Knowledge Graph Reasoning Based on Information Enhancement and Subgraph Alignment.

Area of Science:

• Artificial Intelligence
• Data Mining
• Natural Language Processing

Background:

• Knowledge Graph Reasoning (KGR) is crucial for data mining, aiming to infer new facts for graph completeness and accuracy.
• Large Language Models (LLMs) are increasingly integrated into baseline models, yet their application in KGR requires further exploration.
• Existing LLM-enhanced KGR models present challenges that necessitate innovative solutions.

Purpose of the Study:

• To propose a novel Knowledge Graph Reasoning (KGR) method, LSA, that leverages Large Language Models (LLMs) for enhanced information.
• To improve the accuracy and completeness of knowledge graphs by integrating LLM-generated textual descriptions.
• To address limitations in current LLM-enhanced KGR approaches through a combined strategy of information enhancement and subgraph alignment.

Main Methods:

• LSA utilizes LLMs to generate textual descriptions for graph entities, relationships, and subgraphs.
• Explicit utilization: LLM-generated text features are used as initial features for existing KGR models.
• Implicit utilization: A learning mechanism aligns the structural and textual information of key subgraphs.

Main Results:

• LSA was evaluated on three standard datasets, demonstrating promising performance.
• The method effectively enriches knowledge graphs (KGs) with information derived from LLMs.
• The representation learning model integrated with LSA shows improved expressive capabilities.

Conclusions:

• LSA successfully leverages LLMs to enrich knowledge graphs, leading to more informative representations.
• The subgraph alignment mechanism enhances the integration of structural and textual information.
• The proposed method offers a viable approach for advancing LLM-enhanced Knowledge Graph Reasoning.