DeepEventMine: end-to-end neural nested event extraction from biomedical texts

Area of Science:

• Biomedical Natural Language Processing
• Computational Biology
• Bioinformatics

Background:

• Current neural event extraction methods primarily focus on flat events and general domains.
• Limited research addresses nested and overlapping event detection, often relying on external tools and pre-identified entities.

Purpose of the Study:

• To develop an end-to-end neural model for extracting nested and overlapping events from raw biomedical text.
• To overcome limitations of existing systems that depend on external syntactic tools and pre-identified entities.

Main Methods:

• Proposed DeepEventMine, an end-to-end neural nested event extraction model.
• Leveraged bidirectional encoder representations from transformers (BERT).
• Model detects nested entities, triggers, roles, and event modifications without syntactic parsers.

Main Results:

• Achieved new state-of-the-art performance on seven biomedical nested event extraction tasks.
• Demonstrated promising performance in detecting events directly from raw text, even without gold entities.
• Successfully extracts multiple overlapping directed acyclic graph structures.

Conclusions:

• DeepEventMine offers an effective end-to-end solution for complex nested event extraction in the biomedical domain.
• The model's ability to function without syntactic tools and gold entities enhances its practical applicability.
• The availability of code and models facilitates reproducibility and further research.