Learnability and Cognition: The Acquisition of Argument Structure

412 Pages·1991·55.55 MB·English

by  Steven Pinker

SeriesF oreword
This seriesi n learning, developmen,t and conceptuacl hangew ill include state-
of-the-art referencew orks, seminalb ook-lengthm onographs, and texts on the
developmento f conceptsa nd mental sb' Ucture.s It will span learning in all
domainso f knowledge, from syntaxt o geometryt o the social world, andw ill be
concernedw ith all phaseso f developmen,t from infancy througha dulthood.
The seriesi ntendst o engages uchf undamentaql uestionsa s
Then aturea nd limits of learning and maturation: the influence of the environment
, of initial Sb' Ucture,s ando f maturationacl hangesin the nervouss ystemo n
human developmen;t leamability theory; the problem of induction; domain-
specificc onstraintso n developmen.t
The nature of conceptual change: conceptualo rganization and conceptual
changei n child developmen,t in the acquisitiono f expertise, and in the history
of science.
Lila Gleitman
SusanC arey
ElissaN ewport
ElizabethS pelke
,
Acknowledgments
In developingt he ideasp resentedin this book, I was lucky to havee ncountered
audiencesw ho refusedt o believe them, studentsw ho refused to pretendt o
understandth em, andc hildren who refusedt o behavei n accordancew ith them.
Facingt hesec hallengesle d me to discoveriest hat provided the mosts atisfying
momentso f this research.
JessG ropenh ass haredm y enthusiasmfo r this topic during the entiret ime I
havew orked on the book, and I haveb enefitedg reatly from our discussionso n
every aspec.t His independentp roposals on how to grapple with various
problems were invariably of great help, and the ingenious experimentsh e
developeda nde xecuteda rea crucial part of the research. I am happyt o be able
to thank him for thesei nvaluablec ontributions. Among other graduates tudents
at MIT, PaulB loom andK arin Stromswolda lsop rovidedh elpful commentsa nd
discussions.
Jill Gauldinga ndM arc Light took on ast heir seniorr esearchp rojectst he task
of implementingp artso f the theorya sa computers imulation. Their penetrating
analysiso f the representationafol rmalism and learninga lgorithmsl ed to countless
improvementsin thep recision, economy, anda ccuracyo f thesem echanisms
and in the clarity of the exposition.
Michelle Hollander and Richard Goldberg assistedi n the developmental
researchw ith dedication, intelligence, and skill. Loren Ann Frost, Ronald
Wilson, and Larry Rosend eservet hanksf or their work on earlier experiment.s
I ama lso very gratefult o the child-carec entersi n the Bostona reat hat invited us
in to conductt he research.
I am in debtt o a numbero f researcherws ho haves haredt heir findingsa ndd isagreements
. Melissa Bowermanh as doubted whether constraintso n lexical
rules could get the child out of the learnability paradoxI have addresse.d Lila
'
Gleitmanh asq uestionedh ow mucho f a verbs meaninga child could learnf rom
the situationsi n which it is used. JaneG rimshawa ndJ anetR andallh avew arned
Acknowledgments xiv
againstn eglectingp ropertieso f the grammaticalr epresentationo f argument
structurea sa sourceo f learningc onstraints. JanetF odorh asb eens kepticala bout
'
how productivec hildrens useo f rules really is. KennethW exier has argued
againsta ssumingth att he biologicalm echanismso f languagea cquisitionr emain
unchangedth roughc hildhood. I think that all of thesep eoplea re right about
something, and I have strived toward a theory that is eclectic enough to
encompasasl l of their insightsi n somef orm, thoughn aturally thesep eoplec an
be expectedto continuet o find its weaknesse.s What makest he topic so much
fun to work on is thati t is cleart hat the ultimateb estt heory, thoughe clectic, will
" "
not be a banalr esignationt o the effect that anything can happen. There are
striking regularitiesin argumenst tructurea ndi ts acquisition, andI am glad to be
part of a researchc ommunityt hat is working toward discoveringt hem.
I alsoh aveb eenf ortunatet o havew orkedo n this projecta t MIT during a time
whent he LexiconP rojecta t theC enterf or Cognitive Sciencew as in full swing.
Beth Levin, directoro f the projectt hrough 1987, haso ffered many helpful comments
on this work, and the theoryo wesa greatd eal to her research. Levin, Jay
Keyser, and KennethH ale createda stimulating environmentw ith a seminar
seriesa nd technicalr eportst hat were an important catalyst in the research. A
visiting positioni n the Departmenot f Psychologya t BrandeisU niversity gave
an official statust o my very helpful discussionsw ith Jane Grimshaw, Ray
Jackendof,f Alan Prince, andJ erry Same.t
Lila Gleitman, JaneG rimshaw. RayJ ackendof,f BarbaraL andau, Beth Levin,
and severala nonymousre viewersr eadp ortions of the manuscripta nd offered
many invaluables uggestion.s KatarinaR ice editedt he manuscriptm aste rfully.
Michael Tarr, David Plotkin, and Kyle Cave gave me generousa dvice on
compute-rrelatedm atters, includingw hat to do with a blazing Sun. I am grateful
to all of them.
I thankm y parents, Harry andR oslyn, and my brothera nd sister, Robert and
Susan, for their interesta nd encouragemen.t
AlthoughN ancyE tcoff doesn ot sharem y obsessionw ith verbs, her interests
encompassa ll aspectso f mind, and I have benefited from her insights on
" "
countlessm atters. For these ninsights and for her support, I thank her.
This researchw asf undedb y NIH grantH D 18381a ndb y a grantf rom the Alfred
P. SloanF oundationto the MIT Centerf or Cognitive Science.
Chapter 1
A Leamability Paradox
Some of the most rewarding scientific pursuits begin with the discovery of a
paradox. Nature does not go out of its way to befuddle us, and if some
phenomenon seems to make no sense no matter how we look at it, we are
probably in ignorance of deep and far-ranging principles . For anyone interested
in the human mind, language offers many such opportunities for discovery.
Language is created anew each generation, so details of grammar, even subtle
and intricate ones, are products of the minds of children and bear the stamp of
their learning abilities.
This book is about a paradox in language acquisition. The paradox begins with
a small linguistic puzzle: Why does He gave them a book sound natural, but He
'
donated them a book sound odd? It is complicated by a fact about children s
environment - that they are not corrected for speaking ungrammatically - and
a fact about their behavior - that they do not confine themselves to the verb
phrase structures they have heard other people use. In trying to resolve this
paradox, we must face fundamental questions about language and cognition:
When do children generalize and when do they stick with what they hear? What
is the rationale behind linguistic constraints? How is the syntax of predicates and
arguments related to their semantics? What is a possible word meaning? Do
languages force their speakerst o construe the world in certain ways? Is there a
'
difference between a word meaning and a concept? Why does children s
language seem different from that of adults? The goal of this work is to resolve
the learning paradox and to show how the solution leads to insight into these deep
questions.
The strategy I will follow comes out of what is sometimes called the
learnability approach to language acquisition (Hamburger and Wexier, 1975;
Pinker, 1979; Wexier and Culicover, 1980; Baker and McCarthy, 1981). This
approach focuses on the logical nature of the task facing the child as he or she tries
to learn a language and on the mental representations and proces ses that make
Chapter1
suchl earnings uccessfu. lI will pursueth es olutiont o thel earningp aradoxr elentlessly
'
, trying to createa trail that leadsf rom the prelinguisticc hild to the adults
commando f subtled iscriminationso f linguistic structure. Thoughp artso f the
trail mayb er oughg oing, whati s mosti mportanti s thate achs egmenlti nk up with
'
the next to form an unbrokenp atho f explanationf rom childrens experiencet o
'
adults knowledge.
In this chapterI outlinet hep roblem: first, the specificd omaino f languagea nd
why it is importan,t thent he logic of languagele arningi n genera,l then thej uxtaposition
of the two thatc reatesth e learningp uzzle. Then I considers omeh alf-
dozens implew aysi n which thep roblemm ight bee liminated. All can be shown
to be incorrecto r unsatisfactor.y In my mind this is what elevatest he problem
from a puzzlet o a paradox, which the rest of the book attemptst o solve.
Chapter2 discus ses phenomenath at point to a way out of the paradoxa nd
presentse videncet hat that pathi s the right one. The next threec hapterso utline
a theoryo f adult linguistic knowledget hat is logically capableo f resolvingt he
paradoxw hile providinga ne xplanationfo r thef orm of that knowledge. Chapter
3 tries to makes enseo f the phenomen,a making them fallout of more general
principles. Chapter4 extendst hosep rinciples so that the original linguistic
problemsc anb es olvedi n detail. Chapter5 dealsw ith representatio;n it presents
andj ustifies an explicit descriptiono f the representationasl tructuresf or verb
meaningsa nd rulest hat the theoryn eeds.
The next two chapterst ake up the psychologicalp roces ses for acquiring the
linguistic knowledgeu nderlyingt he solutiont o the paradox. Chapter6 is about
learning; it discus sest hec omputationapl roblemo f how the linguistic structures
area cquiredt hroughi nteractionw ith the environmen,t and it outlinesa proposal
'
for how the child doest his. Chapter7 is aboutc hildrens developmen;t in it I
comparet he factso f child languagew ith the acquisitionp roblemsa nd mechanisms
discussedp reviously. In thec oncludingc hapterI spell out somei nteresting
implications that the solution of the paradox holds for language and
cognition.
Much of this book is aboutw ords, andt his calls for a speciala pology. People
know tens of thousandso f words, no two alike, making the mental lexicon a
domain of immeasurabler ichness. Any theory that tries to find common
organizing principles amongstt his richnessc an be confronted with a huge
numbero f empiricalt ests. While this makesf or lively linguistic argumentation,
at times it canb e overwhelming. In the middle chapters( 3,4, and 5) I describe
a theory of the mental representationo f words and rules whosem achineryi s
outlinede xplicitly andw hich is buttressedw ith manyl inguistic data. I havet ried,
however, to organizet he material so that it can be absorbedb y readersw ith
varyingd egreeosf e xpertisaen din tere,s itncludintgh osew ithl ittleb ackground
in linguistic.s
A LeamabiliPtya radox
The key ideaso f thesem iddle chaptersa re presentedin overviews ectionsa t
the beginninga nd in summarys ectionsa t the end. The first sectiono f chapter3 ,
section3 .1, is a capsuled escriptiono f the theoryd iscussedin that chapte,r and
similarly section4 .1 motivatesa nd previewst he claimso f chapter4 . The final
sectiono f chapter4 spellso ut the relationshipb etweent he two key partso f the
theory, the one presentedin chapter3 and the one presentedin chapter4 . The
generalj ustification for the theoryo f representationin chapter5 is presentedin
sections5 .1t hrough5 .4, andt he accomplishmentosf thet heorya res ummarized
at the end of the chapter. Finally, chapter8 beginsw ith a brief recapitulationo f
everything that went before.
The detailed linguistic discussionsi n the middle of chapters3 -5 are also
modularly organized. In each one I begin with linguistic evidencet hat is
independenot f the problemsI try to solve. Thesec anb e found in sections3 .2,
4.2, and5 .3-5.4. I presentth et heoreticacl laimse xplicitly in sections3 .3,4.3 and
5.5. In the remaining sectionsI apply the theory to each of four linguistic
phenomena, thed ative, causative, locative, andp assivea lternation.s Becauseth e
topic of this book is the psychologyo f languagea cquisition, I havec hosent o
organize the book around issueso f representationa nd learning rather than
aroundt he linguistic phenomen,a and this meanst hat I discusse acho f the four
alternationsa numbero f times. The sectionsin which the individuala lternations
are discusseda re self-contained, labeled, and cross-reference,d and specialists
with an interesti n onea lternationc ans kip or skim theo thers. Readersw ho want
to seet he theory appliedi n detail to one illustrative alternationa ree ncouraged
to track the discussionso f the dative.
But let me get on with the paradox.
1.1 Argument Structure and the Lexicon
Humanl anguagesd o not defines traightforwardm appingsb etweenth oughtsa nd
words. To get a sentenc,e it is not enought o selectt he appropriatew ords and
string them togetheri n an ordert hat conveyst he meaningr elationshipsa mong
them. Verbsa rec hoosy; not all verbsc ana ppearin all sentence,s evenw hent he
combinationsm ake perfects ense, as shown in (1.1).
(1.1) Johnf ell.
. Johnf ell the floor.
Johnd ined.
. Johnd ined the pizza.
Johnd evouredt he pizza.
. Johnd evoured.
Chapter1
Johna te.
Johna tet he pizza.
Johnp ut somethings omewhere.
* Johnp ut something.
*
Johnp ut somewhere.
*Johnp ut.
Thesef actsd emonstrateth e phenomenono ften referredt o as subcategorization
: different subcategorieos f verbsm aked ifferent demandso n which of their
argumentsm ustb e expresse,d which canb e optionally expresse,d and how the
expresseda rgumentsa re encodedg rammatically- that is, as subjects, objects,
or oblique objects( objectso f prepositionso r oblique cases). The propertieso f
verbs in different subcategoriesa re specified by their entries in the mental
lexicon, in data structuresc alled argument structures (also called predicate
arguments tructures, subcategorizationfr ames, subcategorization, scasef rames,
lexical forms, andt hetag rids). Thus the arguments tructureo ffall, dine, andt he
intransitivev ersion of eat would specify that only a subjecti s permi tted. The
arguments tructuresf or devour and the transitive versiono f eat would specify
that a subjecta nd an object are required. The arguments tructuref or put would
call for no more and no lesst han a subjec,t an object, and an oblique object.
Lexical arguments tructures play an extremely important role in modem
theorieso f language. Beginningw ith Aspectso f theT heoryo f Syntax( Chomsky,
1965) andc ontinuingt o thep resent, it hasb ecomea pparentth atm anyo f thef acts
of grammara rec ausedb y propertieso f the particular lexical itemst hat go into
sentence.s Recentt heorieso f grammars pecify rich collectionso f information
in lexical entries and relatively impoverished rules or principles in other
componentso f grammar( e.g., Chomsky, 1981; Bresnan, I 982a). Sentences
conform to the demandso f the words in them becauseo f generalp rinciples( for
example, Chomsky's Theta-Criterion andB resnan's Coherencea ndC ompleteness
Principles) that deema sentencet o be grammaticalo nly if the arguments
'
specifiedb y thev erbs arguments tructurea rea ctually presenta sc onstituentsin
'
the sentencea nd vice versa. Chomskys ProjectionP rinciple specifiesf urther
'
that the demandso f verbsarguments tructuresm ust be satisfieda t every level
of sentencere presentatio,n not just deeps tructure.
'
Sincev erbsarguments tructuresa ssumes ucha largeb urdeni n explainingt he
facts of language, how argument structuresa re acquiredi s a correspondingly
crucial part of the problemo f explaining languagea cquisition. (In fact, Elliott
andW exler, in press, haveg ones o far ast o suggestth at languagea cquisitionm ay
be nothing but the acquisitiono f information aboutt he wordsi n the language.)
How arguments tructuresa re acquiredi s intertwined with the questiono f why
A LeamabilityP aradox
1.2 The Logical Problem of Language Acquisition
Languagea cquisitioni n genera,l andt hea cquisitiono f verb arguments tructures
in panicular, canb e thoughto f in the following te nDS . The child hearsa finite
(cid:0)
numbero f sentencefsro m his or herp arentsd uring the language-learningy ears,
'
which are symbolizedb y th eX s in (1.2). But a languageis an open-endeds et,
not a fixed list, so the child must generalizef rom thesei nputs to an infinite set
of sentencetsh at includest he input sampleb ut goesb eyondi t. This is showni n
(1.2) ast hec ircle with the arrow pointing to it. As in all induction problems, the
hardp art is that an infinite numbero f hypothesesa re consistentw ith the input
sampleb ut differ from eacho ther and from the correct hypothesis( the actual
targetl anguag)e in ways that are not detectableg iven the input samplea lone.
Someo f the incorrecth ypothesesa re depictedb y the other circles in (1.2).
(1.2)
The solution to this (or any other) learning problem works as follows . Constraints
on the learner force him to entertain a restricted set of hypotheses that
includes the correct hypothesis but excludes many others. The learner can then
compare the predictions ofa hypothesis (which sentencesi t generates) with the
input data so that incorrect hypotheses can be rejected.
'
There are four ways in which one of the the child s hypotheses can be incorrect
'
before learning is successful. The child s language can be disjoint from the target
language, as in ( 1.3a).ln this casea ny sentencei n the input is sufficient to infonn
the child that the hypothesis is wrong. Such sentences, called positive evidence,
Chapter1
" "
ared epictedin the figure with a + symbo.l Likewise, if the languageg enerated
'
by thec hilds hypothesisg rammarin tersectsth e targetl anguage, as in (1.3b), or
is a subseto f it, as in (1.3c), positivee videncec onsistingo f input sentencesin
the nonoverlappingre gion of the targetl anguages uffices to impel the child to
rejectt he hypothesi.s However, if the child entertainsa grammarg eneratinga
superseot f the targetl anguage, asi n (1.3d), no amounto f positive evidencec an
strictly falsify the guess. What he or shen eedsi s negativee vidence: evidence
aboutw hichw ord stringsa reu ngrammatica(lt hat is, not in the targetl anguage).
This is shown as the "-" symbolsi n (1.3d). Explaining successfull earning
basically consistso f showing that the learnerc an entertain and stick with a
correcth ypothesisa ndc an falsify any incorrecto nes( see, e.gOsherson, Stob,
and Weinstein, 1985; Pinker, 1979; Wexier and Culicover, 1980).
(1.30) H 0T OH TJ T H
@ @
(a) (b) (c) (d)
The first important question about child languagea cquisition is whether
negative evidence is available. Obviously no one gives children a list of
ungrammaticasl entencest agged with asterisks. The most psychologically
plausiblek ind of negativee videncew ouldb es omes orto f parentalf eedbackt hat
childrenm ight receiver elatedt o whethert heir own utterancesa re grammatical
or not, sucha sc orrectionso r expressionos f approva.l The availablee vidence
suggestst hat children are not provided with such information. Brown and
Hanion (1970) found that parentsd o not differentially expressa pproval or
'
disapprovacl ontingento n whethert he childs prior utterancew as well formed;
nor do they understandw ell-formedq uestionsb ettert han ill-formed ones. As a
resulti t is commonlya ssumedth atc hildrend o not dependo n negativee vidence
to acquire a language. This meanst hat they cannot engage in the sort of
hypothesisfa lsification illustratedi n figure (2d); eithert hey nevere ntertaina ny
hypothesist hat is a superseto f the target language, or, if they do, some
endogenoufso rce must impel themt o abandonit , becauseth e world will never
force them to. On the other hand, children cannots imply stick with the exact
sentencetsh eyh ear, becauseth eym ustg eneralizeto the infinite languageo f their
community. This tension, betweent he needt o generalizea nd the need not to
generates uperset,s characterizesm anyo f the toughestp roblemsi n explaining
human languagea cquisition. Someo f thesea re discussedb y Baker (1979),
Berwick (1986), Bowerman (1987a, 1987b), Braine (1971), Chomsky and
Lasnik (1977), Pinker (1982, 1984), and Wexier and Culicover ( 1980).
A LeamabilityP aradox
'
1.3 Bake Fs Paradox
Now we can juxtapose the facts of arguments tructurew ith the logic of the
learningp roblem. The acquisitiono f the syntacticp ropertieso f verbsi s oneo f
the clearest casesi n which the no-negative-evidencep roblema rises. Though
GeorgiaG reen (1974; pp. 3, 199) first pointedo ut a learningp aradoxb asedo n
it, C. Lee Baker (1979) discussedit in a largerc ontextt hat drewm orea ttention.
Considera child hearings entencep airs sucha s thosei n (1.4) and fonning the
associateda rguments tructures.
( 1.4) John gave a dish to Sam.
give: NP I - NP2 to-NP3
John gave Sama dish.
give: NP .I - NP-] NP-2
Johnp assedth es alamti o Fred.
pass: NP. - NP2 to-NPJ
John passedF red the Salam.i
pass: NP._ NPJN P2
John told a joke to Mary.
tell: NP. - NP2 to-NPJ
John told Mary a joke.
tell: NP. - NPJ NP2
It would seem to be a reasonableg eneralizationt hat any verb with the
NPI - NP2 to-NPJ arguments tructure( prepositionadl ative) could alsoh avea
NP. - NPJ NP2 a rguments tructure( double-objectd ative). This generalization
could be capturedi n, say, a lexical rule sucha s that in (1.5), which would allow
the child to createa double-objectd ativec orrespondingto anyp repositionaol ne
for somen ew verb (e.g., send), eveni f he or sheh adn everh eardt he verb in the
double-object form.
(1.5) NP. - NP2t o-NPJ - > NP,- NP3N P2
The problem is thatn ot all thev erbsw ith thep repositionaal rgumenst tructure
dativize (that is, appeari n both versionso f the alternation), as (1.6) shows.
(1.6) John donateda painting to the museum.
* John donatedt he museuma painting.
John reportedt he accidentt o the police.
* John reportedt he police the acciden.t
But the child hasn o way of knowingt his, given then onavailabilityo f negative
'
evidence. The fact that he or sheh asnt heardt he ungrammaticasl entenceisn
'
(1.6) could simply reflect adultsneverh avingh ada n opportunityt o uttert hem