SemirFL: Boosting Fault Localization via Combining Semantic Information and Information Retrieval

Abstract

Automated fault localization aims to reduce software maintenance's workload during software development's evolution. Applying different features extracted from bug reports and source files can help locate faults. However, these approaches consider programming languages as natural when measuring similarity features, considering only precise term matching and ignoring deep semantic similarity features. Furthermore, existing bug localization approaches need to utilize the structural information extracted from source files, where program languages have unique structural features compared to natural languages. In this paper, we proposed SemirFL, a model combining both Convolutional Neural Network(CNN) and revised Vector Space Model(rVSM), which is feeded with four metadata features (bug-fixing recency, bug-fixing frequency, collaborative filtering score, and class name similarity). SemirFL has been studied on four open-source projects. The experimental results show that SemirFL can significantly outperform the existing representative techniques in locating faults in the buggy source files.