1
Effect of Genre, Speaker, and Word Class on the
Realization of Given and New Information
Interspeech 2006 - Pittsburgh, PA

• Agustín Gravano Julia Hirschberg
• agus, julia_at_cs.columbia.edu

Spoken Language Processing Group Columbia
University
2
Motivation

• Speakers of American English tend to
• accent references to new information, and
• deaccent references to old (or given)
  information.
• (Chafe 1974, Prince 1981 1992, inter alia)
• Variation of prominence in given entities is
  strongly affected by the persistence of
• grammatical function (subject, object, etc.) and
• position in the sentence.
• (Terken Hirschberg, 1994)

3
Motivation

• Possible applications
• Improve naturalness of TTS systems.
• Aid ASR.
• Questions
• What are other sources of variation?
• What is the effect of
• speaker?
• genre?
• word class?

4
Main Results

• Speakers vary the manner in which they realize
  differences in information status.
• Speakers tend to produce given verbs with
  higher intensity than new verbs, both in read
  and spontaneous speech.

5
Overview

• Materials and Methods
• Corpus
• Information status
• Word classes
• Features
• Results
• Nouns
• Verbs
• Discussion
• Conclusions

6
Boston Directions Corpus

• Hirschberg Nakatani 1996
• Spontaneous and read monologues.
• 9 increasingly complex direction-giving tasks
• Describe how to get to MIT from Harvard.
• Method
• Spontaneous speech recorded and transcribed.
• Speakers returned and read.
• 4 speakers (3 male, 1 female).

7
Boston Directions Corpus

• Mean length of tasks
• Spontaneous 111s
• Read 84s
• Excerpt from the spontaneous part of the corpus
• first enter the Harvard Square T stop and
  buy a token then proceed to get on the
  Inbound um Red Line uh subway ...
• Corpus size
• Spontaneous 66m
• Read 50m
• Prosody labeled using the ToBI convention.

8
Information Status

• Prince 1981
• Entities are new when first introduced in the
  discourse.
• Evoked entities are given. They are already in
  the discourse.
• Simple definition
• A word w is given if in the current task there is
  at least one previous occurrence of a word with
  the same stem.
• Otherwise, we say that w is new.

9
Word Classes

• Automatically labeled the part-of-speech of all
  the words in the corpus using the Brill Tagger.
• Categorized words into
• Nouns
• Verbs
• Adjectives
• Adverbs
• Others
• Significant results only for Nouns and Verbs.

10
Features

• Word acoustic features, extracted using Praat
• Max, mean, min pitch
• Max, mean, min intensity
• Pitch and intensity features were also normalized
  with respect to the mean value of
• 1 second around the target word,
• 5 words around the target word,
• the target words Intermediate Phrase.
• Pause before and after the word.

11
Results Nouns
READ READ READ READ SPON SPON SPON SPON
  S1 S2 S3 S4 S1 S2 S3 S4
Max Pitch g
Mean Pitch n n g
Min Pitch n g
Max Pitch / Context Mean Pitch n n n n
Mean Pitch / Context Mean Pitch n n n n
Min Pitch / Context Mean Pitch n n
Max Intensity n n g g
Mean Intensity n n g g
Min Intensity g g g
Max Int / Context Mean Intensity n n n n
Mean Int / Context Mean Intensity n n n g g
Min Int / Context Mean Intensity g
Pause Before n n n n n
Pause After g g g
n mean value for the new words is significantly
larger than for the given words g mean value
for the given words is significantly larger than
for the new words
12
Results Verbs
READ READ READ READ SPON SPON SPON SPON
  S1 S2 S3 S4 S1 S2 S3 S4
Max Pitch n
Mean Pitch g g n
Min Pitch g n
Max Pitch / Context Mean Pitch g n
Mean Pitch / Context Mean Pitch g
Min Pitch / Context Mean Pitch g n
Max Intensity g g g g g g g
Mean Intensity g g g g g g g
Min Intensity g g g
Max Int / Context Mean Intensity g g g g g
Mean Int / Context Mean Intensity g g g g g g g g
Min Int / Context Mean Intensity g g
PauseBefore g g
PauseAfter g
n mean value for the new words is significantly
larger than for the given words g mean value
for the given words is significantly larger than
for the new words
13
Discussion Variation of intensity in verbs

• Examples
• you get out of the T stop you cross
  Massachusetts Avenue ... you wanna cross Mass
  Ave opposite that there's usually a bunch of
  cabs and and people standing around there so
  then once you've crossed it you're you're in
  Harvard Yard proper
• then you're right at the entrance to what is
  called the Infinite Corridor and it's called
  the Infinite Corridor because it's this really
  long place you can walk entirely indoors
• Direct objects of cross and call are either
  deaccented or pronominalized in the second and
  third mentions.
• With no other salient accented items in their
  phrases, the given mentions of these verbs are
  more prominent.

14
Discussion Variation of intensity in verbs

• Example
• so you're going to have to transfer you
  transfer by going to Government Center which is
  inbound
• The increased intensity of the second mention of
  transfer might be due to the change in its verb
  form.
• Similar to Terken Hirschberg, 1994
• Given nouns tend to be accented if they represent
  a different grammatical function from the first
  mention.

15
Conclusions and Future Work

• Evidence of
• Speaker variation in the way they realize
  differences in information status.
• Given verbs tend to be produced with a greater
  intensity than new verbs.
• Nouns and verbs behave very differently.
• Only preliminary results more work needed.
• Future Work
• Repeat and deepen these analyses on larger
  corpora of read and spontaneous speech, and in
  conversation.

16
Effect of Genre, Speaker, and Word Class on the
Realization of Given and New Information
Interspeech 2006 - Pittsburgh, PA

• Agustín Gravano Julia Hirschberg
• agus, julia_at_cs.columbia.edu

Spoken Language Processing Group Columbia
University