PaperlayerView exact revision on arXiv

Source-linked AI summary

Training-Free BioNEL-RAG: Advancing Biomedical Entity Linking through Retrieval-Augmented Generation and Large Language Model

Ziyi Chen, Hui Liu, Ziyang Ling, Ronald X. Xu, Mingzhai Sun

arXiv:2610.14789v1cs.CL

TL;DR

Biomedical entity-linking methods often rely on domain-specific training, and their cross-domain effectiveness can decline. BioNEL-RAG retrieves candidate entities and uses an LLM to select one from mention context; it reports competitive or superior performance without task-specific training.

  • Problem

    Many biomedical entity-linking methods rely on domain-specific training, and their effectiveness can decline when applied across domains.

    Source: §I · p. 1

  • Method

    BioNEL-RAG retrieves candidate entities from an expandable biomedical knowledge base and prompts an LLM to select one using the mention context.

    Source: §I · p. 1

  • Results

    On NCBI-Disease, BioNEL-RAG exceeds SAPBERT-BioSYN by 0.7% in Acc@1, while the compared SOTA baselines were fine-tuned on their respective datasets.

    Source: 1) Comparison of BioNEL-RAG with other… · p. 3

  • Takeaways & Limitations

    BioNEL-RAG achieves competitive or superior performance without task-specific training, including on biomedical entity-linking datasets.

    Source: 1) Comparison of BioNEL-RAG with other… · p. 3, §IV · p. 4

Abstract

from arXiv · show

Biomedical Named Entity Linking (BioNEL) is the task of mapping mentions extracted from biomedical texts to canonical entities or concepts in the standardized knowledge base. Although existing methods have achieved remarkable performance in mapping accuracy, most rely on training with domain-specific datasets, which fundamentally limits their cross-domain adaptability. To address this limitation, we propose BioNEL-RAG, a novel training-free framework that combines retrieval-augmented generation (RAG) with large language models (LLMs). Within this framework, the RAG module focuses on retrieving candidate entities, while the large language model module plays the role of selecting the most appropriate candidate for entity normalization, which is based on the analysis of the mention's context, with the aid of its remarkable text comprehension capabilities. Experimental results demonstrate that BioNEL-RAG achieves performance comparable to existing methods without requiring training on subdomain datasets, whereas these methods necessitate fine-tuning on specific subdomains such as genes, diseases, or chemical compounds.

I. INTRODUCTION

Biomedical entity-linking models can lose effectiveness across domains, while LLMs risk hallucinating in specialized biomedical settings and fine-tuning demands substantial resources. BioNEL-RAG pairs retrieval with LLM-based candidate selection to deliver training-free performance and adapt to updated biomedical resources.

I. INTRODUCTION

  • Fine-tuned biomedical models can lose effectiveness across domains, while LLMs may hallucinate in specialized fields and fine-tuning requires substantial data and computation.The generalization of a fine-tuned model is often confined to its training dataset.Source: §I · p. 1
  • BioNEL-RAG retrieves candidate entities, then prompts an LLM to use mention context to select the normalized entity.Its expandable knowledge base is designed to adapt to updates from biomedical resources such as UMLS.Source: §I · p. 1

A. Datasets

The study covers biomedical entity linking across disease, chemical, gene, and social-media datasets. BioNEL-RAG builds an indexed knowledge base from a biomedical dictionary, retrieves candidates for each mention, and uses sentence context to guide normalization.

A. Datasets

  • The evaluation datasets cover disease, chemical, and gene entities, as well as social-media text.They include NCBI-Disease, BC5CDR-Disease and BC5CDR-Chemical, BC2GM-Gene, and SMM4H-2017.Source: A. Datasets · pp. 1–2
  • The framework separates dictionary handling and knowledge-base construction in Retrieval Augmentation from mention processing in Normalization.The figure caption names these as its two core components.Source: Figure 1 · p. 2

B. BioNEL Retrieval-Augmented Generation

  • BioNEL-RAG indexes biomedical dictionary entries as vectors, retrieves top-k candidates for a mention, and prompts an LLM with the candidates and sentence context to select the final entity.It skips text chunking because dictionary entries are concise and structured.Source: B. BioNEL Retrieval-Augmented Generation · p. 2

1) Task Definition:

BioNEL-RAG predicts a knowledge-base ID for each biomedical mention, using encoded dictionary entries to retrieve candidates and a prompt to select among them. Its key components reduce embedding dimensionality and use inner-product similarity search.

1) Task Definition:

  • BioNEL-RAG maps an input mention to its corresponding identifier, or CUI, while allowing multiple mentions to share one ID.The task is framed as predicting the ID associated with the mention.Source: 1) Task Definition · p. 2

2) Key Components:

  • PCA reduces high-dimensional dictionary embeddings to 384 dimensions, the selected trade-off for subsequent experiments.The method addresses redundancy and computational overhead for concise entity entries.Source: 2) Key Components · pp. 2–3
  • IndexFlatIP performs a linear scan using inner product similarity to retrieve the top-k candidate entities.The candidates are obtained from reduced embeddings of the mention and dictionary synonyms.Source: 2) Key Components · p. 3
  • The prompt asks the LLM to select the candidates most similar in meaning to the mention and follow a specified response format.It supplies an input mention, its context sentence, and a candidate list.Source: 2) Key Components · p. 3

1) Large Language Model:

The framework uses Qwen3-8B as its large language model, selected for its reported IFEval performance and multitask learning relative to similarly sized models.

1) Large Language Model:

  • Qwen3-8B [9] was selected for its reported IFEval performance and effective multitask learning compared with models of similar size.Source: 1) Large Language Model · p. 3

2) Prompt Design:

BioNEL-RAG prompts the LLM to select an entity from retrieved candidates rather than generate an answer from retrieved text. The mention, candidate list, and original context sentence support that selection.

2) Prompt Design:

  • The prompt uses the mention, retrieved candidates, and original context sentence to have the LLM select an entity rather than generate an answer.The passage identifies context as crucial for disambiguating the mention’s meaning.Source: 2) Prompt Design · p. 3

III. EXPERIMENTS AND RESULTS

BioNEL-RAG is competitive with or superior to fine-tuned SOTA baselines across the reported disease and chemical datasets. It requires no training on those datasets and outperforms two baselines on NCBI-Disease Acc@1.

III. EXPERIMENTS AND RESULTS

  • The comparison uses top-k accuracy, following the Acc definition introduced in BioSyn.Source: §III · p. 3

A. Model performance

1) Comparison of BioNEL-RAG with other models:

  • BioNEL-RAG outperforms SAPBERT-BioSYN by 0.7% and SAPBERT-PubMedBERT by 0.9% in NCBI-Disease Acc@1.On both BC5CDR datasets, it is comparable to SAPBERT-BioSYN and consistently outperforms SAPBERT-PubMedBERT; the baselines were fine-tuned on each dataset.Source: 1) Comparison of BioNEL-RAG with other… · p. 3

2) Compare with models on unseen datasets:

BioNEL-RAG generalizes to datasets excluded from baseline training, with reported Acc@1 gains over SAPBERT-BioSYN on both test sets. The ablations also report an embedding-model comparison.

2) Compare with models on unseen datasets:

1) Embedding Model:

  • The ablation study compares the general-purpose all-Mini-LM-v21 with domain-specific SAPBERT-PubMedBERT as embedding models.The supplied results report an 11.3% Acc@1 advantage for the domain-specific model.Source: 1) Embedding Model · pp. 3–4, Table III · p. 4

2) Large Language Models:

The ablation compares the full framework with a version lacking the LLM module. The LLM-equipped model scores higher on both reported accuracy metrics.

3) Embedding Distillation:

Embedding Distillation avoids performance degradation while slightly improving both accuracy metrics and reducing computational and storage costs. The conclusion describes BioNEL-RAG as achieving SOTA performance without training.

3) Embedding Distillation:

  • Embedding Distillation avoids performance degradation, slightly improves Acc@1 and Acc@5, and significantly reduces computational and storage costs.Source: 3) Embedding Distillation · p. 4

IV. CONCLUSION

  • BioNEL-RAG achieves SOTA performance without requiring training.Source: §IV · p. 4

Related papers

Loading 2610.14789v1…