Transcript The time-course of prediction in incremental sentence

The time-course of prediction in
incremental sentence processing:
Evidence from anticipatory eye
movements
Yuki Kamide, Gerry T.M. Altman, and
Sarah L. Haywood (2003)
Incremental processing

The incoming inputs are processed
incrementally on a word-by-word basis;
relevant constraints are applied in parallel
to the analysis of the input as it unfolds.
Prediction of thematic roles

The processor can anticipate thematic role
assignments drawing on different sources
of information:

Lexical information about the verb coupled
with discourse-based information about
available entities (Boland et al., 1995;
Altmann, 1999)
E.g., Which preschool nursery/military base did
Hank deliver the machine guns to _ last week?
 Slower reading times in the ‘preschool nursery’
condition

Prediction of thematic roles

Selectional restrictions (Altmann & Kamide,
1999)
E.g., The boy will eat the cake.
The boy will move the cake.
 More anticipatory eye-movements to the target
in the selective condition.

Overview of the study

The present study explores the extent to which
the incremental analysis of a sentence can lead
to the assignment of thematic roles in advance
of linguistic input at which that assignment is
unambiguously signaled.
 Verb-based
information (in combination with a preverbal argument) in English (Experiments 1 and 2)
 In the absence of the verb, morphosyntactic and
semantic constraints extracted from pre-verbal
arguments in Japanese (Experiment 3)
Experiment 1
Does the processor anticipate information
pertaining to the Goal argument?
 Can anticipatory eye movements be found
during an expression that refers to a
different object in the scene?
 Can anticipatory eye movements be
obtained even if there is no explicit task
other than to look and listen?

Experiment 1
Animate Goal
Agent
Theme
Inanimate Goal
Distractor
Inanimate goal
Animate goal
Methods
64 subjects
 18 experimental pictures each paired with
the animate condition and the inanimate
condition
 SMI EyeLink head-mounted eye-tracker

Results
Region 1: No evidence of
anticipatory eye movements
towards the Goal
Region 2: More anticipatory eyemovements towards the
appropriate Goal
Discussion



The processor can anticipate Goal arguments
even during reference to some other object in
the scene in a ‘look and listen’ task.
Some 3-place Vs with an animate Goal allowed
an alternating constituent order casting doubt on
whether the processor anticipated the Goal to be
referred in the 1st or in 2nd post-verbal position.
It is not sure whether the verb alone, or the
combination of the verb with its direct object led
to the anticipatory eye movements.
Experiment 2
Can the arguments of a verb be predicted
on the basis of combinatory information
derived from the semantics of the Agent in
combination with the verbs’ selectional
restrictions?
 The boy ate … vs. The cat ate …

Experiment 2
Method
64 subjects
 24 scenes with 4 sentential conditions

Analyses
Results
Combinatory effects: More looks to the motorbike in the
man ride condition than in the girl ride condition (Regions 1
and 2) and in the man ride condition than in the man taste
condition (Region 2)
Discussion



The semantic properties of a forthcoming Theme
are predicted on the basis of the combination of
information about the Agent and about the verb.
Relatively small number of looks to the target
objects  overt shifts in attention as evidenced
by eye movements may underestimate the true
extent of attentional shifts
Does the prediction of a verb’s arguments have
to be based on the verb itself?
Experiment 3
In Japanese, all arguments of the verb
appear prior to the verb and each
argument is case-marked.
 E.g., syoojo-ga neko-ni sakana-o yatta.
girl-nom cat-dat fish-acc gave
 Is there pre-head (pre-verb) prediction in
Japanese?

Experiment 3
Dative condition
Accusative condition
Prediction: More anticipatory looks towards the hamburger in the Dative
condition than in the Accusative condition.
Method
24 native speakers of Japanese
 16 experimental scenes

Results
Significantly more looks to the Target in the Dative condition than in
the Accusative condition during the adverb
Discussion

Prediction of forthcoming arguments is possible even in
the absence of the grammatical head.


This prediction was in part based on syntactic information
regarding case-structure in Japanese.
?? Combinatory information derived across different NPs

Evidence against head-driven parsing accounts; support
for incremental pre-head attachment.

Preference for analyzing ‘NP-dat’ as the Goal of a 3-place
verb over analyzing it as the Theme or Goal of a
monotransitive verb


Sensitivity to the statistical distributions of particular verb types
The availability in the concurrent visual scene of an object that
could plausibly take on the Theme role.
General discussion



Experiment 1: verb-based info is not limited to
anticipating Themes, but can also anticipate
Goals.
Experiment 2: In combination with info conveyed
by the verb, the Agent can constrain the
anticipation of a subsequent Theme.
Experiment 3: In a head-final construction in
Japanese, properties of the Theme can be
anticipated on the basis of the combination of
lexico-semantic and case-marking information.
General discussion


Is it the linguistic structure or the visual context
that triggers a predictive process?
The processor uses accruing constraints to
compute the likely thematic relationships
amongst the entities already referred to in the
linguistic input and amongst the entities
concurrently available in the visual context.

Incremental, probabilistic processing drawing on
different sources of information at the earliest possible
opportunity to establish the fullest possible
interpretation of the input at each moment in time.