Biomedical named entity recognition and linking datasets: survey and our recent development.

Area of Science:

• Bioinformatics
• Computational Biology
• Natural Language Processing

Background:

• Natural language processing (NLP) is crucial for extracting information from biomedical literature.
• Existing biomedical named entity recognition (BNER) datasets face challenges with outdated annotations, inconsistency, and low portability.
• The evolution of NLP techniques necessitates updated and robust benchmark datasets.

Purpose of the Study:

• To review common BNER datasets and identify annotation problems.
• To introduce a revised JNLPBA dataset addressing issues of inconsistency and portability.
• To develop an Ensembled Biomedical Entity Dataset (EBED) for multi-task BNER.

Main Methods:

• Reviewed existing BNER datasets for annotation inconsistencies and portability issues.
• Revised the JNLPBA dataset to resolve identified problems.
• Evaluated the revised dataset's portability using state-of-the-art BNER systems across diverse biomedical texts.
• Extended the revised JNLPBA with additional data sources to create the EBED, incorporating gene, disease, and chemical entity annotations.

Main Results:

• Identified significant annotation problems in commonly used BNER datasets.
• The revised JNLPBA dataset demonstrated improved portability across different biomedical literature types.
• The EBED provides a comprehensive multi-task dataset with 85,000 entity mentions, 25,000 with database identifiers, and 5,000 attribute tags.

Conclusions:

• The developed datasets offer improved resources for training and evaluating BNER systems.
• The EBED facilitates multi-task learning for enhanced biomedical entity recognition.
• These datasets contribute to more robust and reliable information retrieval from biomedical publications.