Proceedings of the Annual Meeting of the Cognitive Science Society

Distributional Semantic Models have been successful atpredicting many semantic behaviors. The aim of this paper isto compare two major classes of these models – co-occurrence-based models, and prediction error-driven models– in learning semantic categories from child-directed speech.Co-occurrence models have gained more attention incognitive research, while research from computationallinguistics on big datasets has found more success withprediction-based models. We explore differences betweenthese types of lexical semantic models (as representatives ofHebbian vs. reinforcement learning mechanisms,respectively) within a more cognitively relevant context: theacquisition of semantic categories (e.g., apple and orange asfruit vs. soap and shampoo as bathroom items) from linguisticdata available to children. We found that models that performsome form of abstraction outperform those that do not, andthat co-occurrence-based abstraction models performed thebest. However, different models excel at different categories,providing evidence for complementary learning systems.