Adversarial Augmentation Policy Search for Domain and Cross-Lingual Generalization in Reading Comprehension

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3723–3738 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 3723 Adversarial Augmentation Policy Search for Domain and Cross-Lingual Generalization in Reading Comprehension Adyasha Maharana UNC Chapel HillMohit Bansal {adyasha, mbansal}@cs.unc.edu Abstract Reading comprehension models often overfit to nuances of training datasets and fail at ad- versarial evaluation. Training with adversar- ially augmented dataset improves robustness against those adversarial attacks but hurts gen- eralization of the models. In this work, we present several effective adversaries and auto- mated data augmentation policy search meth- ods with the goal of making reading com- prehension models more robust to adversarial evaluation, but also improving generalization to the source domain as well as new domains and languages. We first propose three new methods for generating QA adversaries, that introduce multiple points of confusion within the context, show dependence on insertion lo- cation of the distractor, and reveal the com- pounding effect of mixing adversarial strate- gies with syntactic and semantic paraphras- ing methods. Next, we find that augmenting the training datasets with uniformly sampled adversaries improves robustness to the adver- sarial attacks but leads to decline in perfor- mance on the original unaugmented dataset. We address this issue via RL and more efficient Bayesian policy search methods for automat- ically learning the best augmentation policy combinations of the transformation probability for each adversary in a large search space. Us- ing these learned policies, we show that adver- sarial training can lead to significant improve- ments in in-domain, out-of-domain, and cross- lingual (German, Russian, Turkish) generaliza- tion.1 1 Introduction There has been growing interest in understanding NLP systems and exposing their vulnerabilities through maliciously designed inputs (Iyyer et al., 2018; Belinkov and Bisk, 2018; Nie et al., 2019; 1We will publicly release all our code, adversarial policy data, and models on our webpage. Gurevych and Miyao, 2018). Adversarial exam- ples are generated using search (Alzantot et al., 2018), heuristics (Jia and Liang, 2017) or gradient (Ebrahimi et al., 2018) based techniques to fool the model into giving the wrong outputs. Often, the model is further trained on those adversarial examples to make it robust to similar attacks. In the domain of reading comprehension (RC), adver- saries are QA samples with distractor sentences that have significant overlap with the question and are randomly inserted into the context. By having a fixed template for creating the distractors and training on them, the model identifies learnable bi- ases and overfits to the template instead of being robust to the attack itself (Jia and Liang, 2017). Hence, we first build on Wang and Bansal (2018)’s work of adding randomness to the template and significantly expand the pool of distractor candi- dates by introducing multiple points of confusion within the context, adding dependence on insertion location of the distractor, and further combining distractors with syntactic and semantic paraphrases to create combinatorially adversarial examples that stress-test the model’s language understanding ca- pabilities. These adversaries inflict up to 45% drop in …

Content truncated to 500 words