A Multidomain and Multilingual Conceptual Data Model

for Online Reviews Representation

Marcirio Silveira Chaves

and Winnie Picoto

Business and Information Technology Research Center (BITREC), Universidade Atlântica,

Fábrica da Pólvora de Barcarena, 2730-036 Barcarena, Portugal

Centro de Investigação Avançada em Gestão do ISEG (ADVANCE),

Instituto Superior de Economia e Gestão – ISEG, Lisboa, Portugal

Keywords: Model-Driven Engineering, Metamodel, UML, Online Reviews.

Abstract: User-Generated Content (UGC) such as online reviews are freely available in the web. This kind of data has

been used to support clients’ and managerial decision making in several industries, e.g. books, tourism or

hospitality. However, the challenge is how to represent this information in a structured way in order to

leverage on the information provided by the use of Web 2.0 applications. To deal with this challenge,

models and metamodels have been used to support a set of concrete applications in several sub-domains into

Computer Science and Information Systems body of knowledge (Karagiannis and Höfferer, 2006). This

paper focuses on the model-driven engineering and introduces a new multidomain and multilingual

conceptual data model to represent UGC. This model is based on a characterization of online reviews and

aims to capture all the facets of these reviews. The characterization of the reviews’ sentences extends

previous models (such as Martin and White, 2007; Ding et al., 2008; Liu, 2010). Applications build on the

model proposed in this paper may allow in-depth analysis of the fine-grained and disperse knowledge

existent in the UGC. Furthermore, as this model is domain-independent it can be used to represent multiple

types of reviews.

1 INTRODUCTION

The usage of Web 2.0 websites to express opinions

has increased in the last years generating a huge

amount of data which has been used to support

decision making. The distillation of knowledge from

this unstructured and dispersed information can be a

key factor to managers improve their products and

services. Commercial Web 2.0 tools such as

Clarabridge (2012), Attensity (2012), SocialMetrics

(2012) and Synthesio (2012) are available, but their

cost still remains inaccessible to Small and Medium

Entreprises (SME). Considering that SME represent

99% of the European businesses (EC, 2012), the

development of a financially affordable solution is

needed.

Moreover, commercial Web 2.0 tools hardly

present the data model they use to represent the

unstructured opinions’ texts into structured ones.

Such a model is omitted or really does not exist in

the literature. This literary gap evidences the few

attention that model-driven engineering has given to

Web 2.0 data.

On the other hand, from the academic point of

view, Liu (2010) presents a characterization of the

opinions expressed thought plain text data, as blogs,

tweets and full-blown service/product review. This

characterization may be extended and used to

develop a model to represent online reviews.

A model is an abstraction of phenomena in the

real world, and a metamodel is yet another

abstraction, highlighting properties of the model

itself (van Gigch, 1991). Models and metamodels

have been used to support a set of concrete

applications in several sub-domains into Computer

Science and Information Systems, as described in

the survey carried out by Karagiannis and Höfferer

(2006). The model-driven approach defines

relationships among concepts in a domain and

precisely specifies the key semantics and constraints

associated with these domain concepts (Schmidt,

2006).

Silveira Chaves M. and Picoto W..

A Multidomain and Multilingual Conceptual Data Model for Online Reviews Representation.

DOI: 10.5220/0004021800140023

In Proceedings of the 7th International Conference on Software Paradigm Trends (ICSOFT-2012), pages 14-23

ISBN: 978-989-8565-19-8

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

A conceptual data model, sometimes called

domain model, is typically used to explore domain

concepts with project stakeholders. It is also used to

map concepts and their relationships in a domain.

We argue that it is well suited to represent

knowledge from opinionated texts.

The main question addressed in the present study

is how to capture the different facets of online

reviews in an usable knowledge representation? This

paper first describes the properties of opinionated

texts which should be taken into account along the

model’s design. The solution approach outlines a

new multidomain and multilingual conceptual data

model for capturing and storing knowledge from

opinionated texts.

The model proposed is within the scope of the

framework described in Chaves, Trojahn and Pedron

(2012). This framework aims to provide an

environment to customer knowledge

management which integrates information from

online reviews, stores it in a knowledge base

and presents it in a user interface. A knowledge

base is been built to implement the conceptual

data model proposed in this paper.

This paper is structured as follows: Section 2

discusses some of the existing work about

modelling. Section 3 presents a characterization of

online reviews, detailing the main elements which

should be explicit and represented in a model.

Section 4 introduces the scenario problem, raising

some questions which can be answered with the help

of a model. Section 5 describes the solution

approach which is based on a conceptual data model

proposed to represent knowledge from opinionated

texts. Section 6 describes manual and automatic

approaches for recognizing facets in online reviews.

In Section 7 the conclusions, final remarks and

further research are discussed.

2 RELATED WORK

UGC modelling is a recent field of knowledge and

related works seem to be rare making difficult a

direct comparison with existing studies. To our

knowledge, this is the first academic work that seeks

to develop a metamodel and a class model to

structure the unstructured UGC. In fact, metamodels

has been widely used in the software engineering to

solve problems from the real world (Karagiannis and

Höfferer, 2006). Related works are concerned with

opinion mining of online reviews and online reviews

classification. Although there have been a stream of

works aiming at capture knowledge from those

unstructured review sentences which stands for the

relevance to address this matter, less attention have

been given to represent that information in a

structured way so reviews’ could be used to support

managerial decisions.

Regarding previous studies that developed a

metamodel, the present work is most closely related

to Chaves, Rodrigues and Silva (2007). They used a

metamodel to represent geographical information in

order to implement a knowledge base. They also

used structured data as the model input. Their model

stores administrative and physical geographic

information which is exported as ontologies. The

ontologies represent knowledge integrated from geo-

administrative and geo-physical domains. Although,

the classes Feature, Source and Feature-Relationship

are reused by the model proposed in this paper, we

deal with unstructured data, i.e. web texts, and do

not export them as ontologies.

Looking at existing works that have dealt with

online reviews, Turney (2002), Bai (2011), Khan et

al. (2010) and Hu and Lu (2004) have developed

different approaches to extract knowledge from that

unstructured data.

Turney (2002) applied a semantic orientation to

unsupervised classification of reviews. He

developed an algorithm for classifying reviews as

recommended or not recommended based on the

average semantic orientation for each review

sentence. The algorithm has an average accuracy of

74% evaluated based on reviews from four different

sectors (automobiles, banks, movies and travel

destinations). Although that work does not propose a

class model to structure those reviews, it presents a

set of attributes to classify reviews that we have also

considered in our model.

In line with that work, Bai (2011) developed a

method for sentiment analysis. Her results show that

the proposed method is capable to identify a

parsimonious set of predictive features suggesting

that sentiments are captured by conditional

dependencies among words.

Khan et al. (2010) presents an automated method

for opinion extraction from customer reviews. The

unstructured review sentences were represented in

such a way that is possible to extract knowledge

from it. They found that only a portion included

opinion-oriented words that needed to be processed

in order to get knowledge from those reviews and

that the most important was the choice of feature set.

Hu and Lu’s (2004) work aimed at mine and

summarize customer reviews of a product which is

argued by the authors to be different from the

traditional ones since they only mined the features

A Multidomain and Multilingual Conceptual Data Model for Online Reviews Representation

on which the customers have expressed some

opinion. Towards the accomplishment of such

objective, they have taken three steps: (1) mining

product features with comments; (2) opinion

sentences identification and orientation; (3) results

summary. The proposed method was validated by

means of an experience which has demonstrated to

be effective.

3 OPINIONS’

CHARACTERIZATION

In order to understand what kind of knowledge to

represent in the content underlying the online

reviews, we build on the existing approaches of the

opinion mining field. One of the most common

approaches to deal with the opinion mining problem

is the lexicon-based one. Lexicon-based approaches

include the use of a list of nouns, verbs, adjectives

and adverbs (Chesley et al., 2006) and a list of

conjunctions and connectives (Liu, 2010). Chesley

et al. (2006) use verbs and adjectives to classify

English opinionated blog texts. Ding et al. (2008)

use all of these parts of speech in a holistic lexicon-

based approach. Khan et al. (2010) use auxiliary

verbs to get features and opinion-oriented words

about products from texts.

Processing textual data allows us to provide

more fine-grained knowledge to the decision

making. Online opinions can be expressed in several

ways, using different words to express positive or

negative feelings or even to be neutral. In order to

capture as many information as possible from online

reviews about a product or a service, first it is

necessary to understand each sentence in the review.

We use and extend the definition proposed by

(Ding et al., 2008; Liu, 2010; Martin and White,

2007) to analyse the reviews’ sentences. Let the

review be r. In the most general case, r is

characterized as a set of the following elements

{O,F,SO,A,R,I,C,SG,H,S,PoS,CC}, where: O:

Object, F: Feature, SO: Semantic Orientation, A:

Attitude, S: Source, SG: Suggestion; R:

Recommendation, I: Intention, C: Complain, H:

Holder, PoS: Part-of-Speech, and CC: Conjunction

and Connective.

1. Object (O): An object is a product or a

service, under review which is composed by a set of

features. Objects are also called entities.

2. Feature (F): A feature is a component or part

of an object. For example, actor and photography

are features of a movie. Features are also called

attributes or facets. According to Ding et al. (2008) a

feature can be explicitly or implicitly mentioned in a

review.

2.1 Explicit Feature: If a feature f appears in

review r, it is called an explicit feature in r. For

example, in the sentence “The hotel is located very

near the city centre”, location is an explicit feature.

2.2 Implicit Feature: If f does not appear in r

but is implied, it is called an implicit feature in r. In

the sentence “Hotel is far from public

transportation”, location is an implicit feature.

3. Sentence-Orientation (SO): A review

consists of a sequence of sentences r=〈 s1, s2, …,

sm〉 (Ding et al., 2008). A sentence can be

evaluated from the following perspectives:

3.1 Objectivity: An objective sentence contains

or mention facts, e.g. “This hotel is far from the

airport, ca. 15km.”, while a subjective sentence does

not mention any fact, e.g. “The parking could be

free”.

3.2 Positivity: It describes the orientation

present in a sentence, i.e. positive, negative, neutral

and irrelevant. For example, the sentence “Free and

fast wifi in the room”, refers to a positive impression

about the feature room in a hotel.

3.3 Intensity: It refers to the strength of the

private state that is being expressed, in other words,

how strong is an emotion or a conviction of belief

(Wilson, 2008). It describes how intense it was the

experience of using a product or service, i.e. very

positive, positive, neutral, negative and very

negative. For example, the sentence “Very kindly

staff” refers to a very positive impression on the staff

service.

3.4 Negation: It is marked if the sentence

contains at least an occurrence of negation, e.g. not

and never. This information is relevevant to support

algortihms in the task of recognizing several kinds

of sentence-orientation (e.g. positivity and intensity).

4. Attitude (A): The evaluation of attitude

expressions is a complex task. We use the Appraisal

Theory (Martin and White, 2007), which is a

systemic-functional approach to analysing how

subjective language is used to express an attitude of

some kind towards some target. Appraisal Theory

specifies three attitude types: affect (personal

emotional state), appreciation (evaluation of

phenomena), and judgment (social or ethical

appraisal of other’s behaviour).

4.1 Affect: It refers to a personal emotional state,

e.g. happiness, sadness or angry, and is the most

explicitly subjective type of appraisal (Whitelaw et

al. 2005). For example, the sentence “I was very

happy for spend the holiday in this comfortable

hotel” expresses a feeling of happiness.

ICSOFT 2012 - 7th International Conference on Software Paradigm Trends

4.2 Appreciation: It refers to the intrinsic object

properties, e.g. dirty, cold or small.

4.3 Judgment: It refers to an opinion formed by

judging other’s behaviour, e.g. “attentive staff” and

“friendly waiter”. Both appreciation and judgment

can be classified as positive or negative.

5. Recommendation (R): A recommendation is

a positive or negative statement which explicitly

mention in words, an action or advice as for example

,“I recommend!” or “if you plan to travel there, find

another hotel closer to the main area, cause this one's

not at all worth it!”. In a recommendation it is also

implicitly present the satisfaction of a reviewer.

Reviews that state some intention can be also

considered an implicit recommendation as for

example, the sentence “If we get the chance to go to

Roma again we will ensure that we stay at the X

Hotel again”.

6. Suggestion (SG): A hint given by the holder

in order to improve the product or service and it

could be an explicit or implicit suggestion:

6.1 Explicit Suggestion: It is a direct mention on

a need of the product or service. For example, in a

hotel service, an explicit suggestion is “The

mattresses of the beds need to be exchanged”.

6.2 Implicit Suggestion: It is an indirect

mention on a need of the product or service, as for

example, “the hotel does not offer Internet service”.

7. Intention: it is a mention in which a holder

explicitly intends (or not) to experiment again some

product or service. It could be positive or negative.

7.1 Positive Intention: For example, “When I

come back to Lisbon, I intend to stay there myself”

and “I intend to come back there”.

7.2 Negative Intention: For example,

“Certainly, we won’t repeat”.

It is important to notice that an intention can be

also considered as an implicit recommendation. The

sentence “If we get the chance to go to Lisbon again

we will ensure that we stay at the X Hotel again” is

an example of this case.

8. Complain (C): It is a phrase expressing a

feeling of dissatisfaction, or resentment. It is

stronger than a suggestion.

9. Opinion Holder (H): The holder of a

particular opinion is the person or the organization

that holds the opinion (Ding et al., 2008). A holder is

identified with demographic characteristics, e.g.

name, city and country. For example, sites such as

tripadvisor.com or booking.com classify holders

according to types (including families with older

children, families with young children, mature

couples, groups of friends, solo travellers and young

couples among others).

10. Source (S): An information source is a web

site which contains a set of reviews. Examples

include sites such as amazon.co.uk, tripadvisor.com

and booking.com.

11. Part-of-Speech (PoS): In order to evaluate a

sentence in a review, we should consider the parts-

of-speech mentioned such as adjectives, adverbs and

verbs. Adjectives are classified as positive (e.g.

good, excellent and clean), negative (e.g. awful,

boring and terrible), neutral (e.g. regular and

indifferent) and dual, which can express positive and

negative opinion (e.g. long). In this characterization,

nouns are represented by concepts of a domain

ontology and mapped as features (described in the

item 2 - Feature).

12. Conjunction and Connective (CC):

Connectives are words that help identifying

additional adjective opinion words and their

orientations. According to Liu (2010), one constraint

is about conjunction (i.e. and), which says that

conjoined adjectives usually have the same

orientation. For example, in the sentence, “This

room is beautiful and spacious.”, if “beautiful” is

known to be positive, it can be inferred that

“spacious” is also positive, since people usually

express the same opinion on both sides of a

conjunction. Rules or constraints are also designed

for other connectives (e.g. or, but, either-or, and

neither-nor). For example, “This hotel is beautiful

but difficult to get there”, the occurrence after the

connective but is an indicator of a negative opinion.

4 PROBLEM SCENARIO

The characterization described in Section 3 extends

the definition proposed by Ding et al. (2008) and Liu

(2010) and allows a better understanding of the

complexity in processing online opinions written in

natural language. An in-depth analysis need to be

performed in order to show the relationships

between each key text segment of this

characterization. In that sense, a tool for supporting

managerial decision making should be able to

answer questions such as:

- What are the verbs more often used to describe

positive opinions on the features A, B and C?

- Which online reviews make recommendations?

Which of them are explicit?

- What types of holders give more (explict and

implicit) suggestions?

- What is the most frequent semantic orientation

in online reviews written in Italian?

A Multidomain and Multilingual Conceptual Data Model for Online Reviews Representation

- What is the co-occurrence between attitudes

and features?

In order to answer these and other questions, it is

necessary to design a conceptual data model to

represent the complex characterization previously

described. This model could be used to develop a

decision support system to help managers to analyse

the huge amount of information available on

dispersed Web 2.0 applications.

However, finding relevant and useful

information on the user generated content is a

challenge in the Web 2.0 context. From the user

point of view, it would be useful if reviews,

comments and posts from multiple information

sources could be summarized on a single view

enabling him or her to make a decision about a

specific product or service. On the other hand, from

the managerial point of view, it is important to know

specific features of a product or a service such as the

best-selling gender of a book or a specific book, and

the main negative reviews or comments about a

hotel.

In both cases, the following requirements are

addressed by the proposed model:

 To support multidomain and multilingual texts;

 To store multiple information sources;

 To support fast and simple generation of

managerial reports;

 To allow a fine-grained storage of the online

reviews’ content.

5 SOLUTION APPROACH

This section presents a Unified Modeling Language

(UML) class-based modelling approach to deal with

the UGC featured in Section 3. It introduces the

metamodel and details the main classes according to

the characterization described in Section 3.

5.1 Opinion Conceptual Data Metamodel

Proliferation of social media applications and their

increasingly complexity and sophistication claims

for the development of a generic and robust

metamodel to capture the tacit knowledge available

in those applications’ reviews. Figure 1 depicts the

metamodel proposed to capture customers’ opinions.

The class Review stores the identifier for the full

content of each review, its date and the language in

which the review is written. Each review is

composed by one or more sentences or clauses

which are stored in the class Sentence. The origin of

each review is stored in the class Source. Each

review has also a holder, who usually provides basic

data such as name, city and country. The class

Holder stores this data.

The class Object stores the name of the object,

e.g. hotel Alpha, movie Beta and book Gama, that is

reviewed. This class allows the metamodel to be

multidomain as it is possible to store reviews in a

wide range of topics. The class Feature stores the

main terms in a domain. These terms can be

provided by a domain ontology, such as staff or

room for accommodations, director or photographer

for a movie, and author or title for a bookstore.

The class PoS (Part-of-Speech) stores the parts-

of-speech, i.e. adjectives, adverbs and verbs,

contained in the sentences. The class Semantic

Orientation captures the main perspectives in the

context of an opinion. Some perspectives regards to

the global object, i.e. recommendation or intention,

while other express opinion on a specific feature, i.e.

positivity, complain, attitude, intensity and

suggestion.

The class Feature-Sentence makes the relationships

among a feature, its sentences and the existing

semantic orientation. The class Co-oc PoS stores

the relationships between parts-of-speech and

features. For example, the feature staff in the

accommodation domain can be more often

Figure 1: Opinion class metamodel.

ICSOFT 2012 - 7th International Conference on Software Paradigm Trends

mentioned with the adjective helpful in the reviews.

The metamodel presented in Figure 1 is detailed in

the next sections.

5.2 Representation of Reviews,

Holders, Objects and Features

Figure 2 presents a more detailed part of Review,

Holder, Object and Feature classes making explicit

their main attributes.

The class Review stores the identifier of each

review, date, language code in which the review is

written and the rate. The class Sentence stores the

reviews itself split in sentences. The class Holder-

Type stores the types of customers, e.g. young

couple, solo traveller and family with young

children. The class Object-Type stores the types of

objects, e.g. hotel, pension and apartment. The class

Object-Relationship captures the relationships

between objects, e.g. a hotel chain is composed by

several individual hotels. The class Feature-

Relationship captures the relationships between the

features, e.g. Swimming Pool is part of Pool.

Figure 2: Review, Holder, Object and Feature Model.

5.3 Representation of Semantic

Orientation of Features in the

Sentences

This model intends to capture several perspectives

regarding to the semantic orientation of a sentence.

Some sentences present opinions about one or more

features. In Figure 2, the classes Sentence and

Feature-Sentence stores the existing indicators to

each feature in each sentence.

The class Sentence is specialized in a

Recommendation and an Intention. An opinion

recognized as recommendation or intention is about

the global object, e.g. hotel, movie and book, rather

than a specific feature. In addition, both can express

a positive or negative opinion.

The semantic orientations of a sentence evolving

specific features are stored in the class Feature-

Sentence. A feature can be explicit or implicit in a

sentence, which is captured in the attribute

exImplicit. Sentences are also classified as the

objectivity. An objective sentence contains or

mentions facts, while a subjective sentence does not

mention any fact. The class Feature-Sentence also

captures a negative occurrence in each sentence.

Each feature in a sentence (Feature-Sentence

class) can be also associated with a feeling related

to:

 Positivity, e.g. positive, neutral and irrelevant;

 Attitude: which includes three types: affect,

appreciation and judgment;

 Intensity, e.g. very positive, neutral and very

negative;

 Complain, e.g. very serious, serious, not

serious;

 Suggestion, e.g. “breakfast could be include

more fruits”.

Figure 3: Representation of the semantic orientation of the

features in the sentences.

Some of these opinions are about a specific feature

of an object and other are generic ones. The classes

Positivity, Attitude, Intensity, Complain and

Suggestion store these kinds of opinion.

5.4 Representation of Co-Occurrences

The Opinion Model should also stores co-

occurrences between features and syntactic

categories (i.e. nouns, adjectives, adverbs, verbs and

connectives) and features. Nouns are qualifiers, e.g.

A Multidomain and Multilingual Conceptual Data Model for Online Reviews Representation

price, design and accessibility, used in the

comments. Connectives are usually conjunctions,

e.g. or, but, either-or, and neither-nor. Co-

occurrences allow managers to use the same words

to advertise the product or service. For example, if

location is more co-related with the noun or qualifier

metro, this is an indicator that customers appreciate

a hotel near a metro station. Figure 4 presents the

part of the model that deal with co-occurrences.

The model also stores co-occurrences between

features and syntactic categories, i.e. nouns,

adjectives, adverbs and verbs, in the classes named

Co-Occurrence-X-Y. The classes Noun, Adjective,

Adverb, Verb and Connective store its own the list of

terms.

It is important to notice that these classes can be

very useful to Marketing managers, since the

frequency of the co-occurrences indicates the

association done by customers to each feature of the

domain. For example, in hospitality industry,

customers associate the adjectives clean and

spacious to the feature room. With this information,

a Marketing manager can explore these words in

advertising campaigns.

Figure 4: Representation of co-occurrences detected in

online reviews.

5.5 Full Conceptual Data Model

Figure 5 presents the full conceptual data model

which integrates all the parts previously described

with four more classes. This model aims to be

sufficiently generic to serve as building blocks for

knowledge bases to store reviews and related data

from different knowledge domains.

The multidomain aspects of this model can be

explained based on the three classes: Review, Object

and Feature. From this point, it is easier to

understand the remain model and how to adapt it to

represent knowledge in different domains.

Regarding the multilingual aspect, Figure 5

shows that each review, object name and feature

name owns a multilingual representation as well as

the specific classes Holder Type and Object Type.

Each name in the syntactic category is also

represented in its multilingual form using the

attribute langCode whenever it is necessary to store

text data.

Finally, using this model, managers can also

explore historical data gathered. For example, it can

be interesting for an accommodation manager or

tourism agent to know how suggestions,

recommendations and complains have evolved over

time.

6 RECOGNIZING FACETS IN

ONLINE REVIEWS

In order to show the relevance of the facets

mentioned along this paper, we are working in

manual and automatic tasks to foster the main

classes presented in Figure 5. In both approaches

features are mapped to concepts of an ontology

which provides a common conceptualisation in a

specific domain.

Considering the manual approach, Chaves,

Gomes and Pedron (2012) analyzed a set of 1500

multilingual (English, Portuguese and Spanish)

online reviews about small and medium hotels in the

Lisbon region. They annotated each review

according to the following criteria: types of holder,

features, qualifiers (the most frequent terms used to

mention features), positivity, intensity, country of

origin of the holder and rate of review (numeric

value assigned by the holder of the review). Their

findings show that the features room and service

were the concepts that guests pay more attention to

in their review and ratings, and also points the main

features which small and medium hotel managers

should prioritize according to the profile of the guest

(classes Holder and Holder-Type in Figure 5).

Moreover, they show the co-relation between

features and their intensity (i.e. strenght of polarity)

in the reviews.

On the other hand, Tromp (2011) proposes a

automatic four-step approach for multilingual

sentiment analysis, which is composed by language

identification, POS tagging, subjectivity detection

and polarity detection. For polarity detection, he

developed an algorithm which uses heuristic rules

that stem from patterns (e.g. positive and negative).

The result of this step could be used to instantiate

the classes Sentence, Feature, Feature-Sentence and

ICSOFT 2012 - 7th International Conference on Software Paradigm Trends

Figure 5: Full conceptual data model.

Positivity in the model proposed in Figure 5.

Chaves et al. (2012) developed an algorithm

named PIRPO (Polarity Recognizer in Portuguese)

which recognizes the polarity (class Positivity in

Figure 5) expressed in each sentence of a review.

PIRPO receives as input online reviews, a list of

adjectives and concepts of an ontology. The output

is a list of sentences with polarity recognized in each

concept detected. This output is currently being used

to instantiate the classes Review, Sentence, Feature,

Feature-Sentence and Positivity. The extension of

PIRPO will allow instantiate the remaining classes,

mainly the semantic orientation of the features in the

sentences.

7 CONCLUSIONS

The representation of the unstructured data in the

web remains a challenge for the most of researchers

and practioners in modelling field. This paper

introduces a multidomain and multilingual

conceptual data model to capture the different facets

of unstructured online reviews. It also presents a

fine-grained characterization of opinionated texts,

which evidences the main aspects of those opinions

that could be relevant to managerial or user decision

making.

It is important to mention that the level of

complexity of this model varies according to the

application requirements. The full conceptual data

model can be partially or totally implemented

depending on the context in which it is going to be

used. Probably, the implmentation of the semantic

orientation of the features in the senteces is

incremental due to the lack of algorithms to

recognize all the aspects designed in the model.

The implementation of this model will also allow

the storage of the content from multiple information

sources, which will facilitate the fast and simple

generation of managerial reports. Moreover, from

the customer point of view, it will be possible to

A Multidomain and Multilingual Conceptual Data Model for Online Reviews Representation

search information about a specific object (e.g.

book, movie or hotel) in a single place.

Applications built based on the proposed model

may allow in-depth analysis of the fine-grained

knowledge dispersed in the web. However, the

success of the implementation of this conceptual

data model is also dependent on an algorithm able to

recognize the facets in the opinionated texts. The

automatic identification of the semantic orientation

of the features in the reviews remains a current

challenge for Computer Science researchers.

To better test the model in the multidomain

aspect, we should use instances from other domain

than accommodations (e.g. books, cars or movies).

Regarding to the multilingual representation, we

should automatically load the reviews annotated by

Chaves, Gomes and Pedron (2012). These

annotations cover most of the facets in the

conceptual data model.

Finally, the model proposed is in its first version

and we know that there is room for improvements.

As future work, the model will be also tested with an

application for information visualization developed

in Carvalho and Chaves (2012).

ACKNOWLEDGEMENTS

This research was partially supported by the national

funds of FCT – the Portuguese Science and

Technology Foundation within the strategic project

PEst-OE/EGE/UI4027/2011.

REFERENCES

Attensity, 2012. Available at http://www.attensity.com.

Last access: January 6, 2012.

Bai, X., 2011. Predicting Consumer Sentiments from

Online Reviews. Decision Support Systems 50(4),

March, Elsevier Science, p. 732-742.

Carvalho, E.; Chaves, M. S., 2012. Exploring User

Generated Data Visualization in the Accommodation

Sector. Proceedings of the 16

International

Conference Information Visualisation, IEEE,

Montpellier, France, 10-13 July.

Casey W., Navendu G., and Shlomo A., 2005. Using

Appraisal Groups for Sentiment Analysis. In

Proceedings of the 14

ACM International Conference

on Information and Knowledge management (CIKM

'05). ACM, New York, NY, USA, 625-631.

Chaves, M. S.; Gomes, R. and Pedron, C., 2012. Decision

making based on Web 2.0 Data: The Small and

Medium Hotel Management. Proceedings of the 20

European Conference on Information Systems,

Barcelona, Spain, 10-13 June.

Chaves, M. S.; Freitas, L. A.; Souza, M. and Vieira, R.,

2012. PIRPO: An Algorithm to deal with Polarity in

Portuguese Online Reviews from the Accommodation

Sector. Proceedings of the 17

International

Conference on Applications of Natural Language

Processing to Information Systems (NLDB),

Groningen, The Netherlands, 26-28 June.

Chaves, M. S.; Rodrigues, C. and Silva, M. J., 2007. Data

Model for Geographic Ontologies Generation.

XATA2007 - XML: Aplicações e Tecnologias

Associadas. Ramalho, José Carlos; Lopes, João

Correia and Carriço, Luís (Eds.). 15-16 February,

Lisbon, Portugal.

Chaves, M. S.; Trojahn, Cássia and Pedron, Cristiane

Drebes, 2012. A Framework for Customer Knowledge

Management based on Social Semantic Web: A Hotel

Sector Approach. In: Customer Relationship

Management and the Social and Semantic Web:

Enabling Cliens Conexus. Colomo-Palacios, Ricardo;

Varajão, João and Soto-Acosta, Pedro (Eds.). p. 141-

157, Hershey, PA: IGI Global. ISBN: 978-161-35-

0044-6

Chesley, P.; Vincent, B.; Xu, L. and Srihari R., 2006.

Using Verbs and Adjectives to Automatically Classify

Blog Sentiment. in AAAI Symposium on

Computational Approaches to Analysing Weblogs

(AAAI-CAAW), 27–29.

Clarabridge, 2012. Sentiment and Text Analytics Software

- Clarabridge. Available at http://clarabridge.com. Last

access: January 6, 2012.

Consoli, D.; Diamantini, C. and Potena, D., 2009.

Affective Algorithm to Polarize Customer Opinions.

Proceedings of the 11

International Conference on

Enterprise Information Systems, Volume HCI, ICEIS

(5), Milan, Italy, May 6-10, 157-160.

Ding, X., Liu, B., and Yu, P. S., 2008. A Holistic Lexicon-

based Approach to Opinion Mining. Proceedings of

the Conference on Web Search and Web Data Mining

(WSDM) - ACM, Palo Alto, California, USA, p. 231-

240.

EC, (2012). European Comission: Enterprise and Industry.

Small and Medium-sized Enterprises (SMEs) Fact and

Figures about the EU´s Small and Medium Enterprise.

Available at http://ec.europa.eu/enterprise/policies/sme

/facts-figures-analysis/index_en.htm. Last access:

January 8, 2012.

Hu, M.; Liu, B., 2004. Mining and Summarizing Customer

Reviews. Proceedings of the 10

ACM SIGKDD

International Conference on Knowledge Discovery

and Data Mining (KDD’04), August 22-25, Seatle,

WA, USA, p. 168-177.

Karagiannis, D.; Höfferer, P., 2006. Metamodels in

Action: An overview. Proceedings of the First

International Conference on Software Paradigm

Trends (ICSOFT), INSTICC Press, Setúbal Portugal,

September 11-14. ISBN: 972-8865-69-4

Khan, K., Baharudin, B. B., Khan, A. and Fazal_e_Malik,

2010. Automatic Extraction of Features and Opinion-

Oriented Sentences from Customer Reviews. World

Academy of Science, Engeneering and Technology,

ICSOFT 2012 - 7th International Conference on Software Paradigm Trends

Issue 62, February. ISSN:1307-6892.

Liu, B., 2010. Sentiment Analysis and Subjectivity. In

Handbook of Natural Language Processing, Second

Edition, Eds: N. Indurkhya and F. J. Damerau), CRC

Press, Taylor and Francis Group, Boca Raton, FL.

Chapter 28.

Martin, J. R. and White, P. R. R., 2007, The Language of

Evaluation, Appraisal in English. Palgrave Macmillan,

First edition, London & New York, 256 pages.

OMG, 1999. Unified Modeling Language Specification

version 1.3. Technical Report, Object Management

Group.

Schmidt, D. C., 2006. Model-driven Engineering. IEEE

Computer 39(2), February, p. 25–31.

Synthesio, 2012. Synthesio. Available at http://synthesio.

com. Last access: April 30, 2012.

SocialMetricx, 2012. Socialmetrix - Social Media

Analytics for serious decision making. Available at

http://www.socialmetrix.com. Last access: January 6,

2012.

Tromp, E., 2011. Multilingual Sentiment Analysis on

Social Media. Master’s Theisis. Department of

Mathematics and Computer Science, Eindhoven

University of Technology.

Turney, P., 2002. Thumbs Up or Thumbs Downs?

Semantic Orientation Applied tio Unsupervised

Classification of Reviews. Proceedings of the 40

Annual Meeting of the ACL, Philadelphia, July, p. 417-

424.

van Gigch, J. P., 1991. System Design Modeling and

Metamodeling. Plenum, First edition. July, 453 pages.

ISBN: 0306437406.

Whitelaw, C.; Garg, N. e Argamon, S., 2005. Using

Appraisal Groups for Sentiment Analysis. In

Proceedings of the 14

ACM International Conference

on Information and Knowledge Management (CIKM

'05). ACM, New York, NY, USA, p. 625631.

Wilson, T., 2008. Fine-Grained Subjectivity Analysis.

PhD Dissertation, Intelligent Systems Program,

University of Pittsburgh.

A Multidomain and Multilingual Conceptual Data Model for Online Reviews Representation