Improving Bug Severity Prediction With Domain-Specific Representation Learning

Automating the process of bug severity assignment can accelerate bug triagers' efficiency in the life-cycle of software maintenance, improving the quality of software products. The mainstream approaches for bug severity prediction mainly use different neural networks due to their automated lear...

Full description

Saved in:
Bibliographic Details
Published in:IEEE access 2023, Vol.11, p.62829-62839
Main Authors: Wei, Ye, Zhang, Chunfu, Ren, Teng
Format: Article
Language:English
Subjects:
Automation
Bug severity prediction
Computer bugs
Context
Debugging
domain-specific pre-training
Learning
Neural networks
pre-trained language model
Predictive models
Representation learning
Semantics
Software
Software maintenance
Task analysis
Training
Transformers
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Automating the process of bug severity assignment can accelerate bug triagers' efficiency in the life-cycle of software maintenance, improving the quality of software products. The mainstream approaches for bug severity prediction mainly use different neural networks due to their automated learning ability. However, there are two problems that make existing approaches fail to predict severities for some bugs: 1) they cannot learn the internal knowledge of bug reports; 2) supervised training is difficult to understand the global context of bug reports. To resolve these two problems, in this paper, we propose a bug severity prediction approach, namely KICL, which combines pre-trained language models and domain-specific pre-training strategies, i.e., Knowledge-Intensified pre-training and contrastive learning pre-training. Specifically, Knowledge-Intensified allows KICL to learn project-specific bug report tokens, deeply understanding internal knowledge of bug reports. As for contrastive learning, it allows KICL to perform sequence-level learning, understanding bug reports from the perspective of the global context. When finishing pre-training, we can fine-tune pre-trained KICL for bug severity prediction. To evaluate the effectiveness of KICL, we choose six baseline approaches and compare their performance on a public dataset. The experimental results show that KICL outperforms all baseline approaches by up to 30.68% in terms of weighted average F1-score, achieving new results for bug severity prediction.
ISSN:2169-3536
2169-3536
DOI:10.1109/ACCESS.2023.3279205