Can similarity-based models of induction handle negative evidence