Anaphoric relations in the clinical narrative: corpus creation

Area of Science:

• Natural Language Processing
• Clinical Informatics
• Computational Linguistics

Background:

• Automated analysis of clinical narratives is crucial for extracting valuable information from electronic medical records.
• Anaphoric relations, where a word or phrase refers to another, are complex linguistic features within clinical text.
• Developing systems to understand these relations requires high-quality annotated data.

Purpose of the Study:

• To establish a gold standard annotated corpus of clinical notes for the automated discovery of anaphoric relations.
• To describe the characteristics and creation process of this cross-institutional clinical narrative corpus.
• To lay the groundwork for an anaphoric relation resolver within a clinical natural language processing system.

Main Methods:

• A standard methodology was employed for annotation guideline development.
• Gold standard annotations were created for a multi-institutional corpus of deidentified clinical notes.
• Inter-annotator agreement (IAA) was calculated to assess annotation consistency and quality.

Main Results:

• The gold standard corpus comprises 7214 markables, 5992 pairs, and 1304 chains of anaphoric relations.
• Reports averaged 40 anaphoric markables, 33 pairs, and seven chains.
• High inter-annotator agreement was observed on the Mayo dataset (0.6607) and moderate agreement on the UPMC dataset (0.4072), indicating a usable corpus.

Conclusions:

• The developed corpus is suitable for the development of systems aimed at resolving anaphoric relations in clinical text.
• The study highlights the importance of diverse annotator backgrounds for robust annotation.
• The deidentified annotated corpus will be made available to researchers to advance clinical NLP research.