Recent advances in neural sequence-to-sequence models have led to promising\nresults for several language generation-based tasks, including dialogue\nresponse generation, summarization, and machine translation. However, these\nmodels are known to have several problems, especially in the context of\nchit-chat based dialogue systems: they tend to generate short and dull\nresponses that are often too generic. Furthermore, these models do not ground\nconversational responses on knowledge and facts, resulting in turns that are\nnot accurate, informative and engaging for the users. In this paper, we propose\nand experiment with a series of response generation models that aim to serve in\nthe general scenario where in addition to the dialogue context, relevant\nunstructured external knowledge in the form of text is also assumed to be\navailable for models to harness. Our proposed approach extends\npointer-generator networks (See et al., 2017) by allowing the decoder to\nhierarchically attend and copy from external knowledge in addition to the\ndialogue context. We empirically show the effectiveness of the proposed model\ncompared to several baselines including (Ghazvininejad et al., 2018; Zhang et\nal., 2018) through both automatic evaluation metrics and human evaluation on\nCONVAI2 dataset.\n