Information Technology Reference
In-Depth Information
used synonymously [
36
]. The task of sentiment analysis may be decomposed in
subjectivity detection and polarity (or orientation) classification. The goal of subjec-
tivity detection is to distinguish objective from opinion-oriented text. While objective
parts solely report facts without any personal assessment or evaluation, opinionated
text parts may contain any type of subjective expressions that reflect the private
state of a holder. This includes personal attitudes, views, statements, feelings, etc.,
expressed by the text's author or by other people mentioned or cited in the text. Hav-
ing detected a subjective text part, the type and orientation of the expressed opinion
must be determined. The most common orientations are
positive
and
negative
.For
example, a literature critic may conclude that a reviewed topic is excellent, which
can be regarded as a positive opinion toward the topic. Other classification schemes
distinguish between supporting and opposing expressions. This variant of sentiment
classification allows to contrast different points of view toward topics or political
issues.
Our definition of opinion is driven by Liu and Zhang [
28
]. Following the authors,
we define opinion as a quintuple consisting of a
•
target entity
(e.g., a product, an individual, a topic)
•
target aspect of the entity
(e.g., product features, subtopic)
•
orientation
(e.g., positive/negative/neutral)
•
holder
(the entity holding the opinion)
•
time
(when the opinion was expressed)
In order to gain a valuable opinion information, not all parts of the quintuple
must necessarily be extracted. The perception of a movie in the media, e.g., can be
inferred without knowing when the review was submitted or by whom. However, the
extraction of a target entity and the valuation of the entity is essential. The opinion
target may be a mentioned (named) entity or concept with a concrete text-anchor, but
also an abstract topic that makes the identification of the opinion target especially
difficult. In contrast to concrete entities, abstract topics are nonlocal information that
have to be mined from the context. In our scenario, which is sentiment analysis based
on newspaper quotations, the structure of quotations already provides an opinion
holder. Knowing who uttered a quotation, we consider the quotation holder being
also the opinion holder.
Our Contribution
. We present work that aims at determining opinion orientation
in German news articles. In contrast to many other sentiment analysis systems we
focus our approach on direct and reported speech identified previously in the news
articles. We assume that quotations are the most subjective parts of news articles
and that they transport the opinions of the cited speakers as they are. We cast the
task of sentiment classification to a three-class problem and label each quotation as
either negative, positive, or neutral. Objective quotations reflecting facts are marked
as neutral as well. We propose a supervised approach where we first search for
subjective quotations and, second, decide the polarity of the subjective quotations.
Both steps are separately solved by a Support Vector Machine classifier. As part of
our work we examine the effectiveness of diverse sentiment classification features
Search WWH ::
Custom Search