Role of Context in Learning-Based Source Code Tasks - V2 скачать в хорошем качестве

Role of Context in Learning-Based Source Code Tasks - V2

As software becomes more complex, reducing manual effort in software development is increasingly important. Artificial intelligence (AI) can help with tasks such as code completion and bug fixing, but current tools often use context in an ad hoc manner, which limits their effectiveness. In this talk, I discuss techniques to improve AI-assisted software development by leveraging different forms of \emph{context}, such as code relationships, relevant examples, and bug-reproduction tests. Software developers often perform code changes that require an understanding of multi-location context. I introduced metrics to characterize the complexity of such multi-location code changes. Empirical evaluation shows that our techniques improve upon the state of the art by systematically selecting context. The proposed metrics may help predict whether AI models are likely to succeed on tasks that require reasoning across multiple code locations.