Toward Adversarial Training on Contextualized Language Representation

Beyond the success story of adversarial training (AT) in the recent text domain on top of pre-trained language models (PrLMs), our empirical results showcase that current AT can appear mediocre or even harmful on certain tasks, e.g. reading comprehension and commonsense reasoning. This paper investigates AT from the perspective of contextualized language representation. We find that the gain from AT does not derive from increasing the training risk, but from deviating the language representation. The fact is that the current AT attack is better at fooling the decoder (i.e. the classifier), but can be trivial to the encoder. Based on the observations, we propose simple yet effective \textit{Contextualized representation-Adversarial Training} (CreAT), in which the attack is explicitly optimized to deviate the contextualized representation and obtains the global worst-case adversarial examples. CreAT is proven to be all-powerful compared to AT, with performance gain covering a wider range of downstream tasks. We apply CreAT to language pre-training. Our CreAT-empowered DeBERTa outperforms naive DeBERTa by a large margin, achieving the new state-of-the-art performances on a wide range of challenging benchmarks, e.g. AdvGLUE (59.1 $ \rightarrow $ 61.1), HellaSWAG (93.0 $ \rightarrow $ 94.9), ANLI (68.1 $ \rightarrow $ 69.3), PAWS (50.3 $ \rightarrow $ 54.5).