Open-domain neural dialogue models have achieved high performance in response\nranking and evaluation tasks. These tasks are formulated as a binary\nclassification of responses given in a dialogue context, and models generally\nlearn to make predictions based on context-response content similarity.\nHowever, over-reliance on content similarity makes the models less sensitive to\nthe presence of inconsistencies, incorrect time expressions and other factors\nimportant for response appropriateness and coherence. We propose approaches for\nautomatically creating adversarial negative training data to help ranking and\nevaluation models learn features beyond content similarity. We propose\nmask-and-fill and keyword-guided approaches that generate negative examples for\ntraining more robust dialogue systems. These generated adversarial responses\nhave high content similarity with the contexts but are either incoherent,\ninappropriate or not fluent. Our approaches are fully data-driven and can be\neasily incorporated in existing models and datasets. Experiments on\nclassification, ranking and evaluation tasks across multiple datasets\ndemonstrate that our approaches outperform strong baselines in providing\ninformative negative examples for training dialogue systems.\n