A Statistical Model of Developmental Changes in Early Word Learning

Young language learners are able to map a word onto its ref-erent from an infinite number of possible word-to-world map-pings at early stage and further this learning ability turns to be much more effective at the later stage. What mechanisms underlie behavioral changes during word learning? This pa-per presents a developmental model of statistical associations, suggesting that as far as human minds are equipped with gen-eral associative learning mechanisms, ambiguity in the word-learning situation can be significantly reduced by recruiting cu-mulative lexical knowledge in statistical computations. Conse-quently, this leads to increasingly fast subsequent learning and abrupt behavioral changes, such as the vocabulary spurt and fast mapping. We provide a formal account of this argument by developing a computational model fed with the data col-lected from a series of picture-book reading sessions to sim-ulate word acquisition in natural contexts. The results show that previously learned lexical knowledge can not only nar-row the search space but also bootstrap the subsequent learn-ing process, which greatly improves learning results without the change of the underlying learning mechanism. Hence, this work suggests that using lexical knowledge accumulated in subsequent statistical learning is a scaffold for vocabulary growth in early language acquisition. 1