GENERATING AND REVISING TEXT - A FULLY KNOWLEDGE-BASED APPROACH

Traditional natural language generation systems use separate knowledge bases with entirely different representational methods for the different aspects and stages of language production. In doing so, they segregate much of their ''knowledge'' from the process, burying it in algorithms, and making it impossible to reason simultaneously with the system's actual working representation of information relating to different parts of the task. Some researchers have adopted this approach out of expediency; but others have argued that it is necessary if the problem is to be rendered tractable. This report argues that revision is important to producing high-quality text, and that the segregated knowledge-base approach is inappropriate for natural language generation systems that incorporate revision. We describe the system Kalos, a natural language generation system that uses revision, with particular attention to its uniform knowledge base. We demonstrate techniques using that knowledge base to improve text generation efficiency. A detailed example from Kalos is included.