Unifying cross-linguistic and within-language patterns of finiteness marking in MOSAIC

MOSAIC, a model that has already simulated cross-linguistic differences in the occurrence of the Optional Infinitive phenomenon, is applied to the simulation of the pattern of finiteness marking within Dutch. This within-language pattern, which includes verb placement, low rates of Optional Infinitives in Wh-questions and the correlation between finiteness marking and subject provision, has been taken as evidence for the view that children have correctly set the clause structure and inflectional parameters for their language. MOSAIC, which employs no built-in linguistic knowledge, clearly simulates the pattern of results as a function of its utterance-final bias, the same mechanism that is responsible for its successful simulation of the crosslinguistic data. These results suggest that both the crosslinguistic and within–language pattern of finiteness marking can be understood in terms of the interaction between a simple resource-limited learning mechanism and the distributional statistics of the input to which it is exposed. Thus, these phenomena do not provide any evidence for abstract or innate knowledge on the part of the child. The Optional Infinitive phenomenon One of the key features of children’s early multi-word speech is that, in many languages, children often produce utterances that contain non-finite verb forms in contexts where a finite verb form is obligatory in the adult language. Thus, English-speaking children may produce utterances such as he go instead of he goes. While the English example may suggest that the child has simply omitted the inflectional morpheme (-es), data from languages such as German and Dutch where the infinitive carries its own morphological marker (-en) make it clear that children actually produce an infinitive instead of a finite verb form. Thus, German-speaking children may produce utterances such as Vater spielen (Daddy play-inf) instead of Vater spielt (Daddy plays-fin). While such Optional Infinitive errors are quite frequent in obligatory subject languages such as English, Dutch and German, they are quite rare in pro-drop languages like Spanish and Italian. An influential theory by Wexler (1994, 1998) explains this cross-linguistic pattern of results (and the relative sparseness of other types of errors) by assuming that children have correctly set all the clause structure and inflectional parameters for their language from a very early age, but optionally under-specify Agreement and/or Tense. Due to cross-linguistic differences between the grammars of pro-drop and obligatory subject languages, this leads to the provision of non-finite verb forms in contexts where a finite verb form is required in obligatory subject languages such as English and Dutch, but not in pro-drop languages such as Spanish. The great strength of Wexler’s account, which is consistent with the Universal Grammar approach to language acquisition (Chomsky, 1981; Pinker, 1984) is that it provides a unified explanation of the cross-linguistic occurrence of Optional Infinitive errors. Recently, however, Freudenthal, Pine and Gobet (2006, submitted) have shown that the cross-linguistic differences in the occurrence of Optional Infinitive errors may reflect sensitivity to differences in the distributional statistics of the input children are subject to. Thus, Freudenthal et al. showed that MOSAIC, a performance limited distributional analyser which receives child-directed speech as input and learns to produce progressively longer utterance-final phrases provides a good fit to the developmental characteristics of the Optional Infinitive phenomenon in English, Dutch, German and Spanish. The key to MOSAIC’s simulation of the differential rates of Optional Infinitive errors across the four languages is its bias towards learning material that occurs near the end of an utterance. Since finite and nonfinite verbs pattern differently in the four languages, utterance-final phrases have radically different proportions of non-finite verbs. MOSAIC was therefore able to provide a good quantitative fit to the cross-linguistic data as a result of the interaction between its utterance-final bias and the distributional properties of the languages to which it was exposed. While MOSAIC’s ability to deal with the cross-linguistic pattern of finiteness marking based on one, relatively simple processing constraint is encouraging, a clear challenge for MOSAIC remains the simulation of the pattern of regularities in finiteness marking that has been found within languages that display the OI phenomenon. Thus, Poeppel and Wexler (1993) have identified a number of constraints relating to verb placement, differential rates of OI-errors in declaratives and Wh-questions, and the correlation between the provision of subjects and the finiteness of an utterance. These regularities, which are particularly noticeable in languages such as German and Dutch, where verb position is dependent on finiteness, place considerable additional constraints on models of children’s early multi-word speech. This is particularly true, as these regularities appear to be consistent with Wexler’s claim that children have correctly set all the clause structure and inflectional parameters for their language from a very early age, whilst being problematic for rival Nativist theories of the Optional Infinitive phenomenon (e.g. Rizzi, 1994). Given the apparent support that it provides for Wexler’s account, the within-language pattern of finiteness marking can be considered a strong test for models like MOSAIC that embody the notion that children’s early multi-word speech reflects an interaction between a performancelimited learning process and the statistical regularities of the input children receive. Within-language regularities in the pattern of finiteness marking in German and Dutch German and Dutch are particularly interesting languages with respect to the development of finiteness marking as verb position in these languages is dependent on finiteness: finite verbs take second position, whereas non-finite verbs take sentence-final position. Poeppel and Wexler (1993) show that (in German) the child overwhelmingly places finite and non-finite verbs in their correct position from a very early age. From this finding, Poeppel and Wexler conclude that the child knows the rules for verb placement. Poeppel and Wexler also show that, while Optional Infinitive errors are quite frequent in German and Dutch children’s early declarative multi-word speech, they are virtually absent from Wh-questions. Wexler (1998) argues that it is a property of V2 languages (like German and Dutch) that Wh-questions should contain a finite verb. By implication, the absence of OI-errors in Dutch Wh-questions provides evidence for the notion that Dutch-speaking children have correctly set the V2 parameter (see also Poeppel & Wexler, 1993). Finally, there is a correlation in German and Dutch between the occurrence of finite verb forms in an utterance and the likelihood that the utterance contains a subject. It is well known (see e.g., Bloom, 1990) that children often produce utterances with missing subjects (e.g. Need a toy). Several authors have shown that utterances that contain finite verbs are more likely to contain a subject than utterances that contain only non-finite verb forms. Poeppel and Wexler (1993) explain this correlation by claiming that non-finite verbs license subject omission; it is grammatical for non-finite verbs to occur without subjects. Thus, in the utterance He wants to go, the non-finite verb go has no subject. Finite verbs, on the other hand, do require a subject. Wexler thus claims that children’s lower rates of subject provision on non-finite verbs reflects children’s knowledge that non-finite verbs can occur without overt subjects, whereas finite verbs cannot. In summary, the Optional Infinitive phenomenon has attracted a great deal of attention due to its occurrence across a range of languages. MOSAIC has already been shown to provide a close quantitative fit to the basic Optional Infinitive phenomenon in English, Dutch, German and Spanish. Nativist accounts such as Wexler’s (1998), explain the Optional Infinitive phenomenon by assuming that children misrepresent a small portion of the grammar, but have largely set the clause structure and inflectional parameters for their language correctly. The withinlanguage regularities are consistent with such a view, and therefore provide considerable constraints for any process model of children’s early multi-word speech. This paper aims to establish whether MOSAIC captures these regularities, thereby providing support for the notion that children’s early multi-word speech can be understood in terms of the interaction between psychologically plausible constraints on learning and the statistical structure of the