### 第三章——基于BERT的中文情感分析实战 #### 任务介绍 对中文进行分类demo，分成0/1/2 ![1610163467782](assets/1610163467782.png) ![1610376739912](assets/1610376739912.png) > 可以理解为0是一般，1是好，2是差。模型代码、数据集都在我的网盘里，链接：https://pan.baidu.com/s/18vPGelYCXGqp5OCWZWz36A > 提取码：de0f 我们使用的是Google官方开源的中文BERT预训练模型 ![1610164248560](assets/1610164248560.png) > vocab.txt里把常用的中文基本覆盖了 #### 读取处理自己的数据集 ~~~python class MyDataProcessor(object): """Base class for data converters for sequence classification data sets.""" def get_train_examples(self, data_dir): """Gets a collection of `InputExample`s for the train set.""" raise NotImplementedError() def get_dev_examples(self, data_dir): """Gets a collection of `InputExample`s for the dev set.""" raise NotImplementedError() def get_test_examples(self, data_dir): """Gets a collection of `InputExample`s for prediction.""" raise NotImplementedError() def get_labels(self): """Gets the list of labels for this data set.""" raise NotImplementedError() ~~~ > 这是完全照搬class DataProcessor的类，只是类名改成MyDataProcessor **读取数据的类get_train_examples** ~~~python class MyDataProcessor(DataProcessor): """Base class for data converters for sequence classification data sets.""" def get_train_examples(self, data_dir): """Gets a collection of `InputExample`s for the train set.""" file_path = os.path.join(data_dir, 'train_sentiment.txt') f = open(file_path, 'r', encoding='utf-8') # 读取数据，并指定中文常用的utf-8 train_data = [] index = 0 # ID值 for line in f.readlines(): # 参考XnliProcessor guid = "train-%d" % index line = line.replace('\n', '').split('\t') # 处理换行符，原数据是以tab分割 text_a = tokenization.convert_to_unicode(str(line[1])) # 第0位置是索引，第1位置才是数据，可以查看train_sentiment.txt label = str(line[2]) # 我们的label里没有什么东西，只有数值，所以转字符串即可 train_data.append( InputExample(guid=guid, text_a=text_a, text_b=None, label=label)) # 这里我们没text_b，所以传入None index += 1 # index每次不一样，所以加等于1 return train_data # 这样数据就读取完成 ~~~ > 参照XnliProcessor ~~~python class XnliProcessor(DataProcessor): """Processor for the XNLI data set.""" def __init__(self): self.language = "zh" def get_train_examples(self, data_dir): """See base class.""" lines = self._read_tsv( os.path.join(data_dir, "multinli", "multinli.train.%s.tsv" % self.language)) examples = [] for (i, line) in enumerate(lines): if i == 0: continue guid = "train-%d" % (i) # 获取样本ID text_a = tokenization.convert_to_unicode(line[0]) text_b = tokenization.convert_to_unicode(line[1]) # 获取text_a和b，我们只有a所以把b去掉 label = tokenization.convert_to_unicode(line[2]) # 获取标签 if label == tokenization.convert_to_unicode("contradictory"): label = tokenization.convert_to_unicode("contradiction") examples.append( InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label)) # 把读进来的东西传到InputExample，这个类可以点进去，里面什么都没做，只不过是模板，我们也照着做 return examples ~~~ **获取label** ~~~ # 也是参考XnliProcessor，把return改成0，1，2即可 def get_labels(self): """Gets the list of labels for this data set.""" return ["0", "1", "2"] ~~~ **以下是完整的** ~~~python class MyDataProcessor(DataProcessor): """Base class for data converters for sequence classification data sets.""" def get_train_examples(self, data_dir): """Gets a collection of `InputExample`s for the train set.""" file_path = os.path.join(data_dir, 'train_sentiment.txt') f = open(file_path, 'r', encoding='utf-8') # 读取数据，并指定中文常用的utf-8 train_data = [] index = 0 # ID值 for line in f.readlines(): # 参考XnliProcessor guid = "train-%d" % index line = line.replace("\n", "").split("\t") # 处理换行符，原数据是以tab分割 text_a = tokenization.convert_to_unicode(str(line[1])) # 第0位置是索引，第1位置才是数据，可以查看train_sentiment.txt label = str(line[2]) # 我们的label里没有什么东西，只有数值，所以转字符串即可 train_data.append( InputExample(guid=guid, text_a=text_a, text_b=None, label=label)) # 这里我们没text_b，所以传入None index += 1 # index每次不一样，所以加等于1 return train_data # 这样数据就读取完成 def get_dev_examples(self, data_dir): """Gets a collection of `InputExample`s for the dev set.""" file_path = os.path.join(data_dir, 'test_sentiment.txt') f = open(file_path, 'r', encoding='utf-8') dev_data = [] index = 0 for line in f.readlines(): guid = "dev-%d" % index line = line.replace('\n', '').split('\t') text_a = tokenization.convert_to_unicode(str(line[1])) label = str(line[2]) dev_data.append( InputExample(guid=guid, text_a=text_a, text_b=None, label=label)) index += 1 return dev_data def get_test_examples(self, data_dir): """Gets a collection of `InputExample`s for prediction.""" file_path = os.path.join(data_dir, 'test_sentiment.txt') # 我们直接用验证集来输出结果 print(file_path) f = open(file_path, 'r', encoding='utf-8') test_data = [] index = 0 for line in f.readlines(): guid = "test-%d" % index line = line.replace('\n', '').split('\t') text_a = tokenization.convert_to_unicode(str(line[1])) label = '0' # 这里的label随机使用即可，只是为了传入 test_data.append( InputExample(guid=guid, text_a=text_a, text_b=None, label=label)) index += 1 return test_data def get_labels(self): """Gets the list of labels for this data set.""" return ["0", "1", "2"] # 参考XnliProcessor，改成返回0，1，2 ~~~ #### 训练BERT中文分类模型 main函数增加运行内容 ~~~python def main(_): tf.logging.set_verbosity(tf.logging.INFO) processors = { "cola": ColaProcessor, "mnli": MnliProcessor, "mrpc": MrpcProcessor, "xnli": XnliProcessor, 'my':MyDataProcessor, # 这是增加的部分，这样运行参数task_name才能对应上 } ~~~ 参数 ~~~ -task_name=my -do_train=true -do_eval=true -data_dir=data -vocab_file=../GLUE/BERT_BASE_DIR/chinese_L-12_H-768_A-12/vocab.txt -bert_config_file=../GLUE/BERT_BASE_DIR/chinese_L-12_H-768_A-12/bert_config.json -init_checkpoint=../GLUE/BERT_BASE_DIR/chinese_L-12_H-768_A-12/bert_model.ckpt -max_seq_length=70 -train_batch_size=32 -learning_rate=5e-5 --num_train_epochs=3.0 -output_dir=my_model ~~~ > task_name：运行的模块，在main里指定了名字对应的类 > > do_train：是否训练 > > do_eval：是否验证 > > data_dir：数据地址 > > vocab_file：词库表 > > bert_config_file：bert参数 > > init_checkpoint：初始化参数 > > max_seq_length：最长字符限制 > > train_batch_size：训练次数 > > learning_rate：学习率 > > num_train_epochs：循环训练次数 > > output_dir：输出路径 ![1610334445172](assets/1610334445172.png) 设置参数完成，run即可 ![1610334464000](assets/1610334464000.png) 最终模型结果 ![1610372590658](assets/1610372590658.png) ![1610376941475](assets/1610376941475.png) #### 预测结果并输出 进行预测的参数 ~~~python -task_name=my -do_predict=true -data_dir=data -vocab_file=../GLUE/BERT_BASE_DIR/chinese_L-12_H-768_A-12/vocab.txt -bert_config_file=../GLUE/BERT_BASE_DIR/chinese_L-12_H-768_A-12/bert_config.json -init_checkpoint=my_model -max_seq_length=70 -output_dir=my_model_predict ~~~ > init_checkpoint：使用的初始化参数已经是我们训练过的了 RUN完后有如下文件 ![1610377383372](assets/1610377383372.png) 打开与原文件对比，是准确的，不过现在是概率，我们转成值 ![1610378614228](assets/1610378614228.png) 添加get_results.py ~~~python import os import pandas as pd if __name__ == '__main__': path = "my_model_predict" pd_all = pd.read_csv(os.path.join(path, "test_results.tsv"), sep='\t', header=None) data = pd.DataFrame(columns=['polarity']) print(pd_all.shape) for index in pd_all.index: neutral_score = pd_all.loc[index].values[0] positive_score = pd_all.loc[index].values[1] negative_score = pd_all.loc[index].values[2] if max(neutral_score, positive_score, negative_score) == neutral_score: data.loc[index+1] = ["0"] elif max(neutral_score, positive_score, negative_score) == positive_score: data.loc[index+1] = ["1"] else: data.loc[index+1] = ["2"] data.to_csv(os.path.join(path, "pre_sample.tsv"), sep='\t') ~~~ 运行完后，同个目录下会出现pre_sample.tsv文件，对比结果 ![1610379269689](assets/1610379269689.png) > 正确 至此，我们完成了中文情感分类实战，写了函数训练、验证，并输出预测结果，BERT也算正式使用了起来，给在做的你点个赞👍。