Knowledge about the rhetorical structure of a text is useful for automatic abstraction. We are interested in the automatic extraction of rhetorical units from the source text, units such as PROBLEM STATEMENT, CONCLUSIONS and RESULTS. We want to use such extracts to generate high-compression abstracts of scientific articles. In this paper, we present an extension of Kupiec, Pealersen and Chen's (1995) methodology for trainable statistical sentence extraction. Our extension addition-ally classifies the extracted sentences according to their rhetorical role.