Biomedical event causal relation extraction with deep knowledge fusion and Roberta-based data augmentation.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Computational Biology

Background:

• Biomedical event causal relation extraction (BECRE) is crucial for understanding biomedical texts.
• Existing BECRE methods struggle with limited features and data imbalance issues.
• Traditional oversampling methods for data imbalance neglect data diversity.

Purpose of the Study:

• To propose a novel BECRE method addressing limitations of existing approaches.
• To enhance the extraction of causal relationships between biomedical events.
• To improve data diversity and overcome imbalance in BECRE datasets.

Main Methods:

• Deep knowledge fusion incorporating structural event representation and entity relation paths.
• Utilizing Graph Convolutional Neural Networks (GCN) and predicated tensor models for event representation.
• Employing a triplet attention mechanism for fusing knowledge representations.
• Implementing a Roberta-based data augmentation strategy for imbalanced datasets.

Main Results:

• The proposed method achieves state-of-the-art performance on the Hahn-Powell's and BioCause datasets.
• Deep knowledge fusion effectively establishes semantic connections between biomedical events.
• Roberta-based data augmentation successfully addresses data imbalance while promoting diversity.

Conclusions:

• The novel BECRE method demonstrates superior performance over current advancements.
• Deep knowledge fusion and advanced data augmentation are effective for BECRE.
• The approach offers a significant improvement for extracting causal relations from biomedical literature.