README.md

---
license: cc-by-nc-sa-4.0
language:
- zh
---

[**Kuaipedia**](https://github.com/KwaiKEG/Kuaipedia) is developed by [KwaiKEG](https://github.com/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. 

![demo](./demo-case.gif)

* **Items** is a set of entities and concepts, such as [Shiba Inu](https://en.wikipedia.org/wiki/Shiba_Inu), [Moon](https://en.wikipedia.org/wiki/Moon) and [Galileo Galilei](https://en.wikipedia.org/wiki/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](https://en.wikipedia.org/wiki/Shiba_Inu), or "formation", "surface conditions", "how-to-draw" of item [Moon](https://en.wikipedia.org/wiki/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]](https://arxiv.org/abs/2211.00732)

## Data

**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.

## Reference
```
@article{Kuaipedia22,
  author    = {Haojie Pan and
               Zepeng Zhai and
               Yuzhou Zhang 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}
}
```