Importance of the Single-Span Task Formulation to Extractive Question-answering

Abstract

Recent progress in machine reading comprehension and question-answering has allowed machines to reach and even surpass human question-answering. However, the majority of these questions have only one answer, and more substantial testing on questions with multiple answers, or multi-span questions, has not yet been applied. Thus, we introduce a newly compiled dataset consisting of questions with multiple answers that originate from previously existing datasets. In addition, we run BERT-based models pre-trained for question-answering on our constructed dataset to evaluate their reading comprehension abilities. Among the three of BERT-based models we ran, RoBERTa exhibits the highest consistent performance, regardless of size. We find that all our models perform similarly on this new, multi-span dataset (21.492% F1) compared to the single-span source datasets (~33.36% F1). While the models tested on the source datasets were slightly fine-tuned, performance is similar enough to judge that task formulation does not drastically affect question-answering abilities. Our evaluations indicate that these models are indeed capable of adjusting to answer questions that require multiple answers. We hope that our findings will assist future development in questionanswering and improve existing question-answering products and methods.