Prose Generation from Expert Systems: An Applied Computational Linguistics Approach

refined in an iterative, incremental fashion. Even when a satisfactory draft evolves, an author frequently makes minor changes on each reading of the paper, particularly if the author sets the manuscript aside between readings. Compared to the manual process of writing a paper, the process of creating a computer program to generate prose can be even more frustrating and complex, particularly if the program must react to multiple conditions and still produce polished prose. Indeed, in a preliminary evaluation of the advice of the Roundsman system described later in this article, most of the corrections made by the domain expert involved rather subtle modifications of the system’s prose expression (Rennels 1987). The PROSENET/TEXTNET approach, designed to facilitate the incremental refinement of an expert system’s prose output, involves three components: (1) sentence-level and phrase-level ATNs to structure sentence-level and phrase-level prose expression, (2) expressive frames to structure paragraph-level content, and (3) paragraph-level ATNs to structure paragraph-level prose expression. This approach has already been used to implement several prototype expert systems in a variety of medical domains, including several systems developed at Yale to critique physician plans (Miller 1986), including ATTENDING, which critiques anesthetic management (Miller 1983a, 1983b, 1984); HT-ATTENDING, which critiques the management of essential hypertension (Miller and Black 1984); s expert systems are more frequently used in diverse areas of life, it becomes increasingly important that the interface between these systems and their users be as effective as possible. One component of this interface for many systems involves the generation of English prose to communicate the expert system’s conclusions and recommendations. A number of different approaches have been taken to the generation of natural language text by computer. As discussed in the next section, these approaches include the straightforward output of “canned” text, the automated translation of if-then inference rules, as well as systems that are more sophisticated from both a linguistic and semantic standpoint. This article describes the PROSENET/TEXTNET approach, which uses a linguistic formalism, the augmented transition network (ATN), to structure the generation of prose whose expression can vary widely depending on its content. The prose is pieced together from prose fragments that can vary in length from single words to whole sentences or even paragraphs, depending of the complexity and variability of the material being discussed. The approach is designed to allow the expert system builder great flexibility in polishing and refining the exact prose that will be output in different circumstances. The rationale for the approach can best be understood by considering the process of writing prose manually. Creating a polished paper usually involves multiple drafts, as the organization and prose expression is The PROSENET/TEXTNET approach is designed to facilitate the generation of polished prose by an expert system. The approach uses the augmented transition network (ATN) formalism to help structure prose generation at the phrase, sentence, and paragraph levels. The approach also uses expressive frames to help give the expert system builder considerable freedom to organize material flexibly at the paragraph level. The PROSENET /TEXTNET approach has been used in a number of prototype expert systems in medical domains, and has proved to be a convenient and powerful tool. A