This is a step-by-step implementation of Chinese Title Entity Recognition via BERT-BiLSTM-CRF Model.

For original codes and tutorials, please visit here.
For Chinese readers, please visit here.

My job lies in the Chinese chat title named entity recognition by fine-tuning the BERT model.
So I changed several lines of codes and extended more Chinese chat title entities to the original dataset.
96.73% accuracy has been achieved via the BERT-BiLSTM-CRF.

The main purpose of the job（examples）:

Input One: 小贾你最近忙什么呢？
Input Two: 贾舒越
Output: 小贾 is 贾舒越

Input One: 建勋师兄你何时来实验室？
Input Two: 邸建勋
Output: 建勋师兄 is 王建勋

Input One: 最近王宇航学习怎么样呀
Input Two: 王海生
Output: There is no match for 王海生.

Input One: 贾泽阳现在回家了嘛
Input Two: 吴泽阳
Output: There is no match for 吴泽阳.

For Chinese readers, you guys could read the 提取聊天对方的称谓 - 方案与deadline.pdf to get into the details.

Step One: configure the tensorflow and bert environment

pip install bert-base==0.0.7 -i https://pypi.python.org/simple
tensorflow >= 1.12.0
tensorflow-gpu >= 1.12.0  # GPU version of TensorFlow.
GPUtil >= 1.3.0  # no need if you dont have GPU
pyzmq >= 17.1.0  # python zmq

Step Two: Download the BERT pre-trained model and training dataset

Download the BERT pre-trained model from here.
Be sure to place the extracted folder "chinese_L-12_H-768_A-12" on "init_checkpoint" folder.
training dataset from here.
Be sure to place "train.txt" on the "data" folder.

Step Three: Train the model via command line

Open the CMD terminal or the Anaconda Prompt and be sure to guide it to the working path and tensorflow environment:
e.g. my working path is /Users/shuyuej/Desktop/Python-Files/Chinese-Chat-Title-NER-BERT-BiLSTM-CRF/.

Then input the command:

bert-base-ner-train -data_dir /Users/shuyuej/Desktop/Python-Files/BERT-BiLSTM-CRF-NER/data/ -output_dir /Users/shuyuej/Desktop/Python-Files/BERT-BiLSTM-CRF-NER/final_output/ -init_checkpoint /Users/shuyuej/Desktop/Python-Files/BERT-BiLSTM-CRF-NER/init_checkpoint/chinese_L-12_H-768_A-12\bert_model.ckpt -bert_config_file /Users/shuyuej/Desktop/Python-Files/BERT-BiLSTM-CRF-NER/init_checkpoint/chinese_L-12_H-768_A-12/bert_config.json -vocab_file /Users/shuyuej/Desktop/Python-Files/BERT-BiLSTM-CRF-NER/init_checkpoint/chinese_L-12_H-768_A-12/vocab.txt -batch_size 8

FYI, be sure to change my "/Users/shuyuej/Desktop/Python-Files/BERT-BiLSTM-CRF-NER/" to your own BERT-BiLSTM-CRF-NER path.

For Windows OS System: I use the following command line:

bert-base-ner-train -data_dir E:\BERT-BiLSTM-CRF-NER\data\ -output_dir E:\BERT-BiLSTM-CRF-NER\final_output\ -init_checkpoint E:\BERT-BiLSTM-CRF-NER\init_checkpoint\chinese_L-12_H-768_A-12\bert_model.ckpt -bert_config_file E:\BERT-BiLSTM-CRF-NER\init_checkpoint\chinese_L-12_H-768_A-12\bert_config.json -vocab_file E:\BERT-BiLSTM-CRF-NER\init_checkpoint\chinese_L-12_H-768_A-12\vocab.txt -batch_size 8

The final trained model will be in the "final_output" folder.

Step Four: Test and enjoy the model

The Test File is "test.py" and could test the "Test-set.xlsx" and get a result.
Before you execute the file, be sure to change the paths of trained BERT model, original pre-trained BERT model, and Test-set.xlsx to your own.
And you could see the results and power of BERT.

Another executive file is "predict-test.py" in which you could input the sentence and name and finally get the match results. Be sure to change the paths same as "test.py" file.