SHF:Medium: Studying and Exploiting the Bimodality of Software

编写的软件有两个受众：运行它的机器和维护它的人。机器理解正式的操作语义;然而，人类通过变量名、注释、编码风格等中隐藏的非正式、嘈杂的信息来理解代码。人类读者当然可以在脑海中“模拟”代码执行，但这并不容易。因此，开发人员（同时总是努力确保代码具有正确的计算语义）尝试编写易于阅读和理解的代码。因此，代码是双峰的，在两个通道中携带信息：1）一个精确的，正式的，将确切的计算意义传递给执行平台，2）一个非正式的“自然”的，为人类读者设计，以促进人类理解的快速，嘈杂的过程。该项目旨在利用这两个渠道及其协同作用，以多种方式改进软件工具。首先，正式渠道将用于为非正式渠道生成训练数据，以训练工具“宽松地”解析由学生开发人员创建或在StackOverflow和其他非正式文档中找到的不成熟的错误代码。第二，正式渠道（例如，类型检查器）将用于训练概率学习器以在重复学习设置中创建类型注释。第三，将使用信号（例如，变量名），以帮助静态分析算法进行精度和速度的权衡;最后，代码的双峰性将被用来创建微调的代码样本，这些代码样本保留了意义（但形式不同），用于旨在提高人类代码理解和生产的心理语言学研究。该项目将开发一个新的入门编程课程，教导初学程序员学习“双峰”思考需要IT解决方案的问题：不仅仅是计算抽象，而且更全面地考虑对社会和环境的影响。该项目是由一个跨学科的团队，包括学者在软件工程，心理语言学，科学和技术研究。这个奖项反映了NSF的法定使命，并已被认为是值得通过评估使用基金会的智力价值和更广泛的影响审查标准的支持。

Software, as written, has two audiences: the machine(s) that run it, and the human(s) who maintain it. Machines understand formal operational semantics; humans, however, understand code via the informal, noisy information latent in variable names, comments, coding styles etc. Human readers can certainly mentally "simulate" code execution, but it’s not easy. Developers therefore (while always striving to ensure that the code has the correct computational semantics) try to write code  that  is easy to read and understand. Code is thus bimodal, carrying information in two channels: 1) a precise, formal one that carries the exact computational meaning to the execution platform, and 2) an informal ``natural" one for human readers that is designed to facilitate the fast, noisy process of human comprehension. This project aims to exploit these two channels, and their synergies, to improve software tools, in several ways. First, the formal channel will be used to generate the training data for the informal channel in order to train tools to "leniently" parse inchoate, erroneous code, as created by student developers or found on StackOverflow and other informal documentation. Second, the formal channel (e.g., a typechecker) will be used to train probabilistic learners to create type annotations in a reinforcement-learning setting. Third, the use of learned models will be studied using signals (e.g., variable names) on the informal channel to help static analysis algorithms make precision-speed tradeoffs; finally, the bimodality of code will be leveraged to create finely-tuned code samples that preserve meaning (but are different in form) for pyscho-linguistics studies aimed at improving human code comprehension and production. The project will develop a new introductory programming curriculum that teaches beginning programmers to learn to think "bimodally" about problems requiring IT solutions: not just in terms of computational abstractions, but also more holistically, about the impacts on society and the environment. The project is enabled by an interdisciplinary team, consisting of scholars in Software Engineering, Psycho-linguistics, and Science and Technology studies.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.