HVQR
收藏arXiv2019-09-23 更新2024-08-06 收录
下载链接：
http://arxiv.org/abs/1909.10128v1
下载链接
链接失效反馈官方服务：
资源简介：
HVQR数据集由中山大学智能系统工程学院创建，包含157,201个问题-答案对，旨在评估视觉问答系统的解释性和高阶推理能力。该数据集通过结合图像场景图和常识知识库构建，每个关系三元组在所有问题中仅出现一次，挑战现有模型处理未见问题和知识事实的能力。HVQR数据集适用于推动复杂问题请求的研究，要求模型具备多步骤推理能力，并提供可解释的评估基准，以增强系统的自诊断能力。
The HVQR dataset was created by the School of Intelligent Systems Engineering, Sun Yat-sen University, and contains 157,201 question-answer pairs. It is designed to evaluate the explainability and high-order reasoning capabilities of visual question answering (VQA) systems. Constructed by combining image scene graphs and common sense knowledge bases, each relational triple appears exactly once across all questions, which challenges existing models' ability to handle unseen questions and knowledge facts. The HVQR dataset is suitable for advancing research on complex question queries that require models to possess multi-step reasoning abilities, and provides an explainable evaluation benchmark to enhance the self-diagnosis capabilities of such systems.
提供机构：
中山大学智能系统工程学院
创建时间：
2019-09-23
