Sentiment Analysis in Brazilian Portuguese Tweets in the Domain of

Calamity: Application of the Summarization Method and Semantic

Similarity in Polarized Terms

Ariana Moura da Silva

, Rodrigo da Matta Bastos

and Ricardo Luis de Azevedo da Rocha

Computer Engineering Department, University of Sao Paulo, Av. Prof. Luciano Gualberto, tv. 3, 158, São Paulo, Brazil

Languages and Adaptive Techniques Laboratory, University of Sao Paulo, São Paulo, Brazil

Keywords: Sentiment Analysis, Natural Language Processing, Polarization, Summarization, Latent Semantic Analysis.

Abstract: This research integrates an interdisciplinary project which mobilizes the areas of Computer Engineering,

Linguistics and Communication to perform the processing of texts in a natural language extracted from

microblogging service Twitter as well as to conduct an analysis and classification of the sentiments mined.

Many proposals have been formulated using the polarization method; however, most projects do not

encompass an automatic classification by semantic proximity. This research aims to evaluate the reaction of

individuals shared in the social network, not only to classify them as positive or negative, but also to

ascertain the semantic similarity of these messages in the same domain. Based on the set of tweets in

Portuguese extracted from a corpus of calamity, we apply three methods: a) the lexical classifier, called

Summarization Method; b) the semantic classifier, called LSA - Latent Semantic Analysis; c) the ASSTPS

classifier - Analysis of Semantic similarity in Polarized and Summarized terms. The results are applied to a

set of 811 tweets of the calamity domain and point out which method obtained the best hit rate and semantic

approximation. In this sense, the classification of sentiments by semantic proximity can help greatly,

performing the sorting of content of relevant messages, discarding unnecessary information, linking

messages with the same theme in common, and even generating Metrics for classifying emotions.

1 INTRODUCTION

The insertion of the World Wide Web proposed by

Berners-Lee in 1990 provided the sharing of

information through the first web browser (Berners-

Lee, 2012). This feature provided the sharing of

documents published on the Internet. Later, there

was the availability of these contents in what is

called the WEB 1.0, considered static. Then Web 2.0

came, and it was considered dynamic, by allowing

end user interaction with the structure and content of

the page, inserting comments, sending photos and

sharing thousands of files. With the evolution of

Web 1.0 to Web 2.0, the user went on to perform

lower download rates and increased the upload rates.

With the increase in the collaboration of users in

Web 2.0, the need to search for such content placed

on the internet rose. Web 3.0 was then established to

facilitate information mining and to create a more

effective language exchange between man and

machine, also called the Semantic Web (Semantic

Web, 2018).

The information and opinions entered by the

users generated interest in the industry, by

companies and financial institutions, among others,

in interpreting the opinion about a product, a subject,

a theme. However, reducing the large mass of data

available on the WEB is a great problem, which has

stirred a new area of computer science studies called

sentiment analysis, which has gained prominence as

from 2000 (Liu, 2012). To conduct research on

sentiments analysis, it is necessary to use Natural

Language Processing - NLP techniques.

Human language is not merely the manifestation

of a physical action by the human being. Words are

like symbols, their meanings indicate an idea or a

thing. Language symbols can be encoded in voice,

gesture, writing and others. NLP has different levels,

from speech processing to semantic interpretation

and discourse processing. NLP aims to design, to

build algorithms capable of helping the machine

Silva, A., Bastos, R. and Rocha, R.

Sentiment Analysis in Brazilian Portuguese Tweets in the Domain of Calamity: Application of the Summarization Method and Semantic Similarity in Polarized Terms.

DOI: 10.5220/0006947802250231

In Proceedings of the 10th International Joint Conference on Computational Intelligence (IJCCI 2018), pages 225-231

ISBN: 978-989-758-327-8

225

understand natural human language (Chaubard et al.,

2018).

An emotion is not simply a state of sentiment

(Plutchik, 2001). Sentiment analysis at the word

level verifies the polarity of this word specifically.

At the sentence level, not only the polarity of words

is considered, but also the relationship between these

words and their grammatical usage. At the document

level, it considers the full context of the document,

leading to a more complex analysis of how the

phrases interact with one another (Liu, 2012).

According to data obtained from the Internet

World Stats (Internet World Stats, 2018), Brazil is

the fourth country in the world ranking of the

number of users accessing the internet until

December 2017. In the third position comes the

USA, in the second position, India, and in the first

position, China. It also tells us that the Portuguese

language is in the fifth position of the most widely

used language on the Internet, following Arabic,

Spanish, Chinese and English in the first position

(Internet World Stats, 2018). Portuguese is the fifth

most used language on Twitter (The Statistics Portal,

2018), which currently has 330 million of active

users in the world (The Statistics Portal; Twitter

NIC, 2018).

With the exponential growth of social media,

users are no longer based on the opinion of people

close to them. With a Web available and easy to

access, users search for and find opinions on the

most diverse subjects from people around the world.

The same is true when you want to find a product

and think about issues from any part of the world.

Thinking about structuring data, the present work is

part of a slice of a larger research project, which

performed the automatic extraction of tweets in

Brazilian Portuguese from Twitter, used NLP

techniques to the notation of the corpus, assembled a

database of the domain of calamity using techniques

for the construction of ontology. With the annotated

corpus, two methods were applied: summarization

methods of the terms (Freitas, 2015) and the LSA-

latent semantic analysis (Catae, 2012).

A method is proposed from the result generated

in the two previous methods. We called it semantic

similarity in polarized and summarized terms -

ASSTPS. The details of this method are shown in

section 2. Section 3 shows works similar to the one

proposed here, sentiment analysis of tweets in

Portuguese. Section 4 shows a set of data in which

the methods have been validated, this set may be of

any other specific domain, in this work data from a

corpus of calamity was used. Section 5 discusses the

results found with the proposed new method and

section 6 explains the conclusions and future works

proposed.

1.1 Contribution

Many works have performed methods of

summarization, polarization and LSA separately; it

is important to emphasize that the method employed

in this work is unprecedented. Here there will be a

proposal of a triangular matrix which indicates the

largest semantic similarity of tweets of a given

domain versus the description of the synset (set of

terms with the same meaning, as defined below)

more semantically similar to the tweet. In addition,

in this work we perform the complete translation of

Sentiwordnet (Esuli et al., 2006), including the

description of the synsets. The details of this method

will be shown in section 2.

1.2 Goals

The objective of this work is to find an effective

method to polarize terms, not in an isolated way, but

rather, in such a way that the crossover of tweets

with synsets and its descriptions of the Sentiwordnet

are as similar as possible so that the classification is

performed automatically without the manual

intervention of the man in the pre-classification of

similar messages. This research aims to evaluate the

reaction of individuals shared in the social network,

and not only to classify them as positive or negative,

but also to ascertain the semantic similarity of these

messages from the same domain.

2 MATERIALS AND METHODS

Messages were extracted automatically from Twitter

and stored in MySQL database, composing the

capture phase. Traditional NLP techniques were

used for the cleanliness and notation of the corpus,

following the main steps as follows: removal of stop

words (words which are not aggregators of

sentiment, in this work), tokenization, stemming,

and tagging; composing the preprocessing phase.

For a third phase, we named the classification of

sentiments between positive, negative and neutral,

polarity of words, using the database of

Sentiwordnet (Esuli et al., 2006). This lexicon

feature has 117.374 entries coming from automatic

annotations from all synsets of WordNet 3.0. Synset

is a set of terms with the same meaning.

Sentiwordnet is a tool which assigns a note to the

degree of positivity, negativity and objectivity of

IJCCI 2018 - 10th International Joint Conference on Computational Intelligence

226

words, available only in English. Compared to the

lexical resources available in Portuguese,

Sentiwordnet contains a greater number of words,

and is a very used resource for the classification of

sentiments. Because of this we made the decision to

translate the whole database of Sentiwordnet to

Portuguese automatically, including the synsets and

the comments which composed it.

2.1 Summarization Method

The Summarization Method or Term Score

Summation (Hamouda and Rohaim, 2011), performs

the sum of the positive and negative items found; for

example, in the sentence below, the positive words

with their values are: flowered - 0,125 and

enchanted - 0,375, totalling 0,5, and the negative

word with its value: polluted - 0,375. In this case, it

would be categorized as a positive phrase.

Algorithm 1 - “term score summation”, from Sousa

(2016, page 26) describes the workings of the

algorithm implemented in this work.

Sample: “the girl drew the polluted river, the

neutral neg.0,375 neutral

flowered garden and the enchanted sky”

pos. 0,125 neutral pos.0,375 neutral

Sousa (2016) applies the same summarization

method to analyze sentiments in texts extracted from

comments from the site Tripadvisor, which is related

to tourism. In his work, he proposes a second

method in which, after doing the sum of the terms

found in the synsets of Sentiwordnet, the average of

the positive values and the average of the negative

values is calculated and this value is assigned to the

word (Sousa, 2016). If we use only the

Summarization method, or the method of

summarization with application of the average, after

the translation of the synset to Portuguese, we will

have a problem, as shown in the Figure 1. After the

translation of the synset, we obtained a result of 12

synset "feliz" for descriptions of distinct synsets in

its original form in English.

Figure 1: Example Synset “feliz” translated.

Thinking about this problem, we will explain

what the LSA method is in the subsection below,

and in section 2.3 it will be shown that the proposed

new method does not perform the average of the

summarization. We solved the problem by the

semantic similarity of the tweet with the detailed

description of the Synset.

2.2 LSA Method

Catae (2012) says that the Latent Semantic Analysis

“is a technique in natural language processing,

which aims to simplify the task of finding words and

sentences similarity. Using a vector space model for

the text representation, it selects the most significant

values for the space reconstruction into a smaller

dimension” (Catae, 2012).

As parameters to entry of data in this algorithm,

811 tweets will be confronted in one space sample,

where the same tweets are lines and columns, so that

we can visualize the similarity among all tweets.

2.3 ASSTPS Method

The ASSTPS method uses the LSA algorithm in the

similarity calculation between the tweet and the

description of the translated synset, to select the

values to use in summarization. For example, in the

case of the word "love", if the description of greater

similarity is "a loved one: used as terms of

affection", the values of the word "love" will be pos

= 0125 and neg = 0.

For each token in the tweet, its synsets are

fetched. The LSA is made between its synsets and

the tweets to generate the semantic space. Within

that space, the similarity between the tweet being

analyzed and the synsets is calculated, allowing the

choice of the most similar synset. Note that a new

semantic space is generated for each token.

In this way, the result has greater accuracy, since

we are not doing the sum or average of the terms,

including those that are not similar. In this new

method we are crossing the synsets which have high

similarity for their description translated into

Portuguese with the tweets of the corpus of calamity.

3 RELATED WORK

In recent years, the interest in building of sentiment

classifiers has increased. Sentiment Analysis using

machine learning algorithms and Bayesian networks

using texts extracted from Twitter in the Spanish

language was proposed by Grigori et al. in 2010

(2010). In 2014, Kanavos et al., (2014) proposed a

model of sentiment analysis which measures the

influence of the user on the network with the other

nodes of that network. In 2013, Porshev et al.,

(2013) proposed a model of predicting the

Sentiment Analysis in Brazilian Portuguese Tweets in the Domain of Calamity: Application of the Summarization Method and Semantic

Similarity in Polarized Terms

227

psychological state of users from the texts of

Twitter, and classified the forecasts using the DJIA

algorithm, classifying them into eight emotions.

There are three lines of Research in the analysis

of sentiment: the lexical approach, the one by

machine learning and the hybrid which is the

junction of the previous two (Wilson et al., 2005).

We are quoting the works which have characteristics

in common with the proposed work. Works which

also use NLP techniques in the Portuguese language

to conduct sentiment analysis. In 2017, Santos

proposed a methodology for identifying polarity in

Texts based on Brazilian law projects and applied it

in a data set which reports on projects of law against

and in favor of the liberalization of abortion in

Brazil (Santos, 2017).

Freitas and Vieira (2015) present a semantic

polarity classifier for the classification of comments

written in Portuguese taken from TripAdvisor, an

online portal of opinions on travel and

accommodations (Freitas, 2015). Duarte (2013) used

NLP techniques to extract mentioned entities of

tweets in Portuguese and performed analysis of

sentiment taking into consideration the grammatical

construction of the message (Duarte, 2013).

Dosciatti and Ferreira (2013) used Support Vector

Machines (SVM) to identify emotions in texts

written in Portuguese in Brazil, the Corpus used in

the experiment is composed of news extracted from

an online newspaper (Dosciatti and Ferreira, 2013).

Lopes Rosa (2015) ranked the intensity of

sentiments in positive, negative and neutral

polarities, by the means of a new dictionary of

words in Portuguese and a new calculation of

sentiments (Lopes Rosa, 2015).

4 CORPUS AND SELECTION OF

TEXTS FOR ANNOTATION

To perform automatic extraction of texts, it is

necessary to first choose the ontology using

techniques of NLP. And for storing the large amount

of information we use computational linguistic

techniques (CL). One of the main types of storage

types in CL is in the singular corpus and in the plural

body. "Corpus is a collection of language portions

that are selected and organized according to explicit

linguistic criteria, in order to be used as a sample of

the language" (Percy et al., 1996).

A slice of data has been removed from a corpus

of calamity. Calamity represents an originally

positive situation and becomes a negative situation.

Actions can be taken, and again it is possible that the

situation which was negative becomes positive

(Segers et al., 2017).

Soon it is believed that sentiment analysis in this

set data becomes better validations. Although the

method can be applied in any set of data, for this

work we chose tweets from September 20, 2017,

which were part of the topTrend '#20deSetembro' or

containing the word 'Mexico' in the description of

the topTrend and should still contain the Word

'earthquake' from the description of the tweet.

Performing such query on the corpus returned 811

distinct tweets.

5 EVALUATION AND RESULTS

Figure 2 and the examples first show the validation

of the Summarization Method alone in the calamity

corpus. A total of 811 tweets, 6.789 tokens were

accounted. The positive value of 284.23 shown in

Figure 2 equates the sum of the lemmas values

found in the automatic translation of Sentiwordnet.

The negative value of 347.54 shown in Figure 2

equates the sum of the lemmas values found in the

automatic translation of Sentiwordnet.

375 tweets were classified for positive polarity

and 428 for negative polarity. It can therefore be

concluded that the issue in question "Earthquake in

Mexico on 20 September 2017" obtained 52.77% of

negativity, 46.24% of positivity.

Figure 2: Results Summarization Method.

Samples from the corpus which were sorted

correctly by the Summarization Method were

extracted below. The samples are in the Portuguese

language.

Sample with 1,2867 positive polarity: “Visitei o

México ano passado, um país sensacional, cheio de

gente feliz, tô mal com a notícia do terremoto”.

Sample with 1,7806 negative polarity: “México é

um belo País! Que tristeza o Terremoto de ontem.

E agora, Porto Rico em perigo novamiente tambem

nao me deixa tranquila aqui!”.

Figure 3 is a sample of the data for the validation

of the LSA Method, where high rates of semantic

similarity were found in tweets compared two to two

with opposite polarities. At 0.53 rate of similarity for

the tweet ID: 4 and 446; and the rate of 0.52

IJCCI 2018 - 10th International Joint Conference on Computational Intelligence

228

similarity to the tweets ID: 671 and 765. Figure 4 is

a sample of the data for the validation of the LSA

Method, where high rates of semantic similarity

were found in tweets compared two to two with

equal polarities. At 0.61 rate of similarity for the

tweet ID: 29 and 47; and the rate of 0.72 similarity

to the tweets ID: 80 and 431.

According to the results from the first two

methods, discrepancy was observed, not allowing a

fundamental conclusion for the event of September

20, 2017 with the natural phenomenon of an

earthquake in Mexico. The results should obviously

be of sadness, unhappiness and anguish, pointing to

a negative polarity analysis.

Figure 3: Results - High Rate Similarity x Polarity

Opposite.

Figure 4: Results - High Rate Similarity x Polarity Equal.

We sought some common synsets in the field of

calamity, for example: Earthquake, volcano and

cyclone. And we realized that most of them had zero

values in the Sentiwordnet database. Figure 5 shows

examples of common synsets in the calamity corpus

with zero values. And one of the synsets

"earthquake" is of positive value of 0.125. This

caused the new applied method to show high values

for positive polarity, since the term "earthquake"

appears in all the tweets analyzed in this work. Two

analyses were performed for validating the ASSTPS

method. The first, as shown in Figure 6, kept the

original value found for the synset "earthquake" in

the Sentiwordnet database; the second analysis, as

shown in Figure 7, changed to 0.25 the negative

value for all the "earthquake" synsets found.

For validating the ASSTPS method, the synset

values found in Sentiwordnet remained; the result of

the method here proposed categorized in the base

449 tweets with positive polarity, representing

55.36% of the base.

The earthquake is a natural disaster event, and its

causalities stir actions of dissatisfaction in the

population. Thus, as a complementary exercise, we

adjusted the negative value of the Synsets

"earthquake" to 0.25. We re-processed the algorithm

and Figure 7 shows the values obtained from 647

tweets classified by proximity semantics in the

description of the synset, representing 79.77%.

Figure 5: Example Calamity Synsets.

Figure 6: Results – ASSTPS Method.

Figure 7: Results – ASSTPS Method with synset

“earthquake” negative value 0.25.

6 CONCLUSIONS AND FUTURE

WORK

One of the main contributions of this work is the

research into sentiment analysis performed in a

Brazilian Portuguese calamity corpus. In this work,

we present semantic classifications of polarity for

analyzing texts. The classifiers used Sentiwordnet to

assign values to words. The entire collection of

Sentiwordnet was automatically translated into

Portuguese; even with translation errors, the

classifiers that used the translated Sentiwordnet

achieved good performance.

The Summarization method is a lexical classifier

that has some mismatches. The LSA method

performs the semantic proximity of the tweets

among other tweets and does not perform any

polarity check. The proposed new method performs

the semantic similarity of the tweets with the synsets

of Sentiwordnet by semantic proximity and performs

the polarity classification. The polarization of

sentiments by semantic proximity associated with

the synsets of Sentiwordnet and their comments can

be of great help in tasks, such as sorting content of

relevant messages, discarding unnecessary

Sentiment Analysis in Brazilian Portuguese Tweets in the Domain of Calamity: Application of the Summarization Method and Semantic

Similarity in Polarized Terms

229

information, linking messages with the same theme

in common, and even generating metrics for

classifying emotions.

As future works, we suggest improvements to the

translation process, performing bi-gram and tri-gram

words as proposed in the work by Lopes Rosa

(2015). Use the new method proposed to classify

emotions at a second level, such as anger, fear, love

and hate; above all, to use a base similar to

Sentiwordnet with values of positivity, negativity

and more accurate objectivity for the dominance of

calamity.

ACKNOWLEDGEMENTS

The authors are grateful to the CNPQ process

141077/2015-8 for the support received.

REFERENCES

Berners-Lee, Tim. The WorldWideWeb browser. World

Wide Web Consortium. Homepage,

https://www.w3.org/People/Berners-Lee/WorldWide

Web.html, last accessed 2018/03/ 01.

Catae, F. S. Classificação Automática de Texto por Meio

de Similaridade de Palavras: um algoritmo mais

eficiente. Master's Dissertation, Universidade de São

Paulo (2012).

Chaubard, F., Fang, M., Genthial, G. et al. Natural

language processing with deep learning. Lecture

Notes: Part I - Winter - 2017. Homepage,

https://tensorflowkorea.files.word press.com/2017/03/

cs224n-2017winter-notes-all.pdf, last accessed

2018/04/15.

Dosciatti, M. M.; Ferreira, E. C. L. P. C. Identificando

emoções em textos em português do brasil usando

máquina de vetores de suporte em solução multiclasse.

ENIAC-Encontro Nacional de Inteligência Artificial e

Computacional. Fortaleza, Brasil (2013).

Duarte, E. S. Sentiment analysis on twitter for the

portuguese language. PhD Thesis, Faculdade de

Ciências e Tecnologia (2013).

Esuli, A., Sebastiani, F. Sentiwordnet: A publicly

available lexical resource for opinion mining. In: In

Proceedings of the 5th Conference on Language

Resources and Evaluation - LREC’06, 417-422

(2006).

Freitas, L. D; Vieira, R. Exploring resources for sentiment

analysis in portuguese language. In: IEEE. 2015

Brazilian Conference on Intelligent Systems (BRACIS)

(2015).

Freitas, L. A. de. Feature-level sentiment analysis applied

to brazilian Portuguese reviews. PhD Thesis,

Pontifícia Universidade Católica do Rio Grande do Sul

(2015).

Grigori, S. et al. Empirical study of machine learning

based approach for opinion mining in tweets.

In Proceedings of the 11th Mexican international

conference on Advances in Artificial Intelligence

(2012).

Hamouda, A., Rohaim, M. Reviews classification using

sentiwordnet lexicon. In: World Congress on

Computer Science and Information Technology

(2011).

Internet World Stats. Top 10 Languages Used in the Web -

December 31, 2017. Homepage,

https://www.internetworldstats.com/stats7.htm, last

accessed 2018/04/01.

Internet World Stats. Top 20 Countries With the Highest

Number of Internet Users - December 31, 2017.

Homepage, https://www.internetworldstats.com/

top20.htm, last accessed 2018/02/01.

Kanavos, A. et al. Conversation Emotional Modeling in

Social Networks, Tools with Artificial Intelligence

(ICTAI), 2014 IEEE 26th International Conference

(2014).

Liu, B. Sentiment Analysis and Opinion Mining

(Synthesis Lectures on Human Language

Technologies. California: Morgan & Claypool

Publishers (2012).

Lopes Rosa, R. Análise de sentimentos e afetividade de

textos extraídos das redes sociais. Tese de Doutorado

– Escola Politécnica da Universidade de São Paulo

(2015).

Percy, C. E. et al. Synchronic Corpus Linguistics – Papers

from the sixteenth International. Conference on

English Language and Research on Computerized

Corpora (ICAME 16), Amsterdam/Atlanta (1996).

Plutchik, R. The nature of emotions. v. 89, n. 4, p. 344–

350 (2001).

Porshnev, A.; Redkin, I.; Shevchenko, A., "Machine

Learning in Prediction of Stock Market Indicators

based on Historical Data and Data from Twitter

Sentiment Analysis," Data Mining Workshops

(ICDMW), 2013 IEEE 13th International Conference

(2013).

Sousa, R. C. C. Identificando Sentimentos de Textos em

Português com o Sentiwordnet traduzido. Monografia

– UFC, curso de Ciência da Computação (2016).

Santos, D. L. B. Metodologia de Identificação de

Polaridade em Textos com Base em Projetos de Lei

Brasileiros. Dissertação de mestrado – UFRJ,

Programa de Engenharia Civil (2017).

Segers, R., Caselli, T., Vossen, P. The Circumstantial

Event Ontology (CEO). Proceedings of the Events and

Stories in the News Workshop. Association for

Computational Linguistics, Vancouver, Canada

(2017).

Semantic Web - W3C. Semantic Web. Homepage,

https://www.w3.org/standards/semantic web/, last

accessed 2018/03/01.

The Statistics Portal. Most-used languages on Twitter as

of September 2013. Homepage,

https://www.statista.com/statistics/267129/most-used-

languages-on-twitter/, last accessed 2018/04/08.

IJCCI 2018 - 10th International Joint Conference on Computational Intelligence

230

The Statistics Portal. Most popular social networks

worldwide as of April 2018, ranked by number of

active users (in millions). Homepage,

https://www.statista.com/statistics /272014/global-

social-networks-ranked-by-number-of-users/, last

accessed 2018/04/08.

Twitter NIC. A Multilingual Social City. Homepage,

http://ny.spatial.ly/, last accessed 2018/04/25.

Wilson, T.; Wiebe, J.; Hoffmann, P. Recognizing

contextual polarity in phrase-level sentiment analysis.

In: Proceedings of the conference on Human

Language Technology and Empirical Methods in

Natural Language Processing. Stroudsburg, PA, USA:

Association for Computational Linguistics (2005).

Sentiment Analysis in Brazilian Portuguese Tweets in the Domain of Calamity: Application of the Summarization Method and Semantic

Similarity in Polarized Terms

231