Dialect topic modeling for improved consumer medical search.

Area of Science:

• Health Informatics
• Natural Language Processing
• Computational Linguistics

Background:

• Consumer access to health information is limited by a language barrier.
• Existing resources often lack coverage of medical slang.
• Bridging this gap is crucial for effective health information retrieval.

Purpose of the Study:

• To present a Bayesian model for aligning documents across different language dialects (slang, common, technical).
• To extract semantic topics from these documents.
• To improve information retrieval accuracy for health-related queries, especially those containing slang.

Main Methods:

• Developed a Bayesian model named diaTM (dialect Topic Model).
• Explicitly modeled dialect mixtures within documents.
• Modeled the joint influence of dialects and topics on word selection.
• Evaluated using consumer questions and a medical document corpus.

Main Results:

• The diaTM model demonstrated improved information retrieval relevance.
• Achieved a 25% improvement in normalized Discounted Cumulative Gain at 5 (nDCG@5) compared to a Latent Dirichlet Allocation (LDA) baseline.
• Showcased effectiveness even when queries included slang medical terms.

Conclusions:

• The diaTM model effectively addresses the language gap in health information access.
• It enables more relevant retrieval of medical information, particularly when using non-technical or slang terminology.
• This approach enhances the utility of health information for consumers.