Biomedical Argument Mining Based on Sequential Multi-Task Learning.

Area of Science:

• Biomedical informatics
• Computational linguistics
• Artificial intelligence

Background:

• Biomedical argument mining extracts argumentative structures from text to understand reasoning in medical decision-making.
• Current methods often overlook sequential dependencies between argument component classification and relation identification.
• Relation identification typically lacks contextual information, limiting its effectiveness.

Purpose of the Study:

• To propose a novel sequential multi-task learning approach for biomedical argument mining.
• To explicitly model the sequential dependency between argument component classification and relation identification.
• To enhance relation identification by incorporating contextual information.

Main Methods:

• A sequential multi-task learning framework was developed.
• An information transfer strategy was used to link argument component classification to relation identification.
• Graph convolutional networks (GCNs) were employed to model dependencies among argument component pairs within their context.

Main Results:

• The proposed method demonstrated superior performance compared to state-of-the-art approaches.
• Experimental results on a benchmark dataset validated the effectiveness of the sequential learning strategy.
• Incorporating sequential dependencies and contextual information significantly improved mining accuracy.

Conclusions:

• The novel sequential multi-task learning approach enhances biomedical argument mining accuracy.
• Explicitly modeling sequential dependencies and contextual information is crucial for robust argument mining.
• This work offers a more effective method for understanding argumentative structures in biomedical literature.