BoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questions
This addresses the problem of improving question-answering systems for complex, non-factoid yes/no questions, with incremental contributions in dataset creation and transfer learning insights.
The paper tackles the challenge of naturally occurring yes/no questions by creating the BoolQ dataset and finds they are unexpectedly difficult, requiring complex inference. The best method using BERT achieves 80.4% accuracy, compared to 90% human accuracy and a 62% baseline.
In this paper we study yes/no questions that are naturally occurring --- meaning that they are generated in unprompted and unconstrained settings. We build a reading comprehension dataset, BoolQ, of such questions, and show that they are unexpectedly challenging. They often query for complex, non-factoid information, and require difficult entailment-like inference to solve. We also explore the effectiveness of a range of transfer learning baselines. We find that transferring from entailment data is more effective than transferring from paraphrase or extractive QA data, and that it, surprisingly, continues to be very beneficial even when starting from massive pre-trained language models such as BERT. Our best method trains BERT on MultiNLI and then re-trains it on our train set. It achieves 80.4% accuracy compared to 90% accuracy of human annotators (and 62% majority-baseline), leaving a significant gap for future work.