A system for coreference resolution for the clinical narrative.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Clinical Informatics

Background:

• Coreference resolution is crucial for understanding clinical narratives.
• Existing computational methods require optimization for the unique characteristics of clinical text.

Purpose of the Study:

• To research and implement advanced computational methods for coreference resolution in clinical narratives.
• To develop and evaluate a system that identifies coreferent mentions (pronouns, noun phrases) within clinical documents.

Main Methods:

• Utilized the Ontology Development and Information Extraction corpus, annotated for coreference.
• Trained and compared classifiers (Support Vector Machines, Decision Trees, Perceptrons) using semantic, syntactic, and surface features.
• Employed feature selection to optimize classifier performance.

Main Results:

• Support Vector Machines with a Radial Basis Function kernel and all features achieved the best performance (MUC=0.352, B(3)=0.690, CEAF=0.486, BLANC=0.596).
• The developed system demonstrated performance comparable to that on general English text.
• Identified sentence distance exceeding 10 sentences as a primary error source.

Conclusions:

• The computational coreference resolution system performs well on clinical narratives.
• Future work should address distant coreferent markables and enhance the utilization of synonymy and ontological knowledge.
• The best performing methods are released as open-source modules for the Clinical Text Analysis and Knowledge Extraction System and Ontology Development and Information Extraction platforms.