参考 基于深度学习的自然语言处理

1. 读取数据

数据文件：

import numpy as np
import pandas as pddata = pd.read_csv("yelp_labelled.txt", sep='\t', names=['sentence', 'label'])data.head() # 1000条数据

# 数据 X 和 标签 y
sentence = data['sentence'].values
label = data['label'].values

2. 数据集拆分

# 训练集 测试集拆分
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(sentence, label, test_size=0.3, random_state=1)

3. 文本向量化

• 训练 tokenizer，文本转成 ids 序列

# 文本向量化
import keras
from keras.preprocessing.text import Tokenizer
tokenizer = Tokenizer(num_words=6000)
tokenizer.fit_on_texts(X_train) # 训练tokenizer
X_train = tokenizer.texts_to_sequences(X_train) # 转成 [[ids...],[ids...],...]
X_test = tokenizer.texts_to_sequences(X_test)
vocab_size = len(tokenizer.word_index)+1 # +1 是因为index 0, 0 不对应任何词，用来pad

• pad ids 序列，使之有相同的长度

maxlen = 100
# pad 保证每个句子的长度相等
from keras.preprocessing.sequence import pad_sequences
X_train = pad_sequences(X_train, maxlen=maxlen, padding='post')
# post 尾部补0，pre 前部补0
X_test = pad_sequences(X_test, maxlen=maxlen, padding='post')

4. 建立CNN模型

from keras import layers
embeddings_dim = 150
filters = 64
kernel_size = 5
batch_size = 64nn_model = keras.Sequential()
nn_model.add(layers.Embedding(input_dim=vocab_size, output_dim=embeddings_dim, input_length=maxlen))
nn_model.add(layers.Conv1D(filters=filters,kernel_size=kernel_size,activation='relu'))
nn_model.add(layers.GlobalMaxPool1D())
nn_model.add(layers.Dropout(0.3))
# 上面 GlobalMaxPool1D 后，维度少了一维，下面自定义layers再扩展一维
nn_model.add(layers.Lambda(lambda x : keras.backend.expand_dims(x, axis=-1)))
nn_model.add(layers.Conv1D(filters=filters,kernel_size=kernel_size,activation='relu'))
nn_model.add(layers.GlobalMaxPool1D())
nn_model.add(layers.Dropout(0.3))
nn_model.add(layers.Dense(10, activation='relu'))
nn_model.add(layers.Dense(1, activation='sigmoid')) # 二分类sigmoid, 多分类 softmax

参考文章：
Embedding层详解
Keras: GlobalMaxPooling vs. MaxPooling

• 配置模型

nn_model.compile(optimizer='adam', loss='binary_crossentropy',metrics=['accuracy'])
nn_model.summary()
from keras.utils import plot_model
plot_model(nn_model, to_file='model.jpg') # 绘制模型结构到文件

Model: "sequential_4"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
embedding_4 (Embedding)      (None, 100, 150)          251400    
_________________________________________________________________
conv1d_8 (Conv1D)            (None, 96, 64)            48064     
_________________________________________________________________
global_max_pooling1d_7 (Glob (None, 64)                0         
_________________________________________________________________
dropout_7 (Dropout)          (None, 64)                0         
_________________________________________________________________
lambda_4 (Lambda)            (None, 64, 1)             0         
_________________________________________________________________
conv1d_9 (Conv1D)            (None, 60, 64)            384       
_________________________________________________________________
global_max_pooling1d_8 (Glob (None, 64)                0         
_________________________________________________________________
dropout_8 (Dropout)          (None, 64)                0         
_________________________________________________________________
dense_6 (Dense)              (None, 10)                650       
_________________________________________________________________
dense_7 (Dense)              (None, 1)                 11        
=================================================================
Total params: 300,509
Trainable params: 300,509
Non-trainable params: 0

5. 训练、测试

history = nn_model.fit(X_train,y_train,batch_size=batch_size,epochs=50,verbose=2,validation_data=(X_test,y_test))
# verbose 是否显示日志信息，0不显示，1显示进度条，2不显示进度条
loss, accuracy = nn_model.evaluate(X_train, y_train, verbose=1)
print("训练集：loss {0:.3f}, 准确率：{1:.3f}".format(loss, accuracy))
loss, accuracy = nn_model.evaluate(X_test, y_test, verbose=1)
print("测试集：loss {0:.3f}, 准确率：{1:.3f}".format(loss, accuracy))# 绘制训练曲线
from matplotlib import pyplot as plt
pd.DataFrame(history.history).plot(figsize=(8, 5))
plt.grid(True)
plt.gca().set_ylim(0, 1) # set the vertical range to [0-1]
plt.show()

输出：

Epoch 1/50
11/11 - 1s - loss: 0.6933 - accuracy: 0.5014 - val_loss: 0.6933 - val_accuracy: 0.4633
Epoch 2/50
11/11 - 0s - loss: 0.6931 - accuracy: 0.5214 - val_loss: 0.6935 - val_accuracy: 0.4633
Epoch 3/50
11/11 - 1s - loss: 0.6930 - accuracy: 0.5257 - val_loss: 0.6936 - val_accuracy: 0.4633
....省略
11/11 - 0s - loss: 0.0024 - accuracy: 1.0000 - val_loss: 0.7943 - val_accuracy: 0.7600
Epoch 49/50
11/11 - 1s - loss: 0.0016 - accuracy: 1.0000 - val_loss: 0.7970 - val_accuracy: 0.7600
Epoch 50/50
11/11 - 0s - loss: 0.0027 - accuracy: 1.0000 - val_loss: 0.7994 - val_accuracy: 0.7600
22/22 [==============================] - 0s 4ms/step - loss: 9.0586e-04 - accuracy: 1.0000
训练集：loss 0.001, 准确率：1.000
10/10 [==============================] - 0s 5ms/step - loss: 0.7994 - accuracy: 0.7600
测试集：loss 0.799, 准确率：0.760

训练集：loss 0.001, 准确率：1.000
测试集：loss 0.799, 准确率：0.760
存在过拟合，训练集准确率很高，测试集效果差

• 随意测试

text = ["i am not very good.", "i am very good."]
x = tokenizer.texts_to_sequences(text)
x = pad_sequences(x, maxlen=maxlen, padding='post')
pred = nn_model.predict(x)
print("预测{}的类别为：".format(text[0]), 1 if pred[0][0]>=0.5 else 0)
print("预测{}的类别为：".format(text[1]), 1 if pred[1][0]>=0.5 else 0)

输出：

预测i am not very good.的类别为： 0
预测i am very good.的类别为： 1