Post on 30-Dec-2015
description
Agustín Gravano1,2
Julia Hirschberg1
(1)(1) Columbia University, New York, USAColumbia University, New York, USA(2) Universidad de Buenos Aires, Argentina(2) Universidad de Buenos Aires, Argentina
Backchannel-Inviting Cuesin Task-Oriented Dialogue
Agustín Gravano Interspeech 2009 2
Interactive Voice Response Systems
• Quickly spreading. Mostly simple functionality.
• “Uncomfortable”, “awkward”.
• ASR+TTS account for most IVR problems.
• As ASR and TTS improve, other problems revealed.• Coordination of system-user exchanges.
• Backchannels.
Introduction
Agustín Gravano Interspeech 2009 3
• Short expressions uttered by listeners to: • Convey that they are paying attention.
• Encourage the speaker to continue.
• Examples: okay, uh-huh, mm-hm, alright.
• Very frequent in task-oriented dialogue.
• Thus, modeling human usage of BC should lead to an improved system-user coordination.
Introduction
Backchannels
Agustín Gravano Interspeech 2009 4
Goal
• Learn when backchannels are likely to occur.
• Find “backchannel-inviting” cues.• Cues displayed by the speaker “inviting” the
listener to produce a backchannel response.
• This could improve the coordination of IVRs:• Speech understanding: Detect points in the user’s
turn where a backchannel would be welcome. • Speech generation: Display cues inviting the user
to produce a backchannel.
Introduction
Agustín Gravano Interspeech 2009 5
Talk Outline
• Previous work• Material• Method• Results• Conclusions
Agustín Gravano Interspeech 2009 6
Previous Work
• Duncan 1972, 1973, 1974, inter alia.• Hypothesized six turn-yielding cues in face-to-face
dialogue.• Several studies continued this line of research, but
always excluded backchannels.
• Ward & Tsukahara 2000.• Region of low pitch lasting 110ms or more.
• Cathcart et al. 2003.• Language model based on pause duration and part-
of-speech tags to predict the location of BC.
Backchannel-Inviting Cues
Agustín Gravano Interspeech 2009 7
Columbia Games Corpus
• 12 task-oriented spontaneous dialogues.• Standard American English.• 13 subjects: 6 female, 7 male.• Series of collaborative computer games.• No eye contact. No speech restrictions.• 9 hours of dialogue.• Manual orthographic transcription, alignment.• Manual prosodic annotations (ToBI).
Material
Agustín Gravano Interspeech 2009 8
Player 1: Describer Player 2: Follower
Material
Columbia Games Corpus
Agustín Gravano Interspeech 2009 9
Backchannel-Inviting Cues
• Cues displayed by the speaker “inviting” the listener to produce a backchannel response.
Agustín Gravano Interspeech 2009 10
Method
• 3 trained annotators identified Backchannels using a labeling scheme described in [Gravano et al. 2007].
• To find BC-inviting cues, we compare:• IPUs preceding Holds,
• IPUs preceding Backchannels.
Backchannel-Inviting Cues
• IPU (Inter Pausal Unit): Maximal sequence of words from the same speaker surrounded by silence ≥ 50ms.
Hold Backchannel
Speaker A:
Speaker B:
IPU1 IPU2
IPU3
IPU4
Agustín Gravano Interspeech 2009 11
Backchannel-Inviting Cues
Individual Cues
1. Final rising intonation: 81% of IPUs before BC end in H-H% or L-H%.
2. Higher pitch level.
3. Higher intensity level.
4. Lower NHR (voice quality).
5. Longer IPU duration (seconds, #words).
6. Final POS bigram:72% of IPUs before BC end in DT NN, JJ NN, or NN NN.
} •entire IPU•final 1.0 sec•final 0.5 sec
Agustín Gravano Interspeech 2009 12
Defining Presence of a Cue
• 2 representative features for each cue:
Final intonation Pitch slope over final 200ms, 300ms.
Intensity level Mean intensity over final 500ms, 1000ms.
Pitch level Mean pitch over final 500ms, 1000ms.
Voice quality NHR over final 500ms, 1000ms.
IPU duration Duration in ms, and in number of words.
Final POS bigram {‘DT NN’, ‘JJ NN’, ‘NN NN’} vs. Rest (binary).
• Define presence/absence based on whether the value is closer to the mean before BC or H.
Backchannel-Inviting Cues
Agustín Gravano Interspeech 2009 13
Top Frequencies of Complex Cues
BC-inviting cues:
1: Final intonation
2: Intensity level
3: Pitch level
4: IPU duration
5: Voice quality
6: Final POS bigram
digit == cue present
dot == cue absent
Agustín Gravano Interspeech 2009 14
Backchannel-Inviting Cues
Combined Cues
Number of cues conjointly displayed
Per
cent
age
of I
PU
s fo
llow
ed b
y a
BC
r 2 = 0.993
Agustín Gravano Interspeech 2009 15
Backchannel-Inviting Cues
IVR Systems
• After each IPU from the user:if estimated likelihood > thresholdthen produce a backchannel
• To elicit a backchannel from the user, if desired:Include as many cues as possible in the system’s final IPU.
Agustín Gravano Interspeech 2009 16
Summary
• Study of backchannel-inviting cues.• Objective, automatically computable.• Combined cues.• Improve turn-taking decisions of IVR systems.
• Results drawn from task-oriented dialogues.• Not necessarily generalizable.• Suitable for most IVR domains.
• SIGdial 2009: Study of turn-yielding cues.
Agustín Gravano Interspeech 2009 17
Special thanks to…• My advisor, Julia Hirschberg• Thesis Committee Members
• Maxine Eskenazi, Kathy McKeown, Becky Passonneau, Amanda Stent.
• Speech Lab at Columbia University• Stefan Benus, Fadi Biadsy, Sasha Caskey, Bob Coyne, Frank
Enos, Martin Jansche, Jackson Liscombe, Sameer Maskey, Andrew Rosenberg.
• Collaborators• Gregory Ward and Elisa Sneed German (Northwestern U);
Ani Nenkova (UPenn); Héctor Chávez, David Elson, Michel Galley, Enrique Henestroza, Hanae Koiso, Shira Mitchell, Michael Mulley, Kristen Parton, Ilia Vovsha, Lauren Wilcox.