CookDial: a dataset for task-oriented dialogs grounded in procedural documents

This work presents a new dialog dataset, CookDial, that facilitates research on task-oriented systems with procedural knowledge understanding. The corpus contains 260 human-to-human dialogs in which an agent, given recipe document, guides the user to cook dish. Dialogs CookDial exhibit two unique features: (i) alignment between flow and supporting document; (ii) complex agent decision-making involves segmenting long sentences, paraphrasing hard instructions resolving coreference context. In addition, we identify three challenging (sub)tasks assumed system: (1) User Question Understanding, (2) Agent Action Frame Prediction, (3) Response Generation. For each of these tasks, develop neural baseline model, evaluate dataset. We publicly release comprising rich annotations both documents, stimulate further domain-specific document-grounded systems.