NER-使用kashgari中的Bert+BiLSTM+CRF

kashgari做Bert+BiLSTM+CRF

kashgari：

• 是一個基於tensorflow的做Bert+LSTM模型的庫

• 庫的源碼可以參考：https://github.com/BrikerMan/Kashgari

• kashgari基於tensorflow2.0的版本里面沒有CRF，所以使用的版本如下

環境：

• conda create --envs myTestNER python==3.6
• pip insall tensorflow==1.14.0
• pip install kashgari==1.1.5`

數據准備：

train_x = []
train_y = []
with open('data_file_name', encoding='utf-8') as f:
	for line in f.readlines():
		cur = line.strip().split()
		train_x.append(list(data[0]))
		train_y.append(list(data[1]))
# train_x = [[seq1], [seq2], [seq3], ...]
# train_y = [[tag1], [tag2], [tag3], ...]
# seq1 = ['我'， '愛', '北', '京']
# tag1 = ['O', 'O', 'B-D', 'I-D']

訓練時候需要下載BERT預訓練模型，從這里下載：https://github.com/ymcui/Chinese-BERT-wwm

這里訓練很簡單，超參數調節可以參考： https://github.com/BrikerMan/Kashgari/blob/v2-trunk/docs/tutorial/text-labeling.md

訓練：

import kashgari
from kashgari.embeddings import BERTEmbedding
from kashgari.tasks.labeling import BiLSTM_CRF_Model
bert = BERTEmbedding(model_folder="chinese_roberta_wwm_ext_L-12_H-768_A-12", sequence_length=256, task=kashgari.LABELING)
model = BiLSTM_CRF_Model(bert)
model.fit(train_x, train_y, x_validate=train_x, y_validate=train_y, epochs=2, batch_size=32)
model.save('my_bert_crf.h5')

預測：

import kashgari
model = kashgari.utils.load_model('my_bert_crf.h5')
predict = model.predict([['我'， '愛', '北', '京']])		# 二維
print(predict)
# [['O', 'O', 'B-D', 'I-D']]