Advancing program representation and leveraging big code for automatic program transformation
This work explores the potential of leveraging the large amount of available code in open-source repositories and community question-answering sites, collectively known as Big Code, to automate and enhance program transformation processes. With billions of lines of code across various programming languages and styles, the opportunity to enhance software development workflows through code reuse is significant. This research focuses on two primary tasks: code translation and code style transfer, aiming to streamline these processes by utilizing automated methods upon the rich resources of Big Code. The core component of this work is a novel code search engine designed to facilitate the retrieval of code snippets that, while functionally similar, differ in language or stylistic features. This engine utilizes a unique feature representation focused exclusively on content features, enabling precise and efficient identification and retrieval of relevant code snippets from the extensive Big Code databases. This approach not only addresses the challenges of program transformation but also capitalizes on the abundance of existing code to enhance development efficiency and effectiveness. This work introduces methodologies for both code translation and code style transfer. For code translation, we propose a retrieval-based unsupervised approach and data augmentation methods for a supervised approach with data derived from Big Code, aimed at improving the accuracy and efficiency of translating code between languages. We enhanced the program feature representation by embedding the features of the target language using auto-encoders. For code style transfer, we develop a system that can either directly retrieve code snippets in the desired style or generate new code snippets through a deep learning-based model, thereby facilitating the adaptation of code to various stylistic guidelines without altering its functionality. We enhanced the program feature representation by encoding the content features of the input code and the style features of the target style individually through a Siamese network. Through extensive experimentation, we evaluate the performance of our systems, demonstrating their effectiveness, efficiency, and scalability. Our findings indicate that leveraging Big Code and enhanced program representation for automated program transformation can significantly reduce manual coding effort and foster the dissemination of coding best practices.
