AFFECTIVE ALGORITHM TO POLARIZE CUSTOMER OPINIONS

Domenico Consoli, Claudia Diamantini and Domenico Potena

Dipartimento di Ingegneria Informatica, Gestionale e dell’Automazione

Universit´a Politecnica delle Marche, Via B. Bianche, 60131, Ancona, Italy

Keywords:

Customer Opinions, Web 2.0, Sentiment Analysis, Affective Analysis, Opinion Mining, Opinion Polarity

Ekman Emotional Indexes.

Abstract:

Human interact with other people and exchange reviews and ideas via web. With the explosion of Web 2.0

platforms such as blogs, discussion forums, peer-to-peer networks, and various other types of social media,

consumers share, on the web, their opinions regarding any product/service. Opinions give information about

how product/service and reality in general is perceived by other people. Emotional needs are associated with

the psychological aspects of product ownership. The customer when writes his reviews on a product/service

transmits emotions in the message that he/she feels ﬁrst and after purchasing the product. For the enterprise

understanding customer emotional needs is vital for predicting and inﬂuencing their purchasing behaviour. In

this paper, we polarize, with original algorithm, customer opinions basing on emotional indexes that are used

for decipher, in affective key, facial expressions and emotional lexicon.

1 INTRODUCTION

Emotions are treated largely in affective computing

(Picard, 1997) that focuses on improving the interac-

tion between user and computer. The affective com-

puting aims to build emotional machines (Khashman,

2008) that can recognize and express emotions. The

computer, in future, will understand the mood of op-

erator and they’ll act accordingly; also the computer,

with its behavior, will arouse emotions in the opera-

tor. We’ll have a friendly computer with technologi-

cal and advanced emotional interfaces. The user af-

fective state is present in many ways, through facial

expressions, speech and text; also the text brings an

intrinsic emotional content. We must not forget that

today computers, for the most part, processing only

textual data and that the text analysis, based on the

sentiments, has increased with the availabilityof large

amount of web pages.

In this work, we focus on the analysis of the opin-

ions that customers express over the Web. Customers

exploit web 2.0 tools (chats, forums, blogs, and so

on) for expressing their opinions about a product and

suggesting solutions for improving it. There are var-

ious web sites that collect free customer reviews:

epinions.com, cnet.com, complaints.com, planetfeed-

back.com, ecomplaints.com, dooyoo.it.

Customer opinions constitute a gold mine for an

enterprise, both for the improvement of products and

for the reinforcement of the customer loyalty. In order

to full exploit the information contained in customer

opinions, it is important to polarize opinion about

(part of) product and in particular to return whether

the opinion is positive or negative.

In our paper we present an original algorithm,

based on emotional indexes, to polarize customer

opinions. This paper is organized as follows: in the

next section we give a brief description of the emo-

tional purchase process, while in the third section, re-

lated works for sentiment analysis are discussed. The

fourth and ﬁfth sections are devoted to present our ap-

proach to customers opinions polarization and the re-

sults. Finally some conclusions are drawn.

2 EMOTIONS IN CUSTOMER

BEHAVIOR

Emotions play a crucial role in all phases of purchase

behavior. The customer when writes a review about

a product/service, expresses also the same emotions

that she felt ﬁrst and after purchasing the product; so

from a enterprise point of view it is important to un-

derstand what drives customer to choose a particular

product/service. Understanding customer emotional

157

Consoli D., Diamantini C. and Potena D. (2009).

AFFECTIVE ALGORITHM TO POLARIZE CUSTOMER OPINIONS.

In Proceedings of the 11th International Conference on Enterprise Information Systems - Human-Computer Interaction, pages 157-160

DOI: 10.5220/0001851601570160

 SciTePress

needs is vital for predicting and inﬂuencing their pur-

chasing behavior.

Many words express a direct affective senti-

ment (“joy”, “sorrow”, “happiness”) other indirect

(“cry”, “smile”, “monster”, “abandoned”). Some

words (e.g. “nail”, “ball”, “table”), if decontextual-

ized, don’t express neither directly or indirectly emo-

tions, also if the sentence, where these words belong

to, has a speciﬁc emotional polarity. For example, the

word “nail” means an object used to ﬁx a picture to

a wall and can remind us positive or negative emo-

tions; it depends from the context. The “ball”, that

easily we associate with a game, has likely an emo-

tional positive contribution.

Many times the emotional/symbolic traits are

highly representative of the speciﬁc identity and

brand. For example, in luxury goods, the emotional

aspects as brand, uniqueness and prestige for purchas-

ing decisions, are more important than rational as-

pects such as technical, functional or price. For the

seller would be important to measure accurately the

value generated by emotions to determine so an ap-

propriate price.

Another factor that inﬂuence purchasing customer

is, for example, the disgust that plays a key role in the

inverse relationship between attitude and intention to

purchase. Customer don’t buy products disgusting.

The disgust is a repugnance toward any object, action

or person. The disgusting is an index of variation of

the intention to purchase.

The relationship that establishes between a brand

and a customer is the the same among people. Be-

tween company and customer establishes an emo-

tional relationship. The creation and the excite-

ment of a series of positive emotions into customers

helps company in the sell. The enterprise can in-

volve customers to increase their sense of belong-

ing to a community that shares the same brand and

the same values with other people. The enterprises

often encourage exchange of opinions, by making

available virtual communities, e.g. Italian Nikon’s

Camera forum, where people review Nikon prod-

ucts (http://www.nital.it/forum/), the blog on Benet-

ton products (http://benettontalk.com), and so on.

3 RELATED WORKS

In the literature, various methods have been proposed

for opinion mining and sentiment analysis(Liu et al.,

2003).

In the Keyword Spotting (Boucouvalas and Zhe,

2002) approach, text is classiﬁed into affective cat-

egories based on the presence of fairly unambigu-

ous affective words like “distressed, “happiness and

“anger.

All terms that describe emotional states represent

the most direct way to communicate emotion by text.

The simplest and most used analysis is based on the

search for keywords (like “happy, “sad, “angry, etc).

The Lexical Afﬁnity (Valitutti et al., 2004) method

assigns to words a probabilistic afﬁnity (trained from

linguistic corpora) for a particular emotion. For ex-

ample, “accident” might be assigned a 80% proba-

bility of being indicating a negative affect, as in “car

accident”, “hurt by accident”.

Esuli and Sebastiani (Esuli and Sebastiani, 2006)

have created SentiWordNet, a lexical resource for

opinion mining, where they assign to each synset (set

of synonyms) of WordNet a sentiment scores: posi-

tivity, negativity and objectivity (i.e. neutral). The

opinion is positive if the positivity of its terms is

higher than negative and objective scores and vicev-

ersa for negative opinion. WordNet Affect (Strappa-

rava and Valitutti, 2004) is a linguistic resource for

a lexical representation of affective knowledge. In

WordNet Affect each synset of WordNet is labeled

by one or more affective-labels, representing the af-

fective meaning of the synset. Examples of affective-

labels are emotion, mood, trait, cognitive state, phys-

ical state, etc.

The original algorithm that we have developed

mainly focuses on six Ekman emotional indexes (Ek-

man, 2007): “anger”, “disgust”, “fear”, “happi-

ness”, “sadness”, and “surprise”. We use these com-

ponents to study emotional lexicon.

4 ALGORITHM TO POLARIZE

CUSTOMER OPINIONS

The goal of our approach is to polarize a customer

opinion about a topic, that is a characteristic of a

(part of) product/service. We use this approach in

the conceptual framework CeC (Consoli et al., 2008)

for gathering customer opinion and improving the

product/service.

We ﬁrst consider customer opinions concerning

a some topic, then we break them into homogenous

sentences, next each sentence is split in words. The

polarity of a sentence is given by estimating the

polarity of any its words.

To divide a sentence in words we make a

preprocess: elimination of stop-words (articles, con-

junctions, prepositions) and division of the sentence

into single words with lemmatization (Berry and

Castellanos, 2007). In the preprocessing we have

used the GATE library with some modiﬁcation in

ICEIS 2009 - International Conference on Enterprise Information Systems

158

order to reduce the complexity of the analysis. The

output of the preprocessing phase is a dataset of

signiﬁcant words of each opinion.

Analysing this dataset, we create a matrix of

word-document occurrences containing the number

of times a word appears in a document. This matrix

is then compressed by SVD algorithm (Okˇsa and

Vajterˇsic, 2001) to obtain the words-words co-

occurrence matrix W, where each word is represented

by a n dimensional statistical vector. Each dimension

of such statistical vector represents the co-occurrence

of the single word with other words of the matrix.

Words in W can be divided in affectives and

non-affectives. The former can be further classiﬁed

into direct affective words, that directly refers to

affective states (e.g. ”fear”, ”cheerful”, ”disgusted”),

and indirect affective words, having an indirect affec-

tive impact that depends on the context, i.e. words

representing moods, situations eliciting emotions, or

emotional responses like ”monster” or ”cry”.

Affective words will be exploited to build up the

model that will allow us to assign a polarity value to

non-affective words. For this reason we call training

set, with an abuse of notation, the set of affective

words, and test set the non-affective words. In the

following, it is described the algorithm for building

the polarity model and to assign a polarity value to a

non-affective word:

1) Assign an affective vector to each word of the

training set.

The components of the affective vector represent the

contribution of the following states in the emotional

deﬁnition of the word: “happy”, “sad”, “angry”,

“surprise”, “fear”, and “disgust”. In our approach,

the affective components have an integer value

between 0 and 10. For example, if we associate to

the word ”stench” the affective vector (0, 2, 2, 0,

0, 6), it means that in the deﬁnition of the affective

meaning of the word stench, the elements sad and

angry contribute with a small value, the element

disgust with a high value and other elements don’t

produce any contribution.

2) For each word t of test set, select the word w

of the training set, maximizing the normalized scalar

product k :

k =

k.kw

where t

and w

are the statistical vectors of t and

w respectively. The resulting w represents the most

similar word to t in the co-occurrence space. Our

idea is that word with similar statistical vector have

also similar affective meaning.

3) Compute the affective vector t

of t:

= k· w

where w

is the affective vector of w and the function

int(x) gives the integer part of x.

4) Compute the polarity value of t:

The word has a positive value if the sum of compo-

nents related to positive affective states (happy, sur-

prise) is greater than the sum of components of neg-

ative affective states (sad, angry, fear, disgusted); the

word polarity is negative in the other case. Finally,

the sum of the affective vectors of all words of an

sentence deﬁnes its affectivity and, consequently, its

polarity value.

5 CASE STUDY

In order to test the validity of our algorithm on the

opinion polarity we have used almost 1000 customer

opinions about a resort in Sharm el-Sheikh and in

particular we selected opinions on some services:

Kitchen, Restaurant, Room Service, and Administra-

tion.

Opinions were collected from various Internet

sites, like alpharooms.com, realholidayreports.com.

After the pre-processing phase, we derive 11900 sig-

niﬁcant words, divided in 2300 affective words and

9600 non-affective ones.

In order to show the results of our algorithm we

consider two typical sentences taken from separate

posts: the ﬁrst expresses globally positive opinion

while second a negative one. The sentences that we

have taken into account are follows:

• Sentence 1. Food was excellent and a great vari-

ety, especially if you are fond of sweets.(positive

opinion)

• Sentence 2. Beware of the dangerous on the bad

beach and disgusting and terrible pasta.(negative

opinion)

The resulting affective vectors of any words in these

sentences, are shown in Table 1 and Table 2. In Tables

words already belong to the training set are in italic.

In the tables there are several zeros. If a word has

a higher value of happy probably it has a low value

(zero) as angry or sad and viceversa.

In Table 1, despite fond (”cherished with great af-

fection” from the Merriam-Webster dictionary) has a

AFFECTIVE ALGORITHM TO POLARIZE CUSTOMER OPINIONS

159

Table 1: Affective vectors of words in Sentence 1.

Happy Sad Angry Surprise Fear Disgust

food 7 0 0 0 0 0

excellent 9 0 0 0 0 0

great 10 0 0 0 0 0

variety 5 0 0 2 0 0

especially 5 0 0 0 0 0

fond 0 6 0 0 4 0

sweets 8 0 0 0 0 0

Phrase 1 44 6 0 2 4 0

positive emotional valence, the algorithm associates

to term a negative affective vector. This does not hap-

pen for other words, whose the affective vectors are

correctly estimated. The sum of happy and surprise

(46) is greater than the sum of other components (12),

thus the sentence globally returns a positive opinion.

Table 2: Affective vectors of words in Sentence 2.

Happy Sad Angry Surprise Fear Disgust

beware 0 6 0 0 4 0

dangerous 0 0 0 0 8 0

bad 0 0 0 0 7 0

beach 9 0 0 0 0 0

disgusting 0 0 0 0 9 0

terrible 0 6 0 0 9 0

pasta 5 0 0 0 0 0

Phrase 14 0 0 0 24 10

In Table 2, the algorithm associates to the adjective

terrible a high score of fear index. Since it refers to

pasta we would associate it with disgusted; fear would

be appropriate in the case of a terrible percept in the

presence of a bleak scene. The disambiguation of the

true sentiment in response to the context is a limit of

the present work and will be investigated in.

6 CONCLUSIONS

Nowaday, for the enterprise, it is important gather-

ing a large amount of customer opinion from the web

2.0 tools. In this paper we illustrate an original ap-

proach to polarize (positiveor negative)opinion based

on Ekman indexes to evaluate emotional value of re-

view about product/service. In our opinion, these in-

dexex, allows to better capture the emotional state of

customers about purchase.

The approach, based on the affective value of

each single word, produces good results on docu-

ments medium-large dimensions but often fails on in-

dividual sentence. For resolving the question, in fur-

ther works, we’ll try to understand the semantics of

the contents of a sentence using a knowledge base of

common sense.

For use these results in the emotional recognition

it is necessary focuses on concepts rather than on

terms, in order to relate the words to emotional states

through a conceptual representation. To this end we’ll

consider a conceptual semantics.

Another improvementto characterize and to struc-

ture emotional concepts, it is to build a detailed tax-

onomy from basic emotions (fear, anger, joy, disgust,

sadness). This can help more easy to capture the emo-

tional sense of a generic word.

REFERENCES

Berry, M. and Castellanos, M., editors (2007). Survey of

Text Mining II: Clustering, Classiﬁcation, and Re-

trieval. Springer.

Boucouvalas, A. and Zhe, X. (2002). Text-to emotion en-

gine for real time internet communication. In Pro-

ceedings of International Symposium on CSNDSP,

pages 164–168, UK. Staffordshire University.

Consoli, D., Diamantini, C., and Potena, D. (2008). Improv-

ing the customer intelligence with customer enterprise

customer model. In Proceedings of the 10th Interna-

tional Conference on Enterprise Information Systems,

pages 323–326, Barcelona, Spain.

Ekman, P. (2007). Emotions Revealed: Recognizing Faces

and Feelings to Improve Communication and Emo-

tional Life. OWL Books, NY.

Esuli, A. and Sebastiani, F. (2006). SENTIWORDNET: A

publicly available lexical resource for opinion min-

ing. In Proceedings of LREC-06, 5th Conference on

Language Resources and Evaluation, pages 417–422,

Genova, IT.

Khashman, A. (2008). A modiﬁed backpropagation learn-

ing algorithm with added emo-tional coefﬁcients. In

Transaction on neural networks, volume 19 of 11,

pages 1896–1909. IEEE.

Liu, H., Lieberman, H., and Selker, T. (2003). A model

of textual affect sensing using real-world knowledge.

In IUI ’03: Proceedings of the 8th international con-

ference on Intelligent user interfaces, pages 125–132,

New York, NY, USA. ACM.

Okˇsa, G. and Vajterˇsic, M. (2001). Systolic block-jacobi

svd algorithm for processor meshes. 185:211–235.

Picard, R. W. (1997). Affective computing. MIT Press,

Cambridge, MA, USA.

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

An affective extension of wordnet. In The 4th Interna-

tional Conference On Language Resources and Eval-

uation, pages 1083–1086.

Valitutti, A., Strapparava, C., and Stock, O. (2004). Devel-

oping affective lexical resources. 2(1):61–83.

ICEIS 2009 - International Conference on Enterprise Information Systems

160