Information Technology Reference
In-Depth Information
remaining elements maintain their default ranking). In the general case, each
constraint would be assigned a probability p ( s )andopine would solve a prob-
abilistic CSP as described in [14]. We simplify the problem by only using con-
straints supported by multiple patterns in Table 2.11 and by treating them as
hard rather than soft constraints. Finding a strength-based ranking of cluster
adjectives amounts to a topological sort of the induced constraint graph. In
addition to the main opinion word, opinion phrases may contain intensifiers
( e.g. , very ). The patterns in Table 2.11 are used to compare the strength of
modifiers ( e.g. , strength ( very ) > strength ( somewhat )) and modifiers which
can be compared in this fashion are retained as intensifiers . opine uses inten-
sifier rankings to complete the adjective opinion rankings ( e.g. , “very nice” is
stronger than “somewhat nice”). In order to measure opine's accuracy on the
opinion ranking task, we scored the set of adjective opinion rankings for the
top 30 most frequent properties as follows: if two consecutive opinions in the
ranking are in the wrong order according to a human judge, we labeled the
ranking as incorrect. The resulting accuracy of opine on this task was 73%.
2.4 Related Work
The review-mining work most relevant to our research is described in [2],
[15] and [7]. All three systems identify product features from reviews, but
opine significantly improves on the first two and its reported precision is
comparable to that of the third (although we were not able to perform a direct
comparison, as the system and the data sets are not available). [2] doesn't
assess candidate features, so its precision is lower than opine's. [15] employs
an iterative semi-automatic approach which requires human input at every
iteration. Neither model explicitly addresses composite (feature of feature) or
implicit features. [7] uses a sophisticated feature extraction algorithm whose
precision is comparable to opine's much simpler approach; opine's use of
meronymy lexico-syntactic patterns is inspired by papers such as [16] and
[17]. Other systems [18, 19] also look at Web product reviews but they do not
extract opinions about particular product features.
Recognizing the subjective character and polarity of words, phrases or
sentences has been addressed by many authors, including [13, 20, 10]. Most
recently, [21] reports on the use of spin models to infer the semantic orienta-
tion of words. The chapter's global optimization approach and use of multiple
sources of constraints on a word's semantic orientation is similar to ours, but
the mechanism differs and the described approach omits the use of syntactic
information. Subjective phrases are used by [1, 22, 19, 9] and others in order to
classify reviews or sentences as positive or negative. So far, opine's focus has
been on extracting and analyzing opinion phrases corresponding to specific
features in specific sentences, rather than on determining sentence or review
polarity. To our knowledge, [7] and [23] describe the only other systems which
address the problem of finding context-specific word semantic orientation. [7]
uses a large set of human-generated patterns which determine the final se-
Search WWH ::




Custom Search