Disfl-QA
收藏资源简介:
Disfl-QA 是针对信息搜索环境中上下文不流畅的目标数据集,即对维基百科段落的问答。 Disfl-QA 基于 SQuAD-v2 数据集,其中开发集中的每个问题都被注释以使用段落作为干扰源来添加上下文不流畅。最终数据集由约 12k(不流利的问题、答案)对组成。超过 90% 的不流畅性是更正或重新开始,这使其成为更难于不流畅性校正的测试集。 Disfl-QA 旨在填补语音和 NLP 研究社区之间的主要空白。我们希望该数据集可以作为基准数据集,用于测试模型对混乱输入的鲁棒性。
Disfl-QA is a targeted dataset for disfluent contexts in information search scenarios, specifically designed for question answering over Wikipedia paragraphs. Built upon the SQuAD-v2 dataset, each question in its development set is annotated with contextual disfluencies by using the source passage as a distractor. The final dataset contains approximately 12k (disfluent question, answer) pairs. Over 90% of the disfluencies take the form of corrections or restarts, making it a more challenging test set for disfluency correction tasks. Disfl-QA aims to fill a key gap between the speech and natural language processing (NLP) research communities. We expect this dataset to act as a benchmark for evaluating model robustness against noisy inputs.




