A Survey on Out-of-Distribution Evaluation of Neural NLP Models

TL;DR

This work addresses the challenge that neural NLP models often falter under out-of-distribution conditions by unifying adversarial robustness, domain generalization, and dataset biases under the umbrella of distribution shift. It clarifies data-generating processes—natural domain shift ($NDS$), debiased data, and adversarially perturbed data—and evaluation paradigms (data-based and method-based), linking each to covariate-shift concepts and shifted features. The authors offer a framework that connects the three OOD lines, highlights the role of semantic versus background features and biased features, and proposes opportunities such as a comprehensive benchmarking suite and cross-line detection approaches. They also discuss challenges, including covariate-shift assumptions, realism of adversarial attacks, and the need to achieve truly generalizable OOD performance across diverse NLP tasks.

Abstract

Adversarial robustness, domain generalization and dataset biases are three active lines of research contributing to out-of-distribution (OOD) evaluation on neural NLP models. However, a comprehensive, integrated discussion of the three research lines is still lacking in the literature. In this survey, we 1) compare the three lines of research under a unifying definition; 2) summarize the data-generating processes and evaluation protocols for each line of research; and 3) emphasize the challenges and opportunities for future work.