Combining relation extraction with function detection for BEL statement extraction.

Area of Science:

• Bioinformatics
• Computational Biology
• Natural Language Processing

Background:

• The BioCreative-V challenge focused on automatically extracting causal relation networks in Biological Expression Language (BEL) from biomedical literature.
• Prior methods often repurposed models from related tasks, failing to fully leverage the available BEL training corpus.

Purpose of the Study:

• To develop a deep learning approach for direct extraction of BEL statements from biomedical text.
• To maximize the utilization of the existing BEL training corpus for improved causal relation network extraction.

Main Methods:

• A deep learning framework decomposing the task into entity relation extraction and entity function detection.
• Utilizing two attention-based bidirectional long short-term memory networks for subtask processing.
• Combining extracted entity relations and functions to form BEL statements, enhanced by threshold filtering for precision.

Main Results:

• The proposed method achieved state-of-the-art performance on the BioCreative-V Track 4 corpus.
• An overall F1-measure of 46.9% was obtained in stage 2 and 21.3% in stage 1.
• Evaluation was conducted both with and without gold standard entities, demonstrating robustness.

Conclusions:

• The deep learning approach effectively extracts BEL statements by directly utilizing the training corpus.
• The method demonstrates superior performance compared to previous approaches for causal relation network extraction.
• This work advances automated knowledge extraction from biomedical literature for BEL representation.