Inside-Out Code Auto-Completion

Auto-complete has become a ubiquitous feature in modern-day typing systems. Traditional systems typically try to predict the next characters or words the user is about to type. In contrast, keyword-based auto-completion systems expect the user to type keywords, then fill in the least informative parts of the sentence, and have been successful for a variety of natural language tasks. In this project, we implement a keyword-based auto-complete technique for programming languages, as auto-complete systems in modern IDEs are still mostly left-to-right. We use three programming languages with different verbosity levels: Python, Haskell and Java. We evaluate different encoding schemes paired with a neural decoder with various trade-offs in translation accuracy and compression, and evaluate their accuracy and robustness.