Study On User Study Process Oriented Query Expansion Methods

For some engines having published parts of their log,query expansion based on log has been more and more concerned.But in real applications,these methods have following problems:1)irrelevant feedbacks have big negative impact on the construction of term-relationship graph,which decreasing the precision of query expansion based on queries in log;2)irrelevant feedbacks have big negative impact on the latent semantic division,which decreasing the precision of diversified query expansion based on queries in log;3)irrelevant feedbacks have big negative impact on the mapping between query terms and document terms,which decreasing the precision of query expansion based on clicked documents;4)query expansion methods have limit applicable scenarios,so they are limited in real applications.The core phenomenon in above problems is that irrelevant feedbacks have big negative impact on descriptions of user intents.By observing the log,this dissertation finds a main reason of irrelevant feedbacks,named tentative clicks.Then,a user intent model has been designed to analize and filter irrelevant feedbacks by modeling tentative clicks,which is called user intent model fused with studying process.Based on this model,three kinds of query expansion methods are designed for decreasing negative impact of irrelevant feedbacks.And at last,an expansion term selection method based on the Term-Relationship graph with path-constrained has been designed to fuse models of different query expansion methods,which expending the application of the method.General achievements of this dissertation are presented as follows,(1)Query intent model fused with studying process(refined as QIMSP).By observing the log,it can be found that the tentative click is a main source of irrelevant feedbacks.Because many users often don't know the characteristics of documents they want,they need to tentatively click some documents and learn the characteristics of documents they want.In view of the irrelativeness between the user intent and the clicked document,tentative clicks are usually irrelevant feedbacks.This dissertation model the processes of users learning and applicating characteristics of documents,and proposes a user intent model fused with studying process.Because the model has fused with irrelevant feedbacks recognizing,the description of queries constructed by this model is more precisely,especially for those having long user study processes.(2)Query expansion based on QIMSP.Sorting expansion terms is a core step of query expansion.Traditional query expansion methods based on queries in log divide queries in log into terms,and then compute the similarity between terms by user intent,and sorting expansion terns by similarity results at last.Because traditional methods cannot effectively decrease the negative impact of irrelevant feedbacks to the similarity computation when log has large amount of irrelevant feedbacks,this dissertation compute the similarity between terms based on QIMSP,which can decrease the negative impact of irrelevant feedbacks,increase the precision of query expansion.(3)Diversified query expansion based on QIMSP.Diversified query expansion methods have been focused by many researchers.The reason of this method been proposed is that queries can often been mapped into multiple latent targets.Current researchers use linear transformation like singular value decomposition to analyze the multiple latent targets of queries.During this process,big amount of irrelevant feedbacks can lead to the unusual flow of energy in the matrix,which leading to the incorrectness of the method describing the latent targets of queries.In view of the problem,this thesis proposes a diversified query expansion method based on QIMSP,which fusing QIMSP into linear transformation.This method transforms the mapping of "term-document" into mapping of "term-user study process",which decreasing the times of irrelative documents transferring energy in the matrix.So it can decrease the negative impact of irrelevant feedbacks,increase the precision of diversified query expansion.(4)Query expansion based on QIMSP and topic model.This is a query expansion method based on clicked documents,aiming at supplementing the current query with document terms in relevant clicked documents,so that the extended query can describe the current user's query intention more clearly.In view of the sparsity of the log,current researchers pay more attention to the topic model to analyze the relativity between query terms and document terms.This method obtains data from training tuple<query,clicked document>,and uses the data to construct the topic model.But the model constructed by data from<query,clicked document>is inconsistent with the user intent when log has big amount of irrelevant feedbacks.For decreasing the negative impact of irrelevant feedbacks,this thesis fuses the training of topic with QIMSP,and proposes a parameter evaluating method for topic model based on user intent.This method decreases the negative impact of irrelevant feedbacks by decreasing the impact of irrelevant document terms in the state transitions of Markov,and then increases the fitting of topic model for user intent.(5)Expansion term selection based on TRGPC.At present,all kinds of log based query expansion methods are designed for specific application scenarios,and the algorithm can get better results when the user query condition is suitable for the application of the algorithm.However,in practical applications,many queries may not be clearly divided into a suitable scene.At this point,it is difficult to choose the appropriate query expansion method to deal with the current query.In order to extend those queries,this dissertation proposes a term-relationship graph with path-consistence(denoted as TRGPC),and designs a new method based on TRGPC.Because TRGPC can take into account the application scenarios of different query expansion methods,it is able to deal with the query that are not clear in the application scene,so as to meet the information needs of the current users.