Trick Me If You Can: Human-in-the-loop Generation of Adversarial Examples for Question Answering

TL;DR

This paper introduces a human-in-the-loop method for generating challenging adversarial questions in QA, revealing diverse weaknesses in models through interactive human guidance and validation.

Contribution

It presents a novel interactive framework for human-guided adversarial question generation, improving the diversity and complexity of adversarial examples for QA models.

Findings

Adversarial questions successfully stump neural and retrieval models.

Questions cover a wide range of reasoning phenomena.

The approach exposes significant robustness challenges in QA systems.

Abstract

Adversarial evaluation stress tests a model's understanding of natural language. While past approaches expose superficial patterns, the resulting adversarial examples are limited in complexity and diversity. We propose human-in-the-loop adversarial generation, where human authors are guided to break models. We aid the authors with interpretations of model predictions through an interactive user interface. We apply this generation framework to a question answering task called Quizbowl, where trivia enthusiasts craft adversarial questions. The resulting questions are validated via live human--computer matches: although the questions appear ordinary to humans, they systematically stump neural and information retrieval models. The adversarial questions cover diverse phenomena from multi-hop reasoning to entity type distractors, exposing open challenges in robust question answering.