schema_guided_dialog

The language presents in the datset is only English. The SGD dataset provides a challenging testbed for a number of tasks in task-oriented dialogue, including language understanding, slot filling, dialogue state tracking and response generation.

For the creation of the SGD dataset, they developed a multi-domain dialogue simulator that generates dialogue outlines over an arbitrary combination of APIs, dialogue states and system actions. Then, they used a crowd-sourcing procedure to paraphrase these outlines to natural language utterances. This novel crowd-sourcing procedure preserves all annotations obtained from the simulator and does not require any extra annotations after dialogue collection.