Distributional Semantic Models have been successful at predicting many semantic behaviors. The aim of this paper is to 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 in cognitive research, while research from computational linguistics on big datasets has found more success with prediction-based models. We explore differences between these types of lexical semantic models (as representatives of Hebbian vs. reinforcement learning mechanisms, respectively) within a more cognitively relevant context: the acquisition of semantic categories (e.g., apple and orange as fruit vs. soap and shampoo as bathroom items) from linguistic data available to children. We found that models that perform some form of abstraction outperform those that do not, and that co-occurrence-based abstraction models performed the best. However, different models excel at different categories, providing evidence for complementary learning systems.