Bi-Encoder-Based Approach to Biomedical Document-Level Entity Recognition and Relation Extraction

Area of Science:

• Biomedical Natural Language Processing
• Computational Biology
• Bioinformatics

Background:

• Accurate biomedical entity and relation extraction is crucial for knowledge discovery.
• Document-level analysis is necessary due to relations spanning sentences.
• Current methods face challenges in speed and handling complex entities.

Purpose of the Study:

• To introduce BioECR, an end-to-end system for document-level biomedical named entity recognition, coreference resolution, and relation extraction.
• To enhance the recognition of complex and nested biomedical entities.
• To improve the efficiency of relation extraction tasks.

Main Methods:

• Utilized a bi-encoder structure with biomedical entity types and descriptions for efficient nested entity recognition.
• Employed a composition graph convolutional neural network to reduce noise and selectively fuse information.
• Integrated entity type clustering to resolve coreference errors across multiple entity types.

Main Results:

• Achieved state-of-the-art performance on all subtasks across three benchmark datasets (CDR, GDA, BioRED).
• Demonstrated significant improvements in recognizing complex entities and relations.
• Reduced inference time by approximately 60% compared to existing methods.

Conclusions:

• BioECR offers a highly effective and efficient solution for document-level biomedical text mining.
• The proposed methods successfully address challenges in nested entities, noisy graph processing, and coreference resolution.
• This advancement facilitates large-scale biomedical knowledge graph construction and downstream applications.