Learning Subjective Adjectives From Corpora

1
Learning Subjective Adjectives From Corpora

• Janyce M. Wiebe
• New Mexico State University

Office of Naval Research grant N00014-95-1-0776.
2
Introduction

• Learning evaluation and opinion clues
• Distributional similarity process
• Refinement with lexical features
• Improved results from both

3
Subjectivity Tagging

• Recognizing opinions and evaluations
• (Subjective sentences) as opposed to
• material objectively presented as true
• (Objective sentences)

Banfield 1985, Fludernik 1993, Wiebe 1994
4
Examples

• At several different levels, its a fascinating
  tale. subjective
• Bell Industries Inc. increased its quarterly to
  10 cents from 7 cents a share.
• objective

5
Types
Complained You Idiot! Terrible product
6
Subjectivity
?
Complained You Idiot! Terrible product
Speculated Maybe
7
Subjectivity

• Same word, different types
• Great majority objective
• Great! positive evaluative
• Just great. negative evaluative

8
Subjectivity

• Multiple types, sources, targets
• Its the best!, he gushed.

-
Writer He It
9
Applications Flame Recognition
10
Review Mining
From HoodoogthoodooBUGZAPPER_at_newnorth.netgt Newsgro
ups rec.gardens Subject Re Garden software I
bought a copy of Garden Encyclopedia from
Sierra. Well worth the time and money.

11
Information Extraction

• Northwest Airlines settled the remaining
  lawsuits, a federal judge said. objective
• The cost of health care is eroding our standard
  of living and sapping industrial strength,
  complains Maher. subjective

12
Other Applications

• Clustering documents by ideology
• Text summarization
• Style in machine translation and generation

13
Overview

• Identify large set of candidate clues
• Existing resources are not sufficient
• Not consistently marked for subjectivity
• Not customized to the genre
• Learn lexical clues from corpora

14
Corpus and Annotation

• Subjectivity tags assigned by multiple annotators
  to 1001 WSJ sentences
• Tags representing consensus opinions obtained
  with EM

Wiebe et al. 1999 Bruce Wiebe 1999
15
Adjectives

• Classifications correlated with adjectives
• Adjectives extracted from annotations
• They promised e 2 yet more for
• e 3 really good e? 1 stuff.
• "It's e? 3 really e- 3 bizarre," says Albert
  Lerman, creative director at the Wells agency.

16
Lins Distributional Similarity
Word R W I R1
have have R2 dog brown R3 dog
. . .
Lin 1998
17
Lins Distributional Similarity
Word1
Word2
R W R W R W
R W R W R W
R W R W
R W R W
R W R W
18
Bizarre
strange similar scary unusual
fascinating interesting curious tragic
different contradictory peculiar silly sad
absurd poignant crazy funny comic
compelling odd
19
Bizarre
strange similar scary unusual
fascinating interesting curious tragic
different contradictory peculiar silly sad
absurd poignant crazy funny comic
compelling odd
20
Bizarre
strange similar scary unusual
fascinating interesting curious tragic
different contradictory peculiar silly sad
absurd poignant crazy funny comic
compelling odd
21
Bizarre
strange similar scary unusual
fascinating interesting curious tragic
different contradictory peculiar silly sad
absurd poignant crazy funny comic
compelling odd
22
Good

• bad better best nice poor
• terrific great decent lousy
• dismal excellent positive exciting
• fantastic marvelous strong
• important dumb fair healthy

23
Good

• bad better best nice poor
• terrific great decent lousy
• dismal excellent positive exciting
• fantastic marvelous strong
• important dumb fair healthy

24
Experiments
25
Experiments
Separate corpus
Distributional similarity
Seeds
26
Experiments
Separate corpus
Distributional similarity
Seeds
Filtering
S gt Adj gt Majority
27
Lexical features

• Polarity and Gradability
• Learned from corpora
• Statistical processing informed by linguistic
  insights
• Different data sets used

28
Gradability
More additional Very additional
Hatzivassiloglou Wiebe 2000
29
Polarity
Corrupt and brutal Corrupt but brutal
Hatzivassiloglou McKeown 1997
30
Experiments
Distributional similarity
Seeds
Filtering
31
Experiments
Distributional similarity
Seeds
Filtering
Lexical Classification
32
Results
Seed 7.5
33
Results
Lex
Seed Lex Pol,Grad 6.4
18.0 Pol-, Grad
19.9 21.4 Pol,- Grad
8.4 18.2
Seed 7.5
34
Future Work

• Apply process to Netnews and Listservs
• Apply word-sense disambiguation techniques to
  potentially subjective expressions
• Flame recognition and review mining

35
Conclusions

• Learning linguistic knowledge from corpora for a
  pragmatic task
• Linguistic information
• Manual annotation
• Linguistic constraints
• Processes improve each other

36
Application 1 Flame recognition
From pattreck_at_aol.com (PattReck) Newsgroups
rec.gardens.roses Subject Re red(as in wine
red) roses My two favorite old reds Cramoisi
Superieure, especially great climbing, and
Francis Dubreuil. Also Prospero does well
in southern California - aren't you on the west
coast? -- Candace
37
Flames (continued)
From Suzanne ltSuzanne_member_at_newsguy.comgt Newsgro
ups rec.gardens.roses Subject Re BILL
WARRINER!!!!
gtgtWow. You guys are really working poor Suzanne
over. gtpo thang. I thank she been workin over
her bottle of Kahlua. Up !!! I've been
working at a job - no Kahlua! You are
a snow-snorting dust-bowl dweller, the dustiest
of the dusties. Bill Bradley has the support of
the "environmentalists" ha ha ha!
38
Likely

• likely possible willing probable
• receptive unlikely able logical
• rumored potential counterproductive
• moot significant hesitant worthy
• unwilling probably desirable
• weak forthcoming imminent