Source-linked AI summary
BERN2: an advanced neural biomedical named entity recognition and normalization tool
Mujeen Sung, Minbyul Jeong, Yonghwa Choi, Donghyeon Kim, Jinhyuk Lee, Jaewoo Kang
TL;DR
Biomedical text mining requires scalable recognition and normalization of diverse entities, but existing tools can be limited in coverage, speed, and morphological handling. BERN2 combines a multi-task NER model with hybrid rule-based and neural normalization, and it outperforms existing tools while providing annotations more efficiently.
Problem
Existing biomedical text-mining tools can support few entity types, be slow or memory-intensive, and fail to normalize morphologically varied mentions with dictionary rules alone.
Method
BERN2 uses a shared-backbone multi-task NER model and combines rule-based normalization with BioSyn for unresolved entities.
Results
BERN2 supports the largest number of entity types, outperforms other tools on most NER types except species, and its hybrid NEN outperforms other tools.
Takeaways & Limitations
BERN2 provides accurate and efficient biomedical annotation through web service or local installation for downstream text-mining systems.
Abstract
from arXiv · showhide
In biomedical natural language processing, named entity recognition (NER) and named entity normalization (NEN) are key tasks that enable the automatic extraction of biomedical entities (e.g. diseases and drugs) from the ever-growing biomedical literature. In this article, we present BERN2 (Advanced Biomedical Entity Recognition and Normalization), a tool that improves the previous neural network-based NER tool by employing a multi-task NER model and neural network-based NEN models to achieve much faster and more accurate inference. We hope that our tool can help annotate large-scale biomedical texts for various tasks such as biomedical knowledge graph construction.
1 Introduction
Biomedical text mining needs scalable NER and NEN, but existing tools cover few entity types, can be slow, and struggle with morphological variation. BERN2 addresses these limitations with broader entity coverage, a multi-task NER model, and hybrid normalization.
- Existing biomedical text-mining tools often support few entity types and rely on sequential single-type NER models that slow inference or require substantial GPU memory.
- Rule-based dictionary normalization can miss morphologically varied mentions such as “oxichlorochine” when the exact form is absent from the dictionary.
- BERN2 supports nine biomedical entity types, uses one multi-task NER model to reduce annotation time, and combines rule-based with neural NEN models.
2 Materials and methods
BERN2 combines parallel multi-task recognition with hybrid normalization in a pipeline that accepts text or PubMed IDs. Its NER shares a biomedical language-model backbone across entity-specific tasks, while BioSyn handles mentions not resolved by rules.
- BERN2 accepts plain text or PubMed IDs, recognizes nine entity types, extracts entity spans and types, and normalizes them to corresponding CUIs.
- For PubMed IDs, precomputed abstract annotations are returned when available, making retrieval faster than annotating the text anew.
- 2.1 Multi-task named entity recognition: The multi-task NER model uses a shared Bio-LM backbone with separate two-layer task-specific BIO classifiers and predicts entity types in parallel.
- 2.1 Multi-task named entity recognition: Parallel task-specific layers add few parameters and keep GPU memory usage comparable to a single pretrained language model.
- 2.2 Hybrid named entity normalization: BioSyn retrieves the dictionary entity whose encoded representation has the highest inner-product similarity with the input mention representation.
- 2.2 Hybrid named entity normalization: BERN2 first applies rule-based normalization, then uses BioSyn for unresolved gene/protein, disease, and drug/chemical mentions.
3 Results
BERN2 outperforms competing tools on most biomedical NER entity types and achieves higher normalization accuracy with its hybrid rule-based and BioSyn system.
- 3.1 Named entity recognition: BERN2 supports the largest number of entity types and outperforms other biomedical text-mining tools on most NER types, except species.
- 3.2 Named entity normalization: BERN2’s hybrid NEN system outperforms other tools on the BC2GN and BC5CDR normalization benchmarks.
4 Conclusion
BERN2 is an accurate and efficient biomedical NER and NEN tool that supports web-service and local installation use. Its multi-task NER and hybrid NEN models outperform existing tools while improving annotation efficiency.
- BERN2 combines multi-task NER and hybrid NEN to provide accurate, efficient biomedical text mining with web-service and local-installation options.