Software Bug Severity Prediction Using Convolutional Neural Network and BiLSTM Models

Creating bug-free software is nearly impossible, but addressing bugs promptly can significantly improve the software development process. Early bug resolution enhances software quality, efficiency, and reliability, although it’s costly during maintenance and development. As software requirements grow and delivery times shrink, ensuring quality becomes challenging. Bugs are a given, as error-free software is unfeasible, and testing is complex and time-consuming. The complexity of software and its development, along with insufficient testing, means developers can’t guarantee bug-free products. Managing bugs is a critical yet arduous part of software maintenance, with bug repositories receiving numerous reports daily, each containing textual and metadata information about the bug. Prioritizing bugs by severity is essential but manually handling the vast number of reports is daunting. High-severity bugs should be addressed before those with lower severity. This paper explores using Convolutional Neural Network (CNN) and Bidirectional LSTM (BiLSTM) with Glove Embedding on Eclipse and Mozilla datasets for automated bug severity prediction. The proposed model’s effectiveness is assessed by accuracy score, confusion matrix, and classification report. The results indicate that BiLSTM performs better than CNN, achieving 91% accuracy on the Mozilla dataset and 89% on the Eclipse dataset.