Khmer unicode text data for unsupervised learning language model

This goal of this repos is serve as a collection of Khmer corpus that can be used to train Khmer language model. These text are already run through word segmentation to add space between words using Conditional Random Fields (CRF).

Corpus so far contains:

• km-wikipedia: Khmer wikipedia downloaded January 2020 (has 7.8 millions words, 110K lines)
• oscar: OSCAR Khmer text (has 10 millions words)
• classification: Khmer news article on traffic accident with labels using Pandas pickle format (has 238K words, 820 articles)

Some of the Khmer news articles can be found here: https://github.com/phylypo/segmentation-crf-khmer/tree/master/data. This repo has 2 sets of 1000 articles segmented using CRF from https://niptict.edu.kh/khmer-word-segmentation-tool/. See more detail post here.

It also contains some sample code to train BERT from scratch. See this Python Notebook. It contains steps by steps to pretrain BERT on Khmer language.

• use Khmer news articles to pretain BERT using word masking
• use classification data to test the model with 99% accuracy for 2 classes