Dense retrieval models, which aim at retrieving the most relevant document\nfor an input query on a dense representation space, have gained considerable\nattention for their remarkable success. Yet, dense models require a vast amount\nof labeled training data for notable performance, whereas it is often\nchallenging to acquire query-document pairs annotated by humans. To tackle this\nproblem, we propose a simple but effective Document Augmentation for dense\nRetrieval (DAR) framework, which augments the representations of documents with\ntheir interpolation and perturbation. We validate the performance of DAR on\nretrieval tasks with two benchmark datasets, showing that the proposed DAR\nsignificantly outperforms relevant baselines on the dense retrieval of both the\nlabeled and unlabeled documents.\n