Unsupervised biomedical named entity recognition: experiments with clinical and biological texts

Area of Science:

• Biomedical Natural Language Processing
• Computational Linguistics
• Bioinformatics

Background:

• Named entity recognition (NER) is vital for biomedical information extraction.
• Current NER tools are often genre/task-dependent, requiring extensive re-annotation or rule development for new data.
• Adapting existing methods to different biomedical text genres poses significant challenges.

Purpose of the Study:

• To propose an unsupervised approach for biomedical named entity recognition.
• To address challenges in entity boundary detection and type classification without annotated data.
• To develop a system that reduces the effort required for adapting NER to new biomedical domains.

Main Methods:

• A stepwise solution combining noun phrase chunking and inverse document frequency filtering for candidate entity extraction.
• Leveraging distributional semantics principles for classifying candidate entities into predefined categories.
• An unsupervised methodology avoiding handcrafted rules, heuristics, or annotated datasets.

Main Results:

• The proposed unsupervised system achieves competitive results on clinical notes and biological literature datasets.
• The entity classification component, in particular, demonstrates strong performance.
• The system outperforms a baseline dictionary match approach in biomedical NER tasks.

Conclusions:

• Unsupervised methods can effectively perform biomedical named entity recognition.
• The developed approach offers a viable alternative to supervised methods, reducing reliance on annotated data.
• Further research can build upon the detailed error analysis to enhance future iterations of the system.