Integrating Knowledge for Subjectivity Sense Labeling Page: 12

                                
WM:

His alarm grew.
alarm, dismay, consternation - (fear resulting from the aware-
ness of danger)
=> fear, fearfulness, fright - (an emotion experienced in an-
ticipation of some specific pain or danger (usually accompa-
nied by a desire to flee or fight))
What's the catch?
catch - (a hidden drawback; "it sounds good but what's the
catch?")
=> drawback - (the quality of being a hindrance; "he
pointed out all the drawbacks to my plan")
The following objective examples are given in WM:
The alarm went off.
alarm, warning device, alarm system - (a device that signals the
occurrence of some undesirable event)
> device - (an instrumentality invented for a particular pur-
pose; "the device is small enough to wear on your wrist"; "a
device intended to conserve water")
He sold his catch at the market.
catch, haul - (the quantity that was caught; "the catch was only
10 fish")
=> indefinite quantity - (an estimated quantity)
WM performed an agreement study and report
that good agreement (r,=0.74) can be achieved be-
tween human annotators labeling the subjectivity of
senses. For a similar task, (Su and Markert, 2008)
also report good agreement.
3 Related Work
Many methods have been developed for automati-
cally identifying subjective (opinion, sentiment, at-
titude, affect-bearing, etc.) words, e.g., (Turney,
2002; Riloff and Wiebe, 2003; Kim and Hovy, 2004;
Taboada et al., 2006; Takamura et al., 2006).
Five groups have worked on subjectivity sense la-
beling. WM and Su and Markert (2008) (hereafter
SM) assign S/O labels to senses, while Esuli and Se-
bastiani (hereafter ES) (2006a; 2007), Andreevskaia
and Bergler (hereafter AB) (2006b; 2006a), and
(Valitutti et al., 2004) assign polarity labels.
WM, SM, and ES have evaluated their systems
against manually annotated word-sense data. WM's
annotations are described above; SM's are similar.
In the scheme ES use (Cerini et al., 2007), senses
are assigned three scores, for positivity, negativity,

and neutrality. There is no unambiguous mapping
between the labels of WM/SM and ES, first because
WM/SM use distinct classes and ES use numerical
ratings, and second because WM/SM distinguish be-
tween objective senses on the one hand and neutral
subjective senses on the other, while those are both
neutral in the scheme used by ES.
WM use an unsupervised corpus-based approach,
in which subjectivity labels are assigned to word
senses based on a set of distributionally similar
words in a corpus annotated with subjective expres-
sions. SM explore methods that use existing re-
sources that do not require manually annotated data;
they also implement a supervised system for com-
parison, which we will call SMsup. The other three
groups start with positive and negative seed sets and
expand them by adding synonyms and antonyms,
and traversing horizontal links in WordNet. AB, ES,
and SMsup additionally use information contained
in glosses; AB also use hyponyms; SMsup also uses
relation and POS features. AB perform multiple
runs of their system to assign fuzzy categories to
senses. ES use a semi-supervised, multiple-classifier
learning approach. In a later paper, (Esuli and Se-
bastiani, 2007), ES again use information in glosses,
applying a random walk ranking algorithm to a
graph in which synsets are linked if a member of
the first synset appears in the gloss of the second.
Like ES and SMsup, we use machine learning, but
with more diverse sources of knowledge. Further,
several of our features are novel for the task. The
LCS features (Section 6.1) detect subjectivity by
measuring the similarity of a candidate word sense
with a seed set. WM also use a similarity measure,
but as a way to filter the output of a measure of distri-
butional similarity (selecting words for a given word
sense), not as we do to cumulatively calculate the
subjectivity of a word sense. Another novel aspect
of our similarity features is that they are particular-
ized to domain, which greatly reduces calculation.
The domain subjectivity LCS features (Section 6.2)
are also novel for our task. So is augmenting seed
sets with monosemous words, for greater coverage
without requiring human intervention or sacrificing
quality. Note that none of our features as we specif-
ically define them has been used in previous work;
combining them together, our approach outperforms
previous approaches.

12