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

TL;DR

This paper presents a systematic, linguistically informed approach to creating adversarial questions for NLP models, demonstrating its effectiveness through the winning team’s high error rate in a challenge.

Contribution

It introduces a novel, systematic method for generating adversarial questions in NLP, emphasizing linguistic insights to challenge models effectively.

Findings

Achieved a 62% error rate on the QA model

First place in DADC 2022 adversarial challenge

Pilot experiments support the approach's effectiveness

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%. 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.