Information Technology Reference
In-Depth Information
of adjective similarity is given below). For example,
A
1
=
{
“intuitive”
}
is
merged with
A
2
=
.
Clusters are labeled with the names of their corresponding properties (see
Table 2.6). The property names are obtained from either WordNet (
e.g.
,
big
is
a value of
size
), or from a name-generation module which adds su
xes (
e.g.
,
“-iness”, “-ity”) to adjectives and uses the Web to filter out non-words and
highly infrequent candidate names. If no property names can be found, the
label is generated based ona djectives: “beIntercontinental,” “beWelcome,”
etc.
Adjective Similarity
The adjective similarity rules in Table 2.5 consist
of WordNet-Based rules and Web-Based rules. WordNet relationships such as
pertain
(
adjSynset, nounSynset
)and
attribute
(
adjSynset, nounSynset
)are
used to relate adjectives to nouns representing properties: if two adjectives
relate to the same property or to related properties, the two adjectives are
similar. In addition to such WordNet-based rules, opine bootstraps a set
of lexical patterns (see 2.3.7 for details) and instantiates them in order to
generate search-engine queries which confirm that two adjectives correspond
to the same property. Given clusters
A
1
and
A
2
, opine instantiates patterns
such as “
a
1
, (*) even
a
2
“with
a
1
∈
{
“understandable”, “clear”
}
A
2
in order to check if
a
1
and
a
2
are similar. For example, hits
(“clear, (*) even intuitive”) >
5, therefore
“clear” is similar to “intuitive.”
A
1
and
a
2
∈
Table 2.5. WordNet-Based and Web-Based Adjective Similarity Rules.
Notation:
s
1
,
s
2
= WordNet synsets
adj
1
and
adj
2
are similar if
∃s
1
,s
2
s.t.
pertain
(
adj
1
,s
1
)
, attribute
(
adj
2
,s
2
)
,isA
(
s
1
,s
2
)
∃s
1
,s
2
s.t.
pertain
(
adj
1
,s
1
)
, pertain
(
adj
2
,s
2
)
,isA
(
s
1
,s
2
)
∃s
1
,s
2
s.t.
attribute
(
adj
1
,s
1
)
, attribute
(
adj
2
,s
2
)
,isA
(
s
1
,s
2
)
∃p ∈{
“[
X
]
,even
[
Y
]
,
“[
X
]
,almost
[
Y
]
, ...}
s.t.
hits
(
p
(
adj
1
,adj
2
))
>t,t
=
threshold
Table 2.6. Examples of Labeled Opinion Clusters
Quality
:
like, recommend, good, very good, incredibly good, great, truly great
Clarity
:
understandable, clear, straightforward, intuitive
Noise
:
quiet, silent, noisy, loud, deafening
Price
:
inexpensive, affordable, costly, expensive, cheap
Given an explicit feature
f
and a set of opinions associated with
f
which
have been clustered as previously described, opine uses the opinion clus-
ters to extract
implicit features
. For example, given
f
=
Room
and opinions
clean, spotless
in the
Cleanness
cluster, opine generates the implicit feature
RoomCleanness
. We evaluated the impact of
implicit feature extraction
in the
Hotels and Scanners domains.
1
Implicit features led to a 2% average increase
1
Hu's datasets have few implicit features and Hu's system doesn't handle implicit
feature extraction.
Search WWH ::
Custom Search