Fine-Tuning UniXcoder for Code Smell Detection in Java Projects

Abstract

Code smells are symptoms of poor software design that hinder maintainability and increase technical debt interest. Automated detection of code smells is crucial in software development industry, where manual inspection is costly and error-prone. Although several studies have explored code smell detection, most existing approaches, both structural/metric-based and textual/transformer-based, still exhibit limitations in capturing the semantic context of source code to handle complex smells effectively. This study investigates the application of UniXcoder, a pre-trained transformer model for source code, to classify Java source code methods across multiple projects as smell or clean, with a particular focus on the Switch Statements smell. The proposed method fine-tunes UniXcoder and evaluates its performance using standard metrics, namely accuracy, precision, recall, and F1-score, in comparison with traditional machine learning models based on structural metrics and textual representations. Experimental results show that the fine-tuned UniXcoder achieves the highest classification accuracy of 80.2%, the highest precision among all baseline methods evaluated, indicating its effectiveness in minimizing false positives. These findings confirm that pre-trained transformer models, particularly UniXcoder, are capable of understanding both the semantic and syntactic structures of code and demonstrate stronger performance than frequency-based textual representation approaches as well as software metrics-based approaches. Furthermore, the results highlight UniXcoder’s potential for integration into industrial software engineering workflows, supporting automated code smell detection as part of continuous code quality assurance.