“It is a fundamental challenge to evaluate whether a model can truly capture the meaning ofsentences. Evaluation of whether a model well captures the meaning of individual words, how-ever, can be effectively achieved by analyzing whether the model encodes words in a vectorspace where semantically similar words form clusters. Inspired by this approach, we propose theSentence-Space Metrics (SSM) to evaluate model interpretation of sentences, and the sentencespace is constructed based on the pairwise entailment relationships between all sentence pairswithin a sentence pool. We use three metrics to evaluate a sentence space, i.e., (1) sparsity, (2)clustering of related sentences, and (3) similarity with the sentence space measured from hu-mans. The SSM is applied to evaluate 20 models, including ChatGPT, 18 BERT-family modelsfine-tuned for Natural Language Inference (NLI) task, as well as SimCSE, a sentence representa-tion model. The SSM reveals dramatic differences among models: Although all models achievehigh accuracy on standard NLI datasets such as MNLI, none of them mirrors the human behaviorunder the SSM. These results demonstrate that, compared with traditional accuracy measures,the SSM considers pairwise relationships between hundreds of sentences and therefore providea more fine-grained evaluation of model interpretation of sentences.Introduction”

“Pre-trained language models are sensitive to adversarial attacks, and recent works have demon-strated universal adversarial attacks that can apply input-agnostic perturbations to mislead mod-els. Here, we demonstrate that universal adversarial attacks can also be used to harden NLPmodels. Based on NLI task, we propose a simple universal adversarial attack that can misleadmodels to produce the same output for all premises by replacing the original hypothesis with anirrelevant string of words. To defend against this attack, we propose Training with UNiversalAdversarial Samples (TUNAS), which iteratively generates universal adversarial samples andutilizes them for fine-tuning. The method is tested on two datasets, i.e., MNLI and SNLI. It isdemonstrated that, TUNAS can reduce the mean success rate of the universal adversarial attackfrom above 79% to below 5%, while maintaining similar performance on the original datasets. Furthermore, TUNAS models are also more robust to the attack targeting at individual samples:When search for hypotheses that are best entailed by a premise, the hypotheses found by TUNASmodels are more compatible with the premise than those found by baseline models. In sum, weuse universal adversarial attack to yield more robust models. Introduction”