Entity linking (EL) for biomedical text is typically benchmarked on English-only corpora with flat mentions , leaving the more realistic scenario of nested and multilingual mentions largely unexplored. We present our system for the BioNNE 2025 Multilingual Biomedical Nested Named Entity Linking shared task (English & Russian), closing this gap with a lightweight pipeline that keeps the original EL model intact and modifies only three task-aligned components: Two-stage retrieval-ranking. We leverage the same base encoder model in both stages: the retrieval stage uses the original pre-trained model, while the ranking stage applies domainspecific fine-tuning. Boundary cues. In the ranking stage, we wrap each mention with learnable [Ms] / [Me] tags, providing the encoder with an explicit, language-agnostic span before robustness to overlap and nesting. Dataset augmentation. We also automatically expand the ranking training corpus with three complementary data sources, enhancing coverage without extra manual annotation. On the BioNNE 2025 leaderboard, our two stage system, bilingual bert (BIBERT-Pipe), ranks third in the multilingual track, demonstrating the efectiveness and competitiveness of these minimal yet principled modifications. Code are publicly available at https://github.com/Kaggle-Competitions-Code/BioNNE-L.
Biomedical entity linking (BEL) – also known as named entity normalization or grounding – is the
task of mapping entity mentions in the text to entries in a reference knowledge base. In the biomedical
domain, EL plays a vital role in text mining by standardizing mentions of diseases, genes, drugs, and
other entities to canonical identifiers [
While early BEL research has made significant progress in English-only settings with flat (nonoverlapping) mentions, real-world biomedical documents often exhibit nested entities and appear in multiple languages—posing persistent challenges that remain under-addressed.
Nested mentions—where one entity is embedded within or overlaps another—are prevalent in
biomedical literature. For example, in “EGFR exon 19 deletion mutation”, the terms “EGFR” and “exon 19 deletion”
refer to distinct concepts, both requiring normalization. Ignoring nested structures can lead to
incomplete or incorrect linking. Meanwhile, the increasing volume of biomedical text in non-English
CLEF 2025 Working Notes, 9 – 12 September 2025, Madrid, Spain
⋆You can use this document as the template for preparing your publication. We recommend using the latest version of the
ceurart style.
* Corresponding author.
languages highlights the importance of multilingual BEL. Studies have shown that models trained in
English exhibit significant performance drops when applied to languages like Spanish or Russian [
Several technical barriers exacerbate these challenges: (i) the lack of annotated multilingual data, especially in low-resource biomedical languages; (ii) inconsistencies in concept coverage across languages in knowledge bases; and (iii) the inherent ambiguity and granularity of biomedical terminology. Existing EL pipelines are typically not equipped to handle these complexities simultaneously.
In this paper, we propose a lightweight, encoder-agnostic pipeline for multilingual, nested biomedical
EL. Our method introduces three key enhancements: (i) a two-stage retrieval-ranking strategy that
leverages the same base encoder model, where the retrieval stage utilizes the original pre-trained model
and the ranking stage benefits from contrastive learning training; (ii) boundary cue tagging, using
learnable tokens ([Ms] / [Me]) to explicitly encode span boundaries, enabling robust modeling of nested
mentions; and (iii) dataset augmentation by incorporating additional complementary data sources,
enriching training coverage without requiring manual annotation. Our approach maintains the original
EL model architecture while significantly improving robustness across languages and nested spans. It
can be seamlessly integrated with several biomedical encoders (e.g., BioLinkBERT [
Our system achieved third place in the BioNNE-L 2025 multilingual track [
To further advance research in biomedical entity linking, BioASQ 2025 [
Multilingual Biomedical Entity Linking. Multilingual BEL is an increasingly important research
direction due to the global nature of biomedical literature. Traditional approaches often rely on
translation to English prior to linking, but this can introduce noise and domain mismatch [
Contrastive and Graph-Based Learning. Contrastive learning has proven efective for biomedical
EL, particularly in bi-encoder architectures. GEBERT [
Large Language Models. LLMs like ChatGPT and GPT-4 have been tested on biomedical entity
link tasks [
Our Contribution. In contrast to prior work, our method is eficient and explicitly designed for both multilinguality and nesting. It requires no architectural change to the encoder and is compatible with any transformer-based biomedical model. Our use of span boundary cues provides strong supervision for nested and cross-lingual linking, while dataset augmentation further improves accuracy.
Our approach follows a two-stage paradigm: (i) dense retrieval to obtain a small set of plausible concepts for each mention, and (ii) cross–encoder ranking to pick the best concept. Although the backbone encoder may vary, the surrounding pipeline remains unchanged and is illustrated in Figure 2.
Let = { 1, . . . , || } be a biomedical knowledge base whose entries are represented by canonical names and concept unique identifiers ( Cuis). Given a document written in language ℓ ∈ {en, ru} that contains a set of (possibly nested) entities mentioned in ℳ() = { 1, . . . , }, The goal is to ifnd a mapping Φ : ( , , ℓ) →− ⋆ ∈ , 1 ≤ ≤ , where ⋆ denotes the concept that is semantically equivalent to the surface form of in its context.
We factor Φ into two components:
Retrieval: ret(, , ℓ) →−
= ⟨︀ 1 , . . . , ⟩︀ , Rank: rank︀( , , , ℓ)︀ →− ˆ ∈ , (1) (2) where ≪ || (we use = 10). Let ⋆ be the gold concept for mention . We report:
Acc@1 = 1 ∑︁ 1[︀ ˆ = ⋆ ]︀ , Acc@ = 1 ∑︁ 1[︀ ⋆ ∈ Top-(ˆ)︀] , ∈ {5, 10}, =1 =1 where ˆ is the ranker–sorted list of the retrieval candidates for the -th mention. Thus, the rank stage simply re-orders the 10 candidates returned by the retrieval stage, and we evaluate whether the gold concept appears within the first positions.
Retrieval stage. We experiment with five publicly available biomedical encoders, including
BioLinkBERT [
Rank stage. After the retrieval stage, we need to rank the Cuis. In particular, we build our ranking model without extra model modification. It is eficient to train the rank model with the retrieval model. We use two types of architecture to feed k candidates from the retrieval stage: 1. Listwise (LTR) (Figure 2 a): the candidates ⟨1 . . . ⟩ are concatenated after the [SEP] token;
One forward pass yields logits z ∈ R trained with a listwise soft margin loss. 2. Contrastive Learning (CL) (Figure 2 b): Each candidate is paired with the mention context and processed independently; every pass produces a binary score ∈ R optimized by cross-entropy.
Since the LTR scheme processes all candidates in a single forward pass, while the CL scheme handles only one candidate per pass, the CL scheme inherently requires times more computation during both training and inference. Despite costing times more computation, the CL scheme eliminates cross-candidate interference. Obviously, each forward pass evaluates a single candidate against the mention context, reducing the problem to an independent binary decision that the model can learn more easily.
Considering the limit of the training set, we also add additional dataset for the RU and BI tracks,
including MedMentions [17], a manually annotated resource for the recognition of English biomedical
concepts, and MCN [
Datasets. We follow the oficial BioNNE 2025 split: the train set with extra dataset is for training the model, the development set is used for model selection; the evaluation set is kept blind for final ranking. We set the retrieval numbers to = 10. The base dataset of this task is NEREL-BIO [18]. Metrics. For retrieval we report Acc@; for ranking we use the cross-validated Acc@1 (CV Acc) on the development folds. Final leaderboard numbers are the organisers’ Acc@1 on the hidden evaluation set.
For the ranking phase we keep the best retriever for each track: SapBERT–PubMedBERT (meantoken) for English, and SapBERT–UMLS–XLMR (large) for both the Russian and Bilingual tracks.
Table 2 analyses two cross-encoder architectures on the English dev set. The Listwise, although computationally cheap (one forward pass), plateaus at 0.5918 CV Acc. Switching to the CL design—i.e. an independent binary decision per candidate—raises accuracy to 0.6604 when additional MedMentions are used. We attribute the gain to two factors: (i) candidates no longer compete inside the softmax, Russian (RU) dev set SapBERT-UMLS-XLMR (base) SapBERT-UMLS-XLMR (large) Bilingual (BI) dev set SapBERT-UMLS-XLMR (base)
SapBERT-UMLS-XLMR (large) thus reducing interference, and (ii) the binary objective is simpler, allowing the model to specialise on ifne-grained lexical cues.
Bilingual results in Table 3 confirm the trend. Incorporating MedMentions and the MCN dataset adds a further 0.7–2.4 pp on Russian and bilingual tracks, but English still benefits the most ( +6.9 pp).
Lang CV(Acc) Approach
EN EN RU RU 5.3. Final result
To assess the contribution of boundary cues, we also perform an ablation study Table 5 on boundary cues, known as special tokens [Ms] and [Me], which indicate the start and end of the target entity. The improvement is most pronounced on the Russian (RU) track, where the Acc@1 increases by 6.60%. We attribute this to the richer morphology of Russian: the explicit [Ms] / [Me] markers help the model to delineate entity spans that may otherwise be obscured by inflectional endings. For the English (EN) and the Bilingual (BI) setting, the gains are more modest 1.20% and 1.24%, respectively - but still positive, confirming that boundary information remains beneficial even in languages with a relatively simpler morphology.
Lang w/ [Ms] and [Me] w/o [Ms] and [Me]
Gain EN RU BI
We present a simple yet efective two–stage pipeline for the BioNNE 2025 Bilingual Nested Entity Linking task. Keeping the base encoder untouched, we obtained competitive performance by addressing three task-specific bottlenecks: (i) explicit mention boundary cues ( [Ms]/[Me]) indicating the position of mention, (ii) eficient rank architecture design for ranking mention and (iii) data augmentation with MedMentions / MCN boosting the final result. On the oficial leaderboard our system ranks 3rd in BI track, with Acc@1 of 0.637 (BI), while training on a single Nvidia 3090.
During the preparation of this work, the authors used ChatGPT and Grammarly, to: Grammar and spelling check, paraphrase, and minor translation. After using these tools, the authors reviewed and edited the content as needed and assume full responsibility for the content of the publication. [14] Q. Chen, Y. Hu, X. Peng, Q. Xie, Q. Jin, A. Gilson, M. B. Singer, X. Ai, P.-T. Lai, Z. Wang, V. K. Keloth, K. Raja, J. Huang, H. He, F. Lin, J. Du, R. Zhang, W. J. Zheng, R. A. Adelman, Z. Lu, H. Xu, Benchmarking large language models for biomedical natural language processing applications and recommendations, Nature Communications 16 (2025) 3280. doi:10.1038/s41467-025-56989-2. [15] Y. Ding, Q. Zeng, T. Weninger, ChatEL: Entity linking with chatbots, in: Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), ELRA and ICCL, Torino, Italia, 2024, pp. 3086–3097. URL: https://aclanthology.org/2024.lrec-main.275/. [16] Y. Gu, R. Tinn, H. Cheng, M. Lucas, N. Usuyama, X. Liu, T. Naumann, J. Gao, H. Poon, Domain-specific language model pretraining for biomedical natural language processing, 2020. arXiv:arXiv:2007.15779. [17] S. Mohan, D. Li, Medmentions: A large biomedical corpus annotated with umls concepts, 2019.
URL: https://arxiv.org/abs/1902.09476. arXiv:1902.09476. [18] N. Loukachevitch, A. Sakhovskiy, E. Tutubalina, Biomedical concept normalization over nested entities with partial UMLS terminology in Russian, in: Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), ELRA and ICCL, Torino, Italia, 2024, pp. 2383–2389. URL: https://aclanthology.org/2024. lrec-main.213/.