Interactive medical word sense disambiguation through informed learning.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Machine Learning

Background:

• Medical word sense disambiguation (WSD) is complex and typically requires extensive expert-labeled data.
• Developing high-quality WSD models often demands significant human effort and domain expertise.

Purpose of the Study:

• To develop an interactive learning algorithm for medical WSD.
• To efficiently leverage expert domain knowledge for building accurate WSD models with minimal human effort.

Main Methods:

• Developed an interactive learning algorithm allowing experts to label instances and features.
• Experts provided supervision by labeling instances, specifying sense-indicative words, and highlighting evidence.
• The algorithm iteratively selected informative instances for labeling, using three WSD corpora (MSH, UMN, VUH) for evaluation.

Main Results:

• The interactive learning algorithm significantly outperformed active learning for medical WSD.
• Achieved 90% accuracy on the MSH corpus with 42% less labeling effort.
• Demonstrated reduced labeling effort for UMN (35% less) and VUH (16% less) corpora.

Conclusions:

• High-quality medical WSD models can be efficiently trained with minimal supervision.
• Expert input through labeling informative instances and contextual features is key to efficient model training.