Early Language Acquisition: Children's Understanding of Relative Clauses, Lecture notes of Construction

Download Early Language Acquisition: Children's Understanding of Relative Clauses and more Lecture notes Construction in PDF only on Docsity! Haskins Laboratories Status Report on Speech Research 1989, SR-99 / 100, 118-136 Competence and Performance in Child Language* Stephen Craint and Janet Dean Fodortt 1. INTRODUCTION This paper presents the results of our recent experimental investigations of a central issue in linguistic theory: which properties of human language are innately determined? There are two main sources of information to be tapped to find the answer t9 this question. First, universal properties ofhuman languages are plausibly (even if not necessarily) taken to be innately determined. In addition, properties that emerge in children's language in the absence of decisive evidence in their linguistic input are reasonably held to be innate. Clearly, it would be most satisfactory if these two diagnostics for what is innate agreed with each other. In some cases they do. For example, there is a universal principle favoring transformational movement of phrases rather than oflexical categories e.g., topicalization of noun phrases but not of nouns. To the best of our knowledge children abide by this principle; they may hear sentences such as Candy, you can't have now~ but they don't infer that nouns can be topicalized. If they did, they would say things like *Vegetables, I won't eat the. But this is not an error characteristic of children. Instead, from the moment they produce topicalized constructions at all, they apparently produce correct NP­ topicalized forms such as The vegetables, I won't eat. In recent years, this happy convergence of results from research on universals and research This research was supported in part by NSF Grant BNS 84.18537, and by a Program Project Grant to Haskins Laboratories from the National Institute of Child Health and Human Development (HD·01994). The studies reported in this paper were conducted in collaboration with several friends and colleagues: Henry Hamburger, Paul Gorrell, Howard Lasnik Cecile McKee, Keiko Murasugi, Mineharu Nakayama, Jay~ Sarma and Rosalind Thornton. We thank them for their permission to gather this work together here. 118 on ~cquisition has been challenged by expenmental studies reporting various syntactic failures on the part of children. The children in these experiments are apparently violating putatively universal phrase structure principles or constraints on transformations. Failure to demonstrate early knowledge of syntactic principles is reported by Jakubowicz (1984), Lust (1981), Matthei (1981, 1982), Phinney (1981), Roeper (1986), Solan and Roeper (1978) Tavakolian (1978, 1981) and Wexler and Chie~ (1985). Some explanation is clearly called for if a syntactic principle is respected in all adult languages but is not respected in the language of children. Assuming that the experimental data accurately reflect children's linguistic competence, there are several possible responses to the unaccom­ modating data. The most extreme would be to give up the innateness claim for the principle in question. One might look for further linguistic data which show that it isn't universaL Or one might abandon the hypothesis that all universal prin~iples are innate. For instance, Matthei (1981) obtamed results that he interpreted as evidence that universal constraints on children's interpretation of reciprocals are learned not innate. However, this approach is plausible o~ly if one ~an offer so~e other explanation (e.g., functIonal explanatIon) for why the constraints should be universal. But this is not always easy; as Ch~msky (1986) has emphasized, many propertIes of natural language are arbitrary and have no practical motivation. A different response to the apparent failure of children to respect constraints believed to be ~nnate.is to argue that the constraints are as yet mapphcable to their sentences. The claim is that as soon as a child's linguistic analyses have reached the level of sophistication at which a The Perception of Phonetic Gestures 117 Stoll, G. (1984). Pitch of vowels: Experimental and theoretical investigation of its dependence on vowel quality. Speech Communication, 3, 137-150. Studdert-Kennedy, M. (1986). Two cheers for direct realism. Journal ofPhonetics; 14, 99-104. Studdert-Kennedy, M., &: Shankweiler, D. (1970). Hemispheric specialization for speech perception. Journal of the Acoustical Society ofAmerica, 48, 579-594. VanDerVeer, N. (1979). Ecological acoustics: Human perception of entlironmental sounds. Unpublished doctoral dissertation, Cornell University. Warren, W. &: Verbrugge, R. (1984). Auditory perception of breaking and bouncing events: A case study in ecological acoustics. Journal of Experimental Psychology: Human Perception and Performance, 10, 704-712. Whalen, D. (1984). Subcategorical mismatches slow phonetic judgments. Perception and Psychophysics, 35, 49-64. Whalen, D. &: Liberman, A. (1987). Speech perception takes precedence over nonspeech perception. Science, 237,169-171. FOOTNOTES "In I. G. Mattingly &: M. Studdert-Kennedy (Eds.), Modularity and the motor theory of speech perception. Hillsdale, NJ: Lawrence Erlbaum Associates, in press. tAlso Dartmouth College, Hanover, New Hampshire. ttAlso University of Connecticut, Storrs. Now at the University of California at Riverside, Department of Psychology. 1There is a small qualification to the claim that listeners cannot tell what contributions visible and audible information each have to their perceptual experience in the McGurk effect. Massaro (1987) has shown that effects of the video display can be reduced but not eliminated by instructing subjects to look at, but to ignore, the display. 2Liberman et al. identify a cipher as a system in which each unique unit of the message maps onto a unique symbol. In contrast, in a code, the correspondence between message unit and symbol is not 1:1. 3Liberman et al. propose to replace the more conventional view of the features of a phoneme (for example, that of Jakobson, Fant &: Halle, 1951) with one of features as "implicit instructions to separate and independent parts of the motor machinery" (p. 446). 4With one apparent slip on page 2: "The objects of speech perception are the intended phonetic gestures of the speaker, represented in the brain as invariant motor commands...." Sane can certainly challenge the idea that listeners recover the very gestures that occurred to produce a speech signal. Obviously there are no gestures at all responsible for most synthetic speech or for "sine-wave speech" (e.g., Remez, Rubin, Pisoni, &: Carrell, 1981) and quite different behaviors underlie a parrot's or mynah bird's mimicking of speech. The claim that we argue is incontrovertible is that listeners recover gestures from speech-like signals, even those generated in some other way. (We direct realists [Fowler, 1986a,b] would also argue that "misperceptions" (hearing phonetic gestures where there are none) can only occur in limited varieties of ways-the most notable being signals produced by certain mirage-producing human artifacts, such as speech synthesizers or mirage­ producing birds. Another, however, possibly, includes signals produced to mimic those of normal speakers by speakers with pathologies of the vocal tract that prevent normal realization of gestures.) 6There are two almost orthogonal perspectives from which perception can be studied. On the one hand, investigators can focus on processes inside the perceiver that take place from the time that a sense organ is stimulated until a percept is achieved or a response is made to the input. On the other hand, they can look outside the perceiver and ask what, in the environment, the organism under study perceives, what information in stimulation to the sense organs allows perception of the things perceived, and finally, whether the organisms in fact use the postulated information. Here we focus on this latter perspective, most closely associated with the work of James Gibson (e.g., 1966; 1979; Reed &:Jones, 1982). 7It is easy to find examples in which perception is heteromorphic with respect to the proximal stimulation and homomorphic with respect to distal events-looming, for example. We can also think of some examples in which perception appears homomorphic with respect to proximal stimulation, but in the examples we have corne up with, they are homomorphic with respect to the distal event as well (perception of a line dfawn by a pencil, for example), and so there is no way to decide whether perception is of the proximal stimulation or of the distal event. We challenge the motor theorists to provide an example in which perception is homomorphic with structure in proximal stimulation that is not also homomorphic with distal event structure. These would prOVide convincing cases of proximal stimulation perception. 8Bregman (1987) considers duplex perception to disconfirm his "rule of disjoint allocation" in acoustic scene analysis by listeners. According to the rule, each acoustic fragment is assigned in perception to one and only one environmental source. It seems, however, that duplex perception does not disconfirm the rule. 9Using a more sensitive, AXB, test, however, we have found that listeners can match the metal door chirp, rather than a wooden door chirp, to the metal door slam at performa~ce levels considerably better than chance. Competence and Perfonnance in Child Language 121 But the monkey drinking milk clause is a relative clause, and Subjacency prohibits the what from moving out of it. Thus the only acceptable structure is (2b), and the only acceptable answer is "a crayon." If these data are interpreted solely in terms of children's grammatical knowledge, then the conclusion would then have to be that knowledge of Subjacency sets in quite late in at least some children. As we noted earlier, Otsu suggested that the innateness of Subjacency could be salvaged by showing that the children who appeared to violate Subjacency had not yet mastered the phrase structure of relative clauses (of sufficient complexity to contain an extractable noun phrase). When he cOftducted an independent test of knowledge of relative clause structure, he found, as predicted, a correlation between phrase structure and Subjacency application in the children's performance. However, the children's performance was still surprisingly poor: 25% of the children who were deemed to have mastered relative clauses gave responses involving ungrammatical Subjacency violating extractions . from relative clauses. We have argued (Crain & Fodor, 1984) for an alternative analysis of Otsu's data, which makes it possible to credit children with knowledge of both phrase structure principles and constraints on transformations from an early age. We claim that children's parsing routines can influence their performance on the kind of sentences used in the Subjacency test; in particular, that there are strong parsing pressures encouraging subjects to compute the ungrammatical analysis of such sentences. Until a child develops sufficient capacity to override these parsing pressures, they may mask his syntactic competence, making him look as if he were ignorant of Subjacency. A powerful general tendency in sentence parsing by adults is to attach an incoming phrase low in the phrase marker if possible. This has been called Right Association; see Kimball (1973), Frazier and Fodor (1978). In sentence (3), for example, the preferred analysis has with NP modifying drinking milk rather than modifying drawing a picture, even though in this case both analyses are grammatically well-formed because there has been no WH-movement. (3) Mary is drawing a picture of a monkey that is drinking milk with NP. To see how strong this parsing pressure is, note how difficult it is to get the sensible interpretation of (3) when a crayon is substituted for NP. This Right Association preference is still present if the NP in (3) is extracted, as in (1). The word with in (1) still coheres strongly with the relative clause, rather than with the main clause. The result is that the analysis of (2) that is most immediately apparent is the ungrammatical (2a) in which what has been extracted from the relative clause. Since this 'garden path' analysis is apparent to most adults, it is hardly surprising if some of Otsu's child subjects were also tempted by it and responded to (1) in the picture verification task by saying "a straw" rather than "a crayon." We conducted several experiments designed to establish the plausibility of this claim that the relatively poor performance of children on sentences like (1) is due to parsing pressures rather than to ignorance of universal constraints. In the first experiment, we tested children and adults on complement-clause questions as in (4). Subjacency does not prohibit extraction from complement clauses, so if there were no Right Association effects this sentence should be fully ambiguous, with both interpretations equally available. (4) What is Bozo watching the dog jump through? That is, given a picture in which Bozo the clown is looking through a keyhole at a dog jumping through a hoop, it would be correct to say either the "the keyhole" or "a hoop." Intuitively, though, the interpretations are highly skewed for adults, with a strong preference for the Right Association interpretation ("hoop") in which the preposition attaches within the lower clause. Our experiment showed that the same is true for children. We tested twenty 3- to 5-year-olds (mean age 4;6) on these sentences using a picture verification task just like Otsu's, and 90% of their responses were in accord with the Right Association interpretation.2 Thus children and adults alike are strongly swayed by Right Association. This is an important result. To the best of our knowledge the question of whether children's parsing strategies resemble those of adults has not previously been investigated. But children certainly should show the same preferences as adults, if the human sentence parsing mechanism is innately structured. And the parsing mechanism certainly should be innately structured, because it would be pointless to be born knowing a lot of facts about language if one weren't also born knowing how to use those facts for speaking and understanding. It is satisfying, then, to have shown that children 122 Crain and Fodor exhibit Right Association. And the fact that they do offers a plausible explanation for why so many of them failed Otsu's Subjacency test-they were listening to their parsers rather than to their grammars. Our other experiments in support of this conclusion were designed to show that even people whose knowledge of Subjacency is not in doubt­ Le., adults-are also tempted to violate Subjacency when it is in competition with Right Association. We ran Otsu's Subjacency test on adults just as he did with the children. The adults gave Subjacency-violating low attachment responses to 21% of these questions. This was not quite as high a rate as for the children, but as we have noted, adults surely have a greater capacity than children do for checking an illicit analysis and shifting to a less preferred but well-formed analysis before they commit themselves to a response. In an attempt to equalize adult self­ monitoring capacities with those of children, we re-ran the Subjacency experiment with an additional distracting task (= listening for a designated phoneme in the stimulus sentence). Under these conditions the adults gave Subjacency violating responses to 29% of the relative clause constructions, a slightly higher rate than the 25% for Otsu's child subjects. Escalating still further, we changed the sentences so that the grammatically well-formed analysis was semantically or pragmatically anomalous, as in (5). (5) What color hat is Barbara drawing a picture of an artist with? Under these circumstances, where the semantics clearly favored the Subjacency-violating analysis, 75% of adults' responses violated Subjacency. This makes it very clear that linguistic competence may not always be revealed by linguistic performance. Finally, we ran another study, in which we asked adults to classify sentences as ambiguous or unambiguous. The sentences were spoken in turn with only a few seconds between them, and there were 72 of them, so the task was fairly demanding. The materials included complement questions like (4) and relative clause questions like (1), as well as ambiguous and unambiguous control sentences of many varieties. The results showed a 62% ambiguity detection rate for the ambiguous control sentences, with a 16% 'false alarm' rate for the unambiguous control sentences. Thus the subjects were able to cope with the task tolerably well, though not perfectly. What was interesting was that the ambiguity of the complement questions was detected only 48% of the time, in line with our claim that Right Association obscures the alternative reading with the prepositional phrase in the main clause. And most interesting of all was that 80% of the relative clause questions were judged to be ambiguous, even though Subjacency prohibits one analysis . and renders them unambiguous. Our explanation for this extraordinary result is that the subjects first computed the Right Association analysis favored by their parsing routines, then recognized that this was unacceptable because of Subjacency, and so rejected it in favor of the analysis with the prepositional phrase in the main clause. We assume that it was this rapid shift from one analysis to the other that gave our subjects such' a strong impression that these sentences were ambiguous. Note that if this misanalysis-with­ revision occurs 80% of the time for adults, only a slight handicap in children's ability to revise would be sufficient to account for their errors. To sum up: we still have no positive proof that Subjacency is innate, but at least now there is no evidence against it. Our experiments make it plausible that children as young as can be tested are like adults both with respect to their knowledge of this universal constraint and with respect to their parsing routines-they are just not very good yet at coping with conflicts between the two. 2.1.2. Backward pronominalization. A fundamental constraint on natural language is the structure dependence of linguistic rules. The innateness hypothesis implies that children's earliest grammars should also exhibit structure dependence-even if their linguistic experience happens to be equally compatible with structure­ independent hypotheses. However, it has been proposed that children initially hypothesize a structure-independent constraint on anaphora, prohibiting all cases of backward pronominalization (Solan, 1983). 3 Backward pronominalization consists of coreference between a noun phrase and a preceding pronoun, as indicated by the indices in (6). (6) That hei kissed the lion made the ducki happy. We will argue that children do in fact permit backward pronominalization, subject to structure­ dependent constraints. We contend that the appearance of a general restriction against backward pronominalization is due to a parsing preference for the alternative 'extra-sentential' Competence and Performance in Child Language 123 reading of the pronoun in certain comprehension tasks. The results of a new comprehension methodology show that children as young as 2;10 admit the same range of interpretations for pronouns as adults do. Two sources of evidence have been cited as evidence that children up to 5 or 6 years uniformly reject backward pronominalization. First, children who are asked to repeat back a sentence such as (7) often respond by converting it into a forward pronominalization construction, as in (8) (Lust, 1981). (7) Because she was tired, Mommy was sleeping. (8) Because Mommy was tired, she was sleeping. The fact that these children took the trouble to exchange the pronoun and its antecedent certainly indicates that they disfavor backward pronom­ inalization in their own productions. But it does not show that the backward pronominalization interpretation is not compatible with the child's grammar, as suggested by Solan (1983). To the contrary, the conversion of (7) to (8) shows that children do accept backward pronominalization in comprehension; for they would think of (8) as an acceptable variant of (7) only if they were interpreting the pronoun in (7) as coreferential with the subsequent lexical noun phrase (Lasnik & Crain, 1985). Second, it has been found that when the acting­ out situation for a sentence like (6) includes a potential referent for he other than the duck (e.g., a farmer), this unmentioned object is usually favored by the children as the referent of the pronoun (Solan, 1983; Tavakolian, 1978). In contrast to the prevailing view, we would attribute this to a parsing preference for the extra­ sentential interpretation of the pronoun; it does not have to be taken as evidence that children have a grammatical prohibition against backward anaphora. Our suggestion, then, is that children's knowledge might be comparable to, that of adults, even if their performance differs. It is particularly important to keep this distinction in mind for potentially ambiguous constructions such as these. When a sentence has more than one possible interpretation, the interpretation that children select can tell us which interpretation they prefer; it cannot show that others are unavailable to them. After all, adults also exhibit biases in connection with ambiguous constructions, but this does not lead us to accuse them of ignorance of alternative interpretations. To establish how much children actually do know, we should look for the factors that might be biassing their interpretations, and also for ways of minimizing this bias so that interpretations which are less preferred but nevertheless acceptable to them have a chance of showing through. The most likely general source of bias against backward pronominalization is the fact that interpretation of the pronoun would have to be delayed until the antecedent is encountered later in the sentence. This retention of uninterpreted items may strain a child's limited working memory. There is some evidence for this speculation. Hamburger and Crain (1984) have noted that children show a tendency to interpret adjectives immediately, without waiting for the remainder of the noun phrase, even in cases where this leads them to give incorrect responses. And Clark (1971) has observed errors attributable to children's tendency to act out a clause immediately without waiting for other clauses in the sentence. The only way to interpret the pronoun immediately in a sentence like (6) is to assign it an extra-sentential referent, as children typically do. If this proposal is correct, it should be that children will accept backward pronominalization in an experimental task that presses subjects to access every interpretation they can assign to a sentence. Crain and McKee (1985) used a true/false paradigm in which subjects judge the truth value of sentences against situations acted out by the experimenter. The sentences were as in (9), where either a coreferential reading or an extra-sentential reading of the pronoun is possible. (9) When he went into the barn, the fox stole the food. On each trial, a child heard a sentence following a staged event acted out by one of two experimenters, using toy figures and props. The second experimenter manipulated a puppet, Kermit the Frog. Following each event, Kermit said what he thought had happened on that trial. The child's task was to indicate whether or not the sentence uttered by Kermit accurately described what had happened. Children were asked to feed Kermit a cookie ifhe said the right thing, that is, if what he said was what really happened. In this way, 'true' responses were encouraged in the experimental situation. But sometimes Kermit would say the wrong thing, if he wasn't paying close attention. When this happened, the child was asked to make Kermit eat a rag. (In pilot 126 Crain and Fodor We noted in section 1 a range of possible explanations of apparently delayed knowledge of linguistic facts. In the present case they would include the following: • the construction does not, after all, belong to the core but is 'peripheral' and hence should be acquired late; • children don't hear this construction until quite late in the course oflanguage development and so could not be expected to know it exists; • the core principles in question undergo maturation and so are not accessible at early stages of acquisition; • the experimental data are faulty and children do indeed have knowledge of this construction. We will argue for this last alternative. And just as in the previously described studies of innate constraints, we will lay the blame for the misleading experimental data on the fact that traditional experimental paradigms do not make sufficient allowance for the limited memory and computational capacities of young children. Once again, our story is that non-linguistic immaturity can create the illusion of linguistic immaturity. 2.2.1. Relative Clauses. Children typically make more errors in understanding sentences containing relative clauses (as in 22) than sentences containing conjoined clauses (as in 23), when comprehension is assessed by a figure manipulation (act-out) task. (22) The dog pushed the sheep that jumped over the fence. (23) The dog pushed the sheep and jumped over the fence. The usual finding that (22) is more difficult for children than (23) up to age 6 years or so has been interpreted as an instance oflate emergence of the rules for subordinate syntax in language development (e.g., Tavakolian, 1981). However, though coordination may be innately favored over subordination, it is also true that subordination is ubiquitous in natural language; relative clause constructions are very close to the 'core.' So ignorance of relative clauses until age 6 would stretch the innateness hypothesis. Fortunately this is not how things stand. Hamburger and Crain (1982) showed that the source of children's performance errors on this task is not a lack of syntactic knowledge. By constructing pragmatic contexts in which the presuppositions of restrictive relative clauses were satisfied, they were able to demonstrate mastery of relative clause structure by children as young as 3 years. There are two presuppositions in (22): (i) that there are at least two sheep in the context, and (ii) that one (but only one) of the sheep jumped over a fence prior to the utterance. The reason why previous studies failed to demonstrate early knowledge of relative clause constructions, we believe, is that they did not pay scrupulous attention to these pragmatic presuppositions. For example, subjects were required to act out the meaning of a sentence such as (22) in contexts in which only one sheep was present. The poor performance by young children in these experiments was attributed to their ignorance of the linguistic properties of relative clause constructions. But suppose that a child did know the linguistic properties, but that he also was aware of the associated presuppositions. Such a child might very well be unable to relate his correct understanding of the sentence to the inappropriate circumstances provided by the experiment. Adult subjects may be able to 'see through' the unnaturalness of an experimental task to the intentions of the experimenter, but it is not realistic to expect this ofyoung children. Following this line of reasoning, Hamburger and Crain (1982) made the apparently minor change of adding two more sheep to the acting out situation for sentence (22), and obtained a much higher percentage of correct responses. The most frequent remaining 'error' was failure to act out the event described by the relative clause, but since felicitous usage presupposes that this event has already occurred, this is not really an error but is precisely the kind of response that is compatible with perfect- comprehension of the sentence. This interpretation of the data is supported by the fact that there was a positive correlation between incidence of this response type and age.5 We have conducted another series of studies on relative clauses, trying several other techniques for assessing grammatical competence. In one study, we employed a picture verification paradigm to see if children could distinguish relative clauses from conjoined clauses, despite the claim of Tavakolian (1981) that they systematically impose a conjoined clause analysis on relatives. In this study, seventeen 3- and 4­ year-oIds responded to relative clause constructions like (24). (24) The cat is holding hands with a man who is holding hands with a woman. (25) The cat is holding hands with a man and is holding hands with a woman. Competence and Performance in Child Language 127 This sentence was associated with a pair of pictures, one that was appropriate to it and one that was appropriate to the superficially similar conjoined sentence (25). 8eventy percent of the 3­ year-oIds' responses and 94 of the 4-year-olds' responses matched sentences with the appropriate picture rather than with the one depicting the conjoined clause interpretation. A second technique we tried used a 'silliness' judgment task (see Hsu, 1981) to establish whether children can differentiate relative clauses from conjoined clauses. Ninety-one percent of the responses of the twelve 3- and 4-year-olds tested categorized as 'silly' sentences such as (26), although sentences such as (27) were accepted as sensible 87% of the time. (26) The horse ate the hay that jumped over the fence. (27) The man watched the horse that jumped over the fence. Notice that sentence (26) would not be anomalous if the that -clause were misinterpreted as an and ­ clause, or if it were interpreted as extraposed from the subject NP; in both cases, the horse would be the understood subject of the relative clause. The results therefore indicate that most children interpret the that-clause in this sentence correctly, i.e., as a subordinate clause modifying the hay. Informal testing of adults suggests that the only respect in which children and adults differ on the interpretation of relative clauses is that the adults are somewhat more likely to accept the extraposed relative analysis as well, though even for adults this analysis is much less preferred. A third experiment, on the phrase structure of relative clause constructions, indicates that children, like adults, treat a noun phrase and its modifying relative clause as a single constituent, inasmuch as they can construe it as the antecedent for a pronoun such as one. In a picture verification study, fifteen 3- to 5­ year-olds responded to the instructions in (28). (28) The mother frog is looking at an airplane that has a woman in it. The baby frog is looking at one too. Point to it. Ninety-three percent of the time the subjects chose the picture in which the baby frog was looking at an airplane with a woman in it, in preference to the picture in which the baby frog was looking at an airplane without a woman in it. That is, the relative clause was included in the noun phrase assigned as antecedent to the pronoun. In short: the weight of evidence now indicates that children grasp the structure and meaning of relative clause constructions quite early in the course of language acquisition, as would be expected in view of the central position of these constructions in natural language.. 2.2.2. Temporal Terms. Another line of research has yielded support for the claim that presupposition failure is implicated in children's poor linguistic performance. These studies employed sentences containing temporal clauses with before and after, as in (29). (29) Push the red car to me before/after you push the blue car. Clark (1971) and Amidon and Carey (1972) have claimed that most normal, 3- to 5-year-olds do not understand these sentences appropriately. 8ince Amidon and Carey established that the children were familiar with concepts of temporal sequence (e.g., as expressed by words like first and last), the implication is that the structure of these adverbial clauses is beyond the scope of the child's grammar at this age. However, the acting-out tasks employed in these studies were once again unnatural ones which ignored the presuppositional content of the test sentences. Felicitous usage of sentence (29) demands that the pushing of the blue car has already been contextually established by the hearer as an intended, or at least probable, future event; but this was not established in these experimental tasks. It is very likely, then, that these studies underestimated children's ability to comprehend temporal subordinate clauses. For example, Amidon and Carey reported that five and six year old children who were not given any feedback frequently failed to act out the action described in the subordinate clause. Johnson (1975) found that four and five year old children correctly acted out commands such as those in (30) only 51% of the time; again, the predominant error was failure to act out the action described in the subordinate clause. (30) a Push the car before you push the truck. (81 before 82) b. Mter you push the motorcycle, push the bus. (After 81, 82) c. Before you push the airplane, push the car. (Before 82, 81) d Push the truck after you push the helicopter. (82 after 81) 128 Crain and Fodor Table 1. Note that the relevant factor is whether subordinate clause information was provided in advance. An analysis of variance confirmed that the mere amount of information provided makes no significant difference. The FI group performed better than the F group by only 6 percentage points, which does not approach statistical significance. And the I group performed just a little worse (non-significantly again) than the NC group. Although our study was not specifically designed to assess age differences, we performed a post hoc breakdown of correct responses by two age groups: under 4;4, and 4;4 and over. The older group, as one would expect, performed somewhat If he chose the bus, for example, a typical command would be (31). (31) Push the car before you push the bus. On the other hand, a subject in the I group who had chosen the bus would be given the command (32). (32) Push the bus before you push the car. Fifty-six children participated in the study, ranging in age from 3;4 to 5;10 (mean 4;5). Each child was assigned to one of the four groups, which were of equal size and approximately matched for age. The 'game' equipment consisted of 6 toy vehicles arranged in a row on a table between the child and the experimenter. The stimulus set consisted of 12 commands spoken by the experimenter which the child was to act out. There were three sentences of each of the four types illustrated in (30) above. We were careful to balance order of choice with order of action and assignment to clause type. The results showed a significant difference between the F and FI groups on one hand, and the I and NC groups on the other. Table 1 shows the percentages of correct responses, where a correct response consisted of performing both actions in the sequence specified by the sentence. Main Clause Information 55% 77% mean 66%65% + FI 80% F 74% I 51% NC 59% + mean Subordinate Clause Information Crain (1982) satisfied the presupposition of the subordinate clause by having the subordinate clause act correspond to an intended action by the subject, and observed a striking increase in children's performance. To satisfy the presupposition, children were asked, before each command, to choose a toy to push on the next trial. The child's intention to push a particular toy was incorporated into the command that was given on that trial. For instance, sentence (30d) could be used felicitously for a child who had expressed his intent to push the helicopter. Correct responses (i.e., responses in which both the main clause and subordinate clause action were performed, and in the correct order) were produced 82% of the time. Crain's interpretation of these results was that the children's improved performance was due to the satisfaction of the presupposition of the subordinate clause. However, we now note that the results of that study are open to another interpretation. It may be that improved performance was not due specifically to the contextual appropriateness of the sentence, but to the fact that the child's task was simplified because he was provided with more advance information concerning what his task would be. In the act-out or 'do-what-I-say' paradigm applied to temporal terms, the child must discern two aspects of the command: (i) which two toys to move, and (ii) in which order to move them. If the child has established his intent to move a particular toy, his task involving (i) is simplified. Thus, improved performance may be due to the satisfaction of presuppositions or it may be due to the additional information the child possesses. Another study was conducted to disentangle these two factors (Gorrell, Crain, & Fodor, 1989). In this study, there were four groups of subjects. One group, the Felicity Group (F), was given commands containing before and after with prior information about the subordinate clause action, just as in the previous experiment. A second group, the Information Group (I) received prior information about the main clause action; note that this does not satisfy the presupposition of the sentence. There was also a third group, the No Context Group (NC), who received no advance information at all, and a fourth group, the Felicity plus Information Group (FI), who received information over and above what would satisfy the felicity conditions since they chose both actions in advance. Consider, for example, a subject in the F group. He would be asked to choose a toy to push. Competence and Performance in Child Language 131 Planning complexity can also be deconfounded from the complexity of semantic constituent processing. Note that semantic considerations lead to the prediction that (39i), would be hardest, because its word meanings have to combine in a way contrary to the surface sequence of words (= the biggest ofJohn's books). The pattern of children's responses supports the predictions of our procedural account, and does not conform to the account based on syntactic or semantic constituency. Children responded correctly to phrases like (39i) 88% of the time. They gave correct answers only 39% of the time for examples like (39ii), and they were only 17% correct for phrase like (39iii). Thus, this experiment provides clear evidence that plans, not linguistic structures (syntactic or semantic) can determine processing success and failure for young children. The second experiment addresses the cognitive difficulty of planning by prefacing the test with a sequence of exercises designed to alleviate the planning difficulty. This activity does not provide any extra exposure to the phrases tested; nevertheless, we anticipate a reduction in errors on these phrases. Consider a phrase such as "the second tallest building." This plan requires the interpreter to identify its referent. The child must integrate sequential pairwise comparisons of relative size. In the pre-test activity the child would be shown a display of several objects of one type (say boxes), but of different sizes, and asked to hand the experimenter the biggest one. Then, once this object was removed from the array, the experimenter asked the child to perform the task again, saying, "Now, find the biggest box in this group." In this way the child would identify the second biggest box without ever hearing the phrase "the second biggest box" uttered. Children's comprehension of the phrase was tested before and after the preparatory task. They gave significantly more correct responses (46%) following the preparatory task than before it (8% in this experiment). This result suggests that their difficulty with phrases of this sort stems from the complexity of the response plans. 3. SENTENCE PRODUCTION To acquire a language is to learn a mapping between potential utterances and associated potential meanings. Successful mastery should reveal itself in both comprehension and production. In the previous section we were concerned with studies of children's comprehension, in which their knowledge is tested by presenting utterances and observing the interpretations that they assign. We now turn to tests of children's competence which proceed in the other direction: the input to the child is a situation, which has been designed to suggest a unique sentence meaning, and the behavior we observe is the utterance by which the child describes that situation. It would have been reasonable to expect that the sorts of nonsyntactic problems that present obstacles for children in comprehension tasks might prove to be as hard or even harder for them to overcome in production tasks. But we have not found this to be the case. The results of recent elicited production studies are dramatically better than those of comprehension studies directed to the same linguistic constructions. For example, Richards (1976) elicited appropriate uses of the deictic verbs come and go from children age 4;0 ­ 7;7, while Clark and Garnica (1974) reported that even 8-year-olds didn't consistently distinguish between come and go in a comprehension task. The disparity between production and comprehension studies is particularly striking because it is the reverse of what one would expect. To find production superior to comprehension in children's language is as surprising as it would be to find production superior to comprehension in adult second-language learning, or to find recall superior to recognition in any psychological domain. It is plausible to argue, therefore, that the superiority of production is only apparent, and is due to differences in the sensitivities of production tests and comprehension tests. And the logic of the situation suggests that it is the comprehension tests that are deficient. After all, success is hard to argue with. With suitable controls, successful production by children is a strong indicator of underlying linguistic competence, as long as their productions are as appropriate and closely attuned to the context as adult utterances are. Because there are so many ways to combine words incorrectly, consistently correct combinations in the appropriate contexts are not likely to come about by accident. On the other hand, failure on any kind of psychological task cannot be secure evidence of lack of the relevant knowledge, since the knowledge may be present but imperfectly exploited. As we saw in the previous section, com­ prehension studies seem to be particularly susceptible to problems of parsing, planning and so forth which impede the full exploitation of linguistic knowledge. Production tasks appear to be less hampered by these extra-grammatical 132 Crain and Fodor factors. This is probably because production avoids non-verbal response planning, which we have seen is a major source of difficulty in act-out comprehension tasks. It is worth noting also that . in constructing contexts to elicit particular utterance types, we have no choice but to attend to the satisfaction of the presuppositions that are associated with the syntactic structures in question, because otherwise the subjects won't utter anything like the construction that is being targeted. In elicited production it is delicate manipulations of the communicative situation that give one control over the subject's utterances. 3.1. Relative Clauses. In section 2 we presented evidence ofyoung children's competence with relative clauses. Further confirmation was obtained by Hamburger and Crain (1982), using an elicited production methodology. Pragmatic contexts were constructed in, which the presuppositions of restrictive relatives were satisfied. It was discovered that children as young as three reliably produce relative clauses in these contexts. A context that is uniquely felicitous for a relative clause is one which requires the speaker to identify to an observer which of two objects to perform some action on. In our experiment, the observer is blindfolded during identification of a toy, so the child cannot identify it to the observer merely by pointing to it or saying this / that one. Also, the differentiating property of the relevant toy is not one that can be encoded merely with a noun (e.g., the guard) or a prenominal adjective (e.g., the big guard) or a prepositional phrase (e.g., the guard with the gun), but involves a more complex state or action (e.g., the guard that is shooting Darth Vader). Young children reliably produce meaningful utterances with relative clauses when these felicity conditions are met. For example: (40) Jabba, please come over to point to the one that's asleep. (3;5) Point to the one that's standing up. (3;9) Point to the guy who's going to get killed. (3;9) Point to the kangaroo that's eating the strawberry ice cream. (3;11) Note that the possibility of imitation is excluded because the experimenter takes care not to use any relative clause constructions in the elicitation . situation. This technique has now been extended to younger children (as young as 2;8), and to the elicitation of a wider array of relative clause constructions, including relatives with object gaps (e.g., the guard that Princess Leia is standing on). 3.2. Passives. Borer and Wexler (1987) have argued that A-chains, which are involved in the derivation of verbal passive constructions, are not available to children in the first few years.8 Borer and Wexler maintain that knowledge of A-chains is innate, but becomes accessible only after the language faculty undergoes maturational change. We were not convinced, however, that this maturation hypothesis is necessitated by the facts. Rather, the facts seem to be consistent with A­ chains being innate and accessible from the outset. One source of data cited in support of the maturation hypothesis is the absence of full passives in the spontaneous speech of young children. But this of course is not incontrovertible evidence that children's grammars are incapable of generating passives. Full passives are rarely observed in adults' spontaneous speech either, or in adult speech to children. But their paucity is not interpreted in this case as revealing a lack of grammatical knowledge. Instead, it is understood as due to the fact that the passive is a marked form which it is appropriate to use only in certain discourse contexts; in most contexts the active is acceptable and more natural, or a reduced passive without a by-phrase is sufficient. That is, the absence of full verbal passives in adult speech· is assumed to be a consequence of the fact that it's only in rare situations that the full passive is uniquely felicitous. But the same logic that explains why adults produce so few full passives may apply equally to children. Perhaps they too have knowledge of this construction, but do not use it except where the communicative situation is appropriate. We have tested this possibility in an experiment with thirty-two 3- and 4-year-old children. (Crain, Thornton, & Murasugi, 1987). One experimenter asked the child to pose questions to another experimenter. The pragmatic context was carefully controlled so that questions containing a full verbal passive would be fully appropriate. The following protocol illustrates the elicitation technique: Adult See, the Incredible Hulk is hitting one of the soldiers Look over here. Darth Vader goes over and hits a soldier. So Darth Vader is also hitting one of the soldiers. Ask Kciko which one. Child to Keiko: Which soldier is getting hit by Darth Vader? Competence and Performance in Child Language 133 Note that the child knows what the correct answer is to his question, and that he cannot expect to elicit this answer from his interlocutor (Keiko) unless he includes the by-phrase. In fact, exactly 50% of responses were passives with full by -phrases). Of course, active constructions are also felicitous in this context (e.g., Which soldier is Darth Vader hitting?), even though the contextual contrast with another agent (the Incredible Hulk) may tend to favor the passive stylistically. And indeed 31% of responses were active questions with object gaps. The other 19% of responses included mostly sentences that were grammatical but not as specific as the context demanded (e.g., passive lacking by-phrases). Using this technique, we were able to elicit full verbal passives from all but three of the thirty-two children tested so far, including ones as young as 3;4. Some examples are shown in (41). (41) She got knocked down by the Smurfie. (3;4) Which girl is pushing, getting pushed by a car? (3;8) He got picked up from her. (3;11) It's getting ate up from Luke Skywalker. (4;0) Which giraffe gets huggen by Grover? (4;9) Note that these utterances contain a variety of morphological and other errors, but they all nevertheless exhibit the essential passive structure (underlying subject in pre-verbal position; agent in post-verbal prepositional phrase). 9 It might be argued that the children's passives elicited in this experiment do not involve true A-chains. However, since they are just like adult passives (disregarding morphological errors), the burden of proof falls on anyone who holds that adult passives involve A-chains and children's passives do not. No criterion has been proposed, as far as we know, which distinguishes adult's and children's passives in this respect. For example, it is true that the children almost always use a form of get in place of the passive auxiliary be, but get is acceptable in adult passives also. Wet is more regular and phonologically more prominent than forms of be, and this may be why it is more salient for children.) Children's considerable success in producing passive sentences appropriate to the circumstances (i.e., their correct pairing of sentence forms and meanings) constitutes compelling evidence of their grammatical competence with this construction. Comparison of these results with the results of testing the same children with two comprehension paradigms (act- out and picture-verification) confirms that, like spontaneous production data, these measures underestimate children's linguistic knowledge. The finding that young children evince mastery of the passive obviates the need to appeal to maturation to account for its absence in early child language. Maturation cannot of course be absolutely excluded; but a maturation account is motivated only where a construction is acquired surprisingly late-where this means later than would be expected on the basis of processing complexity, pragmatic usefulness in children's discourse, and so forth. (Also, as noted in section 1, some important cross-language and cross­ construction correlations need to be established to confirm a maturational approach; see Borer & Wexler, 1987, on comparison of English passives with passive and causative constructions in Hebrew.) The elicited production results suggest that the age at which passive is acquired in English falls well within a time span that is compatible with these other factors, and so maturation does not need to be invoked. 3.3. Wanna contraction. Another phenomenon that can be shown by elicitation to appear quite early in acquisition is wanna contraction in English. The facts are shown in (42) and (43). (42)a. Who do you want to help? b. Who do you wanna help? (43)a. Who do you want to help you? b. Who do you wanna help you? Every adult is (implicitly) aware that contraction is admissible in (42b) but not (43b). However, on the usual assumption that children do not have access to 'negative data' (Le., are not informed of which sentences are ungrammatical) it is difficult to see how this knowledge about the ungrammaticality of sentences like (43b) could be acquired from experience (at any age). So this is yet another candidate for innate linguistic knowledge. (What is known innately would be that a trace between two words prevents them from contracting together. The relevant difference between (42b) and (43b) is that in (43b) the who is the subject of the subordinate clause and has been moved from a position between the want and the to. The trace of this noun phrase that is left behind blocks the contraction. In (42b), by contrast, the trace is in object position after help, and therefore is not in the way of the contraction.) Crain and Thornton (in press) used the elicited production technique to encourage children to ask questions that would reveal violations like (43b) if