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

Trick Me If You Can: Human-in-the-Loop Generation of Adversarial Examples for Question Answering Eric Wallace EE and UMIACS University of Maryland ewallac2@umiacs.umd.edu Pedro Rodriguez, Shi Feng CS, UMIACS University of Maryland {pedro, shifeng} @cs.umd.edu Ikuya Yamada Studio Ousia ikuya@ousia.jp Jordan Boyd-Graber CS, iSchool, UMIACS, LSC University of Maryland jbg@umiacs.umd.edu Abstract Adversarial evaluation stress-tests a model’s understanding of natural language. Because 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 inter- face. We apply this generation framework to a question answering task called Quizbowl, where trivia enthusiasts craft adversarial ques- tions. 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. 1 Introduction Proponents of machine learning claim human par- ity on tasks like reading comprehension (Yu et al., 2018) and commonsense inference (Devlin et al., 2018). Despite these successes, many evaluations neglect that computers solve natural language processing (NLP) tasks in a fundamentally dif- ferent way than humans. Models can succeed without developing ‘‘true’’ language understanding, instead learning superfi- cial patterns from crawled (Chen et al., 2016) or manually annotated data sets (Gururangan et al., 2018; Kaushik and Lipton, 2018). Thus, recent work stress-tests models via adversarial evaluation: elucidating a system’s capabilities by exploiting its weaknesses (Jia and Liang, 2017; Belinkov and Glass, 2019). Unfortunately, whereas adversarial evaluation reveals simplistic model fail- ures (Ribeiro et al., 2018; Mudrakarta et al., 2018), exploring more complex failure patterns requires humaninvolvement(Figure 1): Automatically mod- ifying natural language examples without inval- idating them is difficult. Hence, the diversity of adversarial examples is often severely restricted. Instead, our human–computer hybrid approach uses human creativity to generate adversarial examples. A user interface presents model inter- pretations and helps users craft model-breaking examples (Section 3). We apply this to a question answering (QA) task called Quizbowl, where trivia enthusiasts—who write questions for academic competitions—create diverse examples that stump existing QA models. The adversarially authored test set is none- theless as easy as regular questions for humans (Section 4), but the relative accuracy of strong QA models drops as much as 40% (Section 5). We also host live human vs. computer matches—where models typically defeat top human teams—but observe spectacular model failures on adversarial questions. Analyzing the adversarial edits uncovers phe- nomena that humans can solve but computers cannot (Section 6), validating that our frame- work uncovers creative, targeted adversarial edits (Section 7). Our resulting adversarial data set pres- ents a fun, challenging, and diverse resource for future QA research: A system that masters it will demonstrate more robust language understanding. 387 Transactions of the Association for Computational Linguistics, vol. 7, pp. 387–401, 2019. https://doi.org/10.1162/tacl a 00279. c⃝ 2019 Association for …

Content truncated to 500 words