Transcript Workshop on Constructions and Grammatical

HPSG 2007: Workshop on Constructions and Grammatical Theory
Stanford University
July 21, 2007
The role of default constructions in the
processing of mismatch: the case of
possessive free relatives
Elaine J. Francis
Purdue University
[email protected]
1
Why is canonical form simpler for
processing?

Fewer dependencies involving gaps/traces (e.g.,

Lexical biases of verbs (Gahl et al 2003, Menn 2000,

Canonical sentence templates used for initial
interpretation of clauses (Ferreira 2003, Townsend &
trace-deletion hypothesis in aphasic comprehension,
Grodzinsky 1995)
Menn et al 2003)
 Verbs like kick, chase, break are biased toward an active
interpretation.
 Verbs like elect, injure are biased toward a passive
interpretation.
Bever 2001; similar ideas in Piñango 2000, Jackendoff
2007)
2
Canonical templates and “good-enough”
processing

Canonical sentence template:
Linear order:
Semantic role:



NP
Agent
The dogs
V
Action
destroyed
NP
Patient
the garden.
Townsend and Bever (2001) interpret numerous studies showing
slower processing for non-canonical clauses in terms of violation
of simple templates such as this one.
Ferreira (2003) shows that sentences which violate the template
(passives, object clefts) are misunderstood significantly more
often than sentences that conform to the template (actives,
subject clefts), despite being unambiguous sentences with no
“garden path” structures and no subordinate clauses.
These authors hypothesize that listeners/readers use simple
templates for a rough and ready interpretation of sentences
before full syntactic and semantic processing is complete.
Incorrect interpretations may linger.
3
Canonical templates as linear order
default constructions




Psycholinguistic evidence from studies such as Ferreira
(2003) supports the existence of canonical templates in
grammar.
Although Townsend & Bever (2001) adopt a derivational
approach to syntax, putting canonical templates into a
special level of pseudo-syntax, the idea actually fits more
easily into a parallel-architecture, constructionist view of
grammar.
Canonical templates can be understood as default
constructions that specify basic mappings between linear
order (not hierarchical structure) of constituents and
semantic roles.
Specific constructions such as passive contain the relevant
information to override the defaults, but this information is
not always accessed in time to ensure a correct
interpretation.
4
Motivation for current study

Linear order default constructions appear
to be important for comprehension of
clauses.

Do similar defaults play a role in noun
phrase comprehension?

What consequences does non-canonical
ordering of the head noun have for
sentence comprehension?
5
Free Relatives in English

Whoever said that is a fool.
whoever = ‘the person who’ or ‘anyone who’

Whoever’s idea that was is a fool.
whoever’s = ‘the person whose’ or ‘anyone whose’
6
Possessive free relatives: some examples

“There were rose petals scattered across the floor and
some had writing on them. One said, ‘I'll love you forever’,
and another said, ‘Be mine till the end of time.’ How sweet,
whoever's boyfriend did this is a lucky girl.”


“You're willing to put up with all the bull****, mental
abuse, pain just to ensure that whoever's children you are
going to bear is strong, so your children are strong. It's
worth all that pain, isn't it?”


Story on Quizilla.com, 2-20-2007
Comment on Übersite.com, 12-04-2004
“…as far as the kids on stage behind Roger, I agree with
Basje on this one too - that's pretty unprofessional...I am
pretty sure whoever's kids those were could afford a nanny
or sitter for that night.”

Comment on Queenzone.com, 3-31-2005
7
“I bet whoever's car that is is having a worse day
than you.” Comment on Stereokiller.com, 4-16-2007
8
Canonical templates for NP

Canonical templates for NP order determiner or possessor
before the head noun (bearer of referential index):
Linear Order: Det
Semantic Role: Specifier
The
Linear Order: Possessor
Semantic Role: Specifier
John’s
N
Head
dogs
S
Modifier
that got loose
are in trouble.
N
Head
dogs
S
Modifier
that got loose
are in trouble.
9
Possessive free relatives

Ordinary possessive relative clauses are syntactically and
semantically complex, but still conform to the canonical
template for NP:
Linear Order:
Semantic Role:

Det
Specifier
The
N
Head
guy
S
Modifier
whose dogs got loose
is in trouble.
Possessive free relatives look similar, but violate default
ordering. Interpretation of referential index for NP depends on
possessive pronoun whoever’s:
Linear Order:
Expected Semantic Role:
Actual Semantic Role:
Possessor
Specifier
Head
Whoever’s
N
S
Head
Modifier
Modifier
dogs
got loose
is in trouble.
10
Experiment 1: Verb Decision Task

Prediction: Possessive free relatives should be more
frequently misunderstood and more slowly processed than
similar phrases that conform to the ordering defaults for NP.

Visual presentation of sentences in which main verb is
missing:
Whoever’s dogs got loose __ in trouble.

Participants must press a button choosing “is” or “are” to
complete the sentence. Number specification on the
relevant nouns is counterbalanced.

Accurate response requires identification of the head noun
in the subject of the matrix clause.
11
Design and Stimuli

2x2 repeated measures design
 Possessive vs. non-possessive
 Free vs. normal

4 relative clause types (10 sets)
1.
2.
3.
4.


The guy whose dogs got loose is in trouble.
Whoever’s dogs got loose is in trouble.
The dogs that got loose are in trouble.
Whichever dogs got loose are in trouble.
(normal possessive)
(free possessive)
(normal non-possess)
(free non-possessive)
5 balanced blocks of 40 sentences each (20 test sentences, 20
fillers in each block)
Random ordering of sentences within each block, random
ordering of blocks.
12
Participants

42 Purdue University students




Ages 18-51 (average age 23)
Native speakers of American English
16 male, 26 female
Participants were compensated with a
choice of either $3 or extra credit from
certain instructors, for a 15-20 minute
session.
13
Percent Correct for Verb Decision Task
100%
Percent Correct
95%
90%
85%
80%
75%
70%
65%
60%
normal
possessive
free
possessive
normal
non-possess.
free
non-possess.
Relative Clause Type
Main effects:
Interaction:
Possessive: F(1, 41) = 33.27, p < 0.01
Free: F(1, 41) = 60.29, p < 0.01
Possessive x Free: F(1,41)=26.40, p < 0.01
14
Response Time (ms)
Response Time for Verb Decision Task
4400
4200
4000
3800
3600
3400
3200
3000
2800
2600
2400
2200
2000
normal
possessive
free
possessive
normal
non-possess.
free
non-possess.
Relative Clause Type
Main effects:
Possessive: F(1, 41) = 105.83, p < 0.01
Free: F(1, 41) = 44.95, p < 0.01
Interaction:
Possessive x Free: F(1, 41) = 0.01, p = 0.91 (not significant)
15
Experiment 1 summary




Participants were least accurate and had the
slowest response times with possessive free
relatives.
Results for accuracy appear to confirm initial
hypothesis that non-canonical structure
contributes to more frequent miscomprehension.
Results for response time indicate that
possessives are processed more slowly than nonpossessives, and that free relatives are processed
more slowly than normal relatives.
However free relatives are slower than normal
relatives to about the same degree, regardless of
whether they are possessive or not.
16
Experiment 2: True-False Decision Task


New task to avoid any extraneous factors related
to English subject-verb agreement.
Visual presentation of sentences (with singular
number on all nouns) followed by either a true or
false statement:




Sentence: Whoever’s dog got loose is in trouble.
Statement: Some dog is in trouble. (True or False?)
Participants must press a button choosing “true”
or “false” to indicate the truth of the statement in
relation to the original sentence.
Accurate response requires identification of the
head noun in the subject of the matrix clause.
17
Design and Stimuli

2x2 repeated measures design



4 relative clause types (10 sets)
1.
2.
3.
4.


Possessive vs. non-possessive
Free vs. non-free
The guy whose dog got loose is in trouble.
Whoever’s dog got loose is in trouble.
The dog that got loose is in trouble.
Whichever dog got loose is in trouble.
(normal possessive)
(free possessive)
(normal non-possessive)
(free non-possessive)
4 balanced blocks of 30 sentences each (10 test sentences, 20 fillers
in each block)
Random ordering of sentences within each block, random ordering of
blocks.
18
Participants

21 Purdue University students




Ages 18-23 (average age 20)
Native speakers of American English
5 male, 20 female
Participants were compensated with a
choice of either $6 or extra credit from
certain instructors, for a 35-40 minute
session.
19
Percent Correct
Percent Correct for True-False Task
100%
95%
90%
85%
80%
75%
70%
65%
60%
normal
possessive
free
possessive
normal
non-poss.
free
non-poss.
Relative Clause Type
Main effects:
Possessive: F(1, 20) = 53.10, p < 0.01
Free: F(1, 20) = 21.85, p < 0.01
Interaction:
Possessive x Free: F(1,20) = 53.57, p < 0.01
20
Response Time (ms)
Response Time for True-False Task
2500
2400
2300
2200
2100
2000
1900
1800
1700
1600
1500
normal
possessive
free
possessive
normal
non-poss.
free
non-poss.
Relative Clause Type
Main effects:
Possessive: F(1, 20) = 14.93, p < 0.01
Free: F(1, 20) = 0.74, p = 0.39 (not significant)
Interaction:
Possessive x Free: F(1,20) = 0.55, p = 0.47 (not significant)
21
Experiment 2 summary




Participants were least accurate with possessive free
relatives.
Results for accuracy lend additional support for the
hypothesis that non-canonical structure contributes to more
frequent miscomprehension.
Results for response time indicate that possessive relatives
are processed more slowly than non-possessive relatives.
Although possessive free relatives were again the slowest
overall, the difference between free relatives and normal
relatives was not significant.
22
Conclusions and Implications




Two experiments on possessive free relative clauses
suggest that a simple default construction for NP may play
a role in comprehension of both canonical and noncanonical NPs.
Violation of the default appears to affect basic
understanding of NP meaning, as shown in accuracy
results.
The use of default constructions in processing may help
constrain the occurrence of non-canonical construction
types by making certain kinds of linear order mismatches
especially costly for language users. Thus, we expect to find
explicit linguistic cues in constructions that violate the
default ordering.
One reason possessive free relatives are harder than, for
example, passives, might be that whoever’s is the only
overt cue to the ordering mismatch.
23
Defaults as constructional biases





This analysis is compatible with usage-based models of
grammar and processing. Default constructions can be
understood as constructional biases of clauses or phrases,
akin to lexical biases of verbs.
Constructional biases are based on frequency of certain
linear order-semantic role mappings.
Constructional biases based on linear order defaults are
distinct from the frequencies of particular lexical items or
particular constructions. E.g., infrequent constructions can
still conform to defaults.
Infrequent constructions that conform to the relevant
linear order default are predicted to be easier to
understand than equally infrequent constructions that
violate the default, all else being equal.
Similarly, other kinds of mismatches that affect syntactic
category or hierarchical structure in syntax (but not linear
order) are predicted to incur fewer comprehension errors.
24
Thanks to

Research assistant: Yanhong Zhang

Funding: Purdue University

Experiment participants at Purdue

Abstract reviewers and other colleagues
who provided comments
25
Selected References







Ferreira, Fernanda. 2002. Good-enough representations in
language comprehension. Current Directions in Psychological
Science 11: 11-15.
Ferreira, Fernanda. 2003. The misinterpretation of noncanonical
sentences. Cognitive Psychology 47: 164-203.
Gahl, Susanne, Lise Menn, Gail Ramsberger, Daniel S. Jurafsky,
Elizabeth Elder, Molly Rewega, L. Holland Audrey. 2003. Syntactic
frame and verb bias in aphasia: Plausibility judgments of
undergoer-subject sentences. Brain and Cognition 53: 223-228.
Goldberg, Adele E. and Giulia M. L. Bencini . 2005. Support from
language processing for a constructional approach to grammar.
Language in Use: Cognitive and Discourse Perspectives on
Language and Language Learning, ed. by Andrea Tyler, 3-18.
Georgetown University Round Table on Languages and Linguistics.
Grodzinsky, Yosef. 1995. Trace deletion, theta roles and cognitive
strategies. Brain and Language 51: 469-497.
Grosu, Alexander. 2002. Strange relatives at the interface of two
millennia. Glot International 6: 145-167.
Jackendoff, Ray S. 2007. A parallel-architecture perspective on
language processing. Brain Research 1146: 2-22.
26
References, continued







Love, Tracy E. and David A. Swinney. The influence of canonical
word order on structural processing. Sentence Processing: A
Crosslinguistic Perspective. Syntax and Semantics 31: 153-166.
Müller, Stefan. 1999. An HPSG-analysis for free relative clauses in
German. Grammars 2: 53-105.
Piñango, Maria. 2000. Canonicity in Broca’s sentence
comprehension: the case of psychological verbs. Language and
the Brain, ed. by Yosef Grodzindky, Lewis P. Shapiro, and David
Swinney, 327-350. San Diego: Academic Press.
Sag, Ivan A. 1997. English relative clause constructions. Journal
of Linguistics 33: 431-483.
Sag, Ivan A. and Thomas Wasow. to appear. Performancecompatible competence grammar. Non-Transformational Syntax,
ed. by Kersti Börjars and Robert D. Borsley. Oxford: Blackwell.
Townsend, David J. and Thomas G. Bever. 2001. Sentence
Comprehension: The Integration of Habits and Rules. Cambridge,
MA: MIT Press.
Wright, Abby & Andreas Kathol. 2003. When a head is not a head:
a constructional approach to exocentricity in English. The
Proceedings of the 9th International Conference on HPSG, ed. By
John Bok Kim and Stephen Wechsler, 373–389. Stanford: CSLI 27
Publications.