赞
踩
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.feature_extraction.text import CountVectorizer review_long_clean = pd.read_excel('1_review_long_clean.xlsx') #导入处理后的数据集
导入评价情感词典
#来自知网发布的情感分析用词语集
pos_comment=pd.read_csv('./正面评价词语(中文).txt',header=None,sep='\n',encoding='utf-8')
neg_comment=pd.read_csv('./负面评价词语(中文).txt',header=None,sep='\n',encoding='utf-8')
pos_emotion=pd.read_csv('./正面情感词语(中文).txt',header=None,sep='\n',encoding='utf-8')
neg_emotion=pd.read_csv('./负面情感词语(中文).txt',header=None,sep='\n',encoding='utf-8')
查看一下正负情感词典分布
print('正面评价词语:',pos_comment.shape)
print('负面评价词语:',neg_comment.shape)
print('正面情感词语:',pos_emotion.shape)
print('负面情感词语:',neg_emotion.shape)
正面评价词语: (3743, 1)
负面评价词语: (3138, 1)
正面情感词语: (833, 1)
负面情感词语: (1251, 1)
#将正面和负面的合并
pos=pd.concat([pos_comment,pos_emotion],axis=0)
print('正面合并后:',pos.shape)
neg=pd.concat([neg_comment,neg_emotion],axis=0)
print('负面合并后:',neg.shape)
正面合并后: (4576, 1)
负面合并后: (4389, 1)
#判断在不在词典中
c='点赞'
print(c in pos.values)
d='歇菜'
print(d in neg.values)
False
False
不在就添加进词典
#不在就添加进词典
new_pos=pd.Series(['点赞'])
new_neg=pd.Series(['歇菜'])
positive=pd.concat([pos,new_pos],axis=0)
print(positive.shape)
negative=pd.concat([neg,new_neg],axis=0)
print(negative.shape)
#将情感词存入,Dataframe中,并且正面词赋予权重1
positive.columns=['review']
positive['weight']=pd.Series([1]*len(positive))
positive.head()
#将情感词存入,Dataframe中,并且负面词赋予权重-1
negative.columns=['review']
negative['weight']=pd.Series([-1]*len(negative))
negative.head()
#将正负情感词合并
pos_neg=pd.concat([positive,negative],axis=0)
pos_neg.shape
(8967, 2)
#表联接
#清洗赶紧的数据,赋值给data
data=review_long_clean.copy()
print('data原始shape:',data.shape)
#将data和情感词两个表合并在一块,按照data的word列,情感数据按照review
review_mltype=pd.merge(data,pos_neg,how='left',left_on='word',right_on='review')
review_mltype.shape
print('表联接合并后:',review_mltype.shape)
print('----------完成评论中的情感词赋值------------')
print('review_mltype:')
review_mltype.head()
data原始shape: (25172, 6)
表联接合并后: (25172, 8)
----------完成评论中的情感词赋值------------
review_mltype:
#删除情感值列,同时赋予不是情感词的词,权重为0
review_mltype=review_mltype.drop(['review'],axis=1)
review_mltype=review_mltype.replace(np.nan,0)
review_mltype.head()
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。
#读取否定词词典,新建一列为freq
notdict=pd.read_csv('./not.csv')
notdict.shape
notdict['freq']=[1]*len(notdict)
notdict.head()
#准备一,构建amend_weight列,其值与情感权重值相同
review_mltype['amend_weight']=review_mltype['weight']
#创建,id列,初始化值为0-到最后一个的,顺序值
review_mltype['id']=np.arange(0,review_mltype.shape[0])
review_mltype.head(10)
review_mltype是整理后有情感权重的,全部评论数据
只取出有情感值的文本
# 准备二,只取出有情感值的行
only_review_mltype=review_mltype[review_mltype['weight']!=0]
#只保留有情感值的索引
only_review_mltype.index=np.arange(0,only_review_mltype.shape[0]) #索引重置
print(only_review_mltype.shape)
only_review_mltype.head()
(1526, 10)
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。
#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]] )
review_mltype.shape
review_mltype[(review_mltype['weight']-review_mltype['amend_weight'])!=0] #说明两列值一样
将一条文本中所有情感值计算,在相加到一块。
emotion_value=review_mltype.groupby('index_content',as_index=False)['amend_weight'].sum()
emotion_value.head()
emotion_value.to_csv('./1_emotion_value.csv',index=True,header=True)
#每条评论的amend_weight总和不等于零
content_emotion_value=emotion_value.copy()
content_emotion_value.shape
#取出不等于0的的文本
content_emotion_value=content_emotion_value[content_emotion_value['amend_weight']!=0]
#设置情感倾向
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()
读取原始数据
#读取原始数据
raw_data=pd.read_csv('./reviews.csv')
raw_data.head()
方法缺陷:
#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[6]
raw_data.content[7]
raw_data.content[8]
将情感分析结果,和原始数据集合并:
#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)
与真实值对比下,情感分析结果的准确性:
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
data=data_type[['content_type','ml_type']]
data=data.dropna(axis=0)
print( classification_report(data['content_type'],data['ml_type']) )
precision recall f1-score support
- 1
neg 0.90 0.87 0.88 437
pos 0.89 0.92 0.90 501
accuracy 0.89 938
macro avg 0.90 0.89 0.89 938
weighted avg 0.89 0.89 0.89 938
将文本按照情感分类,然后统计词云
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('./pl.jpg') 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('./pl.jpg') 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()
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。
data=review_mltype.copy()
word_data_pos=data[data['ml_type']=='pos']
word_data_neg=data[data['ml_type']=='neg']
#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])
#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] ]
#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) )
#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
#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
[1,
0.04,
0.006666666666666667,
0.0033333333333333335,
0.0,
0.004,
0.006666666666666666,
0.025000000000000015,
0.03500000000000002,
0.046222222222222234]
[1,
0.02,
0.0,
0.0,
0.0,
0.0026666666666666666,
0.002857142857142857,
0.0035714285714285718,
0.01333333333333334,
0.022222222222222233]
pd.Series(pos_k,index=range(1,11)).plot()
plt.title('正面评论LDA主题数寻优')
plt.show()
<matplotlib.axes._subplots.AxesSubplot at 0x1de4b888>
Text(0.5, 1.0, ‘正面评论LDA主题数寻优’)
pd.Series(neg_k,index=range(1,11)).plot()
plt.title('负面评论LDA主题数寻优')
plt.show()
matplotlib.axes._subplots.AxesSubplot at 0x1d077ac8>
Text(0.5, 1.0, ‘负面评论LDA主题数寻优’)
确定正负面主题数目都为3
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)
[(0,
‘0.085*“安装” + 0.036*“满意” + 0.019*“服务” + 0.018*“不错” + 0.015*“好评” + 0.012*“客服” + 0.011*“人员” + 0.010*“物流” + 0.009*“送” + 0.008*“家里”’),
(1,
‘0.024*“师傅” + 0.021*“送货” + 0.019*“很快” + 0.016*“值得” + 0.013*“售后” + 0.012*“信赖” + 0.012*“东西” + 0.009*“太” + 0.009*“购物” + 0.009*“电话”’)]
[(0,
‘0.089*“安装” + 0.018*“师傅” + 0.015*“慢” + 0.013*“装” + 0.011*“打电话” + 0.011*“太慢” + 0.008*“坑人” + 0.008*“服务” + 0.007*“配件” + 0.007*“问”’),
(1,
‘0.021*“垃圾” + 0.019*“太” + 0.019*“差” + 0.016*“安装费” + 0.015*“售后” + 0.014*“东西” + 0.013*“不好” + 0.013*“客服” + 0.012*“加热” + 0.012*“小时”’)]
Copyright © 2003-2013 www.wpsshop.cn 版权所有,并保留所有权利。