Collaborative Research: SHF: Medium: Toward Understandability and Interpretability for Neural Language Models of Source Code

人工智能（AI）的进步导致软件开发人员开发了几种新型工具，旨在帮助自动化构建和维护软件的软件开发过程的各个部分。然而，复杂的底层深度学习模型和大量训练数据集的结合使得很难解释为什么这些模型以及由它们驱动的开发工具会以这样的方式运行。考虑到这些工具在软件工程（SE）中开始扮演的日益重要的角色，必须开发允许涉众更好地理解和使用这些工具的技术，以便能够维护关键的软件基础结构。该项目将开发一个框架和方法，使构建人工智能开发工具的研究人员和使用这些工具的软件工程师能够解释底层模型做出预测的原因。其目的是让研究人员获得详细的见解，了解为什么模型可能没有按照预期执行，从而允许有针对性的改进和新模型的知情创建。该方法将被集成到人工智能驱动的软件开发工具中，允许软件工程师做出明智的决定，判断工具的建议何时可能是有益的还是有害的，从而建立对其使用的信任。可解释性框架还将支持与这些工具交互的新形式，提供一种随时间改进的自然语言反馈机制。该项目将制作和传播有关构建和使用人工智能编程工具的最佳实践的教育材料。这些材料的目的是将其纳入各级教育现有的计算机知识课程。此外，该项目将侧重于从传统上代表性不足的类别中招募和留住计算机科学专业的学生。这个项目有三个具体目标。首先，它将设计一种自动化方法，用于生成源代码“上下文无关”神经语言模型行为的全局解释。该项目的这一部分将使用因果推理理论将大型语言模型的预测映射到人类可解释的编程语言概念，其中行为的解释将通过因果干预产生。其次，它将通过开发一组可解释性技术，为给定的SE任务（例如，程序修复）生成行为的、基于特征的和文本的解释，从而为代码的上下文化语言模型的局部解释开发自动化技术。最后，该项目将创建技术，使研究人员和开发人员能够根据生成的解释向模型提供反馈。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。

Advances in artificial intelligence (AI) have led to the development of several new types of tools for software developers that aim to help automate various parts of the software development process of building and maintaining software. However, the combination of complex underlying deep-learning models and massive training datasets makes it difficult to interpret why these models, and the developer tools powered by them, behave the way they do. Given the increasingly important role that these tools are beginning to play in software engineering (SE), it is imperative that techniques be developed that allow stakeholders to better understand and work with these tools such that critical software infrastructure can be maintained. This project will develop a framework and methodology that enables both researchers who build AI-powered developer tools, and software engineers who use these tools, to interpret why the underlying models make the predictions they do. The objective is to allow researchers to obtain detailed insights into why a model may not be performing as expected, allowing for targeted improvement and informed creation of new models. The methodology will be integrated into AI-powered software development tools, allowing software engineers to make informed decisions about when a tool’s suggestion may be helpful or harmful, thus building trust in their use. The interpretability framework will also enable new forms of interaction with these tools, providing a mechanism for natural language feedback that improves over time. This project will produce and disseminate educational materials on best practices related to building and using AI-powered programming tools. These materials are intended to be integrated into existing computer-literacy courses at all levels of education. In addition, the project will focus on recruiting and retaining computer science students from traditionally underrepresented categories.This project has three specific goals. First, it will design an automated approach for generating global explanations of the behavior of “context-free” neural language models for source code. This component of the project will map predictions from large language models to human-interpretable programming language concepts using causal inference theory, wherein explanations of behavior will be generated via causal interventions. Second, it will develop automated techniques for local explanations of contextualized language models of code by developing a set of interpretability techniques that generate behavioral, feature-based, and textual explanations defined for given SE tasks (e.g., program repair). Finally, the project will create techniques that enable researchers and developers to provide feedback to models based on generated explanations.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.