We present a computational (LSTM) model that learns to produce English (3sg and -bare) verb inflection when trained on English child-directed speech (CDS). The model is trained on input containing morphemized verbs and learns to predict the next token (word/morpheme) given a preceding sequence of tokens. The model produces the type of error (-bare for -3s) made by English-learning children while avoiding errors that children do not often make (-3s for -bare). The model also shows the same type of sensitivity to input statistics that has been reported in English-learning children. Finally, we manipulated the length of the sequences the model is trained on and show that this results in the delayed acquisition of -3sg forms that is characteristic of English-learning children with Developmental Language Disorder (DLD). Taken together these results suggest that input-driven learning is a major determinant of the patterns observed in both typical and impaired acquisition of English verb inflection.

Elicited imitation is a widely used method for testing a child's knowledge of a language for scientific or clinical purposes. A child hears an utterance and is asked to repeat what they have heard. While it is assumed that their fluency or speed in doing so is contingent on their linguistic competence, little is known about the cognitive mechanisms and/or representations involved. To explore this, we train an encoder-decoder model, consisting of recurrent neural networks, to encode and reproduce a corpus of child-directed speech and then test its performance on the experimental task of Bannard and Matthews (2008). In that study pre-school children were asked to repeat high- and low-frequency four-word sequences in which the first three words were identical (e.g., sit in your chair and sit in your truck) and the final words and bigrams were closely matched for frequency. We find that like those children our model makes more errors on the initial three words when they are part of a low-frequency than a high-frequency sequence, despite the fact that the words being repeated are identical. We explore why this might be and pinpoint some possible similarities between the model and child language processing.

We present a computational model of the acquisition of German case that is evaluated against empirical data obtained from naturalistic speech. The model substitutes nouns into existing contexts, and proceeds through a number of stages that reflect increasing knowledge on the part of a child, both of the determiner-noun sequences that are legal in German, and of the determiner-noun sequences that are appropriate in specific sentential contexts (cases). The model provides a natural account of gender and case errors, the two most common error types produced by children, and shows the highest error rates in dative contexts and lowest error rates in nominative contexts, as is true of children learning German. However, the model’s error rates in the early stages are considerably higher than those shown by children, suggesting that children possess a fairly sophisticated representation of how lexical contexts assign case from a relatively early age.