This dataset and how it came to be, along with some baseline models, are described in this paper.
Each JSON file in the dialogues directory contains one dialogue in the following format:
| Key | Value |
|---|---|
| "AnonymizedUserWorkerID" | String that is unique for each worker but unrelated to the worker's AMT Worker ID |
| "AnonymizedWizardWorkerID" | String that is unique for each worker but unrelated to the worker's AMT Worker ID |
| "BatchID" | We collected dialogues in batches, identified by this ID |
| "CompletionLevel" | Can be "Complete", "EarlyDisconnectDuringDialogue", or "DisconnectDuringDialogue" |
| "DialogueID" | Unique ID of this dialogue |
| "Events" | List of events representing the dialogue |
| "FORMAT-VERSION" | |
| "Scenario" | Dictionary containing information about the scenario of this dialogue |
| "UserQuestionnaire" | List of question/answer pairs for questions given to the user |
| "WizardQuestionnaire" | List of question/answer pairs for questions given to the wizard |
Please use the following bibtex entry if you are using STAR for your research:
@article{mosig2020star,
author = {Johannes E. M. Mosig and Shikib Mehri and Thomas Kober},
title = "{STAR: A Schema-Guided Dialog Dataset for Transfer Learning}",
journal = {arXiv e-prints},
keywords = {Computer Science - Computation and Language},
year = 2020,
month = oct,
eid = {arXiv:2010.11853},
archivePrefix = {arXiv},
eprint = {2010.11853},
primaryClass = {cs.CL},
}