Uso de modelos de linguagem para identificação de code smells em typescript

Abstract

Code smells are indicators of potential quality issues in source code that, although not directly affecting software correctness, negatively impact maintainability, evolution, and readability. Currently, the detection of these issues is predominantly performed using static analysis tools, which rely on structural rules and syntactic metrics. While widely adopted, such approaches present limitations, particularly in cases where code understanding depends on semantic and contextual aspects that are not fully captured by purely static analyses. In this context, this work investigates the use of language models for the automatic detection of code smells in TypeScript source code, comparing this approach with traditional static analysis tools. Initially, public GitHub repositories were collected and filtered based on relevance and language predominance criteria. Subsequently, SonarQube was employed to identify code smells, focusing on the rule Cognitive Complexity of functions should not be too high. The detected instances were then subjected to manual analysis, enabling the construction of a more accurately labeled dataset. Using this dataset, two CodeBERT-based models were trained: one using supervised f ine-tuning only, and another combining Masked Language Modeling with fine-tuning. The trained models were evaluated using standard classification metrics and directly compared with SonarQube’s performance. The results show that the language model achieved better performance in identifying code smells in scenarios where purely static analysis is limited, while SonarQube demonstrated higher precision in identifying clean code. These findings indicate that hybrid approaches combining language models and static analysis tools represent a promising direction for more robust code smell detection.