我也说说中文分词（上：基于字符串匹配）

1. 序

词是句子组成的基本单元，不像英语句子已经分好词了，中文处理的第一步就是中文分词。

分词中面临的三大基本问题

分词规范
分词歧义
未登录词的识别
中文分词算法大概分为两大类

第一类：基于字符串匹配

    即扫描字符串，如果发现字符串的子串和词相同，就算匹配。这类分词通常会加入一些启发式规则，比如“正向/反向最大匹配”, “长词优先” 等策略。

优点：速度快，都是O(n)时间复杂度，实现简单，效果尚可

缺点：就是对歧义和未登录词处理不好

案例：庖丁解牛分词器就是基于字符串匹配的分词。

歧义的例子很简单"长春市/长春/药店"、 "长春/市长/春药/店"
未登录词即词典中没有出现的词，当然也就处理不好

第二类：基于统计以及机器学习的分词方式

      这类分词基于人工标注的词性和统计特征，对中文进行建模，即根据观测到的数据（标注好的语料）对模型参数进行估计，即训练。 在分词阶段再通过模型计算各种分词出现的概率，将概率最大的分词结果作为最终结果。常见的序列标注模型有HMM和CRF。

优点：很好处理歧义和未登录词问题，效果比基于字符串匹配效果好

缺点：需要大量的人工标注数据，较慢的分词速度

案例：Stanford Word Segmenter

 

2. 基于字符串匹配的中文分词（以前向最大匹配为例）

参考代码

def WordSeg(Inputfile, Outputfile):
f = file(Inputfile)
w = file(Outputfile, 'w')
for line in f:
line = line.strip().decode('utf-8')
senList = []
newsenList = []
tmpword = ''
for i in range(len(line)):
if line[i] in StopWord:
senList.append(tmpword)
senList.append(line[i])
tmpword = ''
else:
tmpword += line[i]
if i == len(line) - 1:
senList.append(tmpword)
#Pre
for key in senList:
if key in StopWord:
newsenList.append(key)
else:
tmplist = PreSenSeg(key, span)
for keyseg in tmplist:
newsenList.append(keyseg)
Prewriteline = ''
for key in newsenList:
Prewriteline = Prewriteline + key + ' '

w.write(Prewriteline.encode('utf-8') + '\n')
f.close()
w.close()

def PreSenSeg(sen, span):
post = span
if len(sen) < span:
post = len(sen)
cur = 0
revlist = []
while 1:
if cur >= len(sen):
break
s = re.search(u"^[0|1|2|3|4|5|6|7|8|9|\uff11|\uff12|\uff13|\uff14|\uff15|\uff16|\uff17|\uff18|\uff19|\uff10|\u4e00|\u4e8c|\u4e09|\u56db|\u4e94|\u516d|\u4e03|\u516b|\u4e5d|\u96f6|\u5341|\u767e|\u5343|\u4e07|\u4ebf|\u5146|\uff2f]+", sen[cur:])
if s:
if s.group() != '':
revlist.append(s.group())
cur = cur + len(s.group())
post = cur + span
if post > len(sen):
post = len(sen)
s = re.search(u"^[a|b|c|d|e|f|g|h|i|j|k|l|m|n|o|p|q|r|s|t|u|v|w|x|y|z|A|B|C|D|E|F|G|H|I|J|K|L|M|N|O|P|Q|R|S|T|U|V|W|X|Y|Z|\uff41|\uff42|\uff43|\uff44|\uff45|\uff46|\uff47|\uff48|\uff49|\uff47|\uff4b|\uff4c|\uff4d|\uff4e|\uff4f|\uff50|\uff51|\uff52|\uff53|\uff54|\uff55|\uff56|\uff57|\uff58|\uff59|\uff5a|\uff21|\uff22|\uff23|\uff24|\uff25|\uff26|\uff27|\uff28|\uff29|\uff2a|\uff2b|\uff2c|\uff2d|\uff2e|\uff2f|\uff30|\uff31|\uff32|\uff33|\uff35|\uff36|\uff37|\uff38|\uff39|\uff3a]+", sen[cur:])
if s:
if s.group() != '':
revlist.append(s.group())
cur = cur + len(s.group())
post = cur + span
if post > len(sen):
post = len(sen)
if (WordDic.has_key(sen[cur:post])) or (cur + 1 == post):
if sen[cur:post] != '':
revlist.append(sen[cur:post])
cur = post
post = post + span
if post > len(sen):
post = len(sen)
else:
post -= 1
return revlist

注意几点

首先根据标点切开分成小句子，标点绝对是分割的最佳标志。
句子中的数字（阿拉伯、汉字、阿拉伯+汉字。注意有十白千万亿）自动检测出来，不用再切割了。比如1998年=>1998 年
句子中的英文单词直接识别出来，不用分割了。比如：Happy New Year。
后向最大匹配与前向思路相同，只不过切分方向是从后往前。

 

3. 利用N-gram进行中文分词

      语言模型是根据语言客观事实而进行的语言抽象数学建模，是一种对应关系。语言模型与语言客观事实之间的关系，如同数学上的抽象直线与具体直线之间的关系。

N-gram

      语言模型在自然语言处理中占有重要地位，尤其是在基于统计模型的NLP任务中得到了广泛的应用，目前主要采用的是n元语法模型（N-gram model），这种模型构建简单、直接，但同时也因为数据缺乏而必须采取平滑算法。

      一个语言模型通常构建为字符串s的概率分布p(s)，p(s)试图反映字符串s作为一个句子出现时的频率。例如一个人所说的100个句子中大约有一句是“OK”，那么可以任务P(OK)=0.01。而对于句子“the apple eat an chicken”，可以认为其概率为0，因为几乎没有人这么说。与语言学不同，语言模型与句子是否合乎语法没有关系。对于字串假设有l个基元（基元可以是字、词、短语等）组成句子，那么s
= w1w2...wl，其概率计算公式为：

p(s)=p(w1)p(w2|w1)p(w3|w1w2).......p(wl|w1w2...pl-1)

      把第i个词wi之前的词w1w2....wi-1成为wi的“历史”。随着历史长度的增加，不同的历史数目成指数增长。如果历史长度为i-1，那么就有Li-1种不同的历史（L为词汇集的大小），这样必须在所有历史的基础上得出产生第i个词的概率。这样不可能从训练数据中正确估计出(wi|w1w2...wi-1),并且很多历史不可能从训练数据中出现。其中一种比较实际的做法基于这样的假设：第n个词的出现只与前面n-1个词相关，而与其它任何词都不相关，整句的概率就是各个词出现概率的乘积。这些概率可以通过直接从语料中统计N个词同时出现的次数得到。常用的是二元的Bi-Gram（只与前一个词有关）和三元的Tri-Gram（只与前两个词有关）。

二元模型为例

p(s) = p(w1|<BEG>)p(w2|w1)p(w3|w2)*****p(wl|wl-1)p(<End>|wl)

其中p(wi|wi-1) = p(wi-1wi)/p(wi-1*)

      前边已经利用前向最大匹配和后向最大匹配对句子进行了中文分词。为了提高分词的准确度，可以利用N-gram比较前向、后向哪个分词的得到的概率结果更大，就取相应的分词结果。

 

4. 小试牛刀

1. 前向后向中文分词

数据下载：待分词文件+对应答案+词典

代码

#! -*- coding:utf-8 -*-
import sys
import os
import re
#StopWordtmp = ['。', '，', '！', '？', '：', '“', '”', '‘', '’', '（', '）', '【', '】', '｛', '｝', '-', '－', '～', '［', '］', '〔', '〕', '．', '＠', '￥', '•', '.']
StopWordtmp = [' ', u'\u3000', u'\x30fb', u'\u3002', u'\uff0c', u'\uff01', u'\uff1f', u'\uff1a', u'\u201c', u'\u201d', u'\u2018', u'\u2019', u'\uff08', u'\uff09', u'\u3010', u'\u3011', u'\uff5b', u'\uff5d', u'-', u'\uff0d', u'\uff5e', u'\uff3b', u'\uff3d', u'\u3014', u'\u3015', u'\uff0e', u'\uff20', u'\uffe5', u'\u2022', u'.']

WordDic = {}
StopWord = []
span = 16

def InitStopword():
for key in StopWordtmp:
StopWord.append(key)

def InitDic(Dicfile):
f = file(Dicfile)
for line in f:
line = line.strip().decode('utf-8')
WordDic[line] = 1;
f.close()
print len(WordDic)
print "Dictionary has built down!"

def WordSeg(Inputfile, Outputfile, outputfile2):
f = file(Inputfile)
w = file(Outputfile, 'w')
w2 = file(Outputfile2, 'w')
for line in f:
line = line.strip().decode('utf-8')
senList = []
newsenList = []
tmpword = ''
for i in range(len(line)):
if line[i] in StopWord:
senList.append(tmpword)
senList.append(line[i])
tmpword = ''
else:
tmpword += line[i]
if i == len(line) - 1:
senList.append(tmpword)
#Pre
for key in senList:
if key in StopWord:
newsenList.append(key)
else:
tmplist = PreSenSeg(key, span)
for keyseg in tmplist:
newsenList.append(keyseg)
Prewriteline = ''
for key in newsenList:
Prewriteline = Prewriteline + key + ' '

#Post
newsenList = []
for key in senList:
if key in StopWord:
newsenList.append(key)
else:
tmplist = PostSenSeg(key, span)
for keyseg in tmplist:
newsenList.append(keyseg)
Postwriteline = ''
for key in newsenList:
Postwriteline = Postwriteline + key + ' '
Postwriteline = Postwriteline.strip(' ')

w.write(Prewriteline.encode('utf-8') + '\n')
w2.write(Postwriteline.encode('utf-8') + '\n')
f.close()
w.close()
w2.close()

def PreSenSeg(sen, span):
post = span
if len(sen) < span:
post = len(sen)
cur = 0
revlist = []
while 1:
if cur >= len(sen):
break
s = re.search(u"^[0|1|2|3|4|5|6|7|8|9|\uff11|\uff12|\uff13|\uff14|\uff15|\uff16|\uff17|\uff18|\uff19|\uff10|\u4e00|\u4e8c|\u4e09|\u56db|\u4e94|\u516d|\u4e03|\u516b|\u4e5d|\u96f6|\u5341|\u767e|\u5343|\u4e07|\u4ebf|\u5146|\uff2f]+", sen[cur:])
if s:
if s.group() != '':
revlist.append(s.group())
cur = cur + len(s.group())
post = cur + span
if post > len(sen):
post = len(sen)
s = re.search(u"^[a|b|c|d|e|f|g|h|i|j|k|l|m|n|o|p|q|r|s|t|u|v|w|x|y|z|A|B|C|D|E|F|G|H|I|J|K|L|M|N|O|P|Q|R|S|T|U|V|W|X|Y|Z|\uff41|\uff42|\uff43|\uff44|\uff45|\uff46|\uff47|\uff48|\uff49|\uff47|\uff4b|\uff4c|\uff4d|\uff4e|\uff4f|\uff50|\uff51|\uff52|\uff53|\uff54|\uff55|\uff56|\uff57|\uff58|\uff59|\uff5a|\uff21|\uff22|\uff23|\uff24|\uff25|\uff26|\uff27|\uff28|\uff29|\uff2a|\uff2b|\uff2c|\uff2d|\uff2e|\uff2f|\uff30|\uff31|\uff32|\uff33|\uff35|\uff36|\uff37|\uff38|\uff39|\uff3a]+", sen[cur:])
if s:
if s.group() != '':
revlist.append(s.group())
cur = cur + len(s.group())
post = cur + span
if post > len(sen):
post = len(sen)
if (WordDic.has_key(sen[cur:post])) or (cur + 1 == post):
if sen[cur:post] != '':
revlist.append(sen[cur:post])
cur = post
post = post + span
if post > len(sen):
post = len(sen)
else:
post -= 1
return revlist

def PostSenSeg(sen, span):
cur = len(sen)
pre = cur -  span
if pre < 0:
pre = 0
revlist = []
while 1:
if cur <= 0:
break
s = re.search(u"[0|1|2|3|4|5|6|7|8|9|\uff11|\uff12|\uff13|\uff14|\uff15|\uff16|\uff17|\uff18|\uff19|\uff10|\u4e00|\u4e8c|\u4e09|\u56db|\u4e94|\u516d|\u4e03|\u516b|\u4e5d|\u96f6|\u5341|\u767e|\u5343|\u4e07|\u4ebf|\u5146|\uff2f]+$", sen[pre:cur])
if s:
if s.group() != '':
revlist.append(s.group())
cur = cur - len(s.group())
pre = cur - span
if pre < 0:
pre = 0
s = re.search(u"^[a|b|c|d|e|f|g|h|i|j|k|l|m|n|o|p|q|r|s|t|u|v|w|x|y|z|A|B|C|D|E|F|G|H|I|J|K|L|M|N|O|P|Q|R|S|T|U|V|W|X|Y|Z|\uff41|\uff42|\uff43|\uff44|\uff45|\uff46|\uff47|\uff48|\uff49|\uff47|\uff4b|\uff4c|\uff4d|\uff4e|\uff4f|\uff50|\uff51|\uff52|\uff53|\uff54|\uff55|\uff56|\uff57|\uff58|\uff59|\uff5a|\uff21|\uff22|\uff23|\uff24|\uff25|\uff26|\uff27|\uff28|\uff29|\uff2a|\uff2b|\uff2c|\uff2d|\uff2e|\uff2f|\uff30|\uff31|\uff32|\uff33|\uff35|\uff36|\uff37|\uff38|\uff39|\uff3a]+", sen[pre:cur])
if s:
if s.group() != '':
revlist.append(s.group())
cur = cur - len(s.group())
pre = cur - span
if pre < 0:
pre = 0

if (WordDic.has_key(sen[pre:cur])) or (cur - 1 == pre):
if sen[pre:cur] != '':
revlist.append(sen[pre:cur])
cur = pre
pre = pre - span
if pre < 0:
pre = 0
else:
pre += 1
return revlist[::-1]

if __name__ == "__main__":
if len(sys.argv) != 5:
print("Usage: python wordseg.py Dicfile Inputfile Outfile")
Dicfile = sys.argv[1]
Inputfile = sys.argv[2]
Outputfile = sys.argv[3]
Outputfile2 = sys.argv[4]
InitDic(Dicfile)
InitStopword()
WordSeg(Inputfile, Outputfile, Outputfile2)

对比前向后向结果



评测指标

正确率 = 正确识别的个体总数 /  识别出的个体总数

召回率 = 正确识别的个体总数 /  测试集中存在的个体总数

F值  = 正确率 * 召回率 * 2 / (正确率 + 召回率)

评测程序

from __future__ import division
import os
import sys
import linecache
if __name__ == "__main__":
if len(sys.argv) != 3:
print "Usage: python evaluate.py inputfile goldfile"
exit(0)
infile = sys.argv[1]
goldfile = sys.argv[2]
count = 1
count_right = 0
count_split = 0
count_gold = 0
f = file(infile)
for line in f:
inlist = line.strip().decode('utf-8').split(' ')
goldlist = linecache.getline(goldfile, count).strip().decode('utf-8').split(' ')
count += 1
count_split += len(inlist)
count_gold += len(goldlist)
tmp_in = inlist
tmp_gold = goldlist
for key in tmp_in:
if key in tmp_gold:
count_right += 1
tmp_gold.remove(key)
f.close()
print "count_right", count_right
print "count_gold", count_gold
print "count_split", count_split

p = count_right / count_split
r = count_right / count_gold
F = 2 * p * r /(p + r)
print "p:", p
print "r:", r
print "F:", F

结果



 

2. N-gram中文分词

 数据下载：训练数据集+测试+答案+字典+评测程序+N-gram分词

参考代码

def WordSeg(Inputfile, Outputfile):
f = file(Inputfile)
w = file(Outputfile, 'w')
dic_size = 0
for key in StatisticDic:
for keys in StatisticDic[key]:
dic_size += StatisticDic[key][keys]
for line in f:
line = line.strip().decode('utf-8')
senList = []
newsenList = []
tmpword = ''
for i in range(len(line)):
if line[i] in StopWord:
senList.append(tmpword)
senList.append(line[i])
tmpword = ''
else:
tmpword += line[i]
if i == len(line) - 1:
senList.append(tmpword)
#N-gram
for key in senList:
if key in StopWord:
newsenList.append(key)
else:
Pretmplist = PreSenSeg(key, span)
Posttmplist = PostSenSeg(key, span)
tmp_pre = P(Pretmplist, dic_size)
tmp_post = P(Posttmplist, dic_size)
tmplist = []
if tmp_pre > tmp_post:
tmplist = Pretmplist
else:
tmplist = Posttmplist
#print 'tmplist', tmplist
for keyseg in tmplist:
newsenList.append(keyseg)
writeline = ''
for key in newsenList:
writeline = writeline + key + '  '
writeline = writeline.strip('  ')
w.write(writeline.encode('utf-8') + '\n')
#break

f.close()
w.close()

运行

#! -*- coding:utf-8 -*-
from __future__ import division
import sys
import os
import re
StopWordtmp = [' ', u'\u3000',u'\u3001', u'\u300a', u'\u300b', u'\uff1b', u'\uff02', u'\u30fb', u'\u25ce',  u'\x30fb', u'\u3002', u'\uff0c', u'\uff01', u'\uff1f', u'\uff1a', u'\u201c', u'\u201d', u'\u2018', u'\u2019', u'\uff08', u'\uff09', u'\u3010', u'\u3011', u'\uff5b', u'\uff5d', u'-', u'\uff0d', u'\uff5e', u'\uff3b', u'\uff3d', u'\u3014', u'\u3015', u'\uff0e', u'\uff20', u'\uffe5', u'\u2022', u'.']

WordDic = {}
StopWord = []
StatisticDic = {}
span = 16

def InitStopword():
for key in StopWordtmp:
StopWord.append(key)

def InitDic(Dicfile):
f = file(Dicfile)
for line in f:
line = line.strip().decode('utf-8')
WordDic[line] = 1;
f.close()
print len(WordDic)
print "Dictionary has built down!"

def InitStatisticDic(StatisticDicfile):
StatisticDic['<BEG>'] = {}
f = file(StatisticDicfile)
for line in f:
chunk = line.strip().decode('utf-8').split('  ')
if chunk[0] != '':
if not StatisticDic['<BEG>'].has_key(chunk[0]):
StatisticDic['<BEG>'][chunk[0]] = 1
else:
StatisticDic['<BEG>'][chunk[0]] += 1

for i in range(len(chunk) - 1):
if not StatisticDic.has_key(chunk[i]) and chunk[i] != '':
StatisticDic[chunk[i]] = {}
if chunk[i] != '':
if not StatisticDic[chunk[i]].has_key(chunk[i+1]):
StatisticDic[chunk[i]][chunk[i+1]] = 1
else:
StatisticDic[chunk[i]][chunk[i+1]] += 1
if not StatisticDic.has_key(chunk[-1]) and chunk[-1] != '':
StatisticDic[chunk[-1]] = {}
if chunk[-1] != '':
if not StatisticDic[chunk[-1]].has_key('<END>'):
StatisticDic[chunk[-1]]['<END>'] = 1
else:
StatisticDic[chunk[-1]]['<END>'] += 1

def WordSeg(Inputfile, Outputfile):
f = file(Inputfile)
w = file(Outputfile, 'w')
dic_size = 0
for key in StatisticDic:
for keys in StatisticDic[key]:
dic_size += StatisticDic[key][keys]
for line in f:
line = line.strip().decode('utf-8')
senList = []
newsenList = []
tmpword = ''
for i in range(len(line)):
if line[i] in StopWord:
senList.append(tmpword)
senList.append(line[i])
tmpword = ''
else:
tmpword += line[i]
if i == len(line) - 1:
senList.append(tmpword)
#N-gram
for key in senList:
if key in StopWord:
newsenList.append(key)
else:
Pretmplist = PreSenSeg(key, span)
Posttmplist = PostSenSeg(key, span)
tmp_pre = P(Pretmplist, dic_size)
tmp_post = P(Posttmplist, dic_size)
tmplist = []
if tmp_pre > tmp_post:
tmplist = Pretmplist
else:
tmplist = Posttmplist
#print 'tmplist', tmplist
for keyseg in tmplist:
newsenList.append(keyseg)
writeline = ''
for key in newsenList:
writeline = writeline + key + '  '
writeline = writeline.strip('  ')
w.write(writeline.encode('utf-8') + '\n')
#break

f.close()
w.close()

def P(tmplist, dic_size):
rev = 1
if len(tmplist) < 1:
return 0
'''
print 'tmplist', tmplist
print "tmplist[0]", tmplist[0]
print '-----------'
'''
rev *= Pword(tmplist[0], '<BEG>', dic_size)
rev *= Pword('<END>', tmplist[-1], dic_size)
for i in range(len(tmplist)-1):
rev *= Pword(tmplist[i+1], tmplist[i], dic_size)
return rev

def Pword(word1, word2, dic_size):
#print 'word1:', word1
#print 'word2:', word2
div_up = 0
div_down = 0
if StatisticDic.has_key(word2):
for key in StatisticDic[word2]:
#print 'key:', key
div_down += StatisticDic[word2][key]
if key == word1:
div_up = StatisticDic[word2][key]
return (div_up+1) / (div_down + dic_size)

def PreSenSeg(sen, span):
post = span
if len(sen) < span:
post = len(sen)
cur = 0
revlist = []
while 1:
if cur >= len(sen):
break
s = re.search(u"^[0|1|2|3|4|5|6|7|8|9|\uff11|\uff12|\uff13|\uff14|\uff15|\uff16|\uff17|\uff18|\uff19|\uff10|\u4e00|\u4e8c|\u4e09|\u56db|\u4e94|\u516d|\u4e03|\u516b|\u4e5d|\u96f6|\u5341|\u767e|\u5343|\u4e07|\u4ebf|\u5146|\uff2f]+", sen[cur:])
if s:
if s.group() != '':
revlist.append(s.group())
cur = cur + len(s.group())
post = cur + span
if post > len(sen):
post = len(sen)
s = re.search(u"^[a|b|c|d|e|f|g|h|i|j|k|l|m|n|o|p|q|r|s|t|u|v|w|x|y|z|A|B|C|D|E|F|G|H|I|J|K|L|M|N|O|P|Q|R|S|T|U|V|W|X|Y|Z|\uff41|\uff42|\uff43|\uff44|\uff45|\uff46|\uff47|\uff48|\uff49|\uff47|\uff4b|\uff4c|\uff4d|\uff4e|\uff4f|\uff50|\uff51|\uff52|\uff53|\uff54|\uff55|\uff56|\uff57|\uff58|\uff59|\uff5a|\uff21|\uff22|\uff23|\uff24|\uff25|\uff26|\uff27|\uff28|\uff29|\uff2a|\uff2b|\uff2c|\uff2d|\uff2e|\uff2f|\uff30|\uff31|\uff32|\uff33|\uff35|\uff36|\uff37|\uff38|\uff39|\uff3a]+", sen[cur:])
if s:
if s.group() != '':
revlist.append(s.group())
cur = cur + len(s.group())
post = cur + span
if post > len(sen):
post = len(sen)
if (WordDic.has_key(sen[cur:post])) or (cur + 1 == post):
if sen[cur:post] != '':
revlist.append(sen[cur:post])
cur = post
post = post + span
if post > len(sen):
post = len(sen)
else:
post -= 1
return revlist

def PostSenSeg(sen, span):
cur = len(sen)
pre = cur -  span
if pre < 0:
pre = 0
revlist = []
while 1:
if cur <= 0:
break
s = re.search(u"[0|1|2|3|4|5|6|7|8|9|\uff11|\uff12|\uff13|\uff14|\uff15|\uff16|\uff17|\uff18|\uff19|\uff10|\u4e00|\u4e8c|\u4e09|\u56db|\u4e94|\u516d|\u4e03|\u516b|\u4e5d|\u96f6|\u5341|\u767e|\u5343|\u4e07|\u4ebf|\u5146|\uff2f]+$", sen[pre:cur])
if s:
if s.group() != '':
revlist.append(s.group())
cur = cur - len(s.group())
pre = cur - span
if pre < 0:
pre = 0
s = re.search(u"^[a|b|c|d|e|f|g|h|i|j|k|l|m|n|o|p|q|r|s|t|u|v|w|x|y|z|A|B|C|D|E|F|G|H|I|J|K|L|M|N|O|P|Q|R|S|T|U|V|W|X|Y|Z|\uff41|\uff42|\uff43|\uff44|\uff45|\uff46|\uff47|\uff48|\uff49|\uff47|\uff4b|\uff4c|\uff4d|\uff4e|\uff4f|\uff50|\uff51|\uff52|\uff53|\uff54|\uff55|\uff56|\uff57|\uff58|\uff59|\uff5a|\uff21|\uff22|\uff23|\uff24|\uff25|\uff26|\uff27|\uff28|\uff29|\uff2a|\uff2b|\uff2c|\uff2d|\uff2e|\uff2f|\uff30|\uff31|\uff32|\uff33|\uff35|\uff36|\uff37|\uff38|\uff39|\uff3a]+", sen[pre:cur])
if s:
if s.group() != '':
revlist.append(s.group())
cur = cur - len(s.group())
pre = cur - span
if pre < 0:
pre = 0

if (WordDic.has_key(sen[pre:cur])) or (cur - 1 == pre):
if sen[pre:cur] != '':
revlist.append(sen[pre:cur])
cur = pre
pre = pre - span
if pre < 0:
pre = 0
else:
pre += 1
return revlist[::-1]

if __name__ == "__main__":
if len(sys.argv) != 5:
print("Usage: python wordseg.py Dicfile Inputfile Outfile")
Dicfile = sys.argv[1]
StatisticDicfile = sys.argv[2]
Inputfile = sys.argv[3]
Outputfile = sys.argv[4]
InitDic(Dicfile)
InitStatisticDic(StatisticDicfile)

#print "Dic:", StatisticDic
InitStopword()
WordSeg(Inputfile, Outputfile)

分词结果



可以看出结果不如前向切分，但高于后向切分。原因是没有采取平滑策略，利用+1平滑后结果变为



结果超过前向、后向切分，说明有效。

转载：http://www.cnblogs.com/kaituorensheng/p/3629729.html