The machine-readable dictionary (MRD) is a major resource in the task of building a large-scale lexicon for a natural-language system. Because these dictionaries were written for human users, not computers, they show evidence of inconsistency of treatment, ambiguity within entries, and lack of systematicity in the explicit statement of essential facts. This ‘human bias’ makes dictionary mapping very difficult and of limited worth. Electronic text corpora now allow the evaluation of a dictionary entry by comparing it with actual word use in running text. This is done for several lexical items across five MRDs; corpus evidence is shown to support apparently incompatible semantic descriptions. An outline is given for part of a lexical database entry which could facilitate the mapping of dictionary entries and consequently maximize the useful facts to be extracted from a number of MRDs