File size: 1,565 Bytes
677e378
 
 
 
 
 
 
 
 
 
 
87a8ba1
677e378
 
 
 
 
 
 
86e1f66
 
 
87a8ba1
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
---
language:
- zh
- yue
---

### Dataset Summary
Cantonese has been a low-resource language in NLP. This dataset is a major step towards changing that.

To our knowledge, this is the **first large-scale, properly curated, and deduplicated web dataset built specifically for Cantonese**. It was created by filtering years of Common Crawl data and a Cantonese language detector, followed by a deduplication process using MinHash.

The result is a high-quality collection of **~250K unique documents** containing **~150 million words**. 🚀

### Dataset Curation Process
* Downloaded and processed Common Crawl using [code](https://github.com/jedcheng/c4-dataset-script)
* The resultant traditional Chinese dataset can be found [here](https://huggingface.co/datasets/jed351/Traditional-Chinese-Common-Crawl-Filtered)
* Cantonese text filtered using [CantoneseDetect](https://github.com/CanCLID/cantonesedetect) and the resultant dataset can be found [here](https://huggingface.co/datasets/jed351/Cantonese_Common_Crawl_Filtered)
* Using MinHash deduplication to yield the final web dataset (Wikipedia and LIHKG excluded)

### Dataset Files
[CantoneseDetect](https://github.com/CanCLID/cantonesedetect) has two modes: fully Cantonese text or Cantonese text that only present in quotes. 

Both sets of data were cleaned with MinHash and are present in the dataset.

### Acknowoledgement
Special thanks to [Eons Data Communications Limited](https://eons.cloud/) and [Votee AI](https://votee.ai/) for the compute resources to download and process the Common Crawl Dumps.