Leveraging multi-task learning to fine-tune RoBERTa for self-admitted technical debt identification and classification

Abstract

Self-Admitted Technical Debt (SATD) detection aims to identify whether a code comment explicitly admits technical debt and classify its specific type. Existing research largely treats identification and classification as separate tasks, with classification-focused approaches suffering from Out-Of-Vocabulary (OOV) issues and relatively low macro-averaged $F_1$-score. To address these challenges, this paper presents FRoM, a unified and efficient approach that integrates SATD identification and classification into a single pipeline. Specifically, FRoM employs a byte-level tokenizer to effectively mitigate OOV problems and leverages multi-task learning to fine-tune a pre-trained model for improved classification performance. Additionally, FRoM incorporates a novel undersampling technique to remove semantically similar non-SATD samples, reducing the time required for fine-tuning. Empirical evaluations on two datasets, comprising 38,902 and 2,528 comments respectively, demonstrate that FRoM achieves state-of-the-art performance in both identification and classification tasks. Furthermore, a case study highlights that our deployed tool, FRoMD, exhibits competitive performance compared to ChatGPT-4o. The dataset and the code are available at https://github.com/HduDBSI/FRoM.