Bidirectional long short-term memory with CRF for detecting biomedical event trigger in FastText semantic space

Area of Science:

• Biomedical informatics
• Natural Language Processing
• Computational Biology

Background:

• Biomedical event extraction is vital for understanding biological relationships.
• Trigger detection is a critical step in event extraction.
• Traditional methods struggle with multi-word triggers and require extensive feature engineering.

Purpose of the Study:

• To develop an improved method for biological event trigger detection.
• To address limitations of traditional classification-based approaches.
• To enhance the accuracy and generalization of event extraction models.

Main Methods:

• Utilized a sequence annotation model (LSTM and CRF) for trigger detection.
• Incorporated entity features and combined character-level and word-level embeddings.
• Trained and evaluated the model on the MLEE corpus and other benchmark datasets (BioNLP 2009, 2011, 2013).

Main Results:

• The model achieved an F-score of approximately 78.08% using only LSTM and CRF.
• Incorporating entity features improved the F-score to around 80%.
• The model demonstrated generalization capabilities, achieving over 60% F-scores on other corpora, outperforming comparative experiments.

Conclusions:

• The sequence annotation model simplifies training and avoids complex feature engineering.
• The proposed method effectively identifies multi-word triggers, boosting recognition F-scores.
• Entity details and combined word embeddings are crucial for successful trigger detection.