We implemented a BiLSTM-CRF model for NER, which identifies three types of entities: PERSON, LOCATION and ORGANIZATION.

• Python: 3.6
• TensorFlow: 1.8

Dataset comes from MSRA corpus.

The directory ./data contains:

• the preprocessed data files, train_data and test_data
• a vocabulary file word2id.pkl
• a char embedding file vector.npy #This file is a pre-trained char embeddings. Randomly initializes the embedding when this file does not exist.

The directory ./data/case contains:

• a original file'input.txt',# When you need to use our model for NER, you need to put these sentences into 'input.txt'
• a 'result.json' # save all identified entities

The directory ./model is used to save the trained model file

We use BIO tag scheme, with each character its label for one line. Sentences are splited with a null line.

我	O
们	O
是	O
受	O
到	O
郑	B-PER
振	I-PER
铎	I-PER
先	O
生	O
、	O
阿	B-PER
英	I-PER
先	O
生	O
著	O
作	O
的	O
启	O
示	O

We use Dr. Xiao Han’s bert-as-serviceto obtain the BERT embeddings for NER, which is used as follows:

• 1.Environmental requirements: python version >=3.5, tensorflow version >=1.10
• 2.pip install bert-serving-server;
• 3.pip install bert-serving-client
• 4.Download the trained BERT Chinese modelchinese_L-12_H-768_A-12.zip
• 5.start up bert-as-service : bert-serving-start -model_dir /home/chinese_L-12_H-768_A-12 -num_worker=2
• 6.Get the word vector using getBertVector.py

we use BiLSTM, can efficiently use both past and future input information and extract features automatically.

we use CRF and a Softmax, label the tag for each character in one sentence.

python main.py --model=0

the model can be 0 ,1 or 2. "--model=0" means train "--model=1" means test and "--model=2" means application. If you use "--model=2", you can use the trained model to identify the entities in some original sentences. First, you need to put the original sentence into the ./data/case/input.txt file; Then run the command python main.py --model=2 and the recognized result will be saved to ./data/case/ result.json

The best test performance of the model is:

accuracy:  98.64%; precision:  90.87%; recall:  87.13%; F1:  88.96
        LOC: precision:  93.41%; recall:  90.16%; F1:  91.76  2777
        ORG: precision:  85.64%; recall:  84.67%; F1:  85.15  1316
        PER: precision:  90.78%; recall:  84.38%; F1:  87.46  1844

[1] https://github.com/Determined22/zh-NER-TF

[2] https://github.com/hanxiao/bert-as-service

[3] https://github.com/google-research/bert