Towards a Model of Prediction-based Syntactic Category Acquisition: First Steps with Word Embeddings

We present a prototype model, based on a combination of count-based distributional semantics and prediction-based neural word embeddings, which learns about syntactic categories as a function of (1) writing contextual, phonological, and lexical-stress-related information to memory and (2) predicting upcoming context words based on memorized information. The system is a first step towards utilizing recently popular methods from Natural Language Processing for exploring the role of prediction in childrens’ acquisition of syntactic categories. 1