Statistical Learning for Inference between Implementations and Documentation

Abstract

API documentation is useful for developers to better understand how to correctly use the libraries. However, not all libraries provide good documentation on API usages. To provide better documentation, existing techniques have been proposed including program analysis-based and data mining-based approaches. In this work, instead of mining, we aim to generate behavioral exception documentation for any given code. We treat the problem of automatically generating documentation from a novel perspective: statistical machine translation (SMT). We consider the documentation and source code for an API method as the two abstraction levels of the same intention. We use SMT to translate documentation from source code and vice versa. Our preliminary results show that the direction of statistical learning for inference between implementations and documentation is very promising.