PyTorch solution of NER task Using BiLSTM-CRF model.
This repo contains a PyTorch implementation of a BiLSTM-CRF model for named entity recognition task.
Structure of the code
At the root of the project, you will see:
├── pyner
| └── callback
| | └── lrscheduler.py
| | └── trainingmonitor.py
| | └── ...
| └── config
| | └── basic_config.py #a configuration file for storing model parameters
| └── dataset
| └── io
| | └── data_loader.py
| | └── data_transformer.py
| └── model
| | └── embedding
| | └── layers
| | └── nn
| └── output #save the ouput of model
| └── preprocessing #text preprocessing
| └── train #used for training a model
| | └── trainer.py
| | └── ...
| └── utils # a set of utility functions
| └── test
├── test_predict.py
├── train_bilstm_crf.py
├── train_word2vec.py
Dependencies
- csv
- tqdm
- numpy
- pickle
- scikit-learn
- PyTorch 1.0
- matplotlib
How to use the code
- Download the
source_BIO_2014_cropus.txtfrom BaiduPan(password: 1fa3) and place it into the/pyner/dataset/raWdirectory. - Modify configuration information in
pyner/config/basic_config.py(the path of data,...). - run
python train_bilstm_crf.py． - run
python test_predict.py．
Result
----------- Train entity score:
Type: LOC - precision: 0.9043 - recall: 0.9089 - f1: 0.9066
Type: PER - precision: 0.8925 - recall: 0.9215 - f1: 0.9068
Type: ORG - precision: 0.8279 - recall: 0.9016 - f1: 0.8632
Type: T - precision: 0.9408 - recall: 0.9462 - f1: 0.9435
----------- valid entity score:
Type: T - precision: 0.9579 - recall: 0.9558 - f1: 0.9568
Type: PER - precision: 0.9058 - recall: 0.9205 - f1: 0.9131
