Ontology-based Sentiment Analysis Model for Recommendation Systems

Samreen Zehra, Shaukat Wasi, Imran Jami, Aisha Nazir, Ambreen Khan and Nusrat Waheed

Department of Computer Science, Mohammad Ali Jinnah University, Karachi, Pakistan

Keywords:

Ontology, Sentiment Analysis, Recommendation System.

Abstract:

In this paper, we propose a novel approach towards developing a recommendation system using ontology-

based sentiment analysis. To conduct our study, we have targeted a Facebook closed group which contains

posts/reviews regarding different schools. For elucidating the knowledge domain, a school ontology is man-

ually designed based on a set of extracted post/comment data. Sentiment analysis is consequently performed

on the resulting Data set and the relative sentiment scores are stored back in the ontology for making recom-

mendations in future.

1 INTRODUCTION

Sentiments are very important as they make a founda-

tion for decision making for the stakeholders. Opin-

ion mining (also termed as Sentiment analysis) is a

sub-ﬁeld of web mining that deals with extracting,

analyzing and classifying the sentiments of different

people regarding a target entity and its related at-

tributes, from user-generated reviews(Liu, 2012).

Opinions are studied at three levels of granularity

(document, sentence or feature) (Freitas and Vieira,

2013). In our case, sentence level analysis is be-

ing used to predict user’s sentiment correctly in each

comment/post, that has no underlying semantic con-

text. It is important to understand that sentiment anal-

ysis can only be performed on subjective sentences

that contain some personal sentiment, stance, view, or

belief attributed towards an entity. The task of clas-

sifying a given text being objective or subjective is

referred to as Subjectivity classiﬁcation(Liu, 2012).

In our research, we have selected subjective com-

ments manually from Facebook comments ﬁle. Sen-

tiStrength tool was then employed for discovering the

polarity of a given subjective comment, i.e. neutral,

positive or negative and to assign a numeric score

that expresses the strength/intensity of the sentiment.

Considering the various hurdles in conducting senti-

ment analysis of unstructured texts, the researchers

have started exploring semantic-based approaches for

opinion mining. This is mostly achieved by utilizing

an ontology. Ontology is a formal explicit descrip-

tion of concepts in a domain of discourse in the form

of taxonomic hierarchy (Noy et al., 2001). It devel-

ops a shared understanding of domain by building a

common vocabulary lexicon. An Ontology comprises

of concepts, instances, relationships between the con-

cepts, facets and slots. For implementation of our on-

tology, we have used Prot´eg´e, a JAVA based knowl-

edge base framework that complies with W3C stan-

dards of representing ontology in the semantic web.

For opinion mining, it is very challenging to iden-

tify the target of the opinion in a given text. This

problem is known as Named Entity Recognition in

Natural Language Processing. The task of ﬁnding

the target/feature may be eliminated if the knowledge

domain is modeled as an ontology, based on prede-

termined relationships between domain objects and

its features, having speciﬁc rules and axioms. Dur-

ing the past few years, the social media has emerged

as a pivotal source for opinion mining as different

users share their experiences and feedback on it. It

is worth noting that mining user’s perspective, from

these short texts/comments, regarding a particular tar-

get is a demanding job given the fact that these com-

ments/reviews are characterized as ambiguous, sar-

castic and informal. Presently, so much work has al-

ready been done on sentiment analysis of a normal un-

structured text corpus using lexicon-based approach

and machine learning approach or the combination of

two.

In our proposed work, we are trying to study sen-

timents of a particular Facebook group users towards

different schools to formulate a recommendation sys-

tem based on their reviews. A school ontology has

been developed to model the school knowledge-base

for extraction of relevant opinions and polarity cal-

Zehra S., Wasi S., Jami I., Nazir A., Khan A. and Waheed N.

Ontology-based Sentiment Analysis Model for Recommendation Systems.

DOI: 10.5220/0006491101550160

In Proceedings of the 9th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (KEOD 2017), pages 155-160

ISBN: 978-989-758-272-1

culation. Hence, it is highly likely that a person

will have a recommendation based on its require-

ment. This paper is organized as follows: Section

II presents Related Work; Section III describes The

Proposed Approach; Section IV explains the Method-

ology; Section V highlights Experimental Results and

Section V outlines Conclusions and Future Work.

2 RELATED WORK

In the literature, similar work in this domain is car-

ried out by (Kontopoulos et al., 2013) where the

researchers applied ontology-based sentiment analy-

sis methodology on twitter data. They worked on

semi-automatic ontology generation based on twitter

posts about different mobile phones and their associ-

ated features. Separate Datasets comprising of Users’

reviews were used for ontology creation and senti-

ment analysis. In their work the sentiment scores

were computed using a webservice named Open-

Dover. However, the idea of application of senti-

ment analysis on an ontology-based knowledge do-

main was originally proposed by (Yaakub et al., 2012)

which was further analyzed and extended by (Haider,

2012) to include feature-based sentiment analysis of

customers’ reviews on smartphones having enhanced

product feature set. Both of them classiﬁed the sen-

timent polarity on a 7-point scale ranging from 3 to

-3. They also asserted that the feature level sentiment

scores help in the calculation of the overall sentiment

score for the object. In a similar context, feature level

opinion mining is performed by (Freitas and Vieira,

2013) for Portuguese movie reviews, by making use

of already available movie ontology. Another notice-

able work by (Sam and Chatwin, 2015) suggested that

the polarity of the sentiment word heavily depends on

the context in which it is being used. They, there-

fore, developed two different ontologies. One is emo-

tion ontology for storing emotion words of customers

towards different electronic products with associated

polarities, the other is the ontology for electronic

product domain. (Thakor and Sasi, 2015) applied sen-

timent analysis on negative tweets for a postal service

to ﬁguring out the reason for customer dissatisfaction.

In his research, he practiced sentimental analysis us-

ing SentiStrength tool and data was retrieved using

Prot´eg´e SPARQL from ontology.(Alkadri and ElKo-

rany, 2016) used NLP techniques like tokenization,

De-noise, Stemming and POS-Tagging for data pre-

processing and combined three (3) different Arabic

polarity lexicons to form a large-scale Arabic opinion

lexicon (ArOpL) to identify Polarity in Arabic lan-

guage. As mentioned by (Binali et al., 2009), opin-

ion mining research comprises of both feature-level

opinion mining and its related sentiment classiﬁca-

tion. Therefore, two similar objects can be compared

based on their overall sentiment scores. Likewise,

feature level comparison may also be made between

two objects based on their corresponding feature level

sentiment scores.

In the current state-of-the-art, ontology is being

employed for sentiment analysis but the work is not

being extended to frame a recommendation system.

Hence we claim in this research that our effort is

different from similar work already done in this do-

main as we store the calculated sentiment scores in

the same ontology on corresponding hierarchical lev-

els(i.e. class level and object level) for making rec-

ommendations. In this way, we can avoid recalcula-

tion of sentiment scores each time the user submits

a query. However, it requires ontology modiﬁcation

after its formal conceptualization. It is assumed that

the proposed system may also be applied to different

domains after careful modiﬁcations in ontology and

sentiment lexicon. Table 1 presents a detailed com-

parison of the related work already done in the do-

main with our work.

3 PROPOSED APPROACH

In our proposed approach the ontology is engineered

using user comments. We have deliberately selected

a speciﬁc group where people share their feedback re-

garding different schools. It helped us in simplifying

the task of subjectivity classiﬁcation. After extract-

ing reviews regarding target school branches in a ﬁle,

sentiment analysis is performed and the resulting sen-

timent scores are stored back into the ontology. Later

on, the user can seek a recommendation from the sys-

tem through a user interface where he/she can ﬁnd

the best school based on his/her requirements(like lo-

cation, gender-wise orientation).The back-end of the

interface translates user request into a query to get

the branch level/school level scores from the ontol-

ogy and gives a recommendation after comparison of

sentiment scores.

Figure 1 depicts the proposed approach to our re-

search problem having the sequence below.

1. A school domain ontology is manually created

after reviewing Facebook posts/comments using

Prot´eg´e modeling environment.

2. Using ontology created in previous step, respec-

tive objects/classes are extracted which are targets

for our sentiment analysis.

3. Opinionated comments/posts are extracted from

Table 1: GAP Analysis.

Research paper Ontology creation Sentiment classiﬁ-

cation Approach

Querying Ontology Making Recom-

mendation

(Kontopoulos et al.,

2013)

An ontology en-

gineer developed

ontology manually

based on sample

tweets extracted from

twitter. Another

alternative semi-

automatic approach

is also discussed

using OntoGen tool

OpenDover web ser-

vice is used for calcu-

lating polarity score

Not discussed Not discussed

(Thakor and Sasi,

2015)

Data is extracted

from twitter and

cleansed for input

into GATE software.

Results of GATE

software are used

for manual Ontology

creation in Protege

Only negative tweets

are considered for

Sentiment analysis.

Sentiment polarity

is assigned using

SentiStrength tool.

SPARQL for Prot´eg´e

is used for feature

querying. The query

is written by hand

Not discussed

(Haider, 2012) Ontology is created

manually based on

product reviews for

certain mobile mod-

els

Opinion polarization

score is calculated

manually, from a set

of reviews.

Not discussed Not discussed

(Freitas and Vieira,

2013)

Existing movie ontol-

ogy and hotel ontol-

ogy is utilized

Associated polar-

ity is calculated

in real-time using

OpLexicon

Not discussed Not discussed

(Alkadri and ElKo-

rany, 2016)

Existing hotel ontol-

ogy HONTOLOGY

is transalated into

Arabic and further

extended

Associated polarity is

assigned by expert

annotators by hand

Not discussed Not discussed

(Sam and Chatwin,

2015)

Two ontologies are

created manualy.

Emotion Ontology

is used for listing

emotion words and

Elecronic Product

Ontology is designed

for electronic product

reviews

Using Ngram and

Emotion ontology,

sentiment analysis is

performed

Not discussed Not discussed

This Work Ontology is created

manually based on

Face book users’

comments/posts

Associated polarity

is being calculated

using SentiStrength

tool and stored in

Ontology

SPARQL for Prot´eg´e

is used for feature

querying. The query

is automatically

generated behind the

formal user interface

with the use of a

mapping database

Recommendations

are made using the

previously calculated

sentiment score

Facebook where the required object is found.

4. Before polarity calculation, text is processed to

remove special characters(hashtag,@) and web

URLs from the text. User comments are stored

in a separate ﬁle for each school branch to apply

sentiment analysis.

5. SentiStrength API is used to ﬁnd out the polarity

score of each user comment on an eleven point

scale ranging from -5 to 5 and update the polarity

table as shown in Table 2.

6. Using a polarity table calculate an overall polar-

ity value for the object. The calculated sentiment

score is saved back into the ontology on object

level and class level for future referencing.

Figure 1: Proposed Approach.

For generating a recommendation request, a for-

mal application user interface is designed that queries

the ontology with the help of a mapping database. We

use SPARQL query for the retrieval of desired classes

and objects that fulﬁlls user request from an ontol-

ogy. Each selected instance object is compared with

others, taking into account class level and object level

sentiment scores. As a result, a valid recommendation

is generated.

4 METHODOLOGY

4.1 Ontology Engineering

The school ontology is created manually using

Prot´eg´e. School is the base entity/class for all the

schools. All the schools are subclasses of the school

entity/class. Whereas respective branches of each

school are the real instances of any subclass of school.

This has been done considering each branch of one

school may have different Locations, GenderwiseOri-

entation like Boys Campus, Girls Campus etc. Simi-

larly it is possible that one branch is primary and the

other is secondary and like.

In Figure 2. an abstract school ontology is de-

picted containing fewer schools and their branches.

These schools and their branches are the ones which

are mostly discussed on the target forum and for

which sentiments are mostly expressed. The names

of schools are replaced with generic names for ob-

jective reasons. One may note that the data ﬁeld

SentiScoreMain relates to the sentiment score of a

Class(School), however, SentiScore data ﬁeld is for

Figure 2: School Ontology.

storing the same for an object(School Branch).

4.2 Sentiment Analysis

JENA (a JAVA based API for ontology manipula-

tion), is utilized to fetch all the relevant classes(school

names) and their objects (School Branches) from the

School ontology for which comments are to be re-

trieved from the Facebook. Facebook4J is utilized to

fetch ﬁrst m comments from Facebook. These com-

ments are stored in a comments ﬁle. For each object

o, randomly n subjective sentences are extracted from

the comments list into a separate ﬁle manually, where

object o is the target of opinion.

On completion of pre-processing phase, result-

ing comments are submitted to SentiStrength (another

JAVA API for measuring sentiments on the scale of -

5 to 5). Each sentiment related to an object(School

Branch) is stored at object level and the class(School)

is stored on class level. Algorithm 1 details the logic

for generation of polarity table which will be utilized

for Sentiment score calculation. Here, polarities cor-

responding to each object are being stored in a table

row.

4.3 Polarity Calculation

Given the polarity row for each object, a sentiment

score is the weighted average of all the polarity counts

stored against each polarity value for a given object.

Equation(1) will be used to calculate the sentiment

score F(o) for an object o.

F(o) =

∑

z=−5

(z∗ Polarity(o, z))/n (1)

Where n=Total no of opinionated comments for ob-

ject o and Polarity(o,z)is the frequency of Z polarity

Algorithm 1: Polarity table generation.

Input: Ontology with Classes(E),

Attributes(A) and Objects(O)

Parameters: Empty set of Facebook

Comments (C)

Output: Sentiment polarity table P with each

row corresponds to each object o ∈ O

and having eleven columns

corresponding to each polarity value

z ∈ Z (-5 to 5)

1 foreach e ∈ E which has a direct object do

2 if e has object then

3 foreach o ∈ O which is instance of

Class e do

4 //Retrieve m comments from

Facebook

5 C ← RetrieveComments(m)

6 //Copy ﬁrst n comments to Array S

having object o in it and contains

some user’s sentiment

7 S ← CopyComments(n)

8 Insert a row in P for object o and

initialize cell values to 0

9 foreach s ∈ S do

10 r ← GetPolarity(s)

P(o,z) = P(o,z) + 1 ; such that

r = z

11 end

12 end

13 end

14 end

15 return P;

value for object o in given set of comments where ob-

ject o is the subject of opinion. It should be noted here

that value of n remains the same for all the objects to

fairly calculate the sentiment score.

G =

∑

i=1

F(o

)/n (2)

Similarly, in a scenario where sentiments relating to

objects(School Branches) are expressed only, Equa-

tion(2) can be used to ﬁnd an average sentiment score

G for any class (School) itself, where o

,..., o

are

objects of that class.

5 EXPERIMENTAL RESULTS

We have considered four school branches for experi-

mental purpose. For each school branch, we extracted

randomly 250 opinionated comments from the Face-

book group. A separate ﬁle was created for each

school branch and each ﬁle contains comments sep-

arated by newline character. Each ﬁle is then pro-

cessed through SentiStrength tool, hence, polarity ta-

ble is generated. Each row of the table is associated

with an individual School branch. The resulting Po-

larity table is shown in Table 2.

Table 2: Polarity table.

School (O) Polarity

-5 -4 -3 -2 -1 0 1 2 3 4 5 Polarity Score

School1- Branch 1 0 2 3 4 10 1 25 33 38 54 80 3.144

School1-Branch 2 1 2 5 7 5 0 29 40 44 51 66 2.912

School2 -Branch 1 8 4 34 48 52 3 36 27 20 13 5 -0.316

School3-Branch 1 0 1 4 8 6 1 40 33 41 66 50 2.82

If a user is interested in all school on a speciﬁc

location

, then the comparison will be made among

all the school branches which are located at

and the

one with the highest sentiment score will be recom-

mended.

Table 3: School-level Polarity Table.

School (O) Calculation Polarity

School 1 (Sentiment Score (Branch 1)+ Sentiment Score (Branch 2))/2 (3.144+2.912)/2 = 3.028

School 2 Sentiment Score (Branch 1) -0.316

School 3 Sentiment Score (Branch 1) 2.82

Similarly, if a user is asking that which school is

better than the other, then the overall sentiment score

of the school will be taken into account. Table 3

shows the calculation for different schools.

6 CONCLUSION AND FUTURE

WORK

These days recommender companies are moving to-

wards the semantic-based approaches for developing

efﬁcient recommendation systems (Bernstein et al.,

2016). This motivated us to utilize an ontology for

knowledge domain modeling and to perform senti-

mental analysis on user-generated reviews to develop

a recommendation system. In the proposed system,

calculated sentiment scores are stored in the ontology.

This characteristic distinguishes it from other systems

where scores are not saved anywhere. The scope of

our future work involves considering Facebook likes

into the calculation of sentiment score. Additionally,

we look forward to implementing feature-level senti-

ment analysis between two schools and automate the

task of subjectivity classiﬁcation.

REFERENCES

Alkadri, A. M. and ElKorany, A. M. (2016). Semantic

feature based arabic opinion mining using ontology.

International Journal of Advanced Computer Science

and Applications, 7(5):577–583.

Bernstein, A., Hendler, J., and Noy, N. (2016). A new look

at the semantic web. Commun. ACM, 59(9):35–37.

Binali, H., Potdar, V., and Wu, C. (2009). A state of the art

opinion mining and its application domains. In Indus-

trial Technology, 2009. ICIT 2009. IEEE International

Conference on, pages 1–6. IEEE.

Freitas, L. A. and Vieira, R. (2013). Ontology based fea-

ture level opinion mining for portuguese reviews. In

Proceedings of the 22nd International Conference on

World Wide Web, pages 367–370. ACM.

Haider, S. Z. (2012). An ontology based sentiment analysis:

A case study.

Kontopoulos, E., Berberidis, C., Dergiades, T., and Bassili-

ades, N. (2013). Ontology-based sentiment analysis

of twitter posts. Expert systems with applications,

40(10):4065–4074.

Liu, B. (2012). Sentiment analysis and opinion mining.

Synthesis lectures on human language technologies,

5(1):1–167.

Noy, N. F., McGuinness, D. L., et al. (2001). Ontology

development 101: A guide to creating your ﬁrst ontol-

ogy.

Sam, K. M. and Chatwin, C. (2015). Ontology-based sen-

timent analysis model of customer reviews for elec-

tronic products. In Encyclopedia of Information Sci-

ence and Technology, Third Edition, pages 892–904.

IGI Global.

Thakor, P. and Sasi, S. (2015). Ontology-based sentiment

analysis process for social media content. Procedia

Computer Science, 53:199–207.

Yaakub, M. R., Li, Y., Algarni, A., and Peng, B. (2012). In-

tegration of opinion into customer analysis model. In

Proceedings of the The 2012 IEEE/WIC/ACM Inter-

national Joint Conferences on Web Intelligence and

Intelligent Agent Technology-Volume 03, pages 164–

168. IEEE Computer Society.