Modeling law text as Knowledge Graph Inference rules: Towards explainable reasoning systems

Abstract

With the development of deep learning architectures and Large Language Models (LLMs), various attempts to develop question answering and decision support systems for both general purpose as well as domain-specific applications have been made. One special domain that presents significant challenges for such systems is the legal domain, where systems are required to process documents containing highly complex expressions and logic. Additionally, the high-stakes nature of the domain sets additional requirements for systems not only to be accurate and reliable but also to be explainable and traceable. To address this issue, this thesis explores and proposes a novel information modeling approach, where the target is to extract Knowledge Graph (KG) inference rules from law text. A comprehensive experimental process is conducted to demonstrate and evaluate the compatibility of this approach with LLMs, and thereby determine the applicability of this new information modeling approach. Experimental results show that LLMs (represented by OpenAI's ChatGPT-5.3 model) exhibit reliable performance on the new extraction task, which highlights promising potential for further developments toward large-scale applications.