1993 年 8 巻 5 号 p. 639-648
To generate good text, many kinds of decisions should be made. Many researchers have been engaged in searching for both an architecture that would determine a proper order for these decisions and heuristics that would make an appropriate decision locally at each decision point. However, even if such an architecture or heuristics were found, there are still certain kinds of problems that are difficult to consider during the generation process. Those problems include, for example, structural ambiguity and sentence complexity. Some of them can be easily detected and solved by introducing a revision process after generation. In this paper, we argue the importance of text revision with respect to natural language generation, and adopt a computational model of text revision. We also discuss its implementation issues and introduce dependency directed backtracking in order to realize efficient revisions. Our model can be implemented easily because the revision process allows the system to make a decision at each point without any anticipation of the future decisions. Finally, we evaluate our method on an experimental Japanese text generation system.