README.md

metadata

task_categories:
  - text-classification
language:
  - en
  - th
  - es
pretty_name: Multilingual Task-Oriented Dialog
configs:
  - config_name: default
    data_files:
      - path:
          - en/train-en.parquet
          - es/train-es.parquet
          - th/train-th_TH.parquet
        split: train
      - path:
          - en/test-en.parquet
          - es/test-es.parquet
          - th/test-th_TH.parquet
        split: test
      - path:
          - en/eval-en.parquet
          - es/eval-es.parquet
          - th/eval-th_TH.parquet
        split: eval
  - config_name: en
    data_files:
      - split: train
        path: en/train-en.parquet
      - split: test
        path: en/test-en.parquet
      - split: eval
        path: en/eval-en.parquet
  - config_name: es
    data_files:
      - split: train
        path: es/train-es.parquet
      - split: test
        path: es/test-es.parquet
      - split: eval
        path: es/eval-es.parquet
  - config_name: th
    data_files:
      - split: train
        path: th/train-th_TH.parquet
      - split: test
        path: th/test-th_TH.parquet
      - split: eval
        path: th/eval-th_TH.parquet
license: cc-by-sa-4.0

Multilingual Task-Oriented Dialog Data

Directory structure

This dataset consists of 3 directories:

• en contains the English data
• es contains the Spanish data
• th contains the Thai data

In each directory, you'll find a file for each of the train/dev/test splits as used in our paper.

File format

PYTEXT parquet FORMAT

Each parquet file contains following 5 columns: intent label, the slot annotations in a comma-separated list with the format <start token>:<end token>:<slot type>, untokenized utterance, the language, and the token spans from an in-house multilingual tokenizer.

The "upsampled" files contain the upsampled Spanish/Thai data so that there are roughly equal amounts of English and Spanish/Thai data for training and model selection.

License

Provided under the CC-BY-SA license.

Citation

If you use this dataset in your research, please cite the following paper:

@unpublished{Schuster2018,
  author = {Sebastian Schuster and Sonal Gupta and Rushin Shah and Mike Lewis},
    title = {Cross-lingual Transfer Learning for Multilingual Task Oriented Dialog},
    year = {2018},
    note = {arXiv preprint},
    url = {http://arxiv.org/abs/}
}

Questions

Please contact Sonal Gupta ([email protected]) with questions about this dataset.