GS-KGC: A generative subgraph-based framework for knowledge graph completion with large language models

2024-08-20

Information Fusion

AI Summary

Knowledge graph completion (KGC) faces the challenge of achieving high-accuracy link prediction using only local subgraph information, without relying on global graph structure or extensive labeled data. Method: This paper proposes a generative subgraph modeling framework that introduces, for the first time, a subgraph-level prompt generation mechanism—integrating structural-aware encoding with semantic generation from large language models (LLMs) without fine-tuning. The approach supports zero-shot and few-shot KGC. Contributions/Results: Key innovations include structural-aware subgraph sampling and encoding, joint structure-text representation learning, and generative triple decoding. Evaluated on standard benchmarks (FB15k-237, WN18RR), the method significantly outperforms conventional embedding-based models and LLM fine-tuning approaches, achieves a 3× speedup in inference, and demonstrates strong generalization across unseen relations and sparse entities.

Technology Category

• Data Mining & Knowledge Management: Linked Open Data, Knowledge Graphs & KB Completion
• Natural Language Processing: Generation
• Machine Learning: Graph-based Machine Learning

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Problem

• Large Language Models
• Knowledge Graph Completion
• Limited Data Utilization

Innovation

• Generative Subgraph Strategy
• Question Answering Approach
• Large Language Models Integration