朴素贝叶斯算法简单、高效。接下来我们来介绍其如何应用在《红楼梦》作者的鉴别上。
第一步，当然是先得有文本数据，我在网上随便下载了一个txt（当时急着交初稿。。。）。分类肯定是要一个回合一个回合的分，所以我们拿到文本数据后，先进行回合划分。然后就是去标点符号、分词，做词频统计。
1 # -*- coding: utf-8 -*- 2 import re 3 import jieba 4 import string 5 import collections as coll 6 jieba.load_userdict('E:\\forpython\\红楼梦词汇大全.txt') # 导入搜狗的红楼梦词库 7 8 9 class textprocesser: 10def __init__(self): 11 pass 1213# 将小说分成120个章节并单独保存到txt文件中 14def divide_into_chapter(self): 15 red=open('E:\\forpython\\红楼梦.txt',encoding='utf-8') 16 each_line = red.readline() 17 chapter_count = 0 18 chapter_text = '' 19 complied_rule = pile('第[一二三四五六七八九十百]+回 ') while each_line: 22 if re.findall(complied_rule,each_line): 23 file_name = 'chap'+str(chapter_count) 24 file_out = open('E:\\forpython\\chapters\\'+file_name+'.txt','a',encoding = 'utf-8') 25 file_out.write(chapter_text) 26 chapter_count += 1 27 file_out.close() 28 chapter_text = each_line 29 else: 30 chapter_text += each_line 3132 each_line = red.readline() 3334 red.close() 3536 37# 对单个章节的分词 38def segmentation(self,text,text_count): 39 file_name = 'chap'+str(text_count)+'-words.txt' 40 file_out = open('E:\\forpython\\chapter2words\\'+file_name,'a',encoding='utf-8') 41 delset = string.punctuation 4243 line=text.readline() 4445 while line: 46 seg_list = jieba.cut(line,cut_all = False) 47 words = " ".join(seg_list) 48 words = words.translate(delset) # 去除英文标点 49 words = "".join(words.split('\n')) # 去除回车符 50 words = self.delCNf(words) # 去除中文标点 51 words = re.sub('[ \u3000]+',' ',words) # 去除多余的空格 52 file_out.write(words) 53 line = text.readline() 5455 file_out.close() 56 text.close() 57 58 59# 对所有章节分词 60def do_segmentation(self): 61 for loop in range(1,121): 62 file_name = 'chap'+str(loop)+'.txt' 63 file_in = open('E:\\forpython\\chapters\\'+file_name,'r',encoding = 'utf-8') 6465 self.segmentation(file_in,loop) 6667 file_in.close() 68 69# 去除中文字符函数 70def delCNf(self,line): 71 regex = pile('[^\u4e00-\u9fa5a-zA-Z0-9\s]') 72 return regex.sub('', line) 737475# 去除标点后进行词频统计 76def count_words(self,text,textID): 77 line = str(text) 78 words = line.split() 79 words_dict = coll.Counter(words) # 生成词频字典 8081 file_name = 'chap'+str(textID)+'-wordcount.txt' 82 file_out = open('E:\\forpython\\chapter-wordcount\\'+file_name,'a',encoding = 'utf-8') 8384 # 排序后写入文本 85 sorted_result = sorted(words_dict.items(),key = lambda d:d[1],reverse = True) 86 for one in sorted_result: 87 line = "".join(one[0] + '\t' + str(one[1]) + '\n') 88 file_out.write(line) 8990 file_out.close() 91 92 93 94def do_wordcount(self): 95 for loop in range(1,121): 96 file_name = 'chap'+str(loop)+'-words.txt' 97 file_in = open('E:\\forpython\\chapter2words\\'+file_name,'r',encoding = 'utf-8') 98 line = file_in.readline() 99 100 text = ''101 while line:102 text += line103 line = file_in.readline()104 self.count_words(text,loop)105 file_in.close()106107108 if __name__ == '__main__':109processer = textprocesser()110processer.divide_into_chapter()111processer.do_segmentation()112processer.do_wordcount()
文本分类我个人感觉最重要的是选取特征向量，我查阅了相关文献，决定选取五十多个文言虚词和二十多个在120个回合中均出现过的词汇（文言虚词的使用不受情节影响，只与作者写作习惯有关）。下面是生成
特征向量的代码
1 # -*- coding: utf-8 -*- 2 import jieba 3 import re 4 import string 5 import collections as coll 6 jieba.load_userdict('E:\\forpython\\红楼梦词汇大全.txt') # 导入搜狗的红楼梦词库 7 8 class featureVector: 9def __init__(self): 10 pass 1112# 去除中文字符函数 13def delCNf(self,line): 14 regex = pile('[^\u4e00-\u9fa5a-zA-Z0-9\s]') 15 return regex.sub('', line) 161718# 对整篇文章分词 19def cut_words(self): 20 red = open('E:\\forpython\\红楼梦.txt','r',encoding = 'utf-8') 21 file_out = open('E:\\forpython\\红楼梦-词.txt','a',encoding = 'utf-8') 22 delset = string.punctuation 2324 line = red.readline() 2526 while line: 27 seg_list = jieba.cut(line,cut_all = False) 28 words = ' '.join(seg_list) 29 words = words.translate(delset) # 去除英文标点 30 words = "".join(words.split('\n')) # 去除回车符 31 words = self.delCNf(words) # 去除中文标点 32 words = re.sub('[ \u3000]+',' ',words) # 去除多余的空格 33 file_out.write(words) 34 line = red.readline() 35 36 file_out.close() 37 red.close() 3839# 统计词频 40def count_words(self): 41 data = open('E:\\forpython\\红楼梦-词.txt','r',encoding = 'utf-8') 42 line = data.read() 43 data.close() 44 words = line.split() 45 words_dict = coll.Counter(words) # 生成词频字典 4647 file_out = open('E:\\forpython\\红楼梦-词频.txt','a',encoding = 'utf-8') 4849 # 排序后写入文本 50 sorted_result = sorted(words_dict.items(),key = lambda d:d[1],reverse = True) 51 for one in sorted_result: 52 line = "".join(one[0] + '\t' + str(one[1]) + '\n') 53 file_out.write(line) 5455 file_out.close() 56575859def get_featureVector(self): 60 # 将分词后的120个章节文本放入一个列表中 61 everychapter = [] 62 for loop in range(1,121): 63 data = open('E:\\forpython\\chapter2words\\chap'+str(loop)+'-words.txt','r',encoding = 'utf-8') 64 each_chapter = data.read() 65 everychapter.append(each_chapter) 66 data.close() 6768 temp = open('E:\\forpython\\红楼梦-词.txt','r',encoding = 'utf-8') 69 word_beg = temp.read() 70 word_beg = word_beg.split(' ') 71 temp.close() 7273 # 找出每一个回合都出现的词 74 cleanwords = [] 75 for loop in range(1,121): 76 data = open('E:\\forpython\\chapter2words\\chap'+str(loop)+'-words.txt','r',encoding = 'utf-8') 77 words_list = list(set(data.read().split())) 78 data.close() 79 cleanwords.extend(words_list) 8081 cleanwords_dict = coll.Counter(cleanwords) 82 83 cleanwords_dict = {k:v for k, v in cleanwords_dict.items() if v >= 120} 8485 cleanwords_f = list(cleanwords_dict.keys()) 8687 xuci = open('E:\\forpython\\文言虚词.txt','r',encoding = 'utf-8') 88 xuci_list = xuci.read().split() 89 xuci.close() 90 featureVector = list(set(xuci_list + cleanwords_f)) 91 featureVector.remove('\ufeff') 92 93 # 写入文本 94 file_out = open('E:\\forpython\\红楼梦-特征向量.txt','a',encoding = 'utf-8') 95 for one in featureVector: 96 line = "".join(one+ '\n') 97 file_out.write(line) 9899 file_out.close()100 return(featureVector)101 102 if __name__ == '__main__':103vectorbuilter = featureVector()104vectorbuilter.cut_words()105vectorbuilter.count_words()106vectorbuilter.get_featureVector()
朴素贝叶斯文本分类就是用特征向量的词频作为每个回合的代表（偷个懒，直接截图答辩的ppt）
用特征向量把所有一百二十个回合向量化后，你会得到120×70的一个数组。接下来就简单了。直接挑选训练集，在这我是在前80回中挑选了20至29回标记为第一类（用数字1表示），并将其作为第一类的训练集；在后80回合中挑选了110至119回标记为第二类（用数字2表示），并将其作为第二类的训练集。
1 # -*- coding: utf-8 -*- 2 3 import numpy as np 4 from sklearn.naive_bayes import MultinomialNB 5 import get_trainset as ts 6 x_train = ts.get_train_set().get_all_vector() 7 8 9 10 class result:11def __inti__(self):12 pass1314def have_Xtrainset(self):15 Xtrainset = x_train16 Xtrainset = np.vstack((Xtrainset[19:29],Xtrainset[109:119]))17 return(Xtrainset) 1819def as_num(self,x):20 y='{:.10f}'.format(x)21 return(y)2223def built_model(self):24 x_trainset = self.have_Xtrainset()25 y_classset = np.repeat(np.array([1,2]),[10,10])26 27 NBclf = MultinomialNB()28 NBclf.fit(x_trainset,y_classset) # 建立模型29 30 all_vector = x_train31 32 result = NBclf.predict(all_vector)33 print('前'+str(len(result[0:80]))+'回分类结果为：')34 print(result[0:80])35 print('后'+str(len(result[80:121]))+'回分类结果为：')36 print(result[80:121])37 38 diff_chapter = [80,81,83,84,87,88,90,100]39 for i in diff_chapter:40 tempr = NBclf.predict_proba(all_vector[i])41 print('第'+str(i+1)+'回的分类概率为： ')42 print(str(self.as_num(tempr[0][0]))+' '+str(self.as_num(tempr[0][1])))43 44 45 if __name__ == '__main__':46res = result()47res.built_model()
上面是直接调用了skit-learn的MultinomialNB函数，详细情况我在前一篇中讲过。
得到分类结果：
从最终的分类结果来看，在第82回合左右是有一个比较明显的分界点，这样看来前80回合与后40回合在写作风格上还是有显著的差异的，这个结果和红楼梦学术界的年的推断比较一致。
至于为何在后40回中有8个回合被分到1类中，这8个回合分别是81回、82回、84回、85回、88回、89回、91回还有101回，都是在第80回合附近，这个差异有可能是由于上下文的衔接所导致的，因为本文所使用的《红楼梦》文本是从网上下载得到的，，版本不明，所以也有可能是由于红楼梦的版本所导致的。
代码肯定还有很多可以优化的地方，在这里献丑了。。。。
