ACQUIRED

ACQUIRED数据集由加州大学洛杉矶分校创建，专注于真实世界视频中的反事实问题回答。数据集包含3700个视频，覆盖多种事件类型，并从第一人称和第三人称视角确保真实世界的多样性。每个视频都标注了涉及物理、社交和时间三个推理维度的问答对，旨在全面评估模型在多方面的反事实推理能力。数据集的应用领域包括增强人工智能对因果关系的理解及评估干预或条件变化的影响，为未来研究提供了一个全面且可靠的基准。

The ACQUIRED dataset was developed by the University of California, Los Angeles (UCLA), focusing on counterfactual question answering in real-world videos. It contains 3,700 videos covering a wide range of event types, and ensures real-world diversity via both first-person and third-person perspectives. Each video is annotated with question-answer pairs across three core reasoning dimensions: physical, social, and temporal, aiming to comprehensively evaluate models' multi-faceted counterfactual reasoning abilities. The dataset's application areas include advancing artificial intelligence's understanding of causal relationships and assessing the impacts of interventions or conditional changes, serving as a comprehensive and reliable benchmark for future research.

搜集汇总

背景与挑战

背景概述

ACQUIRED是一个包含3.9K标注视频的数据集，专注于回答现实生活视频中的反事实问题。数据集的问题标注涵盖物理、社会和时间三个推理维度，旨在全面评估模型的反事实推理能力。