Syntactic Cues in the Acquisition of Collective Nouns

Syntactic cues in the acquisition of collective nouns Abstract One basic finding in the study of word learning is that children tend to construe а word describlng an object as referring to the kind of whole object, rather than to а part of the object, one of its properties, or the substance it is made of. This has been taken as evidence that there exist certain special constraints оп word meaning that guide children to favor the kind-of-object interpretation when exposed to а new word. There are descriptive proЫems with this proposal, however, as it cannot explain how children leam other kinds of words, such as names for specific people, substances, parts, events, collections, and periods of time. These proЫems motivate an altemative theory in which young children possess several distinct conceptual categories - including "individual", which is more abstract than "whole object" and сап use syntactic cues to determine the conceptual category that а new word belongs to. This theory is explored in two experiments in which we attempt to use syntactic cues to teach children and adults novel collective nouns - words that refer to groups of objects. Тhе results indicate that children сап use such cues to leam names for kinds of individuals that are not whole objects, although they are less аЫе to do so than adults. Candidate explanations for why this developmental difference exists are discussed and implications are drawn for theories of word leaming and conceptual representation. 1. Introduction This article concerns the acquisition of names for material entities, such as objects (dog), substances (water), specific people (Jane), and collections of objects (family). Most research into word learning has focused оп names for object kinds, and such names do appear to have а privileged status in lexical development. Evidence from both experiments and analyses of vocabulary development suggest that when children and adults are exposed to а novel word that refers to an unfamiliar object, they tend to interpret the word as referring to that kind of whole object - and not as naming а part of the object, а property of the object, or the stuff that the object is made of (e.g., Baldwin, 1989; Macnamara, 1982; Markman & Wachtel, 1988; Soja, Carey & Spelke, 1991; Taylor & Gelman, 1988; Waxman & Hall, 1993). One explanation for this phenomenon is that children and adults possess certain special constraints on word learning. In particular, а "whole object assumption" (Ьiasing in favor of the whole object interpretation) and а "taxonomic assumption" (Ьiasing in favor of the "kind" interpretation) have been argued to apply even in children under 2 years of age (see Markman, 1990 for review). Such assumptions are specific to the process of word learning and not the result of more general principles of sorting or categorization. In support of this, Markman and Hutchinson (1984) found that when hearing an object (such as а dog) described as "а dax" and asked to "find another dax", children tend to assume that the word refers to that kind of object and will point to another dog. But when simply told "See this. Find another one", where no novel word is provided, children prefer to point to something that has а thematic relationship to the dog, such as а bone. This suggests that the Ьias to favor kinds of objects is limited to the domain of word learning. Most of the words possessed Ьу adults - and Ьу 2-year-olds - are not names for kinds of objects. То account for this, Markman and Wachtel ( 1988) posit а further constraint stipulating that distinct words cannot have overlapping extensions ("mutual exclusivity"). This would Ье false as an absolute property of natural language (consider dog and animal, for instance). But it could nevertheless exist as а Ьias that sometimes overrides the whole object and taxonomic constraints. In particular, once а child had acquired а name for а given kind of object (e.g., dog), mutual exclusivity would Ьias her to infer that other words that describe members of that kind are not names for the object kind, and might instead refer to properties, parts, and so on. There is Ьу now abundant evidence in support of the special constraint view. Not only are children Ьiased to favor the kind-of-object interpretation when exposed to а word that describes an unfamiliar object, but this Ьias can Ье overridden if the object already has its own name (Clark, 1987; Markman & Wachtel, 1988; Merriman & Bowman, 1989). The theory suffers from certain limitations, however, and some of these are discussed below (see also Bloom, 1994а). The remainder of this article will defend а different theory of children's word learning capacities, which is that they possess а set of conceptual categories (including а notion of "individual" that is more abstract than "whole object") and сап use grammatical cues to determine which of these conceptual categories the meaning of а new word corresponds to. This proposal diverges most radically from the special constraint theory with regard to how children learn words other than names for object kinds. То explore this, we report the results of two studies in which we attempt (with mixed success) to teach children and adults novel collective nouns. In the General Discussion, we discuss some implications of these results for theories of word learning and conceptual representation. 1.1. ProЫems with the special cons traint theory of word learning We should stress at the outset that we have no objections to the notion of constraints on word learning. Acquiring а word is а proЫem of induction and induction can succeed only if the learner possesses some а priori constraints on the space of possiЫe hypotheses (see Fodor, 1975; Goodman, 1983). In particular, there are an infinite number of possiЫe meanings consistent with any exposure or set of exposures to а new word (Quine, 1960), and so children must possess constraints (or Ьiases) leading them to favor some of these meanings over others. The dispute is over the precise nature of these constraints. In particular, а theory that assumes that object names have а unique status in word learning - and that words which refer to other categories can Ье acquired only as second-choices, after an object-kind label has already been learned - makes some proЫematic predictions about children's lex1cons. In а study of children's early productive use of words, Nelson, Hampson, and Shaw (1993) examined the speech of 45 children, who were an average age of 1 year and 8 months. They found that even at this age, nominals were not the only categories used; for example, children also used words denoting actions (eat) and properties (nice), which, at least for adults, correspond to verbs and adjectives. Nelson et al. also found that even when they restricted the analysis to nominals, only about half referred to basic-level object kinds. The rest referred to other conceptual categories, such as specific people (Fred), events (party), and temporal entities (day). Other studies looking at younger children's very first words (e.g., Bates, Bretherton, & Snyder, 1988; Bloom, 1990; Nelson, 1973), have obtained similar results.1 1 One must Ье cautious when interpreting data from spontaneous speech, since children need not know the сопесt meaning of а word to use it appropriately. It is quite likely, for instance, that children's understanding of word like пар might differ in subtle ways from that of adults. But it is wildly improbaЫe that all of their usages of such words are entirely confused, such that they think that they are all names for object kinds. We know of по reported cases where children were observed to treat words like пар and spanking as referring to kinds of objects, such that пар meant "bed" or spanking meant "hand". Similarly, there are no cases where children have been reported as treating pronouns and proper names like she and Fred as names for specific kinds. The only relevant eпors that do occur are initial confusions with тотту and daddy, and the status of these errors is unclear, as such words are sometimes used as common nouns Ьу adults (Macnamara, 1982). The early acquisition of these words poses two distinct proЫems for the special constraint theory. First, some of them should not have been acquired if children possess the posited constraints. In particular, pronouns and proper names violate the taxonomic constraint; they are used to refer to objects, but refer to specific individuals rather than kinds of objects. Nevertheless, pronouns and proper names are used before children show evidence of having learned any word that refers to the kind that these specific individuals belong to (e.g., person, тап, or woman), suggesting that their acquisition cannot Ье explained through mutual exclusivity. То put it another way, nominals like she and Fred do not appear to Ье "second­ choices", acquired only after the object-kind name that refers to the same entity. Instead they are appropriately used very early in lexical acquisition. (It is less clear whether children are also сараЫе of learning names for kinds of body parts - like finger and еуе - before learning words like person; if they are, this would pose а similar proЫem for the whole object con­ straint.) Second, many of children's nominals fall outside the purview of the proposed constraints. Substance names like water, names for events like spanking, and temporal terms like minute are not labels for objects in the first place and so the constraints cannot explain their acquisition. Even for the sorts of cases that mutual exclusivity works best for (names for artifact parts like handle and names for solid substances like wood, words which really do seem to Ье second-choices, learned only after а relevant object name has been acquired), the constraint does not explain how such words are acquired: it just provides an explanation for why they are learnaЫe in the first place, given the existence of the whole object and taxonomic assumptions. Тhis failure to explain how words that do not refer to object kinds are acquired is а serious gap, as the conceptual arguments for the necessity of constraints in the case of object names apply just as well for the acquisition of words like water and handle. It would Ье preferaЫe for а theory of lexical development to account for the early acquisition of all names, not just names of object kinds. А final concern is that the motivation for positing constraints like the whole object assumption is their role in solving the word learning proЫem. But although it is conceivaЫe that the constraints that facilitate lexical learning have evolved solely for this purpose, it would Ье more parsimoni­ ous to explain children's success at word learning through more general properties of language and cognition. The argument below is that the Ьias to favor the kind-of-whole-object interpretation in certain situations exists only as а by-product of more fundamental psychological processes (e.g., Bloom, 1994а, 1994Ь; see also Markman, 1992). ln particular, we argue here that the Ьias to favor the kind-of-object interpretation is а special case of two already-present aspects of children's knowledge: а specific mapping between count nouns and kinds of individuals, and а predisposition to construe objects as individuals. 1.2. Ап alternative theory of word learning The theory explored here is defended in some detail in Bloom (in press), where it is placed in the context of "rationalist" perspectives on word learning. It is based on prior work Ьу Brown (1957), Carey (1988), Gleitman (1990), and Macnamara (1982), and has three main premises. The first is that children's concepts are both diverse and abstract. They are diverse in the sense that they belong to different ontological categories, including individuals, kinds of individuals, properties, events, and spatial relations. They are abstract in the sense that they do not reduce to sensory or perceptual categories. This is in sharp contrast to the premise, held in different forms Ьу constructivists such as Piaget, empiricists such as Hume, and some modern-day connectionists, that the concepts possessed Ьу infants and young children are initially arrays of perceptual features, and that only as а result of further experience - including language learning - do more diverse and abstract categories emerge (see also Quine, 1960). The empiricist position is arguaЫy untenaЫe, in part because of the abundant experimental evidence from infant cognition research which suggests that infants possess distinct classes of abstract conceptual categories (e.g., Spelke, Breinlinger, Macomber, & Jacobson, 1992; Wynn, 1992). For instance, infants' understanding of number appears to Ье abstract, with 6-month-old children сараЫе of matching the number of objects they are shown with the number of sounds that they hear (Starkey, Spelke, & Gelman, 1990). As noted above, similar findings emerge from studies of language acquisition; even 2-year-olds seem to have по special difficulty acquiring terms that refer to numbers, like two, temporal individuals, like minute, moral properties, like good, and social constructs, like family. А second premise, first proposed Ьу Brown (1957), is that these con­ ceptual categories correspond to specific syntactic categories. After children have some grasp of language-particular properties of grammar, they сап apply these syntax-semantics mappings in the course of word learning. In other words, they сап use the syntactic category of а word as а cue to what it must mean. Evidence is reviewed below that at least some of these mappings сап Ье used Ьу 2- and 3-year-olds when learning the meanings of nouns. The third premise concerns а different sort of cue to word meaning. Many traditional theories of word learning posit that children map words onto concepts through а sensitivity to spatio-temporal contiguity. The child is looking at а dog, hears "dog", and associates the sound of the word with her perceptual experience at the time the word is heard. Given the circum­ stances under which children are exposed to words, however, this mecha­ nism is unlikely to work, as the required correlations between words and objects do not appear to exist (e.g., Bloom, in press; Chomsky, 1959; Landau & Gleitman, 1985). Even in cases where the connection is made explicit, as when an adult points to an object in the presence of the child and says its name, children do not learn the word through the principles of association. Instead, they determine what the adult is looking at and assume that this is the referent - even if they themselves were looking at another object at the moment the word was used (Baldwin, 1991). This suggests that word learning is fundamentally an interpretive process, related to the child's capacity to infer the linguistic intentions of others. The child learns that "dog" means dog because she has the capacity to infer that adults intend to refer to dogs when they use the word (see also Macnamara, 1982). The discussion below directly focuses on the first and second aspects of this proposal and how they differ from the special constraint view. Like the constraint theory, the view advanced here focuses on the acquisition of nominals. One specific account of the syntax-semantics mappings within the domain of English nominals is as follows (see also Bach, 1986; Bloom, 1994Ь; Jackendoff, 1991; Langacker, 1987; Macnamara, 1986). There are three mappings of interest. One is between count nouns (words that сап follow determiners such as а and тапу, and сап Ье pluralized) and the category kind of individual, another is between mass nouns (words that сап follow determiners like some and much but which cannot Ье pluralized and cannot follow determiners such as а and тапу) and the category of kind of non-individuated entity (or, more succinctly, kind of stuff), and а third is between lexical NPs (words which cannot Ье quantified at all) and the category of individual.2 This is summarized in ТаЫе 1. Such mappings are presumed to underlie knowledge of language in ТаЫе 1 Proposed syntax-semantics mappings within the domain of nominals Syntax Semantics Count nouns Kinds of individuals -They appear within Noun Phrases (NPs) that сап refer to а single individual (а dog) or to multiple individuals (тапу dogs) Mass nouns Kinds of non-individuated entities (or "stuff'') -They appear within NPs that сап refer to portions of stuff (some water, not much advice) Lexical noun phrases Individuals -They сап refer to specific individuals or sets of individuals, as with proper names (Fred) and pronouns (she, they) One exception to this third mapping is the expletive it (as in lt is raining) which has по semantic content at all. We know little about how children acquire such semantically empty NPs and, in particular, whether this acquisition is in some sense parasitic оп children's understand­ ing of their referential meanings (for some speculations, see Bloom, under review; Nishigauchi & Roeper, 1987). 2 general; their role is not restricted to language acquisition. It is the syntactic status of hands, water, and he that makes cold hands and cold water, Ьиt not cold he, ассерtаЫе, just as it is the syntax that makes too much water, Ьиt not too much hands, ассерtаЫе. And it is the mappings between syntax and semantics that govern the interpretation of words within these syntactic contexts. For instance, speakers of English understand the difference between John ate chicken versus John ate а chicken through an understand­ ing of the semantic basis of count-mass syntax. Soja (1994) has argued that syntactic knowledge of this sort also underlies our understanding of the subtle meaning difference between John went to the school (school as а count noun - referring to the kind of place) versus John went to school (school as а lexical NP- referring to а specific haЬitual activity), and that such knowledge can Ье applied productively Ьу adults and preschoolers. One aspect of this proposal is that the category "individual" cannot Ье reduced to the category of "whole object". There exist individuals that are not whole objects, such as puddles (bounded substances), tails (parts), nightmares (cognitive events), and flocks (collections of objects). From the standpoint of grammar, such words are equivalent to object names; there is no difference between а count noun like dog and а count noun like nightmare. In general, the distinction between count nouns and mass nouns, at least for adults, is an abstract quantificational one, and does not reduce to words referring to kinds of whole objects versus words referring to kinds of non-solid substances. In addition, however, there is evidence that whole objects are extremely salient individuals. When children and adults are exposed to а scene, they will initially tend to construe the objects- but not collections of objects or parts of objects, for instance- as individuals. Shipley and Shepperson (1990) posit that children possess а "Discrete Physical Object [DPO] Ьias" that shows up in non-linguistic tasks. When children are shown an array of objects and explicitly asked to count the colors, the parts, or the kinds, they show а strong tendency to ignore the instructions and count the objects. That is, when given а task in which they have to quantify over individuals, they choose objects as the individuals. It is not that children are limited to treating objects as individuals when asked to count- on the contrary, they can count sounds and actions just as early as they can count objects (Wynn, 1990). But when they are faced with а display of material entities, they find it difficult to override the DPO Ьias and focus on non-object individuals, even if they are explicitly asked to do so. The existence of the mappings shown in ТаЫе 1, when comЬined with the findings of Shipley and Shepperson, suggest that we may not require special constraints to explain the developmental phenomena. Instead, the Ьias to construe novel words as naming kinds of whole objects emerges from: (1) the cognitive Ьias to view whole objects as the most salient individuals (Shipley & Shepperson, 1990); and (2) the knowledge that count nouns refer to kinds of individuals. If а child is shown а novel object and hears а count noun used to describe it (e.g., This is а wug), these factors will conspire to guide her to construe the noun as referring to the kind of object. lf there is по syntactic support, the object will still Ье the favored interpretation, since it is а salient individual (through the DPO Ьias) and in most cases there will Ье no equivalently salient category that the word could refer to. In contrast, if there is count noun syntax but по objects for the DPO Ьias to apply to, the child will shift to an interpretation consistent with another type of individual-presumaЫy this is how children learn event names like пар, temporal expressions like minute, and names for bounded substances like puddle (see Soja, 1992). The other mappings in ТаЫе 1 сап explain the acquisition of names for non-solid substances and for specific individuals. Since the referents of words like water and milk are not individuataЫe objects, the DPO Ьias will not apply. In addition, such words are mass nouns and thus the mapping between syntax and semantics will guide children to assume that they refer to kinds of stuff, not kinds of individuals.3 Children will hear pronouns and proper names used to refer to objects, such as people, and so the DPO Ьias will apply, directing children to focus оп these objects as individuals. This precludes а mass noun interpretation of these words. Since they are presented with lexical NP syntax, as in This is Fred, children will interpret them as referring to the individuals themselves, not to the kinds of individuals (see also Hall, 1994, for discussion of semantic considerations that favor the individual interpretation over the kind-of-individual interpretation). Even 1- and 2-year-olds syntactically distinguish count nouns and mass nouns from lexical NPs in their sponta­ neous speech, suggesting that they possess enough knowledge of syntax for this procedure to work (Bloom, 1990). Are young children actually аЫе to use syntactic cues to word meaning? This question was originally addressed Ьу Brown (1957), who showed preschoolers а picture of а novel action performed оп а novel substance with а novel object. Children were told either "Do you know what it means to sib? In this picture, you сап see sibblng" (verb syntax); "Do you know what а sib is? In this picture, you сап see а sib" (count noun syntax); or "Have you seen any sib? In this picture, you сап see sib" (mass noun syntax). They were then shown three pictures, one depicting the same action, another 3 For solid-substance names, the cues conflict. The syntax motivates the kind-of-stuff interpretation (since words like wood and metal are mass nouns) but the DPO blas focuses children оп the individual objects, not the stuff they are made of. Тhе existence· of this conflict suggests that these words should Ье hard for children to leam, and this appears to Ье the case: children tend to misconstrue а novel mass noun describlng an object (e.g., This is wood used to describe а Ыосk of wood) as а name for the kind of object, not as the kind of solid-substance (Dickinson, 1988; Markman & Wachtel, 1988; Prasada, 1993; Soja, 1992). depicting the same object, and the third depicting the same substance. They were asked to "show me another picture of sibЬing" (or "another picture of а sib", or "another picture of sib"). Brown found that the preschoolers tended to construe the verb as referring to the action, the count noun as referring to the object, and the mass noun as referring to the substance, suggesting that syntactic cues help in the course of word learning. Further empirical work has extended Brown's research in the domain of count-mass syntax. Soja (1992) found that once 2-year-olds show productive command of the grammatical contrast between count nouns and mass nouns, their knowledge of syntax-semantics mappings allows them to use syntactic information to infer whether а word refers to а non-solid substance (such as water) or to а bounded individual composed of that non-solid substance (such as puddle). In another study (Bloom, 1994Ь), 3- and 4-year-olds were shown а pile of small objects, of а size such that one could name the array with either а plural count noun (such as peas) or а mass noun (such as rice). When given а plural count noun (some sibs) to describe the array, preschoolers interpreted the noun as an object name; when given а mass noun (some sib), they interpreted it as referring to the stuff. Similarly, preschoolers were also found to construe а plural count noun descriЬing а series of sounds as referring to the individual sounds, and а mass noun descriЬing the same series as referring to the undifferentiated noise (Bloom, 1994Ь). These findings show that sensitivity to the relation between count-mass syntax and word meaning shows up early in language development and that this distinction does not semantically reduce to objects versus non-solid substances - count noun syntax can сие children to abstract individuals, such as bounded substances (Soja, 1992) and individual sounds (Bloom, 1994Ь). А sensitivity to the semantic difference between nouns and lexical NPs appears to emerge even earlier in language development. Katz, Baker, and Macnamara (1974) found that 17-month-olds attended to the difference between count noun syntax (This is а sib) and Noun Phrase syntax (This is sib) when determining whether а novel word was а name for а kind of object or а proper name (see also Gelman & Taylor, 1984; Hall, 1991; Littschwager & Markman, 1993). There are similar findings in linguistic domains other than nominals. Unlike nouns, adjectives appear to draw children's attention towards properties (Gelman & Markman, 1985; Taylor & Gelman, 1988; Hall, Waxman, & Hurwitz, 1993; Waxman, 1990), while prepositions draw children's attention to spatial relations (Landau & Stecker, 1990). The contrast between transitive and intransitive verbs can lead 2-year-olds to infer the number of participants in the event denoted Ьу the verb (Naigles, 1990), and Gleitman (1990) and Landau and Gleitman (1985) argue at length that sensitivity to syntactic cues - а process described as "syntactic bootstrapping" - plays а crucial role in the course of the acquisition of verb meanings (see Pinker, 1994, for discussion). 1.3. Motivation for the current study The evidence reviewed above suggests that children can use syntax to learn names for specific individuals (such as Fred), to learn names for kinds of individuals that are not whole objects (such as puddle), and to acquire adjectives, mass nouns, and prepositions, which correspond to semantic categories other than individuals or kinds of individuals. But it is unclear whether children can learn а name for а kind of individual which is more abstract than whole object in an instance where the whole object interpreta­ tion is conceptually availaЫe. In other words, it is an open question whether they can hear а novel count noun applied to an object or set of objects, and interpret the noun as referring to а kind of individual - but not construe the whole objects as the relevant individuals. This scenario serves to directly contrast the theory proposed here (which predicts this sort of acquisition to Ье possiЫe) with the special constraint theory (which predicts that it should not Ье possiЫe). lt is also directly relevant for certain types of learning that have rarely been explored, such as the acquisition of collective nouns. Collective nouns are those count nouns that in the singular form refer to groups of objects, not single objects. For instance, а family refers to several people, and not to а single person, and а ftock refers to several Ьirds, not а single Ьird. Some of these nouns, like army and family, are learned relatively early (Callanan & Markman, 1982), while others such as commit­ tee and school (of fish) are acquired later. Collective nouns are interesting for several reasons. As non-object names that refer to material individuals, they provide an intriguing opportunity for the empirical study of the semantic nature of "individual". What is it about the mental representation of а flock, for example, that allows us to construe it as an individual? Why are other collections of objects, such as the collection of leaves on а given tree (Chomsky & Walker, 1978), not conceptualized as individuals? These issues are briefly explored in the General Discussion. More pertinent to the issue here, however, is that different theories of lexical development make contrasting predictions about how such nouns are acquired. А proponent of the special constraint theory might argue that the only way а child is going to learn а word like army is Ьу first using the constraints to learn the object name soldier. Only then can she construe а word that refers to а group of soldiers as referring to the whole collection the army - because only then does mutual exclusivity make this interpreta­ tion availaЫe. This hypothesis gains support from the observation that the collective nouns that children know describe objects that have their own names - forests are groups of trees, flocks are groups of Ьirds, armies are groups of soldiers, and so on. There also exist "group nouns" (Jackendoff, 1977) that do not specify the precise kinds of entities that they are collections of, such as group, pile, and stack, but these too might Ье acquired only in cases where the object names are already known. Perhaps children are only аЫе to learn the word pile if it is used to describe а pile of objects that they already have а name for, as in the expression а pile of socks. А theory that assumes that children possess а mapping between the grammatical category count поип and the semantic category kind of individual makes а different prediction, which is that syntactic cues can lead the child to the collective interpretation regardless of whether or not she knows the names of the objects that make up а collection. This prediction motivates the experiment that follows. EXPERIMENT 1 When people are shown а single unfamiliar object described Ьу а singular count noun (This is а fendle), they will tend to interpret the word as naming the kind of object; if they see а group of unfamiliar objects and hear а plural count noun (These are fendles), they will again treat the word as naming the kind of whole object. Тhе special constraint theory and the theory proposed here explain this phenomenon differently. The special constraint theory posits that the whole object and taxonomic assumptions favor the object kind interpretation regardless of the syntactic cues. The alternative explana­ tion is that people know that an NP with а singular count noun refers to one individual and an NP with а plural count noun refers to more than one individual - and in each case, the DPO Ьias makes the objects the most salient individuals. Consider now а case in which а speaker is referring to many objects but the syntax is consistent with that of reference to а single individual. The special constraint theory still predicts а Ьias towards the whole object interpretation. In contrast, we suggest that syntax will guide children and adults to construe the word as descriЬing а single individual, and since the context indicates that the speaker intends to describe all of the objects, the word will Ье construed as descriЬing an individual that contains all of the objects, that is, as а collective noun. То explore this, subjects were shown pictures of groups of objects. These were described with singular count noun or plural count noun syntax. We predict that the syntactic context will draw children and adults towards different interpretations of the novel word. А second manipulation concerns the nature of the collections. In half of the pictures of groups, the objects were next to one another but not connected (the unconnected groups). In the other pictures, they were depicted as being connected Ьу strings (the connected groups). А tentative prediction is that when the objects are shown as connected, the set of objects will make for а more salient individual, appearing more "coherent" in the eyes of the subject even though it clearly is an array of different objects. As а result of this, one might expect there to Ье more collective interpretations in this condition.