Agrammatic aphasics demonstrate sensitivity to syntactic structure although they generally fail to make adequate use of structural information in producing and comprehending sentences. Under the 'mapping hypothesis' (cf. Schwartz et at. 1987), the agrammatic performance pattern is taken to reflect impaired mapping between grammatical constituents (subject, object) and thematic roles (agent, theme). We describe a treatment programme ('mapping therapy') directed at the remediation of mapping operations, and report results of a study designed to assess the effects of this intervention on comprehension and production. Eight chronic non-fluent aphasics were trained to identify the verb, agent, and patient/theme in sentences presented in a combined writtenspoken format. Three types of sentences were used: active transitives with action verbs (Type A); active transitives with experiential verbs (Type B), and non-canonical sentence types (e.g. passives) with action verbs (Type C). The first phase of training utilized Type A sentences, the second Type B, the third Type C. Generalization to the untrained sentence types was tested by means of multiple-baseline/generalization probes. Generalization to standard comprehension and production tasks was assessed through a comprehensive Language Assessment Battery. Acquisition and generalization profiles differed across subjects. However, following training most patients showed improved scores on predicted measures of sentence production, and two patients also showed improvement on syntactic comprehension tests. The best outcomes were seen in patients with relatively pure agrammatism. Those with more severe and more complicated impairments had poorer outcomes, in part, we believe, because of the resource demands of this particular training task.