A-OKVQA
收藏资源简介:
A-OKVQA是一个由艾伦人工智能研究所创建的大规模视觉问答数据集,包含约25,000个问题,这些问题需要广泛的世界知识和常识来回答。与现有的知识基础VQA数据集不同,A-OKVQA中的问题不能仅通过查询知识库来回答,而是需要对图像中描绘的场景进行某种形式的常识推理。数据集不仅扩展了代理所需知识的多样性,还增加了回答问题所需的推理系统的复杂性。A-OKVQA旨在推动视觉和语言模型的推理能力,通过提供多样化和挑战性的问题,促进模型在理解和生成答案时结合视觉信息和外部知识的能力。
A-OKVQA is a large-scale visual question answering (VQA) dataset created by the Allen Institute for AI, containing approximately 25,000 questions that require extensive world knowledge and common sense to answer. Unlike existing knowledge-based VQA datasets, the questions in A-OKVQA cannot be answered solely by querying knowledge bases; instead, they require some form of common-sense reasoning over the scenes depicted in the images. This dataset not only expands the diversity of knowledge required for AI agents, but also increases the complexity of the reasoning systems needed to answer these questions. A-OKVQA aims to advance the reasoning capabilities of vision-and-language models, and by providing diverse and challenging questions, it promotes the ability of models to combine visual information and external knowledge when understanding and generating answers.

- 1A-OKVQA: A Benchmark for Visual Question Answering using World Knowledge艾伦人工智能研究所 · 2022年



