UIT-ViWikiQA

越南语处理的总体发展，尤其是机器阅读理解，引起了研究界的极大关注。近年来，有一些大型越南语的机器阅读理解任务的数据集，例如UIT-vipad和UIT-ViNewsQA。但是，数据集的答案并不多样化，无法为研究服务。在本文中，我们介绍了UIT-ViWikiQA，这是第一个用于评估越南语中基于句子提取的机器阅读理解的数据集。UIT-ViWikiQA数据集是从UIT-vipquad数据集转换而来的，包括基于维基百科174越南文章的5.109段落的23.074个问题答案。我们提出了一种转换算法来创建基于句子提取的机器阅读理解数据集，以及越南语基于句子提取的机器阅读理解的三种方法。我们的实验表明，最好的机器模型是xlm-r$ _ Large，它在我们的数据集上实现了85.97% 的精确匹配 (EM) 得分和88.77% 的F1-score。此外，我们根据越南语的问题类型和上下文对MRC模型性能的影响来分析实验结果，从而显示了我们向自然语言处理社区提出的UIT-ViWikiQA数据集的挑战。

The overall development of Vietnamese language processing, especially machine reading comprehension (MRC), has attracted considerable attention from the research community. In recent years, several large-scale Vietnamese MRC datasets have been proposed, such as UIT-viPAD and UIT-ViNewsQA. However, these datasets lack diverse answer types and fail to adequately support cutting-edge research. In this paper, we introduce UIT-ViWikiQA, the first dataset for evaluating sentence extraction-based machine reading comprehension in Vietnamese. The UIT-ViWikiQA dataset is converted from the UIT-viPquad dataset, consisting of 23,074 question-answer pairs across 5,109 paragraphs sourced from 174 Vietnamese Wikipedia articles. We propose a conversion algorithm for constructing sentence extraction-based MRC datasets, alongside three approaches for Vietnamese sentence extraction-based MRC tasks. Our experimental results show that the best-performing model is XLM-R_Large, which achieves an Exact Match (EM) score of 85.97% and an F1-score of 88.77% on our dataset. Furthermore, we analyze the experimental results by examining the impact of Vietnamese question types and contexts on MRC model performance, thereby highlighting the challenges posed by the UIT-ViWikiQA dataset for the natural language processing (NLP) community.

搜集汇总

背景与挑战

背景概述

UIT-ViWikiQA是首个用于评估越南语基于句子提取的机器阅读理解的数据集，由174篇维基百科文章转换而来，包含23,074个问题-答案对。该数据集旨在提升答案多样性，并提供了转换算法和三种方法，实验显示最佳模型xlm-r_Large的EM和F1分数分别达到85.97%和88.77%。