BioELX: Cross-lingual Biomedical Entity Linking via Alias-based Retrieval and LLM Ranking

BioELX is a new unsupervised framework for cross-lingual biomedical entity linking that eliminates the need for costly, task-specific training data. It improves upon existing SapBERT-based methods in two stages: first by enriching retriever training with multilingual aliases from Wikidata to better handle non-English mentions, and second by using a pre-trained LLM ranker for context-aware disambiguation. This approach establishes new state-of-the-art results across multiple benchmarks, demonstrating significant gains especially for low-resource languages.

A Paradigm Shift in Cross-lingual Biomedical Entity Linking

This research paper presents a novel methodological framework, BioELX, designed to overcome two major limitations in cross-lingual biomedical entity linking (BEL): the high cost of expert-annotated multilingual training data and the poor generalization of current retriever models. The core innovation is a two-stage, unsupervised pipeline that strategically leverages existing resources to achieve superior performance without task-specific supervised training. This represents a significant shift from the reliance on manually labeled datasets or English-centric knowledge base aliases that constrain prior systems.

Architectural Innovation: From Multilingual Retrieval to LLM-based Disambiguation

BioELX's framework addresses the failure points of conventional systems through a deliberate decomposition of the linking task.

• Stage 1: Enriched Multilingual Candidate Retrieval. The system enhances the standard SapBERT retriever, which is typically trained on English aliases, by augmenting its training with multilingual aliases derived from Wikidata. This directly targets the generalization gap for unseen non-English mentions.
• Stage 2: Context-aware Disambiguation with a Pre-trained LLM. For ranking and selecting the correct entity from candidates, BioELX employs a large language model (LLM) ranker. This component jointly analyzes the surrounding context of the mention and the candidate entities, moving beyond static string matching. Crucially, this ranking is performed in a zero-shot manner, eliminating the need for a supervised training phase on disambiguation data.

Empirical Validation and Quantifiable Impact

The paper substantiates its claims through experiments on five diverse benchmarks, demonstrating broad and substantial improvements. The results highlight the framework's effectiveness across different domains (general medical, pharmaceutical patents, clinical notes) and languages. The most striking gains are observed for low-resource languages within the XL-BEL benchmark, where BioELX achieves improvements of +30.8 Recall@1 on Thai, +22.1 on Korean, and +21.6 on Turkish. These substantial leaps underscore the method's success in mitigating the data scarcity problem that plagues these languages. Consistent gains on other benchmarks (+12.8 on German medical texts, +6.2 on European Medicines Agency documents, +5.4 on patents) confirm the robustness of the approach.