Transcript Snedeker &Trueswell 2003 Lee Presentation 03_16_06

Using prosody to avoid ambiguity:
Effects of speaker awareness
and referential context
Snedeker and Trueswell (2003)
Psych 526
Eun-Kyung Lee
Prosody in Sentence Processing

The role of prosodic information in either
comprehension or production of syntactically
ambiguous sentences
 Based on the finding of the relation between
syntax and prosody
Previous Findings


Inconsistent
 The use of prosodic cues in syntactic
disambiguation is limited (Allbritton et al, 1996)
vs. reliable (Schafer et al. 2000)
Speaker’s reliable use of prosodic cues
depends on
 Whether other cues disambiguate the structure
 Whether speakers are aware of the potential for
ambiguity
Limitations


Data
 Artificially manipulated prosody
 Obtained mostly from trained speakers with
explicit instruction
No examination of interaction between the speaker
and the listener
The Current Paper



Examines the effect of referential context and
awareness on both the production and
comprehension of prosodic cues to structure
Untrained speakers
Target structure
 Globally ambiguous PP attachment
NP attachment
 Tap the frog
with the flower
VP attachment
Referential Communication Task[1]


Referential context: sets of objects
The speaker and the listener are separated by a
divider
 Allows manipulations of referential effects
independently on the comprehension and the
production task
Referential Communication Task[2]


Procedure
 The experimenter demonstrates an action to the
speaker
 The speaker produces a scripted sentence
describing that action
 The listener performs the action described by the
speaker
The effectiveness of prosodic cues is assessed
depending on how well the listener replicates the
experimenter’s action
Research Questions

How prosodic cues are used by the speaker when
the referential context
 supports both meanings of the target sentence
(Experiment 1)
 Strongly favors the intended meaning of the
utterance (Experiment 2)

When the prosodic information is used by the listener,
based on the eye movement test (Experiment 3)
Experiment 1
Methods[1]





32 pairs of participants
Identical sets of toy animals for the speaker and the
listener
Attribute-possessor relation is demonstrated by a
small object attached to them
Speaker’s utterances are audiotaped and the
listener’s actions are videotaped
Post-experiment interview
 to assess participants’ awareness of the
experimental manipulation and the ambiguity in
the critical items
Methods[2]

Target instrument
Stimuli (on each trial)
 Support both interpretations of the ambiguous
sentence by providing
Unmarked animal
 a potential instrument (large flower)
 two possible direct objects (the frogs) for the
VP attachment
Marked animal
 a potential direct object for the NP attachment
(frog holding flower)
Marked animal
Distractor animal
Distractor object
Methods[3]

4 Conditions





4 counterbalanced presentation lists


Ambiguous, instrument demonstration
 Tap the frog with the flower
Ambiguous, modifier demonstration
 Tap the frog with the flower
Unambiguous, instrument demonstration
 Tap the frog by using the frog
Unambiguous, modifier demonstration
 Tap the frog that has the flower
16 target trials, 30 fillers
4 reverse-order lists
Coding


Listener’s actions
 Instrument responses
 Modifier responses
Speaker’s prosody
 Acoustic analysis: word and pause durations
 Tap
the frog with the flower

Phonological analysis
 Break indices for the break following the verb
and the noun
 Presence or absence of pitch accent on the
preposition
Results[1]
Listener’s actions


Proportion of instrument
responses
 66% for instrument
demonstration
 24% for modifier
demonstration
Reliably lower
performance on
ambiguous structure
compared to
unambiguous structure
Results[2]
Acoustic Analysis
Instrument demo.
 For 68% of the trials
 Lengthening of the
direct object (DO)
 Pause after DO
Modifier demo.
 For 40% of the trials
 Lengthening of the
verb
 Pause after the verb
Results[3]
Phonological Analysis
Modifier demonstration

A relatively frequent IP break after the verb
Instrument demonstration


A relatively frequent IP break after DO
Pitch accent on preposition
Results[4]
Phonological Analysis



68% of the trials with appropriate and disambiguating
phrasing
22% with neutral prosodic phrasing
10% with phrasing that was more appropriate
for the alternate interpretation
 Prosodic cues are a highly effective but imperfect
means of syntactic disambiguation
Results[5]
Awareness of Ambiguity

97% of the speakers and 91% of the listeners were
coded as aware of the ambiguity
 Prosodic disambiguation arises due to the speaker’s
awareness of the ambiguity in the critical items
Experiment 2
Methods[1]

Differences in stimuli from Exp.1
 The speaker’s referential context supports only the
intended meaning of the ambiguous phrase




Listener’s context was the same ambiguous context as in
Exp. 1
Listeners and speakers were told that they would receive
an identical set of objects
The type of Demonstration serves as a betweensubject variable
Syntactically ambiguous conditions only
Methods[2]




32 pairs of participants + additional 10 pairs
(unaware pairs / aware pairs depending on the
speaker’s awareness of the ambiguity)
2 lists
 16 critical sentences, 24 fillers
2 reverse order lists
Coding
 Same as in Exp. 1
Results[1]
Listener’s actions


In Exp. 2, 41%
instrument responses
for instrument Demo.
And 34% for modifier
Demo.
Reliable difference in
the distribution of
responses between the
two experiments
(2(1)=4.04, p>.05)
Results[2]
Acoustic Analysis


No significant effect of
Demonstration in both
word and pause
durations in critical
regions (verb, DO)
Reliable but smaller
effect on duration of the
PP, compared to Exp. 1
Results[3]
Phonological Analysis
Instrument Demo.
 Clear distinction
between the rate of 3
coding categories, but
low proportion of correct
phrasing
But for Modifier Demo.
 Greater rate of incorrect
and ambiguous coding
 The relation between
particular prosodic cues
and syntactic structure
is weak and probabilistic
Results[4]
Awareness of ambiguity
Listeners
 As likely to notice the
ambiguity as those in
Exp. 1
Speakers
 6% of the speakers in
the instrument condition
 56% in the modifier
condition -> but
decreased relative to
Exp. 1
 Due to verb bias:
action verbs, more likely
instrument attachment
Results [5]
Awareness and Listener’s Performance

In Modifier condition
 Instrument responses: no significant difference
between when speakers were aware of the
ambiguity (40%) and unaware (38%)
 the speaker’s awareness alone does not
determine prosodic disambiguation
Results [6]
Awareness and Listener’s Performance

In Instrument condition
 Better performance of listeners in Exp. 1 (66%
instrument actions) than in Exp. 2 (41%)
 Speaker awareness seems to have an effect in
contrast to in the modifier condition
Referential context differs
 Speakers only produce reliable disambiguating
prosody when the context doesn’t do the work for
them

Results [7]
Awareness and duration


Small but reliable
differences between
Aware and Unaware
modifier utterances at
the noun, the noun
pause and prepositional
phrase
 listeners were rarely
sensitive to these
differences
No reliable difference
between the Unaware
modifier and instrument
utterances
Summary of Exp. 1, 2


No reliable prosodic cues (enough for listeners to rely
on) produced by speakers in Exp. 2
Speakers provide reliable prosodic cues only when
the referential context is ambiguous and perhaps
when speakers become aware of this
Experiment 3
Goals
Based on the real-world eye-gaze paradigm combined
with the referential communication task

Sees whether the prosodic cues produced by
speakers could shape online interpretation (the
rapidity with which prosody influences parsing)

Examines when and how early the prosodic
information appears in the utterance
Methods





24 pairs of participants
ISCAN eye-tracking visor
Ambiguous referential contexts both for listeners and
speakers
No unambiguous conditions
2 stimulus lists (8 target items, 24 filler items in each
list) + 4 reversing order lists
Results[1]
Actions, prosody, and awareness



Replicate the findings of Exp. 1
Actions & prosody
 Listeners’ responses to the ambiguous sentences
reflected the intentions of the speaker
 Speakers’ prosody clearly varied with intended
structure
 Significant effect of Demonstration in each of the
critical regions
Awareness of ambiguity
 92% of the speakers, 96% of the listeners
Results[2]
Online interpretation

Re-synchronize the utterances at each word and
conduct the analyses on small time windows
 Direct object noun
 Prepositional object
Results[3]
Proportion of fixation to direct object noun
Significant
difference in
fixation to
unmarked animal
Modifier
(150ms)
Program an
eye movement
Instrument
Time
slice 1
200-300
Time
slice 2
300-400
Time
slice 3
400-500
Results[4]
Proportion of fixation to direct object noun

The reliable effect of Demonstration in the 400500ms time slice
 Within 250ms of the onset of the direct object
noun
 At about the same time as phonologically driven
effects (Animal identification)
Results[5]
Prosody vs. phonologically driven effects
Results[4]
Proportion of fixation to direct object noun


The reliable effect of Demonstration in the 400500ms time slice
 Within 250ms of the onset of the direct object
noun
 At about the same time as phonologically driven
effects (Animal identification)
Prosody influences interpretation prior to the
ambiguous region
 May be due to the difference in a pause after verb
Results[6]
Proportion of fixation to prepositional object
Modifier
Marked animal
Target instrument
Onset of the
prepositional object
Instrument
Marginally
significant
(due to ceiling
effect)
Significant
Summary & Discussion[1]

Referential context affects the speaker’s use of
prosodic cues to disambiguate the sentence
 When the context fails to disambiguate the
sentence
 reliably used
 When the context supports only the intended
meaning
 rarely used
Summary & Discussion[2]

Why conflicting findings with Shafer et al.(2000)?
In Shafer et al. (2000)
 Likelihood that ‘NP with PP’ phrases became
lexicalized (e.g. the square with the triangle)
 Subtler contextual cues to disambiguation
 Longer and more complex sentences
Summary & Discussion[3]


Listeners are sensitive to the existing prosodic cues
Prosodic effects are found prior to the onset of the
ambiguous phrases
 Affect the listener’s initial interpretation of
utterance
 Predict material which has yet to be spoken
Thank you!