Syntax and semantics in the acquisition of locative verbs

However, children must learn which verbs specify which of their arguments as being affected specifically, whether it is the argument whose referent is undergoing a change of location

Syntax and semantics in the acquisition of locative

verbs*

Stanford University

Massachusetts Institute of Technology

University of Maryland

(Received 8 August I989 Revised 19 February I990

ABSTRACT

Children between the ages of three and seven occasionally make e•'rors with

locative verbs like pour and fill, such as * I filled water into the glass and * I

poured the glass with water (Bowerman, I982 ) To account for this pattern of

errors, and for how they are eventually unlearned, we propose that children

use a universal linking rule called OBJECT Ai•FECTEDNESS: the direct object

corresponds to the argument that is specified as 'affected' in some particular

way in the semantic representation of a verb However, children must learn which verbs specify which of their arguments

as being affected specifically,

whether it is the argument whose referent is undergoing a change of location, such as the content argumen t of pour, or the argument whose referent is

undergoing a change of state, such

as the container argument of fill This

predicts that syntactic errors should be associated with specific kinds of

misinterpretations of verb meaning Two experiments

were performed on

the ability of children and adults to understand and produce locative verbs The results confirm that children tend to make syntactic errors with

sentences containing fill and empty, encoding the content argument

as direct [*] We thank Kay Bock, Susan Carey, Eve Clark, Ken Wexler and Carol Tenny for their helpful comments We

are also grateful to the directors, parents arid especially children

of the following centres: Bowen After School Care Program, Inc., Cambridge Nursery School, Central School, Children's Village, Inc., Creative Development Center, KLH Center, Needham Children's Community Center, Newton Community Service Center,

Plowshares Child Care Program, Recreation Place, Rosary Academy Learning Center, Temple Beth Shalom, and the Zervas Program The research reported here is part of the first author's doctoral dissertation Experiment was presented at the Twelfth Annual

Boston University Conference on Language Development This research

was supported

by NIH grant HD I838I to the second author, and by a grant from the Alfred P Sloan Foundation to the MIT Center for Cognitive Science Address for correspondence: Jess

Gropen, Department of Linguistics, Stanford University, Stanford, CA 943o5, USA

5

object (e.g fill the water) As predicted, children also misinterpreted the

meanings of fill and empty as requiring not only that the container be brought

into a full or empty state, but also that the content move in

some specific

manner (by pouring, or by dumping) Furthermore, children who mis-

interpreted the verbs' meanings were more likely to make syntactic errors

with them These findings support the hypothesis that verb meaning and

syntax are linked in precise ways in the lexicons of language learners

INTRODUCTION

Although syntax and semantics interact in the generation of

errors and the

recovery from them in language development, there have been few concrete

demonstrations of how this works in detail The purpose of tfiis work is to

understand this interaction in children's acquisition of locative verbs verbs

such as pour, fill, empty and load Locative verbs express an event involving

the transfer of CONTENT to (I) or from (2) a CONTAINER They are further

subdivided as to whether the content argument (i a, 2a) or the container

argument (i b, 2b) is encoded as the syntactic direct object We shall refer to

these syntactic forms as CONTENT-OBJECT SENTENCES and CONTAINER-OBJECT

SENTENCES, respectively, and to the verbs appearing in them as CONTENT-

OBJECT VERBS and CONTAINER-OBJECT VERBS (see Schwartz-Norman, I976)

Some locative verbs (i c, 2c), which we shall call ALTERNATORS, may accept

either the content or container argument as direct object (We are only

concerned with readings of these sentences in which both postverbal phrases

are arguments of the verb, and not in cases where the PP is taken to be an

embedded modifier of the direct object, e.g Gus dumped the can of garbage,

but not the can of compost.)

a Betty poured water into the cup/*poured' the cup with water

Tom dripped paint onto the floor/*dripped the floor with paint

b Mike filled the cup with water/*filled water into the cup

Lloyd covered the bed with a sheet/*covered a sheet onto the bed

c George loaded the gun with ammo/loaded ammo into the gun

Dan stuffed the hamper with laundry/stuffed laundry into the

hamper

2a Gus dumped garbage from the can/*dumped the ,can of garbage

Tom wiped paint from the brush/*wiped the brush of paint

b Bob ridded the room of bugs/*ridded bugs from the room

The crops depleted the soil of nutrients/*depleted nutrients from the

soil

c Sally emptied the carton of ice cream/emptied ice cream from the

carton

Bob drained the sink of water/drained water from the sink

ii6

Bowerman (i 982) has documented an interesting 13-shaped developmental

pattern in the production of locative sentences by her two children although Christy and Eva initially used these verbs correctly, errors emerge within the

range of three to seven years of age, after which the errors decline Bowerman found that the most frequent errors involved children overextending the content-object form to verbs that ordinarily encode only the container argument as direct object, as in I didn't fill water up to drink it (Eva, 4 I) According to Bowerman, errors of the converse type-involving use of incorrect verbs in the container-object construction-are less frequent Examples of both kinds of errors appear in Table i Pinker (i 989) also found

seven errors involving locative verbs in three different four-year-old children, such as I filled the grain up and And fill the little sugars up in the bowl

TABLE I Examples of overgeneralization (Bowerman, I982)

Errors with the content argument as direct object

E (3;0) I'm going to touch it on your pants

E (4; I) didn't fill water up to drink it; filled it up for the flowers to drink it

C (4; 3) M Simon says, Touch your toes'

C: To what (interprets toes as content, is now looking for container) [note that this is a comprehension error]

E (4; 5) I'm going to cover screen over me

C (4;9) She's gonna pinch it on my foot

E (4; xI) And I'll give you these eggs you can fill up (giving M beads to put into

cloth chicken-shaped container)

E (5;o) Can fill some salt into the bear bear-shaped salt shaker]

E (5 3) Terri said if this rhinestone on a shirt] were diamond then people

would be trying to rob the shirt

C (6; xo) Feel your hand to that

Errors with the container argument as direct object

E (2; 1) Mommy, poured you [M: You poured me ?] Yeah, with water

E (•4; i) don't want it [= toast] because spilled it of orange juice

Bowerman notes that this pattern of development suggests a process of re- organization, driven by the child's discovery that a set of verbs, acquired independently, have a common kind of mapping of semantic arguments onto syntactic roles This process, she points out, is similar to what applies in the familiar example from inflectional morphology In the case of locative verbs, children first use individual verb-specific forms such as fill the glass (cf the morphologically-irregular broke, first learned by rote) Then, they abstract the pattern whereby verbs like load that take the content-object sentence can

also take the container-object sentence This allows them to overgeneralize

the content-object form to container-object verbs, such as infill the water (cf breaked) Bowerman suggests (following Talmy, I972) that the content- object .form is overgeneralized more than the container-object form because

it is the dominant pattern in English for expressing locative events Ac-

CHILD LANGUAGE

cordingly, the overregularization of the container-object form is less

common

(e.g 1spilled it of orange juice) for the same reason that the overregularization

of irregular past tense inflections is

rare (e.g brang on the pattern of sang)

Although there are similarities between late errors with irregular inflection

and locative forms, there are also some dissimilarities that merit further

examination First, whereas irregular morphological forms generally con-

stitute a minority of the lexicon, and are by definition idiosyncratic, neither

is true of the locative verbs that are overgeneralized According to Rappaport

& Levin's (1985) near-exhaustive list of

125 locative verbs involving addition

of content to

a container (not counting two verbs from the list which

were

placed into more than

one syntactic subcategory), the non-alternators

are in the majority (9z), not the minority Second, among the non-alternators there

are almost four times as many container-object verbs (73)

as content-object verbs (19), exactly the opposite pattern to that required to explain the greater

frequency of errors with container-object

verbs Furthermore, the

same

rankings are obtained if

we consider token frequencies: according

to the

frequency analysis performed by Francis

& Kucera (1982) on the basis of a million-word corpus, the

sum of text-frequencies is greater for the

non-

alternators (1295) than for the alternators (658), and greater for the container-

object verbs (944) than for the content-object verbs (351) (Of course, their

frequency analysis fails

to distinguish between different

argument structures

that a verb may take, and it bears only an indirect relationship to adult-to-

child speech.)

More interestingly, neither of the non-alternating

locative forms consists

of idiosyncratic exceptions The principal

difference betwe'en the

content-

object and container-object forms is which

argument gets mapped onto the

role of direct object It can be shown that in broad outline, all locative verbs

conform to a single principle governing the linking of semantic arguments

to

grammatical functions, which

we call the OBJECT AFFECTEDNESS LINKING RULE

An argument is encodable

as the direct object Of a verb if its i-e•re•t ••pe•ifie-d

as being affected in a specific way in the semantic

representation of the verb

According to this rule, the choice of direct object

is governed by whether

the content

or the container must undergo a specified change in order for the

verb"to apply The change, however,

can be a change either of physical

position or a change of state (which

may itself be mentally represented

as a

change of position in an abstract 'state space':

see Jackendoff, 1983) For

example, the meaning of pour specifies the particular way in which the

content is affected: a substance

must move in a cohesive stream, in contrast with, say, dripping or showering But pour does

not specify the change of state

of any container or surface to which the substance

moves one can pour water

into a glass, beside a glass, onto the ground, and so on The linking rule thus

118

LOCATIVE ACQUISITION

specifies that the content argument, but not the container argument, is

encodable as the direct object of pour Other verbs with a specified manner

of motion of a content without a specified change of state of a container are

also restricted to the content-object form, such as spill, drip and shake

In contrast, the meaning of fill specifies the particular way in which the

•container is affected- it undergoes a change of state from being not full to being full but it does not specify the particular manner in which the content

is affected one can fill a glass by pouring water into it, by dripping water into

it, or by dipping a glass into a bathtub The linking rule thus specifies that the container argument, but not the content argument, is encodable as the direct object of fill Similarly, verbs like cover, saturate and stop up specify only a change of state of

a container, and can only encode the container argument as direct object

Finally, the meaning of the verb stuff jointly constrains the particular

change of location that the content undergoes and the particular change of

state that the container undergoes In stuffing clothes into a hamper, for

instance, the clothing must be forced into the hamper (perhaps compressing

the clothing) BECAUSE the hamper is being filled.to a point where its remaining

capacity is too small,, or just barely big enough, relative to the amount of

clothing that is being forced in According to the linking rule, then, the direct object of stuff should be able to encode either the content or container argument, and this is what we find: stuff is an alternator Other alternators also specify actions that are defined in terms of changes simultaneously

specified in terms of content and container, such as brush or dab, where force

is applied pushing the content against the container, or load where a kind of

content dictated by the container enables the container to move or act in the designed way (e.g, a camera, or a gun)

The fact that the direct object always corresponds to the affected entity in

locative verbs manifests itself as a subtle semantic difference between the

versions For example, John loaded the cart with apples implies that the cart

is completely filled with apples, but John loaded the apples into the cart does

not This HOLISTIC INTERPRETATION (Anderson, 197'1) can be seen as a

consequence of the container-object form specifying a state change of a container rather than a location change of a content; the most natural interpretation of a state change is that it is the entire object that undergoes

the change (see Rappaport & Levin, 1985; Pinker, 1989)

The object affectedness rule may be quite general, applying not only to the

iocative alternation in English but to its counterparts in other languages

Both the same kinds of verbs that alternate in English, and the holistic

interpretation accompanying the container-object form, can be found in the

locative alternations of a variety of languages, including many that are

genetically and areally distinct from English (Moravcsik, 1978; Foley & Van

Valin, 1985; Rappaport & Levin, 1986; Gropen, 1989; Pinker, 1989)

119

LANGUAGE Furthermore, the principle can be seen to apply to constructions other than

locative forms across languages Verbs in which

an animate entity (an AGENT)

brings about a direct effect on another entity (a PATIENT), such as verbs of

causation of change of position (e.g causative slide) or state (e.g causative

melt), or verbs of ingestion (e.g eat), are almost invariably transitive

across

languages, with patients

as direct objects In contrast, verbs that fall outside this broad semantic class show

more variation within and across languages, with either argument appearing as direct object, such

as in verbs of emotion (fear vs frighten), or with prepositional objects expressing the non-agentive

argument, such as in verbs of perception (see vs look at) and in verbs of

physical contact without

a change in the contacted surface (hit vs hit at); see

Hopper & Thompson, •98o; Levin, i985; and Talmy,

•985 Furthermore,

the holism effect accompanies the interpretation of direct objects quite

generally: similar semantic shifts

can be seen in the difference between Kurt climbed the mountain and Kurt climbed up the mountain, only the first implying

that the entire mountain has been scaled, and in

a variety of other

constructions discussed in Green, •974; Moravcsik, •978; Hopper &]

Thompson, 980; Levin, i985 and Pinker,

•989

The fact that non-alternating verbs conform to a crosslinguistically

widespread linking pattern, rather than being

a list of idiosyncrasies, suggests

that children may actually use the linking rule in acquiring the locative

alternation,j and that it

may play a role both in the genesis of their errors and

in their recovery from them Though the linking rule appears to be near-

universal, the meanings of individual verbs clearly are not Therefore it is

possible that,mistakes in verb meaning, such as the specification Of which

entity is affected, might be the

source of the syntactic errors reported by

Bowerman (i982)" Specifically, if a child erroneously thought that a con-

tainer-object verb such asfill specified

some specific manner of motion of the

content-pouring, for example- he or she could derive a content-object

form from it using the linking rule, and would produce

errors like fill the

water

Previous research suggests that children are indeed slow in fixing the

standard adult meanings of .verbs compared with the meanings of

nouns (Gentner, •975; •982) In particular, they

have more difficulty acquiring

meaning components relevant

to changes of state than

components relevant

to changes of location (Gentner suggests that this

may be part of

a larger aattern whe"reby functional

components of word meaning are more difficult

:han perceptual/actional components.) For example, Gentner

contrasted the zerb mix, which she suggests specifies a particular change of state ('an

ncrease in homogeneity') but is noncommittal about the kind

of action that

;ffects it, with stir, beat and shake, which are noncommittal about the

esulting state, but which require particular

man•ers of motion Children ,ged five to nine and adults were asked to describe six kinds

of events and to

I20

LOCATIVE ACQUISITION

verify whether each of the four verbs

was appropriate to them: a stirring, beating, or shaking motion performed

on salt and water (which could 'mix ')

or on cream (which, already being a homogeneous substance, could not)

Verbs encoding manners of motion posed

no problem for the children

97 %

of the 5- to 7-year-olds and

93 % of the 7- to 9-year-olds paired the correct

manner-of-motion verb with the appropriate

manner of motion However,

the end-state requirement of mix

was poorly grasped: the 5- to 7-year-olds

used mix on 48 % of the trials where the substance was mixable and

on 46 %

of the trials where it was not

Note that this asymmetry in the acquisition of verb meaning

components, together with the object affectedness rule, offers

us an independently

motivated explanation for why the content-object sentence is produced

and overapplied more frequently than the container-object

sentence, and, unlike

the analogy with overregularization of irregular morphology, it does not

depend on questionable assumptions

about the frequencies of different locative forms in the input Another advantage to the linking theory over the

overregularization theory is that it provides a hypothesis about how the

acquisition, since children

are not corrected or even reliably mis-

comprehended when they make grammatical errors (Brown & Hanlon 97o),

so overgeneral rules are always logically compatible with the child's linguistic

experience (Braine, i97i Baker,

•979; Pinker, •984; •989) Though it has

been suggested that subtle statistical patterns in parental reactions might differentiate ungrammatical from grammatical sentences, it

unlikely that such feedback could be

of much use in unlearning locative

errors, since they are too infrequent to allow aggregation of significant

differences in parental feedback

types (e.g out of

a database of 22,303

utterances in the Brown corpus for Adam

on CHILDES (MacWhinney &

Snow, •985) Pinker (•989) found only

three clear syntactic errors (o'ooo• %) with locative verbs) Furthermore, the

errors occur far later than the ages at

which such feedback might occur (see Bowerman, I987; Pinker' •989)

The standard solution to the no-negative-evidence problem in the

case of

morphology is that irregular forms block the application Of the

regular rule,

so when the child hears broke and realizes that it is nothing more than the past

of break, he or she will avoid saying breaked (see Pinker, x984; Pinker & Prince, 988) But

no such blocking relation exists in the

case of the locative alternation: there is no kind of parental sentence that the child could take

as

reliably indicating that fill the

water is ungrammatical The non-occurrence

of such forms in parental speech cannot impel the child to reject them, because the child must be capable of extending the alternation to true

alternating verbs that happen

not to have been used in both forms in the input, such as, say, daub or spatter In contrast, the

account based

on the object linking rule does

suggest an account of how the syntactic errors are

unlearned Children can learn that a verb likefill does not require a particular

manner of motion of the content as soon as they hear the verb used in a clear

context in which no such motion takes place: for example, when a glass is

filled by means of bailing or dripping When the child processes such inputs

that falsify the erroneous manner component and he or she expunges that

component from the verb's semantic representation, the linking rule will no

longer map the content argument onto the direct object function, and the

errors will cease (see Gropen, 1989; Pinker, 1989)

In two experiments, we tested the hypothesis that children use the object

affectedness linking rule to predict the syntactic privileges of verbs, but that

they must learn what is and is not specified as affected in the semantic

representations of individual verbs First, using a task where children

describe pictures, we tried to confirm that children overgenerate locative

forms in sentences like The man is filling the water and The man is pouring the

glass with water, and to see whether we replicated the asymmetry in favour

of the former type of error, involving container-object verbs in content-

object constructions Second, using a two-alternative forced-choice task in

which children select pictures corresponding to locative verbs, we attempted

to determine whether children misinterpret the meanings of locative verbs,

perhaps thinking that (e.g,)fill specifies the particular manner in which a

substance changes location instead of the particular change of state that a

container undergoes Third, we tested whether syntactic and semantic errors

are associated across individual children- for example, whether children

who misinterpret the meaning of fill to specify the manner in which a

substance moves will be more inclined to express the content argument as the

direct object

EXPERIMENT

In this experiment we tested children's and adults' syntactic and semantic

knowledge of six common locative verbs: pour, fill, dump, empty, stuff and

splash The parts of the experiment dealing with the verbs stuff and splash are

not discussed here, because they are alternating verbs (so syntactic errors,

strictly speaking, are impossible) and because children and adults performed

similarly on them They are discussed in detail in Gropen (1989)

METHOD

Subjects

Sixty-four native speakers of English living in the Boston area participated:

16 children (1 • boys, 5 girls) aged 2; 6 to 3;5 (mean 3;1);16 (7 boys, 9 girls)

aged 3;6 to 4;5 (mean 3;11); i6 (8 boys, 8 girls) aged 4;6 to 5;•1 (mean

5;o); and 16 paid undergraduate and graduate students at MIT The

children were drawn from middle-class day-care and after-school programs Eight children were replaced in the design because of their unwillingness or inability to perform the syntactic task

Materials Twenty-five line drawings were created, each composed of two panels, like

a comic strip The first panel depicted the manner in which a substance changed location during the course of an action; the second depicted the endstate of a container as a result of the action One example is reproduced

in Fig I, where the first panel depicts a woman pouring water from a pitcher i-nto a glass and the second panel depicts an empty glass next to a puddle of water, showing that the woman had spilled the water Fig 2 displays another example: the first panel shows a woman turning on a tap, allowing water to drip from the spigot into a glass; the second panel shows a glass full of water

A subject who knows that the meaning of pour specifies the manner in which

a substance changes location should choose Fig over Fig 2 as the better

Fig x A picture of 'pouring-spilling' in Experiment

Fig z A picture of 'dripping-filling' in Experiment

123

example of 'pouring' Similarly, a subject who knows that the meaning offill

specifies the resulting endstate of the container should choose Fig z over

Fig as the better example of 'filling'

Of the z5 pictures, one depicted a boy hitting a ball with a bat in the first

panel, with the ball breaking a window in the second panel It was used to

ensure that the subjects understood that the drawings depicted two events or

states that were causally related The remaining z4 drawings were used to test

the subjects' semantic and syntactic knowledge of pour, fill, dump and empty

The semantic test was designed to assess what we shall call a BIAS in the

interpretation of a verb's meaning; subjects were forced to choose which of

two pictures, differing in manner and endstate, best represented the meaning

of the verb (e.g whether fill means pouring or fill means filling) Verbs and

drawings were chosen in pairs so that we could test what we thought would

be the most likely misinterpretations of the verbs Specifically, we thought

that a child might interpret fill to specify a pouring manner, pour to specify

a full endstate, empty to specify a dumping manner, and dump to specify an

empty endstate This also provides a built-in control: because the verbs in

these pairs are closely related in meaning, we were able to test subjects'

interpretation of both verbs of a pair using the same sets of pictures (across

subjects), ensuring that subjects' responses were not due to the salience of the

pictures themselves

Twelve pictures were shared for trials with the x•erbs pour andfill Of these

Z, subsets of three pictures depicted the same scenario- that is, the same

agent, container and content- with one picture ambiguous between pouring

and filling (e.g Fig 3), one depicting pouring but not filling (Fig i), and one

depictingfilling but not pouring (Fig z) Similarly, a set of iz pictures was

shared for the verbs dump and empty We used four different picture sets in

testing the meaning of each verb so that the idiosyncrasies of any one scenario

could not explain the results Table • lists descriptions of the pictures used

in the testing of pour/fill and dump/empty, organized by scenario Within

Fig 3- A picture of 'pouring-filling' in Experiment i

I24

TABLE 2 Picture sets used in the testing of 'pour'/'fill' and

'dump '/' empty'

PANEL (Manner) PANEL 2 (Endstate)

Scenario AI (Pour/Fill)

man pours water from bucket into sink

man pours water from bucket into sink

man drips water from tap into sink

Scenario Az (Pour/Fill) girl pours honey from bottle into bowl

girl drips honey from fork into bowl girl pours honey from bottle into bowl

Scenario BI (Pour/Fill)

boy pours paint from can into bucket

boy drips paint from brush into bucket

boy pours paint from can into bucket Scenario Bz (Pour/Fill)

woman pours water from pitcher into glass

woman pours water from pitcher into glass

woman drips water from tap into glass

Scenario CI (Dump/Empty)

man dumps ice cream from carton into bowl

man scoops ice cream from carton into bowl

man dumps ice cream from carton into bowl

Scenario Cz (Dump/Empty)

girl dumps 'Playdo'* from can onto table

girl dumps 'Playdo' from can onto table girl scoops 'Playdo' from can onto table Scenario DI (Dump/Empty)

woman dumps salad from bowl onto plate

woman dumps salad from bowl onto plate

woman scoops salad from bowl onto plate Scenario D2 (Dump/Empty)

boy dumps sand from pail onto towel

boy scoops sand from pail onto towel

boy dumps sand fro m pail onto towel

sink filled with water empty sink/spilled water sink filled with water bowl filled with honey

bowl filled with honey

empty bowl/spilled honey

bucket filled with paint bucket filled with paint

empty bucket/spilled paint

glass filled with water empty glass/spilled water

glass filled with water

empty carton/ice cream in bowl

empty carton/ice cream in bowl

non-empty carton/some ice cream in bowl

empty can/'Playdo' on table

non-empty can/some 'Playdo' on table

empty can/' Playdo' on table

empty bowl/salad on plate

non-empty bowl/some salad on plate empty bowl/salad on plate

empty pail/sand on towel

empty pail/sand on towel

non-empty pail/some sand on towel Note: each line corresponds to a drawing composed of one manner panel andone endstate panel For each subset of three drawings (e.g AI), the first drawing was displayed before the

remaining two; the second drawing was always displayed on the experimenter's right (the

child's left); the third drawing was always displayed on the experimenter's left (the child's

right)

'Playdo' is a registered trademark

subject, two subCsets (six pictures) were used for the testing of each verb;

across subjects in an age group, each subset was used equally often in the testing of either verb of a pair For the pour/fill sets, the manner distractor

alway• depicted dripping and the endstate distractor always depicted an

empty container (the contents were spilled); for the dump/empty sets, the

I25

manner distractor always depicted scooping and the endstate distractor

always depicted a nofi-empty container

Procedure

Adults were tested individually in a single session.; children were tested

individually in two half-hour sessions separated by several days Two

experimenters tested each child, one interacting with him or her and the

other recording responses

Before testing their knowledge of locative verbs, we introduced subjects to

the format of the pictures by presenting them with the drawing of a boy

hitting a ball with a bat and the ball breaking a •vindow Subjects were first

asked to describe each panel separately, and then both panels together If

subjects did not spontaneously use an appropriate causative verb (e.g break,

smash) in describing the complete drawing, the experimenter modelled the

sentence The boy is breaking the window

During the main body of the experimental session, we tested the Verbs

one

at a time For each verb we began with an ambiguous picture in order to

familiarize the subject with the conventions of the drawings (e.g the use of

shading and a waterline to indicate the fullness of the glass) For example, the

experimenter would show Fig 3 to a subject, and say 'Look at the first

picture: there's a woman, a pitcher, water, and a glass Look at the second

picture there's the glass and the water Now look at both pictures when the

woman does THIS (pointing to the first panel), it ends up like THAT (pointing

to the second panel) And it's called FILLING THIS (gesturing towards the

entire drawing) is FILLING' (In other conditions, the very same picture and

commentary would be used as an example of pouring.) The experimenter

would'then remove the first drawing and administer a forced-choice task

involving two non'ambiguous pictures (e.g Figs and 2); this served as the

test of verb semantics The constituents in each picture (i.e pair of panels)

would be introduced, as above, starting with the picture on the experi-

menter's right Neither of these pictures would be labelled as a depiction of

filling; instead, the experimenter would ask, 'Which of THESE (gesturing

towards both drawings) is FILLING .•' Notice that each picture preserves one

panel f rom the ambiguous picture and introduces a new panel, so any

systematic difference in response could not merely reflect global similarity

between the original picture and either of the forced-choice alternatives If a

subject did not clearly indicate either one picture or the other, the ex-

perimenter repeated the question The position of alternative pictures in the

forced choices was balanced within subject so that, for the two forced choices

per verb, a given type of picture (e.g 'pouring-spilling') appeared on the

right as often as it did on the left

Following this procedure, an elicitation task (the test of syntax) was

I26

administered for that verb The picture selected in the forced-choice task (whether it was correct or incorrect by adult standards) was presented to the subject, who was asked to describe what was happening in the picture We

were concerned that children may have an overall response bias to use the content-object or the container-object constructions, masking any tendency they may have to extend a verb to an incorrect form if it was the less-used

one Therefore we elicited sentences in two discourse contexts, one ap- propriate to each form For example, when the chosen picture corresponded

to Fig 2, the experimenter would make the container the topic in the following way: 'Point to the glass; say GLASS say FILLING What is the

woman doing to the GLASS ?' A natural response to this question is, 'She's filling it with water'-where the container argument is encoded as direct object In the same way, a content-topic query will set up a discourse context

favouring a locative response with the content argument as direct object (For

a discussion of the same methodology applied successfully to eliciting different dative forms, see Gropen, Pinker, Hollander, Goldberg & Wilson, I989) Furthermore, in those trials where a subject failed to produce a

sentence with an unambiguous direct object, we followed it up with a prompt: 'filling what', or 'filling ?' If the prompt also did not work, we asked (e.g.) 'Is the woman filling the glass or filling the water ?' (The order

of alternatives in the question was balanced across the two sets of pictures for

a given verb for a given child, and counterbalanced across subjects within each age group.) In this way, we scored three kinds of response •¢arying in

degree of spontaneity in the syntactic task We balanced the order of

query topic used in the testing of each verb across the subjects in an age-

group

Each cycle of semantic and syntactic tests was performed twice per verb,

the second time with a new set of three pictures The order of verbs itself was

b•lanced

across subjects in an age-group under the constraint that verbs with overlapping picture sets (e.g pour and fill) were never tested in consecutive order or within the same session In addition, the combination of verb testing

order, picture s•t combinations, and query order was counterbalanced across subjects within each age group

Scoring

In the semantic test, children's responses

were scored according to whether the chosen picture was consistent or inconsistent with the meaning of the verb as indicated by the tendencies of the adult control subjects In the syntactic task: responses were scored according to whether the direct object

of a locative construction corresponded to the content or the container in the described picture To count as a locative construction, a response had to specify both the appropriate verb and an unambiguous direct object The

I27

CHILD LANGUAGE

kind of question needed to elicit the locative ••esponse was also scored; that

is, whether the subject responded to the original elicitation, the subsequent

prompt (e.g Filling what '), or the forced-choice question (e.g Filling the

glass or filling the water ') Subjects often used pronouns in their responses

for example, in answering What is the woman doing to the glass ?' a subject

might respond, 'She's filling it' These utterances were counted only if the

referent of the pronoun was disambiguated by the presence of an oblique

object or particle (e.g She poured it into the glass or She poured it out), or the

referent could be pinned down via the subsequent prompt Responses which

were undecipherable or not clearly locative (e.g intransitive responses such

as She's pouring with the water) were excluded

Syntax

Table 3 shows the proportion of trials in which subjects produced content-

and container-object sentences with each verb in response to content-topic

and container-topic eliciting questions, Not surprisingly, adults used pour

and dump almost exclusively in content-object sentences (only two subjects

produced dump-container forms, one apiece), and they used fill almost

exclusively in container-object sentences (0nly one fill-content form was

produced by an adult) Empty, in contrast, is clearly an alternator- eight

adults used it in both locative forms, and four adults each used it in only one

or the other form

Children could deviate from adults' responses in two ways: they could use

fill in content-object sentences, pour in container-object sentences, or dump

in container-object sentences Of these potential errors, children were much

more likely to produce the nonstandard fill form than the other two: 3o

children out of 48 produced at least one fill-content form, whereas only two

children produced at least one pour-container form and only six children

produced at least one dump-container form The tendency to produce fill-

content forms declines with age the mean proportion of trials eliciting such

forms goes from o'53 in the young and middle groups to o'34 in the older

group aod o'o3 in adults, and this difference is significant in an ANOVA (F

(3, 6o) 6"63, p < o'ooi) A series of planned one-tailed t-tests shows that,

for each of the child groups compared separately to the adult group, there is

a significant difference in the proportion of fill-content responses (fro m

youngest to oldest, t (3o) 4"5o, P < o'ooi (3o) 4"5o, P < o'ooi t (3o)

3"oi, p < o'oo3) Although the oldest children produced fewer fill-content

forms (I utterances) than the younger children (i 7 utterances for each of the

younger groups), a post-hoc comparison reveals that this difference is not

significant (two-tailed t (46) I'48, p o'I4) As Table 3 shows, errors with

fill were not just responses to follow-up prompts The original request to describe the picture elicited many fill-content sentences, and the statistical comparisons remain significant when restricted to these responses

Since empty is an alternator, strictly speaking children cannot make errors with it However, Table 3 shows that the children had a consistent preference for the content-object form, which contrasts with adults, who used both forms equally often Across the child groups, the difference between the proportion of trials eliciting content-object sentences and container-object sentences (0"23) was significant (t (47) 2"06, p < 0"05)

Children were clearly sensitive to whether the original query focused on the container or on the content For the alternator empty, there was a preference to use content-object sentences in response to content-topic queries (mean difference o'41) and a preference to use container-object sentences in response to container-object queries (mean difference o'o6),

a significant difference (F (I, 60) lO'87, p < o'oo5) The effect can also be

seen with the non-alternating verb dump, for which container-object errors

were produced exclusively in response to container-topic questions (Fill-

content errors were produced equally often in response to content, and

container-topic questions.)

Semantics

In order to assess subjects' interpretations of particular verbs, we adopted the following criterion" if a subject chose the same type of picture (manner or endstate) on both semantic tests for a given verb, then he or she was considered biased towards that aspect of the verb's meaning Table 4 summarizes the number of biased subjects in each age group Totals that are

significantly greater than chance (o'5 x o'5 x 16 4/16 subjects) at p < o'o5

accbrding to the binomial distribution are underlined

The data reveal markedly different performance for pour and dump versus fill and empty •As expected, adults unanimously treated pour and dump as having more todo with a manner of motion than with a change of state, and fill and empty as having more to do with a change of state than a manner of motion The number of children who were biased towards the adult

interpretations of pour and dump is significantly higher than chance for every

age group (p < o'o5) In contrast, forfill and empty we fail to find a significant number for children biased towards the change of state interpretations, and instead find that some groups of children were biased towards INCORRECT meanings for the verbs Specifically, in the semantic test for fill eight of the oldest children (out of 16)consistently chose pictures showing a pouring

manner without a full endstate (p < o,o3) In addition, in every child-group

more children than would be expected by chance (though not significantly so) showed a bias towards choosing incorrect pictures for empty (i.e a dumping

JCL 18

129

CHILD LANGUAGE

T A B L E 3 Experiment Proportion of trials in which subjects used verbs in

content-object and container-object sentences

AGE-GROUP VERB-FORM 2 6 3 5 3 6 4 5 4 6 5 Adult

Pour

Content-object sentences

1"oo 1.oo

(I8/IO/I) (29/3/0) (29/3/0) Container-object sentences

(11210) Content-object sentences

(I9/IO/I) (27121 ° (26/4/I) Container-object sentences

Container-topic query o" z o" 19 o'o6 o" 12

(21010) (I/I/I) ,(I/O/O) Fill

Content-object sentences

Content-topic query

Container,topic query

Mean

Container-object sentences

Content-topic query

Container-topic query

Mean

Empty

Content-object sentences

Content-topic query

Container-topic query

Mean

Container-object sentences

Content-topic query

Container-topic query

Mean

0"50 0"44 0"50 o'oo 0"53 0"53 0"34 0"03

0"44 0"31 0"75 0"94 0"50 0"56 0"44 0"94

(4/10/1) (IO/I/3) (9/9/1)

0"69 0"69 0"69 0"75 0"50 0"56 0"50 0"25 0"59 0"62 0"59 0"50 (6/8/5) (12/7/I) (I3/6/o)

0"31 0"31 o" 31 0"2 5 0"44 0"38 0.50 0"75 0"38 0"34 o'41 0.50

(5/5/2) (4/4/3) (I I/!/i

130

TABLE 4" Experiment i" Number of subjects biased towards manner or

endstate interpretations of verbs

AGE GROUP

Combined

6-3 5 3 6-4; 5 4; 6-5 11 children Adult

(N 16) (N 16) (N 16) (N 16) (N 16)

Manner bias

Endstate bias

Note: Underlined frequencies are significantly greater than chance, at p < 0"05, according to

a binomial test

manner without an empty endstate) altogether i9 out of the 48 children did

so-a significant proportion (p < o'o2, binomial test)

These data suggest that a child may be more likely to think that the meaning of fill involves pouring than that it involves something being made full; likewise that empty involves dumping rather than that it involves- something being made emptyl Presumably this is due to two factors The first is a bias toward picking up manner components of meaning, as suggested

by Gentner (I978) This overall bias can be demonstrated by aggregating data over the set of four verbs, which was balanced for manner-oriented and

endstate-oriented meanings, and comparing the difference between the proportion of manner and endstate responses for adults and children Adults showed no overall bias (the mean difference between manner and endstate

responses was o'I2, not significantly different from zero), whereas children did: their difference scores were all significantly different from zero (young

0"28, (I5) ='3"2o, p < o"oi mid 0"39, (I5) 4"28, p < o'ooi old 0"52, (I5) 5"26, p < o"ooi) The second factor is the statistical pattern

of cause-and-effect sequences in the world, whereby pouring is a typical

means of effecting fullness and dumping is a typical means of effecting

Note: The numerals in parentheses in Table 3 correspond to the frequencies of sentences used

in response to the ,original question, the subsequent prompt, and the forced-choice question, respectively Adults always responded to the original question For the trials on a given verb performed by each group, total proportions may be less than l'OO and total frequencies may

be less than 32 because of discarded responses

LANGUAGE

emptiness Adults know that these causal connections

are facts about the world, not facts about the semantics of fill

or empty

Contingencies between semantic and

syntactic errors

We have suggested that children's syntax

errors are the product of correct

linking rules operating on incorrect semantic representations Consistent

with this suggestion is the finding that the kinds of verbs that children

use in

incorrect syntactic frames in spontaneous speech and in the syntax test (fill

and empty) are the same verbs that children misinterpreted in the semantics

test, with just the kind of bias (towards manner) that would result in incorrect

content-object sentences In contrast,

pour and dump were rarely the

sources

of either syntactic or semantic

errors

A much more stringent test

can b6 conducted by seeing if the tendency towards semantic and syntactic errors is correlated across individual children

For each age group and verb

we constructed a 2 × 2 contingency table, each child contributing one count, classifying children according to their semantic

bias and their syntactic errors For fill, each child was classified

as either biased towards the pouring manner

or biased towards the full endstate based

on performance in the semantics test, and

was independently classified either

as having produced at least

one content-object form among his or her transitive sentences with fill, or as having produced no such errors, in the

syntax test (One child who produced no transitive

sentences with fill was

excluded.) This analysis is shown in Table 5 Only the oldest children

(4;6-5;x•) showed

any tendency toward a contingency in the predicted

direction (i.e being more likely to produce fill-content

errors if they were TABLE 5" Experiment

I" Contingency between bias in the interpretation of 'fill' and grammatical

errors in sentences with 'fill'

Produced at least

one

content-object sentence Produced no

content-object sentences

2;6-3;5:

Manner biased

l•ndstat•e biased

4

3;6-4;5:

Manner biased

Endstate biased

6

Manner biased

5

Endstate biased

Combined children

Manner biased

8

Endstate biased

132

LOCATIVE ACQUISITION

TABLE 6 Experiment i" Contingency between bias in the interpretation of

empty' and kinds of sentences used with empty'

Produced only content-object sentences Produced only

container-object sentences

z;6-3;5:

Manner biased Endstate biased

3;6-4;5:

Manner biased Endstate biased

4;6-5;I•:

Manner biased Endstate biased Combined children Manner biased Endstate biased

3

6

o

o

I4

biased toward a manner interpretation forfi//), but this contingency was only marginally significant in a Z test (Z" (I) 3"o9; p < o-o8) and not significant

by a Fisher Exact Test (one-tailed p o-i2)

Table 6 displays a similar test that was conducted for performance with

empty Because both forms are grammatical for adults, for the syntactic dimension each child is classified in terms of whether he or she produced only

empty-content forms or only empty-container forms Here the contingency is the right direction for all three age groups, and is significant for the combined

group of children according to a Fisher Exact test (one tailed p < o'o2): a child who is biased toward a manner interpretation for empty will tend to

produce more empty-content sentences than empty-container

sentences whereas a child who is biased towards

an endstate interpretation will tend to

produce more container-object sentences

The contingencies between semantic biases and syntactic forms are encouraging, if somewhat less than conclusive, particularly for fi// Apart from the fact that our samples were

very small (two responses per measure

per child) in comparison to the noisiness of children's behaviour, there are

present in children's grammars First, even if children do assign irrelevant

manner features to a verb's meaning, there is no guarantee that we tested for the exact

plausible, candidate but

one alternative is that children think that filling requires the top surface of a substance to move higher and higher during the

course of the action; such a manner of motion would in fact be compatible

with the panel depicting the full-container endstate Second, subjects were

I33