Paper Type
Short
Paper Number
PACIS2026-1353
Description
Retrieval-Augmented Generation (RAG) has emerged as an effective approach for grounding large language models with external knowledge. However, existing biomedical RAG systems primarily rely on semantic similarity and often fail to preserve complex biomedical entity relationships, leading to semantic noise and unstable performance under deeper retrieval settings. To address this limitation, we propose ECR-KG, a structure-aware retrieval framework that combines entity-centric query rewriting with knowledge-graph-based structural reranking. The framework first expands biomedical queries into entity-aware subqueries to improve retrieval coverage and then applies Graph Edit Distance (GED)-based reranking to evaluate relational consistency between queries and retrieved passages. Experiments on MedMCQA and MedQA demonstrate that ECR-KG improves retrieval robustness and downstream QA accuracy compared to standard RAG baselines, particularly under larger retrieval depths.
Recommended Citation
Chang, Che-Jui; Wei, Cheng-Ru; and Lee, Wei-Po, "Entity-Centric Retrieval-Augmented Generation with Knowledge-Graph Re-ranking for Biomedical Question Answering" (2026). PACIS 2026 Proceedings. 4.
https://aisel.aisnet.org/pacis2026/ishealthcare/ishealthcare/4
Entity-Centric Retrieval-Augmented Generation with Knowledge-Graph Re-ranking for Biomedical Question Answering
Retrieval-Augmented Generation (RAG) has emerged as an effective approach for grounding large language models with external knowledge. However, existing biomedical RAG systems primarily rely on semantic similarity and often fail to preserve complex biomedical entity relationships, leading to semantic noise and unstable performance under deeper retrieval settings. To address this limitation, we propose ECR-KG, a structure-aware retrieval framework that combines entity-centric query rewriting with knowledge-graph-based structural reranking. The framework first expands biomedical queries into entity-aware subqueries to improve retrieval coverage and then applies Graph Edit Distance (GED)-based reranking to evaluate relational consistency between queries and retrieved passages. Experiments on MedMCQA and MedQA demonstrate that ECR-KG improves retrieval robustness and downstream QA accuracy compared to standard RAG baselines, particularly under larger retrieval depths.
Comments
14-Healthcare