KG-Adapter: Enabling Knowledge Graph Integration in Large Language Models through Parameter-Efficient Fine-Tuning

Published in Findings of ACL 2024. KG-Adapter is a parameter-efficient fine-tuning approach that integrates knowledge graphs into decoder-only LLMs: an adapter encodes KGs from node- and relation-centric views and performs joint reasoning with the LLM end-to-end. Training only 28M parameters, a 7B model outperforms the previous fully fine-tuned SOTA and matches prompt-based ChatGPT.

ACL Anthology

Recommended citation: Tian, S., Luo, Y., Xu, T., Yuan, C., Jiang, H., Wei, C., & Wang, X. (2024). "KG-Adapter: Enabling Knowledge Graph Integration in Large Language Models through Parameter-Efficient Fine-Tuning." In Findings of the Association for Computational Linguistics: ACL 2024, pages 3813–3828.
Download Paper