[3Xin2] Poster session 1

[3Xin2-109] Conformance judgment and decision reason generation for rule sentences in natural language

〇Kiyoshiro Ariyama¹, Ichiro Kobayashi², Hirotoshi Taira¹ (1.Osaka Institute of Technology, 2.Ochanomizu University)

Keywords:Rule conformance determination, Large Language Model, Use of external knowledge, Explainable AI

Rule conformance judgment is crucial for guarantee that the actions of robots and computers are compliant with rules. In this study, we develop a system that automatically selects an appropriate rule when multiple external rules are given, and makes a rule conformance judgment while generating reasons for the rule conformance judgment, using the Japanese driver's license test as the subject matter. While large language models like GPT-3.5 can make reasoned judgments based on external rules, they face limitations with the number of rules that can be included in a prompt. We used LUKE, a knowledge-enhancing language model, to narrow down the rules before using a large language model for rule conformance judgments and reasoning. This approach improved the accuracy of both the judgments and the reasoning compared to not using LUKE.