Answer Re-ranking with bilingual LDA and social QA forum corpus

One of the most important tasks for AI is to find valuable information from the Web. In this research, we develop a question answering system for retrieving answers based on a topic model, bilingual latent Dirichlet allocation (Bi-LDA), and knowledge from social question answering (SQA) forum, such as Yahoo! Answers. Regarding question and answer pairs from a SQA forum as a bilingual corpus, a shared topic over question and answer documents is assigned to each term so that the answer re-ranking system can infer the correlation of terms between questions and answers. A query expansion approach based on the topic model obtains a 9% higher top-150 mean reciprocal rank (MRR@150) and a 16% better geometric mean rank as compared to a simple matching system via Okapi/BM25. In addition, this thesis compares the performance in several experimental settings to clarify the factor of the result.