Fine-tuning coreference resolution for different styles of clinical narratives

Area of Science:

• Natural Language Processing (NLP)
• Clinical Informatics

Background:

• Coreference resolution (CR) is vital for downstream NLP tasks like summarization and information extraction.
• Existing open-source CR tools exhibit suboptimal performance in the clinical domain.

Purpose of the Study:

• To explore practical solutions for fine-tuning CR models on clinical data.
• To address the substandard performance of current CR tools in the clinical domain.

Main Methods:

• Investigated automatic silver standard generation using an ensemble approach.
• Proposed a semi-automatic annotation process to aid manual annotation.
• Compared active learning with random sampling for cost-effective manual annotation.

Main Results:

• Automatic silver standard generation proved ineffective for CR model fine-tuning.
• Active learning requires cautious application.
• Semi-automatic annotation combined with continued training facilitates rapid CR model transfer in low-resource scenarios.
• An ensemble approach shows potential for enhanced accuracy.

Conclusions:

• Successfully transferred a general CR model to the clinical domain.
• Provided practical suggestions for rapid CR model transfer across diverse clinical narratives.