Interpreting the Robustness of Neural NLP Models to Textual Perturbations

TL;DR

This paper investigates why NLP models vary in robustness to textual perturbations, proposing that learnability of perturbations explains this variability, supported by extensive experiments on multiple models and perturbation types.

Contribution

It introduces the learnability metric to explain model robustness and provides a causal justification, validated through comprehensive experiments.

Findings

Higher learnability correlates with lower robustness.

Models better at identifying perturbations perform worse at ignoring them.

Empirical evidence supports the inverse relationship between learnability and robustness.

Abstract

Modern Natural Language Processing (NLP) models are known to be sensitive to input perturbations and their performance can decrease when applied to real-world, noisy data. However, it is still unclear why models are less robust to some perturbations than others. In this work, we test the hypothesis that the extent to which a model is affected by an unseen textual perturbation (robustness) can be explained by the learnability of the perturbation (defined as how well the model learns to identify the perturbation with a small amount of evidence). We further give a causal justification for the learnability metric. We conduct extensive experiments with four prominent NLP models -- TextRNN, BERT, RoBERTa and XLNet -- over eight types of textual perturbations on three datasets. We show that a model which is better at identifying a perturbation (higher learnability) becomes worse at ignoring such a perturbation at test time (lower robustness), providing empirical support for our hypothesis.