ADPG: Biomedical entity recognition based on Automatic Dependency Parsing Graph.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Text Mining

Background:

• Named entity recognition (NER) is crucial for extracting information from biomedical texts.
• Previous methods for nested biomedical NER rely on external parsers, which can introduce noise and hinder end-to-end performance.
• Biomedical literature often features long-dependent sentences, posing challenges for accurate entity recognition.

Purpose of the Study:

• To propose a novel automatic dependency parsing approach (ADPG model) for end-to-end biomedical entity recognition.
• To effectively fuse syntactic structure information without external tools.
• To improve the performance of nested biomedical entity recognition.

Main Methods:

• The ADPG model utilizes a multilayer Tree-Transformer to extract semantic and syntactic representations from long-dependent sentences.
• A multilayer graph attention neural network (GAT) is employed to extract dependency paths within the syntactic structure.
• The model integrates these components for end-to-end entity recognition.

Main Results:

• The ADPG model achieved state-of-the-art results on three biomedical and one news domain dataset.
• The model demonstrated strong generalization performance across different domains.
• Experimental results validate the effectiveness of the proposed automatic dependency parsing approach.

Conclusions:

• The ADPG model offers an effective end-to-end solution for biomedical named entity recognition.
• Integrating automatic dependency parsing enhances the extraction of complex entities from biomedical literature.
• The proposed method shows promise for advancing information extraction in the biomedical domain.