Existing NLP datasets contain various biases, and models tend to quickly\nlearn those biases, which in turn limits their robustness. Existing approaches\nto improve robustness against dataset biases mostly focus on changing the\ntraining objective so that models learn less from biased examples. Besides,\nthey mostly focus on addressing a specific bias, and while they improve the\nperformance on adversarial evaluation sets of the targeted bias, they may bias\nthe model in other ways, and therefore, hurt the overall robustness. In this\npaper, we propose to augment the input sentences in the training data with\ntheir corresponding predicate-argument structures, which provide a higher-level\nabstraction over different realizations of the same meaning and help the model\nto recognize important parts of sentences. We show that without targeting a\nspecific bias, our sentence augmentation improves the robustness of transformer\nmodels against multiple biases. In addition, we show that models can still be\nvulnerable to the lexical overlap bias, even when the training data does not\ncontain this bias, and that the sentence augmentation also improves the\nrobustness in this scenario. We will release our adversarial datasets to\nevaluate bias in such a scenario as well as our augmentation scripts at\nhttps://github.com/UKPLab/data-augmentation-for-robustness.\n