GEM/ART
收藏资源简介:
GEM/ART数据集是一个用于研究基于语言的溯因推理(Abductive Reasoning)的数据集。溯因推理是指推断出最合理的解释。例如,如果Jenny下班回家发现家里一团糟,并且记得她离开时窗户是开着的,她可以假设有小偷闯入并造成了混乱,作为最合理的解释。该数据集包含两个观察事件和一个解释标签,标签由亚马逊Mechanical Turk平台的众包工作者生成。数据集的结构包括训练集、开发集和测试集,主要用于训练和评估模型生成合理的假设来解释给定的观察事件。数据集的创建者来自Allen Institute for AI,并且数据集被用于GEM基准测试中。
The GEM/ART dataset is a dedicated resource for research on language-based abductive reasoning. Abductive reasoning refers to the process of inferring the most plausible explanation for observed scenarios. For instance, if Jenny returns home from work to find her house in a state of disarray and recalls leaving her window open, she may hypothesize that a burglar broke in and caused the chaos as the most reasonable explanation. Each sample in this dataset comprises two observed events and an accompanying explanation label, with the labels generated by crowdworkers on the Amazon Mechanical Turk platform. The dataset is structured into training, validation, and test splits, and is primarily used to train and evaluate models that generate plausible hypotheses to explain the given observed events. It was developed by researchers from the Allen Institute for AI, and has been incorporated into the GEM benchmark suite.
数据集概述
数据集名称
- 名称: ART
- 全称: Abductive Reasoning Task
数据集描述
- 任务类型: 推理
- 语言: 英语
- 许可证: Apache License 2.0
- 创建方式: 自动创建
- 数据来源: 原始数据
- 主要用途: 研究语言基础的推理能力
数据集结构
- 数据字段:
observation_1: 描述一个观察/事件的字符串observation_2: 描述一个观察/事件的字符串label: 解释observation_1和observation_2可能发生原因的字符串
- 数据分割:
train: 训练实例dev: 开发实例test: 测试实例
数据集创建者
- 创建者: Chandra Bhagavatula (AI2), Ronan Le Bras (AI2), Chaitanya Malaviya (AI2), Keisuke Sakaguchi (AI2), Ari Holtzman (AI2, UW), Hannah Rashkin (AI2, UW), Doug Downey (AI2), Wen-tau Yih (AI2), Yejin Choi (AI2, UW)
- 组织: Allen Institute for AI
数据集使用
- 主要任务: 推理
- 目标使用: 训练和评估模型以生成解释两个给定观察的合理假设
数据集获取
- 网页: Abductive Commonsense Reasoning
- 下载链接: Google Storage
- 论文: OpenReview
- 联系人: Chandra Bhagavatulla (chandrab@allenai.org)
数据集评估
- 评估指标: BLEU, BERT-Score, ROUGE
- 评估结果: 无
数据集维护
- 维护计划: 无




