FQuAD(French Question Answering Dataset)

语言建模领域的最新进展改进了许多自然语言处理任务的最新成果。其中，机器阅读理解任务取得了显着进展。然而，由于其他语言（如法语）可用的标记资源仍然稀缺，因此大多数结果基本上都是用英语报告的。在目前的工作中，我们介绍了法语问答数据集（FQuAD）。 FQuAD 是法语母语阅读理解数据集，由高等教育学生在一组 Wikipedia 文章中创建的 25,000 多个问题组成。类似于 SQuAD 的数据集分析用于评估带注释的问题和答案的性质。为了评估数据集的质量，我们训练了一个基线模型，该模型在测试集上的 F1 得分为 88.0%，精确匹配率为 77.9%。除此之外，还探索了基于问题类型和训练样本数量影响的性能分析。

Recent advances in language modeling have improved state-of-the-art results across numerous natural language processing (NLP) tasks. Notably, the machine reading comprehension (MRC) task has achieved remarkable progress. However, most results are predominantly reported in English, as annotated resources for other languages such as French remain scarce. In this work, we introduce the French Question Answering Dataset (FQuAD). FQuAD is a native French reading comprehension dataset consisting of over 25,000 questions created by post-secondary students across a collection of Wikipedia articles. Dataset analysis analogous to that of the SQuAD benchmark is conducted to characterize the properties of annotated questions and answers. To evaluate the dataset quality, we train a baseline model that achieves an F1 score of 88.0% and an exact match (EM) accuracy of 77.9% on the test set. Additionally, performance analyses based on question types and the impact of training sample sizes are explored.

搜集汇总

背景与挑战

背景概述

FQuAD是一个由高等教育学生基于维基百科文章创建的法语问答数据集，包含超过25,000个问题，用于机器阅读理解任务。该数据集为法语自然语言处理提供了稀缺的标记资源，并通过基线模型在测试集上取得了88.0%的F1分数和77.9%的精确匹配率。