STRATEGYQA
收藏资源简介:
STRATEGYQA是一个专注于隐式多跳推理的布尔QA基准,由特拉维夫大学和艾伦人工智能研究所创建。该数据集包含2,780个问题,每个问题都需要通过隐含的推理步骤来回答,这些问题涵盖了广泛的策略和知识领域。数据集的创建过程结合了基于术语的启发式方法来激发注释者,严格控制注释者群体,并通过对抗性过滤消除推理捷径。此外,每个问题都附有分解成推理步骤的描述和包含每个步骤答案的维基百科段落。STRATEGYQA的应用领域包括评估和开发能够理解和执行复杂推理的AI系统,旨在解决自然语言处理中的推理挑战。
STRATEGYQA is a boolean question answering (QA) benchmark focused on implicit multi-hop reasoning, developed jointly by Tel Aviv University and the Allen Institute for Artificial Intelligence. This dataset comprises 2,780 questions, each of which requires answering via implicit reasoning steps and covers a broad spectrum of reasoning strategies and knowledge domains. The construction of the dataset leverages terminology-based heuristic approaches to guide annotators, imposes strict controls over the annotator pool, and eliminates reasoning shortcuts through adversarial filtering. Additionally, every question is accompanied by a detailed breakdown of the problem into individual reasoning steps, alongside Wikipedia paragraphs that contain the answers for each respective reasoning step. The application areas of STRATEGYQA include evaluating and developing AI systems capable of understanding and executing complex reasoning, with the goal of addressing core reasoning challenges in natural language processing.




