解压后取出以下文件:
训练数据:icwb2-data/training/pku_ training.utf8
测试数据:icwb2-data/testing/pku_ test.utf8
正确分词结果:icwb2-data/gold/pku_ test_ gold.utf8
评分工具:icwb2-data/script/socre
2 算法描述
算法是最简单的正向最大匹配(FMM):
用训练数据生成一个字典
对测试数据从左到右扫描,遇到一个最长的词,就切分下来,直到句子结束
注:这是最初的算法,这样做代码可以控制在60行内,后来看测试结果发现没有很好地处理数字问题, 才又增加了对数字的处理。
3 源代码及注释
#! /usr/bin/env python# -*- coding: utf-8 -*- # Author: minix# Date: 2013-03-20 import codecsimport sys # 由规则处理的一些特殊符号numMath = [u'0', u'1', u'2', u'3', u'4', u'5', u'6', u'7', u'8', u'9']numMath_suffix = [u'.', u'%', u'亿', u'万', u'千', u'百', u'十', u'个']numCn = [u'一', u'二', u'三', u'四', u'五', u'六', u'七', u'八', u'九', u'〇', u'零']numCn_suffix_date = [u'年', u'月', u'日']numCn_suffix_unit = [u'亿', u'万', u'千', u'百', u'十', u'个']special_char = [u'(', u')'] def proc_num_math(line, start): """ 处理句子中出现的数学符号 """ oldstart = start while line[start] in numMath or line[start] in numMath_suffix: start = start + 1 if line[start] in numCn_suffix_date: start = start + 1 return start - oldstart def proc_num_cn(line, start): """ 处理句子中出现的中文数字 """ oldstart = start while line[start] in numCn or line[start] in numCn_suffix_unit: start = start + 1 if line[start] in numCn_suffix_date: start = start + 1 return start - oldstart def rules(line, start): """ 处理特殊规则 """ if line[start] in numMath: return proc_num_math(line, start) elif line[start] in numCn: return proc_num_cn(line, start) def genDict(path): """ 获取词典 """ f = codecs.open(path,'r','utf-8') contents = f.read() contents = contents.replace(u'r', u'') contents = contents.replace(u'n', u'') # 将文件内容按空格分开 mydict = contents.split(u' ') # 去除词典List中的重复 newdict = list(set(mydict)) newdict.remove(u'') # 建立词典 # key为词首字,value为以此字开始的词构成的List truedict = {} for item in newdict: if len(item)>0 and item[0] in truedict: value = truedict[item[0]] value.append(item) truedict[item[0]] = value else: truedict[item[0]] = [item] return truedict def print_unicode_list(uni_list): for item in uni_list: print item, def divideWords(mydict, sentence): """ 根据词典对句子进行分词, 使用正向匹配的算法,从左到右扫描,遇到最长的词, 就将它切下来,直到句子被分割完闭 """ ruleChar = [] ruleChar.extend(numCn) ruleChar.extend(numMath) result = [] start = 0 senlen = len(sentence) while start < senlen: curword = sentence[start] maxlen = 1 # 首先查看是否可以匹配特殊规则 if curword in numCn or curword in numMath: maxlen = rules(sentence, start) # 寻找以当前字开头的最长词 if curword in mydict: words = mydict[curword] for item in words: itemlen = len(item) if sentence[start:start+itemlen] == item and itemlen > maxlen: maxlen = itemlen result.append(sentence[start:start+maxlen]) start = start + maxlen return result def main(): args = sys.argv[1:] if len(args) < 3: print 'Usage: python dw.py dict_path test_path result_path' exit(-1) dict_path = args[0] test_path = args[1] result_path = args[2] dicts = genDict(dict_path) fr = codecs.open(test_path,'r','utf-8') test = fr.read() result = divideWords(dicts,test) fr.close() fw = codecs.open(result_path,'w','utf-8') for item in result: fw.write(item + ' ') fw.close() if __name__ == "__main__": main()
4 测试及评分结果
使用 dw.py 训练数据 测试数据, 生成结果文件
使用 score 根据训练数据,正确分词结果,和我们生成的结果进行评分
使用 tail 查看结果文件最后几行的总体评分,另外socre.utf8中还提供了大量的比较结果, 可以用于发现自己的分词结果在哪儿做的不够好
注:整个测试过程都在Ubuntu下完成
$ python dw.py pku_training.utf8 pku_test.utf8 pku_result.utf8
$ perl score pku_training.utf8 pku_test_gold.utf8 pku_result.utf8 > score.utf8
$ tail -22 score.utf8
INSERTIONS: 0
DELETIONS: 0
SUBSTITUTIONS: 0
NCHANGE: 0
NTRUTH: 27
NTEST: 27
TRUE WORDS RECALL: 1.000
TEST WORDS PRECISION: 1.000
=== SUMMARY:
=== TOTAL INSERTIONS: 4623
=== TOTAL DELETIONS: 1740
=== TOTAL SUBSTITUTIONS: 6650
=== TOTAL NCHANGE: 13013
=== TOTAL TRUE WORD COUNT: 104372
=== TOTAL TEST WORD COUNT: 107255
=== TOTAL TRUE WORDS RECALL: 0.920
=== TOTAL TEST WORDS PRECISION: 0.895
=== F MEASURE: 0.907
=== OOV Rate: 0.940
=== OOV Recall Rate: 0.917
=== IV Recall Rate: 0.966
基于词典的FMM算法是非常基础的分词算法,效果没那么好,不过足够简单,也易于入手,随着学习的深入,我可能还会用Python实现其它的分词算法。另外一个感受是,看书的时候尽量多去实现,这样会让你有足够的热情去关注理论的每一个细节,不会感到那么枯燥无力。