PKG-VQA
收藏资源简介:
PKG-VQA数据集是由新加坡高性能计算研究所和新加坡前沿人工智能研究中心创建的,包含多个选择问题,每个问题都包含一个视频片段、一个问题以及五个选项,其中一个为正确答案。该数据集旨在评估模型利用程序性知识进行回答的能力,涵盖了17种程序性知识问题类型,包括单跳和多跳问题,以及需要推理方法如演绎、概率、情境、因果和反事实推理。数据集的构建利用了Procedural Knowledge Graph (PKG),其中包含了关于步骤、任务和它们之间关系的实例。
The PKG-VQA dataset was developed by the Institute of High Performance Computing Singapore and the Singapore Centre for Frontier AI Research. It comprises multiple-choice questions, each containing a video clip, a question, and five options with one correct answer. This dataset is designed to evaluate a model's capability to answer questions by leveraging procedural knowledge, covering 17 types of procedural knowledge questions, including single-hop and multi-hop questions, as well as questions requiring reasoning approaches such as deductive, probabilistic, situational, causal, and counterfactual reasoning. The dataset is constructed using the Procedural Knowledge Graph (PKG), which includes instances of steps, tasks, and the relationships between them.




