NLP基本任务二：基于深度学习的文本分类

本博客参照了复旦大学计算机科学技术学院邱锡鹏教授的文章https://www.zhihu.com/question/324189960

题目：熟悉Pytorch，用Pytorch重写《任务一》，实现CNN、RNN的文本分类；

1. 参考

   1. https://pytorch.org/
   2. Convolutional Neural Networks for Sentence Classification https://arxiv.org/abs/1408.5882
   3. https://machinelearningmastery.com/sequence-classification-lstm-recurrent-neural-networks-python-keras/

2. word embedding 的方式初始化

3. 随机embedding的初始化方式

4. 用glove 预训练的embedding进行初始化 https://nlp.stanford.edu/projects/glove/

5. 知识点：

   1. CNN/RNN的特征抽取
   2. 词嵌入
   3. Dropout

代码：

注：代码并没有严格参照要求去做。

import nltk
import pandas as pd
from sklearn.feature_extraction.text import CountVectorizer
import matplotlib.pylab as plt
%matplotlib inline

import os
os.environ['CUDA_VISIBLE_DEVICES'] = "0"   #设置GPU

#载入数据，注意参数delimiter='\t'
df_train = pd.read_csv(r'sentiment-analysis-on-movie-reviews/train.tsv',delimiter='\t')
df_test = pd.read_csv(r'sentiment-analysis-on-movie-reviews/test.tsv',delimiter='\t')
df_train.head()

#创建transformer，CountVectorizer是属于常见的特征数值计算类，是一个文本特征提取方法。
word_vectorizer = CountVectorizer(ngram_range = (1,1),analyzer = 'word',stop_words = 'english',min_df = 0.001)  #使用默认的英语停用词表
spare_matric = word_vectorizer.fit_transform(df_train['Phrase'])  #编码文件，将文本转化为稀疏矩阵

# print(spare_matric)
#输出如：
# (0, 480)	1
#   (0, 352)	1
#   (0, 222)	2
#   (0, 451)	1
#   (1, 222)	1
#   (1, 451)	1
#   (2, 451)	1
# print(sum(spare_matric))  
#按列求和，输出如：
# (0, 570)	161
# (0, 28)	213

#统计文本单词出现的频率
# print(spare_matric.shape)  #(156060, 587)
frequency = sum(spare_matric).toarray()[0]  #toarray后如[[ 179  204  176  ]]，所以要加[0]
# print(len(frequency))  #587
# print(frequency)
freq = pd.DataFrame(frequency,index = word_vectorizer.get_feature_names(),columns = ['frequency'])
freq.sort_values('frequency',ascending = False)

#观察标签的分布情况,发现稍微符合正态分布，若不符合正态分类，可以尝试通过log等函数进行转化来使其符合正态分布。
a = df_train.Sentiment.value_counts()  #统计一列中不同种类各有多少个
# a.plot(kind = 'bar')  #这种图默认按大小排序
# print(a.index)
# print(a.values)
plt.bar(a.index,a.values)

#更加好看的图
# a = pd.DataFrame(a)
# a['Rating'] = a.index
# sns.set_style("darkgrid", {"axes.facecolor": ".9"})
# fig, ax = plt.subplots(figsize=(10,6))
# sns.barplot(y='Sentiment', x='Rating', data=a)

#对文本进行预处理
import re
df_train['Phrase'] = df_train['Phrase'].str.lower()
df_train['Phrase'] = df_train['Phrase'].apply(lambda x: re.sub('[^a-zA-Z0-9\s]','',x))
df_test['Phrase'] = df_test['Phrase'].str.lower()
df_test['Phrase'] = df_test['Phrase'].apply(lambda x: re.sub('[^a-zA-Z0-9]','',x))
# print(df_train['Phrase'])

X_train = df_train.Phrase
y_train = df_train.Sentiment

#构造字典和训练数据
from keras.preprocessing.text import Tokenizer  #参考https://blog.csdn.net/lovebyz/article/details/77712003
tokenizer = Tokenizer()
# print(X_train)
#输出如下：
# 0         a series of escapades demonstrating the adage ...
# 1         a series of escapades demonstrating the adage ...
# 2                                                  a series
# 3                                                         a
# 4                                                    series
tokenizer.fit_on_texts(X_train.values)  #使用一系列文档来生成token词典，每个元素为一个文档。

X_train = tokenizer.texts_to_sequences(X_train)  #将多个文档转换为word下标的向量形式
# print(len(X_train))  #156060
# print(X_train[0])  #[2, 304, 3, 15110, 5906, 1, 6499, 9, 51, 8, 49, 13, 1, 3514, 8, 167, 49, 13, 1, 11381, 62, 3, 75, 615, 10453, 19, 576, 3, 75, 2003, 5, 54, 3, 2, 40]
# print(len(X_train[0]))  #35
# print(len(X_train[1]))  #14
# print(len(X_train[2]))  #2
X_test = df_test.Phrase
X_test = tokenizer.texts_to_sequences(X_test)

#将数据集统一长度，一般取最大长度
from keras.preprocessing.sequence import pad_sequences
max_length = max([len(x.split()) for x in df_train['Phrase']])
# print(max_length)  #48
X_train = pad_sequences(X_train,max_length)
X_test = pad_sequences(X_test,max_length)
# print(X_train.shape)  #(156060, 48)
# print(X_test.shape)  #(66292, 48)

#构建深度学习模型
from keras import Sequential
from keras.layers import Embedding,LSTM,Dense

EMBEDDING_DIM = 128
dict_len = len(tokenizer.word_index) + 1
model = Sequential()
model.add(Embedding(dict_len,EMBEDDING_DIM,input_length = max_length))  #参数https://blog.csdn.net/jiangpeng59/article/details/77533309
model.add(LSTM(units = 128,dropout = 0.2,recurrent_dropout = 0.2))  #第一个dropout是x和hidden之间的dropout,第二个是hidden-hidden之间的dropout
model.add(Dense(5,activation = 'softmax'))
model.compile(loss = 'sparse_categorical_crossentropy',optimizer= 'adam',metrics= ['accuracy'])
# print(model.summary())
# Layer (type)                 Output Shape              Param #   
# =================================================================
# embedding_4 (Embedding)      (None, 48, 128)           2099712   
# _________________________________________________________________
# lstm_3 (LSTM)                (None, 128)               131584    
# _________________________________________________________________
# dense_3 (Dense)              (None, 5)                 645       
# =================================================================
# Total params: 2,231,941
# Trainable params: 2,231,941
# Non-trainable params: 0
# _________________________________________________________________
# None

model.fit(X_train,y_train,batch_size= 128,epochs= 7,verbose= 1)
# Epoch 6/7
# 156060/156060 [==============================] - 101s 650us/step - loss: 0.5748 - acc: 0.7544
# Epoch 7/7
# 156060/156060 [==============================] - 101s 644us/step - loss: 0.5448 - acc: 0.7645

#模型预测，提交结果
y_test_pred = model.predict_classes(X_test)
final_pred = pd.read_csv(r'sentiment-analysis-on-movie-reviews/sampleSubmission.csv', sep=',')
final_pred.Sentiment=final_pred
final_pred.to_csv(r'results.csv', sep=',', index=False)

#使用CNN
from keras.layers import Conv1D,Dropout,MaxPooling1D,Flatten
def build_model():
    model = Sequential()
    model.add(Embedding(dict_len,output_dim=32,input_length = max_length))  
    model.add(Conv1D(filters = 32,kernel_size = 3,padding='same',activation='relu'))  
    model.add(MaxPooling1D(pool_size=2))
    model.add(Dropout(0.2))
    model.add(Flatten())
    model.add(Dense(5,activation = 'softmax'))
    model.compile(loss = 'sparse_categorical_crossentropy',optimizer= 'adam',metrics= ['accuracy'])
    model.fit(X_train,y_train,batch_size= 128,epochs= 7,verbose= 1)
    return model

model2 = build_model()
# Epoch 6/7
# 156060/156060 [==============================] - 7s 45us/step - loss: 0.6345 - acc: 0.7340
# Epoch 7/7
# 156060/156060 [==============================] - 7s 43us/step - loss: 0.6068 - acc: 0.7462