longhorns at DADC 2022: How many linguists does it take to fool a Question Answering model? A systematic approach to adversarial attacks.

Abstract

Developing methods to adversarially challenge NLP systems is a promising avenue for improving both model performance and interpretability. Here, we describe the approach of the team "longhorns" on Task 1 of the The First Workshop on Dynamic Adversarial Data Collection (DADC), which asked teams to manually fool a model on an Extractive Question Answering task. Our team finished first, with a model error rate of 62%. 1 We advocate for a systematic, linguistically informed approach to formulating adversarial questions, and we describe the results of our pilot experiments, as well as our official submission.

The State of Human-centered NLP Technology for Fact-checking

“…When we consider such human-centered use of fact-checking technologies, providing an automated veracity prediction without justifying the answer can cause a system to be ignored or distrusted, or even reinforce mistaken human beliefs in false claims (the "backfire effect" (Lewandowsky, Ecker, Seifert, Schwarz and Cook, 2012)). Explanations and justifications are especially important given the noticeable drop in performance of state-of-the-art NLP systems when facing adversarial examples (Kovatchev, Chatterjee, Govindarajan, Chen, Choi, Chronis, Das, Erk, Lease, Li et al, 2022). Consequently, automated fact-checking systems intended for humanconsumption should seek to explain their veracity predictions in a similar manner to that of existing journalistic fact-checking practices (Uscinski, 2015).…”