Information Technology Reference
In-Depth Information
ment of
pos
and
neg
labels to phrases which were found to be opinions (that
is, not
neutral
) after the
word SO label assignment
stage is completed.
We compared opine with
PMI++
and
Hu++
on the tasks of interest.
We found that opine had the highest precision on both tasks at a small loss in
recall with respect to
PMI++
. opine's ability to identify a word's SO label
in the context of a given feature and sentence allows the system to correctly
extract opinions expressed by words such as “big” or “small,” whose semantic
orientation varies based on context.
opine's performance is negatively affected by a number of factors: pars-
ing errors lead to missed candidate opinions and incorrect opinion polarity
assignments; other problems include sparse data (in the case of infrequent
opinion words) and complicated opinion expressions (
e.g.
, nested opinions,
conditionals, subjunctive expressions).
2.3.7 Ranking Opinion Phrases
opine clusters opinions in order to identify the properties to which they refer.
Given an opinion cluster
A
corresponding to some property, opine ranks its
elements based on their
relative strength
. The probabilities computed at the
end of the relaxation-labeling scheme generate an initial opinion ranking.
Table 2.11. Lexical Patterns Used to Derive Opinions' Relative Strength.
a,
(
∗
)
even b a,
(
∗
)
not b
a,
(
∗
)
virtually b a,
(
∗
)
almost b
a,
(
∗
)
near b a,
(
∗
)
close to b
a,
(
∗
)
quite b a,
(
∗
)
mostly b
In order to improve this initial ranking, opine uses additional Web-derived
constraints on the relative strength of phrases. As pointed out in [8], patterns
such as “
a
1
, (*) even
a
2
” are good indicators of how strong
a
1
is relative to
a
2
. To our knowledge, the sparse data problem mentioned in [8] has so far
prevented such strength information from being computed for adjectives from
typical news corpora. However, the Web allows us to use such patterns in
order to refine our opinion rankings. opine starts with the pattern mentioned
before and bootstraps a set of similar patterns (see Table 2.11). Given a cluster
A
, queries which instantiate such patterns with pairs of cluster elements are
used to derive constraints such as:
c
1
=(
strength
(
deafening
)
> strength
(
loud
)),
c
2
=(
strength
(
spotless
)
> strength
(
clean
)).
opine also uses synonymy and antonymy-based constraints, since syn-
onyms and antonyms tend to have similar strength:
c
3
=(
strength
(
clean
)=
strength
(
dirty
)).
The set
S
of such constraints induces a constraint satisfaction problem
(CSP) whose solution is a ranking of the cluster elements affected by
S
(the
Search WWH ::
Custom Search