Toward Fuzzy Reasoning with Large Language Models

Abstract

The objective of this research is to incorporate quantitative inference (especially fuzzy infer-(breakpoint)ence) capabilities for real-world environments into large-scale language models. Given the environment and the state of objects when two objects collide, we show that the post-collision state can be inferred in natural language according to this situation. We used T5 to achieve inference with natural language that captures contextual information by using a natural language description of the object’s state before the collision as input and the post-collision state as output. The quantitative changes in the physical state of objects are described in vague expressions in natural language sentences, and by simultaneously inferring membership functions corresponding to these vague expressions, it became possible to connect the real world and the language. We also confirmed that the accuracy of inference is improved by dividing the learning steps into multiple steps.