CantoNLU

This repository accompanies our preprint: "CantoNLU: A benchmark for Cantonese natural language understanding", where we introduce a general language understanding benchmark in Cantonese, in a collaboration with York Hay Ng, Sophia Chan, Helena Zhao, and Annie En-Shiun Lee.

Download model first

This repository requires a local copy of a BERT model and Wikipedia dataset to run.

To download the resources, simply run

python download.py --lang=yue

where lang can be yue or wuu.

Model pre-training

To continually pre-train on Mandarin BERT, simply run

python run.py --pretrain --lang=yue

where lang can be yue or wuu. Additional flags are available--see run.py.

Fine-tuning

To fine-tune on POS and DEPS, the code requires the Cantonese UD file. Download the CoNLL-U file and place it in data/, then use the conllu_2_pos_dataset() function in utils.py.

Pre-trained model weights

The monolingual and transfer models are available at the following Google Drive links.

• Monolingual model: https://drive.google.com/file/d/1wl4MYqPRxj5FPdHJR8SXC7Z7SZCFsLNw/view?usp=drive_link
• Transfer model: https://drive.google.com/file/d/19QKyw-lzbNmU1_EcBUuDuO4TiEfFFuFF/view?usp=drive_link