CoDraw

该数据集名为CoDraw，包含了大约1万个对话，这些对话中人类玩家在协作图像绘制游戏中交流了约13.8万条信息。数据集的特色在于，其中一位玩家作为描述者（Teller）描绘一个场景，而另一位玩家作为绘制者（Drawer）根据这些描述尝试重现该场景。这些对话是基于抽象场景数据集收集而来的。该数据集的规模大约为1万个对话，其任务是协作绘图与沟通。

This dataset is named CoDraw, which contains approximately 10,000 dialogues. In these dialogues, human players exchanged around 138,000 messages while participating in a collaborative image drawing game. A distinctive feature of CoDraw is that one player acts as the Teller, who is responsible for describing a target scene, while the other player serves as the Drawer, who attempts to recreate the scene based on the descriptions provided by the Teller. These dialogues were collected based on an abstract scene dataset. The core task of this dataset focuses on collaborative drawing and communication.

搜集汇总

背景与挑战

背景概述

CoDraw是一个用于研究基于目标的协作沟通的数据集，通过一个协作绘图游戏收集，涉及两名玩家（讲述者和绘图者）在部分可观察的虚拟环境中使用自然语言沟通以重建抽象场景。数据集包含约9993个对话实例，并具有独特的“窥视机会”机制，以增强视觉信息交流，适用于语言理解和视觉推理任务的研究。