Leveraging Code Language Models for Code Smells Detection

Abstract

Improving software quality remains a major challenge in modern software engineering. Among the different techniques used to address this issue, code smell detection plays an important role, as code smells may negatively affect software maintainability, readability, and evolution. Recent advances in machine learning, particularly pre-trained code representation models, have opened new perspectives for automating this task. In this study, code smell detection is formulated as a multi-label classification problem in order to reflect realistic scenarios where multiple smells may co-occur within the same code fragment. To enable the evaluation of pre-trained code representation models in a multi-label setting, six multi-label datasets containing source code were derived from two existing datasets, namely MLCQ and Fontana, by merging instances sharing common code smells. The study investigates multiple experimental configurations based on CodeBERT, GraphCodeBERT, and CodeT5, including direct transformer-based classification, sliding window processing for long code sequences, and hybrid approaches combining transformer embeddings with classical machine learning classifiers. Among the evaluated configurations, the hybrid approach based on CodeBERT embeddings combined with a Random Forest classifier achieved the best overall performance, reaching an accuracy of 0.91 and a micro-F1 score of 0.90 on the Fontana datasets.