1962)). Thus, the most frequent evaluation of
of a sentence within a text with its degree of in-
the product (which is the most uninformative
formativeness.
one) also should surface at the end of the text
due to the ease of its retrieval, which is pre-
Giora (Giora, 1985, Giora, 1988) defined a
sumably what product review readers would re-
discourse topic (DT) as the least informative
fer to as “the bottom line”.
(most uninformative) yet dominant proposition
To the best of our knowledge, this psycholin-
of a text. The DT best represents the redun-
guistic prediction has not been supported by psy-
dancy structure of the text. As such, this propo-
cholinguistic evidence to date. However, it has
sition functions as a reference point for process-
been somewhat supported by the computational
ing the rest of the propositions. The text posi-
tion which best benefits such processing is text
results of Yang, Lin and Chen (Yang et al.,
initial; it facilitates processing of oncoming
2007a, Yang et al., 2007b) who classified emo-
propositions (with respect to the DT) relative to
tions of posts in blog corpora. Yang, Lin & Chen
when the DT is placed in text final position.
realized that bloggers tend to emphasize their
feelings by using emoticons (such as: ☺," and
Furthermore, Giora and Lee showed (Giora
#) and that these emoticons frequently appear in
and Lee, 1996) that when the DT appears also
at the end of a text it is somewhat information-
final sentences. Thus, they first focused on the
ally redundant. However, functionally, it plays a
last sentence of posts as representing the polarity
role in wrapping the text up and marking its
of the entire posts. Then, they divided the posi-
tive category into 2 sub-categories - happy and
boundary. Authors often make reference to the
DT at the end of a text in order to summarize
joy, and the negative category - into angry and
sad. They showed that extracting polarity and
and deliberately recapitulate what has been
consequently sentiments from last sentences out-
written up to that point while also signaling the
performs all other computational strategies.
end of discourse topic segment.
4 Method
3 Polarity-classification vs. Topic-
extraction
We aim to show that the last sentence of a cus-
tomer review is a better predictor for the polarity
When dealing with polarity-classification (as
of the whole review than any other sentence (as-
with topic-extraction), one should again identify
suming that the first sentence is devoted to pre-
the most uninformative yet dominant proposi-
senting the product or service). To test our pre-
tion of the text. However, given the cognitive
diction, we ran two experiments and compared
prominence of discourse final position in terms
their results. In the first experiment we exam-
of memorability, known as “recency effect” (see
ined the readers’ rating of the polarity of reviews
below and see also (Giora, 1988)), we predict
in their entirety, while in the second experiment
that when it comes to polarity-classification, the
we examined the readers’ rating of the same re-
last proposition of a given text should be of
views based on reading single sentences ex-
greater importance than the first one (contrary
tracted from these reviews: the last sentence or
to topic-extraction).
the second one. The second sentence could have
Based on preliminary investigations, we sug-
been replaced by any other sentence, but the first
gest that the DT of any customer review is the
one, as our preliminary investigations clearly
customer’s evaluation, whether negative or
show that the first sentence is in many cases de-
positive, of a product that s/he has purchased or
voted to presenting the product or service dis-
a service s/he has used, rather than the details of
cussed and does not contain any polarity con-
the specific product or service. The message
tent. For example: "I read Isaac’s storm, by Erik
that customer reviews try to get across is, there-
Larson, around 1998. Recently I had occasion to
fore, of evaluative nature. To best communicate
thumb through it again which has prompted this
this affect, the DT should appear at the end of
review…..All in all a most interesting and re-
the review (instead of the beginning of the re-
warding book, one that I would recommend
view) as a means of recapitulating the point of
highly.” (Gerald T. Westbrook, “GTW”)
the message, thereby guaranteeing that it is fully
understood by the readership.
Bạn đang xem 1962) - BÁO CÁO KHOA HỌC LAST BUT DEFINITELY NOT LEAST ON THE ROLE OF THE LAST SENTENCE IN AUTOMATIC POLARITY CLASSIFICATION PDF