VATEX Adverbs

VATEX副词是最大的，有34个副词出现在135个动作中，形成1,550个独特的动作副词对。动作，副词及其组成的分布是长尾的。与仅包含6个的现有HowTo100M副词相比，每个数据集考虑的dverb更多。我们用200的视频样本来衡量每个数据集的注释的质量。由于新数据集来自人类的书面标题，因此一个人明确选择了副词来描述动作，因此注释的噪声要比HowTo100M副词少得多。The VATEX adverb dataset is the largest, comprising 34 adverbs linked to 135 distinct actions and forming 1,550 unique action-adverb pairs. The distributions of actions, adverbs, and their composite pairs follow a long-tail distribution. Compared with the existing HowTo100M adverb dataset which only includes 6 adverbs, the VATEX dataset covers a notably larger number of adverbs. We assessed the annotation quality of both datasets using a sample of 200 videos. Since the VATEX dataset is derived from human-written video titles, where annotators explicitly selected adverbs to describe the corresponding actions, its annotation noise is substantially lower than that of the HowTo100M adverb dataset.

搜集汇总

背景与挑战

背景概述

VATEX Adverbs是一个大规模副词数据集，包含34个副词应用于135个动作中，构成了1,550个独特的动作副词组合，其分布呈现长尾特征。与现有数据集相比，它涵盖了更多副词，且由于数据源自人类书面标题，注释质量更高、噪声更少。该数据集由阿姆斯特丹大学于2022年发布。