SHF: Medium: Principled Co-Reasoning of Software and Natural-Language Artifacts

软件无处不在。软件通常包含大量用自然语言（NL）编写的工件，包括代码注释、更改日志、手册页、代码中的常量字符串以及变量和函数名。软件NL工件包含代码工件中经常缺少的大量语义信息。在分析NL工件并在广泛的软件工程应用程序中利用它们方面已经有了大量的现有工作。然而，大多数现有的工作都是特别的，并且在通用性上受到限制。现有的工作通常将NL工件视为附加信息的来源，而不是分析操作的一流对象（如程序分析中的变量类型），从而错过了充分利用软件NL工件的机会。因此，该项目开发了代码和NL工件的联合分析，将NL工件视为一级对象。除了提高技术水平之外，项目中开发的原则、基础结构和技术是变革性的，提供了具有教育意义和实用的工具来生成高质量的源代码和软件文档。这些技术改进了程序分析、软件维护、软件可靠性和工程生产力，以降低软件开发成本，改善工作和娱乐生活，在这些方面软件是不可或缺的。该项目开发了一种原则性和复杂的软件推理方法，结合了NL分析和程序分析。它自动对各种NL工件建模和分类，并将它们归为相关的代码元素。因此，它们就像程序分析中的其他经典对象（例如变量和语句）一样，成为一级对象。可以对它们进行推断、传播、更新、关联和形式化推理，以最大限度地利用它们丰富的语义（例如，注释可以传播到以前没有通过程序分析进行注释的代码元素）。项目活动包括(1)建模、分类和归属NL工件，通过开发特定领域的语言模型来处理、建模、分类NL工件并将其归属于相应的代码元素；(2)构建统一的表示、传播和NL工件和代码工件的协同推理；(3)产生高度精确和可扩展的概率推理。通过利用概率图模型来执行代码和NL工件的统一推理，以及(4)在包括软件测试在内的领域探索联合分析的新应用。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Software is pervasive. Software typically contains a large volume of artifacts written in natural languages (NL), including code comments, change logs, manual pages, constant strings in code, and variable and function names. Software NL artifacts contain a wealth of semantic information that is often missing in code artifacts. There has been substantial existing work on analyzing NL artifacts and leveraging them in a wide range of software-engineering applications. However, most existing work is ad hoc, and is limited in its generality. Existing work typically considers NL artifacts as sources for additional information instead of first-class objects on which analysis operates (like variable types in program analysis), missing the opportunity to take full advantage of software NL artifacts. Thus, this project develops co-analysis of code and NL artifacts, which treats NL artifacts as first-class objects. In addition to advancing the state of the art, the principles, infrastructure, and techniques developed in the project are transformative, providing educational and practical tools to generate high-quality source code and software documents. These techniques improve program analysis, software maintenance, software reliability, and engineering productivity, for lower software development cost and better work and recreational lives, where software is indispensable.The project develops a principled and sophisticated software reasoning method that couples NL analysis and program analysis. It automatically models and classifies various kinds of NL artifacts, and attributes them to the related code elements. As such, they become first-class objects just like other classic objects in program analysis (e.g., variables and statements). They can be inferred, propagated, updated, associated, and formally reasoned about, to maximize the utilization of their rich semantics (e.g., comments can be propagated to code elements that are not previously commented through program analysis). The project activities include (1) modeling, classifying, and attributing NL artifacts, through developing domain-specific language models to process, model, classify NL artifacts and attribute them to the corresponding code elements, (2) building uniform representation, propagation, and co-reasoning of NL artifacts and code artifacts, (3) producing highly accurate and scalable probabilistic inference, by leveraging probabilistic graph models to perform the uniform reasoning of both code and NL artifacts, and (4) exploring new applications of co-analysis in domains including software testing.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.