A3-CodGen introduces a novel framework for code generation that leverages information from the code repository. By incorporating local-aware, global-aware, and third-party-library-aware knowledge, the framework aims to generate more accurate and efficient code with fewer errors and redundancies. The research highlights the importance of context-aware code generation tools in addressing evolving demands in software development. The study evaluates different configurations of knowledge extraction to optimize code reuse and effectiveness.
The paper discusses the limitations of existing code generation approaches and emphasizes the need for models to be aware of the working context within a code repository. It explores how A3-CodGen bridges this gap by providing three types of repository knowledge: local functions, global functions, and third-party libraries. The framework systematically mines and utilizes external knowledge to generate code that closely resembles human developers' work.
The experiments conducted aim to investigate the effectiveness of different types of repository-aware knowledge in assisting models with code generation tasks. The study introduces a new benchmark dataset called RepoEval to evaluate A3-CodGen's performance accurately.
다른 언어로
소스 콘텐츠 기반
arxiv.org
더 깊은 질문