Twitter Sentimental Analysis using Deep Learning Techniques

International Journal of Scientific Research in Computer Science, Engineering and Information Technology ISSN : 2456-3307 (www.ijsrcseit.com) doi : https://doi.org/10.32628/IJSRCSEIT Twitter Sentimental Analysis using Deep Learning Techniques Dr. Pamela Vinitha Eric, Anu Priya K R Professor, Student New Horizon College of Engineering, Bengaluru, Karnataka, India ABSTRACT Article Info Volume 6, Issue 4 There is a rapid growth in the domain of opinion mining as well as sentiment analysis which targets to discover the text or opinions present on the disparate Page Number: 381-386 social media plat- forms via machine-learning (ML) with polarity calculations, Publication Issue : sentiment analysis or subjectivity analysis. Sentimental analysis (SA) indicates July-August-2020 the text organization which is employed to cate- gorize the expressed feelings or mindset in diverse manners like favorable, thumbs up, positive, unfavorable, Article History thumbs down, negative, etc. SA is a demanding and notable task that compris- Accepted : 25 July 2020 es i) natural-language processing (NLP), ii) web mining and iii) ML. Also, to Published : 05 Aug 2020 tackle this challenge, the SA is merged with deep learning (DL) techniques since DL models are efficient because of their automatic learning ability. This paper emphasizes re- cent studies regarding the execution of DL models like i) deep neural networks (DNN), ii) deep-beliefnetwork (DBN), iii) convolutional neural networks (CNN) together with, iv) re- current neural network (RNN) model. Those DL models aid in resolving different issues of SA like a) sentiment classification, b) the classification methods of i) rule-based classifiers(RBC), KNN and iii) SVM classification methods. Lastly, the classification methods’ performance is contrasted in respect of accu- racy. Keywords : Sentiment analysis, Opinion mining, Deep learning. I. INTRODUCTION renders the understandable information connected to the public views, as it examines diverse reviews and A. Sentimental Analysis tweets. It is a verified effectual tool for the prediction of numerous imperative events like general elections SA is contextual mining of text which recognizes and and also box office movies [2]. Pub- lic reviews are ex- torts subjective information from the source utilized to assess a specific entity, i.e., product, person material, and it also assists a business to comprehend or location which existson disparate websites like the social sentiment of their service, brand or product Yelp and Amazon. Therefore, SA is utilized for the whilst manages determination of the expressive directions of user sentiments, subjective text, and opinions [1,40]. SA reviews automatically [3]. The requirement for SA is observing online chats. SA Copyright: © the author(s), publisher and licensee Technoscience Academy. This is an open-access article distributed under the terms of the Creative Commons Attribution Non-Commercial License, which permits unrestricted noncommercial use, distribution, and reproduction in any medium, provided the original work is properly cited 381 Dr. Pamela Vinitha Eric et al Int J Sci Res CSE & IT, July-August-2020; 6 (4) : 381-387 elevated owing to the increased requisite of analyzing For the best SA of paragraphs and sentences, ‘Hidden and also structuring of the concealed information Markov model’ [10-12,43-48] is employed. The which comes as of the social media in the sort of un- optimiza- tion of words together with sentences structured data [4]. As imperative resources of real- brings faster learning which enhances data accuracy time opinion, Twitter, texts and the other social for social media. Data to- kenization at word root networks have fascinated substantial interests of the levels assists to create positive and negative facets of research in- dustry and community [5]. SA (opinion data. All those approaches are working harder to mining (OM)) of brief informal texts on social media diminish the errors in OM and SA to attain a better summarizes opinions as a) positive, b) neutral or c) negative statement of the opinion holder [6,41,42]. A level of data accurateness for social media [13]. million numbers of tweets are created daily on C. Techniques for sentimental analysis multifarious issues. Linguistic flexibility in expression SA has 2 categories of techniques, a) ML Approach and Topical diversity in content are 2 notable and b) Lexicon based approach [14-17]. challenges in examining tweets. Numerous twitter sentiment analyzers depend on diverse sentiment Machine Learning Approach lexicons either to feed features to classifier models or ML is the utmost prominent methodology gaining to ascertain sentiment scores [7]. the attention of researchers owing to its accuracy and adapta- bility [18]. In SA, mostly the supervised learning alternatives of this methodology are employed. It encom- passes 4 stages: i) Data collection, ii) Pre-processing, iii) Training data, iv) Classification as well as plotting results. Multiple tagged corpora are proffered on the training data. The Classifier presented numerous feature vectors from the former data. A model is built centered upon the Fig. 1. Diagram for sentimental analysis training data-set which is implemented over the new/hidden text for classification. In the ML technique, the key for classifier accuracy is the B. Features of Sentimental Analysis selection of pertinent features. Normally, i) unigrams (one-word phrases), ii) bi-grams (two successive Sentiments depend upon a certain range of values of fea- tures like bi-grams and also tri-grams with their phrases), iii) tri-grams (three successive phrases) are cho- sen as feature vectors. There are various polarities and also on their combinations [8, 9]. Their proposed features like a) number of negative words influences are iterative and slow in nature. So for continuing the work on the neural network’s hidden and positive words, b) the length of the document, c) layer, a kernel function is be- ing employed which algorithm (Naïve Bayes) [19-22]. Accu- racy differs evaluates the existence of class label. The conditional from 63% to 80% relying on the combination of dependencies betwixt the various edges and nodes of chosen features. Fig.2 delineates the working of an an acyclic graph are executed with the aid of ML approach. SVM (Support Vector Ma- chines), and d) NB ‘Bayesian networks’, which assist in the extortion of data at the contextual level. Volume 6, Issue 4, July-August-2020 | http://ijsrcseit.com 377 Dr. Pamela Vinitha Eric et al Int J Sci Res CSE & IT, July-August-2020; 6 (4) : 381-387 D. Deep Learning ML technology powers several aspects of modern commu- nity i.e. as of web searches, content filtering in social net- works to suggestions in e-commerce websites, in addition, it exists increasingly on consumer products like smartphones and cameras. ML systems are utilized to i) recognize objects in images, ii) match news articles, iii) transcribe speech Fig. 2. General structure for ML approach to text, iv) products or posts with con- sumer’s interests, and v) choose pertinent results of a search. Lexicon-based Approach These applications exploit a class of techniques termed DL. This technique is guided by the utilization of a diction- ary comprising pre-tagged lexicons. The DL is a representation-learning methodology with input text is transmuted to tokens by utilizing the multi-leveled representation, attained by composing Tokenizer. All newly arriving tokens are then sim- pler but non-linear (NL) modules where each matched for the lexicon in the dictionary. If a transmutes the representation in one level (beginning positive match is encountered, the score gets added to from the raw input) to a representation in a higher the total pool of a score for the inputted text e.g. if abstract level. With the compilation of such adequate ‘dramatic’ is positively matched in the dictionary transmutations, excep- tionally complex functions are then increment this text’s total score else decrement learned. DL comprises un- supervised learning or tag that word as negative. Albeit, this technique is together with supervised learning. amateur in nature, its variants are established to be valuable. Fig. 3 delineates the operations of a lexical E. Sentimental Analysis with Deep Learning technique. Recently, DL algorithms delivered impressive performance in NLP applications encompassing SA across numerous datasets. Such models don’t need any pre-defined features which are hand-picked by an engineer, but they could learn sophisticated features as of the dataset by themselves. Alt- hough every single unit in these Neural Networks (NN) is fairly simple, by means of stacking layers of NL units at the back of one another, those models are competent to learn highly sophisticated decision boundaries. Words are signified in a high-dimension vector space, and the feature extortion is left to the Fig. 3. General structure for a Lexicon-based approach NN. As an outcome, those models could map words with identical syntactic as well as seman- tic properties to adjacent locations in their coordinate Volume 6, Issue 4, July-August-2020 | http://ijsrcseit.com 378 Dr. Pamela Vinitha Eric et al Int J Sci Res CSE & IT, July-August-2020; 6 (4) : 381-387 sys- tem, in a way which is evocative of comprised a convolutional layer to extort information comprehending the words’ meaning. Architectures by a large piece of text, so SA with CNN exhibited like RNNs are also compe- tent to effectively that it attained augmented accuracy performance in comprehend the sentences’ structure. twitter sentiment clas- sification when contrasted to some traditional methodolo- gies like the SVM and These make DL models the best fit for tasks like SA. NB methods. II. LITERATURE SURVEY Xiao et.al [24] recommended a hybridized NN model This phase talks about the characteristic research architecture termed LSCNN with data augmentation technology (DAT), which outperformed numerous works related to SA utilizing DL field. SA tasks are single performed effectually by executing disparate models augmented the gener- alization competency of the like DL models, which have been extended recently. recommended Those models encom- pass RNN, CNN, DNN, RBC, exhibited KNN, SVM classifier, along with DBN. This section combination with the NNs model could attain delineates the efforts of disparate researchers toward astounding performance without any handcrafted executing DL models and ML approach for executing traits on SA or brief text classification. It was tested the SA. on a Chinese news headline corpus and Chinese on- NN models. that model. the The recommended Experirecom- ment mended DAT outcomes DAT in line com- ment dataset. It outperformed numerous A. Sentimental Analysis Using Convolutional Neural modern models. Evidence confirmed that the Networks (CNN) recommended DAT could attain more precise approach to distribution representation from data for DL, which comprehend real situations with the SA of a Twitter augmented the generalization traits of the extorted data centered on DL techniques. With the suggested features. The combination of the LSCNN fusion method, it was viable to forecast user satisfaction on a model and DAT was appropriate to brief text SA, product, happiness with a certain environment or destructive situation after disasters. Lately, DL was specifi- cally on the small-scale corpus. competent to resolve problems in voice recognition Jinzhan et.al [25] suggested a methodology for la or computerized vision. CNN worked fine for image beling the words of the sentences via integrating deep analysis together with classification. An impera- tive CNN (DCNN) with the sequential algorithm. Firstly, reason to employ CNN for image analysis and image classification was that the CNN could extort an area the aspects embraced by a) words vectors, b) part of speech vectors, c) dependent syntax vectors was of features as of global information precisely and also extorted to train the DCNN, and then the sequential it was competent to regard the relations amongst algorithm was em-ness of SA, it was suggested to those features. The above solution could attain the construct the tuples em- bracing aspect, sentimental utmost with shifter, sentiment intensity, sentimental words after classification. For NLP, texts’ data features could also attaining the sentimental labels for every word be extorted piece by piece. Regarding the relations existent in the sentence. Then, an algorithm was built amongst those features without considering the for inherent aspect recognition by considering the 2 context or complete sentence might incor- rectly key facets of the aspects as i) a topic- the matching interpret the sentiment. And, it was the most effectu- de- gree of aspects and ii) sentimental words- the al method to perform image classification. CNN human lan- guage habit. The experiment delineated Shiyang et.al [23] accuracy suggested in analysis an together Volume 6, Issue 4, July-August-2020 | http://ijsrcseit.com 379 Dr. Pamela Vinitha Eric et al Int J Sci Res CSE & IT, July-August-2020; 6 (4) : 381-387 that the algorithm could effectually detect the then supplied to a one-dimension CNN separately for inherent aspect. The issue of inherent aspect sentimental classifica- tion. This approach was recognition on SA and sentiment labeling was appraised on 4 sentimental classi- fication datasets resolved. As a fresh tool for SA, this methodology and contrasted with extensive baselines. Experiential could be employed to the enterprise management outcomes information analysis, like a) product online review, b) categorization could augment the performance of product online reputation, c) brand image and d) sentence- level SA; (2) the suggested approach attains consumer preference management, and could also be modern out- comes on numerous benchmarking utilized for the SA of huge text data. datasets. Gichang et.al [26] recommended a methodology for Table 1. Analysis of convolutional neural networks exhibited that: (1) sentence type recognizing keywords differentiating negative and posi- tive sentences by utilizing a weakly supervised learning methodology centered on a CNN. In this model, all words were signified as a continual-valued Researcher Name and year Model Used Purpose Limitations Less vector whereas, all sentences were signified as a matrix whose rows matched to the word vector utilized in the sentence. Subsequently, the CNN was Data Set Yazhi et.al [28] CNN SA Movie Review and IMDB convolutional layer utilized. trained utilizing those sentence matrices as inputs, in 5 datasets that are 1) STSTd data set, 2) SE2014 dataset, addition, the sentiment labels as an output. After training, the word attention scheme was implemented to recognize higher-contributing words to classify outcomes with the class activation map utilizing the weights. To vali- date the recommended 3) STSGd data set, methodology, the classification accurateness and the rate of polarity words amongst higher scoring words was assessed utilizing 2 movie review da- tasets. Experiential outcome confirmed that the recom- Zhao and Gui [29] DCNN mended model could correctly categorize the Twitter sentiment classification 5) SSTd. Paved attention on pretrained word embeddings. STS and MR Gold Dataset Only utilized smaller training dataset. 4) SED, and Comprehend sentence polarity and successfully recognize the matching words with the higher polarity scores. Shiyang et.al [23] CNN situations in the real world. Tao et.al [27] suggested a divide & conquers methodology which initially categorized the B. Sentimental Analysis Using Recurrent Neural sentences into disparate types, then executed the SA Networks (RNN) separately on sen- tences as of each type. Especially, Wenge et.al [30] recommended an SA model it was ascertained that the sentences tend to be centered on RNN, which took a part of a document as utmost intricate if it comprised more sentimental input and then the subsequent parts were utilized to words. Thus, it was suggested to employ an NN forecast the senti- mental label distribution. The centered sequence model to categorize opinionated recommended methodology learned words representation and also the sentimental dis- tribution. sentences into 3 types as per the count of targets transpired in a sentence. Each pool of sentences was Volume 6, Issue 4, July-August-2020 | http://ijsrcseit.com 380 Dr. Pamela Vinitha Eric et al Int J Sci Res CSE & IT, July-August-2020; 6 (4) : 381-387 Experiential studies were executed on commonly datasets like IMDB and SemEval- 2016. Experiential utilized datasets and the outcomes had proved its outcomes delineated that the design outperformed propitious potential. the baseline LSTM by 1%~2% in respect of accuracy and was effectual with notable performance en- Wen et.al [31] suggested an approach termed DRI- hancement over numerous non-RNN latent semantic RCNN (‘Deceptive Review Identification by RCNN’) de- signs (specifically in handling brief texts). It also to recognize deceptive reviews by utilizing DL and integrated the idea to an alternative of LSTM named word contexts. The fundamental idea was that, since the gated re- current unit (GRU) model and attained truthful and deceptive reviews were provided by writers with and without real experience fine performance, which confirmed that this methodology was adequate to augment disparate DL correspondingly, the re- view writers should have models. disparate context knowledge on their targeted goals under description. To distinguish the deceptive and C. Sentimental Analysis Using Deep Belief Networks truthful context knowledge embraced on the online (DBN) reviews, each word of a review was signified with 6 elements as a re-current convolutional vector (RCV). Shusen et.al [33] presented a 2-step SSL (semi- The primary and secondary components were 2 supervised learning) methodology termed fuzzy numerical word vectors attained from training DBNs deceptive reviews, Primarily, the common DBN was trained by the SSL respectively. The 3rd and 4th compo- nents were left by utilizing the training dataset. Then, a fuzzy neighboring truthful and deceptive context vectors membership function (FMF) was designed for all attained by means of training a RCNN on word classes of reviews centered on the DL archi- tecture. vectors and contextual vectors of left words. Also, the As the DBN training maps every review to the DBN 5th and 6th components were right neighboring output space, the dissemination of the entire training truthful and deceptive contextual vectors of right samples on the space was valued as prior knowledge, words. Further- in addition, was encoded by sequences of FMFs. Secondly, grounded on the fuzzy membership more, ReLU (Rectified Linear Unit) and max-pooling functions and the DBN attained in the primary step, filter was employed to transfer RCVs of words on a an FDBN architecture was built and the supervised review to a review vector by extorting positive learning stage was employed to increase the FDBN’s maximal feature elements in RCVs of words in the review. Experiment outcomes on the deception classification performance. FDBN inherited the powerful abstraction competency of DBN and dataset and the spam dataset delineated that the delineated suggested DRI-RCNN approach per- formed better competency for handling sentimental data. To take on considering the modern techniques in deceptive over the upsides of both FDBN and active learning, review recognition. an active FDBN (AFD) SSL method was suggested. together with truthful The (FDBN) the experiential for sentimental attractive validation classification. fuzzy on classification 5 sentimental Fei et.al [32] suggested an LSTM-centered design that classification datasets delineated the ef- fectiveness of was responsive to the words that existed in the AFD and FDBN methods. vocabu- lary; therefore, the keywords influence the semantics of the complete document. The suggested Yong et.al [34] suggested a word positional form model was assessed in a brief-text SA task on 2 together Volume 6, Issue 4, July-August-2020 | http://ijsrcseit.com with a word-to-segment matrix 381 Dr. Pamela Vinitha Eric et al Int J Sci Res CSE & IT, July-August-2020; 6 (4) : 381-387 representation to integrate the position information initialize the NN structure. The RBM layers could to DBNs for senti- mental classification. Subsequently, take probability distribution samples of the inputted the performance was assessed by the total accuracy. data to learn concealed structures for fine higher Therefore, these experien- tial outcomes exhibited level that by including positional infor- mation on ten (‘Classification RBM’) layer which was integrated on small text data sets, the matrix representation was RBM layers was employed to attain the final utmost effective. On considering the linear positional sentimental classification label intended for the posts. contribution form, it further suggested that the Experiential outcomes exhibited that with proper positional information should be regarded for SA or NLP tasks. parameters and structures, the performance of suggested DNN on sentimental classification was features’ representation. A Class RBM better on consider- ing recent surface learning Analysis Using Deep Neural Network (DNN) models like NB or SVM, which confirmed that the suggested DNN model was relevant for shorter Harika et.al [35] presented a scheme to spot the document classification with the suggested feature sentimental online Hindi product’s reviews centered dimension extension methodology. on its multiple modality natures (text together with audio). For every au- dio input, ‘Mel Frequency Shusen et.al [37] suggested an SSL algorithm termed Cepstral Coefficients’ (MFCC) features were extorted. ‘active deep network’ (ADN). Primarily, suggested These features were utilized to build a sentiment the SSL framework of ADN. ADN was built by RBM design utilizing DNN and GMM (Gauss- ian Mixture with un-supervised learning centered on labeled and Models) classifiers. maximal unlabeled reviews. After that, the built structure was modi- fied by means of gradient- From outcomes, it was perceived that DNN classifier descent centered supervised learning having an prof- fered better outcomes in contrast to GMM. exponential loss function. Secondly, in the SSL Further features of text were extorted from the framework, then active learning was employed to transcription of the audio input by utilizing Doc2vec vectors. SVM classifier was utilized to build a recognize reviews that were marked as training data, after that, utilized the chosen labeled and all sentimental model utilizing those textual features. unlabeled reviews for training ADN architecture. From experiential results, it was perceived that Furthermore, to integrate the information density integrating the text and audio features brought with AND suggested IADN (in- formation ADN) enhance- ment in the performance for spotting the sentiment of online products’ reviews. methodology, which could employ the information density of the entire un-labeled reviews in selecting the manually labeled reviews. Experiments on 5 Xiao et.al [36] suggested a contents extension sentimental classification datasets confirmed that structure (i.e), integrating posts and connected IADN and ADN outperformed the classical SSL comments to a microblog conversation intended for algorithms features extortion. A convolution auto-encoder was sentimental classification. and DL techniques employed for employed which could extort contextual information as of microblog conversation which was utilized as D. Sentimental Analysis Using Rule-Based Classifiers features intended for the posts. A custom DNN, [38] presented an effectual OM together with SA of which was integrated with numerous layers of RBM Web reviews utilizing disparate rule centered ML (‘Restricted Boltzmann Machine’), was executed to algorithms. To utilize SentiWordNet that created Volume 6, Issue 4, July-August-2020 | http://ijsrcseit.com 382 Dr. Pamela Vinitha Eric et al Int J Sci Res CSE & IT, July-August-2020; 6 (4) : 381-387 score count words from the 7 categories namely i) strong-positive, ii) posi- Discussion: The above figure 4 [36], contrasted the dis- parate classifier’s performance in respect of tive, iii) weak-positive, iv) neutral, v) weak-negative, accuracy. The accuracy range was varied based on the vi) negative and vii) strong-negative words. The number of com- ments (n). From the above figure, it present- ed approach was tested on online books and was clear that, when n=0, the SVM classifier offered political re- views and delineated the efficacy via 0.62 accuracy and when n= 10, it offered 0.72 Kappa measures, which had 97.4 % accuracy and accuracy. Similarly, KNN offered 0.64 accuracy for n= lesser error rate. The weighted mean of disparate accuracy measures namely Precision, TP-Rate and 0, but for n=10, it offered 0.63 accu- racy. Likewise, DBN offered 0.6 and 0.73 accuracies when n= 0 and Recall depicted higher efficacy rate and less FP-Rate. 10 respectively. Comparative experiments on disparate rule centered ML algorithms were performed via a 10-Fold cross- III. PROBLEM STATEMENT validation training design for sentimental classification. We use the dataset from Kaggle which was crawled E. Sentimental Analysis Using SVM Classifier and labeled positive/negative. The data pro- vided Vo et.al [39] suggested a model utilizing an SVM comes with emoticons, usernames and hashtags algo- rithm with the Hadoop M (Map)/ R (Reduce) which are required to be processed and converted for English document category emotion classification into a standard form. We also need to extract useful in the Cloud era parallel network environment. features from the text such uni- grams and bigrams Cloud era was also a disseminated system. This which is a form of representation of the “tweet”. We English testing dataset (ETD) had 25,000 documents, use various machine learning algorithms to conduct encompassing 12,500 posi- tive and also 12,500 sentiment analysis using the extracted fea- tures. negative reviews. This ETD had 90,000 sentences, However, just relying on individual models did not embracing 45,000 positive sentences together with 45,000 negative ones. This model was exper- imented give a high accuracy so we pick the top few models to generate a model ensemble. Ensembling is a form of on the ETD and attained 63.7% accuracy of senti- meta learning algorithm tech- nique where we mental classification on this ETD. combine different classifiers in order to improve the prediction accuracy. Finally, we report our experimental results and findings at the end. Methodology and Implementation 1.1 Pre-processing Raw tweets scraped from twitter generally result in a noisy dataset. This is due to the casual nature of people’s usage of social media. Tweets have certain Fig. 4. Compare the performance of the different classi- fier in terms of accuracy with the number of comments Volume 6, Issue 4, July-August-2020 | http://ijsrcseit.com special characteristics such as re- tweets, emoticons, user mentions, etc. which have to be suitably extracted. Therefore, raw twitter data has to be normalized to create a dataset which can be 383 Dr. Pamela Vinitha Eric et al Int J Sci Res CSE & IT, July-August-2020; 6 (4) : 381-387 vote over the predictions of 5 of our best models achieving an accuracy of 83.58%. easily learned by various classifiers. We have applied an extensive number of pre-processing steps to standardize the dataset and reduce its size. We first Every twitter user has a handle associated with them. do some general pre-processing on tweets which is as Users often mention other users in their tweets by follows. @handle. We replace all user mentions with the word USER_MENTION. The regular expression used • Convert the tweet to lower case. to match user mention is @[\S]+. • • Replace 2 or more dots (.) with space. Strip spaces and quotes (" and ’) from the ends Emotion of tweet. Users often use a number of different emoticons in URL their tweet to convey different emotions. It is Users often share hyperlinks to other webpages in their tweets. Any particular URL is not important for text classification as it would lead to very sparse features. Therefore, we re- place all the impossible to exhaustively match all the different emoticons used on social media as thenumber is ever increasing. However, we match some common emoticons which are used very frequently. We replace the matched emoticons. URLs in tweets with the word URL. The regular expression used to match URLs IV. CONCLUSION is ((www\.[\S]+)|(https?://[\S]+)). The provided tweets were a mixture of words, emoticons, URLs, hastags, user mentions, and symbols. Before training the we pre-process the tweets to make it suitable for feeding into models. We implemented several machine learning algorithms like Naive Bayes, Maximum Entropy, Decision Tree, Random Forest, XGBoost, SVM, Multi- Layer Perceptron, Recurrent Neural networks and Convolutional Neural Networks to classify the polarity of the tweet. We used two types of features namely unigrams and bigrams for classification and observes that augmenting the feature vector with User Mention bigrams improved the accuracy. Once the feature has been extracted it was represented as either a sparse Neural methods performed better than other vector or a dense vector. It has been observed that classifiers in general. Our best LSTM model achieved presence in the sparse vector representation recorded an accuracy of 83.0% on Kaggle while the best CNN a better performance than frequency. model achieved 83.34%. The model which used features from our best CNN model and classifies using SVM performed slightly better than only CNN. We finally used an ensemble method taking a majority Volume 6, Issue 4, July-August-2020 | http://ijsrcseit.com 384 Dr. Pamela Vinitha Eric et al Int J Sci Res CSE & IT, July-August-2020; 6 (4) : 381-387 [9]. Anil Bandhakavi, Nirmalie Wiratunga, Stewart V. REFERENCES Massie and Rushi Luhar, “Opinion context [1]. Sis using SVMclassifier, 2016 IEEE 20th Int. extraction for aspect sentiment analysis”, Conf. Comput. Support. Coop. Work Des., pp. Proceedings of the Twelfth International AAAI 276281, 2016. Conference on Web and Social Media (ICWSM [2]. M. Haenlein and A.M. Kaplan, “An empirical 2018), 2018. analysis of attitudinal and behavioral reactions [10]. Dattu, Bholane Savita, and Deipali V. Gore, “A toward the abandonment of unprofita- ble survey on sentiment analysis on twitter data customer relationships”, J. Relatsh. Mark., vol. 9, no. 4, pp. 200228, 2010. using different techniques”, International Journal of Computer Science and Information [3]. Nurulhuda Zainuddin, Ali Selamat, and Roliana Ibrahim, “Improv- ing twitter aspect-based Technologies, vol. 6, no. 6, pp. 5358-5362, 2015. [11]. Singh, Jaspreet, Gurvinder Singh, and Rajinder sentiment analysis using hybrid approach”, In Singh, “A review of sentiment analysis Asian Conference on Intelligent Information and techniques for opinionated web text”, CSI Trans- Database Sys- tems, pp. 151-160, Springer, actions on ICT, vol. 4, no. 2-4, pp. 241-247, 2016. [12]. Kaushik, Chetan, and Atul Mishra, “A scalable, Berlin, Heidelberg, 2016. [4]. Zhou Jin, Yujiu Yang, Xianyu Bao, and Biqing lexicon based tech- nique for sentiment Huang, “Combining user-based and global analysis”, International Journal in Foundations of lexicon features for sentiment analysis in twit- Computer Science & Technology (IJFCST), vol. ter”, In Neural Networks (IJCNN), International 4, no. 5, 2014. Joint Con- ference on, pp. 4525-4532, IEEE, [13]. Vohra, S.M., and J.B. Teraiya, “A comparative study of sentiment analysis techniques”, Journal 2016. [5]. Rafeeque Pandarachalil, Selvaraju Sendhilkumar, and G.S. Maha- lakshmi, “Twitter sentiment JIKRCE, vol. 2, no. 2, pp. 313-317, 2013. [14]. Lalji, T., and S. Deshmukh, “Twitter sentiment analysis for large-scale data: an un- supervised analysis using hybrid approach”, International approach”, Cognitive computation, vol. 7, no. 2, pp. 254-262, 2015. Research Journal of Engineering and Technology, vol. 3, no. 6, pp. 2887-2890, 2016. [6]. Abbasi, Ahmed, Stephen France, Zhu Zhang, [15]. Chandni, Nav Chandra, Sarishty Gupta and and Hsinchun Chen, “Selecting attributes for Renuka Pahade, “Senti- mental analysis and its sentiment classification using feature rela- tion challenges”, IJERT, vol. 4, no. 3, 2015. networks”, IEEE Transactions on Knowledge and Data Engi- neering, vol. 23, no. 3, pp. 447-462, [16]. Du, Jingcheng, Jun Xu, Hsingyi Song, Xiangyu Liu, and Cui Tao, “Optimization on machine learning based approaches for senti- ment 2011. [7]. Sarker, Abeed, Diego Mollá-Aliod, and Cécile analysis on HPV vaccines related tweets”, Paris, “Outcome po- larity identification of Journal of Biomedi- cal Semantics, vol. 8, no. 1, medical papers”, 2011. pp. 9, 2017. [8]. S.K. Bharti, B. Vachha, R.K. Pradhan, K.S. Babu [17]. Bhuvaneswari K. and R. Parimala, “Correlation and S.K. Jena, “Sarcastic sentiment detection in based feature selec- tion for movie review tweets streamed in real time: a big data sentiment classification”, International Journal of approach”, Advanced Digital Communications Networks, 2016. and Research in Computer Communication Engineer- ing, vol. 5, no. 7, 2016. Volume 6, Issue 4, July-August-2020 | http://ijsrcseit.com and 385 Dr. Pamela Vinitha Eric et al Int J Sci Res CSE & IT, July-August-2020; 6 (4) : 381-387 and [26]. Gao, Yazhi, Wenge Rong, Yikang Shen, and Mohamed Kissi, “Arabic text classification using Zhang Xiong, “Convo- lutional neural network new stemmer for feature selection and deci- sion based trees”, Journal of Engineering Science and combination”, In Neural Networks (IJCNN), Technology, vol. 12, no. 6, pp. 1475-1487, 2017. 2016 International Joint Conference on, pp. [18]. Bahassine, Said, Abdellah Madani, [19]. Agarwal, Basant, and Namita Mittal, “Categorical probability pro- portion difference (CPPD): a sentiment analysis using Adaboost 1333-1338. IEEE, 2016. [27]. Jianqiang, Zhao, Gui Xiaolin, and Zhang Xuejun, sentiment “Deep convolution neural networks for twitter classification”, In Proceedings of the 2nd Workshop on Sentiment Analysis Where AI sentiment analysis”, IEEE Access, vol. 6, pp. 23253-23260, 2018. feature selection method for Meets Psychology, COLING, pp. 17-26, 2012. [20]. Hasan, Ali, Sana Moin, Ahmad Karim, and Chao Li, and Zhang Xiong, “Structural “Machine information aware deep semi-supervised recur- learning-based sentiment analysis for twitter rent neural network for sentiment analysis”, accounts”, Mathematical and Computational Frontiers of Computer Science, vol. 9, no. 2, pp. Applications, vol. 23, no. 1, pp. 11, 2018. 171-184, 2015. Shahaboddin Sham- shirband, [28]. Rong, Wenge, Baolin Peng, Yuanxin Ouyang, [21]. Liao, Shiyang, Junbo Wang, Ruiyun Yu, Koichi [29]. Zhang, Wen, Yuhang Du, Taketoshi Yoshida, Sato, and Zixue Cheng, “CNN for situations and Qing Wang, “DRI-RCNN: An approach to understanding based on sentiment analysis of deceptive review identification using recurrent twitter data”, Procedia Computer Science, vol. convolutional neural network”, 111, pp. 376-381, 2017. Pro- cessing & Management, vol. 54, no. 4, pp. [22]. Sun, Xiao, and Jiajin He, “A novel approach to Information 576-592, 2018. generate a large scale of supervised data for short [30]. Hu, Fei, Li Li, Zi-Li Zhang, Jing-Yuan Wang, text sentiment analysis”, Multimedia Tools and and Xiao-Fei Xu, “Emphasizing essential words Applications, pp. 1-21, 2018. for sentiment classification based on recurrent [23]. Feng, Jinzhan, Shuqin Cai, and Xiaomeng Ma, “Enhanced sentiment labeling and implicit aspect identification by integration of deep convolution neural network and sequential algorithm”, Cluster Computing, pp. 1-19, 2018. neural networks”, Journal of Computer Science and Technology, vol. 32, no. 4, pp. 785795, 2017. [31]. Zhou, Shusen, Qingcai Chen, and Xiaolong Wang, “Fuzzy deep be- lief networks for semi- [24]. Lee, Gichang, Jaeyun Jeong, Seungwan Seo, CzangYeob Kim, and Pilsung Kang, “Sentiment supervised sentiment classification”, Neurocomputing, vol. 131, pp. 312-322, 2014. classification with word localization based on [32]. Jin, Yong, Harry Zhang, and Donglei Du, weakly supervised learning with a convolutional “Incorporating positional information into deep neural network”, Knowledge-Based Systems, vol. belief networks for sentiment classification”, In 152, pp. 70-82, 2018. Industrial Conference on Data Mining, pp. 1-15. [25]. Tao Chen, Ruifeng Xu, Yulan He, and Xuan Springer, Cham, 2017. Wang, “Improving sen- timent analysis via [33]. Abburi, Harika, Rajendra Prasath, Manish sentence type classification using BiLSTM-CRF Shrivastava, and Suryakanth V. Gangashetty, and CNN”, Expert Systems with Applications, “Multimodal sentiment analysis using deep 2016. neural networks”, In International Conference Volume 6, Issue 4, July-August-2020 | http://ijsrcseit.com 386 Dr. Pamela Vinitha Eric et al Int J Sci Res CSE & IT, July-August-2020; 6 (4) : 381-387 on Mining Intel- ligence and Knowledge Exploration, pp. 58-65. Springer, Cham, 2016. [34]. Sun, Xiao, Chengcheng Li, and Fuji Ren, “Sentiment analysis for Chinese microblog based on deep neural networks with convolution- al extension features”, Neurocomputing, vol. 210, pp. 227-236, 2016. Cite this article as : Dr. Pamela Vinitha Eric , Anu Priya K R, "Twitter Sentimental Analysis using Deep Learning Techniques", International Journal of Scientific Research in Computer Science, Engineering and Information Technology (IJSRCSEIT), ISSN : 24563307, Volume 6 Issue 4, pp. 381-387, July-August 2020. Journal URL : http://ijsrcseit.com/CSEIT206423 Volume 6, Issue 4, July-August-2020 | http://ijsrcseit.com 387