An Approach to Detect Polarity Variation Rules for Sentiment Analysis

Pierluca Sangiorgi

1,2

, Agnese Augello

and Giovanni Pilato

ICAR, Istituto di Calcolo e Reti ad Alte Prestazioni, CNR, Consiglio Nazionale delle Ricerche,

Viale delle Scienze - Ediﬁcio 11, 90128, Palermo, Italy

INAF, Istituto di Astroﬁsica Spaziale e Fisica Cosmica - Palermo, via U. La Malfa 153, 90146, Palermo, Italy

Keywords:

Subjectivity Analysis, Sentiment Analysis, Opinion Mining, Machine Learning.

Abstract:

Sentiment Analysis is a discipline that aims at identifying and extract the subjectivity expressed by authors

of information sources. Sentiment Analysis can be applied at different level of granularity and each of them

still has open issues. In this paper we propose a completely unsupervised approach aimed at inducing a set

of words patterns that change the polarity of subjective terms. This is a very important task because, while

sentiment lexicons are valid tools that can be used to identify the polarity at word level, working at different

level of granularity they are no longer sufﬁcient, because of the various aspects to consider like the context,

the use of negations and so on that can change the polarity of subjective terms.

1 INTRODUCTION

In recent years, with the advent of blogs, forums, so-

cial communities, product rating platforms and so on,

users have become more active in production of large

amount of information. Starting from news as well as

products reviews and any other information systems

that allow on-line user interaction in terms of con-

tents, users provide information through their contri-

butions in the form of opinions, discussions, reviews

and so on.

Companies and organizations are increasingly in-

terested in this type of information because it can be

used as knowledge resource for operations of market

survey, political behavior and, in general, measure-

ment of satisfaction.

Information produced by network users usually

regards what is known as their ”private states”: their

opinions, emotions, sentiments, evaluations and be-

liefs (Quirk et al., 1985) (Banea et al., 2011). The

term subjectivity is usually used in literature as a lin-

guistic expression of private state (Banea et al., 2011).

One of the latest and most challenging research

task is the automatic detection of subjectivity pres-

ence in text, which is named subjectivity analysis.

The task to also identifying, where possible, its po-

larity, by classifying it as neutral, positive or negative

is deﬁned in literature as sentiment analysis.

Sentiment analysis may be realized at several lev-

els of granularity, like word, sentence, phrase or doc-

ument level. Usually each level of analysis exploits

the results obtained by the underlying layers.

What makes these tasks hard to achieve is the am-

biguity of words, the context-sensitivity of subjec-

tive terms, the need of appropriate linguistic resources

for different languages, the presence of negations, the

presence of irony, and so on (Montoyo et al., 2012).

In this paper we propose an approach to auto-

matically discover several sequential patterns (i.e.

a sequence of tokens in the sentence) in a speciﬁc

language that change the polarity of subjective words.

The novelty of the approach is that it is completely

unsupervised, requiring only the use of one single

linguistic resource: a sentiment lexicon, which is

speciﬁc for the language under consideration. This

can be done in order to build a polarity variation

detector to be used in sentiment analysis applications.

2 RELATED WORKS

In sentiment analysis, the comprehension of sen-

tences polarities is a complex task, which involves

the analysis of the composition of the words in the

sentences, considering their prior polarities. Several

sentiment lexicons (Wilson et al., 2005a), (Stone and

Hunt, 1963), (Baccianella et al., 2010), (Strapparava

and Valitutti, 2004), can be accessed in order to ex-

amine the polarity at a “word-level”; in some cases

344

Sangiorgi P., Augello A. and Pilato G..

An Approach to Detect Polarity Variation Rules for Sentiment Analysis.

DOI: 10.5220/0004961903440349

In Proceedings of the 10th International Conference on Web Information Systems and Technologies (WEBIST-2014), pages 344-349

ISBN: 978-989-758-024-6

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

these lexicons can be obtained trough machine learn-

ing approaches (which cluster terms according to their

distributional similarity (Turney and Littman, 2003)),

or by means of bootstrapping methodologies starting

from few term-seeds (Banea et al., 2011)(Pitel and

Grefenstette, 2008).

Polarity of words is highly dependent on the do-

main in which they are used, so that adjectives with

positive polarity could have an opposite or a neutral

polarity in another domain. Moreover, especially in

the context of product reviews users often adopt ab-

breviations and idioms; therefore methods for the au-

tomatic creation of lexicons may be very useful in this

context also to improve existing dictionaries.

At a ”sentence-level” it is necessary consider the

composition of the word, with their prior polarities,

into the phrase. Different compositional models have

been proposed in (Yessenalina and Cardie, 2011),

(Wu et al., 2011) and (Chardon et al., 2013). In

(Hu and Liu, 2004) a set of adjective words (opinions

words) is identiﬁed using a natural language process-

ing method in order to decide the opinion orientation

of a sentence: for each opinion word its semantic ori-

entation is determined, and the opinion orientation is

then predicted by analyzing the predominance of pos-

itive or negative words. In (Moilanen and Pulman,

2007) a composition model based on a syntactic tree

representation has been proposed. Many other factors

have to be considered, especially the presence of po-

larity inﬂuencers (Wilson et al., 2005b) (Polanyi and

Zaenen, 2006). In (Wilson et al., 2005b) a study on

the most important features for recognizing contex-

tual polarity has been performed, and the performance

of these features has been evaluated by using several

different machine learning algorithms. In (Tan et al.,

2012) it has been proposed an automatic approach to

detect polarity pattern rules, based on the extraction

of typed dependency bi-grams of sentences and the

use of Class Sequential Rules (CSRs) to derive po-

larity class rules from sequential patterns within the

bi-grams. In (Tromp and Pechenizkiy, 2013) is pro-

posed an algorithm for polarity detection based on the

use of different heuristic rules.

3 AN APPROACH TO DETECT

POLARITY VARIATION RULES

The proposed approach is thought to work in an unsu-

pervised manner. It can be run just one time in order

to deﬁne the rules. A subsequent re-run of this ap-

proach is not required, excepted for an extension of

the number of rules using other data sets as inputs.

The process can be viewed as a black box that pro-

duces, given a set of documents as input, a set of rules

in a speciﬁc structure that will be described below.

The entire process can be decomposed in a sequence

of steps as reported in ﬁgure 1.

Figure 1: The architecture of the proposed approach.

The main idea is to extract several user reviews

expressed in English language from an on-line mar-

ket store which implements a star rating system. Each

one of the reviews represents a short document, usu-

ally a sentence, that can be classiﬁed as having pos-

itive or negative polarity according to the number of

stars that are compulsorily inserted by the users in or-

der to publish a review.

Given these two classes of documents, after a sim-

ple pre-processing through a tokenizer used for split-

ting their content into words, we scan each word by

exploiting a sentiment lexicon in English language.

If in a document with a given polarity is detected a

word with a different one, the three words to its left

and those to its right are saved in a list.

Repeating this process for all the extracted re-

views, we obtain a collection of words typically sur-

rounding subjective words that, “for some reason”,

are probably responsible for the variation of the po-

larity of the subjective words.

Without concerning why these rules change words

polarity, we can study them in terms of frequency,

support and conﬁdence and select the most promising

ones.

The collection of these sequential patterns will

form a set of polarity variation rules. These rules

could be integrated in a polarity variation detector in

order to check, during other sentiment analysis appli-

cations, if a given sentence matches one of these new

rules and therefore a different polarity for the term

should be considered.

In a nutshell, the approach is based on four main

steps:

AnApproachtoDetectPolarityVariationRulesforSentimentAnalysis

345

1. extraction of a large amount of user reviews from

on-line market store with star rating system;

2. sentiment analysis at word level using sentiment

lexicons;

3. extraction of the left and right context of a subjec-

tive word with inverted polarity respect to that of

the sentence;

4. identiﬁcation of the sequential patterns rules that

change the polarity;

It might seem that an approach like this could be

replaced by a set of rules manually written by pro-

fessional linguists, but this is not completely true.

This because a linguist can deﬁne rules based on the

grammar of the language that not always ﬁts the “lan-

guage” generally used by people on the social media.

Furthermore, this approach can be potentially applied

to any language of interest, since the polarity variation

detection depends only on the tools used.

4 FORMALIZATION OF THE

PROPOSED APPROACH

In our approach three main elements can be identiﬁed:

1. the input resources which consist of on-line user

reviews that we want to analyze;

2. the processing chain;

3. the output which consist of linguistic patterns that

cause the change of polarity of subjective terms.

We deﬁne reviews as input as an ordered sequence of

variable length of words, punctuation and symbols z

= {z

, z

, ..., z

}

and the desired output as two lists P

= {p

} and

−

= {p

−

}, respectively for positive and negative

sentences, of sequential patterns:

= [n

, n

]

composed by an experimentally ﬁxed number of

terms, some of which empty, that represent the

sequence of three terms that are present before (n

, n

) and after (n

, n

) a subjective term, which

can cause its polarity variation.

The details of the entire processing chain will be

described in the next subsections.

4.1 Acquisition and Data Preparation

The ﬁrst step of the processing chain accomplishes

the task of collecting reviews in order to create the

knowledge resource, rich of subjective terms, from

which to extract the polarity variation rules.

To this aim we consider the on-line user reviews

as information sources, this because this type of

text carries out opinions and sentiments expressed by

the users, containing subjectivity terms with an high

probability.

In particular, we are considering to retrieve the re-

views from the Google Play market store, Amazon,

Itunes or other on-line market shop.

This choice is motivated by a fundamental feature

characterizing this kind of stores: they implement a

star rating system, with scores usually from one to

ﬁve, that must be inserted by the users in order to pub-

lish a review. This feature allows us to extract reviews

as classiﬁed in terms of opinion (bad or good) ex-

pression. We make the supposition that reviews with

higher value of stars express positive sentiments of

users while reviews with lower value of stars express

negative sentiments.

The result of this step is to extract several reviews

with their associated stars values for different prod-

ucts and group them by positive or negative expressed

opinion in two large class of documents R

= {r

}

and R

−

= {r

−

}, considering only the stars values.

Once the classes of documents are built, they pass

through a tokenizer chain to split the content of re-

views in separated terms. The splitting must be

done considering space character as well as punc-

tuation, maintaining every remaining alphanumeric

terms without ﬁltering.

No other text pre-processing, like stopwords re-

moval, must be done, this because we are interested

to all terms that compose the reviews and could be de-

terminant in the rules deﬁnition despite they not carry

out particular information. For example think about

terms like ”but”, ”not” and so on.

The result of this process is that every review is

cleaned from spaces, punctuation and symbols. This

lead to a review r

composed only by terms t

= {t

, t

, ..., t

}

4.2 Identiﬁcation of the Variation

Neighborhood Lists

Given the i-th review with positive polarity r

composed by a sequence of terms t

= {t

, t

, ...}

every j-th term t

is checked through a senti-

ment lexicon, and whenever the term is detected as

being subjective with negative polarity, the term is

WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies

346

marked and its neighbor terms are stored in an array

of predetermined size:

= [n

j−3

, n

j−2

, n

j−1

, n

j+1

, n

j+2

, n

j+3

]

where

n = t (if t is a valid terms)

n = ” ” (if t is empty).

We deﬁne as “Variation Neighborhood List” a set

of sequential patterns of ﬁxed size composed by the

terms in the left and in the right context of the sub-

jective terms with discordant polarity, and in this case

is an element of the Variation Neighborhood List

= {N

} relative to the positive reviews.

This operation is executed for both positive and

negative reviews. Generally speaking, we state that

whenever a discordance between the prior polarity of

a subjective term with the polarity of the review that

the term belongs to is detected, a new element of the

Variation Neighborhood List is built.

In order to clarify this statement, let us consider,

for example, three negative reviews after the tok-

enizer process:

−

= {this, app, is, good, when, it, works, but,

ninety, percent, of, the, time, it, will, not, even, open,

on, my, phone}

−

= {my, previous, review, was, good, as, i,

was, satisﬁed, with, the, app, but, now, feel, that, its,

the, worst, app, ever}

−

= {not, good, needs, to, be, ﬁxed}

All the three reviews contain the subjective

term good that have positive prior polarity, which

is however discordant with the overall sentiment

expressed by the reviews.

In this case we build the three arrays:

−

= [this, app, is, when, it, works]

−

= [previous, review, was, as, i, was]

−

= [ , , not, needs, to, be]

Note that the subjective term is discarded and not

stored in the array, this because we are interested to

its context which can affect its polarity as well as the

context of other subjective terms.

The only restriction is to consider only an even

number of terms for each context in order to have al-

ways patterns with the ﬁrst half of terms referring to

the left context of subjective terms and the second half

to the right.

4.3 Rules Extractor

Once the two Neighborhood Lists L

and L

−

respec-

tively for positive and negative reviews are built, they

are singularly processed in order to extract the two

ﬁnal set of polarity variation rules.

In order to reach this goal, the list is expanded

adding all possible combinations of terms of the orig-

inal sequential patterns extracted in the previews step.

Considering the sequential pattern N

−

of the above

example, this means create additional sequential pat-

terns N

2,k

with k ∈ [1, 62], like:

2,32

−

= [previous, , , , , ]

2,16

−

= [ , review, , , , ]

2,8

−

= [ , , was, , , ]

2,39

−

= {previous, , , as, i, was}

and so on.

The k-th sequential pattern is built considering the

binary encoding of k that leave the original terms in

the positions with value equal to 1 and replace the

terms with ” ” in the positions with value equal to 0.

k=0 and k=63 are not considered because they are re-

spectively the empty pattern and the original pattern.

Note that 63 is the maximum number of sequential

patterns for each polarity variation detected, but they

could be less if the original pattern have empty ele-

ments (” ”).

After the expansion process, the next step is to

compare all the sequential patterns of the expanded

list, remove all the duplicates and saving for each pat-

tern its value of frequency.

Since a pattern with few empty elements is strong

and representative of polarity variation, but not prob-

able to occur many times, instead of its frequency we

consider a weight value w associated to each pattern;

w is deﬁned as:

w = ( f -1)*l

where l is the number of non empty elements of

the pattern (l ∈ [1, 6]) and f is the frequency of the

pattern in the list.

In this way patterns that occur only once in the

list have weight w equal to 0 and therefore they are

discarded. Patterns with higher value of l, have an

high conﬁdence, but a very low support.

Considering only the frequency does not work: for

example, let us consider that a pattern N

i,62

with a fre-

quency value of 2, implies the existence of 6 patterns

with the same frequency but l=1 (N

i,1

, N

i,2

, N

i,4

, N

i,8

i,16

, N

i,32

), as well as the other ”sub-patterns” with

all value of l generated from the expansion procedure

of the pattern with l=6.

AnApproachtoDetectPolarityVariationRulesforSentimentAnalysis

347

Once the weight value w is calculated, for each

patterns for the two Neighborhood Lists, using an ap-

propriate value of threshold we remove all the pat-

terns with a value lower than a given threshold .

All the patterns remaining from the cut-off phase

represent the polarity variation rules we are looking

for and stored as elements p

in the two list of rules

= {p

} and P

−

= {p

−

They are expressed as vector of 6 terms, someone

of which may be empty. These vectors, if used in sen-

timent analysis tasks, indicate, with a grade of proba-

bility, which terms must be present before and/or after

a subjective term to cause its polarity variation.

5 CONCLUSIONS AND FUTURE

WORK

We think that the work proposed in this paper could

be a plausible approach to determine and solve the

polarity variation problem in sentiment analysis ap-

plications. We are thus interested to develop this ap-

proach and we are working in this direction consider-

ing Google Play market store as resource for reviews

and SentiWordNet, a lexical resource for opinion min-

ing created by manual assignment of polarity to each

synset of Wordnet, as sentiment checker. This will let

us to deﬁne a set of valid rules to be integrated in a

computer aided system for real time detection of po-

larity variation as support for other systems.

The choice to work in this direction, in fact, is not

casual and arises from a real necessity occurred dur-

ing other sentiment analysis works were we are in-

volved. Just for example, during a sentiment anal-

ysis process done at word level on a large collec-

tion of Google Play market reviews, we have noticed

that SentiWordNet found more positive than negative

words inside reviews classiﬁed as negative with the

star rating system. Find any discordant polarity terms

inside a sentence it’s a typical situation, but not in the

numbers we found in our collection. This prompted

us to investigate and deﬁne a valid approach to solve

the problem.

ACKNOWLEDGEMENTS

This work has been partially supported by the POR

2007/2013 - Regione Siciliana - Misura 4.1.1.1. - IDS

(Innovative Document Sharing) Research Project and

by the PON01 01687 - SINTESYS (Security and IN-

TElligence SYSstem) Research Project.

REFERENCES

Baccianella, S., Esuli, A., and Sebastiani, F. (2010). Senti-

wordnet 3.0: An enhanced lexical resource for sen-

timent analysis and opinion mining. In Chair), N.

C. C., Choukri, K., Maegaard, B., Mariani, J., Odijk,

J., Piperidis, S., Rosner, M., and Tapias, D., editors,

Proceedings of the Seventh International Conference

on Language Resources and Evaluation (LREC’10),

Valletta, Malta. European Language Resources Asso-

ciation (ELRA).

Banea, C., Mihalcea, R., and Wiebe, J. (2011). Multilingual

Sentiment and Subjectivity Analysis. ACL 2012.

Chardon, B., Benamara, F., Mathieu, Y., Popescu, V., and

Asher, N. (2013). Sentiment composition using a

parabolic model. In Proceedings of the 10th In-

ternational Conference on Computational Semantics

(IWCS 2013) – Long Papers, pages 47–58, Potsdam,

Germany. Association for Computational Linguistics.

Hu, M. and Liu, B. (2004). Mining and summariz-

ing customer reviews. In Proceedings of the Tenth

ACM SIGKDD International Conference on Knowl-

edge Discovery and Data Mining, KDD ’04, pages

168–177, New York, NY, USA. ACM.

Moilanen, K. and Pulman, S. (2007). Sentiment composi-

tion. In Proceedings of Recent Advances in Natural

Language Processing (RANLP 2007), pages 378–382.

Montoyo, A., Mart

ınez-Barco, P., and Balahur, A. (2012).

Subjectivity and sentiment analysis: An overview of

the current state of the area and envisaged develop-

ments. Decision Support Systems, 53(4):675 – 679.

Pitel, G. and Grefenstette, G. (2008). Semi automatic build-

ing method for a multidimensional affect dictionary

for a new language. In LREC. European Language

Resources Association.

Polanyi, L. and Zaenen, A. (2006). Contextual valence

shifters. In Shanahan, J., Qu, Y., and Wiebe, J., ed-

itors, Computing Attitude and Affect in Text: Theory

and Applications, volume 20 of The Information Re-

trieval Series, pages 1–10. Springer Netherlands.

Quirk, R., Greenbaum, S., Leech, G., and Svartvik, J.

(1985). A comprehensive grammar of the English lan-

guage. Longman, London.

Stone, P. J. and Hunt, E. B. (1963). A computer ap-

proach to content analysis: studies using the general

inquirer system. In Proceedings of the May 21-23,

1963, spring joint computer conference, AFIPS ’63

(Spring), pages 241–256, New York, NY, USA. ACM.

Strapparava, C. and Valitutti, A. (2004). Wordnet affect:

an affective extension of wordnet. In In Proceedings

of the 4th International Conference on Language Re-

sources and Evaluation, pages 1083–1086.

Tan, L.-W., Na, J.-C., Theng, Y.-L., and Chang, K. (2012).

Phrase-level sentiment polarity classiﬁcation using

rule-based typed dependencies and additional com-

plex phrases consideration. Journal of Computer Sci-

ence and Technology, 27(3):650–666.

Tromp, E. and Pechenizkiy, M. (2013). Rbem: A rule based

approach to polarity detection. In Proceedings of the

Second International Workshop on Issues of Sentiment

Discovery and Opinion Mining, WISDOM ’13, pages

8:1–8:9, New York, NY, USA. ACM.

WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies

348

Turney, P. D. and Littman, M. L. (2003). Measuring praise

and criticism: Inference of semantic orientation from

association. ACM Trans. Inf. Syst., 21(4):315–346.

Wilson, T., Hoffmann, P., Somasundaran, S., Kessler, J.,

Wiebe, J., Choi, Y., Cardie, C., Riloff, E., and Pat-

wardhan, S. (2005a). Opinionﬁnder: A system for

subjectivity analysis. In HLT/EMNLP. The Associa-

tion for Computational Linguistics.

Wilson, T., Wiebe, J., and Hoffmann, P. (2005b). Rec-

ognizing contextual polarity in phrase-level sentiment

analysis. In Proceedings of the Conference on Hu-

man Language Technology and Empirical Methods in

Natural Language Processing, HLT ’05, pages 347–

354, Stroudsburg, PA, USA. Association for Compu-

tational Linguistics.

Wu, Y., Zhang, Q., Huang, X., and Wu, L. (2011). Struc-

tural opinion mining for graph-based sentiment rep-

resentation. In Proceedings of the Conference on

Empirical Methods in Natural Language Process-

ing, EMNLP ’11, pages 1332–1341, Stroudsburg, PA,

USA. Association for Computational Linguistics.

Yessenalina, A. and Cardie, C. (2011). Compositional

matrix-space models for sentiment analysis. In Pro-

ceedings of the Conference on Empirical Methods in

Natural Language Processing, EMNLP ’11, pages

172–182, Stroudsburg, PA, USA. Association for

Computational Linguistics.

AnApproachtoDetectPolarityVariationRulesforSentimentAnalysis

349