MMDialog: A Large-scale Multi-turn Dialogue Dataset Towards Multi-modal Open-domain Conversation

This repository is the official site of ACL'23 paper: MMDialog: A Large-scale Multi-turn Dialogue Dataset Towards Multi-modal Open-domain Conversation

About the dataset

A Dialogue Case of MMDialog:

Statistics:

If you use it in your work, please cite our paper:

@inproceedings{feng-etal-2023-mmdialog,
    title = "{MMD}ialog: A Large-scale Multi-turn Dialogue Dataset Towards Multi-modal Open-domain Conversation",
    author = "Feng, Jiazhan and Sun, Qingfeng and Xu, Can and Zhao, Pu and Yang, Yaming and Tao, Chongyang and Zhao, Dongyan and Lin, Qingwei",
    booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.acl-long.405",
    doi = "10.18653/v1/2023.acl-long.405",
    pages = "7348--7363"
}

Dataset Folder Format:

File: conversations.json

Note:

1. Training set do not contains "negative_candidate_media_keys" and "negative_candidate_texts", which only exists in test and validation set. Each "negative_candidate_xxx" contains 999 negative candidates for retrieval task.
2. All image filenames are in "media_key.jpg" format.
3. Words like :smiling_face_with_smiling_eyes: and :raising_hands: are emotion tokens, please refer to https://github.com/carpedm20/emoji
4. To compute the CLIP scores in metric MM-Relevance, we provide a demo in compute_mmrel.py.
5. We also provide an evaluation example for metrics evaluated within a single modality (e.g., BLEU, Recall) in EvaluationExample.md.

How to get the dataset

To get this dataset, you and your organization require:

1. Who it's for: You are either a master’s student, doctoral candidate, post-doc, faculty, or research-focused employee at an academic institution or university.
2. Non-commercial use: You should only use this access for non-commercial purposes.
3. Clearly Plan: You have a clearly defined research objective, and you have specific plans for how you intend to use and analyze this data from your research.
4. Promise your behavior: You should promise you would not share this dataset without our qualification review and permission.

If you don't meet all of the requirements above, we would not share you the dataset.

We need you to fill in the form below:

Then use your edu or research email account to send the form to [fengjiazhan@pku.edu.cn] for a review, if you meet all the requirements, we would share you a cloud folder which stores the pre-processed dataset within a week.