No abstract

The release of large natural language inference (NLI) datasets like SNLI and MNLI have led to rapid development and improvement of completely neural systems for the task. Most recently, heavily pre-trained, Transformer-based models like BERT and MT-DNN have reached near-human performance on these datasets. However, these standard datasets have been shown to contain many annotation artifacts, allowing models to shortcut understanding using simple fallible heuristics, and still perform well on the test set. So it is no surprise that many adversarial (challenge) datasets have been created that cause models trained on standard datasets to fail dramatically. Although extra training on this data generally improves model performance on just that type of data, transferring that learning to unseen examples is still partial at best. This work evaluates the failures of state-of-the-art models on existing adversarial datasets that test different linguistic phenomena, and find that even though the models perform similarly on MNLI, they differ greatly in their robustness to these attacks. In particular, we find syntax-related attacks to be particularly effective across all models, so we provide a fine-grained analysis and comparison of model performance on those examples. We draw conclusions about the value of model size and multi-task learning (beyond comparing their standard test set performance), and provide suggestions for more effective training data.

No abstract