This article aims to investigate the empirical validity of the types of lexical networks found in the cognitive literature which account for the various senses of a polysemous word by showing how they relate to each other through minimal changes. It sets out from an existing network analysis, namely Norvig and Lakoff's (1987) analysis of the verb 'take', and compares its structure with data obtained from corpora and a sentence production experiment. The assessment of the validity of Norvig and Lakoff's model is both linguistic (does the model reflect frequency as attested in naturally-occurring language?) and cognitive (does the model reflect salience in the ordinary language user's mind?). It is argued that the model Norvig and Lakoff propose does not stand up to empirical validation, but may still have a role to play in a theory of polysemy.