building natural language generation systems（建立自然语言生成系统）.pdfVIP

Building Natural Language Generation Systems Ehud Reiter Department of Computing Science University of Aberdeen King’s College Aberdeen AB9 2UE, BRITAIN 6 email: ereiter@csd.abdn.ac.uk 9 9 1 y 1 Introduction a M Natural Language Generation (NLG) systems generate texts in English (or other human lan- 2 guages, such as French) from computer-accessible data. NLG systems are (currently) most often used to help human authors write routine documents, including business letters [SBW91] and 1 weather reports [GDK94]. They also have been used as interactive explanation tools which v 2 communicate information in an understandable way to non-expert users, especially in software 0 engineering (eg, [RK92]) and medical (eg, [BMF+94]) contexts. 0 From a technical perspective, almost all applied NLG systems perform the following three 5 0 tasks [Rei94]: 6 9 Content Determination and Text Planning: Decide what information should be commu- / nicated to the user (content determination), and how this information should be rhetori- g l- cally structured (text planning). These tasks are usually done simultaneously. p m Sentence Planning: Decide how the information will be split among individual sentences and c paragraphs, and what cohesion devices (eg, pronouns, discourse markers) should be added : to make the text flow smoothly. v i X Realization: Generate the individual sentences in a grammatically correct manner. r a