Project

Application of Recurrent Neural Networks for More Accurate Bug Reports Classifcation

Analyzing and processing bug reports can provide valuable information in software maintenance. However, some bug reports may contain unrelated information or just be a duplicated one. Thus, it will increase the efficiency to flter out those bug reports before performing analysis. In this paper, we propose a new approach for bug report classifcation, based on Recurrent Neural Networks (RNN), which will label bug reports as helpful or unhelpful. This approach achieves the purpose of reducing false positives and then improving the ranking performance by fltering out those unhelpful reports. Our model is tested over 9,000 bug reports from three software projects. The paper starts from how the input data is prepared for the model, how the model works, what the performance of the optimal model is, comparing with other previous models, and the evaluation of results. The evaluation result shows that our model helps improve a state-of-the-art IR-based system’s ranking performance under a trade-off between the precision and the recall. Our comparison experiments show that the RNN achieves the best trade-off between precision and recall than other classifcation models, including Convolutional Neural Network (CNN), multilayer perceptron, and a simple approach that classifies a bug report based on its length. In the situation that precision is more important than recall, our classifcation model helps for bug locating.

Analyzing and processing bug reports can provide valuable information in software maintenance. However, some bug reports may contain unrelated information or just be a duplicated one. Thus, it will increase the efficiency to flter out those bug reports before performing analysis. In this paper, we propose a new approach for bug report classifcation, based on Recurrent Neural Networks (RNN), which will label bug reports as helpful or unhelpful. This approach achieves the purpose of reducing false positives and then improving the ranking performance by fltering out those unhelpful reports. Our model is tested over 9,000 bug reports from three software projects. The paper starts from how the input data is prepared for the model, how the model works, what the performance of the optimal model is, comparing with other previous models, and the evaluation of results. The evaluation result shows that our model helps improve a state-of-the-art IR-based system’s ranking performance under a trade-off between the precision and the recall. Our comparison experiments show that the RNN achieves the best trade-off between precision and recall than other classifcation models, including Convolutional Neural Network (CNN), multilayer perceptron, and a simple approach that classifies a bug report based on its length. In the situation that precision is more important than recall, our classifcation model helps for bug locating.

Relationships

Items