ConvLab/tm3

Dataset Card for Taskmaster-3

• Repository: https://github.com/google-research-datasets/Taskmaster/tree/master/TM-3-2020
• Paper: https://aclanthology.org/2021.acl-long.55.pdf
• Leaderboard: None
• Who transforms the dataset: Qi Zhu(zhuq96 at gmail dot com)

To use this dataset, you need to install ConvLab-3 platform first. Then you can load the dataset via:

from convlab.util import load_dataset, load_ontology, load_database

dataset = load_dataset('tm3')
ontology = load_ontology('tm3')
database = load_database('tm3')

For more usage please refer to here.

Dataset Summary

The Taskmaster-3 (aka TicketTalk) dataset consists of 23,789 movie ticketing dialogs (located in Taskmaster/TM-3-2020/data/). By "movie ticketing" we mean conversations where the customer's goal is to purchase tickets after deciding on theater, time, movie name, number of tickets, and date, or opt out of the transaction.

This collection was created using the "self-dialog" method. This means a single, crowd-sourced worker is paid to create a conversation writing turns for both speakers, i.e. the customer and the ticketing agent. In order to gather a wide range of conversational scenarios and linguistic phenomena, workers were given both open-ended as well as highly structured conversational tasks. In all, we used over three dozen sets of instructions while building this corpus. The "instructions" field in data.json provides the exact scenario workers were given to complete each dialog. In this way, conversations involve a wide variety of paths, from those where the customer decides on a movie based on genre, their location, current releases, or from what they already have in mind. In addition, dialogs also include error handling with repect to repair (e.g. "No, I said Tom Cruise."), clarifications (e.g. "Sorry. Did you want the AMC 16 or Century City 16?") and other common conversational hiccups. In some cases instructions are completely open ended e.g. "Pretend you are taking your friend to a movie in Salem, Oregon. Create a conversation where you end up buying two tickets after finding out what is playing in at least two local theaters. Make sure the ticket purchase includes a confirmation of the deatils by the agent before the purchase, including date, time, movie, theater, and number of tickets." In other cases we restrict the conversational content and structure by offering a partially completed conversation that the workers must finalize or fill in based a certain parameters. These partially completed dialogs are labeled "Auto template" in the "scenario" field shown for each conversation in the data.json file. In some cases, we provided a small KB from which workers would choose movies, theaters, etc. but in most cases (pre-pandemic) workers were told to use the internet to get accurate current details for their dialogs. In any case, all relevant entities are annotated.

• How to get the transformed data from original data:
  • Download master.zip.
  • Run python preprocess.py in the current directory.
• Main changes of the transformation:
  • Remove dialogs that are empty or only contain one speaker.
  • Split each domain dialogs into train/validation/test randomly (8:1:1).
  • Merge continuous turns by the same speaker (ignore repeated turns).
  • Annotate dialogue acts according to the original segment annotations. Add intent annotation (==inform). The type of dialogue act is set to non-categorical if the slot is not description.other or description.plot. Otherwise, the type is set to binary (and the value is empty). If there are multiple spans overlapping, we only keep the shortest one, since we found that this simple strategy can reduce the noise in annotation.
  • Add domain and intent descriptions.
  • Rename api to db_results.
  • Add state by accumulate non-categorical dialogue acts in the order that they appear.
• Annotations:
  • dialogue acts, state, db_results.

Supported Tasks and Leaderboards

NLU, DST, Policy, NLG, E2E

Languages

English

Data Splits

split	dialogues	utterances	avg_utt	avg_tokens	avg_domains	cat slot match(state)	cat slot match(goal)	cat slot match(dialogue act)	non-cat slot span(dialogue act)
train	18997	380646	20.04	10.48	1	-	-	-	100
validation	2380	47531	19.97	10.38	1	-	-	-	100
test	2380	48849	20.52	10.12	1	-	-	-	100
all	23757	477026	20.08	10.43	1	-	-	-	100

1 domains: ['movie']

• cat slot match: how many values of categorical slots are in the possible values of ontology in percentage.
• non-cat slot span: how many values of non-categorical slots have span annotation in percentage.

Citation

@inproceedings{byrne-etal-2021-tickettalk,
    title = "{T}icket{T}alk: Toward human-level performance with end-to-end, transaction-based dialog systems",
    author = "Byrne, Bill  and
      Krishnamoorthi, Karthik  and
      Ganesh, Saravanan  and
      Kale, Mihir",
    booktitle = "Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)",
    month = aug,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.acl-long.55",
    doi = "10.18653/v1/2021.acl-long.55",
    pages = "671--680",
}

Licensing Information

CC BY 4.0