ubuntu 16.04
anaconda创建一个python==3.7的虚拟环境
pip install -r requirements.txt
使用kashgari调用bert预训练权重和下游分类任务
本实验可以完成BiLSTM_CRF_Model, CNN_LSTM_Model, BiLSTM_Model, BiGRU_Model 或 BiGRU_CRF_Model的搭建
训练数据使用的是：source.txt和target.txt
验证数据使用的是：dev.txt和dev-label.txt
测试数据使用的是：test1.txt和test_tgt.txt
运行实验需要替换 BERTEmbedding()中的预训练权重目录，本次实验选用的chinese_L-12_H-768_A-12
/media/ding/Data/model_weight/nlp_weights/tensorflow/google/chinese_L-12_H-768_A-12
权重下载地址chinese_L-12_H-768_A-12
训练模型，主函数修改为：
main(train=True)
设置网络的参数epoch=5;batch_size=64;sequence_length=150;开始训练模型，保存模型到checkpoint文件夹下
测试模型,主函数修改为：
main(predict=True, checkpoint='checkpoint')
开始对模型在测试集上测试，checkpoint为模型保存的文件夹
预测模型，主函数修改为：
main(predict=True, checkpoint='checkpoint')
开始输入一个样本用模型进行预测，checkpoint为模型保存的文件夹
训练模型：
BiLSTM_CRF_Model;epoch=10;batch_size=64;sequence_length=150;
训练方法：
python ner.py
训练过程：
Epoch 1/10
282/282 [==============================] - 172s 610ms/step - loss: 5.8533 - accuracy: 0.9877 - val_loss: 139.6916 - val_accuracy: 0.9901
Epoch 2/10
282/282 [==============================] - 169s 601ms/step - loss: 1.5082 - accuracy: 0.9967 - val_loss: 137.0616 - val_accuracy: 0.9910
Epoch 3/10
282/282 [==============================] - 172s 610ms/step - loss: 1.0474 - accuracy: 0.9975 - val_loss: 134.8197 - val_accuracy: 0.9912
Epoch 4/10
282/282 [==============================] - 165s 585ms/step - loss: 0.7489 - accuracy: 0.9982 - val_loss: 132.7582 - val_accuracy: 0.9917
Epoch 5/10
282/282 [==============================] - 165s 586ms/step - loss: 0.5523 - accuracy: 0.9986 - val_loss: 131.1968 - val_accuracy: 0.9935
Epoch 6/10
282/282 [==============================] - 167s 592ms/step - loss: 0.4096 - accuracy: 0.9990 - val_loss: 129.8139 - val_accuracy: 0.9946
Epoch 7/10
282/282 [==============================] - 164s 582ms/step - loss: 0.3472 - accuracy: 0.9991 - val_loss: 128.5244 - val_accuracy: 0.9946
Epoch 8/10
282/282 [==============================] - 167s 591ms/step - loss: 0.2750 - accuracy: 0.9993 - val_loss: 127.5287 - val_accuracy: 0.9943
Epoch 9/10
282/282 [==============================] - 169s 600ms/step - loss: 0.1996 - accuracy: 0.9995 - val_loss: 126.6077 - val_accuracy: 0.9942
Epoch 10/10
282/282 [==============================] - 165s 584ms/step - loss: 0.1689 - accuracy: 0.9996 - val_loss: 125.9106 - val_accuracy: 0.9942
验证集的测试结果：
precision recall f1-score support
PER 0.9415 0.9699 0.9555 166
ORG 0.8459 0.8543 0.8501 302
LOC 0.8919 0.8967 0.8943 368
micro avg 0.8853 0.8959 0.8906 836
macro avg 0.8851 0.8959 0.8905 836
测试集的测试结果：
precision recall f1-score support
LOC 0.9070 0.9070 0.9070 258
PER 0.9752 0.9899 0.9825 199
ORG 0.9020 0.8976 0.8998 205
micro avg 0.9262 0.9290 0.9276 662
macro avg 0.9259 0.9290 0.9275 662
预测：
'十 一 月 ， 我 们 被 称 为 “ 南 京 市 首 届 家 庭 藏 书 状 元 明 星 户 ” 。'
输出：
'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'B-LOC', 'I-LOC', 'I-LOC', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O'
BiGRU_CRF_Model;epoch=10;batch_size=64;sequence_length=150;
训练方法：
python ner.py
Epoch 1/10
282/282 [==============================] - 172s 609ms/step - loss: 6.2045 - accuracy: 0.9876 - val_loss: 147.0721 - val_accuracy: 0.9933
Epoch 2/10
282/282 [==============================] - 163s 578ms/step - loss: 1.3152 - accuracy: 0.9970 - val_loss: 144.6661 - val_accuracy: 0.9921
Epoch 3/10
282/282 [==============================] - 163s 578ms/step - loss: 0.8577 - accuracy: 0.9980 - val_loss: 142.6533 - val_accuracy: 0.9921
Epoch 4/10
282/282 [==============================] - 163s 579ms/step - loss: 0.5895 - accuracy: 0.9986 - val_loss: 140.9434 - val_accuracy: 0.9924
Epoch 5/10
282/282 [==============================] - 164s 583ms/step - loss: 0.4318 - accuracy: 0.9989 - val_loss: 139.7154 - val_accuracy: 0.9923
Epoch 6/10
282/282 [==============================] - 164s 580ms/step - loss: 0.3175 - accuracy: 0.9992 - val_loss: 138.5307 - val_accuracy: 0.9928
Epoch 7/10
282/282 [==============================] - 162s 575ms/step - loss: 0.2626 - accuracy: 0.9993 - val_loss: 137.6368 - val_accuracy: 0.9924
Epoch 8/10
282/282 [==============================] - 162s 576ms/step - loss: 0.2176 - accuracy: 0.9994 - val_loss: 136.7970 - val_accuracy: 0.9921
Epoch 9/10
282/282 [==============================] - 162s 576ms/step - loss: 0.2019 - accuracy: 0.9995 - val_loss: 135.7567 - val_accuracy: 0.9928
Epoch 10/10
282/282 [==============================] - 163s 577ms/step - loss: 0.1632 - accuracy: 0.9996 - val_loss: 135.0596 - val_accuracy: 0.9926
验证集的测试结果：
precision recall f1-score support
LOC 0.8909 0.9321 0.9110 368
PER 0.9583 0.9699 0.9641 166
ORG 0.8562 0.8874 0.8715 302
micro avg 0.8915 0.9234 0.9072 836
macro avg 0.8918 0.9234 0.9073 836
测试集的测试结果：
precision recall f1-score support
LOC 0.9129 0.9341 0.9234 258
PER 0.9752 0.9899 0.9825 199
ORG 0.9171 0.9171 0.9171 205
micro avg 0.9329 0.9456 0.9392 662
macro avg 0.9329 0.9456 0.9392 662
预测：
'十 一 月 ， 我 们 被 称 为 “ 南 京 市 首 届 家 庭 藏 书 状 元 明 星 户 ” 。'
输出：
'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'B-LOC', 'I-LOC', 'I-LOC', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O', 'O'