We present datasets of conversations between an agent and a simulated user. These conversations are collected using our M2M framework that combines dialogue self-play and crowd sourcing to exhaustively generate dialogues. The dialogue self-play step generates dialogue outlines consisting of the semantic frames for each turn of the dialogue. The crowd sourcing step provides natural language realizations for each dialogue turn. More details are available in this paper. Please cite the paper if you use or discuss these datasets in your work:
@article{shah2018building,
title={Building a Conversational Agent Overnight with Dialogue Self-Play},
author={Shah, Pararth and Hakkani-T{\"u}r, Dilek and T{\"u}r, Gokhan and Rastogi, Abhinav and Bapna, Ankur and Nayak, Neha and Heck, Larry},
journal={arXiv preprint arXiv:1801.04871},
year={2018}
}
We are releasing two datasets containing dialogues for booking a restaurant table and buying a movie ticket. The number of dialogues in each dataset are listed below. The README file within each directory contains further details about the dataset.
Dataset | Slots | Train | Dev | Test |
---|---|---|---|---|
Sim-R (Restaurant) | price_range, location, restaurant_name, category, num_people, date, time |
1116 | 349 | 775 |
Sim-M (Movie) | theatre_name, movie, date, time, num_people |
384 | 120 | 264 |
Sim-GEN (Movie) | theatre_name, movie, date, time, num_people |
100K | 10K | 10K |
The datasets are provided "AS IS" without any warranty, express or implied. Google disclaims all liability for any damages, direct or indirect, resulting from the use of these datasets.
Please email {abhirast, dilekh}@google.com with questions.
Our publication Scalable Multi-Domain Dialogue State Tracking (IEEE ASRU 2017) reports joint goal accuracy on Sim-R and Sim-M datasets. The released version of the datasets includes fixes for some errors in dialogue state and action annotations. This updated version of the paper reports the results on corrected datasets.
Our publication Dialogue Learning with Human Teaching and Feedback in End-to-End Trainable Task-Oriented Dialogue Systems uses Sim-GEN.