在学界一般认为,《红楼梦》后 40 回并非曹雪芹所著。利用机器学习相关算法来进行判断
原理
每个作者写作都有自己的用词习惯和风格,即使是故意模仿也会留下很多痕迹。
在文言文中,文言虚词分布均匀,书中每个回目都会出现很多文言虚词,差别在于出现频率不同,我们把文言虚词的出现频率作为特征。
不只文言虚词,还有其他的词在所有回目中出现频率很多。比如对第 80 回进行词频统计,得到
- 了 172
- 的 142
- 我 70
- 宝玉 65
- 你 61
- 道 54
- 他 51
这些高频词汇也可以作为特征向量。
本文将 20~29 回(诗词曲比较均衡)作为类别 1 的学习样本,将 110~119 回作为类别 2 的学习样本。
将两个类别的特征向量输入到 SVM(支持向量机) 进行训练得出一个分类模型。再对剩余回目进行分类,看它们分别偏向于哪个类别。
分词
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。
#!/usr/bin/python
# -*- coding:utf-8 -*-
import string
import jieba
import sys
import re
class textProcesser(object):
def __init__(self):
pass
# 将书分为章节
def divide_into_chapter(self):
file_in = open('text/redmansions.txt', 'r')
line = file_in.readline()
chapter_cnt = 1
chapter_text = ""
while line:
if '[(' in line:
path_str = 'text/chapter-' + str(chapter_cnt)
file_out = open(path_str, 'a')
file_out.write(chapter_text)
chapter_cnt += 1
file_out.close()
chapter_text = line
else:
chapter_text += line
line = file_in.readline()
file_in.close
# 对一章分词
def divide_into_words(self, document, docID):
path_str = 'text/chapter-words-' + str(docID)
file_out = open(path_str,'a')
line = document.readline()
while(line):
seg_list = jieba.cut(line, cut_all=False)
words = " ".join(seg_list)
file_out.write(words)
line = document.readline()
file_out.close()
# 对所有章节分词
def perform_segmentation(self):
for loop in range(1, 121):
path_str = 'text/chapter-' + str(loop)
file_in = open(path_str, 'r')
self.divide_into_words(file_in, loop)
# 将每个文档去除标点后,再进行词频统计
def count_words(self, document, docID):
result_dict = {}
delset = string.punctuation
line = str(document)
line = line.translate(None, delset) #去除英文标点
line = "".join(line.split('\n')) # 去除回车
line = self.sub_replace(line) #去除中文标点
word_array = []
words = line.split()
for word in words:
if not result_dict.has_key(word):
result_dict[word] = 1
else:
result_dict[word] += 1
path_str = 'text/chapter-wordcount-' + str(docID)
file_out = open(path_str,'a')
# 排序后写入文本
sorted_result = sorted(result_dict.iteritems(), key=lambda d:d[1], reverse = True)
for one in sorted_result:
line = "".join(one[0] + '\t' + str(one[1]) + '\n')
file_out.write(line)
file_out.close()
# 对所有文档进行分词
def perform_wordcount(self):
for loop in range(1, 121):
path_str = 'text/chapter-words-' + str(loop)
file_in = open(path_str, 'r')
line = file_in.readline()
document = ""
while line:
document += line
line = file_in.readline()
self.count_words(document, loop)
file_in.close()
def sub_replace(self, line):
regex = re.compile("[^\u4e00-\u9fa5a-zA-Z0-9\s]")
return regex.sub('', line.decode('utf-8'))
特征选取
- [
‘之’, ‘其’, ‘或’, ‘亦’, ‘方’, ‘于’, ‘即’, ‘皆’, ‘因’, ‘仍’,
‘故’, ‘尚’, ‘呢’, ‘了’, ‘的’, ‘着’, ‘一’, ‘不’, ‘乃’, ‘呀’,
‘吗’, ‘咧’, ‘啊’, ‘把’, ‘让’, ‘向’, ‘往’, ‘是’, ‘在’, ‘越’,
‘再’, ‘更’, ‘比’, ‘很’, ‘偏’, ‘别’, ‘好’, ‘可’, ‘便’, ‘就’,
‘但’, ‘儿’, # 42 个文言虚词
‘又’, ‘也’, ‘都’, ‘要’, # 高频副词
‘这’, ‘那’, ‘你’, ‘我’, ‘他’ # 高频代词
‘来’, ‘去’, ‘道’, ‘笑’, ‘说’ #高频动词
]
选取常用的 42 个文言虚词和通过词频统计得到的高频使用的词作为特征,分别计算它们在各个回目中出现的频率作为特征向量。
代码
class modelBuilder(object):
def __init__(self):
pass
def get_wordnum_of_chapter(self, DocID):
path_str = 'text/chapter-' + str(DocID)
file_in = open(path_str)
text = ""
for line in file_in:
text += "".join(line.split('\n')) # 去除回车
file_in.close
num = len(text.decode("gb18030"))
return num
# 每个文档提取特征向量
def build_feature_vector(self, DocID, label):
path_str = 'text/chapter-wordcount-' + str(DocID)
# function_word_list = ['之', '其', '或', '亦', '方', '于', '即', '皆', '因', '仍',
# '故', '尚', '呢', '了', '的', '着', '不', '乃', '呀',
# '吗', '咧', '啊', '把', '让', '向', '往', '是', '在', '越',
# '再', '更', '比', '很', '偏', '别', '好', '可', '便', '就',
# '但', '儿', # 42 个文言虚词
# '又', '也', # 高频副词
# '这', '那', '你', '我', '他' #高频代词
# '来', '去', '道', '笑'] #高频动词
function_word_list = ['之', '其', '或', '亦', '方', '于', '即', '皆', '因', '仍',
'故', '尚', '呢', '了', '的', '着', '一', '不', '乃', '呀',
'吗', '咧', '啊', '把', '让', '向', '往', '是', '在', '越',
'再', '更', '比', '很', '偏', '别', '好', '可', '便', '就',
'但', '儿', # 42 个文言虚词
'又', '也', '都', '要', # 高频副词
'这', '那', '你', '我', '他' # 高频代词
'来', '去', '道', '笑', '说' #高频动词
]
feature_vector_list = []
for function_word in function_word_list:
find_flag = 0
file_in = open(path_str) #每次打开移动 cursor 到头部
line = file_in.readline()
while line:
words = line[:-1].split('\t')
if words[0] == function_word:
total_words = self.get_wordnum_of_chapter(DocID)
rate = float(words[1]) / total_words * 1000
rate = float("%.6f" % rate)# 指定位数
feature_vector_list.append(rate)
# print words[0] + ' : ' + line
file_in.close()
find_flag = 1
break
line = file_in.readline()
# 未找到词时向量为 0
if not find_flag:
feature_vector_list.append(0)
feature_vector_list.append(label)
return feature_vector_list
def make_positive_trainset(self):
positive_trainset_list = []
for loop in range(20, 30):
feature = self.build_feature_vector(loop, 1) #label 为 1 表示正例
positive_trainset_list.append(feature)
# print positive_trainset_list
np.save('pos_trainset.npy', positive_trainset_list)
def make_negative_trainset(self):
negative_trainset_list = []
for loop in range(110, 120):
feature = self.build_feature_vector(loop, 2) #label 为 0 表示负例
negative_trainset_list.append(feature)
# print negative_trainset_list
np.save('neg_trainset.npy', negative_trainset_list)
def make_trainset(self):
feature_pos = np.load('pos_trainset.npy')
feature_neg = np.load('neg_trainset.npy')
trainset = np.vstack((feature_pos, feature_neg))
np.save('trainset.npy', trainset)
def make_testset(self):
testset_list = []
for loop in range(1, 121):
feature = self.build_feature_vector(loop, 0) #无需 label,暂设为 0
testset_list.append(feature)
# print testset_list
np.save('testset.npy', testset_list)
特征向量含义
下面利用svm训练
# -*- coding: utf-8 -*-
import numpy as np
from sklearn.naive_bayes import MultinomialNB
import get_trainset as ts
x_train = ts.get_train_set().get_all_vector()
class result:
def __inti__(self):
pass
def have_Xtrainset(self):
Xtrainset = x_train
Xtrainset = np.vstack((Xtrainset[19:29],Xtrainset[109:119]))
return(Xtrainset)
def as_num(self,x):
y='{:.10f}'.format(x)
return(y)
def built_model(self):
x_trainset = self.have_Xtrainset()
y_classset = np.repeat(np.array([1,2]),[10,10])
NBclf = MultinomialNB()
NBclf.fit(x_trainset,y_classset) # 建立模型
all_vector = x_train
result = NBclf.predict(all_vector)
print('前'+str(len(result[0:80]))+'回分类结果为:')
print(result[0:80])
print('后'+str(len(result[80:121]))+'回分类结果为:')
print(result[80:121])
diff_chapter = [80,81,83,84,87,88,90,100]
for i in diff_chapter:
tempr = NBclf.predict_proba(all_vector[i])
print('第'+str(i+1)+'回的分类概率为: ')
print(str(self.as_num(tempr[0][0]))+' '+str(self.as_num(tempr[0][1])))
res = result()
res.built_model()
结果如下
1 指该回目属于类别 1,2 指该回目属于类别 2。
可以得出结论
- 前 80 回属于一类,后 40 回属于一类
- 80 回左右是分界点
- 后 40 回风格不同于前 80 回