我也说说中文分词（上：基于字符串匹配）-阿里云开发者社区

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L^i-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(w_i|w₁w_2...w_i-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

其中p(wi|wi-1) = p(w_i-1w_i)/p(w_i-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

 
  
 View Code 

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。

本文转自jihite博客园博客，原文链接：http://www.cnblogs.com/kaituorensheng/p/3629729.html，如需转载请自行联系原作者

我也说说中文分词（上：基于字符串匹配）

热门文章

最新文章

相关电子书