The Chinese BERT tokenizer is widely used among many Chinese NLP models, including the Chinese BART model. However, it cannot be directly applied to Cantonese because it is mainly designed to tokenise Simplified Chinese, whereas Hong Kong Cantonese is mainly written in Traditional Chinese. Moreover, it lacks those Chinese characters that are normally used only in Cantonese, such as '冧', '嚿' and '曱'. Therefore, this project provides a BERT Tokenizer with vocabulary tailored for Cantonese.
To make use of the Chinese BART model, which has already been pre-trained, I plan to conduct a second-stage pre-training of that model on Cantonese datasets. Therefore, the work consists of three parts:
- Convert the tokens of the original tokeniser from Simplified Chinese to Traditional Chinese, while keeping the corresponding embeddings fixed
- Given Cantonese datasets, calculate how many Cantonese-specific characters are missing in the original tokenizer, and add them to the vocabulary
- Randomly initialise new embeddings for new tokens
For the Simplified Chinese-to-Traditional Chinese conversion, the dictionary is from StarCC, which is developed by myself.
For the conversion of the embedding, I designed an algorithm to handle the one-to-many problem. See build_vocab_mapping.py for details.
For the Cantonese datasets I used for calculate Cantonese-specific characters, I select the top 150 charactes from 524288 sentences randomly selected from the LIHKG dataset, as well as the words.hk translation dataset.
sudo apt install -y git-lfs
python3.10 -m venv ./venv
. ./venv/bin/activate
pip install -U pip
pip install -U wheel
pip install "jax[tpu]==0.3.23" -f https://storage.googleapis.com/jax-releases/libtpu_releases.html
pip install -r requirements.txt
huggingface-cli loginwget https://raw.githubusercontent.com/StarCC0/dict/4cf962546a1faf680ac8b15e8ff5d501c98a96ea/STCharacters.txt
wget https://raw.githubusercontent.com/StarCC0/dict/4cf962546a1faf680ac8b15e8ff5d501c98a96ea/TSCharacters.txt
gdown 1F2XlacTo3dTyNItsWY65ocJFgyA9u5CW # lihkg-1-2850000-processed-dedup.csv.xz
xz --decompress lihkg-1-2850000-processed-dedup.csv.xz
# Manually download `yue.txt` from the Cantonese translation datasetpython build_vocab_mapping.py
python add_cantonese_tokens.py
python replace_embedding.py
upload_to_hub.py
# Manually upload `embed_params.dat` to `Ayaka/bart-base-cantonese`