Kucker, S. 1 , McMurray, B. 2, 3 & Samuelson, L. 2, 3
1 The University of Texas at Dallas
2 The University of Iowa
3 The Delta Center
In word learning, a form of statistical learning -word/object co-occurrence probabilities- has been posited as a way to overcome the ambiguity inherent in naming situations. However, it is clear that other mechanisms are at play and the process of mapping words to referents is complex and extended over time. The current study argues for a shift from a purely statistical model toward associative learning; while associative learning is clearly sensitive to statistical regularities between words and objects, it also shows considerable emergent complexity that accounts for the richness of children's lexical behavior. We present a Hebbian learning model that uses cross-situational statistics to acquire word-referent links over time, but also includes real-time competition to make in-the-moment decisions about the most likely referent of a word. This model simulates 18-24 month-old children's behavior during an ostensive naming episode, then during referent selection and retention. The model reveals both building and pruning of word-referent associations after a single naming event but also shows how the degree of change in those associations grows as the new words are subsequently used in referent selection. The results suggest that co-occurrence statistics alone are not enough to capture both in-the-moment and learning processes during lexical development.