命名实体识别（NER）（二）：BERT+CRF模型训练

摘要：上篇介绍了数据的标注过程，接下来就是模型的训练了，本文采用BERT+CRF模型进行训练。

 

采用kashgari模块（一个将Bert封装好的模块，用于快速搭建模型）的Bert模块快速搭建自己模型

在原来bert预训练模型基础上，导入自己的标注数据进一步训练，

bert中文预训练模型就自己下一下吧~

代码比较简单，就是导入数据，在预训练基础上进一步训练

from kashgari.tasks.seq_labeling import BLSTMCRFModel
from kashgari.embeddings import BERTEmbedding
import kashgari
from kashgari import utils
import os
#os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
#os.environ["CUDA_VISIBLE_DEVICES"] = ""

def get_sequence_tagging_data(file_path):

    data_x, data_y = [], []

    with open(file_path, 'r', encoding='utf-8') as f:
        lines = f.read().splitlines()
        # print(lines)
        x, y = [], []
        for line in lines:
            rows = line.split(' ')
            if len(rows) == 4:
                data_x.append(x)
                data_y.append(y)
                x = []
                y = []
            else:
                x.append(rows[0])
                y.append(rows[1])
    return data_x, data_y
#train_x, train_y = get_sequence_tagging_data('training_data_bert_train.txt')

# train_x, train_y = get_sequence_tagging_data('../new_note.txt')
# print(f"train data count: {len(train_x)}")



model training 
embedding = BERTEmbedding('/pvc/train/chinese_L-12_H-768_A-12',40)
model = BLSTMCRFModel(embedding)
model.fit(train_x,
          train_y,
          validation_split = 0.4,
          epochs=10,
          batch_size=32)
print('model_save')
model.save('../model_save/ner_model')

最后附上一段供测试的代码

load_model = BLSTMCRFModel.load_model('../model_save/ner_model')
print(load_model.predict("刘若英语怎么样"))

从训练数据的标注至NER模型训练流程就算简单走一遍了。

至此，NER本节就算完结了~逃