Dynamic Knowledge-Aware LLM for Adverse Drug Reaction Entity Recognition

Yunzhi Qiu, Bo Zhang, Haohao Zhu, Changrong Min, Haifeng Liu, Tongxuan Zhang, Liang Yang, Hongfei Lin

Abstract: Adverse drug reaction (ADR) entity recognition is crucial for ensuring patient medication safety and remains a focus of sustained attention in biomedical research. However, distinct from general-domain entity recognition, ADR entity recognition from social media faces two key challenges: informal linguistic expressions and concept representation diversity. The existing methods mainly focus on introducing domain knowledge into traditional small language models. However, how to incorporate external knowledge into large language models remains a blank. To address these challenges, we propose DKLLM, a Dynamic Knowledge-aware Large Language Model specifically designed for ADR entity extraction. Our approach introduces the Unified Medical Language System (UMLS) as a biomedical knowledge base, where conceptual definitions enhance model understanding of informal social media expressions, while unique Concept Unique Identifiers (CUIs) help mitigate conceptual expression variability. Furthermore, we design a plug-and-play Dynamic Adapter module (D-Adapter) that incorporates gating mechanisms to filter noise from external knowledge. Finally, we explore two integration methods: D-Adapter-aware Attention Network(DAN) and D-Adapter-aware Feedforward Network(DFN). Experimental results demonstrate that DKLLM outperforms all baseline methods while updating fewer than 4% of parameters. The model achieves an average F1-score improvement of 3.27% across two public datasets.

External IDs:dblp:conf/isbra/QiuZZMLZYL25