ABSTRACT
This paper presents an analysis of the role of input size and generativity (ability to produce novel utterances) in simulating developmental data on a phenomenon in first language acquisition. An existing model that has already simulated the basic phenomenon is trained on input sets of varying sizes (13,000 to 40,000 utterances). The ability of the model to produce novel utterances is also manipulated. Both input size and generativity affect the fits for later stages of development. Higher generativity improves fits for later stages, but worsens them for early stages, suggesting generativity is best increased as a function of mean length of utterance (MLU). The effect of training set is variable. Results are discussed in terms of optimal training sets for simulations, and children's developing ability to produce utterances beyond the input they have heard.
