End-to-end learning of recurrent neural networks (RNNs) is an attractivesolution for dialog systems; however, current techniques are data-intensive andrequire thousands of dialogs to learn simple behaviors. We introduce HybridCode Networks (HCNs), which combine an RNN with domain-specific knowledgeencoded as software and system action templates. Compared to existingend-to-end approaches, HCNs considerably reduce the amount of training datarequired, while retaining the key benefit of inferring a latent representationof dialog state. In addition, HCNs can be optimized with supervised learning,reinforcement learning, or a mixture of both. HCNs attain state-of-the-artperformance on the bAbI dialog dataset, and outperform two commerciallydeployed customer-facing dialog systems.
展开▼
