Kuaipedia is developed by Knowledge Engineering Group in Kuaishou (KwaiKEG), collaborating with HIT and HKUST. It is the world's first large-scale multi-modal short-video encyclopedia where the primitive units are items, aspects, and short videos.
- Items is a set of entities and concepts, such as Shiba Inu, Moon and Galileo Galilei, which can be edited at one Wikipedia page. An item may have a title, a subtitle, a summary, attributes, and other detailed information of the item.
- Aspects is a set of keywords or keyphrases attached to items. Those keywords are used to describe specific aspects of the item. For example, "selection", "food-protecting", "color" of item Shiba Inu, or "formation", "surface conditions", "how-to-draw" of item Moon.
- Videos is a set of short-videos whose duration may not exceed 5 minutes. In this paper, we only focus on knowledge videos we detected, Where we follow OECD to define knowledge as:
- Know-what refers to knowledge about facts. E.g. How many people live in New York?
- Know-why refers to scientific knowledge of the principles and laws of nature. E.g. Why does the earth revolve around the sun?
- Know-how refers to skills or the capability to do something. E.g. How to cook bacon in the oven.
Please refer to the paper for more details.
Kuaipedia: a Large-scale Multi-modal Short-video Encyclopedia [Manuscript]
- 2023.06 - Release a subset of [data]
- 2022.11 - 业界首个！快手提出亿级别多模态短视频百科体系快知 (The first! Kuaishou proposed a large-scale multi-modal short-video encyclopedia "Kuaipeida") [机器之心][澎湃] [知乎] [CSDN][51CTO] [IT之家]
- 2022.08 - Obtain "Outstanding Project Award" in Kuaishou AI Day.
We are excited to release a subset of Kuaipedia, featuring the most popular wiki entries for enhanced research opportunities. Along with this, we've also shared our experimental findings. Sample files can be located in the ./data folder, accompanied by a README.md file to clarify each field.
To download the full subset and experimental results of Kuaipedia, please go ahead to huggingface/dataset/kuaipedia, or use the following link:
link: https://pan.baidu.com/s/1yUB97aL2rBVt-Q0c6sYIcw code: kwyw
The raw video can be found by concatenating video_id with the prefix kuaishou.com/short-video. E.g. kuaishou.com/short-video/3xwwuqndapzs6nu.
If you're experiencing any issues with downloading the data file, please don't hesitate to reach out to [email protected] for assistance.
Statistics
| Full Dump | Subset Dump | |
|---|---|---|
| #Items | > 26 million | 51,702 |
| #Aspects | > 2.5 million | 1,074,539 |
| #Videos | > 200 million | 769,096 |
The comparative results with the baseline models are as follows:
| Model | Item P | Item R | Item-Aspect P | Item-Aspect R |
|---|---|---|---|---|
| Random | 87.7 | 49.8 | 36.4 | 49.6 |
| LR | 90.4 | 68.3 | 55.1 | 2.7 |
| T5-small | 93.7 | 76.1 | 79.3 | 58.5 |
| BERT-base | 94.3 | 77.8 | 81.5 | 62.7 |
| GPT-3.5 | 90.5 | 86.4 | 41.8 | 95.7 |
| Ours | 94.7 | 79.7 | 83.0 | 65.7 |
Feel free to explore and utilize this valuable dataset for your research and projects.
@article{Kuaipedia22,
author = {Haojie Pan and
Yuzhou Zhang and
Zepeng Zhai and
Ruiji Fu and
Ming Liu and
Yangqiu Song and
Zhongyuan Wang and
Bing Qin
},
title = {{Kuaipedia:} a Large-scale Multi-modal Short-video Encyclopedia},
journal = {CoRR},
volume = {abs/2211.00732},
year = {2022}
}
Except for the contributers to the paper. We also appreciate efforts and helps from Jingrun Zhang, Yuelei Li, Lijun Mei, Chunguang Pan, Xing Hu, Lingyu Zou, Yang Li, Dexing Yang, Wenzheng Zhao, Guixin Qiu, Lin Yang, Meijuan Yang, Teng Tu, Xinyi Zheng, Yunhui Guo and others who contributes to this project.
If you are insterested in Kuaipedia and more cases, please contact us by e-mail [email protected]