Efficient Knowledge Infusion via KG-LLM Alignment

This paper addresses the limitations of knowledge graph-augmented methods for improving domain-specific performance in large language models (LLMs), particularly in biomedical applications. Key contributions include: 1. A novel LLM-driven approach to construct domain-specific knowledge graphs using a small labeled sample set and a large corpus, addressing knowledge mismatch and poor information compliance. 2. A three-stage KG-LLM alignment strategy that enhances LLMs' ability to utilize knowledge graphs, as demonstrated by improved performance in biomedical question-answering tasks with limited data. 3. The ELPF framework, which fine-tunes LLMs with K-LoRA, a "triples-to-text" adapter, and incorporates KG feedback for enhanced factual accuracy and knowledge compliance. 4. The study evaluates various models, such as K-LoRA fine-tuned LLMs, zero-shot querying, and retrieval-based methods, showing the effectiveness of the proposed method in CMedQA and BioASQ datasets. 5. ELPF outperforms other models in knowledge correctness, diversity, and domain awareness, with K-LoRA and AKGF contributing to its success. However, the model's performance is sensitive to the quality and completeness of the knowledge graph. In conclusion, the paper presents a framework that improves LLMs' performance in domain-specific tasks by constructing and aligning knowledge graphs, and highlights the importance of addressing knowledge mismatch and graph quality for optimal results. Future work will focus on refining the alignment process and addressing generalizability to other domains.