DeMoCo: Developer-Centered, Neural Models of Code

神经软件分析从大型代码语料库中学习预测模型，以解决具有挑战性的软件工程任务。近年来，它的发展势头越来越强劲，补充了传统的程序分析，有时甚至超过了传统的程序分析。这些技术的核心是代码的神经模型，即，深度学习模型可以对程序及其属性进行推理，以做出对开发人员有用的预测。不幸的是，当前的代码神经模型主要由容易获得的数据驱动，例如，阅读成千上万的源代码文件，从第一个到最后一个令牌，他们作出的预测是很难理解的人类，例如，将整个方法归类为缺陷而不作进一步解释。因此，许多当前的技术实现了令人印象深刻的准确性，但对开发人员来说仍然使用有限。该提案将人类开发人员置于代码神经模型的中心，从以数据为中心的范式转变为以开发人员为中心的范式。具体来说，我们计划进行三方面的研究，这将（i）增加我们对人类推理和神经推理如何相互关联的理解，（ii）设计代码的神经模型，模仿开发人员如何推理和探索代码，以及（iii）创建模型，不仅预测代码的属性，而且还向开发人员解释预测。以编译器为中心的代码神经模型将在广泛的软件工程任务中具有潜在的应用前景。作为具体的例子，本建议将它们应用于缺陷检测和故障定位。

Neural software analysis learns predictive models from large code corpora to address challenging software engineering tasks. It has been gaining momentum over recent years, complementing and sometimes even outperforming traditional program analysis. At the core of these techniques are neural models of code, i.e., deep learning models that reason about programs and their properties to make predictions useful to developers. Unfortunately, current neural models of code are mostly driven by what data is easily available, e.g., reading thousands of source code files from the first to the last token each, and they make predictions that are difficult to understand for humans, e.g., by classifying an entire method as buggy without further explanation. As a result, many current techniques achieve impressive accuracy but still remain of limited use to developers. This proposal puts the human developer into the center of neural models of code, shifting from a data-centered paradigm to a developer-centered paradigm. Concretely, we plan to pursue three strands of research, which will (i) increase our understanding of how human reasoning and neural reasoning about programs relate to each other, (ii) design neural models of code that imitate how developers reason about and explore code, and (iii) create models that not only predict properties of code but also explain the predictions to developers. Developer-centered neural models of code will be potentially applicable in a wide spectrum software engineering tasks. As concrete examples, this proposal will apply them to bug detection and fault localization.