Agust

1
The Effect of Contour Type andEpistemic Modality
on theAssessment of Speaker Certainty

• Agustín Gravano1 Stefan Benus2 Julia
  Hirschberg1
• Elisa Sneed German3 Gregory Ward3

1 Columbia University 2 Univerzity Konštantína
Filozofa 3 Northwestern University
2
Overview

• Previous researchers disagree about the role of
  epistemic would in utterance interpretation.
• A Whos the British woman over there?
• B That would be J. K. Rowling.
• Epistemic would conveys...
• Tentativeness (Palmer 1990, Perkins 1983)
• A high degree of speaker certainty (Ward et al.
  2003)

3
Overview

• What is the relation between epistemic would and
  perceived speaker certainty?
• What role does the intonational contour play?
• Two perception experiments
• Textual condition
• Spoken condition

4
Epistemic modality

• Marks the speakers estimation of the likelihood
  that a certain proposition is true in context.
• A Whos the British woman over there?
• B That must be J. K. Rowling.
• That could be J. K. Rowling.
• That might be J. K. Rowling.
• How is the perception of speaker certainty
  affected by the use of epistemic would?
• That would be J. K. Rowling.

5
Perception Study 1 Textual ConditionTask
Overview

• Participants were
• 1) Presented with written dialogues.
• 2) Asked to assess the speaker certainty of a
  target utterance.

6
Perception Study 1 Textual Condition Materials

• Context David is at his desk when a co-worker
  knockson the door.
• Co-worker David, I'm looking for this guy
  named Frank Jackson.
• David Thats the new guy.
• or
• That would be the new guy.

7
Perception Study 1 Textual ConditionExperiment
Design

• Each session contained 60 tokens
• 20 stimuli (only one stimulus from each set)
• 40 fillers without any of the target
  constructions
• Presented in a random order.
• Participants rated the perceived certainty of
  each token on a 5-degree Likert scale
• Very uncertain, Somewhat uncertain, Neither
  certain nor uncertain, Somewhat certain, Very
  certain.

8
Perception Study 1 Textual ConditionComputer
Interface
9
Perception Study 1 Textual ConditionCollected
Data

• 12 participants (8 female, 4 male, mean age
  20.3)
• 240 data points (120 would be, 120 is)
• Participants responses were
• 1) Converted into numeric values
• Very uncertain ? ?2
• Somewhat uncertain ? ?1
• Neither certain nor uncertain ? 0
• Somewhat certain ? 1
• Very certain ? 2
• 2) Normalized using z-scores.

10
Perception Study 1 Textual ConditionResults

• Mean certainty of
• would be tokens ?0.13 1.11
• is tokens ?0.03 1.04
• One-way ANOVA No significant difference.
• No evidence of a difference in perceived
  certainty between modal would and indicative be,
  in a textual condition.

11

• What is the case in a spoken condition?
• How is the perception of speaker certainty
  affected by
• the use of epistemic would?
• the use of a particular intonational contour?

12
Perception Study 2 Spoken ConditionTask Overview

• Participants were
• 1) Presented with written dialogues and a
  recorded target utterance.
• 2) Asked to assess the speaker certainty of each
  target utterance.

13
Perception Study 2 Spoken ConditionComputer
Interface
14
Perception Study 2 Spoken ConditionIntonational
Contours

• Simple declarative contour (H) H L- L
• Downstepped contour H !H (!H) L- L
• Yes-no-question contour (L) L H- H

15
Perception Study 2 Spoken ConditionMaterials

• 20 stimulus sets, each with six variations of the
  same utterance ( 120 files)
• Recorded by a non-professional male speaker of
  American English in a sound-proof booth.

declarative downstepped yn-question
would be
is
16
Perception Study 2 Spoken ConditionExperiment
Design

• Each session contained 60 tokens
• 20 stimuli (only one stimulus from each set)
• 40 fillers (with all 3 contours 13 dec, 13 ds,
  14 yn)
• Presented in a random order.
• Participants rated the perceived certainty of
  each token on the same 5-degree Likert scale
• Very uncertain, Somewhat uncertain, Neither
  certain nor uncertain, Somewhat certain, Very
  certain.

17
Perception Study 2 Spoken ConditionCollected
Data

• 30 participants (24 female, 6 male, mean age
  21.4)
• 600 data points
• Again, participants responses were
• 1) Converted into numeric values.
• 2) Normalized using z-scores.

declarative downstepped yn-question
would be 100 100 100
is 100 100 100
18
Perception Study 2 Spoken ConditionResults

• No interaction between Contour and Modality.
• For all 3 contours would be gt is
• For both modalitiesdownstepped gtdeclarative
  gtgtyn-question
• (All stat. significant.)

19
Conclusions

• Epistemic would conveys...
• Tentativeness.
• A high degree of speaker certainty.
• would be gt is
• However, the choice of intonational contour has a
  stronger impact on perceived certainty.
• downstepped gt declarative gtgt yn-question

20
Future Work

• Production study
• Before the textual perception study, the same 12
  participants recorded each target utterance.
• What contours were used to convey different
  degrees of speaker certainty?

21
The Effect of Contour Type andEpistemic Modality
on theAssessment of Speaker Certainty

• Agustín Gravano1 Stefan Benus2 Julia
  Hirschberg1
• Elisa Sneed German3 Gregory Ward3

1 Columbia University 2 Univerzity Konštantína
Filozofa 3 Northwestern University
22
Extra slideSample Stimuli

• A I think the kids are tired of the water park.
  Maybe we should take them someplace else.
• B What's the Six Flags theme park located in
  Gurnee?
• A That is, would be Great America.
• A What a great party!
• B Yeah, but we're stuck cleaning up all the
  crap.
• A Hey, somebody left their iPod out on the
  floor.
• B That is, would be my roommate.

23
Extra slideCertainty Mean and StDev
24
Extra slideFillers
tokencount certainty mean stdev
downstepped 390 0.667 0.435
declarative 390 0.605 0.459
yn-question 420 ?1.299 0.392
ANOVA Significant difference (F(2, 1197)
2778.2, p0) Tukey test Difference is
significant (95) for dsgtyn and decgtyn, and
approaches significance for dsgtdec.
25
Extra slideEpistemic would Form

• Restricted to intransitive sentences
• SUBJECT would VERB POST-VERBAL CONSTITUENT
• Corpus study (Birner et al. 07)
• 246 naturally-occurring tokens, from oral and
  written sources
• Most frequent subjects are demonstratives (79)
• Nearly all verbs are be (98)
• Post-verbal constituent is typically, but not
  necessarily, a noun phrase.