Research On Query Feature Context Aware Information Retrieval Model

Nowadays Information retrieval (IR) technology has been widely employed toexplore information from the Internet. There are several important factors in achievinghighly efficient information retrieval process. One of them is well-structured andgood-meaningful user query since it is the objective representations of user subjectiveinformation requirements. Normally the user query consists of several keywords, andthen documents matching the keywords best will be retrieved. To improve the accuracyof retrieval results, the user query should be the most exact representation of userinformation needs. However, in most cases user information needs cannot be totallyexpressed through their queries. The inaccurate and incomplete expression of userinformation needs makes it difficult for search engine to improve its performance.Recently, the applications of context in the domain of information retrieval haveattracted a lot of attentions in the research community. Generally, context covers allkinds of information relating to query or user during the retrieval. It provides a lot ofvalued information, which facilitates to understand user query intentions. It has beenproved that the performance of retrieval can be improved with the consideration ofcontext.The process of query construction is the process of information filtering. Largeamounts of information often hide behind the simple query and always reflect useroriginal information needs to some extent. Therefore, in this thesis a newunderstanding of query context is explored from viewpoint of the query itself. Fromthe perspective of the process of query construction, three query-related features aredefined, which are respectively query type, the inner dependency relationships amongquery terms and the exterior conditions of query term occurrences. These featuresreflect user implicit information needs and can give the reasonable explanations of theappearances of terms in queries. Information relating to query features constructs thecontext proposed in this thesis.Based on the idea above, a novel Query Features Context based InformationPetrieval (QFCIR) model is proposed in this thesis. In this model, context informationis described in the form of sentence. There are four parts in the proposed QFCIRmodel, three of which focus on the recognitions of query-related sentences from theperspective of each query feature and the other one is to realize the optimization of retrieval results based on query context.Query type based sentence retrieval model is proposed to explore contextsentences involving information of particular types. Therefore, sentences that areexpected to be relevant should meet the requirements both in query terms and querytypes. To obtain such information about query types, a hybrid classification model isproposed to improve the accuracy of the recognition of query type effectively.Markov Random Field (MRF) based sentence retrieval model is proposed toexplore context reflecting the inner dependency relationships among query terms.Different forms of dependency among query terms are considered in this model.Furthermore, special feature functions are defined, with the purpose to utilizeassociation features between query terms in sentence to evaluate the relevancy ofsentence.Bayesian network based sentence retrieval model is proposed to discover theexterior conditions of query term occurrences, which are the beneficialcomplementarities of the original query. Based on the topology of the simple Bayesiannetwork, this thesis further introduces term contextual associations into the inferenceprocess. The efficiency of Bayesian network in dealing with information uncertaintyand the fiexibility of term associations in reflecting information relevancy guaranteethe accuracy of the inference of query-related context.Based on the information of query-related context, relevancy of each top-rankeddocument is re-evaluated for document re-ranking. A sentence based translation modelis then proposed to evaluate the similarity between query context and each retrieveddocument. This model has solid theoretical foundation for application and is a feasiblesolution in dealing with document relevancy in the condition of query context.