Human Adversarial QA: Did the Model Understand the Paragraph?

Abstract

Recently, adversarial attacks have become an important means of gauging the robustness of natural language models as training and testing set methodology has proved inadequate. In this paper we explore an evaluation based on human-in-the-loop adversarial example generation. These adversarial examples aid us in finding the loopholes in the models and give insights into their working. In the published work on adversarial question-answering, perturbations are made on the questions without changing the background context on which the question is based. In the current work, we examine the complementary idea of perturbing the background context while keeping the question constant. We analyze the state-of-the-art language model BERT for the task of question-answering on SQuAD dataset using novel adversarial examples crafted by humans exposing the weaknesses of the model. We present the typology of the successful attacks here as a baseline for stress-testing QA systems.