Computational Models of Discourse Analysis

1. Computational Models of Discourse Analysis Carolyn Penstein Rosé Language Technologies Institute/ Human-Computer Interaction Institute

2. Computational Approaches • Two steps • Step 1: Metaphor recognition • Step 2: Metaphor interpretation • Does this paradigm cover everything that Lakoff and Johnson place under the heading of metaphor? Examples from the paper: • Lakoff’s concept: • Metaphors structure how we think about an event or state. • The way we think affects: • what we expect to happen, • what we do, • how we respond to what occurs during an event, • and how we talk about what we and others are doing

3. Announcements! • Questions about presentations for next time? • Rearranged syllabus slightly: see Drupal • Posted responses to posts • Readings for next unit + most of rest of semester posted • Next unit focuses on Sentiment Analysis • Product review dataset will be ready by next Monday for Assignment 3 • Note we won’t meet during Spring Break • Unit 3 has a break too! • We won’t meet on Wed, March 30 since several of us will be away

4. MIP: Metaphor Identification Procedure

6. Growing Interest? #References Automatic Approaches

7. Recent Approaches to Detection • Peters and Peters 2000: Mined wordnet for abstract concepts that share word forms such as publication-publisher • Mason 2004: Mine an internet corpus for domain specific selectional restriction differences • Birke and Sarkar 2006: Start with seed sentences that have been annotated with figurative versus literal, and then do something like an instance based learning approach • Gedigan et al. 2006: extract frames for MOTION and CURE from FrameNet, then extract sentences related to these from PropBank. Annotate by hand for metaphoricity. Use a maximum entropy classifier. • Krishnakumaran and Zhu 2007: Look for sentences with “be” verb. Check for hyponymy using WordNet. If not there, look at bigram counts of subj-obj. If not high, then might be metaphorical.

8. What would Fass say? • Problem with selectional restrictions as evidence: • Will detect all kinds of nonliteral and anomalous language regardless if it is metaphorical or not • Common metaphorical sense (i.e., “dead metaphors”) will fail here • Some statements can be interpreted either way: “All men are animals”

9. Recent Approaches to Interpretation • Metaphor based reasoning framework – reason in a source domain and apply reasoning to the target domain using a conceptual mapping • Narayan’s KARMA 2004: parsed text as input • Barnden and Lee’s ATT-Meta 2007: logical forms as input • Talking Points 2008: uses WordNet, then uses minimal edits to bridge concepts • Makeup is the Western burqa • Shutova 2010: uses a statistical paraphrase approach

10. Shutova’s Take Away Message • Approaches from the 80s and 90s were rule based • Knowledge engineering bottleneck • Shutova’s work give some evidence that metaphor can be handled using a more contemporary (i.e., machine learning) paradigms • Cast the metaphor interpretation problem as a paraphrase problem so you can use statistical machine translation approaches

11. Does paraphrase “cut it”?

12. Do you see a metaphor here? * How much of the problem can be solved by paraphrase?

13. Do you see metaphor here? • Evey: Who are you?V: Who? Who is but the form following the function of what and what I am is a man in a mask.Evey: Well, I can see that.V: Of course you can, I’m not questioning your powers of observation, I’m merely remarking upon the paradox of asking a masked man who he is.Evey: Oh.V: But on this most auspicious of nights, permit me then, in lieu of the more commonplace soubriquet, to suggest the character of this dramatis persona. • [pauses for a few seconds] • Voila! In view humble vaudevillian veteran, cast vicariously as both victim and villain by the vicissitudes of fate. This visage, no mere veneer of vanity, is a vestige of the “vox populi” now vacant, vanished…

14. Data’s Identity Note: The focus of the work of Shutova and others who have self-identified as working on metaphor is on uncovering the literal meaning of expository text. • We see evidence of how Data is framing his identity. • Do we see metaphor here? • Lakoff’s concept: • Metaphors structure how we think about an event or state. • The way we think affects: • what we expect to happen, • what we do, • how we respond to what occurs during an event, • and how we talk about what we and others are doing

15. Another spin on Metaphor Recognition • Perspective modeling work • Liberal versus Conservative • Pro or Against • Sentiment analysis more generally • Different computational approach • Skips step 1 – assumes all language represents perspective • Simplifies step 2 – goal is to recognize a category rather than rephrase • Usually models are based on word distributions • Word vectors with weights • Topic models • We’ll explore this in the next unit

16. Framing an Event in Progress • Where does the paradigm for understanding metaphors break down with examples like this? • Step 1: recognize metaphor • Step 2: map to literal meaning • *** Still understanding a concept/situation by comparison with another one

17. Breaking the Paradigm • What can we do with conversational data? • How do we recognize that a metaphor is in play? • What would it mean to do the interpretation?

18. Questions?