WOW++

This README describes WOW++, an augmented version of the WOW dataset that contains multiple knowledge sentences for the last turn within a dialogue.

We hope releasing this dataset will benefit the community for research in knowledge selection for open-domain dialog.

The original WOW dataset can be found here.

WOW++ contains three different splits of data: train/random (test seen)/topic (test unseen). The three files in the dataset include:

• full_test_random_augmented_wow.json (198 dialogues)
• full_test_topic_augmented_wow.json (200 dialogues)
• full_train_augmented_wow.json (400 dialogues)

Conversation Format

Each file consists of a mapping from dialogue id to a dict with the following info:

• turns: The dialogue text
• topic: The topic selected for the original dialogue
• start_speaker: Which speaker began the dialogue
• id: The dialogue id
• knowledges: The knowledges provided to the Turker in the original dialogue
• gold_sentence: The chosen knowledge sentence by the Turker in the original WOW
• annotated_sentences: The annotated WOW++ knowledge sentences with confidence levels (percentage of Turkers that agreed in picking that sentence).

Citation

If you use the dataset in your work please cite the following

@inproceedings{eric2021multi,
  title={Multi-Sentence Knowledge Selection in Open-Domain Dialogue},
  author={Eric, Mihail and Chartier, Nicole and Hedayatnia, Behnam and Gopalakrishnan, Karthik and Rajan, Pankaj and Liu, Yang and Hakkani-Tur, Dilek},
  booktitle={Proceedings of the 14th International Conference on Natural Language Generation},
  pages={76--86},
  year={2021}
}