Comparing Knowledge Graph Completion and Question Answering Accuracy in Legal Information Portals

Abstract

The rapid digitization of legal documents has necessitated the development of advanced information retrieval systems, specifically legal information portals, to facilitate efficient access to statutes, case law, and legal precedents. Knowledge graphs have emerged as a foundational technology within these portals, enabling semantic search and complex reasoning capabilities. However, legal knowledge graphs inherently suffer from incompleteness and noise, which directly impacts the accuracy of downstream applications such as question answering. This paper presents a comprehensive benchmark study evaluating the performance of various knowledge graph completion algorithms and their subsequent effect on question answering accuracy within the context of legal information portals. We construct a specialized legal dataset encompassing thousands of entities and relations derived from real-world judicial records and legislative texts. By systematically evaluating translation-based, tensor factorization, and neural network-based completion models, we establish a baseline for link prediction in the legal domain. Furthermore, we investigate how improvements in knowledge graph completeness correlate with the precision and recall of natural language question answering systems deployed over these graphs. The findings offer critical insights into the bottleneck of current legal question answering architectures and provide a standardized evaluation framework for future research in legal informatics.