Lexical, Prosodic, and Syntactics Cues for Dialog Acts

1
Lexical, Prosodic, and Syntactics Cues for Dialog
Acts
2
Purpose

• Investigating the particular class of dialogue
  acts called backchannels (aka "acknowledgement
  tokens"). Those give feedback to the speaker.
• Continuers an utterance indicating that the
  other speaker should go on speaking.
• incipient speakership
• agreements marks the degree to which a user
  accepts the speakers opinion, statement, etc.
• yes-answers
• Assessments ascriptive sentences expression of
  a certain property ("good," "bad").

3
Questions

• Does it have any importance as cue-phrases for
  the structure?
• Does prosodic knowledge have any importance?
• Is there any difference in syntactic realization
  of certain dialog acts, compared to others?

4
Who Cares?

• Part of a project to "automate discourse
  structure for speech recognition and
  understanding tasks"
• Create models of speakers and dialog.

5
Materials and Methods

• -SWBD-DAMSL dialog tagset
• an adaptation of DAMSL tag-set
• 60 labels that could be combined in different
  dimensions.
• -7 CU-Boulder linguistics grad students labeling
  switchboard conversations of human-to-human
  interaction

6
Results

• 220 unique tags for 205,000 utterances, where
  each utterance received exactly one tag.
• Good agreement rate between labelers (84), but
  too many tags for statistical analysis ?
  everything was condensed into 42 final tags

7
On the lexical level, do words like "yeah" have
any importance as cue-phrases for the structure?"

• Hirschberg Litman "now" and "well."
• Here words with overlap meanings
• Jurafskys Findings
• utterances may be ambiguous, BUT through the
  lexical form, we could find the word's true
  "identity."

8
"Yeah" Vs. "uh-huh"

• Yeah" (and variations) ambiguous ("agreements"
  at 36, at "incipient speaker" at 59", and 86
  at "yes-answers")
• "uh-huh" (with its variations) a "continuer"
  (45 vs. "yeah" at 27).

9
Next Question "Does prosodic knowledge have any
importance in determining DA's meaning?"

• Hirschberg and Litman intonational phrasing and
  pitch accent disambiguates cue phrases and help
  determine discourse structure.
• Jurafsky agrees.
• Runs experiment Text transcription vs. complete
  speech

10
MATERIALS AND METHODS

• 44 random, but previously labeled conversations.
• Conversations were put in full context as well as
  fully transcribed AND the original labeling was
  also available.
• RESULTS 98 agreement!

11
2

• The shifts
• 38 of the 2 continuers ? agreements
• 19 of the 2 opinions ? statements
• 15 of the 2 statements ? opinions
• Etc.

12
Why the Shifts?

• due to cascading changes (continuer ? agreement,
  preceding statement opinion).
• Nature of labeling instruction ambiguous cases
  were labeled as continuers
• Identical lexical form some speakers
  distinguished agreement from continuer based on
  prosody, while others used lexical form.

13
Take Home Message

• Continuers are, in most cases (vs. agreements)
• shorter in duration
• less intonationally marked
• long pauses before speaking

14
3rd QuestionIs there any difference in syntactic
realization of certain dialog acts, compared to
others?

• Goodwin and Goodwin
• Pro Term Copula (Intensifier) Assessment
  Adjective
• Formula happens frequently, where
• pro term mostly equals "that"
• intensifiers are rare, and are usually "really,"
  and "pretty"
• assessment adjective has small range of
  adjectives great, good, nice, wonderful, etc.

15
Syntactic Conclusion

• "micro-syntax" restricted grammatical
  production than would have thought. That is,
  certain dialog acts have their own syntactic
  patterning.