DROP: A Reading Comprehension Benchmark Requiring Discrete Reasoning\n Over Paragraphs

Dheeru Dua, Yizhong Wang, Pradeep Dasigi, Gabriel Stanovsky, Sameer Singh, Matt Gardner

arXiv (Cornell University) · 2019 · 179 citations · 0 references

DOIFull text

Open access

Concepts

Abstract

Reading comprehension has recently seen rapid progress, with systems matching\nhumans on the most popular datasets for the task. However, a large body of work\nhas highlighted the brittleness of these systems, showing that there is much\nwork left to be done. We introduce a new English reading comprehension\nbenchmark, DROP, which requires Discrete Reasoning Over the content of\nParagraphs. In this crowdsourced, adversarially-created, 96k-question\nbenchmark, a system must resolve references in a question, perhaps to multiple\ninput positions, and perform discrete operations over them (such as addition,\ncounting, or sorting). These operations require a much more comprehensive\nunderstanding of the content of paragraphs than what was necessary for prior\ndatasets. We apply state-of-the-art methods from both the reading comprehension\nand semantic parsing literature on this dataset and show that the best systems\nonly achieve 32.7% F1 on our generalized accuracy metric, while expert human\nperformance is 96.0%. We additionally present a new model that combines reading\ncomprehension methods with simple numerical reasoning to achieve 47.0% F1.\n