Interpreting the Robustness of Neural NLP Models to Textual Perturbations
This work addresses the problem of understanding and predicting robustness in NLP models for researchers and practitioners, though it is incremental as it builds on existing knowledge of model sensitivity.
The paper investigates why NLP models vary in robustness to different textual perturbations, finding that a model's ability to learn a perturbation (learnability) inversely correlates with its robustness to that perturbation at test time, with empirical support from experiments on four models and eight perturbation types across three datasets.
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.