Automated Prediction of Bug Reports' Severity using Attention-based Convolutional Neural Network
Myint Theingi, Naw Lay Wah · 2025
Bug report severity attribute is critical for developers in deciding the level of impact or seriousness of a bug. Developers must accurately specify the severity level because the impact of bugs affects the Bug Triage process. This process is essential in maintaining and developing software. Researchers proposed many related works with bug severity prediction to improve the software maintenance process. Most of the previous studies use many different techniques in the determination of severity levels. However, these works still need to improve the prediction performance. This work attempts to use an attention-based Convolutional Neural Network (CNN) to estimate bug reports' severity levels efficiently. Firstly, the unstructured text for each bug report from Eclipse and Mozilla projects of the Bugzilla Bug Tracking System (BTS) is preprocessed by using Natural Language Processing techniques. Then, the preprocessed text is represented as embedded vectors using the word2vec (skip-gram model). Finally, the vectorized inputs feed into the attention-based CNN model to extract features and the model predicts the levels of severity for each bug report. The proposed model develops the prediction performance by adding a scaled dot-product attention mechanism into a simple CNN model. In this paper, accuracy, precision, recall, and f1-score are used to assess the performance using a confusion matrix. The experimental findings proved that the attention-based CNN model enhances the prediction metrics by up to 1.09%, 3.93%, 1.09%, and 3.29% respectively by using an attention layer in the conventional CNN model.