Vietnamese NER using BERT (bert-vn-ner)

Code by Trong-Dat Ngo.

Introductions

Pytorch implementation of BERT models to perform named entity recognition (NER) for Vietnamese. Besides, the features in word are also used to represent words. This system using pretrained model from transformers

Installation

Requirements:

• Python 3.6+
• Pytorch 1.4+

Install dependencies:

pip3 install -r requirements.txt

Dataset format

The input data's format of bert-vn-ner follows CoNLL-2003 format with four columns separated by a tab character, including of word, pos, chunk, and named entity. Each word has been put on a separate line and there is an empty line after each sentence. Note that word segmentation is not used to match BERT's pretrain (bert-base-multilingual-cased). For details, see sample data in 'data' directory. The table below describes an example Vietnamese sentence in dataset.

Run the code

Train

python3 train.py --data_dir data/ --model_name_or_path bert-base-multilingual-cased --output_dir outputs --logs logs --cuda

Note: file structure of the dataset

├── data
|  └── train.csv
|  └── dev.csv
|  └── test.csv

Predict

python3 predict.py --pretrain_dir pretrains/ --max_seq_length 128 --batch_size 4 --cuda

Note: Using Underthesea NLP Toolkit for sentence segmenttation to pre-process input raw text.

API

python3 api.py

Example api result

{
  "result": [
    {
      "entity": "Nguyễn Đức Vinh",
      "label": "PER"
    },
    {
      "entity": "Sở Nông nghiệp và phát triển nông thôn Hà Giang",
      "label": "ORG"
    },
    {
      "entity": "TP Hà Giang",
      "label": "LOC"
    },
    {
      "entity": "Vinh",
      "label": "PER"
    },
    {
      "entity": "TP Hà Giang",
      "label": "LOC"
    }
  ]
}

VLSP 2016 result

The overall performance of bert-base-multilingual-cased on test set:

References

• Google's BERT
• HuggingFace's Transformers
• Underthesea