LIQUID: A Framework for List Question Answering Dataset Generation

Question answering (QA) models often rely on large-scale training datasets, which necessitates the development of a data generation framework to reduce cost manual annotations. Although several recent studies have aimed generate synthetic questions with single-span answers, no study has been conducted creation list multiple, non-contiguous spans as answers. To address this gap, we propose LIQUID, an automated for generating QA datasets from unlabeled corpora. We first convert passage Wikipedia or PubMed into summary and extract named entities summarized text candidate This allows us select answers that are semantically correlated in context is, therefore, suitable constructing questions. then create using off-the-shelf question generator extracted original passage. Finally, iterative filtering answer expansion performed ensure accuracy completeness Using our data, significantly improve performance previous best by exact-match F1 scores 5.0 MultiSpanQA, 1.9 Quoref, 2.8 averaged across three BioASQ benchmarks.