Dialect Topic Modeling for Improved Consumer Medical Search

摘要: Access to health information by consumers is hampered a fundamental language gap. Current attempts close the gap leverage consumer oriented information, which does not, however, have good coverage of slang medical terminology. In this paper, we present Bayesian model automatically align documents with different dialects (slang, common and technical) while extracting their semantic topics. The proposed diaTM enables effective retrieval, even when query contains words, explicitly modeling mixtures in joint influence topics on word selection. Simulations using questions retrieve from corpus show that achieves 25% improvement retrieval relevance nDCG@5 over an LDA baseline.