=Paper=
{{Paper
|id=Vol-1748/paper-06
|storemode=property
|title=A Comparison between Preprocessing Techniques for Sentiment Analysis in Twitter
|pdfUrl=https://ceur-ws.org/Vol-1748/paper-06.pdf
|volume=Vol-1748
|dblpUrl=https://dblp.org/rec/conf/kdweb/AngianiFFFIMM16
}}
==A Comparison between Preprocessing Techniques for Sentiment Analysis in Twitter==
https://ceur-ws.org/Vol-1748/paper-06.pdf
A Comparison between Preprocessing
Techniques for Sentiment Analysis in Twitter
Giulio Angiani, Laura Ferrari, Tomaso Fontanini, Paolo Fornacciari, Eleonora
Iotti, Federico Magliani, and Stefano Manicardi
Dipartimento di Ingegneria dell’Informazione
Università degli Studi di Parma
Parco Area delle Scienze 181/A, 43124 Parma, Italy
{giulio.angiani,laura.ferrari,tomaso.fontanini,paolo.fornacciari,
eleonora.iotti,federico.magliani,stefano.manicardi}@studenti.unipr.it
Abstract. In recent years, Sentiment Analysis has become one of the
most interesting topics in AI research due to its promising commercial
benefits. An important step in a Sentiment Analysis system for text min-
ing is the preprocessing phase, but it is often underestimated and not
extensively covered in literature. In this work, our aim is to highlight the
importance of preprocessing techniques and show how they can improve
system accuracy. In particular, some different preprocessing methods are
presented and the accuracy of each of them is compared with the oth-
ers. The purpose of this comparison is to evaluate which techniques are
effective. In this paper, we also present the reasons why the accuracy
improves, by means of a precise analysis of each method.
Keywords: Sentiment Analysis, Preprocessing, Naive-Bayes Multinomial.
1 Introduction
The subjective analysis of a text is the main task of Sentiment Analysis (SA),
also called Opinion Mining. One of the basic tasks in SA is to predict the polar-
ity of a given sentence, to find out if it expresses a positive, negative or neutral
feeling about a certain topic [16]. Furthermore, in recent research works, SA goes
beyond the concept of polarity, trying to identify the emotional status of a sen-
tence, such as anger, sadness, happiness, etc., according to various classifications
of affective knowledge [4, 6, 13, 17, 21]. The application of SA ranges over several
domains, from movie reviews to social networks, which also are proliferating in
both usage and architectures [9,10]. The demand for new techniques of SA is con-
tinuously growing, due to their inherent capacity of automatic evaluation, from
both the academic and industrial points of view. In the last few years, Opinion
Mining has become a popular research field, which brings together several differ-
ent areas. Due to its heterogeneity, many different techniques were analyzed and
implemented, in order to get increasingly accurate systems for a certain problem
statement. Most of such techniques involve the use of Machine Learning (ML)
classification algorithms—in particular Supervised Learning Algorithms—, i.e.,
methods that are used to train a classifier, whose aim is the association of an
input with its related class, chosen from a certain set of classes. The training is
done by providing the classifier with several examples of inputs and their related
classes. Then, the system extracts a set of features (or attributes) from each of
them, in order to become capable of recognizing the class of generic data, which
can be of different types [14]. The performance of a classifier could be evaluated
by different metrics, such as the accuracy, which is a measure of the correctness
of a classifier. Similarly, Weka provides the confusion matrix of a simulation,
which is useful for the identification of the errors in the model classification. As
detailed in Figure 1, these methods often include some preprocessing corpora,
which make assumptions and choices on the inclusion of features in text repre-
sentations, and that are used for training a classifier and also for evaluating its
performance. In fact, Machine Learning algorithms need to work on data, appro-
priately processed by a set of operations which make assumptions and choices
on the inclusion of features in text representations. This phase is a fundamental
step in order for the whole system to obtain good results. Normally it includes
methods for data cleaning and feature extraction and selection. A good overview
of the steps and the most known algorithms for each step is explained in [12].
Thus, given a corpus of raw data sets, the first step of SA is the preprocessing
of those data. Preprocessing involves a series of techniques which should improve
the next phases of elaboration, in order to achieve better performances.
As illustrated in [11], online texts usually contain lots of noise and uninfor-
mative parts, such as HTML tags. This raises the dimensionality of the problem
and makes the classification process more difficult. The algorithms which are
most used to polish and prepare data that comes from Twitter messages include
the removal of punctuation and symbols, tokenization, stemming, and stopword
as showed, for example in [5] and [18].
Some of these techniques are exposed in the work of A. Balahur [2, 3], which
concerns the problem of classification of Twitter posts, that is, short sentences
which refer to one topic. She uses a series of interesting preprocessing modules
(such as emoticon replacement, tokenization, punctuation marks, word normal-
ization, etc.) and she shows these methods in detail. However, such methods are
collected together before data classification, and the emphasis of her work is not
on why or how each of these modules helps improve the classifier accuracy. In
fact, her work focuses on the classification of many types of sentiments, from
positive, negative and neutral, to anger, disgust, fear, joy, sadness and surprise,
rather than on the effectiveness of the presented preprocessing techniques. In
our research, we have also collected data sets from Twitter and we have imple-
mented some of Balahur’s preprocessing ideas, finding them useful when applied
to such a problem statement. However, our work focuses on their effectiveness
and their performance in terms of accuracy, and such techniques are evaluated
by analysing each one separately.
The work of A. Agarwal et al. [1], also based on Twitter data sets, proposes
the use of emoticons as features and uses a dictionary of 8000 words associ-
ated with a pleasantness score from 1 (negative) to 3 (positive). Emoticons are
divided into five categories (extremely-positive, positive, neutral, negative and
extremely-negative) and they gain a score, like other words. Then, all scores are
added up per sentence and divided by 3. If the result is less than 0.5, then the
sentence is classified as negative. If, on the contrary, it is greater than 0.8, then
the sentence belongs to the positive class. In all other cases, a neutral class is
used. Basic cleaner, slang conversion and negation replacement are also used. In
particular, we analyse emoticon replacement and other techniques, but we avoid
giving a score to each word, leaving the task of assigning weights to features to
the ML algorithm.
In the context of SemEval (Task 4)1 , for SA in Twitter, N. F. Silva et al. [20]
analyse how much the accuracy of classification changes, using various algo-
rithms: Naive-Bayes Multinomial (NBM), Support Vector Machine (SVM), Ad-
aBoost with SVM, and AdaBoost with NBM. In this paper, we focus on prepro-
cessing methods for a fixed classification algorithm: in fact, the only one utilized
is NBM.
In [7] and [8] Twitter was analyzed as a communication medium in which it
is possible to recognize certain features that identify a sort of Twitter culture.
The techniques in the polish phase were chosen while taking into account this
peculiar nature of tweets.
The main goal of the present work is to analyse and compare different pre-
processing steps found in literature, and to actually define the best combination
of the considered methods. In fact, preprocessing is often seen as a fundamental
step for SA, but rarely is it carefully evaluated, thus leaving the open question
of why and to what extent does it increase the accuracy of the classifier.
The data set used in this work is the one provided by SemEval [19] for
Sentiment Analysis in Twitter, which is often used in many other works, such
as [2, 15, 20], in order to make our results comparable with the others. The tool
used to test the accuracy of the classification is Weka2 .
The paper is structured as follows. In Section 2, a brief introduction to Ma-
chine Learning steps is provided. In addition, the section describes the techniques
and algorithms used for classification and features selection. In Section 3, the
considered preprocessing techniques are presented in detail. Section 4 shows the
performance of the obtained classifier for each preprocessing method, on two
different data sets, and discusses the effectiveness of each of such techniques. In
Section 5, a brief discussion of the proposed work and an analysis of the achieved
results conclude the paper.
1
http://alt.qcri.org/semeval2016/task4/
2
http://www.cs.waikato.ac.nz/ml/weka
2 Algorithms and Techniques
This section describes the different preprocessing modules that have been used
in this work3 . All of them are built in Python4 and they work with version 2.7.
The pipeline of the project is organized in the following way. Firstly, we ob-
tain the 2015 and 2016 data sets (both training and test) of Twitter Sentiment
Analysis from SemEval. The training sets are subjected to the various prepro-
cessing techniques analysed in this work. After the text of each instance of a set
has been preprocessed, the resulting sentences (the cleaned tweets) become the
instances of a new training set. Then, such a data set is used for training a classi-
fier and the corresponding test set is classified by Weka. Finally, the accuracies of
the classifiers obtained from different preprocessing modules are compared with
each other, in order to evaluate the efficiency and effectiveness of each technique.
Fig. 1. Steps for training a classifier for sentiment analysis. Firstly, data have to be
prepared in order to obtain a data set – namely, the training set – by means of prepro-
cessing and feature selection methods. Then, such a data set is involved in the learning
step, which uses ML algorithms and yields a trained classifier. Finally, the classifier
has to be tested on a different data set – namely, the test set.
The classifier is made by using Naive-Bayes Multinomial (NBM) method,
i.e., a ML algorithm that gives rise to a probabilistic classifier, which works
on the basis of the Bayes Theorem, with the strong assumption that features
are mutually independent. Let X = (x1 , . . . , xn ) be the feature vector of an
instance in the data set, that is, a binary vector that takes into account the
presence of a feature in that instance, and let C1 , . . . , CK be the possible outputs
(classes). The problem is to gain the posterior probability of having the class
Ck as output, given the feature vector X, and given the prior probability p(Ck )
for each class. Thanks to the Bayes Theorem and the independence between
features, the probability that needs to be estimated is the conditional p(X|Ck ),
and then a classifier is trained with a decision rule, such as the Maximum a
3
https://github.com/fmaglia/SA_cleaners
4
http://www.python.org
Posteriori (MAP) rule. In summary, the probabilistic model of NBM can be
expressed in terms of the following formula:
P
( i xi )! Y xi
p(X|Ck ) = Q pki
i xi ! i
where X = (x1 , . . . , xn ) is the feature vector, pi is the probability that the
feature i appears, Ck is a class and pki is the probability that feature i occurs
in the class Ck . Then, Information Gain (IG) is the algorithm used for feature
selection [REF]. It evaluates the presence or absence of a feature in a document
by measuring its probability of belonging to a class. The amount of information
needed to exactly classify an instance D is defined recursively as follows.
v
X |Dj |
Info A (D) = − · Info(Dj )
j=1
|Di |
when the instance D is divided by some feature attribute A = {a1 , . . . , av } into
sub-instances D1 , . . . , Dv .
3 Preprocessing Phases
3.1 Basic Operation and Cleaning
This first module manages basic cleaning operations, which consist in removing
unimportant or disturbing elements for the next phases of analysis and in the
normalization of some misspelled words. In order to provide only significant
information, in general a clean tweet should not contain URLs, hashtags (i.e.
#happy) or mentions (i.e. @BarackObama). Furthermore, tabs and line breaks
should be replaced with a blank and quotation marks with apexes. This is useful
in order to obtain a correct elaboration by Weka (i.e. not closing a quotation
mark causes a wrong read by the data mining software causing a fatal error
in the elaboration). After this step, all the punctuation is removed, except for
apexes, because they are part of grammar constructs such as the genitive.
The next operation is to remove the vowels repeated in sequence at least three
times, because by doing so the words are normalized: for example, two words
written in a different way (i.e. cooooool and cool ) will become equals. Another
substitution is executed on the laughs, which are normally sequences of “a” and
“h”. These are replaced with a “laugh” tag.
The last step is to convert many types of emoticons into tags that express
their sentiment (i.e. :) → smile happy). The list of emoticons is taken from
Wikipedia5 .
Finally, all the text is converted to lower case, and extra blank spaces are
removed.
5
http://en.wikipedia.org/wiki/List\_of\_emoticons
All the operations in this module are executed to try to make the text uni-
form. This is important because during the classification process, features are
chosen only when they exceed a certain frequency in the data set. Therefore,
after the basic preprocessing operations, having different words written in the
same way helps the classification.
3.2 Emoticon
This module reduces the number of emoticons to only two categories: smile positive
and smile negative, as shown in Table 1.
Table 1. List of substituted Emoticons
smile positive smile negative
0:-) >:(
:) ;(
:D >:)
:* D:<
:o :(
:P :|
;) >:/
Table 2. Likelihood of some Emoticon
Features P (X|Cpos ) P (X|Cneg )
:) 0.005107 0.000296
:( 0.000084 0.001653
:* 0.001055 0.000084
;) 0.000970 0.000084
Table 3. Likelihood of Smile Positive and Smile Negative
Features P (X|Cpos ) P (X|Cneg )
smile positive 0.007320 0.000718
smile negative 0.000336 0.002283
This is done to increase the weight of these features in the classification phase
and to reduce the complexity of the model. In Table 2 and Table 3, it is possible
to notice how much the likelihood of a feature changes.
3.3 Negation
Dealing with negations (like “not good”) is a critical step in Sentiment Analysis.
A negation word can influence the tone of all the words around it, and ignoring
negations is one of the main causes of misclassification.
In this phase, all negative constructs (can’t, don’t, isn’t, never etc) are re-
placed with “not”.
This technique allows the classifier model to be enriched with a lot of negation
bigram constructs that would otherwise be excluded due to their low frequency.
Table 4 shows some examples of extracted features and their likelihood.
Table 4. Example of Features Extracted
Features p(X|Cpos ) p(X|Cneg )
not wait 0.002345 0.000304
not miss 0.000651 0.000043
not like 0.000004 0.000391
3.4 Dictionary
This module uses the external python library PyEnchant6 , which provides a
set of functions for the detection and correction of misspelled words using a
dictionary.
As an extension, this module allows us to substitute slang with its formal
meaning (i.e., l8 → late), using a list. It also allows us to replace insults with
the tag “bad word”.
The motivation for the use of these functions is the same as for the basic
preprocessing operation, i.e., to reduce the noise in text and improve the overall
classification performances.
3.5 Stemming
Stemming techniques put word variations like “great”, “greatly”, “greatest”,
and “greater” all into one bucket, effectively decreasing entropy and increasing
the relevance of the concept of “great”. In other words, Stemming allows us to
consider in the same way nouns, verbs and adverbs that have the same radix.
6
http://pythonhosted.org/pyenchant
This method is already implemented in Weka and the algorithm in use is
IteratedLovinsStemmer 7 .
As in the case of emoticons, with the use of this technique it is possible to
combine features with the same meaning and reduce the entropy of the model.
3.6 Stopwords
Stop words are words which are filtered out in the preprocessing step. These
words are, for example, pronouns, articles, etc. It is important to avoid having
these words within the classifier model, because they can lead to a less accurate
classification.
4 Results
The data set is composed of the training and test sets.
Table 5. Data set table
Data set Positive Negative Total
Training set 1339 1339 2678
Test set 2016 623 169 792
Test set 2015 343 174 517
The training sets are those provided by SemEval, with a little revision: neu-
tral sentences are removed, in order to focus only on positive and negative ones.
Furthermore, in the training set there are more positive sentences than negative
ones. Excess positive ones have been eliminated, because they distort the Bayes
model.
In the executed tests, the features collected have a minimum presence in the
text that is greater than or equal to 5. The Ngrams used are only one-grams
and bi-grams. Before starting the simulation with the test set, a 10-fold cross-
validation is carried out. In particular, we searched for the optimal length of
N-grams to potentially consider as features. In Figure 2, it can be observed that
accuracy nearly peaks at N-gram = 2. Longer sequences increase the complexity
of the training phase, without giving a significant improvement of the result.
Also, we analysed the total number of features to consider. This parameter does
not provide a monotonic improvement to the classifier quality. Instead, it peaks
out at around 1500 features.
At first, the executed simulations compare no preprocessed file vs. basic
cleaned file. As shown in Table 5, the resulting accuracy is strongly increased.
Given the importance of the basic cleaner, we decided to use it in every case,
7
weka.sourceforge.net/doc.dev/weka/core/stemmers/IteratedLovinsStemmer.
html
Fig. 2. Optimization of system parameters.
together with another preprocessing module, in order to evaluate their contri-
bution together.
Stemming increases the performance, because it groups words reduced to
their root form. It allows many words to be selected as useful features for the
classification phase. In fact, it modifies the weight of a feature, usually increasing
it.
Stopword removal enhances the system because it removes words which are
useless for the classification phase. As a common example, an article does not
express a sentiment but it is very present in the sentences.
Table 6. Result classification table
Technique # Max Features CV folds 10 [%] Test 2016 [%] Test 2015 [%]
No preprocess 1800 78,08 65,65 69,05
Basic 2000 80,05 65,40 74,08
Basic + Stemming 2200 80,84 68,68 76,40
Basic + Stopwords 1800 80,32 65,27 74,85
Basic + Negation 2000 80,40 65,65 75,04
Basic + Emoticon 2000 80,13 65,98 74,66
Basic + Dictionary 2000 78,00 64,39 75,82
All 2000 80,40 64,89 75,82
All Without Dictionary 2100 80,76 65,78 75,04
As a notable result, it is interesting that using a dictionary did not enhance
the performance in our tests, but it increased the elaboration-time needed for
cleaning raw data.
There is also an improvement in the accuracy of the classifier between the
two SemEval test-sets: 2016 and 2015. However, this is only due to there being
fewer of sentences in the last test-set, with a corresponding lower probability for
the classifier to make mistakes.
5 Conclusions
Text preprocessing is an important phase in all relevant applications of data
mining. In Sentiment Analysis, in particular, it is cited in virtually all available
research works. However, few works have been specifically dedicated to under-
standing the role of each one of the basic preprocessing techniques, which are
often applied to textual data.
Arguably, having a more precise measure of the impact of these basic tech-
niques can improve the knowledge of the whole data mining process. This work
adopts a straightforward methodology: it basically applies each one of the most
known filters, independently, to the raw data. However, given the importance of
the basic cleaner, we decided to use the basic cleaner in every case, together with
another single preprocessing module, and then evaluate their joint contribution.
As an interesting result, it is worth noting that using a dictionary did not
enhance the performances in our tests, but it increased the elaboration-time
needed for cleaning raw data. All other techniques, however, provided signifi-
cant improvements to the classifier performances. Some of the techniques simply
removed useless noise in the raw data, while others increased the relevance of
some concepts, reducing similar terms and expression forms to their most basic
meaning.
This research has been conducted over data which originated from Twitter.
In our opinion, a similar analytical work should be performed on different kinds
of data sets, to have a more comprehensive understanding of the different pre-
processing filters. The decision to mix some of these filters together is often
correct. However, it should be better motivated by empirical data and result
evaluations for various application domains and the peculiar nature of their
textual data.
References
1. Agarwal, A., Xie, B., Vovsha, I., Rambow, O., Passoneau, R.: Sentiment Analysis
of Twitter Data. Computer Science - Columbia University (New York, USA) (2011)
2. Balahur, A.: Sentiment Analysis in Social Media Texts. European Commission
Joint Research Center (Varese, Italy) (2013)
3. Bao, Y., Quan, C., Wang, L., Ren, F.: The role of pre-processing in twitter senti-
ment analysis. In: International Conference on Intelligent Computing. pp. 615–624.
Springer (2014)
4. Cambria, E., Olsher, D., Rajagopal, D.: Senticnet 3: a common and common-sense
knowledge base for cognition-driven sentiment analysis. In: Twenty-eighth AAAI
conference on artificial intelligence (2014)
5. Duncan, B., Zhang, Y.: Neural networks for sentiment analysis on twitter. In:
Cognitive Informatics & Cognitive Computing (ICCI* CC), 2015 IEEE 14th In-
ternational Conference on. pp. 275–278. IEEE (2015)
6. Esuli, A., Sebastiani, F.: Sentiwordnet: A publicly available lexical resource for
opinion mining. In: Proceedings of LREC. vol. 6, pp. 417–422. Citeseer (2006)
7. Fornacciari, P., Mordonini, M., Tomaiuolo, M.: A case-study for sentiment analysis
on twitter. In: Proceedings of the 16th Workshop ”From Objects to Agents”-WOA
(2015)
8. Fornacciari, P., Mordonini, M., Tomaiuolo, M.: Social network and sentiment anal-
ysis on twitter: Towards a combined approach. In: Proceedings of the 1st Interna-
tional Workshop on Knowledge Discovery on the WEB –KDWEB 2015 (2015)
9. Franchi, E., Poggi, A., Tomaiuolo, M.: Blogracy: A peer-to-peer social network.
International Journal of Distributed Systems and Technologies (IJDST) 7(2), 37–
56 (2016)
10. Franchi, E., Tomaiuolo, M.: Distributed social platforms for confidentiality and
resilience. Social Network Engineering for Secure Web Data and Services p. 114
(2013)
11. Haddi, E., Liu, X., Shi, Y.: The role of text pre-processing in sentiment analysis.
Procedia Computer Science 17, 26–32 (2013)
12. Kotsiantis, S., Kanellopoulos, D., Pintelas, P.: Data preprocessing for supervised
leaning. International Journal of Computer Science 1(2), 111–117 (2006)
13. Liu, B.: Sentiment analysis: Mining opinions, sentiments, and emotions. Cambridge
University Press (2015)
14. Matrella, G., Parada, G., Mordonini, M., Cagnoni, S.: A video-based fall detector
sensor well suited for a data-fusion approach. Assistive Technology from Adapted
Equipment to Inclusive Environments, Assistive Technology Research Series 25,
327–331 (2009)
15. Neethu, M.S., Rajasree, R.: Sentiment Analysis in Twitter using Machine Learning
Techniques. Department of Computer Science - College of Engineering (Trivan-
drum, India) (2013)
16. Pang, B., Lee, L.: Opinion mining and sentiment analysis. Foundations and trends
in information retrieval 2(1-2), 1–135 (2008)
17. Poria, S., Cambria, E., Winterstein, G., Huang, G.B.: Sentic patterns: Dependency-
based rules for concept-level sentiment analysis. Knowledge-Based Systems 69, 45–
63 (2014)
18. Saif, H., Fernández, M., He, Y., Alani, H.: On stopwords, filtering and data sparsity
for sentiment analysis of twitter (2014)
19. Semeval: Semeval 2016 task 4: Sentiment analysis in twitter.
http://alt.qcri.org/semeval2016/task4 (2016)
20. Silva, N.F.F., Hruschka, E.R., Hruschka, E.R.: Biocom Usp: Tweet Sentiment Anal-
ysis with Adaptive Boosting Ensemble. University of Sao Paulo and Federal Uni-
versity of Sao Carlos (Sao Carlos,Brasil) (2014)
21. Strapparava, C., Valitutti, A., et al.: Wordnet affect: an affective extension of
wordnet. In: LREC. vol. 4, pp. 1083–1086 (2004)