Named Entity Recognition

Features

• Use the pretrained Korean BERT (KoBERT) from SKT Brain
• The tokenizer used is also from KoBERT with vocab size 8002
• The data is provided from 한국해양대학교 자연언어처리 연구실, with 23032 training sentences and 931 for validation
• BIO-tagging scheme is applied with 10 distinct entity labels (hence total 21 possible labels)
• Experimented with two models, one with KoBERT-FF-CRF and the other KoBERT-FF-LSTM-CRF
• The result for iterating 12 epochs

	Training Acc	Validation Acc
with LSTM	97.2%	93.3%
without LSTM	98.0%	94.1%

Python version: This code is in Python3.6

Updates: An option for removing 조사 within an entity text (eg. 홍길동은(PER)->홍길동(PER)) is provided based on Komoran pos tagger

Steps for Training Model

Please note that all configurations for training model (hyper_parameters, save_path, load_path, etc) is defined at constant.py

Data Process

python preprocess.py 

By default, all raw data is located at ./data/raw_data and the processed data after running preprocess.py will be saved at ./data/processed_data in the form of .pt files

Model Training

python train.py -model_type bert-lstm-crf -log ./logs/bert-lstm-crf.log

The model_type option should be either bert-crf or bert-lstm-crf depending your choice for the model. Logs for training and evaluation will be recorded at ./logs/bert-lstm-crf.log

Please note that you can change batch_size, epochs etc from constant.py

Inference

python inference_cmd.py -model bert-lstm-crf

This will load a saved model with the highest accuracy from ../models/bert-lstm-crf.