Transcript Sentence Comprehension Lecture 100610

More on predicting word properties in context
Dikker, Rabagliati, Farmer, & Pylkkanen (2010) Psych Science
• MagnetoEncephaloGraphy (MEG) study
– MEG better than EEG for source localization
– BESA (Brain Electrical Source Analyzer) software
• Words have form properties correlated with syntactic
category (Farmer et al., 2006; Monaghan et al., 2007)
– Function morphemes, but also phono/ortho properties
– Words with properties more typical of their category are read faster
• Dikker et al. (2009) found M100 sensitive to function
morphology on content words
– M100 (100-130 msec) larger when content words include salient
unexpected/ungrammatical function morpheme
– e.g., The discovery was in the report / reported.
– Brain source = occipital (visual) cortex!!!
10/6/10
Psyc / Ling / Comm 525 Fall 10
Dikker, et al (2010) cont’d
Is salient category-typical morphology required?
Example Stimuli
- Nouns w/ derivational noun morph: princess, farmer
- Nouns w/ non-morph-marked but typical form: soda
- Nouns w/o form-correlated properties: infant
- (Form typicality derived from large corpus analysis by Farmer et
al., 2006 and Monaghan et al., 2007)
Presented in context that predicts:
- Noun: The tasteless soda …
- Participle: The tastelessly soda …
10/6/10
Psyc / Ling / Comm 525 Fall 10
In participle-predicting contexts:
M100 bigger for morphologically marked nouns
- Replicates Dikker et al. (2009)
M100 also bigger for nouns with more noun-like
phonology/orthography
- So, salient function morphology not necessary
No M100 difference for neutral nouns
- Also replicates Dikker et al. (2009)
10/6/10
Psyc / Ling / Comm 525 Fall 10
Probabilistic Word Prediction
The more the form properties of a word are typical for its syntactic category,
the bigger the M100 in a context that predicts a word from a different category
- By 100-130 msec (!) in visual regions sensitive to letter/non-letter distinctions (!)
- Dikker argues it’s visual word form properties that don’t match prediction, NOT
syntactic category per se
10/6/10
Psyc / Ling / Comm 525 Fall 10
Issues in Sentence Comprehension
• Parallel vs Serial
– Parallel = Construct multiple interpretations & wait for evidence
about which is right
– Serial = Construct only 1 interpretation at first, & then try
another only if that turns out to be wrong (two stages)
• Which makes more sense depends on whether easier to
– Construct multiple interpretations & hold onto them long enough
to avoid making a mistake
– Or construct just one interpretation & hope it’s right
– If first choice usually right, serial approach could be more efficient
10/6/10
Psyc / Ling / Comm 525 Fall 10
• Modular vs Interactive
– Modular = Take only syntactic properties of words (N, V,
Prep, etc.) & phrase structure options into account at first
– Interactive = Take word meaning, context, & general
knowledge into account all along
• Again, which makes more sense depends on whether
it’s easier to
– Make quick first guesses based on just syntactic categories
of words & some basic phrase structure patterns
– Or combine syntax, meaning, context, & knowledge fast
enough in first place
10/6/10
Psyc / Ling / Comm 525 Fall 10
Garden Path Model
(Frazier, Rayner, Clifton, F. Ferreira)
• Influential serial (two-stage) modular
parsing model
– Name comes from claim that we frequently gardenpath during comprehension
• But only become aware of it occasionally
• Minimal Attachment Strategy
– Whenever multiple structural options, start with
simplest one
• i.e. one requiring adding fewest nodes to phrase
structure tree at that point in sentence
10/6/10
Psyc / Ling / Comm 525 Fall 10
Structural Ambiguity
Attachment ambiguity
S
S
Nonminimal
VP
Minimal
VP
NP
NP
V
NP
PP
A thief shot the cop in the park.
10/6/10
NP
V
NP
PP
A thief shot the cop in the park.
Psyc / Ling / Comm 525 Fall 10
Prediction of GP Model
• All sentences with Nonminimal Attachment (NMA)
structures should be harder than all sentences with
Minimal Attachment (MA) structures
– Because always have to revise a wrong first guess in NMA
sentences
– That should take some time
• Though often not enough to become aware of problem
10/6/10
Psyc / Ling / Comm 525 Fall 10
Rayner, Carlson, & Frazier (1983)
• Measured how long people spent reading different regions of
sentences using an eyetracker
• Stimuli:
– Minimally Attached:
The doctor examined the patient with the stethoscope,
but he couldn’t figure out what was wrong.
– Non-minimally Attached:
The doctor examined the patient with the headache,
but he couldn’t figure out what was wrong.
• Prediction of GP Model:
– People should read headache & the words after it more slowly than
stethoscope & the words after it
10/6/10
Psyc / Ling / Comm 525 Fall 10
Results – First Pass Times
10/6/10
Psyc / Ling / Comm 525 Fall 10
Interpretation
• Results support GP Model
– Slower reading times after headache suggest people
garden-pathed & had to reinterpret
• But, do Rayner et al.’s stimuli provide fair test of
MA Strategy?
– Temporary ambiguity starts at with
– Is there anything earlier in sentence that might bias
interpretation one way or other?
– The verb examined, maybe?
• Does it lead to expectations for certain kinds of
phrases following it? (Bresnan)
10/6/10
Psyc / Ling / Comm 525 Fall 10
Verb Argument Structure
• Verbs are the most important words in sentences (in English)
– All other words are interpreted relative to the verb
– The librarian put the book on the shelf.
• put requires all 3 of these arguments to be in the sentence,
• i.e. all 3 arguments are obligatory
• Subject / Agent
= librarian
• Object / Patient (Theme)
= book
• Location
= shelf
* The librarian put.
* The librarian put the book.
• The librarian put on the shelf.
• Most verbs have some optional arguments
– The doctor examined the patient.
– The doctor examined the patient with a stethoscope.
– * The doctor examined.
10/6/10
Psyc / Ling / Comm 525 Fall 10
Verb Bias
• Verbs differ in how often they are used with their
optional arguments
– examine is probably often used with an optional
Instrument
• Maybe Rayner et al. happened to use lots of verbs
that often take the kinds of arguments that were
present in MA versions?
– Would bias results in favor of MA sentences
10/6/10
Psyc / Ling / Comm 525 Fall 10
Taraban & McClelland (1988)
• Measured reading times using self-paced moving
window technique
• Stimuli:
– Used Rayner et al.’s materials
– Plus just as many more like:
MA: John read the article in the bathtub
while he was waiting for a call.
NMA: John read the article in the magazine
while he was waiting for a call.
• Where John read the article in the ... seems to lead to
expectation of
– something modifying article (= NMA) rather than a location
where the reading event took place (= MA)
10/6/10
Psyc / Ling / Comm 525 Fall 10
Results
10/6/10
Psyc / Ling / Comm 525 Fall 10
Taraban & McClelland’s Conclusions
• No general Minimal Attachment Strategy
• People use knowledge about particular words & how
they’re most likely to be used to guide their
interpretation
– Especially verbs?
• Serial vs Parallel
– Supports serial model – should be no GP if fully parallel
• Modular/Interactive
– If verb bias is syntactic, results have no consequence
– If word meaning or extra-sentential context had an
immediate effect, that would clearly support Interactive
10/6/10
Psyc / Ling / Comm 525 Fall 10
Another Kind of
Temporary Structural Ambiguity
“Marge Schott, managing partner of the Cincinnati Reds,
at first did not want to apologize for her remark that
Hitler‘was good at the beginning but he just went too
far’. Under pressure, she finally said that she regretted
her remarks ‘offended many people’.”
- NY Times, 7/21/96,
D. Tannen, I’m sorry, I won’t apologize
10/6/10
Psyc / Ling / Comm 525 Fall 10
that
The referees warned the spectators against heckling the other team.
would probably get too rowdy.
^
Temporary ambiguity about relationship
between the Verb & the Noun after it
- Is the noun a
- Direct Object (DO) (= Minimal Attachment), or
- Subject of an Embedded Clause (= NMA)?
- Ambiguity arises because that is optional
(in English – obligatory in German, Dutch)
10/6/10
Psyc / Ling / Comm 525 Fall 10
Verb Bias
• Verbs differ in how often they’re used in particular
sentence structures
– Compare The referees warned the spectators ...
with The bus driver worried the passengers ...
warned = Direct-Object Biased Verb
worried = Clause-Biased Verb
Does a verb's most likely use guide initial choices
about the interpretation of words following it?
10/6/10
Psyc / Ling / Comm 525 Fall 10
Plausibility of Noun
as Direct Object
• Plausibility of particular Verb + Noun combos in
particular relationships vary
– Compare The referees warned the spectators ...
with The referees warned the game ...
– Meaning of game completely rules out DO possibility
• Does that lead people to expect a verb following game because it
must be the beginning of an embedded clause?
• If yes, how long does it take to develop that expectation?
10/6/10
Psyc / Ling / Comm 525 Fall 10
Example Stimuli
(Garnsey et al., 1997)
DO-Bias Verbs (warned, discovered, heard, …)
The referees warned (that) the spectators would probably get too rowdy.
The referees warned (that) the game
would probably go into overtime.
Clause-Bias Verbs (worried, realized, suspected, …)
The bus driver worried (that) the passengers were starting to get annoyed.
The bus driver worried (that) the tires
were starting to go flat.
Equi-Bias Verbs (regretted, predicted, knew, …)
The senior senator regretted (that) the decision had ever been made public.
The senior senator regretted (that) the reporter had ever seen the report.
10/6/10
Psyc / Ling / Comm 525 Fall 10
Stimulus Norming & Control
Verb Properties
Verb Type
Freq
Length
DO-Bias
Cl-Bias
that-Pref
DO-Bias
146
8.3
14%
75%
89%
Equi-Bias
178
7.6
36%
38%
71%
Clause-Bias
128
7.7
79%
12%
67%
Ambiguous Noun Properties
Plausible as DO
Implausible as DO
Freq
Len
DO
Rate
Cl
Rate
Freq
Len
DO
Rate
Cl
Rate
DO-Bias
132
6.3
6.5
5.7
166
5.5
2.9
5.4
Equi-Bias
106
6.5
6.7
5.9
62
6.3
2.9
5.7
Clause-Bias
67
6.7
6.9
6.3
47
6.3
2.6
6.1
Verb Type
10/6/10
Psyc / Ling / Comm 525 Fall 10
Reading Time Results
(Same pattern in button-pushing moving window & eyetracking first pass)
DO-Bias Verbs
The referees warned the spectators would probably get too rowdy.
The referees warned the game
would probably go into overtime.
read slowly
Clause-Bias Verbs
The bus driver worried the passengers were starting to get annoyed.
The bus driver worried the tires
were starting to go flat.
read slowly
Equi-Bias Verbs
The senior senator regretted the decision had ever been made public.
The senior senator regretted the reporter had ever seen the report.
10/6/10
Psyc / Ling / Comm 525 Fall 10
Verbs Rule
• No immediate effect of Plausibility of V + N
combination when V had a strong bias
• But Plausibility did have an immediate effect
when no V-bias
• So, the two factors interact, with Verb Bias
dominating the interaction
• Consistent with Constraint-Satisfaction Models,
with some kinds of constraints being more influential
10/6/10
Psyc / Ling / Comm 525 Fall 10
Event-Related Brain Potentials (ERPs)
& Language
• N400 (Plausibility, Contextual fit)
– Negative, 400 msec
– Centro-parietal maximum, sometimes larger on right
• P600 (Garden-paths, grammar errors, complexity,…)
– Positive, >500 msec
– Centro-parietal maximum
10/6/10
Psyc / Ling / Comm 525 Fall 10
Stimulus Presentation
in ERP Study
Were the referees
expecting
overtime.
probably
referees
warned
would
game
The
into
the
go 1 a long game?
READY
QUESTION
10/6/10
Psyc / Ling / Comm 525 Fall 10
Waveforms Starting at Temporarily Ambiguous Noun
Pz
DO-Bias Verbs (warned)
Final
word
-6
Plausible Noun:
The referees warned the spectators ...
-3
0
3
6
0
650
1300
1950
2600
Msec
spectators
10/6/10
Psyc / Ling / Comm 525 Fall 10
Question
Waveforms Starting at Temporarily Ambiguous Noun
N400
Pz
DO-Bias Verbs (warned)
-6
Plausible Noun:
The referees warned the spectators ...
Implausible Noun
The referees warned the game ...
-3
0
3
6
0
650
1300
Msec
spectators/
game
10/6/10
Psyc / Ling / Comm 525 Fall 10
1950
2600
Waveforms Starting at Temporarily Ambiguous Noun
DO-Bias Verbs (warned)
Plausible Noun:
The referees warned the spectators ...
Implausible Noun
The referees warned the game ...
N400
Pz
-6
-3
0
3
6
0
650
1300
1950
2600
1950
2600
Msec
spectators/
game
Clause-Bias Verbs (worried)
-6
-3
Plausible Noun:
The bus driver worried the passengers ... 0
3
Implausible Noun:
6
The bus driver worried the tires ...
0
650
1300
Msec
passengers/
tires
10/6/10
Psyc / Ling / Comm 525 Fall 10
Waveforms Starting at Temporarily Ambiguous Noun
DO-Bias Verbs (warned)
Plausible Noun:
The referees warned the spectators ...
Implausible Noun
The referees warned the game ...
N400
Pz
-6
-3
0
3
6
0
650
1300
1950
2600
1950
2600
Msec
spectators/
P600
game
Clause-Bias Verbs (worried)
-6
-3
Plausible Noun:
The bus driver worried the passengers ... 0
3
Implausible Noun:
6
The bus driver worried the tires ...
0
650
1300
Msec
passengers/
tires
10/6/10
Psyc / Ling / Comm 525 Fall 10
Verbs still rule. Why?
• Principled reasons:
– They provide the most useful information
– Information about them may be simply retrieved,
while plausibility requires combining information from
multiple words
• More accidental reasons:
– They precede the critical nouns in these sentences
[ but, see Trueswell (1996) ]
– They generally appear early in English sentences,
leading English speakers to rely on them
10/6/10
Psyc / Ling / Comm 525 Fall 10
Back to Usual Issues
• Effect of plausibility supports Interactive ConstraintSatisfaction Models
• With some kinds of info outweighing &/or available earlier than
others
• Any evidence about Serial vs Parallel?
• Reliable correlations between RT at disambiguation & strength of
bias toward less preferred structure when plausibility supported
the less-preferred structure
• Clause-bias V + Plausible-as-DO N:
- Difficulty increased as DO-bias strength of V increased (r = +.53)
• DO-bias V + Implausible-as-DO N:
- Difficulty decreased as Cl-bias strength of V increased (r = -.58)
• Strongly suggests weighing both options = parallelism
10/6/10
Psyc / Ling / Comm 525 Fall 10
GP Modelers’ Response
• Studies show various factors make easier the
structures predicted by GP Model to be hard
– So, all of these effects COULD be during 2nd stage
reanalysis, since should only try harder option when
forced to reanalyze
• To be definitive, need to show that same factors can
make harder the structures predicted by GP Model to
be easier
– Then effects can’t be due to reanalysis, since no
reanalysis should be necessary according to GP Model
10/6/10
Psyc / Ling / Comm 525 Fall 10
Wilson & Garnsey (2009)
• Are sentences with DO structures made harder by
having a Clause-bias verb?
• Sentences with either Clause or DO structures
DO-bias V
The ticket agent admitted the mistake might not have been caught.
The ticket agent admitted the mistake because she had been caught.
Clause-bias V
The CIA director confirmed the rumor could mean a security leak.
The CIA director confirmed the rumor when he testified to Congress.
10/6/10
Psyc / Ling / Comm 525 Fall 10
Moving Window RT Results
at disambiguating region
80 a
C
Clause with "that"
B
Clause without "that"
60
J
Sentence with DO
B
Sentence with Clause
B
B
40
J
B
B
20
b
C
C
Clause-bias
DO-bias
J
0
Clause-bias
Verb Bias
10/6/10
Psyc / Ling / Comm 525 Fall 10
DO-bias
Verb Bias
Eyetracking First Pass Times
at disambiguating region
50 a
C
Clause with "that"
B
Clause without "that"
b
J
Sentence with DO
B
Sentence with Clause
25
B
B
0
-25
B
B
J
C
C
J
-50
Clause-bias
DO-bias
Clause-bias
Verb Bias
10/6/10
Psyc / Ling / Comm 525 Fall 10
DO-bias
Verb Bias
Regression-Path Times
at disambiguating region
(also called Right-Bounded or Go-Past Times)
100
B
75
50
J
25
0
J
Sentence with DO
B
Sentence with Clause
B
Time from when:
J
-25
-50
Clause-bias
DO-bias
Verb Bias
10/6/10
- First enter a region
- Until leave that region with
forward-going saccade
-Includes time regressing
back to previous regions
Psyc / Ling / Comm 525 Fall 10
Back to Usual Issues
• So, evidence supports both Parallelism &
Interaction of multiple within-sentence
constraints
• And shows Verb Bias effects in sentences that
should not require any reanalysis according to GP
– But Verb Bias effects could be part of 1st stage, since may
be a kind of syntactic knowledge
– And N+V plausibility effects are only testable in sentences
with Clause structure, so could be due to reanalysis
• What about constraints from outside sentence?
10/6/10
Psyc / Ling / Comm 525 Fall 10
The Role of Prosody
• Embedded Clause / Direct Object sentences can
be disambiguated with prosodic phrasing
• Acoustic correlates:
–
–
–
–
–
10/6/10
Pause
Pre-boundary lengthening
Pitch contour
Pitch reset
…
Psyc / Ling / Comm 525 Fall 10
How does this kind of prosodic
boundary marking influence sentence
interpretation?
• Example of DO Prosody
– The basketball star accepted the contract…
…because it paid so well.
Example of Clause Prosody
- The basketball star accepted …
… the contract required him to play every game.
10/6/10
Psyc / Ling / Comm 525 Fall 10
Do speakers produce different prosody
in DO and Clause structures?
Gahl & Garnsey (2004)
10/6/10
Psyc / Ling / Comm 525 Fall 10
Boundaries marked more strongly when
Structure not consistent with Verb Bias
10/6/10
Psyc / Ling / Comm 525 Fall 10