Modelling the Development of Dutch Optional Infinitives in MOSAIC

This paper describes a computational model which simulates the change in the use of optional infinitives that is evident in children learning Dutch as their first language. The model, developed within the framework of MOSAIC, takes naturalistic, child directed speech as its input, and analyses the distributional regularities present in the input. It slowly learns to generate longer utterances as it sees more input. We show that the developmental characteristics of Dutch children’s speech (with respect to optional infinitives) are a natural consequence of MOSAIC’s learning mechanisms and the gradual increase in the length of the utterances it produces. In contrast with Nativist approaches to syntax acquisition, the present model does not assume large amounts of innate knowledge in the child, and provides a quantitative process account of the development of optional infinitives. The Optional Infinitive Stage One phenomenon which has received considerable attention in the area of syntax acquisition is the socalled Optional Infinitive (OI) stage (Wexler, 1994, 1998). Children in the OI stage of development use a high proportion of (root) infinitives, that is, verbs which are not marked for tense or agreement. In English, root forms such as go, or eat are infinitive forms, whereas ate or goes are marked for tense and agreement + tense respectively. Verbs which are marked for agreement or tense are known as finite verbs. (Technically, infinitives are a subclass of the class of non-finite verb forms, which also includes past participles and progressive particles). Another feature of the OI stage is that children often omit subjects from their sentences. That is, children will produce utterances such as throw ball from which the subject (I) is absent. While the proportion of infinitives is (considerably) higher than for adult speech, children in the OI stage do show competence regarding other syntactic attributes of the language. Typically, children will not make errors in the basic verb-object order. English-speaking children, for instance, will say throw ball, but not ball throw. One puzzling feature of the OI stage is that children produce both inflected and uninflected forms in contexts requiring the inflected form, but do not produce finite forms in nonfinite contexts. The fact that children use both inflected and uninflected forms shows that it is not the case that they simply don’t know the inflected forms. The optional infinitive stage has been shown to occur in many different languages, which can differ considerably in their underlying syntactic properties, and children do show competence regarding these syntactic properties. Different languages also differ with respect to how pronounced the OI stage is. Since most verb forms in English are not distinguishable from non-finite forms, it is relatively difficult to distinguish optional infinitives from grammatically correct utterances. In other languages (e.g. Dutch), the number of unambiguously finite forms is larger, and as a result the optional infinitive stage is more pronounced. Wexler (1998) has proposed a Nativist account of why children in the optional infinitive stage produce a large number of non-finite forms. In accordance with Chomsky’s theory of Universal Grammar (Chomsky 1981), he theorizes that children in the optional infinitive stage actually know the full grammar of the language. The only thing they do not know is that Agreement and Tense are obligatory. This approach accounts for the fact that children produce both correct finite forms and incorrect (optional) infinitives. It also explains why children rarely produce other types of errors. Finally, its great strength is that it unifies across languages where children clearly use optional infinitives despite differences in their underlying grammar. However, there are also a number of problems with Wexler’s account. Firstly, Wexler’s theory does not give a process account of developmental change in the use of optional infinitives. He assumes this to be due to maturation. Secondly, the theory makes very limited quantitative predictions. It only predicts that the optional infinitive stage occurs, and that children will stop making optional infinitive errors at some point. It makes no specific predictions regarding the time course of this development, or related changes in other attributes. Thirdly, the theory assumes a large amount of innate knowledge in the child (the theory assumes that the child does not know that inflection is obligatory, but otherwise knows the full grammar of the language). An obvious alternative to Wexler’s theory is that children learn the grammar of a language through exposure to that language. Wexler discounts this kind of learning-based approach on the grounds that the grammar is too difficult to learn, that the optional infinitive stage lasts too long (years), and that, although children produce both correct and incorrect forms, when they use finite forms, they use them correctly (Wexler, 1994). In this paper, we aim to show that the dynamics of the optional infinitive phenomenon can be simulated using a simple learning mechanism which performs a distributional analysis of naturalistic input. Earlier versions of the model have already been shown to simulate the basic optional infinitive phenomenon in both English (Croker, Pine & Gobet, 2001) and Dutch (Freudenthal, Pine & Gobet, 2001). Whereas the earlier versions modelled one specific stage in development, the present model aims to simulate the developmental change that is apparent in the use of optional infinitives. There are a number of reasons for choosing Dutch as the target language. Firstly, as was mentioned, in adult speakers’ Dutch, unambiguous finite forms are far more frequent than they are in English. In English, in the present tense, only the third person singular can be distinguished from the infinitive form. In Dutch, the first, second and third person singular are unambiguously finite. If, for instance, an English speaking child produced I throw ball, it would be unclear whether the verb throw was an infinitive form. The Dutch equivalent ik gooi bal would be classified as a finite form, because gooi is different from the infinitive gooien. Thus, the number of unambiguously finite forms is larger in Dutch than in English. (This suggests that developmental change in the use of optional infinitives is likely to be more pronounced in Dutch than it is in English, which makes the simulation of Dutch child language more informative as a modelling exercise.) A second reason for using Dutch is that detailed data regarding this development are available. Wijnen, Kempen & Gillis (2001) have analysed the corpora of two Dutch speaking children and have shown that the proportion of root infinitives decreases from around 90% to roughly 10% between the ages 1;6 and 3;0. By comparison, root infinitives are used in less than 10% of adults’ utterances. Wijnen et al. concluded that the frequency of occurrence of optional infinitives in the child’s speech was related to frequency, and utterance position, as well as lexical transparency. A third reason for choosing Dutch as the target language is that Dutch grammar is relatively complex when considering finiteness of verb forms. Dutch is what is known as an SOV/V2 language. This means that the verb in Dutch can take one of two positions, depending on its finiteness. A non-finite verb takes the sentence final position, whereas finite verbs take the second position. Therefore, in the sentence Ik gooi een bal (1) (I throw a ball) the verb gooi (throw) is finite and takes second position. In the construction Ik wil een bal gooien (2) (I want a ball throw/ I want to throw a ball) the verb gooien is a non-finite form, and takes sentence final position. (The auxiliary wil is finite and takes second position.) In English, which is an SVO language, verb position is not dependent on the finiteness of the verb. If a model is to learn from the distribution of naturalistic speech input, then the production of a large number of infinitives while respecting the overall grammar would appear to represent a greater challenge in Dutch than in English.