The horse raced by…

advertisement
Sentence Comprehension
Where are we?
Levels of processing in language processing
– Speech/letter perception
– Word recognition
• Lexical access vs. lexical selection
– Sentence comprehension
• Syntactic (is this a module?)
• Semantic analysis
– Discourse/Message level
• Pragmatics
Word Recognition and
Sentence Comprehension
 Word recognition is an essential input to
sentence comprehension.
 What does this mean?
 In both reading and listening, we begin to
develop syntactic & semantic representations
of the sentence as soon as we hear the first
words (recall Marslen-Wilson, 1973).
• The professor explained…
• The horse raced…
Garden Path Sentences
• The horse raced past the barn.
[ NP
][
VP
]
• The horse raced past the barn fell.
[ NP
[
relative clause
] ] [ VP ]
 We build sentence-level structures and
interpretations incrementally.
 There is a lot of temporary ambiguity.
 Sometimes we choose the wrong alternative.
Incremental Parsing &
Interpretation
Garden path sentences (Bever, 1970)
– The horse raced past the barn fell.
(Main Clause/Relative Clause ambiguity)
GP depends upon whether first NP is a good
agent/experiencer, and on frequency of past
participle form of verb (Trueswell, 1996).
– The thief (who was) searched by the police had weapons.
[good agent, infreq participle]
– The award (that was) accepted by the man was
impressive. [poor agent, freq participle]
• Support for Constraint-based Lexicalist Account of
syntactic processing (parsing)
Two critical components of
Sentence Comprehension
• Syntactic Analysis
• Semantic Interpretation
• Within a language, syntactic structure largely
determines meaning (or vice versa)
• Put the ball in the box on the table.
 Can semantic interpretation guide syntactic
analysis in a top-down manner?
 Constraint-based lexicalist – YES
 Frazier’s Garden Path model -- NO
Sentence Comprehension difficulty is not
always related to ambiguity/garden path!
The paper that the pencil that the chalk touched
marked burned.
Even without ambiguity, it is difficult to identify
the hierarchical structure. Once you figure out
the structure, the meaning becomes clear...
The paper
burned.
that the pencil
marked __
that the chalk touched __
Do semantic cues help?
• The water that the boy that the bee stung
carried spilled.
The water
spilled.
that the boy
carried __
that the bee stung __
Frequently, processing difficulty is
caused by a structural ambiguity.
The cotton shirts are made from grows in Georgia.
– (The cotton shirts) (are made from...)
– (The cotton (shirts are made from)) (grows in...)
• Two syntactic structures are possible. If you
select the wrong one, you “garden path.”
• But ambiguity doesn’t necessarily cause
processing difficulty; we’re usually unaware of
structural ambiguities.
• Theories of sentence comprehension should be
able to predict processing difficulty in both
ambiguous and unambiguous sentences.
• Structurally ambiguous sentences have been
especially important in testing theories, because
an important distinction among theories is the
role of top-down contextual information in
syntactic ambiguity resolution.
Frazier’s Garden Path Theory
1. As you recognize each word,
syntactic category is input to
“parsing module”.
2. Syn cat inputs trigger
operation of PS rules. When
input is consistent with
multiple PS rules…
a) Add new constituent
using minimal number of
nodes. “MAS”
b) Given 2 equally minimal
options, attach to current
constituent. “LC”
3. Pass structure to semantic
interpreter.
Is this the
Minimal
Attachment?
Predictions
• Processing difficulty at point of
disambiguation whenever
minimal attachment (or late
closure) turns out to be wrong.
 Potential Problem: A lot hinges
on how you draw the trees (i.e.,
exactly what are the PS rules?)
A well-specified
(but very
specific)
Constraintbased model
The horse raced by…
This model predicts local
difficulty at the verb in a
reduced relative/main
clause ambiguity.
Using multiple constraints to
resolve syntactic ambiguities
• Frazier’s approach ignores relevant information in the
interests of reducing processing load
• Constraint-based approaches assume that we have
the capacity to use constraints from any level of
representation immediately
– The horse raced past the barn fell.
– The guests invited to the party arrived.
 Is initial NP a better subject or direct object of first verb?
 Is there a unique referent for the NP in the discourse, or do you
need a modifier to tell you which horse/guest?
 Is the past tense or the participle more frequent, for this verb?
How do we get from the words to a
sentence-level representation?
• Word could trigger rules in a phrase structure
grammar, which generate hierarchical structure
(as in Frazier’s Minimal Attachment approach):
The = Det; Det + N = NP
Horse = N
Raced = V; V = VP, V + NP = VP
• The lexicon might actually encode structure, so
that pieces of PS trees are accessed during
word recognition and simply pieced together at
matching nodes. (See Figure 3 of F&B,2004)
Most words are associated with multiple “trees”
I like math.
eat like a pig
friends like mine
John ate an apple like...
Which one do you use?
Syntactic ambiguity
resolution as lexical
ambiguity resolution.
John ate an apple like...
Sentence understanding so far…
• Syntactic and semantic representations built up
incrementally
– Garden path sentences
– Difficult, but unambiguous structures
• Frazier’s parser is automatic & dumb
– Ignoring context/semantics, build simplest structure
allowed by PS rules
• Constraint-based lexicalist parser is interactive
and smart
– Use any relevant knowledge to choose best structure
at point of ambiguity
Do Chinese readers use verb transitivity bias to
immediately interpret EC?
Hsieh & Boland, in progress
Judge help ___ rob…
… and steal. (Judge help PRO rob and steal)
…de poor people. (Judge help [__i rob RC] peoplei)
At “rob”, you must posit an EC, but “help” allows
for either PRO (implausible) or trace.
Transitivity bias of “help” is cue for analyzing EC
– Intransitive/Subject Control reading = PRO
– Transitive reading = Trace
Manipulate transitivity of 1st Verb
PRO Ambig (Intransitive-biased)
那個法官幫忙搶劫和偷竊
The judge help ? rob and steal
“The judge helps to rob and steal.”
PRO: transitivity bias;
immediate interp
Trace: plausibility
Anomaly
PRO Unambig (Oblig Intransitive)
那個法官打算搶劫和偷竊
The judge intend PRO rob and steal
“The judge intends to rob and steal.”
PRO: immediate interp
Trace Ambig (Transitive-biased)
Trace: transitivity bias;
那個法官記得搶劫的窮人
plausibility
The judge remember ? rob RC poor people
“The judge remembers the poor people who robbed.”
No Anomaly
Trace Unambig (Oblig Transitive)
那個法官同情搶劫的窮人
The judge sympathize-with __ rob RC poor people
“The judge sympathizes with the poor people who robbed.”
Results
PROAmbig
PROUnambig
TraceAmbig
TraceUnambig
First Fixation Dur
350
250
200
150
T he
judge
V
rob
PROAmbig
PROUnambig
TraceAmbig
TraceUnambig
Total Time
Transitivity Bias
guides analysis
of EC
700
600
RT
RT
300
500
400
300
The
judge
V
rob
Implications
As in English (e.g., subject/object
ambiguity), verb transitivity bias is used to
guide parsing.
– Verb transitivity cue outweighed plausibility
cue
• Because it was more rapidly available?
• Possible that both alternatives were considered in
ambiguous conditions
Syntactic Parallelism Debate
• Lexical & semantic knowledge guide selection of
best analysis at syntactic ambiguity.
– Trueswell (1996)
The award accepted… vs. The thief searched…
• Processing costs (garden path) observed if
structure is later disambiguated in another
manner.
• Support for the hypothesis that ONE syntactic
structure selected & maintained [Serial
Account]?
– Such findings do rule out full parallelism/delayed
interpretation
Syntactic Parallelism Debate
• Serial accounts must explain why some re-analyses are
costly and some are not.
– The horse raced past the barn… fell.
– Ron believes Rex …is a threat. [Pritchett, 1992]
– Reanalysis costs linked to similarity of potential analyses or
informativeness of disambiguating cue [Abney, 1989; Fodor & Inoue, 1994;
Konieczny, 1996; Lewis,1998, Pritchett, 1992]
• Ranked parallelism w/costs for re-ranking [Gibson, 1991;
Gorrell, 1987; Hickok, 1993; Tabor & Hutchins, 2004]
– Re-ranking costs predicted by…
• Relative support for each analysis
• “Digging in” effects (Tabor & Hutchins, 2004)
• Distinguish between Serial+Reanalysis accounts &
Limited, Ranked Parallel accounts
Hsieh et al. Paper
Sharon & Lucy
Hagoort (2008)
• Is brain imaging of use for cognitive theory?
– Fodor: does not help to know it happens somewhere
north of the neck
– Critics focus on fmri
• Cog Neuro toolkit includes ERP & MEG, TMS
• Should we be interested in the machinery that
instantiates human language, or just cognitive
architecture?
• How does ERP research support the Immediacy
Assumption (all types of info brought to bear
immediately)?
Syntactic & Semantic Processing
• Structure and Interpretation co-vary:
– Put (the book) (in the box on the table).
– Put (the book in the box) (on the table).
• Does structural analysis necessarily
license interpretation?
Syntactic & Semantic
Processing in Tandem
• Constraint-based Lexicalist models (e.g.,
MacDonald et al., 1994)
• Boland’s Concurrent model (1997)
• Kuperberg (2007)
The thief (who was) searched by the police had weapons.
[good agent, infreq participle]
The award (that was) accepted by the man was
impressive. [poor agent, freq participle]
Initial attachment of verb to PS tree is influenced by
thematic fit and frequency of alternative verb forms
Boland’s Concurrent Theory
Syntactic
Selection
The award accepted…
Syntactic
Generation
Semantic
Interpretation
subcat, number,
thematic roles
Accepted –active,
case. past
Accepted –passive,
participle
award-Theme
Agent accepted…
Word Recog
accepted
Theme accepted… (by…)
Syntactic Primacy Theories
•
•
Frazier’s Garden Path model
(1978 & later)
Friederici’s 3-Stage Model (e.g.,
2002)
1. Attach each new constituent to
developing phrase structure tree
2. Lexico-semantic processing & thematic
role assignment
3. Syntactic integration & revision
Trueswell’s ‘good agent/freq participle’
condition would have an advantage here
Semantics
Syntax
Adj + N = NP
Adj, N…
Word Recognition
Testing Parsing Theories
• To what extent does semantic
interpretation depend upon successful
syntactic analysis?
Semantics
Syntactic
Selection
Syntactic
Generation
Semantic
Interpretation
subcat, number,
case.
thematic roles
Word Recog
Syntax
Adj + N = NP
Adj, N…
Word Recognition
Primacy of Syntactic
Processing
• Friederici & Weissenborn (2007, p. 54) “semantic
integration of words does not take place for words
which are not syntactically licensed (have wrong word
category).”
• Investigate with ERP’s, because different
components elicited by syntactic &
semantic
anomalies.
• A syntactic category violation…
Der Priester wurde vom *gebaut…
The priest was by-the *built …
[read German/French sentence, then judge acceptability]
…elicits a LAN, followed by a P600, but no N400
effects (Friederici et al., 1999, 2004; Hahne & Friederici, 2002,
Experiment 1; Hahne & Jescheniak, 2001; Isel et al., 2007)
Boland’s Concurrent Theory
Der Priester wurde vom *gebaut
The priest was by-the *built …
Syntactic
Selection
Syntactic
Generation
Semantic
Interpretation
subcat, number,
case.
<NP +dative; -fem>
thematic roles
[Agent]
Word Recog
vom
Counter-evidence ?
Ainsworth-Darnell, Shulman, & Boland (1998)
entrusted:
< NP(Theme) PP(Recipient) >
Control:
Syntactic Anomaly:
Semantic Anomaly:
Double Anomaly:
Jill entrusted the recipe to friends…
Jill entrusted the recipe friends…
P600
Jill entrusted the recipe to platforms… N400
Jill entrusted the recipe platforms…
N400 + P600
Is Syntactic Primacy Universal?
Zhang, Yu, & Boland (under revision)
In German, syntactic category is well-marked (nouns capitalized, casemarkings on determiners and adjectives, verb tense affixes,
agreement morphology, etc.)
Der Priester wurde vom Arbeiter. (expected N)
Der Priester wurde vom gebaut. (ungrammatical V)
N400 observed if told to focus on meaning of sentence
In Chinese, syntactic category is not marked by morphology on the
words themselves
李薇把新鲜的鸭梨慢慢地削了两个。 (expected V “peeled”)
李薇把新鲜的鸭梨慢慢地刀子了两个。(ungrammatical N “knife”)
Some theories assume syntactic primacy (e.g., Frazier, 1978,
Friederici, 2002), others do not (e.g., Boland, 1997; Kuperberg, 2007;
MacDonald et al., 1994).
Could Syntactic Primacy
be a LG-specific strategy?
Strongest evidence for Syntactic Primacy comes from German.
In German, syntactic category is well-marked (nouns capitalized, casemarkings on determiners and adjectives, verb tense affixes,
agreement morphology, etc.)
Der Priester wurde vom Arbeiter… (expected N)
Der Priester wurde vom gebaut… (ungrammatical V)

N400 is observed if told to focus on meaning of sentence
Zhang, Yu, & Boland (under revision)
In Chinese, syntactic category is not marked by inflectional morphology
on the words themselves
李薇把新鲜的鸭梨慢慢地削了两个。 (expected V “peel”)
李薇把新鲜的鸭梨慢慢地刀子了两个。(ungram N “knife”)
Stimuli & Results
Correct
女孩/买/了/裙子/和/手套。 The girl/ bought/ le (PERF)/ skirt/ and/ glove.
(The girl bought a skirt and a glove.)
Syntactic violation
LAN + P600
女孩/买/了/很/裙子/和/手套。The girl/ bought/ le (PERF)/ very/ skirt/ and/ glove.
(The girl bought a very skirt and a glove.)
Semantic violation
N400
女孩/吃/了/裙子/和/手套。
The girl/ ate/ le (PERF)/ skirt/ and/ glove.
(The girl ate a skirt and a glove.)
Syntactic & Semantic violation
LAN + P600 & N400
女孩/吃/了/很/裙子/和/手套。The girl/ ate/ le (PERF)/ very / skirt/ and/ glove.
(The girl ate a very skirt and a glove.)
•
•
PERF = perfective marker
Critical word in bold.
Suggests that syntactic & semantic
processing go on in parallel
Manipulate degree of Semantic
violation at syntactic anomaly
Condition
Correct
Semantic violation
Syntactic category &
weak semantic
violation
Syntactic category &
strong semantic
violation)
Example
李薇/把/新鲜的/鸭梨/慢慢地/削/了/两个。
Wei Li/ ba (PREP)/ fresh/ pears/ slowly/ peeled/ le (PERF)/ two.
(Wei Li peeled two fresh pears slowly.)
李薇/把/新鲜的/鸭梨/慢慢地/胁迫/了/两个。
Wei Li/ ba (PREP)/ fresh/ pears/ slowly/ intimidated/ le (PERF)/ two.
(Wei Li intimidated two fresh pears slowly.)
李薇/把/新鲜的/鸭梨/慢慢地/刀子/了/两个。
Wei Li/ ba (PREP)/ fresh/ pears/ slowly/ knife/ le (PERF)/ two.
(Wei Li knife two fresh pears slowly.)
李薇/把/新鲜的/鸭梨/慢慢地/钢琴/了/两个。
Wei Li/ ba (PREP)/ fresh/ pears/ slowly/ piano/ le (PERF)/ two.
(Wei Li piano two fresh pears slowly.)
Larger N400 for strong semantic
violation compared to weak
semantic violation, despite syntactic
anomaly
Difference Map Showing the Scalp Distributions of
the Strong-Weak Effects
View of top of head; the anterior is on top
Ye, Luo,
Friederici, &
Zhou (2006)
“ba” construction;
auditory
N400
sustained anterior
negativity (no P600)
sustained anterior
negativity (no P600)
+ “N400-like effect”
Summary Across LGs
• English evidence is mostly consistent with
syntactic & semantic processing in tandem
• French & German readers switch between
“syntactic primacy” mode & “interpret anyway”
mode, depending on instructions.
• Chinese stay in “interpret anyway” mode.
– Can you force Chinese (or English) readers into
“syntactic primacy mode”?
– Does Chinese lead to a greater reliance on semantic
processing than German?
• No reason to favor sentence processing theories
that maintain syntactic primacy.
Download