BoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questions

TL;DR

This paper introduces BoolQ, a challenging dataset of naturally occurring yes/no questions, and analyzes transfer learning methods, finding entailment-based transfer from MultiNLI most effective, yet leaving room for improvement.

Contribution

The creation of the BoolQ dataset and the analysis of transfer learning strategies for yes/no question answering in natural settings.

Findings

Transfer from entailment data outperforms paraphrase and extractive QA.

Pre-training on MultiNLI followed by fine-tuning yields 80.4% accuracy.

Significant gap remains between model performance and human accuracy.

Abstract

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.