| The purpose of the text matching task is to extract the inherent semantic characteristics of two pieces of text and predict the correlation or contradiction between the texts.As one of the most important tasks in the field of natural language processing,text matching has always been the focus of many researchers in the field.Benefiting from its advantages of easy deployment,fast response,small model and strong versatility,it occupies a core position in the fields of intelligent QA,search engine and mobile phone assistant.However,there are still some problems in this field,which hinder the further development of text matching.In order to improve the accuracy of the model,researchers propose to interact with texts to strengthen the relevance between texts.Yet for general data,the results remain poor.In addition,there is no fixed solution to the problem of too much noise information in long form text matching.For the text matching task,this paper focuses on the existing problems in the field to carry out algorithm research,experimental verification and practical application.The main work and contributions of this paper can be summarized in the following three points:(1)In response to the excessive number of parameters in current text matching models and the limited prediction ability of the models for general data sets,this paper proposes a text matching framework based on an improved residual module and a simple attention mechanism,which can take into account both the accuracy of the model and the size of the number of parameters.The improved residual module designed in the framework can greatly preserve the low-level propagation features of the text and enhance the extraction of text feature information by the network.For the features after text encoding,the framework adopts a simple attention mechanism to achieve feature alignment among texts.The parallel processing property of the attention mechanism is utilized to not only improve the computational speed of the model but also enhance the accuracy of the network prediction.(2)Most of the popular long form text matching algorithms use pre-trained models with large number of parameters and then perform similarity calculation between texts.Although excellent accuracy is available,the training speed is slow and the number of parameters is huge.In order to reduce the number of parameters of the model,this paper proposes a solution.Firstly,based on Bang Liu et al.’s work,this paper uses the idea of graph division to reduce the number of vertices in the graph by reducing the number of vertices and increasing the amount of key information contained in each vertex,thus reducing the number of parameters of the model and improving the model computation speed.Then the key matching information is aggregated by graph neural network.Finally,this paper extracts the global features of documents from multiple perspectives and unifies the local and global features in aggregation to improve the accuracy of algorithm prediction.(3)In order to verify the effectiveness of the algorithm proposed in this paper on general data and its practical value in industries such as intelligent question and answer systems,search engines and cell phone assistants in people’s daily life.This paper designs and implements an intelligent question and answer system based on the Chinese Wikipedia corpus.The system implements question and answer matching on the Chinese corpus.The system matches the exact answer in the corpus based on the question information inputted by the user,and feeds back the result with the highest score on the client side.Users can intuitively feel the feasibility of the algorithm proposed in this paper during the query process.It also demonstrates the practical value of this paper’s model in industries such as intelligent Q&A,search engines,and chat-bots. |
