In-depth exploration of software defects and self-admitted technical debt through cutting-edge deep learning...

Area of Science:

• Software Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Existing research often addresses Self-Admitted Technical Debt (SATD) or bug detection in isolation.
• Current methodologies frequently overlook advanced deep learning techniques for concurrent SATD and defect analysis.
• There is a need for integrated approaches to understand and manage software quality issues holistically.

Purpose of the Study:

• To develop and evaluate an innovative deep learning-based method for the concurrent identification and classification of SATD and software defects.
• To enhance the understanding and localization of defects within software comments associated with SATD.
• To improve software quality assessment and maintenance processes through a unified approach.

Main Methods:

• Utilized deep learning architectures including LSTM, BI-LSTM, GRU, BI-GRU, and Transformer models (BERT, GPT-3).
• Trained models on diverse datasets from software repositories like Apache, Mozilla Firefox, and Eclipse.
• Employed data analysis, preprocessing, and machine learning techniques for model development and evaluation.
• Incorporated transfer learning with GPT-3 for enhanced performance.

Main Results:

• Deep learning models achieved high performance metrics, with accuracy and precision around 0.98.
• Transformer models, particularly GPT-3, demonstrated slightly superior performance, achieving an accuracy of 0.984.
• Transfer learning with GPT-3 resulted in an accuracy of 0.96 and F1-Score of 0.96, outperforming other models.
• The proposed method significantly improved upon existing methodologies.

Conclusions:

• The developed deep learning approach effectively identifies and classifies both SATD and software defects concurrently.
• The method offers significant implications for software engineering, improving quality assessment, maintenance prioritization, and resource allocation.
• This research facilitates the creation of more reliable and sustainable software systems.