A2Dre+ (Extension of A2D sentences where trivial cases where filtered)

A2Dre 是 A2D 测试集中的一个子集，包括 $433$~\textit{non-trivial} REs。由于其在 $7$~ 语义类别中的高度不平衡分布，我们选择 $4$~ 主要类别 \textsl{外观、位置、运动和静态}。这四个类别的共同点是，在大多数情况下，对于给定的所指对象，可以提供一个表示某个类别的 RE，而一个不提供的 RE。我们使用这些类别通过额外的 RE 来增强 A2Dre，这会根据每个类别的存在或不存在而有所不同。具体来说，基于我们对原始 RE 的分类，对于每个 RE~$re$ 和 category~$C$，我们通过稍微修改 $re$ 来生成一个额外的 RE~$re'$，使其可以（或不可以）快递~$C$。例如，对于Figure~\ref{fig:a2d-images}中的最后一个RE，即\emph{站在女人旁边的黄色连衣裙的女孩}，可以分类为\textit{外观}，\textit{位置} ，没有 \textit{motion} 和 \textit{static}，我们为每个类别生成新的 RE：\emph{站在女人旁边的女孩}（没有 \textit{外观}），\emph{穿着黄色连衣裙的女孩}（没有 \textit{location})、\emph{穿黄色裙子走路的女孩} (\textit{motion}) 和 \emph{穿着黄色裙子的女孩靠近女人} (没有 \textit{static})。我们不对 \textsl{category} 应用这个过程，因为它在几乎所有的 RE 中都有表达，并且在许多情况下它的去除可能很困难。我们将此扩展数据集命名为 A2Dre+。

A2Dre is a subset of the A2D test set, comprising $433$ non-trivial REs. Due to its highly imbalanced distribution across $7$ semantic categories, we select four primary categories: Appearance, Location, Motion, and Static. A shared characteristic of these four categories is that, for most given referent objects, there exist both a RE that conveys a specified category and a RE that fails to do so. We use these categories to augment A2Dre with additional REs, which vary based on whether each category is included or excluded. Specifically, based on our classification of the original REs, for each RE $re$ and category $C$, we generate an additional RE $re'$ via minor modifications to $re$, such that it either expresses $C$ or does not. For instance, the final RE in Figure~
ef{fig:a2d-images}, *the girl in the yellow dress standing next to the woman*, can be categorized as expressing Appearance and Location, but not Motion and Static. We generate new REs for each category as follows: *the girl standing next to the woman* (omitting Appearance), *the girl wearing a yellow dress* (omitting Location), *the girl walking in a yellow dress* (expressing Motion), and *the girl in the yellow dress near the woman* (omitting Static). We do not apply this process to the category, as it is expressed in nearly all REs, and removing it can be difficult in many cases. We name this augmented dataset A2Dre+.

搜集汇总

背景与挑战

背景概述

A2Dre+是A2D测试集的一个扩展子集，通过过滤简单案例并基于外观、位置、运动和静态四个主要语义类别，对原始指称表达进行修改以生成额外表达，从而增强数据集的多样性和平衡性。