A resampling-based method to evaluate NLI models

The recent progress of deep learning techniques has produced models capable of achieving high scores on traditional Natural Language Inference (NLI) datasets. To understand the generalization limits of these powerful models, an increasing number of adversarial evaluation schemes have appeared. These works use a similar evaluation method: they construct a new NLI test set based on sentences with known logic and semantic properties (the adversarial set), train a model on a benchmark NLI dataset, and evaluate it in the new set. Poor performance on the adversarial set is identified as a model limitation. The problem with this evaluation procedure is that it may only indicate a sampling problem. A machine learning model can perform poorly on a new test set because the text patterns presented in the adversarial set are not well represented in the training sample. To address this problem, we present a new evaluation method, the Invariance under Equivalence test (IE test). The IE test trains a model with sufficient adversarial examples and checks the model's performance on two equivalent datasets. As a case study, we apply the IE test to the state-of-the-art NLI models using synonym substitution as the form of adversarial examples. The experiment shows that, despite their high predictive power, these models usually produce different inference outputs for equivalent inputs, and, more importantly, this deficiency cannot be solved by adding adversarial observations in the training data.