Extracting Product Features and Opinions from Reviews - Natural Language Processing and Text Mining - page 13

each domain). The inter-annotator agreement was 82%. opine's recall on the
set of 179 features on which both annotators agreed was 73%.

Table 2.3. Precision Comparison on the Explicit Feature Extraction Task.
OPINE's precision is 22% better than Hu's precision; Web PMI statistics are respon-
sible for 2/3 of the precision increase. All results are reported with respect to Hu's.
Data Hu Hu Hu OPINE OPINE

Assess(Reviews) Assess(Reviews,Web) (Reviews)

D 1 0.75 +0.05 +0.17 +0.07 +0.19

D 2 0.71 +0.03 +0.19 +0.08 +0.22

D 3 0.72 +0.03 +0.25 +0.09 +0.23

D 4 0.69 +0.06 +0.22 +0.08 +0.25

D 5 0.74 +0.08 +0.19 +0.04 +0.21

Avg 0.72 +0.06 +0.20 +0.07 +0.22

Table 2.4. Recall Comparison on the Explicit Feature Extraction Task.
OPINE's recall is 3% lower than the recall of Hu's original system (precision level
= 0.8). All results are reported with respect to Hu's.

Data Hu Hu Hu OPINE OPINE

Assess(Reviews) Assess(Reviews,Web) (Reviews)

D 1 0.82 -0.16 -0.08 -0.14 -0.02

D 2 0.79 -0.17 -0.09 -0.13 -0.06

D 3 0.76 -0.12 -0.08 -0.15 -0.03

D 4 0.82 -0.19 -0.04 -0.17 -0.03

D 5 0.80 -0.16 -0.06 -0.12 -0.02

Avg 0.80 -0.16 -0.07 -0.14 -0.03

2.3.4 Finding Implicit Features

We now address the extraction of implicit features . The system first extracts
opinion phrases attached to explicit features, as detailed in 2.3.5. Opinion
phrases refer to properties ( e.g. , “clean” refers to “cleanliness”). When the
property is implicit ( e.g. , “clean room”), the opinion is attached to an ex-
plicit feature ( e.g. , “room”). opine examines opinion phrases associated with
explicit features in order to extract implicit properties. If the opinion phrase
is a verb, noun, or adverb, opine associates it with Quality ; if the opinion
phrase is an adjective, opine maps it to a more specific property. For instance,
if “clean” and “spacious” are opinions about hotel rooms, opine associates
“clean” with Cleanness and “spacious” with Size .
The problem of associating adjectives with an implied property is closely
related to that of finding adjectival scales [8]. opine uses WordNet synonymy
and antonymy information to group the adjectives in a set of initial clusters.
Next, any two clusters A 1 and A 2 are merged if multiple pairs of adjectives ( a 1 ,
a 2 )existsuchthat a 1 ∈ A 1 , a 2 ∈ A 2 and a 1 is similar to a 2 (an explanation