Abstract: The purpose of query suggestion is to explore the query intent of search engine users and provide relevant query suggestion.Traditional query suggestion methods mainly rely on manually extracting relevant features of queries,such as query frequency,query time,user clicks and dwell time,etc.,and use statistical learning algorithms or ranking algorithms to give query suggestion.In recent years,deep learning methods have been widely used in query suggestion problems.The existing deep learning methods for query recommendation are mostly based on recurrent neural networks,which predict the next query of the user by modeling the semantic features of all queries in the query log.However,the existing deep learning methods have poor context awareness of query suggestion,it is difficult to accurately capture user query intentions,and the influence of time factors on query suggestion is not fully considered,and it lacks timeliness and diversity.In response to the above problems,this paper proposes a query suggestion model combining autoencoder and reinforcement learning(Latent Variable Hierarchical Recurrent Encoder-Decoder with Time Information of Query and Reinforcement Learning,VHREDT-RL).VHREDT-RL introduces a reinforcement learning joint training generator and discriminator,thereby enhancing the context awareness of generating query suggestion,using latent variable hierarchical recursive autoencoders that integrate query time information as a generator,and making query suggestion better time-sensitive and diversity.The experimental results on the AOL data set show that the VHREDT-RL model proposed in this paper achieves better accuracy,robustness and stability than the benchmark method.

Key words: Query intention, Query suggestion, Reinforcement learning, Time information, Variable hierarchical recursive autoencoder