Cross-type biomedical named entity recognition with deep multi-task learning.

Area of Science:

• Biomedical Natural Language Processing
• Machine Learning for Bioinformatics

Background:

• Biomedical Named Entity Recognition (BioNER) systems traditionally rely on handcrafted features, which are time-consuming to develop and specific to entity types like genes, chemicals, and diseases.
• While neural network models have reduced manual feature engineering, their performance in BioNER is often constrained by limited training data for individual entity types.

Purpose of the Study:

• To develop a novel multi-task learning framework for BioNER that leverages collective training data from diverse entity types.
• To improve the performance of BioNER systems by enabling the sharing of information across different entity categories.

Main Methods:

• A multi-task learning framework was proposed to train BioNER models simultaneously on multiple datasets representing different entity types.
• The framework utilizes shared character- and word-level information across relevant biomedical entities found in various labeled corpora.

Main Results:

• The proposed multi-task BioNER model demonstrated substantially improved performance compared to state-of-the-art BioNER systems and baseline neural sequence labeling models across 15 benchmark datasets.
• Analysis confirmed that performance gains are attributable to the effective sharing of linguistic features among related biomedical entities.

Conclusions:

• Multi-task learning offers a powerful approach to overcome data scarcity limitations in BioNER.
• The developed framework effectively enhances BioNER performance by enabling cross-entity data utilization and feature sharing.