One Technique to Smell Them All: A Textual-based Technique for Code and Test Smell Detection

Bad smells are symptoms of poor design or implementation choises that may affect both production and test code. Several previous studies demonstrated the huge impact of these smells on change- and fault-proneness, comprehensibility and, more in general, on maintainability. These are the reasons why researchers spent a lot of effort on the definition of metrics-based techniques able to detect them by exploiting structural properties of the source code. However, code and test smell detection have always been treated as separate problems, since for detecting different smells, different tools relying on a different ah-hoc set of metrics are required. To overcome such limitations, we propose a textual-based smell detector able to detect both code and test smells through IR methods measuring the probability that a code component is affected by a given smell. We run the proposed approach, named TACO (Textual Analysis for Code Smell Detection), on 20 open source projects, comparing its performances with existing state-of-the-art smell detectors purely based on structural analysis. Our findings indicate that TACO’s precision ranges between 57% and 75%, while its recall ranges between 68% and 85%. Moreover, we observed that the proposed detector often outperforms competitive structural approaches, being able to identify smell instances that can not be detected by solely using structural analysis.