用 Python分析朋友圈好友的签名

发布时间:2019-05-22 22:11:50编辑:auto阅读(2262)

    需要用到的第三方库:

    numpy:本例结合wordcloud使用

    jieba对中文惊进行分词

    PIL: 对图像进行处理(本例与wordcloud结合使用)

    snowlp对文本信息进行情感判断

    wordcloud生成词云
    matplotlib:绘制2D图形

    # -*- coding: utf-8 -*-
    """
    朋友圈朋友签名的词云生成以及
    签名情感分析
    想要学习Python?Python学习交流群:984632579满足你的需求,资料都已经上传群文件,可以自行下载!
    """
    import re,jieba,itchat
    import jieba.analyse
    import numpy as np
    from PIL import Image
    from snownlp import SnowNLP
    from wordcloud import WordCloud
    import matplotlib.pyplot as plt
    itchat.auto_login(hotReload=True)
    friends = itchat.get_friends(update=True)
    def analyseSignature(friends):
        signatures = ''
        emotions = []
        for friend in friends:
            signature = friend['Signature']
            if(signature != None):
                signature = signature.strip().replace('span', '').replace('class', '').replace('emoji', '')
                signature = re.sub(r'1f(\d.+)','',signature)
                if(len(signature)>0):
                    nlp = SnowNLP(signature)
                    emotions.append(nlp.sentiments)
                    signatures += ' '.join(jieba.analyse.extract_tags(signature,5))
        with open('signatures.txt','wt',encoding='utf-8') as file:
             file.write(signatures)
    
        # 朋友圈朋友签名的词云相关属性设置
        back_coloring = np.array(Image.open('alice_color.png'))
        wordcloud = WordCloud(
            font_path='simfang.ttf',
            background_color="white",
            max_words=1200,
            mask=back_coloring, 
            max_font_size=75,
            random_state=45,
            width=1250, 
            height=1000, 
            margin=15
        )
        
        #生成朋友圈朋友签名的词云
        wordcloud.generate(signatures)
        plt.imshow(wordcloud)
        plt.axis("off")
        plt.show()
        wordcloud.to_file('signatures.jpg')#保存到本地文件
    
        # Signature Emotional Judgment
        count_good = len(list(filter(lambda x:x>0.66,emotions)))#正面积极
        count_normal = len(list(filter(lambda x:x>=0.33 and x<=0.66,emotions)))#中性
        count_bad = len(list(filter(lambda x:x<0.33,emotions)))#负面消极
        labels = [u'负面消极',u'中性',u'正面积极']
        values = (count_bad,count_normal,count_good)
        plt.rcParams['font.sans-serif'] = ['simHei'] 
        plt.rcParams['axes.unicode_minus'] = False
        plt.xlabel(u'情感判断')#x轴
        plt.ylabel(u'频数')#y轴
        plt.xticks(range(3),labels)
        plt.legend(loc='upper right',)
        plt.bar(range(3), values, color = 'rgb')
        plt.title(u'%s的微信好友签名信息情感分析' % friends[0]['NickName'])
        plt.show()
    analyseSignature(friends)

    效果图

     

关键字