eQASC

该数据集包含 QASC 数据集中问题的 98k 2-hop 解释，注释指示它们是有效（~25k）还是无效（~73k）解释。该存储库通过提供三个新数据集解决了当前缺乏用于区分有效多跳解释和无效的训练数据的问题。主要的 eQASC 包含多跳问答数据集 QASC 的 98k 解释注释，并且是第一个为每个答案注释多个候选解释的方法。第二个数据集 eQASC-perturbed 由 QASC 中解释子集的众包扰动（同时保持其有效性）构建，以测试解释预测模型的一致性和泛化性。第三个数据集 eOBQA 是通过向 OBQA 数据集添加解释注释来构建的，以测试在 eQASC 上训练的模型的泛化性。

This dataset contains 98k 2-hop explanations for questions from the QASC dataset, with annotations indicating whether each explanation is valid (~25k) or invalid (~73k). This repository addresses the current shortage of training data for differentiating valid multi-hop explanations from invalid ones by releasing three novel datasets. The flagship eQASC dataset includes 98k annotated explanations for the QASC multi-hop question answering dataset, and it is the first resource to annotate multiple candidate explanations for each answer. The second dataset, eQASC-perturbed, is constructed via crowdsourced perturbations of a subset of explanations from QASC while maintaining their validity, to evaluate the consistency and generalization of explanation prediction models. The third dataset, eOBQA, is developed by adding explanation annotations to the OBQA dataset, designed to test the generalization of models trained on eQASC.

搜集汇总

背景与挑战

背景概述

eQASC是一个包含98k个2-hop解释的文本推理数据集，用于区分有效和无效的多跳解释，由卡内基梅隆大学和艾伦人工智能研究所在2020年发布。该数据集填补了多跳解释训练数据的空白，并支持解释预测模型的开发和测试。