Punctuation Restoration for Vietnamese

Our Implementation of seq2punct. Our problem is restoring missing puncts in Vietnamese sentences. We consider this problem with [just comma and period].

Framework in project

• Python3
• Pytorch
• Flask
• Matplotlib

Result

Dataset format

Input: Seq of words

Output: Seq of punct

Ex: Text "Một đêm nọ tôi nằm mơ thấy em " has the label: " 0 1 0 0 0 0 0 2" with [0, 1, 2] denotes space, comma, period respectively.

You can download preprocessed data in: https://drive.google.com/drive/u/2/folders/1NfpLGRQAJPlURQa-3G6RFR2cKd1KKavt?fbclid=IwAR1-JV8NOTNGcbVtj7BkUbtNTAixxhisg4y1-qqeeOjzZAwaGcUlzAk3jtg

For using

• Clone this repo
• Pip install requirements
• Create dumps/ folder for saving model checkpoint
• You can download pretrained model in: https://drive.google.com/drive/folders/1pKeP6YGsYveJNiAhl9O8vdG_OAoIx1Gk?usp=sharing

Train

To train by yourselft

python main.py --model [RNN] --n_layers [2] -- embedding_size [256] --hidden_dim [256]

Inference

To infere: Change the path to pretrained model and run command

from infere import pipeline

text = 'Tôi là Khang tôi là sinh viên trường Bách Khoa'

res = pipeline(text)

Inference with bert

We use pretrained phobert to extract the features of text, then add RNN and MLP to classify punct. The result ~ 95% accuracy

Train with training_bert.ipynb Infere with infere_bert.py

Run demo with flask

python app.py