“契合”
用英语怎么说？

Moving on to question answering, Qasper is very interesting. NLP practitioners wrote questions after reading only the title and abstract, while others annotated the answers. Yes, and it contains abstractive, extractive, and yes/no questions. It's a great way to test out models on real research papers. And NarrativeQA uses books and movie scripts. Annotators were given summaries from Wikipedia to generate question-answer pairs, so they had to rely on their own words instead of copying directly. That plays into our goal of zero-shot long text understanding. But QuALITY is probably the hardest dataset here. Half of its questions are labeled as hard. That's because they conducted a speed validation process. Annotators had to skim through the document quickly, and most chose the wrong answer under time pressure.

接下来是问答，Qasper 非常有趣。NLP 从业者在仅阅读标题和摘要后撰写了问题，而另一组 NLP 从业者则在通读全文后标注了答案。是的，它包含抽象性、抽取性以及是/否问题。这是在真实研究论文上测试模型的一个好方法。此外，NarrativeQA 使用了书籍和电影剧本。标注者获得了来自维基百科的摘要来生成问题-答案对，因此他们不得不依赖自己的语言，而不是直接复制。这契合我们零样本长文本理解的目标。但 QuALITY 可能是这里最难的数据集。它有一半的问题被标注为困难。这是因为他们进行了速度验证过程。标注者不得不快速略读文件，大多数人在时间压力下选择了错误的答案。