PRESTO

PRESTO: A Multilingual Dataset for Parsing Realistic Task-Oriented Dialogs

Introduction

PRESTO is a dataset of over 550K contextual multilingual conversations between humans and virtual assistants. PRESTO contains a diverse array of challenges that occur in real-world NLU tasks such as disfluencies, code-switching, and revisions. It is the only large-scale human generated conversational parsing dataset that provides structured context such as a user's contacts and lists for each example.

The dataset can be downloaded here. This README documents the dataset structure and other important information about the dataset.

Dataset Structure

PRESTO is published with the following files:

├── presto_dataset.jsonl
├── presto_dev.jsonl
├── presto_test.jsonl
├── presto_train.jsonl
└── test_partitions
    ├── de-DE
    │   ├── de-DE_code-mixing
    │   │   └── test.jsonl
    │   ├── de-DE_disfluency
    │   │   └── test.jsonl
    │   ├── de-DE_non_contextual
    │   │   └── test.jsonl
    │   ├── de-DE_no_phenomena
    │   │   └── test.jsonl
    │   ├── de-DE_revisions
    │   │   └── test.jsonl
    │   └── test.jsonl
    ... (more languages)

presto_dataset.jsonl contains all examples in the dataset. The provided train, dev, and test sets are those used in the paper. Each test partition contains a subset of the examples from the full test set.

Each JSONL file contains a list of examples. An example has the following fields:

• inputs: The last utterance in the user's conversation with the virtual assistant. E.g. Make a movie list.
• targets: The annotated semantic parse of the user's utterance. E.g. Create_list ( label « movie » ).
• metadata: An object with the following fields:
  • example_id: A unique identifier for the example.
  • locale: The locale of the conversation. E.g. en-US.
  • split: The split of the example. One of train, dev, or test.
  • context: Whether the example's context was present in the original conversation or added synthetically. One of human or synthetic.
  • linguistic_phenomena: The linguistic phenomenon associated with the example. E.g. disfluency. If the example is not associated with any labeled phenomenon, this field is the empty string.
  • previous_turns: A list of the previous turns in the conversation, chronologically ordered. Each turn is an object with the following fields:
    • user_query: The user's utterance in the turn. E.g. Add an item to my shopping list.
    • response_text: The virtual assistant's response to the user's utterance. E.g. What do you want to add?.
  • seeded_lists: A list of the lists present in the user's virtual environment. Each list is an object has the following fields:
    • name: The name of the list. E.g. Shopping.
    • items: A list of the items in the list. E.g. ["milk", "eggs", "bread"].
  • seeded_notes: A list of the notes present in the user's virtual environment. Each note is an object has the following fields:
    • name: The name of the note. E.g. Workout routine.
    • text: The text of the note. E.g. 10 pushups, 10 situps, 10 squats.
  • seeded_contacts: A list of the contacts present in the user's virtual environment. Each contact is a string containing the name of the contact. E.g. ["John Doe", "Jane Doe"].

Citation

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

@misc{https://doi.org/10.48550/arxiv.2303.08954,
  doi = {10.48550/ARXIV.2303.08954},
  url = {https://arxiv.org/abs/2303.08954},
  author = {Goel, Rahul and Ammar, Waleed and Gupta, Aditya and Vashishtha, Siddharth and Sano, Motoki and Surani, Faiz and Chang, Max and Choe, HyunJeong and Greene, David and He, Kyle and Nitisaroj, Rattima and Trukhina, Anna and Paul, Shachi and Shah, Pararth and Shah, Rushin and Yu, Zhou},
  title = {PRESTO: A Multilingual Dataset for Parsing Realistic Task-Oriented Dialogs},
  publisher = {arXiv},
  year = {2023},
}

License

PRESTO is licensed under CC BY 4.0.