Python_LDA实现方法详解

LDA(Latent Dirichlet allocation)模型是一种常用而用途广泛地概率主题模型。其实现一般通过Variational inference和Gibbs Samping实现。作者在提出LDA模型时给出了其变分推理的C源码（后续贴出C++改编的类），这里贴出基于Python的第三方模块改写的LDA类及实现。

#coding:utf-8
import numpy as np
import lda
import lda.datasets
import jieba
import codecs
class LDA_v20161130():
 def __init__(self, topics=2):
   self.n_topic = topics
   self.corpus = None
   self.vocab = None
   self.ppCountMatrix = None
   self.stop_words = [u'，', u'。', u'、', u'（', u'）', u'·', u'！', u' ', u'：', u'“', u'”', u'\n']
   self.model = None
 def loadCorpusFromFile(self, fn):
   # 中文分词
   f = open(fn, 'r')
   text = f.readlines()
   text = r' '.join(text)
   seg_generator = jieba.cut(text)
   seg_list = [i for i in seg_generator if i not in self.stop_words]
   seg_list = r' '.join(seg_list)
   # 切割统计所有出现的词纳入词典
   seglist = seg_list.split(" ")
   self.vocab = []
   for word in seglist:
     if (word != u' ' and word not in self.vocab):
       self.vocab.append(word)
   CountMatrix = []
   f.seek(0, 0)
   # 统计每个文档中出现的词频
   for line in f:
     # 置零
     count = np.zeros(len(self.vocab),dtype=np.int)
     text = line.strip()
     # 但还是要先分词
     seg_generator = jieba.cut(text)
     seg_list = [i for i in seg_generator if i not in self.stop_words]
     seg_list = r' '.join(seg_list)
     seglist = seg_list.split(" ")
     # 查询词典中的词出现的词频
     for word in seglist:
       if word in self.vocab:
         count[self.vocab.index(word)] += 1
     CountMatrix.append(count)
   f.close()
   #self.ppCountMatrix = (len(CountMatrix), len(self.vocab))
   self.ppCountMatrix = np.array(CountMatrix)
   print "load corpus from ％s success!"％fn
 def setStopWords(self, word_list):
   self.stop_words = word_list
 def fitModel(self, n_iter = 1500, _alpha = 0.1, _eta = 0.01):
   self.model = lda.LDA(n_topics=self.n_topic, n_iter=n_iter, alpha=_alpha, eta= _eta, random_state= 1)
   self.model.fit(self.ppCountMatrix)
 def printTopic_Word(self, n_top_word = 8):
   for i, topic_dist in enumerate(self.model.topic_word_):
     topic_words = np.array(self.vocab)[np.argsort(topic_dist)][:-(n_top_word + 1):-1]
     print "Topic:",i,"\t",
     for word in topic_words:
       print word,
     print
 def printDoc_Topic(self):
   for i in range(len(self.ppCountMatrix)):
     print ("Doc ％d:((top topic:％s) topic distribution:％s)"％(i, self.model.doc_topic_[i].argmax(),self.model.doc_topic_[i]))
 def printVocabulary(self):
   print "vocabulary:"
   for word in self.vocab:
     print word,
   print
 def saveVocabulary(self, fn):
   f = codecs.open(fn, 'w', 'utf-8')
   for word in self.vocab:
     f.write("％s\n"％word)
   f.close()
 def saveTopic_Words(self, fn, n_top_word = -1):
   if n_top_word==-1:
     n_top_word = len(self.vocab)
   f = codecs.open(fn, 'w', 'utf-8')
   for i, topic_dist in enumerate(self.model.topic_word_):
     topic_words = np.array(self.vocab)[np.argsort(topic_dist)][:-(n_top_word + 1):-1]
     f.write( "Topic:％d\t"％i)
     for word in topic_words:
       f.write("％s "％word)
     f.write("\n")
   f.close()
 def saveDoc_Topic(self, fn):
   f = codecs.open(fn, 'w', 'utf-8')
   for i in range(len(self.ppCountMatrix)):
     f.write("Doc ％d:((top topic:％s) topic distribution:％s)\n" ％ (i, self.model.doc_topic_[i].argmax(), self.model.doc_topic_[i]))
   f.close()

算法实现demo：

例如，抓取BBC川普当选的新闻作为语料，输入以下代码：

if __name__=="__main__":
 _lda = LDA_v20161130(topics=20)
 stop = [u'!', u'@', u'#', u',',u'.',u'/',u';',u' ',u'[',u']',u'$',u'％',u'^',u'&',u'*',u'(',u')',
     u'"',u':',u'<',u'>',u'?',u'{',u'}',u'=',u'+',u'_',u'-',u'''''']
 _lda.setStopWords(stop)
 _lda.loadCorpusFromFile(u'C:\\Users\Administrator\Desktop\\BBC.txt')
 _lda.fitModel(n_iter=1500)
 _lda.printTopic_Word(n_top_word=10)
 _lda.printDoc_Topic()
 _lda.saveVocabulary(u'C:\\Users\Administrator\Desktop\\vocab.txt')
 _lda.saveTopic_Words(u'C:\\Users\Administrator\Desktop\\topic_word.txt')
 _lda.saveDoc_Topic(u'C:\\Users\Administrator\Desktop\\doc_topic.txt')

因为语料全部为英文，因此这里的stop_words全部设置为英文符号，主题设置20个，迭代1500次。结果显示，文档148篇，词典1347词，总词数4174，在i3的电脑上运行17s。
Topic_words部分输出如下：

Topic: 0
to will and of he be trumps the what policy
Topic: 1 he would in said not no with mr this but
Topic: 2 for or can some whether have change health obamacare insurance
Topic: 3 the to that president as of us also first all
Topic: 4 trump to when with now were republican mr office presidential
Topic: 5 the his trump from uk who president to american house
Topic: 6 a to that was it by issue vote while marriage
Topic: 7 the to of an are they which by could from
Topic: 8 of the states one votes planned won two new clinton
Topic: 9 in us a use for obama law entry new interview
Topic: 10 and on immigration has that there website vetting action given

Doc_Topic部分输出如下：

Doc 0:((top topic:4) topic distribution:[ 0.02972973 0.0027027 0.0027027 0.16486486 0.32702703 0.19189189
0.0027027 0.0027027 0.02972973 0.0027027 0.02972973 0.0027027
0.0027027 0.0027027 0.02972973 0.0027027 0.02972973 0.0027027
0.13783784 0.0027027 ])
Doc 1:((top topic:18) topic distribution:[ 0.21 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.11 0.01 0.01 0.01
0.01 0.01 0.01 0.01 0.01 0.01 0.31 0.21])
Doc 2:((top topic:18) topic distribution:[ 0.02075472 0.00188679 0.03962264 0.00188679 0.00188679 0.00188679
0.00188679 0.15283019 0.00188679 0.02075472 0.00188679 0.24716981
0.00188679 0.07735849 0.00188679 0.00188679 0.00188679 0.00188679
0.41698113 0.00188679])

当然，对于英文语料，需要排除大部分的虚词以及常用无意义词，例如it, this, there, that...在实际操作中，需要合理地设置参数。

换中文语料尝试，采用习大大就卡斯特罗逝世发表的吊唁文章和朴槿惠辞职的新闻。

Topic: 0
的 同志 和 人民 卡斯特罗 菲德尔 古巴 他 了 我
Topic: 1 在 朴槿惠 向 表示 总统 对 将 的 月 国民
Doc 0:((top topic:0) topic distribution:[ 0.91714123 0.08285877])
Doc 1:((top topic:1) topic distribution:[ 0.09200666 0.90799334])

还是存在一些虚词，例如“的”，“和”，“了”，“对”等词的干扰，但是大致来说，两则新闻的主题分布很明显，效果还不赖。

来源：http://blog.csdn.net/liuph_/article/details/53406609