Using API-Embedding for API-Misuse Repair

Application Programming Interfaces (APIs) are a way to reuse existing functionalities of one application in another one. However, due to lacking knowledge on the correct usage of a particular API, developers sometimes commit misuses, causing unintended or faulty behavior. To detect and eventually repair such misuses automatically, inferring API usage patterns from real-world code is the state-of-the-art. A contradiction to an identified usage pattern denotes a misuse, while applying the pattern fixes the respective misuse. The success of this process heavily depends on the quality of the usage patterns and on the code from which these are inferred. Thus, a lack of code demonstrating the correct usage makes it impossible to detect and fix a misuse. In this paper, we discuss the potential of using machine-learning vector embeddings to improve automatic program repair and to extend it towards cross-API and cross-language repair. We illustrate our ideas using one particular technique for API-embedding (i.e., API2Vec) and describe the arising possibilities and challenges.