Research On Severity Prediction Of Multimodal Software Defects Based On Sentiment Probability

Software systems generate a large number of software defects during their lifecycle,and timely detection and fixing of these defects is a key issue to improve software quality and maintain software security.However,due to the different background knowledge between users and developers,defect reports submitted by software users often contain incorrect severity levels,and the daily submission of a large number of defect reports relying solely on manual review mode can no longer meet the demand for timely software defect fixing.Therefore,automated prediction of defect report severity has become a popular research direction to reduce the workload of software maintenance personnel.However,in the existing research on software defect severity prediction,there are problems of coarse prediction granularity;low utilization of defect report information;and inability to be applied to cross-projects.Therefore,this paper proposes a multimodal software defect severity prediction method based on sentiment probability.Specifically,this paper collects hundreds of thousands of complete historical data from the three most commonly used defect tracking repositories,maps them uniformly with different severity levels,and constructs a software defect multi-attribute dataset to address the model’s problems in cross-project prediction.Meanwhile,to address the problem that current sentiment analysis tools are only trained based on non-technical and generic social domains,resulting in language models that are not strong in capturing the sentiment of technical software terms,this study uses the golden annotation criterion of sentiment tree hierarchy classification to construct a manually annotated sentiment dataset of software engineering corpus.Second,this paper proposes the concept of defect knowledge augmentation,which aims to improve the ability of language models to characterize defective texts.The goal of our research is to exploit the domain knowledge required for the defect severity prediction task by customizing the language model to explore its potential help in severity prediction.In addition,we further utilize the augmented language model to quantify the sentiment information in the defect description text to reflect the differences in the impact of defects of different severity levels on users through sentiment probabilities.These sentiment features will be used to enhance the next stage of the defect severity prediction task.Finally,we propose design choices for supervised learning using multimodal defect datasets containing textual,numeric,and category features instead of the single defective textual features in previous studies.We also optimize the structure of the multimodal dataset based on empirical analysis of the severity of the impact of defect report attributes to achieve efficient utilization of defect reports.The key aspects of the approach used in this paper for defect severity prediction include robust data processing of heterogeneous defect data,a complementary integrated learning framework that incorporates large linguistic models and traditional tabular models,and a powerful model integration strategy based on a novel combination of multilayer stacking and cyclic k-fold integration bagging.We compare the method proposed in this paper with the latest defect severity prediction methods,and make and evaluate the validity and reliability of the method with statistical indicators Recall,Precision and F-Measure.The experimental results show that the method has good results in defect severity prediction,which can help maintenance personnel reduce the dependence on users to provide accurate information and achieve general and reliable automatic defect severity prediction.