Computational Models of Discourse Analysis

1
Computational Models of Discourse Analysis

• Carolyn Penstein Rosé
• Language Technologies Institute/
• Human-Computer Interaction Institute

2
Pre-WarmUp Discussion

• What can we do about jargon?
• This paper for Wednesday is so jargon-ridden, I'm
  not sure if it actually makes sense or not. An
  example "Essentially we find the transitive
  closure of the coreference and meronymy relations
  on the initial set of mentions" (first page,
  second column end of first full paragraph).
  ...and this is before any of the technical
  details!

3
Remember that one of the instructional goals
of this course is to teach you how to read this
literature.
4
Warm Up Discussion

• How comprehensive is this table when we consider
  sentiment expressions and targets in our
  Appraisal theory analysis?
• Look at the examples in the table and identify
  whether one of the paths would link the sentiment
  expression to its target.
• Which ones dont work? How would the approach
  need to be extended?

5
Unit 3 Plan

• 3 papers we will discuss all give ideas for using
  context (at different grain sizes)
• Local patterns without syntax
• Using bootstrapping
• Local patterns with syntax
• Using a parser
• Rhetorical patterns within documents
• Using a statistical modeling technique
• The first two papers introduce techniques that
  could feasibly be used in your Unit 3 assignment

6
What can be evaluated?

• Also, from the definition, it seems that
  'mentions' are just any noun or possessive
  pronoun (or features of these that can be
  evaluated). I guess these are the only things
  that can be evaluated, although I'm not sure of
  the possessive pronouns (my, its, his, etc).

7
Dependency Relations
What is the potential downside of using
dependency relations as features?
8
Why its tricky
9
Why dependency relations are important for
sentiment

• A big candy bar versus a big nose
• A deep thought versus a deep hole
• Hard wood floor versus hard luck
• Cold drink versus cold hamburger
• Furry cat versus furry food
• Ancient wisdom versus ancient hardware

10
Possibly unintuitive attributions

• What sentiment is expressed by this sentence
• I broke the handle
• They argue that the speaker expresses regret
  about his own actions
• Comes from Wilson and Wiebes work
• Does this seem reasonable? Why or why not?
• Consistent with Appraisal theory?

11
Student Comment

• I think, like suggestions for the other paper,
  this paper could possibly include the
  positive/negative dimension of Appraisal Theory,
  but I'm not sure how often these situations
  actually come up. Example (7) on page 96 shows
  one example, but I'm not sure if this genre of
  ambiguity is common.

12
Annotation
13
Is there a problem here?

• Explain how this sentiment propagation graph
  would be used in sentiment analysis.
• Can you see a problem that would occur if you
  apply this to movie reviews?

14
Alternative Approaches

• Proximity pick the closest target
• Heuristic Syntax shortest path
• Bloom hand crafted dependency paths
• RankSVM learn weights on types of evidence for
  ranking targets

Not clear how much advantage from types of
features versus the supervised learning approach.
15
Results
What questions are left unanswered and what
follow up experiments would you do? What ideas
does this paper give you for Assignment 3?
16
Tips for Mondays Reading Assignment

• Skip Section 4 and the Appendix the first time
  you read the paper
• Then skim through section 4, skipping over any
  sentences you dont understand
• Focus on the initial paragraphs in
  sections/subsections, as these tend to give a
  high level idea of what the message is
• Keep in mind that their Latent Sentence
  Perspective Model is just Naïve Bayes with one
  twist can you find what that one twist is?

17
Questions?