Personalized Search Framework With Joint Learning Of Document Ranking And Query Suggestion

With the rise of the internet industry,the number of network information resources is huge and rapidly increasing.Information overload has become an important challenge in the era of big data.Information retrieval can help users quickly locate information resources related to their needs from a large amount of noisy data.With the maturity and development of information retrieval technology,adapting to the needs of interaction,intelligence and personalization has become a new trend of current development.Session search system came into being.It uses the session mechanism to track and record the browsing interaction behavior of a user,based on which infer the query intent,and then predict the search results.Document ranking and query prediction are two core tasks in session search scenarios.During a search session,a user not only interacts with the ranked list of documents presented to him result in clicks and skips,but also repeatedly revise a query(e.g.,adding or removing terms)so as to clarify his information need.When concurrently performing these search activities,a user is motivated by the same search intent.Therefore,sharing the intent to jointly optimize these two tasks can promote each other.However,existing approaches to this joint optimization problem only consider a user's short-term search interest in an ongoing session;the user's long-term personal preferences as exhibited during previous sessions are neglected.As different users have different personalized characteristic,similar queries may express different intents.Intuitively,personalized search strategies can alleviate this ambiguity problem by modeling user profiles and provide more precise results for individual users.We believe that there are two kinds of available personalized information buried in a user's historical sessions at least:(1)previous sessions with similar search intent;(2)the long-term search roles and goals that do not change frequently in a user's search history.However,existing long-term personalized approaches focus on optimizing a single task,such as ranking documents using long-term search logs.These approaches still have several shortcomings when modeling long-term information:(1)it is hard to capture long-term search intent;(2)a lack of fine-grained historical information reduces search performance;(3)low efficiency limits the potential applicability in real-world settings.In order to deal with the above problems,we consider a personalized mechanism for learning a user's profile from their short-term and long-term search behaviors to simultaneously enhance the performance of document ranking and next query prediction in an ongoing session.We propose a personalized joint learning framework based on deep learning,called Long short-term session search,Network(LostNet),that jointly learns to rank documents for the current query and predict the next query.Lost Net consists of three modules:(1)a hierarchical session-based attention mechanism,(2)a personalized multi-hop memory network,and(3)joint learning of document ranking and next query prediction.Specifically,the hierarchical session-based attention mechanism models the finegrained short-term search intent within a session.The personalized multi-hop memory network tracks a user's dynamic profile information from their prior search sessions so as to infer their personal search intent.The joint learning network apply a multi-task learning strategy to jointly optimize the ranked list of documents and the query prediction results by sharing personalized search intent.The article uses two large-scale benchmark datasets in session search domain.They are query log data on real search engines.We conduct extensive experiments on two datasets.The results show that the long short-term session search network proposed in this paper achieves significant improvements over state-of-the-art baselines on both document ranking task and query prediction task.The results also confirm that modeling long-term user profile helps to enhance the search performance in the current session.Moreover,we also verify the effectiveness of the personalized memory encoder and multi-task learning framework in LostNet through a series of analysis experiments.