Enhancing structural knowledge in code smell identification: A fusion learning framework combining AST-based metrics with semantic embeddings

Abstract

Identifying code smells is a crucial task in software engineering that aims to uncover potential problems and bad practices in source code. Existing learning-based approaches have achieved good results in identifying code smells by learning features such as software code metrics, syntax, and semantics. However, some gaps in existing research still need to be addressed: (1) Software code metrics are challenging to extract and vary across different levels of code granularity; (2) Highly abstract code semantics rarely convey the structural details of the code. To address these issues, we propose using Abstract Syntax Tree (AST)-based metrics to replace software code metrics for identifying code smells. The proposed AST-based metrics are easy to extract, treat code at all granularity levels uniformly, and precisely describe the structural details of the code. Additionally, we propose a fusion learning framework that combines AST-based metrics and semantic embeddings to identify code smells and their severity. Extensive experimental results reveal that our proposed AST-based metrics have the potential to replace software code metrics in identifying code smells, and the proposed fusion learning framework outperforms state-of-the-art approaches on the same dataset.