Boundary-Aware Dual Biaffine Model for Sequential Sentence Classification in Biomedical Documents.

Area of Science:

• Biomedical Natural Language Processing
• Computational Linguistics
• Medical Informatics

Background:

• Accurate rhetorical role assignment (e.g., background, intervention, outcome) in biomedical texts aids clinical decision-making.
• Existing sequence and span-based methods have limitations in capturing document semantic structure and handling sentence boundaries.

Purpose of the Study:

• To develop an advanced model for precise sentence labeling in biomedical documents.
• To address the semantic coherence and boundary information challenges in rhetorical role assignment.

Main Methods:

• Proposed a boundary-aware dual biaffine model incorporating document semantic structure and boundary information.
• Utilized a dual biaffine layer for explicit semantic modeling.
• Implemented a dynamic programming algorithm for optimal decoding and to minimize prediction errors.

Main Results:

• The boundary-aware dual biaffine model achieved state-of-the-art performance on the PubMed 20 k RCT and PubMed-PICO datasets.
• Demonstrated superior performance compared to strong baseline methods.
• Achieved competitive results on the NICTA-PIBOSO dataset.

Conclusions:

• The proposed model effectively captures document semantic structure and boundary information for improved sentence role labeling.
• This approach enhances the efficiency of evidence retrieval for physicians.
• The model represents a significant advancement in biomedical text analysis for clinical applications.