Biomedical named entity recognition using BERT in the machine reading comprehension framework.

Area of Science:

• Computational Biology
• Bioinformatics
• Natural Language Processing

Background:

• Biomedical Named Entity Recognition (BioNER) is crucial for extracting knowledge from unstructured text.
• Conventional sequence labeling methods for BioNER often underutilize semantic information.
• Existing approaches may require complex decoding processes like Conditional Random Fields (CRF).

Purpose of the Study:

• To develop a novel BioNER approach by formulating it as a Machine Reading Comprehension (MRC) problem.
• To leverage prior knowledge through well-designed queries, enhancing BioNER performance.
• To eliminate the need for traditional decoding steps in BioNER.

Main Methods:

• The study proposes a paradigm shift from sequence labeling to a Machine Reading Comprehension (MRC) framework for BioNER.
• This MRC formulation allows for the integration of external knowledge via targeted queries.
• The method bypasses the necessity of post-processing steps such as CRF decoding.

Main Results:

• The proposed MRC-based BioNER method was evaluated on six diverse datasets (BC4CHEMD, BC5CDR-Chem, BC5CDR-Disease, NCBI-Disease, BC2GM, JNLPBA).
• State-of-the-art (SOTA) performance was achieved across all tested datasets.
• Specific F1-scores include 92.92% (BC4CHEMD), 94.19% (BC5CDR-Chem), 87.83% (BC5CDR-Disease), 90.04% (NCBI-Disease), 85.48% (BC2GM), and 78.93% (JNLPBA).

Conclusions:

• Reformulating BioNER as an MRC task is a highly effective strategy.
• The MRC approach surpasses traditional sequence labeling methods in performance and efficiency.
• This work sets a new benchmark for biomedical knowledge extraction from text.