ECQA
收藏资源简介:
ECQA数据集(Explanation CommonsenseQA的缩写)源自CommonsenseQA。经过人工标注后,它包含了从CommonsenseQA中提取的11K问答对的正向和负向属性,以及英文解释。其目标是为CommonsenseQA中的基于知识的问答任务提供解释,深入理解与各种选项相关的通用属性。
The ECQA dataset (an abbreviation for Explanation CommonsenseQA) originates from CommonsenseQA. After being manually annotated, it encompasses 11K question-answer pairs with their positive and negative attributes extracted from CommonsenseQA, along with English explanations. Its aim is to provide explanations for knowledge-based question-answering tasks within CommonsenseQA, delving into the general attributes related to various options.
ECQA-Dataset 数据集概述
数据集基本信息
- 名称: Explanations for CommonsenseQA (ECQA)
- 来源论文: Explanations for CommonsenseQA: New Dataset and Models (ACL-IJCNLP 2021)
- 论文地址: https://aclanthology.org/2021.acl-long.238/
- 原始数据来源: CommonsenseQA (arXiv:1811.00937 [cs.CL])
- 许可证: Community Data License Agreements Sharing license 1.0
数据集内容
-
核心数据文件:
ecqa.jsonl- 格式: 每行一个JSON对象
- 字段说明:
id: 字符串类型,对应CQA数据的样本IDpositives: 字符串列表,正确选项的正面属性negatives: 字符串列表,错误选项的负面属性explanation: 字符串类型,自由形式的解释文本
-
数据划分文件 (位于
author_split目录):train_ids.txt: 训练集ID列表val_ids.txt: 验证集ID列表test_ids.txt: 测试集ID列表
数据整合说明
-
依赖数据: CommonsenseQA官方数据文件
- 必须下载文件:
train_rand_split.jsonl和dev_rand_split.jsonl - 下载地址: https://www.tau-nlp.org/commonsenseqa
- 必须下载文件:
-
生成完整数据的步骤:
- 创建
cqa目录并放入CommonsenseQA数据文件 - 安装依赖库:
json和pandas - 运行
generate_data.py脚本
- 创建
-
生成文件:
cqa_data.csv: 完整数据cqa_data_train.csv: 训练集cqa_data_val.csv: 验证集cqa_data_test.csv: 测试集
生成文件字段说明
- 无标题列: 索引(从0开始)
- q_no: CQA数据ID
- q_concept: 问题概念
- q_text: 问题主干文本
- q_op1-q_op5: 选项A-E
- q_ans: 正确答案
- taskA_pos: 正面属性列表(换行分隔)
- taskA_neg: 负面属性列表(换行分隔)
- taskB: 自由形式的解释文本
引用信息
bibtex @inproceedings{aggarwaletal2021ecqa, title={{E}xplanations for {C}ommonsense{QA}: {N}ew {D}ataset and {M}odels}, author={Shourya Aggarwal and Divyanshu Mandowara and Vishwajeet Agrawal and Dinesh Khandelwal and Parag Singla and Dinesh Garg}, booktitle="Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)}", year = "2021", address = "Online", publisher = "Association for Computational Linguistics" }




