Comparing neural models for nested and overlapping biomedical event detection.

Area of Science:

• Computational Biology
• Bioinformatics
• Natural Language Processing

Background:

• Biomedical event extraction frequently involves complex nested and overlapping event structures.
• Current neural models often overlook these structures or rely on external tools for detection.
• This limits the accuracy and completeness of extracted biomedical events.

Purpose of the Study:

• To develop and compare two novel neural network models, EXhaustive Neural Network (EXNN) and Search-Based Neural Network (SBNN).
• To enhance the detection of nested and overlapping events in biomedical text.
• To address the limitations of existing models in handling complex event structures.

Main Methods:

• Development of two novel neural network architectures: EXNN and SBNN.
• Evaluation of models in isolation and within a pipeline for event detection.
• Comparison against the state-of-the-art Turku Event Extraction System (TEES).

Main Results:

• Both EXNN and SBNN demonstrated superior performance in detecting nested and overlapping events compared to TEES.
• The proposed models achieved higher accuracy in event detection tasks.
• Performance improvements were observed even when using predicted named entities in a pipeline setting.

Conclusions:

• EXNN and SBNN are effective neural models for biomedical event extraction.
• These models significantly improve the detection of complex event structures.
• The proposed approach offers a more robust solution for extracting information from biomedical literature.