An Approach to Extracting Inductive Inference Processes on Natural Language

Abstract

In natural language reasoning, conclusion is derived from a premise through a reasoning process. In doing so, we want to draw conclusions based on the knowledge recalled from the premises. In this study, we propose a natural language inference method that draws better conclusions from premises by considering the knowledge recalled to draw conclusions as sentences themselves that can express complex contents. Specifically, we develop a method to generate inference processes that are possible with a large language model based on a question-and-answer corpus, and to learn the model so that it can derive correct answers. In the learning phase, data is collected on what inference process led to the conclusion, based on the probability of generating the language model. This is used to build a model that generates knowledge from premises and a natural language inference model that uses premises and generated knowledge to draw conclusions.