Reproduce “Enhancing Large Language Model for Knowledge Graph Completion via Structure-Aware Alignment-Tuning” on OpenResearch

About this paper

Yu Liu, Yanan Cao, Xixun Lin, Yanmin Shang, Shi Wang, Shirui Pan

Structure-aware Alignment-Tuning (SAT) was developed to enhance Large Language Models for Knowledge Graph Completion by bridging the representational gap between graph embeddings and natural language through hierarchical alignment and a unified structural instruction tuning framework. It achieved up to 29.8% relative Hits@1 improvement in link prediction and 5.1% relative F1-score improvement in triple classification on benchmark datasets.

Key insights

• Aligning structured graph representations with natural language semantics at both local (node) and global (subgraph) levels is crucial for effective LLM reasoning on KGs.
• A unified instruction tuning framework, which incorporates structured graph information directly into prompts, can efficiently adapt LLMs to various KGC tasks.
• Providing rich structural context via query subgraphs enables LLMs to infer complex, implicit relationships more accurately than relying solely on verbalized or surface-level textual cues.

Summary by alphaXiv.