Dialogue systems powered by large pre-trained language models (LM) exhibit an\ninnate ability to deliver fluent and natural-looking responses. Despite their\nimpressive generation performance, these models can often generate factually\nincorrect statements impeding their widespread adoption. In this paper, we\nfocus on the task of improving the faithfulness -- and thus reduce\nhallucination -- of Neural Dialogue Systems to known facts supplied by a\nKnowledge Graph (KG). We propose Neural Path Hunter which follows a\ngenerate-then-refine strategy whereby a generated response is amended using the\nk-hop subgraph of a KG. Neural Path Hunter leverages a separate token-level\nfact critic to identify plausible sources of hallucination followed by a\nrefinement stage consisting of a chain of two neural LM's that retrieves\ncorrect entities by crafting a query signal that is propagated over the k-hop\nsubgraph. Our proposed model can easily be applied to any dialogue generated\nresponses without retraining the model. We empirically validate our proposed\napproach on the OpenDialKG dataset against a suite of metrics and report a\nrelative improvement of faithfulness over dialogue responses by 20.35% based on\nFeQA (Durmus et al., 2020).\n