Study On Session Search User Behavior And Information Retrieval Technology

As users’ information needs become complex,they may interact with the search engine for several rounds.This scenario is normally referred to as session search by academia.However,most existing retrieval models need to be improved in these scenarios.Fully understanding user interaction patterns in the session search process and then exerting the corresponding behavioral rules to improve the performance of each module in the search system is one of the core problems in Information Retrieval.Therefore,this thesis aims to comprehensively analyze users’ multi-round search behavior,introduce contextual information to enhance session-level user intent modeling,and design better session search models.Specifically,this thesis mainly focuses on the following three aspects:Pre-trained language model tailored for ad hoc search: Firstly,we design an axiominspired pre-training method tailored for the ad hoc search task,aiming to improve the basic document ranking ability of the session search system.Based on the investigations of previous studies,we design a novel axiomatically regularized pre-training method that achieves better ranking performance in the ad hoc search task.On the one hand,introducing reasonable IR axioms can help the system achieve better performance and higher robustness in low-resource scenarios.On the other hand,an intuitive case study indicates that the proposed method has learned knowledge about relevance matching summarized by human experts,improving the interpretability of the session search system.User query reformulation behavior analyzing and satisfaction modeling: Secondly,for the scenario that the system cannot handle within a single search round,we conduct an in-depth investigation on user query reformulation behavior patterns.Based on the analysis of the field study data,we summarize several patterns for fine-grained query reformulating actions,leverage abundant behavioral features to construct a prediction model and provide suggestions for designing better search result pages.According to the discovered clues in session data,we further regard query reformulating action as a proxy of user intent and introduce this factor into the construction of existing search evaluation metrics.Metrics designed with this idea can correlate better with users’ perceived satisfaction,which is beneficial for accurately optimizing system performance.Session search system optimization based on contextual information: Finally,we attempt to introduce multiple contextual factors to improve the performance of each submodule in the session search system.To facilitate the development of the related domain,we release a large-scale session dataset based on a real-world search log.By combining query history and click-through behavior,we construct a context-aware click model which enhances the system’s ability on click prediction and relevance estimation.Furthermore,we design a session-level multi-tasking learning framework by integrating intra-session and cross-session contexts.The framework performs significantly better in document ranking and query suggestion tasks.Focusing on users’ multi-round search process,this thesis has systematically studied various aspects,including dataset construction,user behavior analysis,module performance optimization,and satisfaction modeling,which are of great significance for facilitating the implementation and development of search engines,as well as improving users’ search experience.