Collaborative Research: SHF: Medium: Towards More Human-like AI Models of Source Code

合作研究：SHF：Medium：迈向更人性化的 AI 源代码模型

• 批准号: 2211429
• 负责人: Yu Huang
• 金额: $ 43.19万
• 依托单位: Vanderbilt University
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2022
• 资助国家: 美国
• 起止时间: 2022-06-15 至 2026-05-31
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2211429&HistoricalAwards=false
• 关键词: Collaborative; Research; SHF; Medium; Towards

The research objective of this project is to design novel artificial intelligence-based models of software that learn from and are informed by human behavior. The frontier of many areas of Software Engineering (SE) research involves applications of AI-based models to SE tasks. Many tasks in SE research rely on the same basic underpinning technologies, often a neural representation of source code that is trained to find features in code, which are then used for various tasks e.g., to predict words for a document or areas of code likely to contain a bug. While the first applications of recurrent neural network-based encoder-decoder models were a paradigm shift over the manually-crafted heuristics and rules that the neural models replaced, subsequent changes have yielded less improvement despite increased sophistication.The vision of this project is to achieve a breakthrough in more human-like neural models of source code. Its aim is to advance a broad spectrum of SE research tasks that rely on neural models, by improving the neural models of code that underpin many downstream tasks. The research plan is three-fold: First, the project will characterize human behavior during different SE tasks via eye-tracking and IDE-based experiments. Second, the project will design models that predict or even mimic human behavior. Third, the project will use those models to augment and improve neural representations of source code, and evaluate these new representations in a variety of SE tasks.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.

该项目的研究目标是设计新型的基于人工智能的软件模型，这些模型可以从人类行为中学习并了解人类行为。 软件工程（SE）研究的许多领域的前沿涉及基于AI的模型在SE任务中的应用。SE研究中的许多任务都依赖于相同的基本支撑技术，通常是源代码的神经表示，经过训练可以找到代码中的功能，然后用于各种任务，例如，预测文档中的单词或可能包含错误的代码区域。 虽然基于递归神经网络的编码器-解码器模型的第一个应用是对人工制作的算法和规则的范式转变，但神经模型取代了这些规则，随后的变化尽管增加了复杂性，但改进较少。该项目的愿景是在源代码的更像人类的神经模型方面取得突破。 其目的是通过改进支持许多下游任务的代码神经模型，推进依赖神经模型的广泛SE研究任务。 该研究计划有三个方面：首先，该项目将通过眼动跟踪和基于IDE的实验来描述人类在不同SE任务中的行为。其次，该项目将设计预测甚至模仿人类行为的模型。第三，该项目将使用这些模型来增强和改进源代码的神经表示，并在各种SE任务中评估这些新的表示。该奖项反映了NSF的法定使命，并被认为值得通过使用基金会的智力价值和更广泛的影响审查标准进行评估来支持。