Generating concise natural language summaries

Abstract

Summaries typically convey maximal information in minimal space. In this paper, we describe an approach to summary generation that opportunistically folds information from multiple facts into a single sentence using concise linguistic constructions. Unlike previous work in generation, how information gets added into a summary depends in part on constraints from how the text is worded so far. This approach allows the construction of concise summaries, containing complex sentences that pack in information. The resulting summary sentences are, in fact, longer than sentences generated by previous systems. We describe two applications we have developed using this approach, one of which produces summaries of basketball games (STREAK) while the other (PLANDOC) produces summaries of telephone network planning activity; both systems summarize input data as opposed to full text. The applications implement opportunistic summary generation using complementary approaches. STREAK uses revision, creating a draft of essential facts and then using revision rules constrained by the draft wording to add in additional facts as the text allows. PLANDOC uses discourse planning, looking ahead in its text plan to group together facts which can be expressed concisely using conjunction and deleting repetitions. In this paper, we describe the problems for summary generation, the two domains, the linguistic constructions that the systems use to convey information concisely and the textual constraints that determine what information gets included.