A transfer learning model with multi-source domains for biomedical event trigger extraction.

Area of Science:

• Computational Biology
• Natural Language Processing
• Machine Learning

Background:

• Automatic extraction of biomedical events is crucial for rapidly updating scientific discoveries.
• Trigger word recognition is a key step in event extraction, heavily reliant on accurate machine learning models.
• Existing datasets often suffer from insufficient and imbalanced annotations, hindering model performance in broad event domains.

Purpose of the Study:

• To develop an improved multi-source transfer learning framework for biomedical trigger detection.
• To enhance knowledge sharing between multiple source domains and a wide-coverage target domain.
• To address the challenge of insufficient and imbalanced annotations in biomedical event extraction datasets.

Main Methods:

• Proposed an improved multi-source domain neural network transfer learning architecture and training approach.
• Extended adversarial networks to extract common features across multiple source and target domains.
• Designed multiple feature extraction channels to capture both global and local common features, enhancing diversity and transferability.

Main Results:

• The proposed approach demonstrated improved recognition performance compared to traditional adversarial networks on the MLEE corpus.
• Experimental results showed competitive performance against other leading systems for wide-coverage trigger recognition.
• The model effectively utilized four diverse corpora as source datasets to improve performance on the target MLEE dataset.

Conclusions:

• The developed Multi-Source Transfer Learning-based Trigger Recognizer (MSTLTR) significantly enhances trigger recognition performance.
• The model's ability to represent common features globally and locally improves generalization on the target domain.
• This approach offers a robust solution for handling insufficient and imbalanced data in biomedical event extraction tasks.