ShARC (Shaping Answers with Rules through Conversation)

机器阅读的大部分工作都集中在问答问题上，其中答案直接在要阅读的文本中表达出来。然而，许多现实世界的问答问题需要阅读文本，并不是因为它包含文字答案，而是因为它包含一个配方，可以与读者的背景知识一起得出答案。我们将这项任务形式化并制定众包策略来收集 32k 任务实例。

Most research in machine reading has focused on question answering (QA) tasks where the answer is directly expressed in the target reading text. However, many real-world QA tasks require reading text not because it contains the explicit textual answer, but because it contains a reasoning recipe that can be combined with the reader's background knowledge to derive the final answer. We formalize this task and develop a crowdsourcing strategy to collect 32k task instances.

搜集汇总

背景与挑战

背景概述

ShARC数据集是一个用于机器阅读理解任务的数据集，专注于需要结合文本和背景知识进行推理的问答问题，包含32k任务实例，发布于2018年。