On the Efficacy of Adversarial Data Collection for Question Answering: Results from a Large-Scale Randomized Study

In adversarial data collection (ADC), a human workforce interacts with a\nmodel in real time, attempting to produce examples that elicit incorrect\npredictions. Researchers hope that models trained on these more challenging\ndatasets will rely less on superficial patterns, and thus be less brittle.\nHowever, despite ADC's intuitive appeal, it remains unclear when training on\nadversarial datasets produces more robust models. In this paper, we conduct a\nlarge-scale controlled study focused on question answering, assigning workers\nat random to compose questions either (i) adversarially (with a model in the\nloop); or (ii) in the standard fashion (without a model). Across a variety of\nmodels and datasets, we find that models trained on adversarial data usually\nperform better on other adversarial datasets but worse on a diverse collection\nof out-of-domain evaluation sets. Finally, we provide a qualitative analysis of\nadversarial (vs standard) data, identifying key differences and offering\nguidance for future research.\n