This is a Pytorch implementation of Named Entity Recognition (NER) for Vietnamese. Pretrained model from BERT Hugging Face and Underthesea Toolkit for sentence segmentation to pre-process input raw text is used.

Requirements:

• Python 3.7
• Pytorch 1.7

Install dependencies:

pip install -r requirements.txt

You can review the requirements.txt's comment for an appropriate Pytorch version.

The input data is from VietAI Advanced NLP-02 class.

Its format follows CoNLL-2003 format with four columns separated by a tab character, including of Word, POS tag, Chunking tag, Named-Entity tag, and nested Named-Entity. However, I did not use the nested named entity column due to the requirement of VietAI Advanced NLP-02 class.

The data consists of 4 entities:

• PER: person.
• ORG: organization.
• LOC: location.
• MISC: miscellaneous.

and

• O: others

Note: Structure of the dataset

├── data
|  └── raw_data
|   └── test
|   └── train

Each word is put on a separate line and there is a start-of-sentence and end-of-sentence labels before and after each sentence, respectively. The table below describes an example Vietnamese sentence in dataset.

1. Create data in .csv format: python create_data.py (The train dataset is split into Train-70% and Validation-30%)

2. Train the model: python train.py --data_dir new_data/ --model_name_or_path bert-base-multilingual-cased --output_dir outputs

3. Convert the outputs of the model to the required format for submission: python create_output.py

You can visualize the prediction: python predict.py --pretrain_dir outputs/

The overall performance of the model on validation dataset: