cod

Cross-lingual Outline-based Dialogue (COD) is a dataset comprised of manually generated, localized, and cross-lingually aligned Task-Oriented-Dialogue (TOD) data that served as the source of dialogue prompts.

COD enables natural language understanding, dialogue state tracking, and end-to-end dialogue modeling and evaluation.

Majewska et al. (2022) create COD using a novel outline-based annotation pipeline for multilingual TOD by Majewska et al. (2022).

English Schema-Guided Dialogue (SGD; Shah et al., 2018; Rastogi et al., 2020) dataset is automatically sampled and mapped into outlines. The outlines are then paraphrased and adapted to the local target domain by human subjects.

Dataset Usage

Run pip install nusacrowd before loading the dataset through HuggingFace's load_dataset.

Citation

@article{majewska2022cross,
  title={Cross-lingual dialogue dataset creation via outline-based generation},
  author={Majewska, Olga and Razumovskaia, Evgeniia and Ponti, Edoardo Maria and Vuli{'c}, Ivan and Korhonen, Anna},
  journal={arXiv preprint arXiv:2201.13405},
  year={2022}
}

License

Unknown

Homepage

https://github.com/cambridgeltl/COD

NusaCatalogue

For easy indexing and metadata: https://indonlp.github.io/nusa-catalogue