用Bert实现中文文本分类的代码

以下是用BERT实现中文文本分类的代码，包括数据预处理、模型构建和训练等步骤：

1. 数据预处理

import pandas as pd
import numpy as np
import tensorflow as tf
import os
from sklearn.model_selection import train_test_split
from transformers import BertTokenizer

# 加载数据
df = pd.read_csv('data.csv')

# 标签映射
labels = df['label'].unique()
label2id = {label: index for index, label in enumerate(labels)}
id2label = {index: label for index, label in enumerate(labels)}

# 文本处理
tokenizer = BertTokenizer.from_pretrained('bert-base-chinese')
max_length = 128

def create_inputs_targets(data):
    input_ids = []
    attention_masks = []
    targets = []
    for _, row in data.iterrows():
        text = row['text']
        label = row['label']
        # 分词
        encoded_dict = tokenizer.encode_plus(text,
                                              add_special_tokens=True,
                                              max_length=max_length,
                                              pad_to_max_length=True,
                                              return_attention_mask=True,
                                              return_tensors='tf')
        input_ids.append(encoded_dict['input_ids'])
        attention_masks.append(encoded_dict['attention_mask'])
        targets.append(label2id[label])
    return np.array(input_ids), np.array(attention_masks), np.array(targets)

# 划分数据集
train_data, test_data = train_test_split(df, test_size=0.2, random_state=42)

train_inputs, train_masks, train_targets = create_inputs_targets(train_data)
test_inputs, test_masks, test_targets = create_inputs_targets(test_data)

2. 模型构建

from transformers import TFBertForSequenceClassification

# 加载预训练模型
model = TFBertForSequenceClassification.from_pretrained('bert-base-chinese', num_labels=len(labels))

# 定义优化器和损失函数
optimizer = tf.keras.optimizers.Adam(learning_rate=1e-5)
loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)

# 编译模型
model.compile(optimizer=optimizer, loss=loss, metrics=['accuracy'])

3. 模型训练

batch_size = 32
epochs = 3

# 训练模型
history = model.fit([train_inputs, train_masks], train_targets, batch_size=batch_size, epochs=epochs, validation_split=0.1)

4. 模型评估

test_loss, test_accuracy = model.evaluate([test_inputs, test_masks], test_targets, batch_size=batch_size)

print('Test Loss:', test_loss)
print('Test Accuracy:', test_accuracy)

以上是用BERT实现中文文本分类的完整代码，其中数据集需自行准备。