QASnowball: An Iterative Bootstrapping Framework for High-Quality Question-Answering Data Generation

• DOI: 10.48550/arxiv.2309.10326
• Publication Date: 2023-01-01
• Authors: Kunlun Zhu; Shihao Liang; Xu Han; Zhi Zheng; Guoyang Zeng; Zhiyuan Liu; Maosong Sun
• Keywords: Bootstrapping (finance); Data set

Recent years have witnessed the success of question answering (QA), especially its potential to be a foundation paradigm for tackling diverse NLP tasks. However, obtaining sufficient data build an effective and stable QA system still remains open problem. For this problem, we introduce iterative bootstrapping framework augmentation (named QASnowball), which can iteratively generate large-scale high-quality based on seed set supervised examples. Specifically, QASnowball consists three modules, answer extractor extract core phrases in unlabeled documents as candidate answers, generator questions filter out data. Moreover, self-enhanced by reseeding fine-tune itself different iterations, leading continual improvements generation quality. We conduct experiments high-resource English scenario medium-resource Chinese scenario, experimental results show that generated facilitate models: (1) training models achieves comparable using data, (2) pre-training fine-tuning achieve better performance. Our code will released advance further work.