Information Retrieval Techniques Based On Multi-dimensional Queries

The advent of the big data era has brought huge development opportunities for the country and society,but at the same time has also raised many challenges.How to effectively use big data information to improve economic and social benefits has become a key factor affecting the nation’s livelihood.Information retrieval,which aims to accurately obtain relevant information from redundant and complex data in response to existing queries,is a key research area in the context of big data.In fact,information retrieval has a wide range of application tasks and scenarios,including web search,recommendation systems,product search,question and answer systems,etc.This dissertation investigates the key problems existing in a variety of information retrieval tasks from the perspective of query dimensionality.This dissertation mainly focuses on information retrieval techniques for multidimensional queries,and investigates typical problems in each of the three dimensional query scenarios.Specifically,for meta-dimensional query-based recommender systems,this dissertation investigates a generalized metric learning method and applies it to a factorization machine recommender system algorithm.For onedimensional query-based product retrieval,this dissertation refines the user’s retrieval preferences in product search from both temporal and spatial perspectives,respectively.For two-dimensional query-based visual question answering,this dissertation investigates three different aspects,namely data analysis,model training,and model evaluation,in an attempt to solve the language prior problem in visual question answering.The main work and innovation of this dissertation contains the following three aspects.（1）Generalized metric learning recommendation system algorithm based on metadimensional queryThe traditional factorization machine approach only models the inter-attribute relationships,without considering the feature interactions within the attributes.The interaction relations within attributes are also extremely important to the factorization machine model,which can be classified into two types of relations,linear and complex nonlinear.In this dissertation,a generalized metric learning method is proposed and applied to the factorization machine method,so as to design a recommendation system model combining generalized metric learning to model the internal interaction relations of attributes,and an efficient algorithm is proposed to reduce the time complexity of the model with certain feasibility.（2）One-dimensional query-based global preference modeling method for product searchExisting product search methods only model the local preferences of users and thus perform ineffectively in retrieval.This dissertation proposes a global preference modeling method from two directions.Specifically,in terms of temporal perspective,this dissertation integrates users’ long-term preferences with short-term preferences for the first time.Considering that different factors in the two preferences have different degrees of influence on the current query results,this dissertation proposes to use the attention mechanism to differentiate the different factors in them.In addition,in terms of spatial perspective,this dissertation introduces users’ preference for visual modality into commodity retrieval,combines it with users’ preference for linguistic modality,and achieves the retrieval goal with the help of a translation-based model.（3）Language prior problem solving for visual question answering based on twodimensional queryThis dissertation explores and researches the language prior problem in visual question answering from three directions.First,from the perspective of data analysis,this paper proposes to explain the language prior problem from the perspective of class imbalance,and designs a method to overcome the class imbalance and applies it successfully;from the perspective of model training,this paper designs a loss function to recognize and solve the language prior problem from the perspective of answer feature space learning.Finally,in terms of model evaluation,this dissertation proposes an evaluation metric to measure the langauge prior effect in visual question answering models,which takes into account both the answer distribution in the training dataset and the prediction results of the current model.In addition,this paper designs a regularization method that can both alleviate the language prior problem generated by the model and improve the accuracy of the model answer retrieval.