A controlled greedy supervised approach for co-reference resolution on clinical text.

Area of Science:

• Computational Linguistics
• Natural Language Processing (NLP)
• Biomedical Informatics

Background:

• Co-reference resolution is crucial for NLP tasks like event linking but remains challenging.
• Existing methods struggle with specialized domains such as clinical records.
• Ambiguity in evaluation metrics complicates co-reference resolution research.

Purpose of the Study:

• To present an improved mention-pair model for co-reference resolution.
• To enhance performance on clinical text by addressing limitations of current methodologies.
• To develop a system that performs well across various entity types and evaluation metrics.

Main Methods:

• A variant of the mention-pair model for co-reference resolution was developed.
• Linguistically and semantically motivated constraints were used to refine training and test instances.
• Aggressive greedy strategies were incorporated into the chain clustering process.

Main Results:

• The proposed approach achieved an unweighted average F1 score of 0.895 on the i2b2/VA 2011 challenge corpus.
• Performance was comparable to the best systems participating in the challenge.
• High average F1 scores were obtained for individual chain types: Test (0.897), Person (0.852), PROBLEM (0.855), and TREATMENT (0.884).

Conclusions:

• The proposed co-reference resolution system demonstrates strong performance, particularly in the clinical domain.
• The method effectively balances performance across different evaluation metrics (MUC, B(3), CEAF) and entity types.
• This approach offers a robust solution for co-reference resolution challenges in specialized text domains.