Today we are more and more provided with information expressing opinions about different topics. In the same way, the number of Web sites giving a global score, usually by counting the number of stars for instance, is also growing extensively and this kind of tools can be very useful for users interested by having a general idea. Nevertheless, sometimes the expressed score (e.g. the number of stars) does not really reflect what it is expressed in the text of a review. Actually, extracting opinions from texts is a problem that have been extensively addressed in the last decade and very efficient approaches are now proposed to extract the polarity of a text. In this presentation we focus on a topic related with opinion but rather than considering the full text we are interested with the opinions expressed on specific criteria. First we show how criteria can be automatically learnt. Second we illustrate how opinions are extracted. By considering criteria we illustrate that it is possible to propose new recommender systems but also to evaluate how opinions expressed on the criteria evolve over time.
Extracting opinions that are expressed in a text is
a topic that have been addressed extensively in the
last decade (e.g.
Actually this review has been scored quite well: 4 stars over 5. Any opinion mining tools will show that the review is much more negative than positive. Let us go deeper on this exemple. Even if We are here on a Saturday night and the food and service was amazing. We brought a group back the next day and we were treated so poorly by a man with dark hair.
He ignored us when we needed a table for 6 to the point of us leaving to get takeaway.
Embarrassing and so disappointing. the review is negative it clearly illustrates that the reviewer was mainly disappointed by the service: he was in the Restaurant and found it amazing. We could imagine that, at that time, the service was not so bad. This exemple illustrates the problem we address in the presentation: we do not focus on a whole text rather we would like to extract opinions related to some specific criteria. Basically, by considering a set of user-specified criteria we would like to highlight (and obviously extract opinions) only on the relevant parts of the reviews focusing on these criteria. The paper is organized as follows. In Section 2 we give some ideas on how to automatically learn terms related to a criterium. We give also some clues for extracting opinions to the criteria in Section 3. Finally Section 4 concludes the paper.
First of all we assume that the end user is
interested in a specific domain and some criteria.
Let us imagine that the domain is movie and the
two criteria are actor and scenario. For each
criterium we only need to have several keywords or
terms of the criterium (seed of terms). For instance
in the movie domain: Actor= {actor, acting,
casting, character, interpretation, role, star} and
Scenario={scenario, adaptation, narrative,
original, scriptwriter, story, synopsis}. Intuitively two
different sets may exist. The first one
corresponding to all the terms that may be used for a
criterium. Such a set is called a class. The second
one corresponds to all the terms which are used in
the domain but which are not in the class. This
set is called anti-class. For instance the term
theater is about movie but is not specific neither to
the class actor nor scenario. Now the problem
is to automatically learn the set of all terms for
a class. Using experts or users to annotate
documents is too expensive and error-prone. By the
way there are many documents available on the
internet having the terms of the criteria that can be
learned. In a practical way by using a research
engine it is easy and possible to get these documents.
For instance, the following query expressed in
Google: ”+movie +actor -scenario adaptation
narrative original -scriptwriter story -synopsis” will
extract a set of documents of the domain movie
(character +), having actor in the document and
without (character ) scenario, adaptation, etc. In
other terms we are able to automatically extract
movie documents having terms only relative to the
class actor. By performing some text
preprocessing and taking into account a frequency of a term
in a specific window of terms (see
A quite similar process may be adapted for
extracted terms used to express opinions:
adjectives, verbs and even grammatical patterns such as
<adverb + adjective > in order to automatically
learn positive and negative expressions. Then by
using the new opinion lexicon extracted we can
easily detect the polarity of a document. In the
same way by using the segmentation performed in
the previous step it is now possible to focus on
criteria and then extract the opinion for a specific
criterium. Interested reader may refer to
In the presentation we will present more in
detail the main approach. Conducted experiments
that will be presented during the talk will show
that such an approach is very efficient when
considering Precision and Recall measures.
Furthermore some practical aspects will be addressed:
how many documents? how many seed terms?
the quality of the results for different domains?
We will also show that such lexicons could also
be very useful for recommending systems. For
instance we are able to focus on the criteria that
are addressed by newspapers and then recommend
the end user only with a list of newspapers he/she
could be interested in. In the same way, evaluating
how opinions evolve over time on different criteria
is of great interest for many different applications.
Interested reader may refer to
The presented work has been done mainly during the Ph.D of Dr. Benjamin Duthil and in collaboration with Ge´rard Dray, Jacky Montmain, Michel Plantie´ from the Ecole des Mines d’Ale`s (France) and Mathieu Roche from the University of Montpellier (France).