python机器学习-乳腺癌细胞挖掘（博主亲自录制视频）https://study.163.com/course/introduction.htm?courseId=1005269003&utm_campaign=commission&utm_source=cp-400000000398149&utm_medium=share
机器学习，统计项目合作QQ：231469242
http://blog.****.net/u010718606/article/details/50148261参考
NLTK中对于很多自然语言处理应用有着开箱即用的api，但是结果往往让人弄不清楚状况。
下面的例子使用NLTK进行命名实体的识别。第一例中，Apple成功被识别出来，而第二例并未被识别。究竟是什么原因导致这样的结果，接下来一探究竟。
In [1]: import nltk
In [2]: tokens = nltk.word_tokenize('I am very excited about the next generation of Apple products.')
In [3]: tokens = nltk.pos_tag(tokens)
In [4]: print tokens
[('I', 'PRP'), ('am', 'VBP'), ('very', 'RB'), ('excited', 'JJ'), ('about', 'IN'), ('the', 'DT'), ('next', 'JJ'), ('generation', 'NN'), ('of', 'IN'), ('Apple', 'NNP'), ('products', 'NNS'), ('.', '.')]
In [5]: tree = nltk.ne_chunk(tokens)
In [6]: print tree
(S
I/PRP
am/VBP
very/RB
excited/JJ
about/IN
the/DT
next/JJ
generation/NN
of/IN
(GPE Apple/NNP)
products/NNS
./.)
In [7]: tokens = nltk.word_tokenize('I bought these Apple products today.')
In [8]: tokens = nltk.pos_tag(tokens)
In [9]: print tokens
['I', 'bought', 'these', 'Apple', 'products', 'today', '.']
In [10]: tree = nltk.ne_chunk(tokens)
In [11]: print tree
(S I/PRP bought/VBD these/DT Apple/NNP products/NNS today/NN ./.)最大熵算法
注意到在上述两个例子Apple这个词被词性标注为NNP（NNP是宾夕法尼亚大学树图资料库II为专有名词，单数）。另外，这两个单词都以大写字 母开始。为什么Apple在1例中被标记为GPE（地缘政治实体），而2例未标记？另外，为什么Apple标记为GPE，而不是ORG（组织机构）？
NLTK的命名实体识别是通过使用的MaxEnt分类器。MaxEnt分类器工作有两个原则：1.总是试图保持均匀分布(即最大化熵)，2.保持其 统计概率与经验数据一致。经验数据来源于语料库，通过手动标记，所以大多数标记数据并不是免费的。NLTK不提供其训练命名实体识别器的语料库（训练数据 来自ACE（自动内容抽取））。NLTK所提供的是一个pickle文件（在nltk_data/chunkers/目录下），而这个pickle文件， 就是训练好的MaxEnt分类器实例。
➜ maxent_ne_chunker tree
.
├── english_ace_binary.pickle
├── english_ace_multiclass.pickle
└── PY3
├── english_ace_binary.pickle
└── english_ace_multiclass.pickle要训练良好的监督学习的算法基于良好的特征。在命名实体识别中，特征可能是这个词是否包含一个大写字母。所以NLTK使用的特征有哪些呢？下面我列出他们：
- 词的形状（是否包含数字/首字母大写/包含符号）
- 词的长度
- 词的前三个字母
- 词尾三个字母
- 词性标签
- 词本身
- 该词是否存在
- 该词前面词的词性（前面是否有名词）
- 前词词性
- 后词词性
- 前词本身
- 后词本身
- …
下面的代码可以列出NLTK中所使用的标签
import nltk
# 载入序列化对象
chunker = nltk.data.load('chunkers/maxent_ne_chunker/english_ace_multiclass.pickle')
# 最大熵分类器
maxEnt = chunker._tagger.classifier()
def maxEnt_report():
maxEnt = chunker._tagger.classifier()
print 'These are the labels used by the NLTK\'s NEC...'
print maxEnt.labels()
print ''
print 'These are the most informative features found in the ACE corpora...'
maxEnt.show_most_informative_features()
def ne_report(sentence, report_all=False):
# 词性标记
tokens = nltk.word_tokenize(sentence)
tokens = nltk.pos_tag(tokens)
tags = []
for i in range(0, len(tokens)):
featureset = chunker._tagger.feature_detector(tokens, i, tags)
tag = chunker._tagger.choose_tag(tokens, i, tags)
if tag != 'O' or report_all:
print '\nExplanation on the why the word \'' + tokens[i][0] + '\' was tagged:'
featureset = chunker._tagger.feature_detector(tokens, i, tags)
maxEnt.explain(featureset)
tags.append(tag)下面的输出报告中列出了NLTK所使用的标签，”I-“，”B-“, “O”前缀的含义为包含/开始/例外（inside/begin/others）标记。当一块开始，第一个词是前缀“B”来表示这个词是一个块的开始。下 一个单词，如果它属于同一块，将以”I-“前缀，表示这是块的一部分，而不是开始。如果一个词不属于一块，贴上“O”，这意味着它是在外面。
➜ test python dd.py
These are the labels used by the NLTK's NEC...
['I-GSP', 'B-LOCATION', 'B-GPE', 'I-ORGANIZATION', 'I-PERSON', 'O', 'I-FACILITY', 'I-LOCATION', 'B-PERSON', 'B-FACILITY', 'B-GSP', 'B-ORGANIZATION', 'I-GPE']
These are the most informative features found in the ACE corpora...
10.125 bias==True and label is 'O'
6.631 suffix3=='day' and label is 'O'
-6.207 bias==True and label is 'I-GSP'
5.628 prevtag=='O' and label is 'O'
-4.740 shape=='upcase' and label is 'O'
4.106 shape+prevtag=='<function shape at 0x8bde0d4>+O' and label is 'O'
-3.994 shape=='mixedcase' and label is 'O'
3.992 pos+prevtag=='NNP+B-PERSON' and label is 'I-PERSON'
3.890 prevtag=='I-ORGANIZATION' and label is 'I-ORGANIZATION'
3.879 shape+prevtag=='<function shape at 0x8bde0d4>+I-ORGANIZATION' and label is 'I-ORGANIZATION'Note:
- GPE is Geo-Political Entity
- GSP is Geo-Socio-Political group
例1输出：
Explanation on the why the word 'Apple' was tagged:
Feature B-GPE O B-ORGAN B-GSP
--------------------------------------------------------------------------------
prevtag=='O' (1) 3.767
shape=='upcase' (1) 2.701
pos+prevtag=='NNP+O' (1) 2.254
en-wordlist==False (1) 2.095
label is 'B-GPE' (1) -2.005
bias==True (1) -1.975
prevword=='of' (1) 0.742
pos=='NNP' (1) 0.681
nextpos=='nns' (1) 0.661
prevpos=='IN' (1) 0.311
wordlen==5 (1) 0.113
nextword=='products' (1) 0.060
bias==True (1) 10.125
prevtag=='O' (1) 5.628
shape=='upcase' (1) -4.740
prevpos=='IN' (1) -1.668
label is 'O' (1) -1.075
pos=='NNP' (1) -1.024
suffix3=='ple' (1) 0.797
en-wordlist==False (1) 0.698
wordlen==5 (1) -0.449
prevword=='of' (1) -0.217
nextpos=='nns' (1) 0.104
prefix3=='app' (1) 0.089
pos+prevtag=='NNP+O' (1) 0.011
nextword=='products' (1) 0.005
prevtag=='O' (1) 3.389
pos+prevtag=='NNP+O' (1) 1.725
bias==True (1) 0.955
en-wordlist==False (1) 0.837
label is 'B-ORGANIZATION' (1) 0.718
nextpos=='nns' (1) 0.365
wordlen==5 (1) -0.351
pos=='NNP' (1) 0.174
prevpos=='IN' (1) -0.139
prevword=='of' (1) 0.131
prefix3=='app' (1) -0.126
shape=='upcase' (1) -0.084
suffix3=='ple' (1) -0.077
prevtag=='O' (1) 2.925
pos+prevtag=='NNP+O' (1) 2.213
shape=='upcase' (1) 0.929
en-wordlist==False (1) 0.891
bias==True (1) -0.592
label is 'B-GSP' (1) -0.565
prevpos=='IN' (1) 0.410
nextpos=='nns' (1) 0.399
pos=='NNP' (1) 0.393
prevword=='of' (1) 0.184
wordlen==5 (1) 0.177
---------------------------------------------------------------------------------
TOTAL: 9.406 8.283 7.515 7.366
PROBS: 0.453 0.208 0.122 0.110最后一行中列出的概率加起来加起来是0.893，而非1。这是因为只输出概率最大的四类标签。
例2输出：
Explanation on the why the word 'Apple' was tagged:
Feature O B-GPE B-ORGAN B-LOCAT
--------------------------------------------------------------------------------
bias==True (1) 10.125
prevtag=='O' (1) 5.628
shape=='upcase' (1) -4.740
label is 'O' (1) -1.075
pos=='NNP' (1) -1.024
suffix3=='ple' (1) 0.797
en-wordlist==False (1) 0.698
prevpos=='DT' (1) 0.585
wordlen==5 (1) -0.449
nextpos=='nns' (1) 0.104
prefix3=='app' (1) 0.089
prevword=='these' (1) -0.024
pos+prevtag=='NNP+O' (1) 0.011
nextword=='products' (1) 0.005
prevtag=='O' (1) 3.767
shape=='upcase' (1) 2.701
pos+prevtag=='NNP+O' (1) 2.254
en-wordlist==False (1) 2.095
label is 'B-GPE' (1) -2.005
bias==True (1) -1.975
pos=='NNP' (1) 0.681
nextpos=='nns' (1) 0.661
prevpos=='DT' (1) -0.181
wordlen==5 (1) 0.113
nextword=='products' (1) 0.060
prevtag=='O' (1) 3.389
pos+prevtag=='NNP+O' (1) 1.725
bias==True (1) 0.955
en-wordlist==False (1) 0.837
label is 'B-ORGANIZATION' (1) 0.718
prevpos=='DT' (1) -0.494
nextpos=='nns' (1) 0.365
wordlen==5 (1) -0.351
pos=='NNP' (1) 0.174
prefix3=='app' (1) -0.126
shape=='upcase' (1) -0.084
suffix3=='ple' (1) -0.077
prevword=='these' (1) 0.067
prevtag=='O' (1) 2.682
label is 'B-LOCATION' (1) -2.038
pos+prevtag=='NNP+O' (1) 1.724
shape=='upcase' (1) 1.275
prefix3=='app' (1) 1.169
bias==True (1) 0.747
prevpos=='DT' (1) 0.745
pos=='NNP' (1) 0.616
en-wordlist==False (1) -0.309
nextpos=='nns' (1) 0.151
wordlen==5 (1) 0.041
---------------------------------------------------------------------------------
TOTAL: 10.730 8.171 7.095 6.802
PROBS: 0.697 0.118 0.056 0.046由此：1和2中在GPE识别中唯一的区别在于下面三行：
prevword==’of’ (1) 0.742
prevpos==’IN’ (1) 0.311
prevpos==’DT’ (1) -0.181
可见，1中
1中的Apple被识别为B-GPE，而2中的Apple被识别为O。
引用：
[1] http://www.nltk.org/book/ch07.html
[2] http://spark-public.s3.amazonaws.com/nlp/slides/Information_Extraction_and_Named_Entity_Recognition_v2.pdf
[3] http://www.mattshomepage.com/#/blog/feb2013/liftingthehood
https://study.163.com/provider/400000000398149/index.htm?share=2&shareId=400000000398149（博主视频教学主页）
