Using domain knowledge and domain-inspired discourse model for coreference resolution for clinical narratives.

Area of Science:

• Natural Language Processing
• Clinical Informatics
• Computational Linguistics

Background:

• Coreference resolution is crucial for understanding clinical narratives.
• Existing systems often overlook discourse aspects, relying on lexical and structural cues.
• Accurate entity clustering in clinical text remains a challenge.

Purpose of the Study:

• To present a novel coreference resolution system tailored for clinical narratives.
• To improve the accuracy of clustering mentions to coherent entities within documents.
• To demonstrate the impact of domain knowledge and discourse modeling on coreference resolution performance.

Main Methods:

• Employed a knowledge-intensive approach incorporating domain-specific lists.
• Developed a knowledge-intensive mention parsing technique to create semantic representations (SR).
• Introduced a task-informed discourse model specifically for 'person' type mentions in clinical narratives.

Main Results:

• Evaluated on four datasets from the 2011 i2b2/VA coreference challenge.
• Achieved unweighted average F1 scores ranging from 84.2% to 88.1% across different metrics (B-cubed, MUC, CEAF).
• Demonstrated the effectiveness of domain knowledge for various mention types across all tested datasets.

Conclusions:

• Extensive use of domain knowledge significantly enhances coreference resolution accuracy.
• Recall errors can be addressed by incorporating additional domain knowledge.
• Precision errors highlight the need to model mention relationships for robust systems.