1997 年 12 巻 6 号 p. 861-869
Programs that deal with high level knowledge processing tend to get large and complicated, and it is difficult to begin by constructing a system that has all the required functions. Hence, it is necessary to progressively improve the system by gradually adding to its capabilities. The improvements are not always new functions but may be new reasoning methods or faster processing schemes. The problem is added cost. To completely rewrite the system or make major changes becomes very expensive. Therefore, such a construction method is important that permits the addition of capabilities or an improvement in efficiency by a small change of programs. However, various existing program construction methods and programming languages may not be suitable for this kind of successive program improvement. In this paper, we propose a system construction method based on equivalent transformation as a knowledge processing system construction method by successive improvement. We use a computational framework called "Rule Based Equivalent Transformation". This uses the formalism of translating a given problem into a declarative program and solving the problem by repeatedly applying equivalent transformation rules to the program. We constructed three natural language understanding systems for the domain of Japanese chess: first a prototype system "A" , and then two improved systems "B" and "C". We improved the efficiency of these systems by adding new rules, and the results clearly showed that, if rule based program transformation is used as a design method, the system's modularity is increased and, thus, system can be easily improved.