A Deep Generative Approach to Search Extrapolation and Recommendation

摘要

Related search query recommendation is a standard feature in many modern search engines. Interesting and relevant queries often increase the active time of users and improve the overall search experience. However, conventional approaches based on tag extraction, keywords matching or click graph link analysis suffer from the common problem of limited coverage and generalizability, which means the system could only make suggestions for a small portion of well-formed search queries. In this work, we propose a deep generative approach to construct a related search query for recommendation in a word-by-word fashion, given either an input query or the title of a document. We propose a novel two-stage learning framework that partitions the task into two simpler sub-problems, namely, relevant context words discovery and context dependent query generation. We carefully design a Relevant Words Generator (RWG) based on recurrent neural networks and a Dual- Vocabulary Sequence-to-Sequence (DV-Seq2Seq) model to address these problems. We also propose automated strategies that have retrieved three large datasets with 500K to 1 million instances, from a search click graph constructed based on 8 days of search histories in Tencent QQ Browser, for model training. By leveraging the dynamically discovered context words, our proposed framework outperforms other Seq2Seq generative baselines on a wide range of BLEU, ROUGE and Exact Match (EM) metrics.