Item-based Constructions and the Logical Problem Brian MacWhinney Department of Psychology Carnegie Mellon University Pittsburgh, PA 15213 [email protected] still, when such evidence is provided, children ap- Abstract pear to ignore it. Chomsky’s (1957) views about the degeneracy The logical problem of language is of the input did not stand up well to the test of grounded on arguments from poverty time. As Newport, Gleitman & Gleitman (1977) of positive evidence and arguments reported, ‘the speech of mothers to children is from poverty of negative evidence. Careful analysis of child language cor- unswervingly well-formed.’ More recently, Sagae, pora shows that, if one assumes that Lavie & MacWhinney (2004) examined several of children learn through item-based con- the corpora in the CHILDES database and found structions, there is an abundance of that adult input to children can be parsed with an positive evidence. Arguments regarding accuracy level parallel to that for corpora such as the poverty of negative evidence can the Wall Street Journal database. also be addressed by the mechanism of This evidence for well formedness of the input conservative item-based learning. When conservativism is abandoned, children did not lead to the collapse of the ‘argument from can rely on competition, cue construc- poverty of stimulus’ (APS). However, it did place tion, monitoring and probabilistic iden- increased weight on the remaining claims regard- tification to derive information from ing the absence of relevant evidence. The overall positive data to recover from overgener- claim is that, given the absence of appropriate alization. positive and negative evidence, no child can ac- quire language without guidance from a rich set of 1. The Logical Problem species-specific innate hypotheses. Some refer to the argument from poverty of stimulus as the ‘logi- Chomsky (1957, 1980) has argued that the child’s cal problem of language acquisition (Baker, 1979), acquisition of grammar is ‘hopelessly underdeter- while others have called it ‘Plato’s Problem,’ mined by the fragmentary evidence available.’ He ‘Chomsky’s Problem,’ ‘Gold’s Problem,’ or attributed this indeterminacy to two major sources. ‘Baker’s Paradox.’ The first is the degenerate nature of the input. Ac- cording to Chomsky, the sentences heard by the 2. Absence of Negative Evidence child are so full of retracing, error, and incomple- In the 1970s, generativist analyses of learnability tion that they provide no clear indication of the (Wexler & Hamburger, 1973) relied primarily on possible sentences of the language. Coupled with an analysis presented by Gold (1967). Gold’s this problem of input degeneracy is the problem of analysis contrasted two different language-learning unavailability of negative evidence. According to situations: text presentation and informant presen- this view, children have a hard time knowing tation. With informant presentation, the language which forms of their language are acceptable and learner can receive feedback from an infallible in- which are unacceptable, because parents fail to formant regarding the grammaticality of every provide consistent evidence regarding the un- candidate sentence. This corrective feedback is grammaticality of unacceptable sentences. Worse called ‘negative evidence’ and it only requires that 53 Proceedings of the Second Workshop on Psychocomputational Models of Human Language Acquisition, pages 53–68, Ann Arbor, June 2005. c 2005 Association for Computational Linguistics ungrammatical strings be clearly identified as un- rameter setting (Wexler, 1998), which could itself acceptable. Whenever the learner formulates an introduce some error. Thus, we would expect er- overly general guess about some particular linguis- ror-free learning to occur primarily for those as- tic structure, the informant will label the resulting pects of the grammar that are completely universal structure as ungrammatical and the learner will use and not parameterized. Parameterized features, this information to restrict the developing gram- such as subject pro-drop, could still be guided by mar. Based on initial empirical results reported by universal grammar. However, their learning would Brown & Hanlon (1970), Gold argued that nega- not necessarily be error-free. tive evidence is not available to the child and that language learning cannot be based on informant 3.1. Structural dependency presentation. Marcus (1993) has argued that the feedback that The paradigm case of error-free learning is the parents provide does not discriminate consistently child’s obedience to the Structural Dependency between grammatical and ungrammatical construc- condition, as outlined by Chomsky in his formal tions. As a result, children cannot rely on simple, discussion with Jean Piaget (Piattelli-Palmarini, overt negative evidence for recovery from over- 1980). Chomsky notes that children learn early on generalization. Although I will argue that parents to move the auxiliary to initial position in ques- provide positive evidence in a form that solves the tions, such as ‘Is the man coming?’ One formula- logical problem (Bohannon et al., 1990), I agree tion of this rule is that it stipulates the movement with the observation that this evidence does not of the first auxiliary to initial position. This formu- constitute overt grammatical correction of the type lation would be based on surface order, rather than envisioned by Gold. structural relations. However, if children want to question the proposition given in (1), they will 3. Absence of Positive Evidence never produce a movement such as (2). Instead, they will always produce (3). Beginning about 1980, generative analyses of 1. The man who is running is coming. learnability began to shift away from an emphasis 2. Is the man who __ running is coming? on the unavailability of negative evidence to argu- 3. Is the man who is running __ coming?’ ments based on the unavailability of positive evi- In order to produce (3), children must be basing the dence. This conceptual shift led to a relative movement on structure, rather than surface order. decline in attention to recovery from overgenerali- Thus, according to Chomsky, they must be in- zation and an increase in attention to reported cases nately guided to formulate rules in terms of struc- of error-free learning. For example, Chomsky’s ture. (1980) statement of the logical problem relies on In the theory of barriers (Chomsky, 1986), the the notion of error-free learning without positive repositioning of the auxiliary in the tree and then in evidence. The argumentation here is that, if a surface structure involves a movement of INFL to structure is never encountered in the input, correct COMP that is subject to the head movement con- use of this structure would have to indicate innate straint. In (2) the auxiliary would need to move knowledge. around the N’ of ‘man’ and the CP and COMP of Researchers have claimed that the child pro- the relative clause, but this movement would be duces error-free learning without receiving positive blocked by the head movement constraint (HMC). evidence for structures such as: structural depend- No such barriers exist in the main clause. In addi- ency, c-command, the binding conditions, subja- tion, if the auxiliary moves as in (2), it leaves a gap cency, negative polarity items, that-trace deletion, that will violate the empty category principle nominal compound formation, control, auxiliary (ECP). Chomsky’s discussion with Piaget does not phrase ordering, and the empty category principle. rely on these details. Chomsky simply argues that In each of these cases, it is necessary to assume the child has to realize that phrasal structure is that the underlying universal is a part of the non- somehow involved in this process and that one parameterized core of universal grammar. If the cannot formulate the rule of auxiliary movement as dimension involved were parameterized, there ‘move the first auxiliary to the front.’ would be a need for some form of very early pa- 54 Chomsky claims that, ‘A person might go movement of the auxiliary from the relative clause. through much or all of his life without ever having Unfortunately, this elicitation procedure encour- been exposed to relevant evidence, but he will ages children to treat the relative clause (‘the boy nevertheless unerringly employ the structure- who is watching Mickey’) as an imitated chunk. dependent generalization, on the first relevant oc- Despite the serious methodological limitation in casion.’ A more general statement of this type pro- this particular study, it seems reasonable to believe vided by Hornstein & Lightfoot (1981) who claim that four-year-old children are beginning to behave that, ‘People attain knowledge of the structure of in accordance with the Structural Dependency their language for which no evidence is available condition for sentences like (2) and (3). But does in the data to which they are exposed as children.’ this mean that they reach this point without learn- In order to evaluate these claims empirically, we ing? need to know when children first produce such There is another type of sentence that provides sentences and whether they have been exposed to equally useful positive evidence regarding auxil- relevant examples in the input prior to this time. In iary movement. These are wh-questions with em- searching for instances of relevant input as well as bedded relative clauses. It turns out that there are first uses, we should include two types of sen- hundreds