Research On Text Semantic Matching Based On Deep Learning

As the arrival of digital era,there is an increasing demand in the field of artificial intelligence,such as question answering system,intelligent customer service,information retrieval,text retelling,machine translation and so on.In order to provide high-quality and efficient services,numerous scholars have devoted themselves to the research of text semantic matching.As a basic task and research hotspot in the field of natural language processing,the previous research on text semantic matching is mainly based on statistical machine learning technology.This kind of semantic matching model needs to consume a lot of manpower cost to mine the potential features of the text.With the development of deep learning,the feature extraction of text data is no longer a difficult problem,and more and more scholars have focused on the research of text semantic matching based on this technology.In this thesis,some mainstream matching models which based on deep learning are compared and analyzed.In view of the shortcomings of these models,this thesis proposes the multi-information cross-fusion text semantic matching model and the optimal weight model fusion algorithm.The main work of this thesis includes:First of all,this thesis proposes a multi-information cross-fusion text semantic matching model(MICF),which is improved on the basis of multi-granularity semantic cross model.In order to solve the problem of the loss of partial semantic feature in the interactive matching process,the model constructs a multi-information cross-fusion embedding layer,a semantic cross-layer and a feature extraction layer to extract multidimensional feature information of text data,such as word importance,context information,word granularity,word matching importance and word location.The experimental results show that,compared with other semantic matching models,MICF has a better semantic matching effect on two Chinese semantic matching datasets.Secondly,this thesis proposes the optimal weight model fusion algorithm(OWBlending),which is improved on the basis of Blending ensemble algorithm.In order to solve the problem that the Blending ensemble algorithm can not comprehensively consider the learning performance of each base learner and the loss of features of the original text data,the OWBlending ensemble algorithm uses a combination strategy of "weighted learning method based on neural network" to combine the output of each base learner in the ensemble algorithm.The algorithm predicts the optimal weight combination of each base learner by building a neural network module to learn the characteristic information of the original text data.The experimental results show that the semantic matching effect of OWBlending on two Chinese semantic matching datasets is better than other ensemble learning algorithms such as Bagging,Stacking,Blending and so on.