#!/usr/bin/env python
# coding: utf-8
# # —— 基于电商产品评论数据情感分析 ——
# ### 1.案例简介
#
# 1、利用文本挖掘技术,对碎片化、非结构化的电商网站评论数据进行清洗与处理,转化为结构化数据。
# 2、参考知网发布的情感分析用词语集,统计评论数据的正负情感指数,然后进行情感分析,通过词云图直观查看正负评论的关键词。
# 3、比较“机器挖掘的正负情感”与“人工打标签的正负情感”,精度达到88%。
# 4、采用LDA主题模型提取评论关键信息,以了解用户的需求、意见、购买原因、产品的优缺点等。
#
# ### 2.框架
#
# 工具准备
#
# 一、导入数据
# 二、数据预处理
# (一)去重
# (二)数据清洗
# (三)分词、词性标注、去除停用词、词云图
# 三、模型构建
# (一)决策树
# (二)情感分析
# (三)基于LDA模型的主题分析
# ## 工具准备
# In[ ]:
import os
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
from matplotlib.pylab import style #自定义图表风格
style.use('ggplot')
from IPython.core.interactiveshell import InteractiveShell
InteractiveShell.ast_node_interactivity = "all"
plt.rcParams['font.sans-serif'] = ['Simhei'] # 解决中文乱码问题
import re
import jieba.posseg as psg
import itertools
#conda install -c anaconda gensim
from gensim import corpora,models #主题挖掘,提取关键信息
# pip install wordcloud
from wordcloud import WordCloud,ImageColorGenerator
from collections import Counter
from sklearn import tree
from sklearn.model_selection import train_test_split
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.metrics import classification_report
from sklearn.metrics import accuracy_score
import graphviz
# #### 注意
#
# 以下方法,是为了帮助我们直观查看对象处理的结果。是辅助代码,非必要代码!
# .head()
# print()
# len()
# .shape
# .unique()
# ## 一、导入数据
# In[ ]:
raw_data=pd.read_csv('data/reviews.csv')
raw_data.head()
# In[ ]:
raw_data.info()
# In[ ]:
raw_data.columns
# In[ ]:
#取值分布
for cate in ['creationTime', 'nickname', 'referenceName', 'content_type']:
raw_data[cate].value_counts()
# ## 二、数据预处理
# ### (一)去重
#
# 删除系统自动为客户做出的评论。
# In[ ]:
reviews=raw_data.copy()
reviews=reviews[['content', 'content_type']]
print('去重之前:',reviews.shape[0])
reviews=reviews.drop_duplicates()
print('去重之后:',reviews.shape[0])
# ### (二)数据清洗
# In[ ]:
# 清洗之前
content=reviews['content']
for i in range(5,10):
print(content[i])
print('-----------')
# In[ ]:
#清洗之后,将数字、字母、京东美的电热水器字样都删除
info=re.compile('[0-9a-zA-Z]|京东|美的|电热水器|热水器|')
content=content.apply(lambda x: info.sub('',x)) #替换所有匹配项
for i in range(5,10):
print(content[i])
print('-----------')
# ### (三)分词、词性标注、去除停用词、词云图
# (1)分词
# 目标
#
# 输入:
# - content、content_type
# - 共有1974条评论句子
# 输出:
# - 构造DF,包含: 分词、对应词性、分词所在原句子的id、分词所在原句子的content_type
# - 共有6万多行
#
# 非结构化数据——>结构化数据
# 
# In[ ]:
#分词,由元组组成的list
seg_content=content.apply( lambda s: [(x.word,x.flag) for x in psg.cut(s)] )
seg_content.shape
len(seg_content)
print(seg_content[0])
# In[ ]:
#统计评论词数
n_word=seg_content.apply(lambda s: len(s))
len(n_word)
n_word.head(6)
# In[ ]:
#得到各分词在第几条评论
n_content=[ [x+1]*y for x,y in zip(list(seg_content.index),list(n_word))] #[x+1]*y,表示复制y份,由list组成的list
# n_content=[ [x+1]*y for x,y in [(0,32)]]
index_content_long=sum(n_content,[]) #表示去掉[],拉平,返回list
len(index_content_long)
# n_content
# index_content_long
# list(zip(list(seg_content.index),list(n_word)))
# In[ ]:
sum([[2,2],[3,3,3]],[])
# In[ ]:
#分词及词性,去掉[],拉平
seg_content.head()
seg_content_long=sum(seg_content,[])
seg_content_long
type(seg_content_long)
len(seg_content_long)
# In[ ]:
seg_content_long[0]
# In[ ]:
#得到加长版的分词、词性
word_long=[x[0] for x in seg_content_long]
nature_long=[x[1] for x in seg_content_long]
len(word_long)
len(nature_long)
# In[ ]:
#content_type拉长
n_content_type=[ [x]*y for x,y in zip(list(reviews['content_type']),list(n_word))] #[x+1]*y,表示复制y份
content_type_long=sum(n_content_type,[]) #表示去掉[],拉平
# n_content_type
# content_type_long
len(content_type_long)
# In[ ]:
review_long=pd.DataFrame({'index_content':index_content_long,
'word':word_long,
'nature':nature_long,
'content_type':content_type_long})
review_long.shape
review_long.head()
# (2)去除标点符号、去除停用词
# In[ ]:
review_long['nature'].unique()
# In[ ]:
#去除标点符号
review_long_clean=review_long[review_long['nature']!='x'] #x表示标点符合
review_long_clean.shape
# In[ ]:
#导入停用词
stop_path=open('data/stoplist.txt','r',encoding='UTF-8')
stop_words=stop_path.readlines()
len(stop_words)
stop_words[0:5]
# In[ ]:
#停用词,预处理
stop_words=[word.strip('\n') for word in stop_words]
stop_words[0:5]
# In[ ]:
#得到不含停用词的分词表
word_long_clean=list(set(word_long)-set(stop_words))
len(word_long_clean)
review_long_clean=review_long_clean[review_long_clean['word'].isin(word_long_clean)]
review_long_clean.shape
# (3)在原df中,再增加一列,该分词在本条评论的位置
# In[ ]:
#再次统计每条评论的分词数量
n_word=review_long_clean.groupby('index_content').count()['word']
n_word
index_word=[ list(np.arange(1,x+1)) for x in list(n_word)]
index_word_long=sum(index_word,[]) #表示去掉[],拉平
len(index_word_long)
# In[ ]:
review_long_clean['index_word']=index_word_long
review_long_clean.head()
# In[ ]:
review_long_clean.to_csv('data/1_review_long_clean.csv')
# (4)提取名词
# In[ ]:
n_review_long_clean=review_long_clean[[ 'n' in nat for nat in review_long_clean.nature]]
n_review_long_clean.shape
n_review_long_clean.head()
# In[ ]:
n_review_long_clean.nature.value_counts()
n_review_long_clean.to_csv('data/1_n_review_long_clean.csv')
# (5)词云图
# In[ ]:
#txt = "life is short, you need python"
#review_long_clean.word.values.dtype
#print("okok+"+review_long_clean.word.values.dtype)
##print("okok+"+review_long_clean.word.values)
#wordcloud.generate(Counter(review_long_clean.word.values))
#wordcloud.generate_from_frequencies(Counter(review_long_clean.word.values))
#wordcloud.to_file('data/ct.png')
# In[ ]:
#font=r"C:\Windows\Fonts\msyh.ttc"
font=r"C:\Windows\Fonts\msyh.ttc"
background_image=plt.imread('data/1.png')
background_image = background_image.astype(np.uint8)
wordcloud = WordCloud(font_path=font, max_words = 100, mode='RGBA' ,background_color='white',mask=background_image)
#wordcloud = WordCloud(font_path=font, max_words = 100, background_color='white',mask=background_image) #width=1600,height=1200, mode='RGBA'
#wordcloud = WordCloud(font_path=None,max_words = 100, background_color='white', mode='RGBA')
#print("okok+"+review_long_clean.word.values)
wordcloud.generate_from_frequencies(Counter(review_long_clean.word.values))
wordcloud.to_file('data/1_分词后的词云图.png')
plt.figure(figsize=(20,10))
plt.imshow(wordcloud)
plt.axis('off')
plt.show()
# In[ ]:
font=r"C:\Windows\Fonts\msyh.ttc"
background_image=plt.imread('data/1.png')
background_image = background_image.astype(np.uint8)
wordcloud = WordCloud(font_path=font, max_words = 100, mode='RGBA' ,background_color='white',mask=background_image) #width=1600,height=1200
wordcloud.generate_from_frequencies(Counter(n_review_long_clean.word.values))
wordcloud.to_file('1_分词后的词云图(名词).png')
plt.figure(figsize=(20,10))
plt.imshow(wordcloud)
plt.axis('off')
plt.show()
# ## 三、模型构建
# ### (一)基于决策树的情感分类
# In[ ]:
#第一步:构造特征空间和标签
Y=[]
for ind in review_long_clean.index_content.unique():
y=[ word for word in review_long_clean.content_type[review_long_clean.index_content==ind].unique() ]
Y.append(y)
len(Y)
X=[]
for ind in review_long_clean.index_content.unique():
term=[ word for word in review_long_clean.word[review_long_clean.index_content==ind].values ]
X.append(' '.join(term))
len(X)
X
Y
# In[ ]:
#第二步:训练集、测试集划分
x_train,x_test,y_train,y_test=train_test_split(X,Y,test_size=0.2,random_state=7)
#第三步:词转向量,01矩阵
count_vec=CountVectorizer(binary=True)
x_train=count_vec.fit_transform(x_train)
x_test=count_vec.transform(x_test)
#第四步:构建决策树
dtc=tree.DecisionTreeClassifier(max_depth=5)
dtc.fit(x_train,y_train)
print('在训练集上的准确率:%.2f'% accuracy_score(y_train,dtc.predict(x_train)))
y_true=y_test
y_pred=dtc.predict(x_test)
print(classification_report(y_true,y_pred))
print('在测试集上的准确率:%.2f'% accuracy_score(y_true,y_pred))
# In[ ]:
#第五步:画决策树
#这个决策数图,需要在网上下载graphviz 包,并且配置环境变量
from graphviz import Digraph
import os
#C:\\Program Files\\Graphviz\\bin
#C:\\Users\\huangcong\\Desktop\\202203教案\\分享资料3\\1.png
os.environ["PATH"] += os.pathsep + 'D:\\software\\Graphviz\\bin'
dot_data=tree.export_graphviz(dtc ,feature_names=count_vec.get_feature_names())
graph=graphviz.Source(dot_data)
graph
# ### (二)情感分析
# 数据预处理的思路与目标
#
# 
# (1)导入评价情感词
# In[ ]:
#来自知网发布的情感分析用词语集
#raw_data=pd.read_csv('C:\\Users\\huangcong\\Desktop\\202203教案\\分享资料3\\reviews.csv') 注意各路径 \ /
#务必这种写法'C:\\Users\\huangcong\\Desktop\\202203教案\\分享资料3\\正面评价词语(中文).txt',header=None,sep='/n',engine='python'
pos_comment=pd.read_csv('data/正面评价词语(中文).txt',header=None,sep='/n',engine='python')
neg_comment=pd.read_csv('data/负面评价词语(中文).txt',header=None,sep='/n',engine='python')
pos_emotion=pd.read_csv('data/正面情感词语(中文).txt',header=None,sep='/n',engine='python')
neg_emotion=pd.read_csv('data/负面情感词语(中文).txt',header=None,sep='/n',engine='python')
# In[ ]:
pos_comment.shape
neg_comment.shape
pos_emotion.shape
neg_emotion.shape
# In[ ]:
pos=pd.concat([pos_comment,pos_emotion],axis=0)
pos.shape
neg=pd.concat([neg_comment,neg_emotion],axis=0)
neg.shape
# (2)增加新词
# In[ ]:
c='点赞'
c in pos.values
d='歇菜'
d in neg.values
# In[ ]:
new_pos=pd.Series(['点赞'])
new_neg=pd.Series(['歇菜'])
positive=pd.concat([pos,new_pos],axis=0)
positive.shape
negative=pd.concat([neg,new_neg],axis=0)
negative.shape
# In[ ]:
positive.columns=['review']
positive['weight']=pd.Series([1]*len(positive))
positive.head()
# In[ ]:
negative.columns=['review']
negative['weight']=pd.Series([-1]*len(negative))
negative.head()
# In[ ]:
pos_neg=pd.concat([positive,negative],axis=0)
pos_neg.shape
# (3)合并到review_long_clean中
# In[ ]:
#表联接
data=review_long_clean.copy()
review_mltype=pd.merge(data,pos_neg,how='left',left_on='word',right_on='review')
review_mltype.shape
review_mltype=review_mltype.drop(['review'],axis=1)
review_mltype=review_mltype.replace(np.nan,0)
review_mltype.head()
# (4)修正情感倾向,
#
# 如有多重否定,那么奇数否定是否定,偶数否定是肯定
#
# 看该情感词前2个词,来判罚否定的语气。如果在句首,则没有否词,如果在句子的第二次词,则看前1个词,来判断否定的语气。
# In[ ]:
notdict=pd.read_csv('data/not.csv')
notdict.shape
notdict['freq']=[1]*len(notdict)
notdict.head()
# In[ ]:
#准备一
review_mltype['amend_weight']=review_mltype['weight']
review_mltype['id']=np.arange(0,review_mltype.shape[0])
review_mltype.head()
# In[ ]:
# 准备二,只保留有情感值的行
only_review_mltype=review_mltype[review_mltype['weight']!=0]
only_review_mltype.index=np.arange(0,only_review_mltype.shape[0]) #索引重置
only_review_mltype.shape
only_review_mltype.head()
# In[ ]:
i=4
review_i=review_mltype[review_mltype['index_content']==only_review_mltype['index_content'][i]]
review_i#第i个情感词的评论
# In[ ]:
#看该情感词前2个词,来判罚否定的语气。如果在句首,则没有否词,如果在句子的第二次词,则看前1个词,来判断否定的语气。
index=only_review_mltype['id']
for i in range(0,only_review_mltype.shape[0]):
review_i=review_mltype[review_mltype['index_content']==only_review_mltype['index_content'][i]] #第i个情感词的评论
review_i.index=np.arange(0,review_i.shape[0])#重置索引后,索引值等价于index_word
word_ind = only_review_mltype['index_word'][i] #第i个情感值在该条评论的位置
#第一种,在句首。则不用判断
#第二种,在评论的第2个为位置
if word_ind==2:
ne=sum( [ review_i['word'][word_ind-1] in notdict['term'] ] )
if ne==1:
review_mltype['amend_weight'][index[i]] = -( review_mltype['weight'][index[i]] )
#第三种,在评论的第2个位置以后
elif word_ind > 2:
ne=sum( [ word in notdict['term'] for word in review_i['word'][[word_ind-1,word_ind-2]] ] ) # 注意用中括号[word_ind-1,word_ind-2]
if ne==1:
review_mltype['amend_weight'][index[i]]=- ( review_mltype['weight'][index[i]] )
# In[ ]:
review_mltype.shape
review_mltype[(review_mltype['weight']-review_mltype['amend_weight'])!=0] #说明两列值一样
# (5)计算每条评论的情感值
# In[ ]:
review_mltype.tail()
# In[ ]:
emotion_value=review_mltype.groupby('index_content',as_index=False)['amend_weight'].sum()
emotion_value.head()
emotion_value.to_csv('./1_emotion_value',index=True,header=True)
# (6)查看情感分析效果
# In[ ]:
#每条评论的amend_weight总和不等于零
content_emotion_value=emotion_value.copy()
content_emotion_value.shape
content_emotion_value=content_emotion_value[content_emotion_value['amend_weight']!=0]
content_emotion_value['ml_type']=''
content_emotion_value['ml_type'][content_emotion_value['amend_weight']>0]='pos'
content_emotion_value['ml_type'][content_emotion_value['amend_weight']<0]='neg'
content_emotion_value.shape
content_emotion_value.head()
# In[ ]:
#每条评论的amend_weight总和等于零
#这个方法其实不好用,有一半以上的评论区分不出正、负情感。
content_emotion_value0=emotion_value.copy()
content_emotion_value0=content_emotion_value0[content_emotion_value0['amend_weight']==0]
content_emotion_value0.head()
raw_data.content[20]
raw_data.content[21]
raw_data.content[26]
# In[ ]:
#合并到大表中
content_emotion_value=content_emotion_value.drop(['amend_weight'],axis=1)
review_mltype.shape
review_mltype=pd.merge(review_mltype,content_emotion_value,how='left',left_on='index_content',right_on='index_content')
review_mltype=review_mltype.drop(['id'],axis=1)
review_mltype.shape
review_mltype.head()
review_mltype.to_csv('./1_review_mltype',index=True,header=True)
# In[ ]:
cate=['index_content','content_type','ml_type']
data_type=review_mltype[cate].drop_duplicates()
confusion_matrix=pd.crosstab(data_type['content_type'],data_type['ml_type'],margins=True)
confusion_matrix
# In[ ]:
data=data_type[['content_type','ml_type']]
data=data.dropna(axis=0)
print( classification_report(data['content_type'],data['ml_type']) )
# (7)制作词云图
# - 只看情感词
# In[ ]:
data=review_mltype.copy()
data=data[data['amend_weight']!=0]
word_data_pos=data[data['ml_type']=='pos']
word_data_neg=data[data['ml_type']=='neg']
# In[ ]:
#按照以上修改,显示信息
font=r"C:\Windows\Fonts\msyh.ttc"
background_image=plt.imread('data/1.png')
wordcloud = WordCloud(font_path=font, max_words = 100, mode='RGBA' ,background_color='white',mask=background_image) #width=1600,height=1200
#wordcloud = WordCloud(max_words = 100, mode='RGBA' ,background_color='white') #width=1600,height=1200
wordcloud.generate_from_frequencies(Counter(word_data_pos.word.values))
plt.figure(figsize=(15,7))
plt.imshow(wordcloud)
plt.axis('off')
plt.show()
# In[ ]:
#font=r"C:\Windows\Fonts\msyh.ttc"
font=r"C:\Windows\Fonts\msyh.ttc"
background_image=plt.imread('data/1.png')
#background_image=plt.imread('./p6sad.jpg')
wordcloud = WordCloud(font_path=font, max_words = 100, mode='RGBA' ,background_color='white',mask=background_image) #width=1600,height=1200
#wordcloud = WordCloud(max_words = 100, mode='RGBA' ,background_color='white') #width=1600,height=1200
wordcloud.generate_from_frequencies(Counter(word_data_neg.word.values))
plt.figure(figsize=(15,7))
plt.imshow(wordcloud)
plt.axis('off')
plt.show()
# - 看所有词
# In[ ]:
data=review_mltype.copy()
word_data_pos=data[data['ml_type']=='pos']
word_data_neg=data[data['ml_type']=='neg']
font=r"C:\Windows\Fonts\msyh.ttc"
background_image=plt.imread('data/1.png')
wordcloud = WordCloud(font_path=font, max_words = 100, mode='RGBA' ,background_color='white',mask=background_image) #width=1600,height=1200
wordcloud.generate_from_frequencies(Counter(word_data_pos.word.values))
plt.figure(figsize=(15,7))
plt.imshow(wordcloud)
plt.axis('off')
plt.show()
background_image=plt.imread('data/1.png')
wordcloud = WordCloud(font_path=font, max_words = 100, mode='RGBA' ,background_color='white',mask=background_image) #width=1600,height=1200
wordcloud.generate_from_frequencies(Counter(word_data_neg.word.values))
plt.figure(figsize=(15,7))
plt.imshow(wordcloud)
plt.axis('off')
plt.show()
# ### (三)基于LDA模型的主题分析
#
# 优点:不需要人工调试,用相对少的迭代找到最优的主题结构。
# (1)建立词典、语料库
# In[ ]:
data=review_mltype.copy()
word_data_pos=data[data['ml_type']=='pos']
word_data_neg=data[data['ml_type']=='neg']
# In[ ]:
#建立词典,去重
pos_dict=corpora.Dictionary([ [i] for i in word_data_pos.word]) #shape=(n,1)
neg_dict=corpora.Dictionary([ [i] for i in word_data_neg.word])
# In[ ]:
print(pos_dict)
# In[ ]:
#建立语料库
pos_corpus=[ pos_dict.doc2bow(j) for j in [ [i] for i in word_data_pos.word] ] #shape=(n,(2,1))
neg_corpus=[ neg_dict.doc2bow(j) for j in [ [i] for i in word_data_neg.word] ]
# In[ ]:
len(word_data_pos.word)
len(pos_dict)
len(pos_corpus)
pos_corpus #元素是元组,元组(x,y),x是在词典中的位置,y是1表示存在。
# (2)主题数寻优
#
# In[ ]:
#构造主题数寻优函数
def cos(vector1,vector2):
'''
函数功能:余玄相似度函数
'''
dot_product=0.0
normA=0.0
normB=0.0
for a,b in zip(vector1,vector2):
dot_product +=a*b
normA +=a**2
normB +=b**2
if normA==0.0 or normB==0.0:
return None
else:
return ( dot_product/((normA*normB)**0.5) )
# In[ ]:
#主题数寻优
#这个函数可以重复调用,解决其他项目的问题
def LDA_k(x_corpus,x_dict):
'''
函数功能:
'''
#初始化平均余玄相似度
mean_similarity=[]
mean_similarity.append(1)
#循环生成主题并计算主题间相似度
for i in np.arange(2,11):
lda=models.LdaModel(x_corpus,num_topics=i,id2word=x_dict) #LDA模型训练
for j in np.arange(i):
term=lda.show_topics(num_words=50)
#提取各主题词
top_word=[] #shape=(i,50)
for k in np.arange(i):
top_word.append( [''.join(re.findall('"(.*)"',i)) for i in term[k][1].split('+')]) #列出所有词
#构造词频向量
word=sum(top_word,[]) #列车所有词
unique_word=set(word) #去重
#构造主题词列表,行表示主题号,列表示各主题词
mat=[] #shape=(i,len(unique_word))
for j in np.arange(i):
top_w=top_word[j]
mat.append( tuple([ top_w.count(k) for k in unique_word ])) #统计list中元素的频次,返回元组
#两两组合。方法一
p=list(itertools.permutations(list(np.arange(i)),2)) #返回可迭代对象的所有数学全排列方式。
y=len(p) # y=i*(i-1)
top_similarity=[0]
for w in np.arange(y):
vector1=mat[p[w][0]]
vector2=mat[p[w][1]]
top_similarity.append(cos(vector1,vector2))
# #两两组合,方法二
# for x in range(i-1):
# for y in range(x,i):
#计算平均余玄相似度
mean_similarity.append(sum(top_similarity)/ y)
return mean_similarity
# In[ ]:
#计算主题平均余玄相似度
pos_k=LDA_k(pos_corpus,pos_dict)
neg_k=LDA_k(neg_corpus,neg_dict)
pos_k
neg_k
# In[ ]:
pd.Series(pos_k,index=range(1,11)).plot()
plt.title('正面评论LDA主题数寻优')
plt.show()
# In[ ]:
pd.Series(neg_k,index=range(1,11)).plot()
plt.title('负面评论LDA主题数寻优')
plt.show()
# In[ ]:
pos_lda=models.LdaModel(pos_corpus,num_topics=2,id2word=pos_dict)
neg_lda=models.LdaModel(neg_corpus,num_topics=2,id2word=neg_dict)
pos_lda.print_topics(num_topics=10)
neg_lda.print_topics(num_topics=10)