Seml: A Semantic LSTM Model for Software Defect Prediction

Software defect prediction can assist developers in finding potential bugs and reducing maintenance cost. Traditional approaches usually utilize software metrics (Lines of Code, Cyclomatic Complexity, etc.) as features to build classifiers and identify defective software modules. However, software metrics often fail to capture programs' syntax and semantic information. In this paper, we propose Seml, a novel framework that combines word embedding and deep learning methods for defect prediction. Specifically, for each program source file, we first extract a token sequence from its abstract syntax tree. Then, we map each token in the sequence to a real-valued vector using a mapping table, which is trained with an unsupervised word embedding model. Finally, we use the vector sequences and their labels (defective or non-defective) to build a Long Short Term Memory (LSTM) network. The LSTM model can automatically learn the semantic information of programs and perform defect prediction. The evaluation results on eight open source projects show that Seml outperforms three state-of-the-art defect prediction approaches on most of the datasets for both within-project defect prediction and cross-project defect prediction.