DeepNER
An Easy-to-use, Modular and Prolongable package of deep-learning based Named Entity Recognition Models.
This repository contains complex Deep Learning models for named entity recognition.
Requirements
Tensorflow-gpu=1.14.0
jieba=0.37
Train Models
- Train Transformer with CRF
python test_Transformer_CRF.py --num_blocks 2
- Train Bilstm with CRF
python test_BiLSTM_CRF.py
- fintuning bert with freezing bert Variables
python test_BiLSTM_CRF.py --freeze_bert True
- fintuning bert Variables simultaneously
python test_BiLSTM_CRF.py --freeze_bert False
Simply Use Bert Sentence Embeddings
Please have a look at tests/predict_bert_sentence_emb.py
bc = BertVector(config)
embs = bc.encode(["这个项目好用吗", "希望被多多点赞！", "修改配置请看config字典"])
Using Bert for text classification
Please have a look at tests/bert_classification.py
model = BertClassification(config)
model.train()
Performances Comparison
| models | Precision | Recall | F1-Score |
|---|---|---|---|
| Transformer-CRF(2 Layers) | 67.56% | 62.88% | 65.14% |
| BiGRU-CRF | 91.66% | 89.85% | 90.75% |
| BiLSTM-CRF | 91.90% | 89.85% | 90.87% |
| Bert-BiLSTM-CRF(freeze) | 94.56% | 95.09% | 94.82% |
| Bert-BiLSTM-CRF(fintuning bert simultaneously) | 95.33% | 94.69% | 95.01% |
TODO
- Lexicon enhance
- Label Attention Network for fine-gained NER
- Nested NER
