04_用SimpleRNN鉴定评论文本

博文配套视频课程：自然语言处理与知识图谱

---

前言

前面的理论和原理介绍完了，接下来回到鉴定文本的情感属性这个案例。可以先把文本向量化，然后在通过Keras构建最简单循环神经网络SimpleRNN层。最后得出结论哪些客户是好评，哪些客户是差评。

读入评论数据集

该项目涉及到的代码和数据文件可以通过百度云盘进行下载，提取密码：lsqz。同学们也可以通过Kaggle网站通过关键字Product Comments 搜索该数据集进行下载。

import pandas as pd

df = pd.read_csv("../data/Clothing Reviews.csv")
df.head()

df['Review Text'] = df['Review Text'].astype(str)
x_train = df['Review Text'].values  # 训练集
y_train = df['Rating'].values

加载数据呈现如下：

用Tokenizer给文本分词

1. 评论句子已经被分解为单词
2. 每个单词已经被分配一个唯一的词典索引
3. x_train_tokenized 是目前列表类型数据

from tensorflow.keras.preprocessing.text import Tokenizer
# 创建字典索引,设定词典大小为2w
dict_size = 20000
tokenizer = Tokenizer(num_words=dict_size)
# jieba 停用词，标点符号, 词性....
tokenizer.fit_on_texts(x_train)  # 使用训练集创建词典索引
print(len(tokenizer.word_index))

# 把文本转化为序列编码 (特征工程)
x_train_tokenized = tokenizer.texts_to_sequences(x_train)
# list.shape ?
print(len(x_train_tokenized))
for v in x_train_tokenized[:50]:
    print(len(v),end="\t")

构建长度统一词向量

发现评论有长有短,但是后续我们要转化为向量必须是等长的，如何处理呢? 可以通过可视化(直方图)展示每个评论的长度。取一个均衡值例如120，长度超过120则截断，小于120，将填充为无意义的0值

# 通过可视化直方图展示每个评论的长度
import matplotlib.pyplot as plt  # 导入matplotlib
import numpy as np
word_per_comment = [len(comment) for comment in x_train_tokenized]
plt.hist(word_per_comment,bins=np.arange(0,200,10))
plt.show()

# 就是把不等长的list变成等长
from tensorflow.keras.preprocessing.sequence import pad_sequences
max_comment_length = 120 # 设定评论输入长度为120
x_train = pad_sequences(x_train_tokenized,maxlen=max_comment_length)

for v in x_train[:50]:
    print(len(v), end="\t")

构建RNN神经网络

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense
from tensorflow.keras.layers import SimpleRNN,Embedding
import tensorflow as tf

rnn = Sequential()
# (None, 120, 60) ==> (评论数,每条评论长度,词向量)  ==> (评论数,词向量)
# [batch,feature len] @ [feature len, hideen len] ==> [batch,hidden len]
# [batch,hidden len] @ [hidden len  hidden len] ==> [batch,hideen len]

#  [b,60] @ [60,100] ==> [b,100]  +  [b,100] @ [100,100] ==> [b,100]
rnn.add(SimpleRNN(units=100))  # 第二层构建了100个RNN神经元
# [b,100] @ [100,10] ==> [b,10]
rnn.add(Dense(units=10,activation=tf.nn.relu))
# [b,10] @ [10,5] ==> [b,5]
# 如果设置为5 则未来的数据也需要进行转化
rnn.add(Dense(units=5,activation=tf.nn.softmax))  # 输出分类的结果
rnn.compile(loss='sparse_categorical_crossentropy',optimizer="adam",metrics=['accuracy'])
print(rnn.summary())
# epochs = 1 测试随机数  ==> tf.random.set_seed(1)
# 修改为 14848 （与20000结果相同）
result = rnn.fit(x_train,y_train,validation_split=0.3,epochs=10,batch_size=64)

print(result)
print(result.history)

生成神经网络层如下：

Model: "sequential_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
embedding (Embedding)        (None, 120, 60)           1200000   
_________________________________________________________________
simple_rnn_1 (SimpleRNN)     (None, 100)               16100     
_________________________________________________________________
dense_2 (Dense)              (None, 10)                1010      
_________________________________________________________________
dense_3 (Dense)              (None, 5)                 55        
=================================================================
Total params: 1,217,165
Trainable params: 1,217,165
Non-trainable params: 0
_________________________________________________________________

完整代码如下

import pandas as pd

df = pd.read_csv("../data/Clothing Reviews.csv")
print(df.info())

df['Review Text'] = df['Review Text'].astype(str)
x_train = df['Review Text']
y_train = df['Rating']

from tensorflow.keras.preprocessing.text import Tokenizer
# 创建词典的索引,默认词典大小20000
dict_size = 20000
tokenizer = Tokenizer(num_words=dict_size)
# 词与索引映射关系
print(len(tokenizer.word_index),tokenizer.index_word)

# 把评论的文本转化序列编码
x_train_tokenized = tokenizer.texts_to_sequences(x_train)
print(x_train_tokenized)

# 通过指定长度,把不等长list转化为等长
from tensorflow.keras.preprocessing.sequence import pad_sequences
max_comment_length = 120
x_train = pad_sequences(x_train_tokenized,maxlen=max_comment_length)
for v in x_train[:10]:
    print(v,len(v))

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense
from tensorflow.keras.layers import SimpleRNN,Embedding
import tensorflow as tf

rnn = Sequential()
# 词嵌入操作学习到的结果为向量长度为60的密集矩阵
rnn.add(Embedding(input_dim = dict_size,output_dim = 60,input_length=max_comment_length))
rnn.add(SimpleRNN(units=100))  # 构建100 RNN循环神经元
rnn.add(Dense(10,activation=tf.nn.relu))  # 全连接层
rnn.add(Dense(5,activation=tf.nn.softmax))   # 全连接层输出分类的结果
rnn.compile(loss='sparse_categorical_crossentropy',optimizer='adam',metrics=['accuracy'])
print(rnn.summary())
# history = rnn.fit(x_train,y_train,validation_split=0.3,epochs=10,batch_size=64)