Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Biomedical Entity Linking for Dutch: Fine-tuning a Self-alignment BERT Model on an Automatically Generated Wikipedia Corpus

Fons Hartendorp, Tom Seinen, Erik van Mulligen, Suzan Verberne

TL;DR

This work delivers the first evaluated biomedical entity linking model for Dutch by leveraging SapBERT-style self-alignment pretraining on a curated Dutch biomedical ontology, followed by fine-tuning on a weakly labeled WALVIS corpus derived from Wikidata and Dutch Wikipedia. The approach yields notable gains on the Dutch Mantra GSC corpus, achieving 54.7% classification accuracy and 69.8% 1-distance accuracy, while the case study on patient forums reveals that downstream performance is constrained by upstream named-entity recognition quality. A key contribution is WALVIS, a reproducible weakly labeled BEL dataset for Dutch, generated automatically without manual annotation. The results demonstrate both the potential and the current limitations of Dutch BEL, and point to improvements via larger corpora and more context-aware models for real-world patient-generated text analysis.

Abstract

Biomedical entity linking, a main component in automatic information extraction from health-related texts, plays a pivotal role in connecting textual entities (such as diseases, drugs and body parts mentioned by patients) to their corresponding concepts in a structured biomedical knowledge base. The task remains challenging despite recent developments in natural language processing. This paper presents the first evaluated biomedical entity linking model for the Dutch language. We use MedRoBERTa.nl as base model and perform second-phase pretraining through self-alignment on a Dutch biomedical ontology extracted from the UMLS and Dutch SNOMED. We derive a corpus from Wikipedia of ontology-linked Dutch biomedical entities in context and fine-tune our model on this dataset. We evaluate our model on the Dutch portion of the Mantra GSC-corpus and achieve 54.7% classification accuracy and 69.8% 1-distance accuracy. We then perform a case study on a collection of unlabeled, patient-support forum data and show that our model is hampered by the limited quality of the preceding entity recognition step. Manual evaluation of small sample indicates that of the correctly extracted entities, around 65% is linked to the correct concept in the ontology. Our results indicate that biomedical entity linking in a language other than English remains challenging, but our Dutch model can be used to for high-level analysis of patient-generated text.

Biomedical Entity Linking for Dutch: Fine-tuning a Self-alignment BERT Model on an Automatically Generated Wikipedia Corpus

TL;DR

This work delivers the first evaluated biomedical entity linking model for Dutch by leveraging SapBERT-style self-alignment pretraining on a curated Dutch biomedical ontology, followed by fine-tuning on a weakly labeled WALVIS corpus derived from Wikidata and Dutch Wikipedia. The approach yields notable gains on the Dutch Mantra GSC corpus, achieving 54.7% classification accuracy and 69.8% 1-distance accuracy, while the case study on patient forums reveals that downstream performance is constrained by upstream named-entity recognition quality. A key contribution is WALVIS, a reproducible weakly labeled BEL dataset for Dutch, generated automatically without manual annotation. The results demonstrate both the potential and the current limitations of Dutch BEL, and point to improvements via larger corpora and more context-aware models for real-world patient-generated text analysis.

Abstract

Biomedical entity linking, a main component in automatic information extraction from health-related texts, plays a pivotal role in connecting textual entities (such as diseases, drugs and body parts mentioned by patients) to their corresponding concepts in a structured biomedical knowledge base. The task remains challenging despite recent developments in natural language processing. This paper presents the first evaluated biomedical entity linking model for the Dutch language. We use MedRoBERTa.nl as base model and perform second-phase pretraining through self-alignment on a Dutch biomedical ontology extracted from the UMLS and Dutch SNOMED. We derive a corpus from Wikipedia of ontology-linked Dutch biomedical entities in context and fine-tune our model on this dataset. We evaluate our model on the Dutch portion of the Mantra GSC-corpus and achieve 54.7% classification accuracy and 69.8% 1-distance accuracy. We then perform a case study on a collection of unlabeled, patient-support forum data and show that our model is hampered by the limited quality of the preceding entity recognition step. Manual evaluation of small sample indicates that of the correctly extracted entities, around 65% is linked to the correct concept in the ontology. Our results indicate that biomedical entity linking in a language other than English remains challenging, but our Dutch model can be used to for high-level analysis of patient-generated text.
Paper Structure (19 sections, 1 equation, 3 figures, 4 tables)

This paper contains 19 sections, 1 equation, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Preliminaries
  4. Unified Medical Language System
  5. Self-alignment pretraining BERT
  6. Methods
  7. Enhancing the UMLS
  8. Corpus compilation
  9. Self-alignment pre-training
  10. Fine-tuning
  11. Inference
  12. Evaluation data and metrics
  13. Case study on patient-support forum
  14. Results and analysis
  15. Quality of WALVIS-corpus
  16. ...and 4 more sections

Figures (3)

  • Figure 1: The task of biomedical entity linking. An entity recognition model identifies entities in free text that are then passed to the biomedical entity linking (BEL) model. The BEL model associates the new, unseen mention with its corresponding concept from an ontology.
  • Figure 2: Flow diagram of ontology enhancement. The remaining number of entries are denoted in italic.
  • Figure 3: WALVIS corpus compilation. All Wikidata entries with a linked Dutch Wikipedia page and a UMLS CUI that are in the ontology are retrieved using SPARQL. Then, all sentences from the Wikipedia dump are parsed and selected if they contain a hyperlink to one of the collected Wikipedia pages.