Query Auto-Completion For Personalized Information Service

With the deepening of information technology revolution in military affairs,military information systems tend to integrate into each other.The explosion of military data makes it difficult to dealing with information overload.Query Auto-Completion(QAC)facilitates user query composition by offering a list of completions that start with the query prefix inputs.This popular feature of systems is employed to promote the efficiency of information service.As query tend to be short and ambiguous,and users often differ in their search interests,it is hard to predict users' intended queries.To address this problem,we should learn users' interests and intentions by mining their search history,and customize query completion lists to satisfy users' particular information needs.The main purpose of this thesis is to implement personalized military information service,which is based on user data mining and ranking model construction.Targeting at addressing the theoretical and technical problems in QAC,we conduct in-depth research in time patterns exploitation,location preference computation,topic interests modelling,search task analysis,neural network application,and propose ranking algorithms accordingly.We enumerate our main contributions as follows:(1)We propose a personalized QAC method based on query's temporal patterns.As the previous works ignore the burst pattern of aperiodic queries,they are unable to suggest trending queries timely.Therefore,we propose a time-sensitive hybrid QAC ranking model that combines the periodicity pattern and the burst pattern of query popularity.We first employ the Discrete Fourier Transform to detect the periodicity of a query's past popularity to forecast its future popularity.We then examining the burst trend of each query's popularity by Weighted Moving Average to predict query's future burst amplitude.Finally,we rank query completions by combing the forecasted popularity and the burst amplitude,thus,making full use of query's temporal patterns.The empirical experiments on a real-world query log show that our proposal can improve the ranking performance over the baselines in terms of Mean Reciprocal Rank(MRR).(2)We propose a personalized QAC method based on location sensitivity.As the special semantics and constraints of geographic queries are under-studied in the existing works,we propose a location-sensitive personalized QAC ranking model that considers users' geographic search intent.First,we investigate and extract two types of geographic queries.Then,we analyze the geographic queries in the clicked documents and the whole search log to obtain the location preference distribution of each user.Finally,we rank query completions by integrating query's future frequency,context information and user location preference altogether.Extensive experiments are conducted on a real-world search log.The results show that our proposal can satisfy users' geographic information needs with the scores of MRR and Success Rate @ top k(SR@k)exceeding the baselines.(3)We propose a personalized QAC method based on user topic interest.As the personalized QAC models often suffer from data sparseness problem,we propose the construction and application of cohorts to address the context sparsity and enhance QAC personalization.We build an individual's interest profile by learning his topic preference through topic models and the aggregate users who share similar profiles.As conventional topic models are unable to learn cohorts automatically,they have to employ general clustering methods to group similar users.However,those methods allocate each user to exactly one cohort,which is fail to model the diversity of users' interests.Therefore,we propose two Cohort Topic Models that assign each user to multiple cohorts with a probability associated with each cohort.The experimental results show that our proposal can remedy the sparseness problem and yield significant relevance improvement over competitive baselines.(4)We propose a personalized QAC method for multi-session tasks.As most of the existing works analyze users' search intent by investigating users' search history within a session,they are unable to handle multi-session tasks that straddle several sessions.Therefore,we explore the potential of task-related features for improving QAC effectiveness under complex search tasks scenarios.First,we list the definitions of session and task,then we identify search tasks by combining queries' text and semantic similarity.After that,we describe our supervised framework for QAC personalization,where four levels of task-related features are considered separately and synthetically,including history-level,task-level,session-level,and query-level.Extensive experiments are conducted on a large-scale query log and the corresponding results suggest that our proposal significantly outperforms the competitive baselines.(5)We propose a personalized QAC method based on recurrent neural network.As the learning-to-rank QAC models are subject to the handcrafted features and are unable to model complicated user-query relationship,we propose a new supervised framework that adopts the Recurrent Neural Network(RNN)to solve the QAC problem.Specifically,we devise three RNN-based QAC ranking models.The first model is a basic RNN structure(BRNN)for session-based QAC which models users' behavior within a session.The second model is a personalized RNN-based model(PRNN)which adds a user-level RNN on the first model.PRNN aims to model the short-term context and longterm history simultaneously.The third model is an attentive personalized RNN structure(A-PRNN)that employs attention mechanism to capture users' main focus.We evaluate the performance of the proposed three models on a real-world query log.The significant improvement over the compared baselines verifies the effectiveness of PRNN and A-PRNN.