Contextual rule learning for knowledge graphs using large language models

Abstract

Knowledge Graphs are crucial in developing knowledge-based AI systems. They infer knowledge using logical rules that describe the data in the knowledge graph. Rule miners aim to extract rules from knowledge bases using various techniques, primarily relying on the KB structured information to find co-occurrences and correlations between facts. Most rule miners, such as AMIE, cannot utilize the context of the facts and incorporate domain knowledge when mining rules, and they also face challenges with high resource utilization, leading to missed correlations and suboptimal rule discovery. This thesis explores the integration of Large Language Models (LLMs) into the rule mining process of AMIE to enhance rule quality, contextual awareness, and computational efficiency. The approach systematically integrates LLM usage to realize LLM-based rule initialization, evaluation, and refinement using Google Gemini. The integration balances rule quality and quantity improvements with computational feasibility. Extensive experiments and evaluations are conducted to assess the impact of the LLM on rule mining, specifically on the accuracy, completeness, and efficiency. Additionally, the experiments are designed to measure the impact of the LLM optimization and the prompt. The results show that incorporating LLMs leads to more rules generated, improving overall coverage of the rule mining process. However, LLM optimization and prompt clearness play a crucial role in the rule quality and computational efficiency.