Dialog act tagging using MemoryBased Learning

1
Dialog act tagging usingMemory-Based Learning

• Term project
• Mihai Rotaru

2
Outline

• Dialog act tagging
• Memory-based learning and why
• Feature selection
• Results
• Conclusion

3
Dialog act (DA) tagging

• Shallow semantic analysis of utterances
• Express the illocutionary force
• Usefulness
• Detection of dialog game boundaries
• In meeting summarizer
• Speech recognizer

4
Automatic DA tagging

• Previous work
• Word substrings
• Cue phrases - Hirshberg and Litman (1993)
• Word n-grams Reithinger and Klesen (1997)
• Dialog act cues - Samuel et al (1998, 2000)
• Statistical approach
• HMM - Shriberg et all (1998) and Stolcke et all
  (2000)

5
Outline

• Dialog act tagging
• Memory-based learning and why
• Feature selection
• Results
• Conclusion

6
Memory-based learning (MBL)

• Memory-based reasoning theory
• Similarity between situations

7
Why MBL?

• Advantages
• handles exception and sub-regularities
• Daelemans (1999) and Van den Bosch (2001)
• Grapheme-to-phoneme conversion
• Part of speech tagging
• Prepositional phrase attachment
• Base noun phrase chunking
• Communication problems
• Disadvantages
• Memory/computational demand

8
Outline

• Dialog act tagging
• Memory-based learning and why
• Feature selection
• Results
• Conclusion

9
Feature selection - corpus

• Switchboard corpus
• 1200 dialogs
• 210,000 utterences
• SWBD-DAMSL coding
• 42 DA plus DA
• Results 71 HMM with trigrams
• DA ignored in study

10
Feature selection - corpus

• Example
• nn _at_B.14 utt1 No, /
• sde _at_B.14 utt2 I don't, I don't
  have any kids. /
• sd _at_B.14 utt3 I, F uh, my
  sister has a, she just had a baby, /
• sd _at_B.14 utt4 he's about five months
  old /
• sd _at_B.14 utt5 C and she was worrying
  about going back to work and what she was
  going to do with him and --
• b A.15 utt1 Uh-huh. /
• B.16 utt1 -- the different, - /
• qy B.16 utt2 do you have kids? /
• na A.17 utt1 I have three. /
• bh B.18 utt1 F Oh, really? /

11
Feature selection

• Previous 3 DA, time since last change of speaker.
• Presence of DA cues? KILLING
• Word bigrams
• Capture pattern of expression
• Set features not supported in TiMBL !!!
• Solution? Hash bigram to feature slots.

12
Feature selection continued
13
Outline

• Dialog act tagging
• Memory-based learning and why
• Feature selection
• Results
• Conclusion

14
Results

• Compare to
• Majority baseline 35
• 71 from Stolckes experiment
• Three dimension to analyze
• Hash size (10, 20 and 30)
• Different/same value for empty hash slots
• Usage of DA

15
Results - dimensions
16
Results - neighbors
17
Results confusion matrix
18
Conclusion

• Promising results from MBL for the task of DA
  tagging
• Future work
• Why 3 neighbor works better?
• Implement set feature in TiMBL