Knowledge Graph Reasoning Based on Information Enhancement and Subgraph Alignment.

Abstract

Knowledge graph reasoning (KGR) is an important task in data mining. It aims to mine the logical rules based on the existing facts and further infer new facts, which makes the graph complete and accurate. Currently, with the development of large language models (LLMs), they are widely integrated with different baseline models for better performance. A few works are proposed on LLM-enhanced KGR models, which leaves many issues to be addressed. Inspired by the efficiency and accuracy of LLM in generating text semantic information, this article proposes a KGR method based on LLM information enhancement and subgraph alignment (LSA). LSA first utilizes LLM to generate textual descriptions corresponding to graph entities, relationships, and subgraphs. Then, it utilizes the generated textual attribute in both explicit and implicit ways: 1) explicit utilization, treating LLM-generated text features as the initialized features for the previous KGR model; and 2) implicit utilization, aligning the structural and textual information of key subgraphs via a learning mechanism. Finally, LSA is evaluated on three typical datasets. The promising performances demonstrate that our LSA leverages LLM to make the KG for richer information, and the representation learning model is empowered with better expressive ability.