Learning Biomedical Entity Linking from Translated Abstracts, Not Swapped Words

Entity linking means taking a phrase in a text, such as "myocardial infarction" or its Spanish form "infarto de miocardio," and mapping it to one fixed identifier in a medical knowledge base. The standard knowledge base is UMLS, a large medical ontology whose concept identifiers are shared across languages. The best multilingual medical linkers today learn mostly from two sources. The first is lists of synonyms, which tell the model nothing about the sentence a phrase appears in. The second is English sentences in which one medical term has been swapped for a foreign-language synonym, so the model only ever sees foreign words inside English grammar. Neither source shows the model a German, Korean or Vietnamese medical sentence whose terms are labeled with the right concept. That matters most when a phrase is ambiguous and only the surrounding sentence can settle it.

We propose to build that missing training signal cheaply. We machine-translate whole English biomedical abstracts that already carry gold concept labels, carry each label across to its translated span, and train an encoder to link mentions in context: inside complete, untouched target-language abstracts, with no code-switching and no sentence shuffling. The claim is specific and can be tested. A large part of the remaining cross-lingual linking error, especially on ambiguous mentions, comes from the model never seeing target-language context. Translated whole abstracts fix this better than the mention-swapped or code-switched English that current systems train on.

Core Claim and Position Relative to Prior Work

Three lines of work frame the claim. The first supplies the starting encoder and also shows the gap. Learning Domain-Specialised Representations for Cross-Lingual Biomedical Entity Linking pulls together the embeddings of all UMLS synonyms of a concept across languages, a method it calls self-alignment pretraining. It released the 10-language XL-BEL benchmark and reported that general multilingual encoders score near zero without such tuning. Its authors state that the method does not use surrounding context. We start from their encoder and add exactly that missing piece.

The second line is the closest competitor. Controllable Contrastive Generation for Multilingual Biomedical Entity Linking (Con2GEN) does use context. It is a sequence-to-sequence model that generates the concept name together with its semantic type and language, and it is trained with hard negatives that share a name but differ in identifier. However, its multilingual training contexts are English MedMentions abstracts with the mention swapped for a UMLS synonym. Con2GEN reports an average gain of 1.3 points over the synonym-only encoder and cuts "right name, wrong identifier" errors from 6.9% to 4.1%. Our proposal keeps its hard-negative idea but changes where the training contexts come from. That change is the variable we isolate.

The third line constrains the design. KBioXLM: A Knowledge-anchored Biomedical Multilingual Pretrained Language Model and EntityCS: Improving Zero-Shot Cross-lingual Transfer with Entity-Centric Code Switching align languages by code-switching entities inside sentences. KBioXLM does this only for English–Chinese and relies on a machine-translated UMLS. Meanwhile, BioBART: Pretraining and Evaluation of A Biomedical Generative Language Model found that sentence permutation, a standard way of corrupting text during pretraining, hurts biomedical models because the flow of scientific abstracts carries meaning. We read this as a warning that inserting foreign terms into English sentences may teach a distorted picture of target-language text. We therefore keep each translated abstract intact and coherent.

The data recipe borrows from En-ViMedNER: An English-Vietnamese Parallel Biomedical Corpus with UMLS Semantic Type Annotations. That project translated MedMentions-derived abstracts into Vietnamese, had medical experts post-edit them, and used an LLM to carry entity labels across, with 98.32% exact-match accuracy on a human audit. It stopped at entity spans and semantic types and never linked to concept identifiers. It also reported that target-language supervision beats English-only transfer by over 16 F1 points, a gap our approach is designed to close without human annotation.

Research Program

Building the training corpus. We start from MedMentions, English PubMed abstracts annotated with UMLS concept identifiers. Each abstract is translated whole into the nine non-English XL-BEL languages with NLLB-200, the open translation model that En-ViMedNER benchmarks. Mention spans are carried across using the LLM-based label projection that En-ViMedNER validated. Projection quality is audited on a few hundred sentences per language. Because each projected span keeps the English gold identifier, every translated abstract becomes a concept-labeled target-language document at no annotation cost.

Training the encoder. We initialize from the multilingual self-alignment encoder of Liu et al. A mention is represented by pooling its tokens as the encoder reads the full abstract. A concept is represented by the same encoder applied to the concept's UMLS names. Training uses three objectives:

  • Contrastive linking loss. This pulls each in-context mention toward its concept, against in-batch negatives and against hard negatives that share a surface name but have a different identifier, as in Con2GEN.
  • Cross-lingual agreement loss. This pulls together the English mention and its translated counterpart within the same abstract pair.
  • Light masked-span infilling. The model reconstructs masked spans of the translated abstracts, with no permutation, so that it does not lose general language ability.

The original synonym loss stays on with a small weight, so that string-level alignment, which is what lets Thai work at all, is not forgotten.

Isolating the key ingredient. The central analysis holds the objective, the number of mention examples and the compute fixed, and varies only the source of the context. The three conditions are:

  • mention-swapped English, as in Con2GEN;
  • code-switched English, as in KBioXLM and EntityCS;
  • whole translated abstracts, which is our proposal.

On Vietnamese we add a fourth condition: the expert post-edited En-ViMedNER translations. This measures how much machine-translation noise costs compared with human-quality target text.

Baselines

Naive reference points. The weakest point is vanilla XLM-RoBERTa with nearest-neighbor search over UMLS names, which Liu et al. report as near zero. The more meaningful point is character n-gram TF-IDF candidate matching of the kind used by Cross-Lingual UMLS Named Entity Linking using UMLS Dictionary Fine-Tuning. Any neural gain has to clear this string-matching floor.

Self-alignment encoder (SapBERT-style, context-free). This is the multilingual synonym-aligned encoder of Liu et al., including the general bitext pseudo-synonyms they use for Thai. It is the starting point of our model, so every improvement over it is attributable to the added context training. The only reported figure on Lacuna is Thai P@1 rising from 11.5% to 30.9% with bitext transfer; full per-language tables were not found on Lacuna. We will rerun it on identical candidate sets and UMLS versions.

Con2GEN. This is the strongest existing contextual method. It is reported at +1.3 average accuracy over SAPBERT and mGENRE across 12 languages, with a 4.1% ambiguous-error rate, on XL-BEL and Mantra GSC. No absolute per-language accuracy was found on Lacuna, so we will reproduce it. Its mBART backbone will be fine-tuned on (a) its original mention-swapped data and (b) our translated abstracts. This tells us whether any gain comes from the data or from our encoder architecture.

Ablations. The three context-source conditions above share the same encoder, the same number of training mentions and the same hyperparameter search budget. Their comparison isolates the proposal's key ingredient.

What counts as beating the baselines. The proposal succeeds if the translated-abstract model beats both Con2GEN and the self-alignment encoder on XL-BEL average accuracy, and lowers the ambiguous-error rate below Con2GEN's reproduced figure, with differences outside the seed-to-seed variance across three runs.

Why not just…?

Why not just train a contextual reranker on top of the synonym encoder, as Galperin et al. do? Galperin et al. retrieve candidates by string similarity and then rerank them in context with general-domain mBERT. On informal Hebrew forum posts this lifted precision from 77% to 98%. Their reranker, however, needs labeled target-language data, which almost no language has. Our translated corpus can supply exactly that supervision. We will therefore also run the two-stage version: the synonym encoder followed by a reranker trained on our translated abstracts. If that two-stage system matches the single encoder, the contribution is the data rather than the encoder design, and we will report it that way.

Why not just translate the test sentence into English and use an English contextual linker? This "translate-test" approach is a real threat, because UMLS is about 70% English and English linkers are strongest. There are reasons for doubt. The HIKE retrieval paper (Mind the Gap: Cross-Lingual Information Retrieval with Hierarchical Knowledge Enhancement) documents translation errors propagating on short technical text, and En-ViMedNER found that fluent translations often break entity boundaries. Neither study measured concept linking, though. We will run translate-test with the same NLLB system and an English contextual linker trained on MedMentions. If it wins on languages with good translation quality, our approach is still justified only where translation quality is poor, such as Thai.

Why not just reuse Con2GEN with more mention-swapped data? If the bottleneck were data volume, scaling the swapped data would close the gap. The equal-mention-count ablation, plus a version with twice as much swapped data, tests this directly. If swapped context at double scale matches translated context, then the language of the context is not the cause, and our claim is refuted.

Why not just prompt a frontier LLM? On Vietnamese biomedical NER, En-ViMedNER reports Gemini 3.1 Flash at 12.70 F1 zero-shot and 36.47 F1 with three examples, against 52.70 F1 for a domain encoder. That evidence concerns NER, not linking. A small LLM-reranking baseline over our top-10 candidates is cheap enough to include.

Proposed Experiments and Decisive Checks

Benchmarks. The main evaluation is XL-BEL, which covers 10 languages and uses Wikipedia contexts. These contexts are deliberately out of domain relative to our PubMed-derived training data, which tests transfer rather than memorization. Mantra GSC adds European drug labels and patents. We also build a new concept-level Vietnamese test set by joining En-ViMedNER's expert-edited test spans to their MedMentions concept identifiers; abstracts that appear in that test split are excluded from training. The metric is Precision@1 (the share of mentions whose top-ranked concept is correct), reported overall, on a slice of ambiguous mentions (surface names that map to more than one UMLS identifier), and separately for concepts seen and unseen in training.

Decisive check 1: the ambiguous slice. If context is the missing ingredient, gains should be concentrated on ambiguous mentions. Uniform gains would instead point to more training data or better string alignment.

Decisive check 2: the context-source ablation. Whole translated abstracts must beat both mention-swapped and code-switched English at equal mention counts. If code-switching does as well, the BioBART-inspired concern about corrupting discourse does not carry over to linking.

Decisive check 3: translation quality. We plot the per-language gain against translation quality and against projection-audit accuracy. The method should help less as translation degrades, with Thai as the hardest case. The Vietnamese comparison of post-edited versus raw machine translation sets an upper bound on what better translation would buy.

Register check. Finally, we run Galperin et al.'s Hebrew consumer-health forum corpus (Camoni) as an explicit out-of-register test. Galperin et al. report a +8.87 F1 gain over the earlier MDTEL system and 98% precision there. We expect our literature-trained model to be weaker on informal text unless it is combined with their dictionary fine-tuning, and we will report that result whichever way it falls.