DANA MCDANIEL AND HELEN SMITH CAIRNS

THE PROCESSING AND ACQUISITION OF CONTROL STRUCTURES BY YOUNG CHILDREN

In order to interpret a sentence involving control, the hearer must identify a referent for the phonetically null element PRO. It is, therefore, of interest to investigate how children acquire this ability. This paper reports some data that we have obtained on the acquisition of control principles by young children. This work is part of a larger research project including investigation of the development of the binding principles. We have some suggestions about the relationship between processing and the acquisition of control that are very tentative, but we think quite promising. Our work has dealt with control in both complement and adjunct clauses. We are following an analysis of control such that there is a rule in Universal Grammar indicating that PRO is controlled by the closest c-commanding NP. We are aware of the fact that there are control structures that are problematic for this analysis. We adopt it, however, because it is a plausible account of the structures we are studying within the constraints of UG. We feel that it is impossible to trace the development of an aspect of grammar unless our conception of that aspect of grammar is informed by a linguistic analysis. Should other analyses appear to be more plausible in the future, then they should be required to provide an account of our developmental data. The rule that PRO is controlled by the closest c-commanding NP leads to object control in the case of VP attached complements, such as (1), and to subject control for S attached adjuncts, such as (2).1 (1) (2) Cookie Monster tells Grover to jump over the fence. Cookie Monster touches Grover after jumping over the fence.

The typical finding is that children perform correctly on the complements at a fairly early age. (Although Tavakolian (1981) has reported some subject control errors in complements among very young children, and difficulty with exceptional verbs such as 'promise' is well known.) The adjuncts are a different story, however, producing errors
313
Lyn Frazier and Jill de Villiers (eds.), Language Processing and Language Acquisition, 313-325. 1990 Kluwer Academic Publishers.

314

DANA McDANIEL AND HELEN SMITH CAIRNS

of interpretation (from the point of view of the adult) until early school age. Hsu et al. (1985), based on a series of act-out tasks, have demonstrated a developmental sequence in the acquisition of control in adjuncts that goes like this. First, the child lacks the rule for control based on c-command and uses a strategy to determine the controller of PRO. The most primitive strategy is to select the subject (the wellknown first noun strategy), followed by a Minimal Distance Strategy, which results in the selection of the object as controller. Next, the child develops the rule for control, but erroneously attaches the adjunct clause to the VP, so the rule, correctly applied, yields object control. Hsu et al. distinguish between children who are using an object strategy and those who are basing object control on VP attachment by presenting sentences such as (3). (3) Cookie Monster stands near Grover after jumping over the fence.

The strategy users will select the closest NP, Grover, as the controller, but the more advanced children select Cookie Monster, since the object of the preposition does not c-command PRO. Next is a transitional stage during which the child begins to attach the adjunct to the S, but does not do so reliably, so mixed responses are observed, with some children vacillating between subject and object control. Finally, the child reliably attaches to the S and exhibits essentially adult-like behavior. This developmental sequence has been verified by a number of experimental studies and also by a small longitudinal study conducted by Hsu and Cairns (to appear). Two aspects of the Hsu et al. (1985) work and analysis troubled us. First, the claim that the children initially lack the c-command rule for control suggests that for a period the child has a non-human grammar, since we are assuming that the rule is part of UG. Such a state of affairs would force us to a developmental view of the acquisition of universal principles. We, however, view Pinker's Continuity Hypothesis as being a more restrictive, and, hence, a more a priori attractive hypothesis. This is precisely because it does not admit grammars for children that are unconstrained by adult grammatical principles. The second problem with the work of Hsu and her colleagues is that all the experiments required the children to act out their interpretation of the relevant sentences. One aspect of this problem is that enactment strategies essentially intervene between the child's interpretation of the

PROCESSING AND ACQUISITION OF CONTROL STRUCTURES 315

sentence and the datum that demonstrates that interpretation. Further, an act-out task only illustrates one interpretation of a sentence. In developing grammars (indeed, in any grammar) there may be multiple possible interpretations, while one is preferable. This is one reason why linguists do not use enactment tasks when investigating the grammars of adults. Instead, they elicit from informants judgments regarding the grammaticality and referential properties of sentences of interest. Judgments are the best way we know of to evaluate an individual's competence. Enactment tasks are more appropriate if one wishes to obtain information about performance strategies and preferences. Since we are concerned with questions of competence, rather than performance, we decided to attempt to elicit linguistic judgments from the children in our studies. We developed a procedure for giving children training and practice in making the kind of judgments we needed. The procedure involves seeing each child two times for practice in giving judgments. During the first session, the linguist engages the child in dialogue about the nature of language, the fact that people speak different languages, etc., and gives the child practice in judging the well-formedness of sentences. In the second session there is more dialogue about the nature of language and practice reporting on the referential properties of sentences. Then, in a third session, children are asked to judge the sentences of interest. The interviews have a very informal tone of engaging the child in conversations about language (as with an adult informant). The children do very well with this task. They like being able to assist us in a scientific investigation. We present their participation to them as an important and serious enterprise, not as a game. The methodology is reported in full in McDaniel and Cairns (1987). We have done several studies using this methodology. The one reported here involved 20 children between the ages of 3: 9 and 5: 4. We elicited judgments regarding sentences relevant to the development of both binding and control and also asked them to act the sentences out (before they judged them) so that we would be able to compare actout and judgment responses. The full report of this study is given in McDaniel, Cairns, and Hsu (1990, forthcoming). We are also currently two-thirds through a longitudinal study involving a group of fourteen children who were 4: 1 through 4: 10 at the beginning of the year. Among the sentences we gave the children for judgment and act-out were control sentences such as (1) and (2). A major finding is that there

316

DANA McDANIEL AND HELEN SMITH CAIRNS

are children (usually very young) who lack control in such sentences. That is, they tell us that anyone can be jumping over the fence in (1) and (2); hence, for these children PRO is not controlled. In our first study, there were two children who lacked control in both complements and adjuncts and three more who had control in the complement, but not in the adjunct. (No one had acquired control in the adjunct before the complement.) Everyone of these five children acted out all six of the adjunct sentences using the object as the actor. Another seven children apparently had control, reporting that Grover had to be the jumper in both (1) and (2) and acting the sentences out accordingly. A crucial point here is that if we were only relying on act-out evidence, the children who have object control of PRO would have been indistinguishable from those who lack control, yet have an enactment strategy indicating a preference for an interpretation identifying the object as the subject of the adjunct clause verb. We also observed children in the mixed and adult stages of development, their act-out responses comporting with their judgments in almost every instance. The stages are further verified in our other studies. In our current study, we are not using an act-out task, but we do get information about the children's preferred interpretation by their first answer to the question "Who would jump over the fence?". We have also added a sentence such as (3) and have verified Hsu's finding that object control children do tend to shift to subject control for these sentences. We consider our most important finding in this range of data to be that there is a stage, previously unattested as far as we know, during which children lack control. Below, we will give our analysis of the structural representations generated by the various developing grammars. The major point here is that during the period in which children have no control, they will behave as though they are violating the c-command rule for control. This is an erroneous interpretation, however, for they cannot violate the rule for control if they do not have a requirement for control. Rather, these children have no controller for PRO (for them it is free) and they have a strategy to determine its preferable referent. This analysis is quite in concert with Pinker's Continuity Hypothesis. As we will show below, we have an account of why these children lack control, and this account is consistent with the claim that every developing grammar is constrained by UG, i.e. is a possible human grammar. Before presenting our analysis, we make precise two distinct aspects

PROCESSING AND ACQUISITION OF CONTROL STRUCTURES 317

of processing that will enter into our account. First is the act of processing that produces a representation of a sentence for a child to act on. To produce such a representation the processor will make use of linguistic knowledge (syntactic and semantic) and will be constrained by whatever limitations exist for the processing system. The representation is then acted upon in whatever manner is requested. In our studies it is either enacted or its grammatical and referential properties are reported. As Otsu's (1981) work suggests, if a processor is unable to produce an appropriate representation of a sentence, the relevant grammatical constraints cannot be operative. In his studies, he accounted for children's apparent violation of universal linguistic principles by demonstrating that they were unable to process the relevant structures. It is this sort of account of developmental phenomena, as we have mentioned, that allows us to maintain the Continuity Hypothesis. There is also another aspect of processing that will interest us here. We refer to the strategies that the child uses to interpret a sentence for which the grammar does not provide a unique interpretation. This is the aspect of processing that allows a child to decide, for example, if PRO is free, what its preferred referent will be. The "First Noun" and "Minimal Distance" strategies are processing strategies of this latter type. To make things easier to say, we propose to adopt the following terminology: We will refer to the first type of processing as "processing" and the latter as "interpretive strategies." Turning now to our analysis of the data, there are two questions that must be addressed. The first is why children initially do not have control in their grammars, and the second is why children do not correctly attach the adjunct clause when they do develop control. We will address them in turn. Recall that we assume that the rule determining the controller of PRO is that the controller is the closest c-commanding NP if there is one and that this rule is part of UG. Given these assumptions, along with the Continuity Hypothesis, children's grammars should correctly determine the controller of PRO as soon as they correctly analyze the construction. This means that if the clause containing PRO is subordinated to a higher clause as a complement or as an adjunct, the grammar should require control. Since young children do not appear to have PRO controlled in these cases, we suggest that in the representation generated by the processor the two clauses are coordinated, rather than one being subordinate to the other. 2 In this way, no NP in the first

318

DANA McDANIEL AND HELEN SMITH CAIRNS

clause c-commands PRO, so that PRO will have arbitrary reference (as is true of the adult control rule). This is similar to Tavakolian's (1981) Conjoined Clause Hypothesis and to Lebeaux's (this volume) account of the acquisition sequence. Our claim raises three questions. First, what are the details of the coordinate structures under discussion? Second, what limits the representation? Is it the result of a limitation of the grammar or of the processor? And, third, what causes the change from coordination to subordination? Turning first to the construction, the second clause of the coordinate structures in question would contain an empty subject and an untensed verb. Assuming that the empty subject is PRO, the requirement that it be ungoverned is met. However, the child's grammar is then apparently allowing an untensed verb to appear in a solitary clause. This is not necessarily ruled out by UG. In some languages, the subjunctive, which is untensed, may stand on its own. Further research could help to specify the children's exact analysis of these constructions and to test the accuracy of the Conjoined Clause Hypothesis in general. We have noted that in correcting the ungrammatical practice sentence "Was climbing the tree" several children gave the sentence "Climbing the tree." It would be interesting to ask the children who do not have control whether sentences like "Climbing the tree" and 'To climb the tree" are grammatical. It is possible, however, that children would allow these constructions only as part of a coordinate structure. There are also constructions that would test the Conjoined Clause Hypothesis. The proposed analysis predicts that children who have only coordination should find wh-movement out of such constructions impossible. That is, a sentence like "Who did John tell Mary to hit?" should be ungrammatical for them, since extraction is out of one part of a conjunction. In addition, such children should not have Exceptional Case Marking, since this could obtain only in a structure of subordination. They should, therefore, find sentences like "John wants Mary to leave" ungrammatical (since Mary is without Case) or they would misanalyze the sentence (possibly making Mary part of the higher clause and PRO the subject of the lower clause). We now turn to the question of why children's representations are initially limited to coordination. Assuming the Continuity Hypothesis and that X-Bar Theory is part of UG, subordination should be available to the child at the outset. There are two possible explanations for this limitation, one related to the processor, the other to the grammar. It is

PROCESSING AND ACQUISITION OF CONTROL STRUCTURES 319

well known that coordinated representations are easier to process than subordinated ones. The former involves little more than the sequential processing of each member of the coordinate. Subordination, on the other hand, involves holding the first clause open so that the second clause can be embedded into it. It is likely that such processing skills require time and adequate memory resources to develop. If the processor is unable to produce subordinated representations, then the child cannot take advantage of their availability in the grammar. Another possibility is that the grammar will not yet generate subordinate structures because the lexical and semantic preconditions for such structures are not yet met. In order to know that a clause is subordinate to another clause, something must be known about the nature of the semantic link between them. If sub-categorization restrictions, in the case of complements, and the nature of subordinating conjunctions, in the case of adjuncts, are not part of the child's lexical knowledge, the semantic link will be ignored and the subordinate structure will not be required. We do not mean to suggest that these two accounts are entirely unrelated, since it is known that the processor responds to the semantics of connectors and to the sub-categorization properties of verbs in its operation. Turning now to the question of why representations change from coordinated to subordinated, there are two possible accounts. A purely processing account would simply suggest that as the limitations on the processor are relaxed developmentally, full subordinated structures are produced in response to the grammatical requirements. A purely semantic account would suggest that as the semantic links are made available by increased lexical knowledge, subordination will follow. We think that a combination of these two explanations is most plausible. For subordination to begin, both the processing and the semantic conditions must be met. The child's processor must achieve sufficient maturity to allow it to respond as the semantic pre-requisites for subordination are met. It is reasonable to assume that this will happen first with complementation, as the semantic relation between verbs and their complements are very salient and learned quite early. Subordination for the adjuncts must await the lexical distinction between coordinating and subordinating conjunctions. It seems quite plausible that the relation between a subordinating conjunction and an adjunct is more subtle than that between a verb and its complement. Control develops in complements before it develops in adjuncts, then, simply

320

DANA McDANIEL AND HELEN SMITH CAIRNS

because the requirement for subordination is recognized earlier for the former than for the latter; but neither can take place until the processor reaches some threshold level of maturation. Turning now to the question of attachment, it is clear that the initial correct attachment of the complement to the VP is explained by the very information that triggered subordination in the first place, viz, its relationship to the verb. We are assuming that subcategorization information identifies a class of complements that are semantically related to the verb, hence will always attach to the VP. These might be recognizable as a class since they do not contain a connector and have to in Infl. It is more difficult to recognize the semantic relation between an adjunct and its matrix sentence. Further, attachment of the adjunct is not determined by the simple realization that it is affiliated with a subordinating conjunction. Our hypothesis is the following. Before the adjunct is construed as a subordinate (hence control) structure, many children have adopted an interpretive strategy to determine the referent of PRO. They have, then, an opinion as to what such sentences should mean. When the confluence of semantic and lexical developments requires that the adjunct be subordinated, the child selects the attachment site that will result (by the universal principle of control by the closest c-commanding NP) in the interpretation that (s)he has come to prefer on strategic grounds. Thus, those children who had an object strategy will attach the adjunct to the VP, so that the controller of PRO will be the object. Children who had a subject strategy, on the other hand, will correctly attach the adjunct to the S, so that the subject will be the closest c-commander of PRO. (Such children will, of course, never go through an object controlled or mixed grammatical stage.) Children who had no strategy or who had a strategy preferring either of the internal NPs to an outside referent might attach the adjunct to either the S or to the VP, or allow both possibilities. Such children would never go through an object controlled stage, but would move immediately to a mixed grammar. The existence of all these different type of grammars based on interpretive strategy is plausible from our investigations. In our first study, all the children without control in adjuncts had an object strategy. In our subsequent investigations, we also found children without control who had a subject strategy or who chose the subject or object with equal frequency, but never chose a referent outside the sentence. In addition, in our current longitudinal study, we have observed children without control who did not acquire

PROCESSING AND ACQUISITION OF CONTROL STRUCTURES

321

control, but appeared to change their strategy. Thus, we have demonstrated that children without control adopt a variety of strategies. We have also found two pilot subjects who, within a month, moved from no control with an apparent subject strategy to a grammar with adult control. According to this account, then, the grammars that children develop with respect to control in adjuncts depends on the interpretive strategy they happened to be using at the time their semantic knowledge became sophisticated enough for them to realize that the adjunct clause must be subordinate. Some children, those with a subject strategy at this time, will immediately have an adult grammar. Others, those who had an object strategy, must undergo further development before reaching the adult grammar. These children will pass through the stages from VP attachment to mixed to S attachment. This progression of stages is diagrammed below. Attachment sequence Object strategy Mixed strategy Subject strategy VP mixed mixed
S S S

We will not attempt to account for the progression through these stages. Semantic development, involving better understanding of the subordinating conjunctions, probably plays a role. In our first study we also had sentences with the "in order to" connector, such as (4). (4) Grover touches Bert in order to jump over the gate. Three children required object control for these, although they had subject control for other adjuncts. This suggests that the semantics of the connectors are acquired individually. Perhaps the fact that sentences with "in order to" contained the morpheme "to" persuaded the children to maintain their analysis as analogous to complements. This would suggest that S attachment is associated with a class of lexically specified subordinating conjunctions. The children also could hear sentences with adjuncts where the subject is clearly intended as the controller of PRO. Sentences with the adjunct preposed might also constitute positive evidence for S attachment, since complements of the verb do not easily prepose.

322

DANA McDANIEL AND HELEN SMITH CAIRNS

This account is highly speculative and is not easy to test empirically. Whereas a longitudinal study can determine which stages a child passes through, no study can definitively determine that a child has not passed through a certain stage. However, if in one interval of a longitudinal study, a large number of children were observed to move from no control and a subject strategy to adult control, our account would be supported. This is especially true if a large number of children with an object strategy in the session before they had control were observed to pass through the stages before reaching the adult grammar. It would be best to study both the strategies and the grammatical development of a large number of children without control. The children should be given both an act-out task and asked for judgments at short intervals. We are planning such a study for the future. There is another aspect of our data that we would like to discuss, although we are not sure whether it should be understood within a theory of the acquisition of control. During our first study we discovered that many children (12 of the 20, in fact) required the pronoun in sentences like (5) to have internal reference. (5) Grover touches Bert before he jumps over the fence. Further, the identity of the required referent correlated perfectly with the child's grammar with respect to control of PRO. Thus, if the child were object controlled, (s)he would insist that "he" had to be "Bert"; the adult children required it to be Grover; and the mixed children would allow either. Children who did not have control for PRO thought that the pronoun could be free. The children who required a particular referent for the pronoun in sentences such as (5) were not generally confused about pronouns. They knew that the pronoun is free in the analogous sentence (6). (6) Grover told Bert that he would jump over the fence. Thus, it appears that this phenomenon is limited to adjuncts and is intimately connected with the child's acquisition of control. In our current longitudinal study, we have found two such children among 14. It occurred to us that whatever was occurring in the grammars of these children might extend beyond pronouns. We asked some of the children in the first study and all the children in the current one for judgments on the sentences of (7). It turns out that children

PROCESSING AND ACQUISITION OF CONTROL STRUCTURES 323

(7)

Bert hit Cookie Monster before jumped over the fence. Grover

~he

who require pronouns to have a referent in adjuncts find such sentences either ill-formed or of uncertain grammaticality ("*" or "??"). It appears, then, that such children have a requirement in their grammar that there be a referential link between the matrix and adjunct clauses in such sentences. Either PRO or a pronoun with the appropriate features could fulfill this requirement. In either case, the element will be controlled, although the controller is determined independently. If there is no such element, the sentence is ungrammatical. There are certain findings in the literature that suggest our conclusion is correct. Goodluck (in press), Tavakolian (1978), Lust et al. (1986) and others have noticed that children's preference for pronominal reference in unrestricted cases tends to parallel those for control. What these studies have not told us is whether those were only strong preferences or whether some grammatical principle required restricted reference. Another supportive study is by Smith and Van Kleeck (1986), who gave children ages 3: 6 to 6: 0 an act-out task and an imitation task using the types of construction in question. They were actually interested in assessing the effect of syntactic complexity on the children's performance, rather than in their grammatical knowledge. They selected the construction with PRO in an adjunct clause as a complex construction, since the interpretation involves finding an antecedent, and used sentences like (7), with an overt subject in the adjunct clause, as the simpler counterpart to the PRO construction. They expected the sentences with PRO to be more difficult to act out, due to their complexity, but easier to imitate, since they were shorter. Their results were not as anticipated, however. On the act-out task they found little difference between the two constructions in the number of errors. The most common type of error for the PRO construction was to choose the wrong controller, which comported with typical findings. In the construction with an overt subject, the most common error was to eliminate this subject, treating the construction as though the subject were PRO. Our hypothesis easily accounts for this otherwise surprising result. The sentences with an overt subject, as opposed to being simple, would have been ungrammatical for many of

324

DANA McDANIEL AND HELEN SMITH CAIRNS

the children. One solution was to act them out using the closest grammatical counterpart. The results of the imitation task were as Smith and van Kleeck had expected, the constructions with an overt subject being more difficult. However, they were surprised to obtain the same results on a second imitation task in which they controlled for length. The construction with an overt subject was still the more difficult. Again, this is a plausible consequence of the ungrammaticality of such sentences. A goal of our on-going research is to see how pervasive this type of grammar is among children. Seven of our longitudinal subjects still lack control in adjuncts, so we hope to be able to trace not only their acquisition of control, but also their extension of it to other elements in the adjunct clause. If some children go through such a stage and others do not, we need to find out what distinguishes the two types of child. At the present time we have few hypotheses about this or about why any children have such a restriction at alP

NOTES
I There are some adults who allow either the subject or the object to be the controller in (2). These adults probably have grammars like those of the children in the mixed control stage described below. 2 Greg Carlson (p.c.) has suggested that rather than having a coordinate structure, children might be analyzing the lower clause as an NP. Sentences like (1) and (2) would be interpreted as (i) and (ii) respectively.

(i) (ii)

Cookie Monster tells Grover [about] the jump over the fence. Cookie Monster touches Grover after the jumping over the fence.

Since the agent of the action nominal is undetermined, the child would interpret reference to be free in (1) and (2). Whereas we feel that this account is plausible, we will not pursue it here. It may be relevant for our account that languages exist in which parataxis is an alternative to subordination in structures of complementation. The paratactic construction is treated as a single sentence in terms of semantics and intonation, but is structurally identical to two separate clauses. Typically, the second clause of the paratactic construction lacks an overt subject. Noonan (1985; 55) gives the following example from Lango. (iii) Dako bkbbbi ico bkwJrJ kai woman told (3 SG DAT) man sifted (3SG) millet 'The woman said it to the man, [he] sifted millet: (The woman told the man to sift millet (and he did

PROCESSING AND ACQUISITION OF CONTROL STRUCTURES

325

Note that the structure in (iii) is not identical to the one we posit for children, since the subject of the second clause in (iii) is pro (Lango being a null subject language), whereas in the children's grammars it is PRO. The important point here is that there are languages in which complementation can be expressed through parataxis. 3 It has also been suggested that these children have a system of switch reference in their grammars. One problem with this is that children also restrict object pronouns in adjunct clauses (having an object pronoun obligatorily refer to an NP in the higher clause.) Switch reference systems, on the other hand, refer to the relationship between the subjects of the two clauses. However, we plan to investigate this issue further in future research.

BIBLIOGRAPHY Goodluck, H.: 'Children's interpretations of pronouns and null NP's', in B. Lust (ed.), Studies in the Acquisition of Anaphora, vol. II: Applying the Constraints, Reidel, Boston, (in press). Hsu, 1. R and Cairns, H. S.: Interpreting PRO: From strategy to structure, Annals of the New York Academy of Sciences, Linguistics Section, (to appear). Hsu, 1. R., Cairns, H. S., and Fiengo, R. W.: 1985, The development of grammars underlying children's interpretation of complex sentences', Cognition 20,25-48. Lebeaux, D.: The structure of the acquisition sequence, this volume. Lust, B., Solan, L., Flynn, S., Cross, c., and Schaetz, E.: 1986, 'A comparison of null and pronominal anaphora in first language acquisition', in B. Lust (ed.), Studies in the Acquisition of Anaphora, Vol. I: Defining the Constraints, Reidel, Boston. McDaniel, D. and Cairns, H. S.: 1987, The child as informant: Eliciting intuitions from young children, (unpublished ms.). McDaniel, D., Cairns, H. S., and Hsu, 1. R: 1990, 'Binding principles in the grammars of young children', Language Acquisition I, 121-139. McDaniel, D., Cairns, H. S. and Hsu, 1. R: Control principles in the grammars of young children, forthcoming. Noonan, M.: 1985, 'Complementation', in T. Shopen (ed.), Language typology and syntactic description, Volume II: Complex Constructions, Cambridge University Press, Cambridge. Otsu, Y.: 1981, Universal grammar and syntactic development in children: Toward a theory of syntactic development, MIT doctoral dissertation. Smith, C. and Van Kleeck, A.: 1986, 'Linguistic complexity and Performance', Journal of Child Language 13,389-408. Tavakolian, S. L.: 1978, 'Children's comprehension of pronominal subjects and missing subjects in complicated sentences', in H. Goodluck and L. Solan (eds.), Papers in the Structure and Development of Child Language, vol. 4, University of Massachusetts, Amherst. Tavakolian, S. L.: 1981, The conjoined-clause analysis of relative clauses', in S. L. Tavakolian (ed.), Language Acquisition and Linguistic Theory, MIT Press, Cambridge.