| Crowdsourced testing is a new software testing mode.Based on crowdsourcing,it distributes the testing tasks originally performed by professional testers in the company to the crowdworkers through crowdsourced platform,and the crowdworkers give feedback the results by submitting bug reports.Compared with traditional software testing,crowdtesting has the characteristics of short test cycle and low test cost,so it is widely used.Because the nature of Crowdtesting depends on a large amount of human resources,and the crowdworkers are independent of each other,there are often a large number of duplicates in the submitted bug reports,the research on the processing of these duplicate bug reports becomes an important task.In this thesis,we focus on the duplicate bug reports processing for crowdtesting scenarios.The main contributions include the following aspects:(1)Duplicate bug reports detection based on text analysis.For test reports in textual form,there will be sparse text problem.From the perspective of text clustering,a duplicate bug reports detection model based on deep clustering framework was proposed.It can be divided into two parts: 1)text representation layer;2)Clustering layer.In the text presentation layer,the Bi-LSTM is used as the feature extractor of the report text,and the reconstruction loss of the autoencoder is introduced as the loss of the network layer.In the clustering layer,K-means clustering algorithm is used to cluster the extracted text features,KL divergence and maximum partition confidence are combined as the clustering loss.Experimental results on a real crowdtesting data set show that the performance of the proposed method is better than the benchmark method.(2)Duplicate bug reports detection by integrating information from text and image.In some scenarios,the bug report has multi-modal data forms,such as the scene of crowdtesting of software based on mobile application,the bug reports submitted by crowdworkers often presents in the form of less text and more screenshots.For bug reports with text and image,it is difficult of how to consider the information of text and screenshots at the same time.In the second work of this thesis,the idea of modal fusion is used for reference.Based on the model in the first work,the image feature extraction layer and the modal fusion layer are added.The image presentation layer is mainly composed of VGG16 network with SEnet module,and the modal fusion layer is used to fuse the image features and text features.The loss of the reconstruction of the autoencoder and the loss of DCCA are used as the loss of the network layer.The experimental results show that this method is superior to the contrast method in several evaluation indexes.(3)Fusion of bug duplicate reports.Although duplicate bug reports describe the same bug,there is complementary information between them.Therefore,the goal of duplicate reports fusion is to remove redundant information and extract effective information.From the perspective of abstractly textual summarization,In this thesis models it as a sentence selection problem and proposes a centroid based abstractly textual summarization model,which constructs the centroid by TF-IDF value and Word Embedding.Meanwhile it constructs the sentence vector by SIF technology and performs spectral clustering in the sentence graph constructed based on sentence similarity.Finally,the sentence closest to the center of mass from each class is selected to generate a bug summary.Experimental results show that the performance of this method is significantly improved compared with the existing methods. |
