Six groups of subjects rated word pairs for the degree to which they exemplified one of six semantic relationships. The relationships that subjects were instructed to rate were antonymy, synonymity, subordination, superordination, coordination, and similarity. Stimulus pairs represented antonyms, synonyms, subordinates, superordinates, and coordinates. The pairs representing each stimulus relationship varied across four levels of typicality, ranging from good examples of the relationship to unrelated pairs. The highest rating in each group was given to the stimulus relationship corresponding to the relationship being judged (e.g., antonyms received the highest rating under antonym judgment instructions). This interaction was strongest for high-typicality pairs and decreased across the levels of typicality. Semantic decision models cannot explain these results unless the models are modified so that decisions are based on relationship similarity, the degree to which a stimulus pair exemplifies the relationship subjects are instructed to judge.