Multitask learning for biomedical named entity recognition with cross-sharing structure.

Area of Science:

• Biomedical informatics
• Computational biology
• Natural Language Processing

Background:

• Biomedical Named Entity Recognition (BioNER) is crucial for biomedical literature mining.
• Traditional BioNER models require extensive feature engineering or hand-crafted rules.
• Neural networks offer automatic feature learning, and multi-task learning enhances performance by leveraging multiple datasets.

Purpose of the Study:

• To propose a novel multi-task model for BioNER utilizing a cross-sharing structure.
• To enhance the performance of BioNER models through effective feature sharing.
• To provide guidance on selecting optimal dataset pairs for multi-task training.

Main Methods:

• Developed a novel multi-task learning model for BioNER with a cross-sharing architecture.
• Trained and evaluated the model on gene, protein, and disease datasets.
• Conducted experiments to identify optimal dataset pairings and analyze entity type influences.

Main Results:

• The proposed cross-sharing multi-task model outperformed existing multi-task models on gene, protein, and disease datasets.
• Analysis identified effective dataset pairs for multi-task training and the impact of different entity types.
• The model demonstrated robust performance even when dataset sizes were reduced.

Conclusions:

• The novel cross-sharing multi-task model significantly improves BioNER performance.
• Feature sharing across datasets via the cross-sharing structure is beneficial.
• The study offers valuable insights for optimizing multi-task learning strategies in BioNER.