Query2vec: Learning deep intentions from heterogeneous search logs

The success of deep learning has been applied to many natural language processing applications such as machine translation, parsing, sentiment analysis and so on. The recent development of neural language models such as Skip-Gram improves accuracy of word embedding with much lower computational cost. The Skip-Gram, however, is not designed to extract user intentions from search queries due to sparseness of queries and different types of search logs such as query-clicks, session, and so on. In this paper, we propose Query-Gram that simultaneously learns query-clicks, session and documents to extract deep intentions of users. Our models outperforms the existing models and baselines: XX% on semantic embedding and XX% on syntactic embedding. Moreover, in vertical search engine such as Naver, our model shows YY% improvement of collection ranking.