How do people come to assign symbolic labels to continuous dimensions? Previous work has shown that prediction-error-driven models are sensitive to the order of labels and exemplars during training; similar patterns of learning are found found in adult learners trained to associate labels with discrete visual stimuli. Here we provide further evidence in support of the hypothesis that an error-driven mechanism underlies word learning, using continuous stimuli to explore the interactions of temporal structure, stimulus frequency, and distinctiveness in shaping associative learning. We conclude that learning to use features of exemplars to predict labels results in over-representation of diagnostic information, as shown by improved associative performance on stimuli near category boundaries. This is consistent with an error-driven model of label acquisition, and highlights the importance of the associative and prediction-based (rather than exclusively syntactic) aspects of symbolic cognition.