Large Language Model and Knowledge Graph Entangled Logical Reasoning

Recently, the integration of large language models (LLMs) and knowledge graphs (KGs) is a research focal point, since they have complementary strengths and weaknesses for logical reasoning. To be specific, LLMs exhibit strong semantic reasoning capacities and have proven to be effective in various reasoning tasks, but they are prone to producing hallucinations due to the lack of world knowledge and structured reasoning abilities. Although KGs alleviate this problem by providing extensive factual knowledge to enhance reasoning accuracy, they are limited in language understanding and reasoning flexibility. Thus, in this paper, we propose a novel framework LKLR that entangles both LLMs and KGs for synergistic reasoning. Firstly, hypothesizing LLMs contain logical structures, we decompose the question into a grounded logical query over the KG by leveraging the chain-of-thought reasoning ability of the LLM. It enables the seamless integration of LLMs and KGs. Then we traverse the query sequentially on the KG to ground each reasoning step in a factual knowledge graph, while maintaining the reasoning flow guided by the LLM. Finally, we enhance the inference ability of the LLM by supplying the grounded knowledge, combining the robust knowledge of KGs with the semantic reasoning of LLMs. Our approach synergistically integrates neural and symbolic reasoning to achieve hybrid reasoning capabilities. We conduct experiments on several question answering benchmarks, including WebQSP, QALD10-en, WebQuestions and T-Rex, our method respectively improves the performance on the Hits@1 metric by 4.5%, 6.6%, 12.3%, and 5.7%, compared to ToG which performs reasoning based on the LLM and the retrieved knowledge from the KG. These results demonstrate that our proposed framework can provide transparent and reliable reasoning.