Proceedings of the Annual Meeting of the Cognitive Science Society

This paper investigates how children may acquire an incremental grammar that supports word-by-word speech production via computational modeling. We formalize language acquisition as the learning of a probabilistic, bidirectional mapping between utterances and intended meanings. Our first contribution is the design of a precise incremental grammar grounded in Synchronous Hyperedge Replacement Grammar. The grammar constrains syntactico-semantic analyses to be strictly left-branching, reflecting the incremental nature of human language comprehension. We then conduct statistical production-oriented grammar induction over naturalist data. For the meaning-to-text generation task, the incremental grammar yields slightly lower semantic constituent recognition accuracy but unexpectedly higher language generation quality, compared to a hierarchical grammar. It also exhibits a small and consistent accuracy gap of semantic constituent recognition relative to a supervised oracle, an effect not observed for the hierarchical grammar. Together, these results suggest a new method for comparing alternative grammatical architectures in silico.