Conversation Graph: Data Augmentation, Training and Evaluation for Non-Deterministic Dialogue Management

Task-oriented dialogue systems typically rely on large amounts of\nhigh-quality training data or require complex handcrafted rules. However,\nexisting datasets are often limited in size considering the complexity of the\ndialogues. Additionally, conventional training signal inference is not suitable\nfor non-deterministic agent behaviour, i.e. considering multiple actions as\nvalid in identical dialogue states. We propose the Conversation Graph\n(ConvGraph), a graph-based representation of dialogues that can be exploited\nfor data augmentation, multi-reference training and evaluation of\nnon-deterministic agents. ConvGraph generates novel dialogue paths to augment\ndata volume and diversity. Intrinsic and extrinsic evaluation across three\ndatasets shows that data augmentation and/or multi-reference training with\nConvGraph can improve dialogue success rates by up to 6.4%.\n