Sentence Extraction-Based Machine Reading Comprehension for Vietnamese

The development of natural language processing (NLP) in general and machine reading comprehension particular has attracted the great attention research community. In recent years, there are a few datasets for tasks Vietnamese with large sizes, such as UIT-ViQuAD UIT-ViNewsQA. However, not diverse answers to serve research. this paper, we introduce UIT-ViWikiQA, first dataset evaluating sentence extraction-based language. UIT-ViWikiQA is converted from dataset, consisting comprises 23.074 question-answers based on 5.109 passages 174 Wikipedia articles. We propose conversion algorithm create three types approaches Vietnamese. Our experiments show that best model XLM-R $$_{Large}$$ , which achieves an exact match (EM) 85.97% F1-score 88.77% our dataset. Besides, analyze experimental results terms question type effect context performance MRC models, thereby showing challenges