A Study Using N-Gram Features For Text Categorization
A Study Using N-Gram Features For Text Categorization
A Study Using N-Gram Features For Text Categorization
Johannes Fürnkranz
Austrian Research Institute for Artificial Intelligence
Schottengasse 3, A-1010 Wien, Austria
E-mail: [email protected]
Abstract
In this paper, we study the effect of using n-grams (sequences of words of length n) for
text categorization. We use an efficient algorithm for generating such n-gram features in two
benchmark domains, the 20 newsgroups data set and 21,578 REUTERS newswire articles.
Our results with the rule learning algorithm R IPPER indicate that, after the removal of stop
words, word sequences of length 2 or 3 are most useful. Using longer sequences reduces
classification performance.
1 Introduction
After Lewis’ influential thesis (Lewis 1992c), the use of Machine Learning techniques for Text
Categorization has gained in popularity (see, e.g., (Hearst and Hirsh 1996; Sahami 1998)). One
requirement for the use of most Machine Learning algorithms is that the training data can be
represented as a set of feature vectors. A straight-forward approach for r epresenting text as
feature vectors is the set-of-words approach: A document is represented by a feature vector that
contains one boolean attribute for each word that occurs in the training collection of documents.
If a word occurs in a particular training document, its corresponding attribute is set to 1, if not it
is set to 0. Thus, each document is represented by the set of words it consists of.1
In this paper, we study the effect of generalizing the set-of-words approach by using word
sequences, so-called n-grams, as features. We describe an algorithm for efficient generation and
frequency-based pruning of n-gram features in section 2. In section 3 we present the results on
two benchmark tasks, Ken Lang’s 20 newsgroups data set and the 21,578 REUTERS newswire
articles. The results indicate that word sequences of length 2 or 3 usually improve classification
1
A related approach, the bag-of-words approach, uses the frequencies of occurrence of the individual words as
feature values. The differences between both approaches in the context of naive Bayes classifiers were studied by
McCallum and Nigam (1998).
1
performance, while longer sequences are not as useful. They also show that moderate frequency-
based pruning of the feature set is useful, while heavy frequency-based pruning results in a
performance decrease on the studied datasets.
This property can be exploited in order to obtain a simple but efficient algorithm. The n-gram
features are generated by n different passes over the documents. In each pass, the number of
occurrences of each feature is counted, and a user-specified threshold is used to prune infrequent
features. In order to avoid the combinatorial explosion in the feature space, we can use the sub-
sequence property for pruning the search space: We only have to count sequences of n words
for which the sequences of the first n ? 1 and the last n ? 1 words have previously passed the
frequency threshold. Other sequences can be ignored.
Figure 1 shows the resulting algorithm. It takes three parameters: the collection of Docu-
ments, the maximum length of the features (MaxNGramSize), and a lower bound on the number
of occurrences of a feature (MinFrequency). The algorithm then computes all Features of length
at most MaxNGramSize that occur at least MinFrequency times in the Documents.
For computing this result, it performs MaxNGramSize passes over the document collection,
one for each possible feature length. In principle, however, one pass over the database would
be sufficient. Instead of merely counting the occurrences of each word, the algorithm has to
keep pointers to the positions where each feature in the text occurs. After computing this list of
2
procedure G ENERATE F EATURES(Documents,MaxNGramSize,MinFrequency)
Features[0] = f;g
for n = 1::MaxNGramSize
Candidates = ;
Features[n] = ;
foreach Doc 2 Documents
foreach NGram 2 NGrams(Doc,n)
InitialGram = NGram ? LastWord(NGram)
FinalGram = NGram ? FirstWord(NGram)
if InitialGram 2 Features[n ? 1]
and FinalGram 2 Features[n ? 1]
CounterfNGramg = CounterfNGramg +1
Candidates = Candidates [ NGram
foreach NGram 2 Candidates
if CounterfNGramg MinFrequency
Features[n] = Features[n] [ NGram
return Features
pointers in the first pass over the documents, the feature set of length n + 1 can be computed
from the feature set of length n by the following algorithm:
1. Find pairs of features that intersect (e.g. find pairs of and pairs of features)
2. For each such pair, compute the intersection of the position pointers of the two features.
This is defined as the subset of the position pointers of the first feature for which a pointer
to the immediately following position is contained in the set of position pointers of the
second feature.
3. Discard all features for which the number of associated position pointers is below the
frequency threshold.
This algorithm is inspired by the AP RIORIT ID algorithm, which is also described in (Agrawal
et al. 1995). It only has to read the documents once, but the memory requirements are much
higher than for the algorithm of figure 1 because it has to store a list of position pointers for
each feature (instead of using only a counter). For each iteration, the number of accesses to the
hash table that stores these position pointers is quadratic in the number of features found in the
previous iteration, while it is linear in the size of the document collection for the AP RIORI-based
algorithm. Consequently, we have found that additional passes over the document collection are
cheaper if the number of features is large. Only for higher n-gram sizes, when the size of the
feature sets becomes small (ca. 500), the use of position pointers begins to pay off.
We have implemented both algorithms in perl. The implementation has an additional pa-
rameter that can be used to specify with which iteration the mode should switch from making
3
additional passes through the document collection to using position indices. Another parameter
allows the user to not only specify a minimum term frequency (number of times a feature occurs
in the collection) but also a minimum document frequency (minimum number of documents in
which a feature must appear). A feature will be accepted if it is above both thresholds.
3 Experimental Results
We used the inductive rule learning algorithm R IPPER for experiments in two domains: the
21578 REUTERS newswire data and Ken Lang’s 20 newsgroups data set. In the following, we
briefly describe R IPPER, our experimental setup, and the results in both domains.
3.1 R IPPER
William Cohen’s R IPPER2 (Cohen 1995) is an efficient, noise-tolerant rule learning algo-
rithm based on the incremental reduced-error-pruning algorithm (Fürnkranz and Widm er 1994;
Fürnkranz 1997). What makes R IPPER particularly well-suited for text categorization problems
is its ability to use set-valued features (Cohen 1996). For conventional machine learning al-
gorithms, a document is typically represented as a set of boolean features, each encoding the
presence or absence of a particular word (or n-gram) in that document. This results in a very
inefficient encoding of the training examples because much space is wasted for specifying the
absence of words in a document. R IPPER allows to represent a document as a single set-valued
feature that simply lists all the words occurring in the text. Conceptually, this does not differ
from the use of boolean features in conventional learning algorithms, but R IPPER makes use of
some clever optimizations. In the remainder of this paper, we will frequently continue to refer to
each n-gram as a separate boolean feature.
4
3.3 20 Newsgroups
The first dataset we experimented with was Ken Lang’s 20-newsgroups data. This is a collection
of 20,000 netnews articles, about 1,000 from each of 20 different newsgroups. The dataset
is available from http://www.cs.cmu.edu/afs/cs/project/theo-3/www/. The task is to
identify to which newsgroup an article belongs to.
We evaluated R IPPER with various feature sets using its built-in cross-validation. Because of
the complexity, we chose to use only 5 folds. Note, however, that this procedure is problematic
because of the characteristics of newsgroup articles: It happens quite frequently that portions
of an articles are quoted in several subsequent articles of the same newsgroup. As such related
articles may appear in both, training and test sets, there is a danger of over-optimistic accuracy
estimates. However, we believe that the estimates are good enough for comparing different
versions of the same learning setup.
Table 1 shows the results. The first column shows the pruning parameters. We measured the
average error rate, the average run-time for the learning algorithm in CPU seconds (this does not
include the time needed for generating the feature set), and the (cumulative) number of generated
features for several different settings of the algorithm’s parameters, and for several different
maximal n-gram sizes. DF and TF stand for minimum document frequency and minimum term
frequency, respectively. The set-of-words setting refers to the conventional text learning setting
where each word is treated as a separate boolean feature.
The best results could be obtained with fairly moderate frequency-based pruning (all features
that occur at least 5 times in at least 3 documents are admitted) and the use of sequences with
maximum size 3. In all groups with identical pruning parameters (except for the ones with very
heavy pruning), the use of n-grams improves the results. However, sequences of length > 3 do
no longer improve the results (and make them worse in some cases). Frequency-based pruning
works well if the parameter settings are fairly low, but the results get worse with increasing
amounts of pruning. Obviously, several good features have a fairly low coverage and are thrown
away with higher settings of the pruning parameters.
A look at the highest ranked features shows that they are not very indicative of any of the
classes. The top ten features and their frequencies are shown in figure 2.
Obviously, none of the words are predictive of any of the classes. The first
word that seems to be predictive for some classes (soc.talk.religion.misc,
soc.religion.christian, and alt.atheism) is god, which is ranked 31 with 4550
occurrences. For higher n-gram sizes, the situation is similar. These problems could be alleviated
by tailoring the stop list to the domain specifics. However, this not only requires a considerable
effort but it also does not solve all problems: The repetitive nature of this domain (entire para-
graphs may be repeated in several documents) may lead to overfitting. For example the fragment
”closed roads mountain passes serve ways escape” produced the 4 highest ranked 4-grams that do
not contain any numerical patterns or special characters, each one of them occurring 153 times.
Most likely, an article that contains this passage has been quoted 152 times.
5
Pruning n-grams Error rate CPU secs. No. Features
set-of-words 47.07 0.92 n.a. 71,731
1 46.18 0.94 12686.12 36,534
DF: 3 2 45.28 0.51 15288.32 113,716
TF: 5 3 45.05 1.22 15253.27 155,184
4 45.18 1.17 14951.17 189,933
1 45.51 0.83 12948.31 22,573
DF: 5 2 45.34 0.68 13280.73 44,893
TF: 10 3 46.11 0.73 12995.66 53,238
4 46.11 0.72 13063.68 59,455
1 45.88 0.89 10627.10 13,805
DF: 10 2 45.53 0.86 13080.32 20,295
TF: 20 3 45.58 0.87 11640.18 22,214
4 45.74 0.62 11505.92 23,565
1 48.23 0.69 10676.43 n.a.
DF: 25 2 48.97 1.15 8870.05 n.a.
TF: 50 3 48.69 1.04 10141.25 n.a.
4 48.36 1.01 10436.58 n.a.
5 48.36 1.01 10462.65 n.a.
1 51.54 0.60 8547.43 n.a.
DF: 50 2 49.71 0.53 8164.27 n.a.
TF: 100 3 51.21 1.26 8079.59 n.a.
4 51.21 1.26 8078.55 n.a.
5 51.21 1.26 8147.75 n.a.
1 52.59 0.71 6609.05 n.a.
DF: 75 2 52.83 0.25 6532.80 n.a.
TF: 150 3 52.36 0.48 6128.49 n.a.
4 52.36 0.48 6128.49 n.a.
5 52.36 0.48 6119.27 n.a.
6
Feature Frequency
ax 62063
D 61603
DD 48247
DDD 31188
19484
DDDD 18331
writes 14684
article 12567
dont 10255
like 10047
representations it seems to be the case that the use of bigrams results in the highest recall and
the lowest precision. In terms of F1 and predictive accuracy, bigrams have a clear advantage at
moderate pruning, while with more heavy pruning, the unigrams representation seems to catch
up.
It is also obvious that precision is correlated to the number of features. Unigrams give higher
precision (but lower recall) than multi-grams, and an increase in the minimum frequency require-
ments also increases precision. For interpreting these results, it should be remembered that this
domain is fairly simple, and for many of the classes the occurrence of a single word is sufficient
to classify many of the articles.
A look at the features is not much different from the results in the 20 newsgroups domain:
the most frequent features seem to bear no obvious relationship to any of the classes. Interesting
is a comparison of the number of features: Although REUTERS contains only slightly more
than 12,000 articles compared to the 20,000 of the 20 newsgroups dataset, the number of found
features differs an order of magnitude. We think that the reasons for this phenomenon are that
newsgroups articles are slightly longer on average, originate from a variety of authors and thus
use a diverse vocabulary, the diversity of the topics of the newsgroups, and the repetitiveness
of newsgroups articles which produces many n-grams by repetition of entire paragraphs of an
article.
However, both, tables 1 and 3, exhibit a sub-linear growth of the number of features. Thus,
the algorithm effectively avoids the super-linear growth of the number of features (see section 2).
4 Related Work
Feature generation and feature selection are important topics in information retrieval. Lewis
(1992c) has emphasized their importance and studied several techniques on the REUTERS
newswire data. Contrary to our results with n-gram features (in particular bigrams), Lewis
(1992a) concludes that in the REUTERS dataset phrasal features (as well as term clustering)
7
Pruning n-grams Recall Precision F1 Accuracy No. Features
set-of-words 76.71 83.42 79.92 99.5140 n.a.
1 77.22 83.55 80.26 99.5211 9,673
DF: 3 2 80.34 82.03 81.18 99.5302 28,045
TF: 5 3 77.56 82.74 80.07 99.5130 38,646
4 78.18 82.31 80.19 99.5130 45,876
1 77.19 83.65 80.29 99.5221 6,332
DF: 5 2 80.05 82.06 81.04 99.5278 13,598
TF: 10 3 77.96 82.29 80.07 99.5106 17,708
4 78.21 82.13 80.12 99.5106 20,468
1 76.92 83.99 80.30 99.5241 4,068
DF: 10 2 79.06 82.04 80.52 99.5177 7,067
TF: 20 3 77.32 82.67 79.91 99.5096 8,759
4 76.98 82.91 79.84 99.5096 9,907
5 Discussion
We presented a simple but efficient algorithm for generating n-gram features and investigated
their utility in two benchmark domains. The algorithm is based on the AP RIORI-algorithm for
discovering frequent item subsets in databases. A similar adaptation of the algorithm has been
independently developed and studied by Mladenić and Grobelnik (1998). In both studies, the
results seem to indicate that the addition of n-grams to the set-of-words representation frequently
used by text categorization systems improves performance. However, sequences of length n > 3
are not useful and may decrease the performance.
Note that the results in this paper were obtained using a simple frequency-based feature
subset selection. Although there is some evidence that frequency based pruning of feature sets
is quite competitive in text categorization domains (Yang and Pedersen 1997; Mladenić 1998),
it might be worth-while to study the use of more sophisticated pruning techniques that take the
8
class information into account. On the other hand, Yang and Pedersen (1997) and Lewis (1992b)
report that heavy pruning may improve performance, which is not consistent with our results.
The main reason for our choice of frequency-based pruning was that it can be easily integrated
into the AP RIORI-based feature generation algorithm. In principle, however, any other feature
subset selection technique could be used as a post-processor to the algorithm. Furthermore, some
techniques could be directly integrated into the algorithm. The only condition that the algorithm
imposes is that if a feature is acceptable to the pruning criterion, all its subsequences have to be
acceptable as well. For some measures that do not implement this condition, upper and/or lower
bounds on the measures could be implemented that allow to weed out unpromising candidates
(such as, e.g., the techniques that are used for pruning candidate conditions with unpromising
information gain bounds in C4.5 (Quinlan 1993) and F OIL (Quinlan 1990)). Extending the
feature generation techniques used in this paper into that direction is subject to further research.
Acknowledgements
This work was performed during the author’s stay at Carnegie Mellon University, which was funded by the
Austrian Fonds zur Förderung der Wissenschaftlichen Forschung (FWF) under grant number J1443-INF
(Schrödinger-Stipendium).
References
AGRAWAL , R., H. MANNILA , R. S RIKANT, H. TOIVONEN , & A. I. V ERKAMO (1995). Fast discov-
ery of association rules. In U. Fayyad, G. Piatetsky-Shapiro, P. Smyth, and R. Uthurusamy (Eds.),
Advances in Knowledge Discovery and Data Mining, pp. 307–328. AAAI Press.
C OHEN , W. W. (1995). Fast effective rule induction. In A. Prieditis and S. Russell (Eds.), Proceedings
of the 12th International Conference on Machine Learning (ML-95), Lake Tahoe, CA, pp. 115–123.
Morgan Kaufmann.
C OHEN , W. W. (1996). Learning trees and rules with set-valued features. In Proceedings of the 13th
National Conference on Artificial Intelligene (AAAI-96), pp. 709–716. AAAI Press.
F ÜRNKRANZ, J. (1997). Pruning algorithms for rule learning. Machine Learning 27(2), 139–171.
F ÜRNKRANZ, J., T. M ITCHELL , & E. R ILOFF (1998). A case study in using linguistic phrases for text
categorization on the WWW. In M. Sahami (Ed.), Learning for Text Categorization: Proceedings
of the 1998 AAAI/ICML Workshop, Madison, WI, pp. 5–12. AAAI Press. Technical Report WS-
98-05.
F ÜRNKRANZ, J. & G. W IDMER (1994). Incremental Reduced Error Pruning. In W. Cohen and
H. Hirsh (Eds.), Proceedings of the 11th International Conference on Machine Learning (ML-94),
New Brunswick, NJ, pp. 70–77. Morgan Kaufmann.
H EARST, M. A. & H. H IRSH (Eds.) (1996). Proceedings of the AAAI Spring Symposium on Machine
Learning in Information Access. AAAI Press. Technical Report SS-96-05.
L EWIS , D. D. (1992a). An evaluation of phrasal and clustered representations on a text categorization
task. In Proceedings of the 15th Annual International ACM SIGIR Conference on Research and
Devlopment in Information Retrieval, pp. 37–50.
9
L EWIS , D. D. (1992b). Feature selection and feature extraction for text categorization. In Proceedings
of a Workshop on Speech and Natural Language, Harriman, NY, pp. 212–217.
L EWIS , D. D. (1992c, February). Representation and Learning in Information Retrieval. Ph. D. thesis,
University of Massachusetts, Department of Computer and Information Science, MA.
L EWIS , D. D. (1997, September). Reuters-21578 text categorization test collection. README file (V
1.2), available from http://www.research.att.com/~lewis/reuters21578/README.txt.
M C C ALLUM , A. & K. NIGAM (1998). A comparison of event models for naive bayes text classifica-
tion. In M. Sahami (Ed.), Learning for Text Categorization: Proceedings of the 1998 AAAI/ICML
Workshop, Madison, WI, pp. 41–48. AAAI Press.
M LADENI Ć , D. (1998). Feature subset selection in text-learning. In Proceedings of the 10th European
Conference on Machine Learning (ECML-98). Springer-Verlag.
M LADENI Ć , D. & M. G ROBELNIK (1998). Word sequences as features in text learning. In Proceedings
of the 17th Electrotechnical and Computer Science Conference (ERK-98), Ljubljana, Slovenia.
IEEE section.
Q UINLAN , J. R. (1990). Learning logical definitions from relations. Machine Learning 5, 239–266.
Q UINLAN , J. R. (1993). C4.5: Programs for Machine Learning. San Mateo, CA: Morgan Kaufmann.
R ILOFF , E. (1995). Little words can make a big difference for text classification. In Proceedings of the
18th Annual International ACM SIGIR Conference on Research and Development in Information
Retrieval, pp. 130–136.
S AHAMI, M. (Ed.) (1998). Learning for Text Categorization: Proceedings of the 1998 AAAI/ICML
Workshop. AAAI Press. Technical Report WS-98-05.
YANG , Y. & J. O. P EDERSEN (1997). A comparative study on feature selection in text categorization.
In D. Fisher (Ed.), Proceedings of the 14th International Conference on Machine Learning (ICML-
97), pp. 412–420. Morgan Kaufmann.
10