Daniel Freudenthal

Modeling the Developmental Patterning of Finiteness Marking in English, Dutch, German, and Spanish Using MOSAIC

In this study, we apply MOSAIC (model of syntax acquisition in children) to the simulation of the developmental patterning of childrens optional infinitive (OI) errors in 4 languages: English, Dutch, German, and Spanish. MOSAIC, which has already simulated this phenomenon in Dutch and English, now implements a learning mechanism that better reflects the theoretical assumptions underlying it, as well as a chunking mechanism that results in frequent phrases being treated as 1 unit. Using 1, identical model that learns from child-directed speech, we obtain a close quantitative fit to the data from all 4 languages despite there being considerable cross-linguistic and developmental variation in the OI phenomenon. MOSAIC successfully simulates the difference between Spanish (a pro-drop language in which OI errors are virtually absent) and obligatory subject languages that do display the OI phenomenon. It also highlights differences in the OI phenomenon across German and Dutch, 2 closely related languages whose grammar is virtually identical with respect to the relation between finiteness and verb placement. Taken together, these results suggest that (a) cross-linguistic differences in the rates at which children produce OIs are graded, quantitative differences that closely reflect the statistical properties of the input they are exposed to and (b) theories of syntax acquisition need to consider more closely the role of input characteristics as determinants of quantitative differences in the cross-linguistic patterning of phenomena in language acquisition.

Modeling the Development of Children's Use of Optional Infinitives in Dutch and English Using MOSAIC

In this study we use a computational model of language learning called model of syntax acquisition in children (MOSAIC) to investigate the extent to which the optional infinitive (OI) phenomenon in Dutch and English can be explained in terms of a resource-limited distributional analysis of Dutch and English child-directed speech. The results show that the same version of MOSAIC is able to simulate changes in the pattern of finiteness marking in 2 children learning Dutch and 2 children learning English as the average length of their utterances increases. These results suggest that it is possible to explain the key features of the OI phenomenon in both Dutch and English in terms of the interaction between an utterance-final bias in learning and the distributional characteristics of child-directed speech in the 2 languages. They also show how computational modeling techniques can be used to investigate the extent to which cross-linguistic similarities in the developmental data can be explained in terms of common processing constraints as opposed to innate knowledge of universal grammar.

Explaining quantitative variation in the rate of Optional Infinitive errors across languages: A comparison of MOSAIC and the Variational Learning Model

In this study, we use corpus analysis and computational modelling techniques to compare two recent accounts of the OI stage: Legate & Yangs (2007) Variational Learning Model and Freudenthal, Pine & Gobets (2006) Model of Syntax Acquisition in Children. We first assess the extent to which each of these accounts can explain the level of OI errors across five different languages (English, Dutch, German, French and Spanish). We then differentiate between the two accounts by testing their predictions about the relation between childrens OI errors and the distribution of infinitival verb forms in the input language. We conclude that, although both accounts fit the cross-linguistic patterning of OI errors reasonably well, only MOSAIC is able to explain why verbs that occur more frequently as infinitives than as finite verb forms in the input also occur more frequently as OI errors than as correct finite verb forms in the childrens output.

Simulating the referential properties of Dutch, German and English Root Infinitives in MOSAIC

Children learning many languages go through an optional infinitive (OI) stage in which they produce nonfinite verb forms in contexts in which a finite verb form is required (e.g., “That go there” instead of “That goes there”). MOSAIC (Model of Syntax Acquisition in Children) is a computational model of language learning that successfully simulates the developmental patterning of the OI phenomenon in English, Dutch, German, and Spanish (Freudenthal, Pine, Aguado-Orea, & Gobet, 2007). In the present study, MOSAIC is applied to the simulation of certain subtle but theoretically important phenomena in the cross-linguistic patterning of the OI phenomenon that are typically assumed to require a more complex formal analysis. MOSAIC is shown to successfully simulate (a) the Modal Reference Effect, the finding that Dutch and German children tend to use root infinitives (RIs) in modal contexts; (b) the Eventivity Constraint, the finding that Dutch and German RIs refer predominantly to actions rather than static situations; and (c) the absence or reduced size of these effects in English. These results provide strong support for input-driven explanations of the Modal Reference Effect as well as MOSAICs mechanism for producing RIs, and the wider claim that it is possible to explain key aspects of childrens early multiword speech in terms of the interaction between a resource-limited distributional learning mechanism and the surface properties of the language to which children are exposed.

On the resolution of ambiguities in the extraction of syntactic categories through chunking

In recent years, several authors have investigated how co-occurrence statistics in natural language can act as a cue that children may use to extract syntactic categories for the language they are learning. While some authors have reported encouraging results, it is difficult to evaluate the quality of the syntactic categories derived. It is argued in this paper that traditional measures of accuracy are inherently flawed. A valid evaluation metric needs to consider the well-formedness of utterances generated through a production end. This paper attempts to evaluate the quality of the categories derived from co-occurrence statistics through the use of MOSAIC, a computational model of syntax acquisition that has already been used to simulate several phenomena in child language. It is shown that derived syntactic categories that may appear to be of high quality quickly give rise to errors that are not typical of child speech. A solution to this problem is suggested in the form of a chunking mechanism that serves to differentiate between alternative grammatical functions of identical word forms. Results are evaluated in terms of the error rates in utterances produced by the system as well as the quantitative fit to the phenomenon of subject omission.

Avoiding dative overgeneralisation errors: semantics, statistics or both?

How do children eventually come to avoid the production of overgeneralisation errors, in particular, those involving the dative (e.g., *I said her “no”)? The present study addressed this question by obtaining from adults and children (5–6, 9–10 years) judgements of well-formed and over-general datives with 301 different verbs (44 for children). A significant effect of pre-emption—whereby the use of a verb in the prepositional-object (PO)-dative construction constitutes evidence that double-object (DO)-dative uses are not permitted—was observed for every age group. A significant effect of entrenchment—whereby the use of a verb in any construction constitutes evidence that unattested dative uses are not permitted—was also observed for every age group, with both predictors also accounting for developmental change between ages 5–6 and 9–10 years. Adults demonstrated knowledge of a morphophonological constraint that prohibits Latinate verbs from appearing in the DO-dative construction (e.g., *I suggested her the trip). Verbs’ semantic properties (supplied by independent adult raters) explained additional variance for all groups and developmentally, with the relative influence of narrow- vs broad-range semantic properties increasing with age. We conclude by outlining an account of the formation and restriction of argument-structure generalisations designed to accommodate these findings.

Do young children have adult-like syntactic categories? Zipf's law and the case of the determiner

Generativist models of grammatical development assume that children have adult-like grammatical categories from the earliest observable stages, whereas constructivist models assume that childrens early categories are more limited in scope. In the present paper, we test these assumptions with respect to one particular syntactic category, the determiner. This is done by comparing controlled measures of overlap in the set of nouns with which children and their caregivers use different instances of the determiner category in their spontaneous speech. In a series of studies, we show, first, that it is important to control for both sample size and vocabulary range when comparing child and adult overlap measures; second, that, once the appropriate controls have been applied, there is significantly less overlap in the nouns with which young children use the determiners a/an and the in their speech than in the nouns with which their caregivers use these same determiners; and, third, that the level of (controlled) overlap in the nouns that the children use with the determiners a/an and the increases significantly over the course of development. The implication is that children do not have an adult-like determiner category during the earliest observable stages, and that their knowledge of the determiner category only gradually approximates that of adults as a function of their linguistic experience.

Understanding the Developmental Dynamics of Subject Omission: The Role of Processing Limitations in Learning

P. Blooms (1990) data on subject omission are often taken as strong support for the view that child language can be explained in terms of full competence coupled with processing limitations in production. This paper examines whether processing limitations in learning may provide a more parsimonious explanation of the data without the need to assume full competence. We extended P. Blooms study by using a larger sample (12 children) and measuring subject omission phenomena in three developmental phases. The results revealed a Verb Phrase-length effect consistent with that reported by P. Bloom. However, contrary to the predictions of the processing limitations account, the proportion of overt subjects that were pronominal increased with developmental phase. The data were simulated with MOSAIC, a computational model that learns to produce progressively longer utterances as a function of training. MOSAIC was able to capture all of the effects reported by P. Bloom through a resource-limited distributional analysis of child-directed speech. Since MOSAIC does not have any built-in linguistic knowledge, these results show that the phenomena identified by P. Bloom do not constitute evidence for underlying competence on the part of the child. They also underline the need to develop more empirically grounded models of the way that processing limitations in learning might influence the language acquisition process.

Why computational models are better than verbal theories: the case of nonword repetition.

Tests of nonword repetition (NWR) have often been used to examine childrens phonological knowledge and word learning abilities. However, theories of NWR primarily explain performance either in terms of phonological working memory or long-term knowledge, with little consideration of how these processes interact. One theoretical account that focuses specifically on the interaction between short-term and long-term memory is the chunking hypothesis. Chunking occurs because of repeated exposure to meaningful stimulus items, resulting in the items becoming grouped (or chunked); once chunked, the items can be represented in short-term memory using one chunk rather than one chunk per item. We tested several predictions of the chunking hypothesis by presenting 5-6-year-old children with three tests of NWR that were either high, medium, or low in wordlikeness. The results did not show strong support for the chunking hypothesis, suggesting that chunking fails to fully explain childrens NWR behavior. However, simulations using a computational implementation of chunking (namely CLASSIC, or Chunking Lexical And Sub-lexical Sequences In Children) show that, when the linguistic input to 5-6-year-old children is estimated in a reasonable way, the childrens data are matched across all three NWR tests. These results have three implications for the field: (a) a chunking account can explain key NWR phenomena in 5-6-year-old children; (b) tests of chunking accounts require a detailed specification both of the chunking mechanism itself and of the input on which the chunking mechanism operates; and (c) verbal theories emphasizing the role of long-term knowledge (such as chunking) are not precise enough to make detailed predictions about experimental data, but computational implementations of the theories can bridge the gap.

Preemption versus Entrenchment: Towards a Construction-General Solution to the Problem of the Retreat from Verb Argument Structure Overgeneralization.

Participants aged 5;2-6;8, 9;2-10;6 and 18;1-22;2 (72 at each age) rated verb argument structure overgeneralization errors (e.g., *Daddy giggled the baby) using a five-point scale. The study was designed to investigate the feasibility of two proposed construction-general solutions to the question of how children retreat from, or avoid, such errors. No support was found for the prediction of the preemption hypothesis that the greater the frequency of the verb in the single most nearly synonymous construction (for this example, the periphrastic causative; e.g., Daddy made the baby giggle), the lower the acceptability of the error. Support was found, however, for the prediction of the entrenchment hypothesis that the greater the overall frequency of the verb, regardless of construction, the lower the acceptability of the error, at least for the two older groups. Thus while entrenchment appears to be a robust solution to the problem of the retreat from error, and one that generalizes across different error types, we did not find evidence that this is the case for preemption. The implication is that the solution to the retreat from error lies not with specialized mechanisms, but rather in a probabilistic process of construction competition.