Scholarlink

<p>Deep neural network (DNN)-based approaches have rapidly gained prominence as a leading method in automating bug assignment, a task that is both time-intensive and critical for effective bug triage in software development. However, DNNs have been shown to be vulnerable - subtle perturbations to their inputs can lead these models to generate unpredictable and erroneous outputs. To mitigate this issue, contrastive learning (CL) has been increasingly adopted. CL is designed to learn discriminative representations by contrasting similar and dissimilar data points. Although CL has demonstrated its effectiveness in domains such as computer vision (CV) and natural language processing (NLP), its application in automating bug assignments remains unexplored. In this paper, we propose a fixer-level supervised contrastive learning method specifically tailored for automated bug assignment, aiming to enhance the robustness and effectiveness of DNN-based bug assignment approaches. Our approach calculates a similarity score between two bug fixers by assessing the semantic similarity of their historically resolved bug reports. By contrasting bug reports resolved by similar fixers (positive examples) against those addressed by different fixers (negative examples) within the same batch, this approach aims to learn robust bug report representations for bug assignment. We conducted an empirical study to evaluate the effectiveness of our approach against two widely used strategies for improving the robustness of neural networks. This evaluation also includes a baseline scenario where no specific strategy is applied, <i>i.e.</i>, the original network. For this comparative analysis, we utilized three distinct neural network architectures, <i>i.e.</i>, Bidirectional Long Short-Term Memory(Bi-LSTM) using Embeddings from Language Models (ELMo), Bi-LSTM with attention using ELMo, and Bidirectional Encoder Representations from Transformer (BERT), as well as three widely used datasets and a combined dataset composed of the above three datasets. We also investigated the impact of varying important hyperparameters on the considered approaches. Our experimental results show that the proposed approach has varying degrees of improvement over the original networks and those with baselines in terms of top-k (k=1, 5, 10) accuracy and Mean Reciprocal Rank (MRR), ranging from 0.83% to 11.14%. Moreover, all three studied neural networks with the proposed approach have lower degradations than the original networks and those with baselines against the adversarial examples generated by the Projected Gradient Descent (PGD) algorithm. This indicates that our approach can better enhance the robustness of all the networks studied than the baselines. Furthermore, the proposed approach achieves better results in small-size learning settings (5%, 10%, and 15% of the labelled bug reports of each dataset were used for model training).</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Fixer-level supervised contrastive learning for bug assignment

  • 期刊论文

  • 发表日期：

    2025年3月3日

查看全文
  • Rongcun Wang,
  • Xingyu Ji,
  • Yuan Tian,
  • Senlei Xu,
  • Xiaobing Sun,
  • Shujuan Jiang

摘要

Deep neural network (DNN)-based approaches have rapidly gained prominence as a leading method in automating bug assignment, a task that is both time-intensive and critical for effective bug triage in software development. However, DNNs have been shown to be vulnerable - subtle perturbations to their inputs can lead these models to generate unpredictable and erroneous outputs. To mitigate this issue, contrastive learning (CL) has been increasingly adopted. CL is designed to learn discriminative representations by contrasting similar and dissimilar data points. Although CL has demonstrated its effectiveness in domains such as computer vision (CV) and natural language processing (NLP), its application in automating bug assignments remains unexplored. In this paper, we propose a fixer-level supervised contrastive learning method specifically tailored for automated bug assignment, aiming to enhance the robustness and effectiveness of DNN-based bug assignment approaches. Our approach calculates a similarity score between two bug fixers by assessing the semantic similarity of their historically resolved bug reports. By contrasting bug reports resolved by similar fixers (positive examples) against those addressed by different fixers (negative examples) within the same batch, this approach aims to learn robust bug report representations for bug assignment. We conducted an empirical study to evaluate the effectiveness of our approach against two widely used strategies for improving the robustness of neural networks. This evaluation also includes a baseline scenario where no specific strategy is applied, i.e., the original network. For this comparative analysis, we utilized three distinct neural network architectures, i.e., Bidirectional Long Short-Term Memory(Bi-LSTM) using Embeddings from Language Models (ELMo), Bi-LSTM with attention using ELMo, and Bidirectional Encoder Representations from Transformer (BERT), as well as three widely used datasets and a combined dataset composed of the above three datasets. We also investigated the impact of varying important hyperparameters on the considered approaches. Our experimental results show that the proposed approach has varying degrees of improvement over the original networks and those with baselines in terms of top-k (k=1, 5, 10) accuracy and Mean Reciprocal Rank (MRR), ranging from 0.83% to 11.14%. Moreover, all three studied neural networks with the proposed approach have lower degradations than the original networks and those with baselines against the adversarial examples generated by the Projected Gradient Descent (PGD) algorithm. This indicates that our approach can better enhance the robustness of all the networks studied than the baselines. Furthermore, the proposed approach achieves better results in small-size learning settings (5%, 10%, and 15% of the labelled bug reports of each dataset were used for model training).