Computational Intelligence - February 2014 - 21

common sense computing, which exploits both computbearing sentiments in the text are classified. Sentence-based SA
er and social sciences to better recognize, interpret, and
can be fine-grained when specific contexts within a sentence are
process opinions and sentiments over the Web."
considered. We observe that fine grain enables a third variety of
One of the goals of sentic computing is to bridge the
SA: aspect-based, where topics are associated to different sentisemantic gap between word-level natural language data and the
ments according to the considered aspect of a topic.
concept-level opinions conveyed by these so as to provide a
SA can be generalized as the analysis of opinions: an opinion
more efficient passage from unstructured textual information
is an intentional statement by somebody (holder) on some fact
to structured machine-processable data.
(topic) that is expressed with a possible sentiment. Given a docuFollowing the same direction, this article presents a novel
ment or a set of documents, they may contain opinions about
computational intelligence method for SA, focused on repremultiple topics that are interesting to users: in this scenario it is
senting the semantics of a sentence by modeling the roles
very important to be able to correctly detect topics, and to sepaplayed by its elements with respect to a model of opinion senrate the opinions associated with each topic [9].
tences. We show, by means of a set of examples and an evaluaOne problem that occurs in current SA approaches is the
tion, that this representation allows SA to precisely identify
incorrect identification of opinion topics. In many recent SA
opinion holders and topics, hence providing the basis for
systems such as [1; 13; 30], the user is prompted to just insert a
improving the accuracy of sentence- and aspect-based SA. Our
topic in order to obtain a sentiment. Then the system returns a
method combines NLP with knowledge representation, but we
list of opinions on the specified topic. However, some of the
also envision the application of fuzzy logic, combinatorial optireturned results are useless to the user because they are out of
mization, and analogical reasoning to arrive at a full-fledged
the domain the user is interested in. For example, one user
sentic computing solution.
could look for opinions about Berlusconi, and get several results
Role assignment in our opinion model relies on the exploitarelated to Berlusconi in politics, business, criminal investigation of fine-grained semantics, combined with frame-based lexitions, or football. A different example based on a similar ratiocal and syntactic analysis. It has been proved in [36; 21; 10] that
nale is a user asking about opinions on the Italian Article 18
the quality of SA algorithms improves when considering seman(Italy's labour code). In this case, opinions and sentiments can
tic features. However, in these
be expressed about e.g. its aboliworks (as in many others that
tion, or in general about the
focus on the analysis of
Article itself.
tweets), topic detection is
A SA system should
performed by relying
be able to differentiate
on named entity recthose possibilities as
ognition or by assumwell. This is practicable
ing that sentences are
only if a system is able
topic-based, which can
to properly distinguish
prove efficient only in
the main topics from
such specific settings.
the sub-topics of an
Existing approaches are
opinion, together with its
mainly based on supervised
holders. For both examples, it
© wikimedia commons
techniques. Although they prove
would be important to distinguish
very effective in many tests, they have three
opinions that are expressed on Berlusconi or on
main weaknesses: reliance on manually labelled samples, such as
Article 18 as main topics, from those expressed about events or
movies and products reviews; need of opinion expressed with
situations involving Berlusconi or Article 18, or other accessory
explicit sentiment; limited portability, as their performance can
topics. In the latter case, they would be sub-topics, hence the
change significantly when a same method is applied to a differopinion considered indirect. For example, if one expressed a
ent domain. On the contrary, our method is domain-indepenpositive opinion on the abolition of Article 18, this would be
dent, meaning that it is completely automatic and unsupervised:
indirectly a negative opinion on the Article 18. Most existing
it does not need a training phase. It performs a semantic interSA systems would however recognize Article 18 as a topic with
pretation of sentences according to a formal opinion model (cf.
a positive opinion, hence providing the wrong analysis for the
Section 3), and is able to identify relevant entities of an opinion
sentiment, due to the incorrect identification of the main topic.
regardless of the presence of explicit expressions of sentiment in
In fact, several works [28; 12; 39; 21] show that considering
a text. We show that our method is able to detect opinion holdtopics jointly with sentiment features improves the perforers with an accuracy of 95%, opinion main topics with 68%, and
mance of SA systems, although those works do not focus on
opinion subtopics with 78% (F1 measure).
topic detection as a specific task. A first step toward addressing
Our contribution can be summarized as follows:
this issue has been done by sentic computing, defined as [6]:
"...A multi-disciplinary approach to sentiment analy❏ a new challenging task: detection of opinion holders, main
sis at the cross-roads between affective computing and
topics and sub-topics;

February 2014 | Ieee ComputatIonal IntellIgenCe magazIne

Table of Contents for the Digital Edition of Computational Intelligence - February 2014