Josef Ruppenhofer, Swapna Somasundaran, Janyce Wiebe University of Pittsburgh

1
Finding the sources and targets of subjective
expressions

• Josef Ruppenhofer, Swapna Somasundaran, Janyce
  WiebeUniversity of Pittsburgh

2
What is Subjectivity?

• The linguistic expression of somebodys opinions,
  sentiments, emotions, evaluations, beliefs,
  speculations (private states)
• This pleased the mainly female audience.
• Source the person who experiences a private
  state
• Target what the private state is about or
  directed towards

3
Motivation

• Sentiment analysis is a fast-growing field with
  many applications (e.g. Question Answering,
  Product review mining, Information Extraction)
• In many kinds of texts we find opinions
  attributed to several different sources, and/or
  opinions about multiple targets

4
Opinion Question Answering

• Q What is the international reaction to the
  reelection of Robert Mugabe as President of
  Zimbabwe?
• A African observers generally approved of his
  victory while Western Governments strongly
  denounced it.

5
Product review mining

• The computer is very good and very easy to use.
  It has a built in camera, bluetooth the all
  singing and dancing machine. Love it. The only
  glitch is the scrolling pad is not as smooth as
  my last Toshiba notebook. One other thing is that
  Vista is a nightmare...

6
Motivation

• Sentiment analysis is a fast-growing field with
  many applications (e.g. Question Answering,
  Product review mining, Information Extraction)
• In many kinds of texts we find opinions
  attributed to several different sources, and/or
  opinions about multiple targets
• Challenge is to associate sources, opinions, and
  targets correctly

7
Roadmap

• Here we discuss some of the challenges that an
  automatic system needs to be able to deal with
• We take the use of Automatic Semantic Role
  Labeling (ASRL) systems as our starting point
• Based on our work in corpus annotation, we show
  that we need additional capabilities beyond ASRL

8
Automatic Semantic Role Labeling
9
Semantic roles

• This pleased the mainly female audience.

PropBank FrameNet
Arg1 Experiencer
Arg0 Stimulus

• We fear an early death much more.

PropBank FrameNet
Arg0 Experiencer
Arg1 Content
10
Mapping opinion roles to semantic roles

• This pleased the mainly female audience.

Opinion roles PropBank FrameNet
Source Arg1 Experiencer
Target Arg0 Stimulus

• We fear an early death much more.

Opinion roles PropBank FrameNet
Source Arg0 Experiencer
Target Arg1 Content
11
Annotation scheme
12
Private states (Quirk et al. 1985)

• States such as emotions, evaluations,
  speculations, etc.
• States that are not open to objective observation
  or verification.
• States that involve a particular persons point
  of view
• Private states involve sources holding attitudes,
  typically towards targets.

13
Ways of evoking private states

• Explicit mentionsHe was boiling with anger.
• Speaking events expressing private statesThe
  papers editors attacked the new House Speaker.
• Expressive subjective elements (Banfield
  1982)That doctor is a quack.
• Objective speech eventsThe bus leaves at 4,
  Bill said.

DSEs
ESEs
14
Nesting of private states

• The US fears a spill-over, said Xirao-Nima.

15
Nesting of private states

• The US fears a spill-over, said Xirao-Nima.

ltwritergt
said Xirao-Nima
The US fears a spill-over
16
Challenges beyond role labeling
Sources Targets
Attribution
Referent Identification Referent Identification
Inferences
Arguing Attitudes
17
Attribution
18
Attribution

• Expressive subjective elements (ESEs) dont have
  a semantic role for their sourceSenior Mike
  Sheehy said , It was a blast .

19
Attribution

• The source for an ESE is not always at the same
  level. CompareSenior Mike Sheehy said, It was
  a blast.
• She loves that idiot.

20
ltwritergt
Senior Mike Sheey said, It was a blast
21
Attribution

• Some expressions function both as ESEs and as
  DSEs

22
Attribution

• Some expressions function both as ESEs and as
  DSEsIt is a shame that there is no jury that
  can mete out justice for a city he has slandered
  for far too long.

source target
DSE He city
ESE ltwritergt he
23
Attribution

• Attribution and content of a private state may
  be presented separately

24
Attribution

• Attribution and content of a private state may
  be presented separately
• Chris Moyles is a brilliant broadcaster, the
  saviour of Radio 1, a comedian, a best-selling
  author, and, in fact, a genius. Or so he says.

25
Attribution
ltwritergt
Chris Moyles is a brilliant broadcaster, the
saviour of Radio 1, a comedian, a best-selling
author, and, in fact, a genius.
26
Attribution
ltwritergt
ltChris Moylesgt
Chris Moyles is a brilliant broadcaster, the
saviour of Radio 1, a comedian, a best-selling
author, and, in fact, a genius.
Or so he says
27
Attribution

• An attribution may apply to several utterances
  without being explicitly signaled each time.

And I went ahead and mailed it in
thinking uh I wont get the scholarship. Who
cares? I dont, just so I can work in the school
and Ill be happy.
But one day I came in and I looked at my mail and
I was accepted.
28
Reference Identification
29
Reference Identification

• Overt referentsYou think about it and then let
  me know.Some people say the Steelers are
  contenders but Im not convinced of it.

30
Reference Identification

• Zero referentsSource Think about it and then
  let me know.Target Some people say the
  Steelers are contenders but Im not convinced.

31
Reference Identification

• Exophora referents are present only in the
  physical contextSorry.Oopsy-daisy.

32
Inferences
33
Inferences

• For some events about which opinions are
  expressed, we can infer additional attitudes
  towards affected or causing participants

34
Inferences

• For some events about which opinions are
  expressed, we can infer additional attitudes
  towards affected or causing participantsI
  think people are happy because Chavez has fallen.

35
Inferences

• For some events about which opinions are
  expressed, we can infer additional attitudes
  towards affected or causing participantsI
  think people are happy because Chavez has fallen.
• I think people are happy because Chavez has
  fallen.

36
Inferences

• The targets occurring in a discourse are often
  interrelated such that opinions about local
  targets contribute to the overall assessment of a
  global target.
• The computer is very good and very easy to
  use. It has a built in camera, bluetooth the
  all singing and dancing machine. Love it. The
  only glitch is the scrolling pad is not as
  smooth as my last Toshiba notebook. One other
  thing is that Vista is a nightmare...

37
Targets of arguing
38
Arguing attitudes

• What is the case or not
• From this it follows that mechanisation is not
  economic unless it can produce higher yields of
  crops than these older methods.
• What should be done or not
• We strongly recommend that all Firefox users
  upgrade to this latest release.

39
Targets of arguing

• Interpretation of arguments made by causal and
  conditional constructions is very
  context-dependent.
• Your presentation will be better if you put
  this on the first slide
• You will want to vote YES if you want to keep
  the cost of government in Lewiston low

40
Targets of arguing

• Interpretation of arguments made by causal and
  conditional constructions is very
  context-dependent.
• Hypothetical Your presentation will be better
  if you put this on the first slide
• Implicit assertion You will want to vote YES if
  you want to keep the cost of government in
  Lewiston low

41
Targets of arguing

• Easy finder with the a whistle function or
  something, or rechargeable station because its
  a pain when you run out of batteries.
• If youre not a good cook, then taking your
  girlfriend out to an expensive restaurant might
  be the next best romantic date idea. ... Youll
  feel good because youve made her happy with a
  romantic date.

42
Targets of arguing

• The prototypical targets that we annotate are
  entities.
• For arguing, we could also annotate the entities
  that arguments are about.
• However, we also recognize that the logical
  targets of arguing are propositions.

43
Targets of arguing

• Clinton should be the presidential candidate.
• Clinton should be the running mate.
• The prototypical targets that we annotate are
  entities.
• For arguing, we could also annotate the entities
  that arguments are about.
• However, we also recognize that the logical
  targets of arguing are propositions.

44
Targets of arguing

• Clinton should be the presidential candidate.
• Clinton should be the running mate.
• The prototypical targets that we annotate are
  entities.
• For arguing, we could also annotate the entities
  that arguments are about.
• However, we also recognize that the logical
  targets of arguing are propositions.

45
Targets of arguing

• Clinton should be the presidential candidate.
• Clinton should be the running mate.
• The prototypical targets that we annotate are
  entities.
• For arguing, we could also annotate the entities
  that arguments are about.
• However, we also recognize that the logical
  targets of arguing are propositions.

46
Targets of arguing

• Clinton should be the presidential candidate.
• Clinton should be the running mate.
• The prototypical targets that we annotate are
  entities.
• For arguing, we could also annotate the entities
  that arguments are about.
• However, we also recognize that the logical
  targets of arguing are propositions.

47
Conclusion

• Semantic role labeling is needed for finding
  sources and targets
• But we also need
• ways of establishing levels of attribution
• capabilities for dealing with zero references
• lexical information to support inferences
• deal with the full variety of attitudes and their
  sources and targets

48
Thanks!josefr_at_cs.pitt.edu
49
References Annotation scheme

• Banfield, Ann. 1982. Unspeakable Sentences
  Narration and Representation in the Language of
  Fiction. Routledge Kegan Paul, Boston.
• Quirk Randolph, Greenbaum Sidney, Leech Geoffrey,
  and Svartvik Jan. 1985. A Comprehensive Grammar
  of the English Language. Longman, New York, NY.
• Janyce Wiebe M. 1994. Tracking point of view in
  narrative. Computational Linguistics 20 (2)
  233-287.

50
References Annotation scheme

• Janyce Wiebe, Theresa Wilson , and Claire
  Cardie. 2005. Annotating expressions of opinions
  and emotions in language. Language Resources and
  Evaluation, volume 39, issue 2-3, pp. 165-210.

51
References Role Labeling

• Penn Discourse Treebank http//www.seas.upenn.edu/
  pdtb/PDTBAPI/pdtb-annotation-manual.pdf
• PropBankhttp//verbs.colorado.edu/mpalmer/projec
  ts/ace.html
• FrameNethttp//framenet.icsi.berkeley.edu/

52
References Role Labeling

• Y. Choi, E. Breck, and C. Cardie. 2006. Joint
  Extraction of Entities and Relations for Opinion
  Recognition. In Proc. of EMNLP 2006.
• S. Kim and E. Hovy. 2006. Extracting Opinions,
  Opinion Holders, and Topics Expressed in Online
  News Media Text. In ACL Workshop on Sentiment and
  Subjectivity in Text.

53
References Belief spaces

• Dyer, Michael G. 1983. In-Depth Understanding A
  Computer Model of Integrated Processing for
  Narrative Comprehension. MIT Press, Cambridge,
  MA.
• Fauconnier, Gilles. 1985. Mental Spaces Aspects
  of Meaning Construction in Natural Language. MIT
  Press, Cambridge, MA.

54
References Belief spaces

• Rapaport, William J. 1986. Logical Foundations
  for Belief Representation. Cognitive Science.
• Wilks, Yorick and Bien, Janusz. Beliefs, Points
  of View, and Multiple Environments. Cognitive
  Science 7 95-119.

55
References Literary theory

• Chatman, Seymour. 1978. Story and Discourse
  Narrative Structure in Fiction and Film. Cornell
  University Press, Ithaca, NY.
• Cohn, Dorrit. 1978. Transparent Minds Narrative
  Modes for Representing Consciousness in Fiction
  Princeton University Press, Princeton, NJ.
• Dolezel, Lubomir. 1973 .Narrative Modes in Czech
  Literature. University of Toronto Press, Toronto,
  Canada.

56
References Literary theory

• Hamburger Käte. 1973. M.J. Rose, Trans., The
  Logic of Literature. Indiana University Press,
  Bloomington, Indiana.
• Kuroda, S.-Y. 1976. Reflections on the
  Foundations of Narrative Theory--From a
  Linguistic Point of View. In van Dijk, T.A.,
  Ed., Pragmatics of Language and Literature, North
  Holland, Amsterdam.
• Uspensky, Boris. 1973. A Poetics of Composition.
  University of California Press, Berkeley, CA.

57
References Discourse

• Allen, James F. and Perrault, C. Raymond. 1980.
  Analyzing Intention in Utterances. Artificial
  Intelligence 15 143-178.
• Fillmore, Charles. 1974. Pragmatics and the
  Description of Discourse. In Fillmore, Charles,
  Lakoff, George, and Lakoff, Robin, Eds., Berkeley
  Studies in Syntax and Semantics I. University of
  California Dept. of Linguistics and Institute of
  Human Learning, Berkeley, CA V1-V21.

58
Extra slides
59
Nesting of private states

• The US fears a spill-over, said Xirao-Nima.

Level Private state span source
0 ltgt ltwritergt
1 said Xirao-Nima
2 fears the US
60
Nesting of private states

• The US fears a spill-over, said Xirao-Nima.

Level Private state span source
0 ltgt ltwritergt
1 said Xirao-Nima
2 fears the US
61
Nesting of private states

• The US fears a spill-over, said Xirao-Nima.

Level Private state span source
0 ltgt ltwritergt
1 said Xirao-Nima
2 fears the US