| In recent years,the benefits of data,algorithms,computing power and other elements,artificial intelligence has advanced from technical theory to product application,continuously penetrated into many industries,reshaped the traditional industry model,enabled industrial upgrading,and human society has stepped forward to an intelligent society.During this period,the requirements of computer software are complicated,the scope continues to expand,the demand changes are more frequent and uncertain,and the reconstruction effect of large complex system is not satisfactory.Therefore,the correct classification of the requirements of software or drivers and other applications has become a basic task of software engineering.Text classification of software requirements is not only helpful to the analysis reference in the early stage of software development,but also helpful to prevent the risk of requirement analysis,and more conducive to the recognition and acceptance of software by users.But,domestic and foreign demand for software text feature extraction and figure modeling study is less,also is not clearly defined for the division of functions,etc.,in order to solve these problems,this thesis figure convolution neural network model and Bert,Bert and convolution proposed fusion depth integrated learning model,and applied to the classification of software requirements text data.Through the analysis of software requirement text data,it is concluded that this kind of data has the characteristics of highly discrete,fuzzy and noisy.In addition,the functional characteristics and dependency attributes of software requirements are rarely considered in current research.There is a lack of research on software requirements.This thesis uses crawler technology to obtain a large number of software requirements text data and carried out data analysis and visualization,etc.,and then constructed heterogeneous graph of requirement text data to capture the relationship between requirement words and requirement documents.Firstly,this thesis uses the single model algorithm of deep learning to classify the software requirement text,such as Bert,Text CNN,Text GCN,Text RCNN,DPCNN and other basic models.Meanwhile,Bert combined CNN,CNNPlus,LSTM,ATT,RCNN,DPCNN and other models and used them for software requirement classification experiment.This thesis explores different combinations of deep learning algorithms to explore the relationship between software requirement texts.Then,this thesis proposes a new model BTM Topic Residual Graph Convolutional Neural Network(BTRGCN),the model using the rich world word co-occurrence information,revealing the theme and the potential for building software requirement text heterogeneous figure,better capture the relationship between the demand and the requirements document and theme,and residual learning network degradation.The combination of BTRGCN and Bert can better classify the software requirement text and improve the accuracy of the model.Then,this article explores different figure convolution neural network to the software requirements text classification experiment contrast,and combines different figure convolution and Bert after research on classification of software requirements,effective for different models of the demand for this article data classification effect,and integrate a variety of deep learning algorithm,and the method of weight set voting can improve the classification performance of the software requirements classification model in this thesis.Finally,all the models are analyzed and studied experimentally and applied to the software requirements text data set.The weighted average was 97.05% on Windows＿a data set and 95.77% on Windows＿b data set is achieved,which is better than Bert’s fusion models.It is better than the graph convolution and Bert fusion model,which proves The effectiveness of deep ensemble learning software requirement classification based on Bert and graph convolution is demonstrated,provides corresponding reference for software engineering development,and helps researchers better obtain requirements from other fields. |
