Productivity and constraints in the acquisition of the passive.

The acquisition of the passive in English poses a learnability problem. Most transitive verbs have passive forms (e.g., k i c k / w a s k i c k e d b y ) , tempting the child to form a productive rule of passivization deriving passive participles from active forms. However, some verbs cannot be passivized (e.g. c o s t / * w a s c o s t b y ) . Given that children do not receive negative evidence telling them which strings are ungrammatical, what prevents them from overgeneralizing a productive passive rule to the exceptional verbs (or if they do incorrectly passivize such verbs, how do they recover)? One possible solution is that children are conservative: they only generate passives for those verbs that they have heard in passive sentences in the input. We show that this proposal is incorrect: *Portions of this paper were presented at the Sixth (1981), Ninth (1984), and Tenth (1985) Annual Boston University Conferences on Language Development. This research was supported by NSF grants BNS 80-24337, 82-16546, 82-09450, by NIH grant RO1 HD 18381, by MIT's Undergraduate Research Opportunities Program, and by a grant to the MIT Center for Cognitive Science from the Alfred P. Sloan Foundation. We are grateful to Melissa Bowerman for her extensive and incisive critical comments on an earlier draft, and to T. John Rosen for his generous advice about the statistical analyses. We also thank Rosen and Karin Stromswold for their comments on an earlier draft and Brian MacWhinney for providing computer-readable transcripts under the auspices of the ChiLDES project. Finally, we thank the staff of the participating child care centers and schools for their cooperation: Allston/Brighton YMCA, Cambridge Montessori School, Cambridge YMCA, CEOC After School Program, Clinton Path Preschool, Dexter Park Day Care Center, Elizabeth Peabody House, Fletcher Community School, Harvard Law School Day Care Center, Henry Buckner School, KLH Child Development Center, MIT Summer Day Camp, Mystic Learning Center, Radcliffe Day Care Center, Roberts Day School, Somerville Boys' and Girls' Club, Somerville YMCA, Stride-Rite Children's Center, Technology Day Care Center, and the Tobin Community School. Reprints may be obtained from Steven Pinker, Department of Brain and Cognitive Sciences, El0-018, MIT, Cambridge, MA 02139, U.S.A. 0010-0277 /87 /$22 .40 © 1987, E l s e v i e r S c i e n c e P u b l i s h e r s B . V . 196 S. Pinker et al. in children's spontaneous speech, they utter passive participles that they could not have heard in parental input, and in four experiments in which 3-8-year-olds were taught novel verbs in active sentences, they freely uttered passivized versions of them when describing new events. An alternative solution is that children at some point come to possess a semantic constraint distinguishing passivizable from nonpassivizable verbs. In two of the experiments, we show that children do not have an absolute constraint forbidding them to passivize nonactional verbs of perception or spatial relationships, although they passivize them somewhat more reluctantly than they do actional verbs. In two other experiments, we show that children's tendency to passivize depends on the mapping between thematic roles and grammatical functions specified by the verb: they selectively resist passivizing made-up verbs whose subjects are patients and whose objects are agents; and they are more likely to passivize spatial relation verbs with location subjects than with theme subjects. These trends are consistent with Jackendoff s "Thematic Hierarchy Condition" on the adult passive. However, we argue that the constraint on passive that adults obey, and that children approach, is somewhat different: passivizable verbs must have object arguments that are patients, either literally for action verbs, or in an extended abstract sense that individual languages can define for particular classes of nonactional verbs. When children learn to speak, they generalize from a finite sample of adult speech to an infinitely large language. How they succeed is the most important problem in language acquisition research. It is a particularly difficult problem because children have no systematic access to negative evidence: information about which strings of words are not grammatical sentences in the language. Children are neither corrected nor miscomprehended more often when they speak ungrammatically (Brown & Hanlon, 1970; HirshPasek, Treiman, & Schneiderman, 1984), and anecdotal evidence suggests that they pay no attention when they are corrected (Braine, 1971; McNeill, 1966). The absence of negative evidence makes acquisi.tion difficult for the following reason: if the child hypothesizes a rule system generating a language that is a superset of the target language, the input, strictly speaking, can never tell the child he or she is wrong (Gold, 1967; Osherson, Stob, & Weinstein, 1985; Pinker, 1979). The no-negative-evidence problem arises in concrete terms in cases where the child is faced with inputs that tempt him or her to form some overinclusive generalization, such as a productive rule that has exceptions, for some lexical items. In a seminal paper, C.L. Baker (1979) has outlined several intriguing examples where this learnability problem arises in the acquisition of certain Acquisition of the passive 197 lexicosyntactic alternations in English that have traditionally been attributed to grammatical transformations (Chomsky, 1965). Baker pointed out that alternations like dativization (John gave a painting to the museum / gave the museum a painting and raising It is likely that John will win / John is likely to win) apply to many different verbs, apparently justifying a rule that could apply productively to any structure of the required form, regardless of the verb it contained; for example, a rule that converted a V NP~ to NP2 phrase into a V NP2 NP1 phrase. The problem is that these alternations do not extend to all predicates; sometimes they exclude a predicate that is a near-synonym of one that does alternate (cf., John donated a painting to the museum / *donated the museum a painting; It is probable that John will win / *John is probable to win). One must explain either how the child unlearns the overgeneralized forms created by the productive rule, or how the child avoids making the generalization in the first place. In this paper we examine the passive construction, which involves most transitive verbs in English (e.g., 1) but not all of them (e.g., 2). (1) a. John owns three bicycles. Many people misunderstand the argument. Dr. Caron weighed the patient. b. Three bicycles are owned by John. The argument is misunderstood by many people. The patient was weighed by Dr. Caron. (2) a. John has three bicycles. The argument escapes many people. Tiny weighs 210 pounds. The coming decade will see many changes. This bottle contains a deadly poison. Tom resembles Gene. b. *Three bicycles are had by John. *Many people are escaped by the argument. "210 pounds are weighed by Tiny. *Many changes will be seen by the coming decade. *A deadly poison is contained by this bottle. *Gene is resembled by Tom. Given these facts, one could imagine three types of learning strategies. The child could coin a general productive rule allowing him or her to create passive participles for transitive verbs, such as those in (lb) on the basis of having heard their active counterparts (la) in the parental input. Let us call 198 S. Pinker et al. this hypothesis unconstrained productivity. The problem for this type of hypothesis, Baker argues, is that it leaves us with the problem of explaining how the child avoids applying such a rule to exceptional verbs such as those in (2a), which would result in the production of ungrammatical (and, because of the no-negative-evidence problem, unexpungeable) forms such as those in (2b). Or, to put it another way, we would have to explain how the exceptions have managed to survive in the language. The alternative that Baker proposed for the dative is that children are conservative learners: in the case of the passive, they would have to hear a verb passivized in the input before they add the passive participle form to their lexicons. In this account, children would be incapable of creating a rule that added nonwitnessed passive participles to their lexicons; they would just store lists of active and passive forms. They would not overgeneralize because they would not generalize to begin with: ungrammatical passives would never appear in parents' speech, and so the child would never learn those participles. Of course, active verb forms and their participles do not form unrelated lists, but on this account the relatedness of corresponding active and passive forms would be captured in a lexical redundancy rule (see Jackendoff, 1975). Such a rule might save space in memory by minimizing redundancy; it could hasten the learning of passive forms heard in the input; or it could dictate what a passive form must look like if one exists for a particular verb. But it would not sanction the addition of new passive participles to the lexicon unless the form was heard in the input. Fodor (1985) has also articulated and defended a version of this position. A third possibility is that despite initial appearances, there are no productive rules with arbitrary lexical exceptions. Rather, productive rules apply to subtle but well-defined classes of verbs defined by phonological and semantic constraints. Possessing representations of these constraints, speakers productively apply passivization to verbs that obey the constraints while avoiding the verbs that disobey them. Mazurkewich and White (1984) argue that a version of this hypothesis, which we will call constrained productivity, can explain how the dative alternation is applied correctly in ,the absence of negative evidence. Verbs are dativizable only if they follow Anglo-Saxon phonological patterns (more precisely, they must constitute a metrical foot; Grimshaw, 1985), and if the indirect object is a prospective possessor of the direct object. The phonological constraint accounts for the contrast between give and donate; the semantic constraint accounts for the contrast between She baked a cake for me / She baked me a cake and She stirred a cake for me / *She stirred me a cake. If there is a constraint that divides the passivizable verbs from the nonpassivizable ones, a speaker who possessed that constraint could be productive but could still avoid or recover from overgeneraliAcquisition of the passive 199 zations. Jackendoff (1972) has in fact proposed a constraint on passivizability, which we will discuss later in some detail. Despite the many studies of children's comprehension of the passive (e.g., Bever, 1970; Maratsos, 1974; de Villiers & de Villiers, 1973), there is little evidence relevant to how children acquire the correct domain of application of passivization in English in the absence of negative evidence. The only relevant data come from two papers whose authors argue that children are not conservative passivizers. Each paper reports examples of children spontaneously using verbs in the passive that they were most unlikely to have ever heard passivized in parental speech. Wasow (1981) reported his daughter saying I don't like being faUed down on, and Bowerman (1983) reported her daughters saying Both are going to be go-ened in!, referring to two toilets, and I f you don't put them in for a very long time they won't get staled, referring to crackers in a bread box. In addition, passivization appears to be productive for adults (Wasow, 1981): given a novel transitive active verb (e.g., to clothesline, meaning to stop a charging athlete by extending an arm across his neck), adults can surely create their passive versions (e.g., Kurt was clotheslined by Kevin). Thus, Wasow points out, even if children are conservative, we would still need an explanation of how they become productive as adults. These facts, of course, do not settle the conservatism issue: it is unknown how representative of children's language the three reports of children's productive passives are (Fodor, 1985), nor what their precise grammatical status is (e.g., whether they are simply phonetic confusions of legitimate passive forms; whether they are adjectival or verbal passives; see Wasow, 1977, 1980). There have also been some suggestions as to what children's constraint on passivization might be if indeed they have a rule of passivization. Maratsos, Kuczaj, Fox, and Chalkley (1979) found that 4-5-year-old childrenwere able to understand both the active and passive versions of verbs denoting actions (e.g., kick, tickle), but only the active versions of verbs denoting perceptual or cognitive relations (e.g., see, hear, know, want). This phenomenon has been replicated by de Villiers, Phinney, and Avery (1982), by Maratsos, Fox, Becker, and Chalkley (1985), and by Gordon and Chafetz (1986). In their first report, Maratsos et al. (1979) suggested that children's passive rule was semantic, not syntactic; it applied not to grammatical subjects and objects, but to arguments bearing the semantic roles agent and patient. A variant of this account is that children's rule of passivization does apply to subjects and objects, but only to subjects and objects that are agents and patients respectively (Pinker, 1982). This is not the correct constraint for adult English (cf., A parcel was received by John; Tabs were kept on subversives; John was thought to be arrogant by his colleagues), but it might be a first approximation, 200 S. Pinker et al. because i t /s true that all agent-patient verbs passivize but not vice versa. However, the relevance of these studies to the current question is unclear. Maratsos et al. tested comprehension of existing English verbs whose degree of exposure to children in parental speech was uncontrolled. Passives of all sorts are rare in parental speech (Brown, 1973; Hochberg & Pinker, in preparation; Stromswold, Pinker, & Kaplan, 1985), and passivized perception verbs involving an experiencer as subject and a stimulus as object are even rarer (Maratsos et al., 1985). Thus children's performance in these experiments may simply have reflected conservative learning of the combinations of verb and voice exemplified in parental input. Some evidence that this may be so comes from Gordon and Chafetz (1986), who found a high correlation between the individual verbs that children had difficulty with in a test of comprehension of passive sentences, and the verbs that they had trouble with in a retest one week later. In any case, successful comprehension is at best an imperfect index of whether a sentence is within a speaker's language: adults, for example, can fail to comprehend legitimate passives (e.g., The horse raced past the barn fell) and can easily comprehend passive strings that are clearly outside their language (e.g., *Shampoo is contained by the bottle). In the experiments reported in this paper, we test these three hypotheses (conservatism, unconstrained productivity, and constrained productivity) using two different methods. Both methods are addressed to the existence and domain of productive passivization in children, not just the ability to master phrases heard in the input, and neither of them relies on comprehension abilities. First, we examine transcripts and reports of spontaneous speech to see if children regularly produce passives that they could not have heard in the input. Second, we expose children to novel verbs in either the active or passive voice, and test the children's ability to produce each verb in the voice in which they heard the verb and, more importantly, in the voice in which the verb had never been heard. Third, in the experiments we vary the semantics of the verb and thus ascertain whether children's tendency to passivize, if it exists, is sensitive to the kinds of constraints that delineate the domain of passivizability in adults.