Leveraging Neural Language Model for Automated Code Quality Issue Identification

Anonymous

Abstract: The usefulness of machine learning techniques for understanding source code and assisting with software engineering tasks have been demonstrated by recent progress in the community. More specifically, language models (LMs) on code have been originally developed based on n-grams. One of the limitations of such models is the Out-Of-Vocabulary issue which abounds in code-related applications. Recent advances in deep learning have provided new and more powerful tools for source code modeling. For instance, RNN and transformer based models have been built to learn embeddings of codes for various downstream tasks such as code completion, summarization and translation. However, the power of neural language models has not been well leveraged for improving the quality of code.