Spoken-SQuAD

在 SpokenSQuAD 中，文档是口语形式，输入问题是文本形式，每个问题的答案始终是文档中的一个跨度。以下程序用于从原始 SQuAD 数据集生成语音文档。首先，使用 Google 文字转语音系统在 SQuAD 中生成文章的语音版本。然后 CMU Sphinx 被起诉以生成相应的 ASR 转录。 SQuAD 训练集用于生成 Spoken SQuAD 的训练集，SQuAD 开发集用于生成 Spoken SQuAD 的测试集。如果相关文章的 ASR 转录中不存在问题的答案，则从数据集中删除问答对，因为这些示例对于现阶段的听力理解机器来说太难了。

In SpokenSQuAD, the documents are in spoken form, the input questions are in text format, and the answer to each question is always a span extracted from the corresponding document. The following procedure is used to generate spoken documents from the original SQuAD dataset. First, the Google Text-to-Speech system is employed to generate spoken versions of the articles in SQuAD. Next, CMU Sphinx is utilized to produce the corresponding ASR transcriptions. The SQuAD training split is used to construct the training set for SpokenSQuAD, while the SQuAD development split serves as the source for the SpokenSQuAD test set. If the answer to a given question does not exist in the ASR transcription of the associated article, the question-answer pair will be removed from the dataset, as these examples are overly challenging for current speech comprehension machine models.

搜集汇总

背景与挑战

背景概述

Spoken-SQuAD是一个用于音频文本阅读理解的数据集，通过将SQuAD数据集中的文章转换为语音文档并生成ASR转录来构建。它基于原始SQuAD的训练和开发集划分，并移除了ASR转录中不存在答案的问答对，以降低听力理解任务的难度。