Structure Building and Unificationling.umd.edu/~colin/research/papers/phillips_cunygc2001.pdf ·...
Transcript of Structure Building and Unificationling.umd.edu/~colin/research/papers/phillips_cunygc2001.pdf ·...
Colin Phillips, Supper @ CU
NY
, 3/27/01
1
Structure Building and Unification
Colin Phillips, University of M
arylandcolin@
glue.umd.edu
The Standard View
• There are m
ultiple (related) structure-building systems
Gram
mar
A recursive characterization of the structural descriptions of all possible expressions of a
language.
When w
e speak of a gramm
ar generating a sentence with a certain structural description, w
em
ean simply that the gram
mar assigns this structural description to the sentence. W
hen we say
that a sentence has a certain derivation with respect to a particular generative gram
mar, w
e saynothing about how
the speaker or hearer might proceed, in som
e practical or efficient manner,
to construct such a derivation. These questions belong to the theory of language use – to thetheory of perform
ance.” (Chomsky 1965, p.9)
ParserA
system for increm
ental recovery of structural descriptions from spoken/w
ritten input.
‘…parsing routines exist m
erely for efficient implem
entation of the competence gram
mar…
’(Fodor, 1998: 290)
AN
Dperhaps a structural notion of sim
plicity, for use in resolving ambiguity
•
Comm
only assumed that the parser im
plements the gram
mar increm
entally in thesam
e way in all languages, i.e. w
hat the parser adds to the gramm
ar is Universal
ProducerA
system for (possibly) increm
ental structuring of sentences for speaking.
Competence/Perform
ance
1. Behavior, e-language‡
2. Dedicated system
s for parsing, production, etc.3. Capacity of gram
mar w
ith bounded vs. unbounded resources4. Form
al distinction:what a function computes vs. how it does so
Questions about U
nification?
• H
ow do perform
ance-systems recruit the inform
ation in the gramm
ar?•
How
are gramm
aticality judgments m
ade? If the parser is a ‘covering gramm
ar’ forthe com
petence gramm
ar, is it a perfectly accurate implem
entation of the gramm
ar?•
H
ow does learning occur, given that the gram
mar is ‘hidden’ behind the parser?
(error signals in learning are only indirectly related to the changes that they must
trigger)
Alternative A
pproach
(2)A
speaker’s gramm
atical knowledge takes the form
of an incremental
generative system, w
hich subsumes production, com
prehension andgram
maticality judgm
ent.
Gram
mar
is an
algorithm
which
generates sentence
structures/interpretations in an incremental fashion.
Production: incremental generation of sentence structures.
Gram
maticality: increm
ental generation of structure for word string; if
generation either blocks completely or m
akes an illegal step, returnsungram
maticality. [often w
ith idealization of infinite resources]Com
prehension: incremental generation of structures, term
inal string must
match input (analysis-by-synthesis com
prehension system).
•
We know
that an incremental structure-building system
exists. Hypothesis is that
that is the only structure-building system.
Unification Q
uestions Again
•
Relation between gram
matical know
ledge and the gramm
atical component of
performance system
s becomes straightforw
ard, since they are identical•
Processes underlying gramm
aticality judgment becom
e more transparent
• In learning, error signals m
ore closely related to resulting changes
Evidence
•
Is this the right way to think about linguistic know
ledge? What evidence can be
brought to bear?
…the available evidence varies across different dom
ains…
a. ‘Proof of Concept’ evidence: can the approach at least be show
n to be feasible?‡
plenty to be done in this regard
b. Syntactic Theory: substantial change in shape of structural derivations‡
scope for a variety of empirical argum
ents
c. G
ramm
aticality Judgment Processes: largely uncharted dom
ain‡
arguments involve ‘proof of concept’,
‡ possible electrophysiological corroboration
d. Parsing &
Production: system doesn’t look so different …
but it has less roomfor error or naiveté‡
evidence is more am
biguous
e. Learning: alm
ost no explicit theories of change in language acquisition‡
arguments should again involve ‘proof of concept’ dem
onstrations
Colin Phillips, Supper @ CU
NY
, 3/27/01
2
a. Proof of Concept
Syntactic Theory can the gram
mar work like this?
Computational im
plementation
really, no cheating!Behavioral Experim
entationis this what people do?
Cognitive Neuroscience
b. Syntactic Theory
Some consequences of left-to-right derivations
Classic form of argum
ents for derivations: surface form show
s that rule X has applied,
although the environment for rule X
is not present in the surface form. Solution: assum
ethat the environm
ent for rule X w
as present at an earlier derivational step, but was later
destroyed by another rule.
How
syntactic constituency tests are supposed to work:
(1)a.
Grom
it [likes cheese] and [hates cats](coordination)
b.G
romit [likes cheese] and W
allace does too(deletion/ellipsis)
c.[Like cheese] though G
romit does ___, he can’t stand Brie.
(movem
ent)d.
Wallace and G
romit like each other.
(reciprocal binding)e. *
Each other like Wallace and G
romit.
(illicit reciprocal binding)
Grom
itlikescheese
NP
S
VP
VN
P
A.
Different constituency tests pick out different sets of strings as constituents.
Although this does not contradict standard assum
ptions about constituency, it doesnot follow
from them
.
Coordination is a very ‘liberal’ diagnostic
(2)a.
Wallace [visited W
endolene] and [bought some w
ool].b.
Wallace gave [G
romit a biscuit] and [Shaw
n some cheese] for breakfast.
c.[W
allace designed] and [Grom
it built] an enormous tin m
oon-rocket.d.
Alice [knew
that Fred wanted to w
rite to] and [hoped that he wanted to
talk to] the president.
Some strings that can be coordinated cannot undergo m
ovement:
(3)a. *
[Grom
it a biscuit] Wallace gave ___ for breakfast.
b. *W
allace gave ___ for breakfast [his favorite pet beagle an enormous
chewy dog-biscuit.]
Some strings that can be coordinated cannot undergo ellipsis:
(4)a.
Alice [knew
that Fred wanted to talk] and [hoped that he w
anted to argue]w
ith the president.
b. *A
lice [knew that Fred w
anted to talk] with the queen and Ethel did w
iththe president.
B.In som
e instances, constituency tests pick out overlapping strings as constituents,w
hich is at odds with standard assum
ptions about constituency.
Coordination of overlapping strings
(5)a.
Wallace gave [G
romit a biscuit] and [Shaw
n some cheese] for breakfast.
b.W
allace gave Grom
it [a biscuit in the morning] and [som
e cheese justbefore bedtim
e.]
VP-fronting in conflict w
ith anaphor-binding possibilities (Pesetsky 1995)
(6)a.
...and [give the books to them in the garden] he did ___ on each other’s
birthdays.b.
...and [give the books to them] he did ___ in the garden on each other’s
birthdays.
C.A
lthough there are excellent characterizations of the range of constituents identifiedby individual structural diagnostics, there is no general theory of w
hy particularstructural tests pick out the strings that they do as constituents, and not others.
Proposal
•
Conflicts between syntactic constituency tests are ‘opacity’ effects, w
hich arisebecause different tests apply at different points in a derivation.
• D
ifferences between tests can be understood if w
e assume that structures are built up
incrementally, from
left-to-right.
Incrementality hypothesis
Sentence structures are built incrementally from
left-to-right, i.e. in the order in which
terminal elem
ents are pronounced.
IP
I'N
P
VP
infl
IP
I'N
P
VP
inflVN
P
IP
I'N
P
VP
inflV
NP
VP
V'
VPP
Wallac
esaw
Grom
it
in the kitchen
Wallac
e
saw
Wallac
e
sawGrom
it
(7a)(7b)
(7c)
Consequences:i.
Since constituency changes systematically as new
material is added to structure, and
different structural diagnostics apply at different points in a derivation, certainsystem
atic differences between constituency tests can be explained.
Colin Phillips, Supper @ CU
NY
, 3/27/01
3
ii.M
akes it possible to maintain the assum
ption that syntactic operations manipulate
constituents, and that sentences have a single derivation.iii.
Since all tests apply to same structural derivation, w
e can explain why conflicts
between constituency tests are m
ore restricted than is predicted by flexibleconstituency theories.
iv.Com
patible with increm
entality of comprehension and production.
A. C
oordinationPredicted to be extrem
ely liberal, because coordinated strings are imm
ediately adjacent;hence no risk of first conjunct losing its constituency before second conjunct built.Should, in principle, be possible for all constituents of a derivation.
(8)W
allace will give G
romit crackers for breakfast.
a.[W
allace]W
allace and Wendolene gave G
romit crackers for breakfast.
b.[W
allace will]
Wallace w
ill and Wendolene probably w
on’t give Grom
it crackers forbreakfast.
c.[W
allace [will give]]
Wallace w
ill give and Wendolene w
ill send their favorite beagle some
crackers for his birthday.W
allace will design but w
on’t complete an exciting new
invention for hisdog’s birthday.
d.[W
allace [will [give G
romit]]]
Wallace w
ill give Grom
it and Wendolene w
ill give Preston a shining newcollar for w
alking about town.
Wallace w
ill give Grom
it and send Preston a shining new collar for
walking about tow
n.
e.[W
allace [will [give [G
romit crackers]]]]
Wallace w
ill give Grom
it crackers and Wendolene w
ill give Preston dogfood for breakfast.W
allace will give G
romit crackers and Preston dog food for breakfast.
f.[W
allace [will [give [G
romit [crackers for breakfast]]]]]
Wallace w
ill give Grom
it crackers for breakfast and Wendolene w
ill givePreston dog food for dinner.W
allace will give G
romit crackers for breakfast and toast for lunch.
...of course, some strings are never constituents:
(9)a. *
The man [w
ho built the rocket has] and [who studied robots designed] a
dog.b. *
Wallace gave his [dog half a dozen] and [sheep a handful of] crackers for
breakfast.c. *
After W
allace fed [his dog the postman] and [his sheep the m
ilkman]
arrived.
B. Coordination vs. Pseudogapping
Predicted to pick out a narrower range of constituents: the string that licenses ellipsis
must be a constituent both w
hen it is first built and when the ellipsis site is built &
licensed, i.e. constituent must be present beyond com
pletion of sentence which is
antecedent of ellipsis.
• V
erb–preposition strings may be coordinated:
(10)a.
John talked to and gossiped about the kid who sprayed paint on his car.
b.The cat looked at and then slept on the rug in the m
iddle of the livingroom
.•
But they cannot be deleted in ellipsis constructions, a.k.a. ‘pseudogapping’(Postal1986, Baltin &
Postal 1996).
(11)a.
Helen talked to Jonathan, and A
lice did ___ *(to) Matthew
.b.
The cat slept on the mat, and the dog did ___ *(on) the chair.
(12)a.
Helen talked to Jonathan m
ore often than Alice did ___ *(to) M
atthew.
b.The cat slept on the m
at more often than the dog did ___ *(on) the chair.
•Reason: pseudogapping requires licensing of ellipsis site after V
+P constituent hasbeen destroyed by addition of N
P.
(13)a.
b.c.
Helen
talkedto
S
VP
NP
P(P)V
Helen
talked
to
S
VP
NP
VPP
PN
PJonathan
Helen
talked
to
S
VP
NP
VPP
PN
PJonathan
Helen
S
VP
NP
did
Sand
C. Binding vs. Predicate Fronting
Prediction: conflicts between different tests can arise w
hen tests can apply at differentstages in the increm
ental derivation of a sentence.
• O
verlapping strings may be coordinated, because they are constituents at different
points in an incremental derivation.
(14)a.
Wallace gave [G
romit a biscuit] and [Shaw
n some cheese] for breakfast.
b.W
allace gave Grom
it [a biscuit in the morning] and [som
e cheese justbefore bedtim
e.]
(15)a.
[give]b.
[give Grom
it]c.
[give [Grom
it [a biscuit]]]d.
[give [Grom
it [[a biscuit] [for breakfast]]]]
• Conflict betw
een movem
ent/ellipsis and binding results is possible, because bindingrelations established after m
ovement/ellipsis relations in increm
ental derivation.
Colin Phillips, Supper @ CU
NY
, 3/27/01
4
(16)a.
to
giveV
PV
NP
V
V'
PP
PN
P
VP
give
the book
them
IP
NP
IP
Ihe
did
b.
to
giveV
PV
NP
V
V'
PP
PN
P
VP
give
the book
them
IP
NP
IP
I
he
did
I'
to
giveVP
V
NP
V
V'
PP
PNP
VP
give
the book
them
to
giveV
PV
NP
V
V'
PP
PN
P
VP
give
the book
them
IP
NP
IP
I
he
did
I'
to
giveVP
V
NP
V
V'
PP
P
NP
VP
give
the book
them
oneach other's birthdays
PN
P
PP
P'
C.
D. M
ovement vs. Ellipsis
Prediction: constituents become unavailable to syntactic processes once they have been
destroyed.
• Both V
P-fronting (VPF) and V
P-ellipsis (VPE) can m
ove/delete strings beginning atthe left-edge of V
P.
(17)a.
... and [give candy to children in libraries on weekends] he did.
b.... and [give candy to children in libraries] he did on w
eekends.
c.... and [give candy to children] he did in libraries on w
eekends.d.
... and [give candy] he did to children in libraries on weekends.
e. *...and [to children in libraries] he did give candy on w
eekends.f. *
...and [in libraries on weekends] he did give candy to children.
(18)a.
John gives candy to children in libraries on weekends, and M
ary does(too).
b.John gives candy to children in libraries on w
eekends and Mary does on
federal holidays.c.
John gives candy to children in libraries on weekends and M
ary does inurban parks on federal holidays.
• Internal to the fronted/elided portion of V
P, both VPF and V
PE pass tests for right-branching V
P-structure (i.e. tests from Barss &
Lasnik 1986, Larson 1988 etc.).
(19)a.
... and [introduce the children to each other] the teacher proceeded to do.b.
... and [congratulate everybody on his birthday] he did.
(20)a.
The principal introduced the children to each other, and then the teacherdid (too).
b.The boss congratulated everybody on his birthday, and the receptionistdid (too).
• W
hen only part of the VP is fronted, V
PF still passes diagnostics of right-branchingV
P-structure. Conflict betw
een VP-fronting and binding is possible, because
movem
ent chain is complete before binding chain is constructed.
(21)a.
John said he would give books to them
,... and give books to them
he did [on each other’s birthdays].(reciprocal binding)
b.M
ary said she would congratulate every boy,
... and congratulate every boy she did [at his graduation].(variable binding)c.
John said he would read every book,
... and read every book he did [at breakneck speed].(collective and distributive scope readings)
• In contrast, w
hen only part of VP undergoes ellipsis, diagnostics of right-branching
structure fail:
(22)a. *
John gave books to them on each other’s birthdays,
and Mary did [on each other’s first day of school].
(reciprocal binding)b. *
Mary congratulated every boy at his graduation,
and Sue did [at his 21st birthday party].(variable binding)
(23)M
ary read all the books quickly.(collective and distributive scope readings)
Collective reading: the total time required to read the books w
as fastD
istributive reading: the time required to read each individual book w
asfast
(24) M
ary read all the books quickly, and John did slowly.(collective reading only)
Colin Phillips, Supper @ CU
NY
, 3/27/01
5
• Explanation: ellipsis m
ust be licensed in second conjunct after adverbial has beenadded in first conjunct. Creation of right-branching structure in first conjunctdestroys the V
+NP constituent, w
hich therefore can no longer license ellipsis.
(25)a.
b.
all the books
quicklyA
dv
V'
VP
V
NP
read
VP
Vall the books
quicklyA
dvV
'
VP
VN
Pread
Similar effect in Japanese
(26)a.
John-ga isoide dono hon-mo yonda.
-nom quickly all books-acc read
‘John read all the books quickly.’(collective &
distributivereadings available)
b.John-ga dono hon-m
o isoide yonda. -nom
all books-acc quickly read‘John read all the books quickly.’
(distributive reading only)
(27)a.
John-wa isoide dono hon-m
o yonda, (sosite) Mary-w
a yukkuri soosita.John-top quickly all books-acc read, (and) M
ary-top slowly did-so
‘John read all the books quickly, and Mary did slow
ly.’(collective reading only)
b. *John-w
a dono hon-mo isoide yonda, (sosite) M
ary-wa yukkuri soosita.
John-top all books-acc quickly read, (and) Mary-top slow
ly did-so‘John read all the books quickly, and M
ary did slowly.’
(both scopes impossible)
•
Control: Comparative Ellipsis (H
eim 1985, D
iesing 1992) allows m
ore minim
alcontrast than V
PF/VPE.
Adverbial phrase m
ay be present in either or bothconjuncts.
(28)a.
John read as many books as Bill did on Thursday.
b.John read as m
any books on Tuesday as Bill did on Thursday.
Prediction: diagnostics of right-branching VPs w
ill succeed when non-elided adverbial is
present in just one conjunct (29a-b), and fail when it is present in both conjuncts (29c).
(29)a.
John read as many books as Bill did in a w
eek.(collective and distributive readings both ok)
b.John read as m
any books in a week as Bill did.
(collective and distributive readings both ok)c.
John read as many books in a w
eek as Bill did in a month.
(collective reading ok, distributive reading impossible)
(30)a. (?)
The provost met as m
any studentsi as the dean did when theyij w
ere firstentering the university.
b. *The provost m
et as many studentsi w
hen theyi were first entering the
university as the dean did when theyij w
ere graduating.
(31)a.John
baked
NP
V
VPN
P
S
Mary
VPN
P
S
S
VPV
V'
PPin an hour
baked
did
as many cakes
as
b.
John
bakedN
PV
VPN
P
S
Mary
VPN
P
S
S
V'PP
in an hour
as many cakes
as
did
Note: the basic loss-of-scope-in-ellipsis facts are not the evidence for increm
entalstructure building: rather it is the contrast betw
een the ellipsis and fronting facts, and theparallel contrast in com
parative ellipsis examples.
E. Multiple D
ependencies (Richards, 1999)
(32)a. *W
hich book did the senator deny [NP the rum
or that he wanted to ban __]
b.W
hich senator __ denied [NP the rum
or that he wanted to ban W
ar &Peace]
c.W
hich senator __ denied [NP the rum
or that he wanted to ban w
hich book]
•
What is surprising about this paradigm
is the fact that the short wh-dependency
appears to license the longer wh-depenency that crosses an island. (The sam
e factshold in Japanese (no w
h-movem
ent) and Bulgarian (multiple w
h-movem
ent)
(33)a.
Which senator __ said that [N
P the rumor that he w
anted to ban which
book] had been spread by Comm
unists.b. *
Which book did the senator say that [N
P the rumor that he w
anted to ban__] had been spread by w
hom.
•
Interpretation: the well-form
ed dependency must precede the island-violating
dependency that it licenses.
(34)a.
Who __ w
onders what w
ho bought __b. *
John wonders w
hat who bought __
Colin Phillips, Supper @ CU
NY
, 3/27/01
6
•
What is surprising about (a) is that the the em
bedded what can cross over the
embedded who.. This is im
possible in (b).
• Richards: this is only possible w
hen there is a well-form
ed matrix w
h-dependency ina higher clause. This im
plies that the construction of the matrix CP precedes the
construction of the embedded CP, as is expected in a left-to-right / top-dow
nderivation.
• A
dditional arguments based on expletive-associate dependencies (e.g. There seem
sto be a m
an in the room etc.)
• K
empson, M
eyer-Viol &
Gabbay (2001): argue that left-to-right derivations offer an
improved treatm
ent of various kinds of anaphoric dependencies, includingrelativization, w
eak crossover.•
O
ther work on left-to-right gram
matical derivations: Richards (1999) on w
h-m
ovement; Boeckx &
Stjepanovic (1999) on wh-m
ovement and clitics; Boeckx
(1999) on island constraints and A-m
ovement; G
uimaraes (1999) on prosody-syntax
interface; Drury (1999) on island constraints; Legate (1999), Schlenker (1998) on
morphological processes
c. Gram
maticality Judgm
ent Processes
still largely unexplored
d. Parsing
• M
ain concern is parsing of unambiguous sequences
Quite varied assum
ptions made in this area:
• IF: Parser is a perfect covering gram
mar
‡ it’s hard to find parsing evidence to distinguish this from
the single systemapproach
• IF: Parser is a poor covering gram
mar, as evidenced by the errors that it m
akes(older version: Fodor, Bever &
Garrett, 1974; nerw
er versions: Townsend &
Bever,2001; Ferreira, Bailey &
Ferraro, 2001)‡
evidence of this kind is more troubling
(i) Increm
ental Analysis of Parasitic G
aps (Phillips & W
ong, 2000)
• Increm
ental structure-building creates look-ahead problems of its ow
n:
(35) *What did the plan to rem
ove ___ ultimately dam
age the building?(36)
What did the plan to rem
ove the building ultimately dam
age ___?
(37)W
hat did the plan to remove ___ ultim
ately damage ___?
(38) *What did the plan that rem
oved ___ ultimately dam
age ___?
•
Problem: parasitic gaps in com
plex subject NPs pose a look-ahead problem
–form
ing a wh-dependency involves crossing a syntactic island (C
ondition onExtraction D
omains): H
uang 1982), but the island violation may be rescued at a
later point in the sentence.•
Fully increm
ental approach to parsing parasitic gaps (cf. Schneider 1999): formdependencies across islands that m
ay involve a parasitic gap, but do not mark the
wh-dependency as com
pleted.
(39)a.
b.
Meanw
hile…
Controversy in parsing literature over whether gram
matical search respects island
constraints.
(40)Y
es!
a. Stowe 1986
[self-paced reading]PP-in-subject
b. Traxler & Pickering 1996
[eye-tracking]RC-in-subject
c. Kluender &
Kutas 1993ab
[ERP violation paradigm]
wh-island
d. McElree &
Griffith 1998
[speeded gramm
aticality]RC-in-object
e. Bourdages 1992[self-paced reading]
RC-in-objectf. Berw
ick & W
einberg 1984 [com
putational arguments]
f. Phillips & W
ong 2001[self-paced reading]
RC-in-subject
Colin Phillips, Supper @ CU
NY
, 3/27/01
7
(41)N
o!
a. Freedman &
Forster 1985[sentence-m
atching]N
P w/ possessor
b. Neville et al. 1991
[ERP violation paradigm]
NP w
/ possessorc. Stevenson 1993
[comprehension, m
atching]com
plex NP
d. Kurtzm
an et al. 1990[speeded-gram
maticality]
infin.-in-subjecte. Phillips &
Wong 2001
[self-paced reading]infin.-in-subject
Studies with m
arginal results…f. Pickering et al. 1994
[self-paced, eye-tracking]RC-in-subject
g. Clifton & Frazier 1989
[end-of-sentence rating]RC-in-object
•
The kinds of island contexts examined in the different studies appear to be fairly
good predictors of whether or not the island w
as imm
ediately respected by theparser.
Suggested explanation (Phillips 1995):
(i) parser posits gaps inside islands w
hich involve extraction from just one offending
category (e.g. subject, adjunct, specific NP)
(ii) this is the kind of environm
ent that supports parasitic gaps!
• Prelim
inary: are parasitic gaps acceptable to anybody other than linguists? Yes!
GRA
MM
ATICA
LITY JU
DG
MEN
T TA
SK
Participants: 50 undergraduatesM
aterials: Sets of 84 sentences rated on 1 (bad) – 5 (good) scale. Includes 24 sets ofsentences constructed from
materials used in P-gap on-line experim
ent; 60 othersentences show
ing variety of wh-extractions (gram
matical &
ungramm
atical), includingsentences based on structures tested in earlier experim
ents on islands.
Figure 1: Mean G
ramm
aticality Ratings for PG-related sentences
RESU
LTS: sentences with both a parasitic gap and an object gap w
ere rated almost exactly
the same as sentence containing only the object gap (P-gap filled by an overt N
P), when
the P–gap was inside an infinitival clause.
P–gap without O
-gap rated much w
orse; combination of P-gap and O
-gap rated lower
when P–gap is in finite relative clause.
SELF-P
ACED
REA
DIN
G S
TUD
Y
• Q
uestion: will subjects show
evidence of actively creating a wh-dependency inside
an island in exactly the environments w
here P-gaps will be licensed?
(42)--- ---- -- ----- ----- ------ -- --- -----The ---- -- ----- ----- ------ -- --- -------- rain -- ----- ----- ------ -- --- -------- ---- in ----- ----- ------ -- --- -------- ---- -- Spain ----- ------ -- --- -----
Participants: 56 U. of D
elaware students (10 excluded for low
comprehension scores)
Materials: 2 x 2 design (±finite x ±gap) 24 sets of 4 conditions, 96 filler sentences;
comprehension question after all trials
Note: in the experim
ent subjects never saw a parasitic gap – they m
erely sawenvironm
ents where P-gaps could potentially occur.
(43)a. The outspoken environm
entalist worked to investigate w
hat the local campaign to
preserve the important habitats had actually harm
ed in the area that the birds onceused as a place for resting w
hile flying south. [infinitive, gap]
b. …w
hether the local campaign to preserve…
[infinitive, no gap]c. …
what the local cam
paign that preserved…[finite, gap]
d. …w
hether the local campaign that preserved …
[finite, no gap]
Restricted set of nouns as heads of complex subject N
P: plan, scheme, request, cam
paign, attempt,
bid, endeavor, lawsuit, temptation, power, evidence, efforts, idea
Critical verb imm
ediately preceding parasitic gap site was alw
ays obligatorily transitive: cheapen,em
barrass, remove, nom
inate, alter, order, aid, preserve, entertain, heal, secure, destroy,rehabilitate, reclaim
, liberate, remove, upset, im
press, uplift, convict, instruct, astonish, develop,overprice.
RESU
LTSCom
prehension accuracy: 87.8% on com
prehension questions (~33 word sentences!)
Analyses exclude trials on w
hich comprehension question answ
ered incorrectly; outlierresidual reading tim
es trimm
ed to 1000ms.
Figure 2: Residual Reading Times for PG
Self-paced Reading Experiment
to investigate what the local cam
paign to preserve the important habitats
5 6 7 8 9 10 11 12 13 14 15
Colin Phillips, Supper @ CU
NY
, 3/27/01
8
At verb inside com
plex subject (region 12):
Infinitive conditions: significant effect of gap/no-gap: (F1 (1,45) = 4.02, p < 0.05;
F2 (1,35) = 5.01, p < 0.05)
Finite conditions: no difference between gap/no-gap conditions.
• Slow
down in reading tim
es at verb reflects wh-dependency formation, but this only
occurs in the infinitive conditions, where a P-gap is possible.
•
Wh-dependencies are form
ed across islands – but in exactly the ways that the
gramm
ar allows.
• C
ON
CLUSIO
N: Structural analysis is fully incremental, and quite sophisticated!
•
Caveat: this experiment found slow
down due to w
h-dependency formation at the
verb where the wh-dependency is created, rather than w
hen it needs to be retractedat the follow
ing NP. A
lthough this is not a ‘filled-gap’ effect per se, it is consistentw
ith other results in the literature (e.g. Pickering et al. 1994).
(ii) M
onotonic Structure Building (Schneider & Phillips, 2001)
QU
ESTION
S: (i) how
is the gramm
ar searched for gramm
atical parses;(ii) is reanalysis a last resort? (Fodor &
Frazier 1980; cf. Frazier 1990,Frazier &
Clifton 1998) or are reanalyses part of the parser’s initial searchspace (G
ibson 1991; Stevenson 1998; Fodor & Inoue 1998)?
…In other w
ords, does the parser attempt to keep to the sam
e derivation?
• Evidence: contrast betw
een strong evidence for locality and strong evidence againstlocality in parsing.
• M
any examples of locality biases in structural am
biguity resolution (Kim
ball 1973,Fodor &
Frazier 1980, Gibson 1991, 1998; m
any others)
(44)a.
John said Bill left yesterday.b.
During his vacations, W
allace likes to swim
very much.
c.John saw
the daughter of the lawyer w
ho he had met at the party in the
mall.
(45)Locality > Structural Sim
plicity (Phillips & G
ibson, 1997ab)
Although Erica hated the house her father had ow
ned ...
IP
V IPCP
VP
NP
IP
NP
I'
Infl
C
NP
IP
CP
C'
NPOp
C
NP
matrix clause
attachmentrelative clauseattachm
ent
AlthoughErica
hated
she
she
the house
(46)W
allace knows G
romit likes cheese.
NP
Wallace
S
Grom
it
VP
knows
likes
• Parsing the em
bedded verb likes requires reanalysis of the preceding NP.
•
This reanalysis gives rise to a measurable slow
down in reading, but little or no
conscious difficulty.•
Much easier than certain other kinds of reanalysis (Sturt et al. 1999a; Pritchett, 1988,
1992; Gibson 1991; papers in Fodor &
Ferreira 1998.)
(47)W
hile Wallace w
as eating the cheese was being stolen from
under his nose.
• Ease of reanalysis in (46) taken by som
e to indicate that (46) does not in fact involvereanalysis (e.g., W
einberg, 1993; Gorrell, 1995); adding the em
bedded verb justrequires addition of new
structural dominance statem
ents.
STR
ATEG
Y: By embedding structures like (46), m
ake a non-local attachment of the verb
available, which requires no reanalysis. If reanalysis is a last resort, the non-local
attachment w
ill be detected and selected instead of reanalysis.
(48)N
P
S
NP
The surprised wom
an
VP
Vdiscovered
S'
who
i
NP
locked the drunk m
an
Colin Phillips, Supper @ CU
NY
, 3/27/01
9
PRED
ICTION: com
bination of locality and easy reanalysis predicts embedded attachm
entto be easy; reanalysis as a last resort predicts m
atrix attachment bias.
(49)N
P
S
NP
The surprised wom
an
VP
locked the front doorlocked
the drunk man
S'
who
Vdiscovered
VP
S'
S
NP
himself
NP
S
NP
The surprised wom
an
VP
locked the front door S'
who discovered V
P
S
NP
herself
the drunk man
Low A
ttachment
H
igh Attachm
ent
Subjects: 56 undergraduate students. (7 additional subjects excluded for scoring <75% on
comprehension questions in experim
ental items or <80%
on filler items.)
Stimuli: 2 x 2 x 2 design: attachm
ent site (high, low), am
biguity, verb class (strongly NP-biased,
neutral or weakly N
P-biased). 48 sets of 4 items.
Verb classes based on Truesw
ell et al. (1993), Garnsey (unpublished com
pletion norms)
Strongly NP-biased verbs: hear, discover, acknowledge, appreciate, warn, understand
Weakly N
P-biased verbs: know, mention, doubt, notice
Disam
biguation with him
self, herself; gender counterbalanced across high/low sites.. A
ll possibleantecedents of reflexives are anim
ate NPs. 100 filler item
s
(50)a. em
bedded, ambiguous
The bilingual man w
ho knows the w
ell-traveled wom
an translated the travel booksherself w
ithout any extra help was quite im
pressed with the result.
b. embedded, unam
biguousThe bilingual m
an who know
s that the well-traveled w
oman translated the travel
books herself without any extra help w
as quite impressed w
ith the result.c. m
atrix, ambiguous
The bilingual man w
ho knows the w
ell-traveled wom
an translated the travel bookshim
self without any extra help from
a dictionary.d. m
atrix, unambiguous
The bilingual man w
ho knows her translated the travel books him
self without any
extra help from a dictionary.
Comprehension question accuracy: low
ambiguous – 80%
; high ambiguous 90%
[High-span readers]
The surprised wom
an who discovered/ that/ the drunk m
an / locked / the / front / door /
1 2 3 4 5 6 7
himself / w
ith / the / spare / key / was / am
azed he could even walk.
8 9 10 11 12 13 14
Figure 4: Residual Reading Times, self-paced reading, M
onotonicity Study 1
Embedded: Significant slow
down due to am
biguity at region 9 in embedded conditions
(F1 (1,24)=34.95, p<.0001; F
2 (1,42)=33.57, p<.0001M
atrix: No significant slow
down due to am
biguity at region 9 in matrix conditions:
all Fs<1.
• V
ery similar findings in Sturt et al. 2001; consistent w
ith data in Schafer et al. (n.d.:cited in Frazier &
Clifton 1998)
• A
lternative explanations of results:
(i) avoid embedding!
(ii) avoid long subject NPs!
(51)Experim
ent 2: manipulation of statistical verb com
plement bias
NP-biased verbs, hear, acknowledge, appreciate, discover, warn, understand
S-complem
ent biased verbs, claim, believe, suspect
Colin Phillips, Supper @ CU
NY
, 3/27/01
10
[High Span Readers]
Figure 5: Residual Reading Times, self-paced reading, M
onotonicity Study 2
•
Results: NP-com
plement bias verbs replicate results of Experim
ent 1. Pattern ofdifficulty is reversed in S-com
plement bias verbs.
•
Conclusion: Structure building is m
onotonic, i.e. reanalysis is a last resort.Syntactic search pursues a sim
ple algorithm – avoid reanalysis – but this is not
necessarily the most m
emory-efficient approach.
(iii) Distinguishing A
ccess and Search in Dative A
lternations (Phillips, Edgar &K
abak, 2000)
•
Question: Can the parser be fooled into creating bogus am
biguities with dative
verbs?
•
Much recent parsing w
ork attempts to assim
ilate parsing to lexical access (e.g.Tanenhaus &
Trueswell 1995, M
acDonald et al. 1994)
•
With lexicalized gram
mars, various errors should not occur, e.g. no confusion of
transitive & intransitive verbs, or N
P vs. S-complem
ent verbs (Boland, Trueswell)
• Tw
o limitations of previous w
ork on argument structure usage:
--overlap of syntactic and semantic selection: c-selection vs. s-selection
--even if lexical information is accessed, fruitful search not guaranteed
Dative/D
ouble-object Alternation
(52)The m
illionaire gave the painting to the museum
.[dative]
(53)The m
illionaire donated the painting to the museum
.
(54)The m
illionaire gave the museum
the painting.[double object]
(55) *The millionaire donated the m
useum the painting.
Other pairs: bring/carry, build/construct, assign, distribute, buy/purchase, get/obtain,
hand/deliver, mail/address, m
ake/create, offer/suggest, sell/contribute, show/display
Figure 6: Gram
maticality Ratings for sentences based on experim
ental materials (below
)
A garden-path involving double object verbs (56). If syntactic subcategorization
information is fully accessed and exploited, then (57) should present no difficulty.
(56)The duke bought the m
useum the elaborate antique chandeliers had gracefully
adorned for his son(57)
The duke purchased the museum
the elaborate antique chandeliers hadgracefully adorned for his son
Colin Phillips, Supper @ CU
NY
, 3/27/01
11
Control 1: compare reading tim
es to unambiguous conditions w
ith complem
entizerControl 2: add conditions w
ith monotransitive verbs, to gain independent m
easure of thedifficulty of analyzing unam
biguous V N
P NP sequences.
Materials: 36 sets of 6 conditions, 84 fillers
Fillers: more closely m
atched in length and complexity than in prelim
inary experiment
Participants: 54 undergraduate students
(58)a. The duke bought the m
useum the elaborate antique chandeliers had
gracefully adorned for his son in the Netherlands.
b. …bought the m
useum that the elaborate antique chandeliers…
c. …purchased the m
useum the elaborate antique chandeliers…
d. …purchased the m
useum that the elaborate antique chandeliers…
e. …w
anted the museum
the elaborate antique chandeliers…f. …
wanted the m
useum that the elaborate antique chandeliers…
RESU
LTS: alternator and non-alternator verbs clearly distinguished, but, non-alternatorconditions are by no m
eans easy!
The duke bought the museum
(that) the elaborate antique chandeliers had gracefully1 2 3 4 5 6 7 8 9 10 11 12adorned for his son … 13 14 15 16
ALTERN
ATO
R VERBS – M
ain effect of ambiguity
NP2 (regions 7-10): F1(1,53) = 2.48, n.s., F2(1,35) = 2.42, n.s.
Rel. Clause (regions 11-13): F1(1,37) = 27.53, p < 0.0001, F2(1,35) = 26.61, p< 0.0001
The duke purchased the museum
(that) the elaborate antique chandeliers had gracefully1 2 3 4 5 6 7 8 9 10 11 12adorned for his son … 13 14 15 16N
ON-A
LTERNA
TOR V
ERBS – Main effect of am
biguityN
P2 (regions 7-10): F1(1,53) = 30.07, p< 0.0001, F2(1,35) = 30.12, p< 0.0001Rel. Clause (regions 11-13):
F1(1,37) = 18.03, p < 0.0001, F2(1,35) = 17.93, p< 0.0001
MO
NO
TRAN
SITIVE V
ERBS – Main effect of am
biguityN
P2 (regions 7-10): F1(1,53) = 16.34, p< 0.0001, F2(1,35) = 16.41, p< 0.0001Rel. Clause (regions 11-13): F1(1,37) = 4.44, p = 0.04, F2(1,35) = 4.21, p= 0.04
Alternator and non-alternator verbs clearly contrast in this experim
ent, but…
(i) Why did w
e find the opposite of this in our pilot?(ii) W
hy are non-alternator conditions still difficult, since they’re unambiguous?
• Relative clauses in m
aterials: 12 trials with m
onotransitive verbs + relative clauses,30%
of filler sentences contain relative clause
• Full access to subcategorization inform
ation may have occurred in both experim
ents
Colin Phillips, Supper @ CU
NY
, 3/27/01
12
• Lexical access is not enough to parse successfully – gram
matical search m
ust alsosucceed.
• The experim
ent (unwittingly) created an environm
ent which provided a good deal of
support for parsing of V N
P NP sequences.
•
Support for this analysis: separate analysis of first half and second half ofexperim
ent shows that reading tim
e profiles for alternator and monotransitive verb
classes are stable across the experiment; for dative verbs, how
ever, there issubstantial change over the course of the experim
ent – the slowdow
n at the relativeclause disam
biguation is confined to the first half of the experiment, and in the
second half of the experiment the reading-tim
e profile is most sim
ilar to them
onotransitive verbs.•
More support for this analysis: w
e ran an additional experiment, w
hich was identical
to the previous one, except that we rem
oved the items that m
ay have been helpingaccurate parsing of V
NP N
P sequences with non-alternators, i.e. m
onotransitiveconditions and fillers containing relative clauses w
ere replaced with sentences
lacking relative clauses. Result: practice effects disappeared.
• C
ON
CLUSIO
N: gramm
atical search is a resource-limited process (perhaps in contrast
to lexical access); full and imm
ediate access to gramm
atical information associated
with lexical heads does not guarantee successful parsing.
(iv)…
but what about real am
biguities?
These are a less central concern, since there’s not the clear redundancy between
the parser-implem
entation and the gramm
ar-implem
entation of the same
linguistic knowledge.
One can, of course, try to show
that notions of ‘structural simplicity’ govern
(i) preferences in selecting between alternative interpretations
(ii) rigid constraints in selecting between alternative representations of a single
interpretation (i.e. structural economy conditions)
e. Learning
consequences largely unexplored
f. Overall
• If the approach is feasible, then one m
ight argue that parsimony m
akes it the nullhypothesis.
•
It is, of course, preferable to have stronger empirical argum
ents in favor of oneapproach or another.
g. Challenges
• W
hat if the parser does behave stupidly?•
What about head-final languages?
• N
eed further development of the im
plemented m
odel•
How
could this be implem
ented in the brain?•
etc.
References
Baltin, Mark &
Paul Postal. (1996). More on reanalysis hypotheses. Linguistic Inquiry 27,
127–145.Berw
ick, R. & A
. Weinberg. (1984). The G
ramm
atical Basis of Linguistic Performance. Cam
bridge,M
A: M
IT Press.Bever, T.G
. (1970) The Cognitive Basis for Linguistic Structures. In: J.R. Hayes (ed), Cognition and
the Developm
ent of Language, pp.279-352. New
York: W
iley.Boeckx, C. (1999). Program
matic notes for a top-dow
n syntax. Ms., U
niversity of Connecticut.Boeckx, C. &
S. Stjepanovic. (1999). The wh/clitic connection in Slavic. Paper presented at the
Formal Approaches to Slavic Linguistics conference, U
. of Pennsylvania.Bourdages, J. (1992). Parsing com
plex NPs in French. In H
. Goodluck &
M. Rochem
ont (eds.),Island Constraints: Theory, Acquisition &
Processing. Dordrecht: K
luwer.
Chomsky, N
. (1965). Aspects of the theory of syntax. Cambridge, M
A: M
IT Press.Crain, S. &
J.D. Fodor. (1985a) H
ow can gram
mars help parsers? In D
.R. Dow
ty, L. Karttunen &
A.
Zwicky (eds.), N
atural language parsing: Psychological, computational and theoretical
perspectives, 94-128. Cambridge, U
K: Cam
bridge University Press.
Crain, S. & J.D
. Fodor. (1985b). Rules and constraints in sentence processing. NELS 15, 87-104.Crain, S. &
J.D. Fodor. (1987). Sentence m
atching and overgeneration. Cognition 26, 123-169.D
rury, J. (1998). The promise of derivations. G
roninger Arbeiten zur Germ
anistischen Linguistik 48,63–107.
Ferreira, F., Bailey, K. G
. D. &
Ferraro, V. (2001). G
ood Enough representations in LanguageCom
prehension. Current Directions in Psychological Science, in press.
Fodor, J. A., Bever, T. G
., & G
arrett, M. (1974) The Psychology of Language. N
ew Y
ork: McG
rawH
ill.Fodor, Janet &
Fernanda Ferreira. (1998). Reanalysis in Sentence Processing. Dordrecht: K
luwer.
Fodor, Janet & Lyn Frazier. (1980). Is the hum
an sentence parsing mechanism
an ATN
? Cognition,8, 417–459.
Fodor, J. D. (1998). Learning to Parse. Journal of Psycholinguistic Research, 27, 285-319.
Forster, K. &
B. Stevenson. (1987). Sentence matching and w
ell-formedness. Cognition, 26, 171-
186.Fram
pton, J. (1990). Parasitic gaps and the theory of wh-chains. Linguistic Inquiry, 21, 49-77.
Frazier, L. & C. Clifton. (1989). Sucessive cyclicity in the gram
mar and the parser. Language &
Cognitive Processes, 4, 93-126.Freedm
an, S. & K
. Forster. (1985). The psychological status of overgenerated sentences. Cognition,19, 101-131.
Guim
araes, M. (1999). D
eriving prosodic structure from dynam
ic top-down syntax. M
s., University
of Maryland.
Kem
pson, R., Meyer-V
iol, W., &
Gabbay, D
. (2001). Dynam
ic Syntax: The Flow of LanguageU
nderstanding. Oxford, U
K: Blackw
ell.K
imball, J. (1973). Seven principles of surface structure parsing in natural language. Cognition, 2,15-47.
Kluender, R. &
M. K
utas. (1993a). Subjacency as a processing phenomenon. Language &
CognitiveProcesses, 8, 573-633.
Colin Phillips, Supper @ CU
NY
, 3/27/01
13
Kluender, R. &
M. K
utas. (1993b). Bridging the gap: Evidence from ERPs on the processing of
unbounded dependencies. Journal of Cognitive Neuroscience, 5, 196-214.K
urtzman, H
. & L. Craw
ford. (1990). Processing parasitic gaps. Proceedings of NELS 21. 217-231.Larson, Richard. (1988). O
n the double object construction. Linguistic Inquiry 19, 335–392.Legate, J. (1999). The m
orphosyntax of Irish agreement. m
s., MIT.
Levelt, Willem
. (1974). Formal G
ramm
ars in linguistics and psycholinguistics (3 volumes). The
Hague: M
outon.M
cElree, B & T. G
riffith. (1998). Structural and lexical constraints on filling gaps during sentencecom
prehension: A tim
e-course analysis. Journal of Experimental Psychology: Learning,
Mem
ory & Cognition, 24, 432-460.
McK
innon, R. & L. O
sterhout. (1996). Constraints on movem
ent phenomena in sentence processing:
Evidence from event-related brain potentials. Language &
Cognitive Processes, 11, 495-523.N
eville, H., J. N
icol, A. Barss, K
. Forster, & M
. Garrett. (1991). Syntactically based sentence
processing classes: Evidence from event-related brain potentials. Journal of C
ognitiveNeuroscience, 3(2), 151-165.
Pesetsky, David. (1995). Zero Syntax: Experiencers &
Cascades. Cambridge, M
A: M
IT Press.Phillips, C. (1995). Right association in parsing and gram
mar. M
ITWPL, 26, 37-93.
Phillips, C. (1996). Order and Structure. Ph.D
. dissertation, MIT.
Phillips, Colin & E. G
ibson. (1997). On the strength of the local attachm
ent preference. Journal ofPsycholinguistic Research, 23, 323–346.
Phillips, Colin & E. G
ibson. (1997b). Local attachment and com
peting constraints. 10th A
nnualCU
NY
Conference on Sentence Processing, Santa Monica, CA
.Phillips, C. &
K. W
ong. (2001). Parsing parasitic gaps: how the parser solves a look-ahead problem
.Presented at the LSA
Phillips, Colin. (1998, to appear). Linear Order and Constituency. Linguistic Inquiry.
Phillips, C., E. Edgar & B. K
abak. (2000). Lexical access and syntactic search. ms., U
niversity ofD
elaware.
Pickering, M, S. Barton &
R. Shillcock. (1994). Unbounded dependencies, island constraints and
processing complexity. In C
lifton, Frazer, & R
ayner (eds.), Perspectives on SentenceProcessing (pp. 199-224). London: Erlbaum
.Postal, Paul. (1986). Studies of Passive Clauses. A
lbany, NY
: SUN
Y Press.
Richards, N. (1999). D
irectionality of tree construction. MITW
PL #33.Richards, N
. (2000). Very-local A
-bar movem
ent in a root-first derivation. ms., M
IT.Schlenker, P. (1999). La flexion de l’adjectif en allem
and: la morphologie de haut en bas. [G
erman
adjective inflection: top-down m
orphology.] Recherches Linguistiques de Vincennes 28.Schneider, D
. (1999). Parsing and Incrementality. Ph.D
. dissertation, U. of D
elaware.
Schneider, D. &
Phillips, C. (2001). Gram
matical search and reanalysis. Journal of M
emory &
Language, in press.Stevenson, B. (1993). U
ngramm
aticality and Stages in Sentence Processing. Ph.D. dissertation,
Monash U
niversity.Stow
e, Laurie. (1986). Evidence for on-line gap location. Language and Cognitive Processes 1,227–245.
Sturt, P., M. Pickering, C. Scheepers &
M. Crocker. (2001, in press). The preservation of structure in
language comprehension: Is reanalysis the last resort? Journal of M
emory and Language.
Sturt, P., Costa, F., Lombardo, F., &
Frasconi, P. (2001). Learning first-pass structural attachment
preferences with dynam
ic gramm
ars and recursive neural networks. m
s. Universities of
Glasgow
, Florence & Eastern Piedm
ont.Tow
nsend, D. &
Bever, T. G. (2001). Sentence Com
prehension. Cambridge, M
A: M
IT Press.Traxler, M
. & M
. Pickering. 1996. Plausibility and the processing of unbounded dependencies: An
eye-tracking study. Journal of Mem
ory and Language 35, 454-475.
Colin PhillipsD
ept. of Linguistics / Cogn. Neurosci. of Language Lab.
University of M
aryland1401 M
arie Mount H
allCollege Park, M
D 20742
d.eduhttp://w
ww
.ling.umd.edu/colin