Deakin University
Browse

File(s) under permanent embargo

Detecting multi-type self-admitted technical debt with generative adversarial network-based neural networks

journal contribution
posted on 2023-03-31, 00:17 authored by Jiaojiao Yu, Xu Zhou, Xiao LiuXiao Liu, Jin Liu, Zhiwen Xie, Kunsong Zhao
Context: Developers often introduce the self-admitted technical debt (SATD), i.e., a compromised solution to satisfy the delivery of the current goals, in code comments but do not eliminate them timely in the following software development and maintenance process. Automatically identifying the SATDs to reduce potential harm to software has attracted the attention of researchers. However, existing approaches only identified SATDs at a coarse-grained level, which impacts developers to locate and remove them. Objective: This paper proposes a novel model named GCF, which is a deep learning method to enhance the performance of multi-type SATD classification based on generative adversarial network. Method: The GCF model employs the JSD Generative Adversarial Network to solve the imbalance problem, utilizes CodeBERT to fuse information of code snippets and natural language for initializing the instances as embedding vectors, and introduces the feature extraction module to extract the instance features more comprehensively. Results: The experimental results show that, the GCF model obtains better performance compared with the state-of-the-art method. Moreover, experiments on the GCF model variants and others with different GAN models show the superiority of the GCF model. Conclusion: Our proposed GCF model effectively solves the problem of imbalanced types of SATD, fuses the information of code snippets and natural language, and extracts key features to achieve outstanding performance in detecting multi-type SATD. Therefore, the GCF model is an effective method for detecting multi-type SATD.

History

Journal

Information and Software Technology

Volume

158

Article number

107190

Pagination

1-15

Location

Amsterdam, The Netherlands

ISSN

0950-5849

Language

eng

Publication classification

C1 Refereed article in a scholarly journal

Publisher

Elsevier

Usage metrics

    Research Publications

    Exports

    RefWorks
    BibTeX
    Ref. manager
    Endnote
    DataCite
    NLM
    DC