Cost-Effective Detection of Self-Admitted Technical Debt: A Collaborative Framework of Proxy Active Learning and Large Language Models
Abstract During software evolution, developers often document design compromises through code comments, a phenomenon known as Self-Admitted Technical Debt (SATD). Precise SATD identification is vital for reducing maintenance costs and enhancing reliability. Although deep learning has improved detection accuracy, its performance is heavily constrained by the extreme sparsity and class imbalance of SATD instances in real-world scenarios, making large-scale expert annotation cost-prohibitive. Existing active learning attempts struggle to balance deep semantic feature extraction with the exorbitant computational overhead of Large Language Models (LLMs).To address these challenges, this paper proposes a deep active learning approach for SATD detection. First, we design a hybrid query strategy combining uncertainty and representativeness with domain-specific keywords to resolve cold-start and sampling redundancy issues. Second, we introduce a decoupled framework that utilizes BERT as a computationally efficient proxy query engine to localize key samples, which are then used for full-parameter training of the Flan-T5 model. Empirical results demonstrate that our method achieves performance comparable to full-supervision baselines using only 30% of the labeled data, significantly lowering the practical barrier for SATD identification.
Authors
- Tie Bao (ORCID: https://orcid.org/0000-0003-1547-5407)
- Tianxin Lei (ORCID: https://orcid.org/0009-0003-5592-8340)
- Zhicheng Wang
Institutions
- Jilin University (CN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-12
- DOI
- https://doi.org/10.21203/rs.3.rs-9952731/v1
- Primary Topic
- Software Engineering Research
- Type
- preprint