Computational models of child language learning: an introduction.

This special issue showcases recent work on the computational modeling of child language acquisition. Together, these nine papers provide testimony to the scientific value of corpus-driven computational modeling. Each of the papers presents a clear, mechanistic model that can be tested, refined or rejected on the basis of publicly available data and/or replicable experiments. However, for many readers, the multiple formalisms and technicalities involved in this type of work can serve as barriers to evaluating the nature of the contributions being made. Therefore, it is my goal in this introduction to summarize what I see as the important take-home messages delivered by each of the nine projects. Hopefully, this overview will encourage the reader to turn then to the details of each of the nine contributions. Of the nine papers included here, eight base their analysis on the corpora of spontaneous adult–child interactions made available through the Child Language Data Exchange System (CHILDES). These CHILDES corpora provide two types of information crucial to the modeling enterprise. First, they document in detail the naturalistic development of language in the child. Second, these corpora provide a good sampling of the adult speech that serves as input to the child’s language learning mechanisms. Given these two empirical bases, the job of the computational modeler is to determine a set of algorithms that can take the child-directed speech (CDS) as input and produce the learner’s output (LO) at successive developmental levels. We can refer to this approach as input–output (I–O) modeling. In its simplest form, I–O modeling tends to view language learning as an emergent, data-driven process. However, there is ample room within this same computational framework for precise statements regarding the operation of non-emergentist innate constraints, parameters, principles and universals. Furthermore, complex features of situational context, social understandings and recent dialog history can, in principle, be quantified and coded as features of the input. More generally, as long as long as all these pieces of the Language Acquisition Device (LAD; Chomsky 1965) are fully specified, all