a hashtag to identify the comment. For this reason, we decided
to use news on Facebook about marriage between people of the same sex, publications about women’s
triumphs, or reports of physical abuse and news about Donald Trump’s wall. Besides, Facebook is the
most-used social network in Mexico [58]. We believe that a gathering of more relevant news comments
such as abortion, adoption by same-sex couples, feminicides and other related news, not only from
Mexico but also from Latin America, as well as the inclusion of experts who study the Spanish language
and colloquialisms, can improve the performance of the classifier.
This paper describes the development of a model to classify cyber-aggression cases applying
Random Forest and OneR. We seek to initially impact Mexico, where there has been a wave of hate
crimes. Our contribution is as follows: (1) We created a data set with cyber-aggression cases from
social networks. (2) We focused on cyber-aggression cases in our native language, i.e., Spanish.
(3) Specifically, we were interested in the most representative types of cyber-aggression in our country
of origin, Mexico. (4) We identified the most relevant terms for the detection of cyber-aggression cases
included in the study. (5) We created an automatic detection model of cyber-aggression cases with
high precision and interpretable by the human being (rule-based).
The results obtained in this work with Random Forest support the identification of relevant
features to classify offensive comments into three cases: racism, violence based on sexual orientation,
and violence against women. Nevertheless, OneR outperformed Random Forest in identifying types
of cyber-aggression, in addition to providing a simple classification rule with the most relevant terms
for each type of aggression. Even when we obtained high-performance classification models with
this particular data, it is essential to highlight that the classifiers used in this study have a better
performance classifying offensive comments against the LGBT population and racists comments.
Therefore the exploration of other machine-learning techniques and the continuous update of the
offensive data set may not be ruled out in future work, thus allowing the analysis of another kind of
cyber-aggression case, e.g., those suffered by children. Also, it is important to continue improving
the feature-selection process. On the other hand, building an automatic labeling system for offensive
comments made by social networks users, and thus minimizing human error, will be of great help.
Finally, we will seek to identify the victims of cyber-aggression to provide them with psychological
attention according to the case of harassment that they suffer.
Author Contributions: Conceptualization—G.O.G.-E.; Methodology—G.O.G.-E., J.H.-T.; Software—G.O.G.-E.;

Supervision—M.V.-A.; Validation—G.O.G.-E.; formal analysis—J.H.-T.; writing—original draft preparation—G.O.G.-E.,
M.V.-A.; writing—review and editing—G.O.G.-E., M.V.-A.
Funding: This research received no external funding.
Acknowledgments: We want to extend our appreciation to Consejo Nacional de Ciencia y Tecnología (CONACYT)
(National Council for Science and Technology) under the ’Catedras CONACYT’ programme, No. 1591.
Conflicts of Interest: The authors declare no conflicts of interest.
138
Appl. Sci. 2019, 9, 1828
References
1. Kowalski, R.M.; Limber, S.P.; Limber, S.; Agatston, P.W. Cyberbullying: Bullying in the Digital Age; John Wiley
& Sons: Hoboken, NJ, USA, 2012.
2. Grigg, D.W. Cyber-aggression: Definition and concept of cyberbullying. J. Psychol. Couns. Sch. 2010,
20, 143–156. [CrossRef]
3. Kopeckỳ, K.; Szotkowski, R. Cyberbullying, cyber aggression and their impact on the victim—The teacher.
Telemat. Inform. 2017, 34, 506–517. [CrossRef]
4. Corcoran, L.; Guckin, C.; Prentice, G. Cyberbullying or cyber aggression?: A review of existing definitions of
cyber-based peer-to-peer aggression. Societies 2015, 5, 245–255. [CrossRef]
5. Watkins, L.E.; Maldonado, R.C.; DiLillo, D. The Cyber Aggression in Relationships Scale: A new
multidimensional measure of technology-based intimate partner aggression. Assessment 2018, 25, 608–626.
[CrossRef] [PubMed]
6. Bauman, S. Cyberbullying: What Counselors Need to Know; John Wiley & Sons: Hoboken, NJ, USA, 2014.
7. Fredstrom, B.K.; Adams, R.E.; Gilman, R. Electronic and school-based victimization: Unique contexts for
adjustment difficulties during adolescence. J. Youth Adolesc. 2011, 40, 405–415. [CrossRef]
8. Bosse, T.; Stam, S. A normative agent system to prevent cyberbullying. In Proceedings of the 2011
IEEE/WIC/ACM International Conferences on Web Intelligence and Intelligent Agent Technology-Volume
02. IEEE Computer Society, Lyon, France, 22–27 August 2011; pp. 425–430.
9. Kowalski, R.M.; Giumetti, G.W.; Schroeder, A.N.; Lattanner, M.R. Bullying in the digital age: A critical
review and meta-analysis of cyberbullying research among youth. Psychol. Bull. 2014, 140, 1073. [CrossRef]
[PubMed]
10. Williams, M.L.; Pearson, O. Hate Crime and Bullying in the Age of Social Media. 2016. Available online:
http://orca.cf.ac.uk/88865/1/Cyber-Hate-and-Bullying-Post-Conference-Report_English_pdf.pdf (accessed on
25 May 2018).
11. Del Rey, R.; Casas, J.A.; Ortega, R. The ConRed Program, an evidence-based practice. Comunicar 2012,
20, 129–138. [CrossRef]
12. Rybnicek, M.; Poisel, R.; Tjoa, S. Facebook watchdog: a research agenda for detecting online grooming
and bullying activities. In Proceedings of the 2013 IEEE International Conference on Systems, Man,
and Cybernetics (SMC), Manchester, UK, 13–16 October 2013; pp. 2854–2859.
13. DitchTheLabel.org. The Annual Cyberbullying Survey. 2013. Available online: https://www.ditchthelabel.
org/wp-content/uploads/2016/07/cyberbullying2013.pdf (accessed on 25 May 2018).
14. Turan, N.; Polat, O.; Karapirli, M.; Uysal, C.; Turan, S.G. The new violence type of the era: Cyber bullying
among university students: Violence among university students. Neurol. Psychiatry Brain Res. 2011, 17, 21–26.
[CrossRef]
15. Van Royen, K.; Poels, K.; Daelemans, W.; Vandebosch, H. Automatic monitoring of cyberbullying on social
networking sites: From technological feasibility to desirability. Telemat. Inform. 2015, 32, 89–97. [CrossRef]
16. Smith, P.K. Cyberbullying and cyber aggression. In Handbook of School Violence and School Safety; Routledge:
London, UK, 2012; pp. 111–121.
17. Ringrose, J.; Gill, R.; Livingstone, S.; Harvey, L. A Qualitative Study of Children, Young People and ’Sexting’:
A Report Prepared for the NSPCC; National Society for the Prevention of Cruelty to Children: London,
UK, 2012.
18. Walker, S.; Sanci, L.; Temple-Smith, M. Sexting: Young women’s and men’s views on its nature and origins.
J. Adolesc. Health 2013, 52, 697–701. [CrossRef]
19. CONAPRED. National Survey on Discrimination. 2010. Available online: https://www.conapred.org.mx/
userfiles/files/ENADIS-2010-Eng-OverallResults-NoAccss.pdf (accessed on 1 March 2019).
20. FUNDACIONARCOIRIS. 2a Encuesta Nacional sobre Violencia Escolar basada en la Orientación Sexual,
Identidad y Expresión de Género hacia Estudiantes LGBT en México. 2018. Available online: www.fundacio
narcoiris.org.mx (accessed on 19 February 2019).
21. INMUJERES. Ciberacoso. 2018. Available online: https://www.gob.mx/inmujeres/articulos/ciberacoso?id
iom=es (accessed on 19 February 2019).
139
Appl. Sci. 2019, 9, 1828
22. INEGI. Microdatos del Modulo sobre Ciberacoso (MOCIBA). 2015. http://www.beta.inegi.org.mx/conten
idos/proyectos/investigacion/ciberacoso/2015/doc/mociba2015_principales_resultados.pdf (accessed on
16 November 2018).
23. INEGI. Módulo de Movilidad Social Intergeneracional (MMSI). 2016. https://www.inegi.org.mx/program
as/mmsi/2016/ (accessed on 11 February 2019).
24. American Psychological Association. Guidelines for psychological practice with lesbian, gay, and bisexual
clients. Am. Psychol. 2012, 67, 10. [CrossRef]
25. Baker, J.R.; Gamberger, D.; Mihelcic, J.R.; Sabljic, A. Evaluation of Artificial Intelligence Based Models for
Chemical Biodegradability Prediction. Molecules 2004, 9, 989–1003. [CrossRef] [PubMed]
26. Li, M.a.; Liu, H.n.; Zhu, W.; Yang, J.F. Applying Improved Multiscale Fuzzy Entropy for Feature Extraction
of MI-EEG. Appl. Sci. 2017, 7, 92. [CrossRef]
27. Li, F.; Li, T.; Wang, H.; Jiang, Y. A Temperature Sensor Clustering Method for Thermal Error Modeling of
Heavy Milling Machine Tools. Appl. Sci. 2017, 7, 82. [CrossRef]
28. Kuo, C.C.; Liu, C.H.; Chang, H.C.; Lin, K.J. Implementation of a Motor Diagnosis System for Rotor Failure
Using Genetic Algorithm and Fuzzy Classification. Appl. Sci. 2017, 7, 31. [CrossRef]
29. Wang, Q.; Liu, F.; Xing, S.; Zhao, X. A New Approach for Advertising CTR Prediction Based on Deep Neural
Network via Attention Mechanism. Comput. Math. Methods Med. 2018, 2018, 1–11. [CrossRef]
30. Hernández-Ocaña, B.; Pozos-Parra, M.D.P.; Mezura-Montes, E.; Portilla-Flores, E.A.; Vega-Alvarado, E.;
Calva-Yáñez, M.B. Two-Swim Operators in the Modified Bacterial Foraging Algorithm for the Optimal
Synthesis of Four-Bar Mechanisms. Comput. Intell. Neurosci. 2016, 2016, 1–18. [CrossRef]
31. Raisi, E.; Huang, B. Cyberbullying identification using participant-vocabulary consistency. arXiv 2016,
arXiv:1606.08084.
32. Van der Zwaan, J.M.; Dignum, V.; Jonker, C.M. A conversation model enabling intelligent agents to give
emotional support. In Modern Advances in Intelligent Systems and Tools; Springer: Berlin/Heidelberg, Germany,
2012; pp. 47–52.
33. Reynolds, K.; Kontostathis, A.; Edwards, L. Using machine learning to detect cyberbullying. In Proceedings
of the 2011 10th International Conference on Machine learning and applications and workshops (ICMLA),
Honolulu, HI, USA, 18–21 December 2011; Volume 2, pp. 241–244.
34. Ptaszynski, M.; Dybala, P.; Matsuba, T.; Masui, F.; Rzepka, R.; Araki, K. Machine learning and affect analysis
against cyber-bullying. In Proceedings of the 36th AISB, Leicester, UK, 29 March–1 April 2010; pp. 7–16.
35. Dinakar, K.; Reichart, R.; Lieberman, H. Modeling the detection of Textual Cyberbullying. Soc. Mob. Web
2011, 11, 11–17.
36. Nandhini, B.S.; Sheeba, J. Online social network bullying detection using intelligence techniques.
Procedia Comput. Sci. 2015, 45, 485–492. [CrossRef]
37. Chen, Y.; Zhou, Y.; Zhu, S.; Xu, H. Detecting offensive language in social media to protect adolescent
online safety. In Proceedings of the 2012 International Conference on Privacy, Security, Risk and Trust and
2012 International Confernece on Social Computing, Amsterdam, The Netherlands, 3–5 September 2012;
pp. 71–80.
38. Sood, S.O.; Antin, J.; Churchill, E.F. Using Crowdsourcing to Improve Profanity Detection. In Proceedings of
the AAAI Spring Symposium: Wisdom of the Crowd, Palo Alto, CA, USA, 26–28 March 2012; Volume 12, p. 6.
39. Rosa, H.; Pereira, N.; Ribeiro, R.; Ferreira, P.; Carvalho, J.; Oliveira, S.; Coheur, L.; Paulino, P.; Simão, A.V.;
Trancoso, I. Automatic cyberbullying detection: A systematic review. Comput. Hum. Behav. 2019, 93, 333–345.
[CrossRef]
40. Potha, N.; Maragoudakis, M. Cyberbullying detection using time series modeling. In Proceedings of the 2014
IEEE International Conference on Data Mining Workshop, Shenzhen, China, 14 December 2014; pp. 373–382.
41. Ptaszyński, M.; Leliwa, G.; Piech, M.; Smywiński-Pohl, A. Cyberbullying Detection–Technical Report 2/2018,
Department of Computer Science AGH, University of Science and Technology. arXiv 2018, arXiv:1808.00926.
42. Breiman, L. Random forests. Mach. Learn. 2001, 45, 5–32. [CrossRef]
43. Hjerpe, A. Computing Random Forests Variable Importance Measures (VIM) on Mixed Numerical and Categorical
Data; KTH, School of Computer Science and Communication (CSC): Stockholm, Sweden, 2016.
44. Han, H.; Guo, X.; Yu, H. Variable selection using mean decrease accuracy and mean decrease gini based on
random forest. In Proceedings of the 2016 7th IEEE International Conference on Software Engineering and
Service Science (ICSESS), Beijing, China, 26–28 August 2016; pp. 219–224.
140
Appl. Sci. 2019, 9, 1828
45. Breiman, L.; Cutler, A. Breiman and Cutler’s Random Forests for Classification and Regression. 2018.
Available online: https://cran.r-project.org/web/packages/randomForest/randomForest.pdf (accessed on
12 April 2019).
46. Breiman, L. Out-Of-Bag Estimation; University of California: Berkeley CA, USA, 1996. Available online:
https://www.stat.berkeley.edu/~breiman/OOBestimation.pdf (accessed on 12 April 2019).
47. R Core Team. R: A Language and Environment for Statistical Computing; R Foundation for Statistical Computing:
Vienna, Austria, 2013.
48. Liaw, A.; Wiener, M. Classification and regression by randomForest. R News 2002, 2, 18–22.
49. Witten, I.; Frank, E.; Hall, M. Practical Machine Learning Tools and Techniques; Morgan Kaufmann: Burlington,
MA, USA, 2011.
50. Cohen, J. A coefficient of agreement for nominal scales. Educ. Psychol. Meas. 1960, 20, 37–46. [CrossRef]
51. Ben-David, A. About the relationship between ROC curves and Cohen’s kappa. Eng. Appl. Artif. Intell. 2008,
21, 874–882. [CrossRef]
52. Matsumoto, M.; Nishimura, T. Mersenne twister: A 623-dimensionally equidistributed uniform
pseudo-random number generator. ACM Trans. Model. Comput. Simul. 1998, 8, 3–30. [CrossRef]
53. Oshiro, T.M.; Perez, P.S.; Baranauskas, J.A. How many trees in a random forest? In International Workshop
on Machine Learning and Data Mining in Pattern Recognition; Springer: Berlin/Heidelberg, Germany, 2012;
pp. 154–168.
54. Torruco, J.H. Descriptive and Predictive Models of Guillain-Barré Syndrome Based on Clinical Data
Using Machine Learning Algorithms. Ph.D. Thesis, Universidad Juárez Autónoma de Tabasco, Tabasco,
México, 2015.
55. Sanchez, H.; Kumar, S. Twitter bullying detection. Ser. NSDI 2011, 12, 15.
56. Hamouda, A.E.D.A.; El-taher, F. Sentiment analyzer for arabic comments system. Int. J. Adv. Comput.
Sci. Appl. 2013, 4. [CrossRef]
57. Van Hee, C.; Jacobs, G.; Emmery, C.; Desmet, B.; Lefever, E.; Verhoeven, B.; De Pauw, G.; Daelemans, W.;
Hoste, V. Automatic detection of cyberbullying in social media text. PLoS ONE 2018, 13, e0203794. [CrossRef]
[PubMed]
58. INFOTEC. 13o. Estudio sobre los Hábitos de los Usuarios de Internet en México. 2018. Available online:
https://www.infotec.mx/work/models/infotec/Resource/1012/6/images/Estudio_Habitos_Usuarios
_2017.pdf (accessed on 11 February 2019).
141
MDPI
St. Alban-Anlage 66
4052 Basel
Switzerland
Tel. +41 61 683 77 34
Fax +41 61 302 89 18
www.mdpi.com
Applied Sciences Editorial Office

E-mail: [email protected]
www.mdpi.com/journal/applsci
MDPI
St. Alban-Anlage 66
4052 Basel
Switzerland
Tel: +41 61 683 77 34
Fax: +41 61 302 89 18
www.mdpi.com ISBN 978-3-03928-573-0

Sentiment Analysis For Social Media

Uploaded by

Copyright:

Available Formats

Sentiment Analysis For Social Media

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Sentiment Analysis For Social Media

Uploaded by

Copyright:

Available Formats

What is the topic of the document?

What is the topic of the document?

What are some of the articles included in the special issue discussed?

What are some of the articles included in the special issue discussed?

Sentiment Analysis

for Social Media

Special Issue Editors

ISBN 978-3-03928-572-3 (Pbk)

About the Special Issue Editors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

Carlos A. Iglesias and Antonio Moreno

Hannah Kim and Young-Seob Jeong

Mohammed Jabreel and Antonio Moreno

Eline M. van den Broek-Altenburg and Adam J. Atherly

Sunghee Park and Jiyoung Woo

Guadalupe Obdulia Gutiérrez-Esparza, Maite Vallejo-Allende and José Hernández-Torruco

Antonio Moreno (Professor). Dr. Moreno is a Full Professor of Artiﬁcial Intelligence at

Keywords: sentiment analysis; emotion analysis; social media; affect computing

2. New Paths in Sentiment Analysis on Social Media

Appl. Sci. 2019, 9, 5037; doi:10.3390/app9235037 1 www.mdpi.com/journal/applsci

3. Applications of Sentiment Analysis in Social Media

Funding: This research received no external funding.

1.1. Background and Purpose of Research

Appl. Sci. 2019, 9, 2419; doi:10.3390/app9122419 5 www.mdpi.com/journal/applsci

1.2. Research Process

Figure 2. Research framework.

2. Materials and Methods

2.1. Sentiment Analysis and Visualization

2.2. Movie Recommendation

2.3. Network Visualization and Layouts

3. Sentiment Analysis Data Processing

3.1. Sentiment Words Collection

3.2. Sentiment Words Reﬁnement

j = number of words in sentimental group t

N = total number of documents in the corpus N = |D|

TFIDF(t, d, D) = t f (t, d) ∗ id f (t, D) (3)

Clustering Characteristics Sentiment Words

3.3. Movie Data Collection

4.1. Heatmap Visualization

4.2. Sentiment Movie Network

Algorithm 1. Separate nodes–avoid crowding neighbors

4.3. Constellation Visualization

Table 2. List of candidate graphics linked to a cluster of sentiment words.

Cluster Name List of Candidate Graphics

Constellation visualization helped us naturally perceive characteristics of a node without having

5. CosMovis Network Analysis Scenario

6.1. Evaluation 1: User Awareness Level Test on Network Location

Table 4. Results of independent t-test analysis.

Equal Variance P-Value/Alternative

Table 5. Details regarding the two test groups.

Question 95% Conﬁdence Before After

Figure 12. Comparison of correct answer rate between test groups.

6.2. Evaluation 2: Test on Usability Impact of the Heatmap Function

Cronbach’s α Cronbach’s α Total

Equal Variance P-Value/Alternative

Question 95% Conﬁdence Provide Non-Provide

6.3. Evaluation 3: Reaction Time Test on Adding Constellation Visualization Function

Table 9. The basic structure of visualization (1 to 3).

Visualization Type Basic Structure

Table 10. Polygon-related questions.