Learning joint representation for community question answering with tri-modal DBM

One of the main research tasks in Community question an- swering (CQA) is to find most relevant questions for a given new query, thereby providing useful knowledge for the users. Traditionally used methods such as bag-of-words or latent semantic models consider queries, questions and answers in a same feature space. However, the correlations among queries, questions and answers imply that they lie in differ- ent feature spaces. In light of these issues, we proposed a tri- modal deep boltzmann machine (tri-DBM) to extract unified representation for query, question and answer. Experiments on Yahoo Answers dataset reveal using these unified rep- resentation to train a classifier judging semantic matching level between query and question outperforms models using bag-of-words or LSA representation significantly.