Proceedings of the Annual Meeting of the Cognitive Science Society

Do childrens early utterances reflect abstract syntactic knowledge or slot-filler formulae developed through word imita-tion? This study compares development of part-of-speech (POS) sequences with word sequences using language models(LMs) trained on mothers utterances (N=1,272,139) from CHILDES English corpora, in which POS tags are automaticallyassigned by MOR and POST programs (MacWhinney, 2000). Word-based and POS-based LM probabilities for childrensmulti-word utterances in the Providence corpus (Brschinger et al., 2013, 15-36 months, Nchildren=6, Nutterances=50,717)were calculated as a function of age. Word-based LM probability of childrens multi-word utterances first increases withage and then levels off after 23 months. By contrast, POS-based probability remains high and stable across all ages. Thissuggests children have adult-like syntactic knowledge even at a very early age when their word sequences are still notadult-like. The pattern of results supports the abstract syntax view. Additional studies will use more accurate POS-taggersand larger datasets.