学习笔记——基于条件随机场的医疗电子病例的命名实体识别

from sklearn import preprocessing
feature = [[0,1], [1,1], [0,0], [1,0]]
label= ['yes', 'no', 'yes', 'no']
lb = preprocessing.LabelBinarizer() #构建一个转换对象
Y = lb.fit_transform(label)
re_label = lb.inverse_transform(Y)
print(Y)
print(re_label)

(一)代码使用方法及介绍
基于条件随机场的医疗电子病例的命名实体识别


medical_ner_crfsuite是[CCKS2017全国知识图谱与语义大会](http://www.ccks2017.com/),医疗电子病例命名实体识别评测任务的一个可执行demo,采用的方法是条件随机场(CRF),实现CRF的第三方库为[python-crfsuite](https://github.com/scrapinghub/python-crfsuite)。目前该demo准确率为68%,召回率为62%,F1值为64.8%。

 (二)MOODLE


1. 数据预处理。调用reader.py中的text2nerformat方法,将data中的数据集转换成NER任务中常用的数据格式


import jieba.posseg
import re
import codecs




htmltag = ['症状和体征', '检查和检验', '治疗', '疾病和诊断', '身体部位']
englishtag = ['SYMPTOM', 'CHECK', 'TREATMENT', 'DISEASE', 'BODY']


def readFileUTF8(filename):
    fr = codecs.open(filename, 'r', 'utf-8')
    text = ''
    for line in fr:
        text += line.strip()
    return text


def extract_tag_information(text):
    res = {}
    for i, html in enumerate(htmltag):
        res[englishtag[i]] = []
        pattern = re.compile(r'<' + html + '>(.*?)', re.S)
        contents = pattern.findall(text)
        for content in contents:
            content = re.compile('<[^>]+>', re.S).sub('', content)
            res[englishtag[i]].append(content)
    return res


def extract_all_information(text):
    pattern = re.compile('<(.*?)>(.*?)', re.S)
    contents = pattern.findall(text)
    ans = ''
    for content in contents:
        content = re.compile('<[^>]+>', re.S).sub('', content[1])
        ans += content
        print (content)
    return ans


def getType(type):
    if type == '症状和体征':
        return 'SIGNS'
    elif type == '检查和检验':
        return 'CHECK'
    elif type == '疾病和诊断':
        return 'DISEASE'
    elif type == '治疗':
        return 'TREATMENT'
    elif type == '身体部位':
        return 'BODY'
    else:
        return 'OTHER'


def split(text):
    """以标签数据分割成list"""
    res = []
    start = 0
    end = 0
    while end < len(text):
        if text[end] == '<':
            # < 前面的信息写入
            if start != end:
                res.append(text[start: end])
                start = end + 1
            else:
                start += 1
            # <>中的信息
            end = go(text, start)
            res.append(text[start: end])
            start = end + 1
            end = start
        else:
            end += 1
    if start != end:
        res.append(text[start: end])
    return res


# 将标签数据集转换成ner格式的标准数据集
def text2nerformat(text):
    # 过滤掉所有的标签
    # content = re.compile('<[^>]+>', re.S).sub('', text)
    segment = jieba.posseg.cut(text)
    # 采用BIOSE方式
    # B: 开始,I:中间,O:无关词,S:单个词,E:结尾
    # 将训练数据转换为标准的ner格式的数据
    start = 0
    type = ''
    stack = []
    flag = 0
    features = []
    pieces = split(text)
    pre = 0
    for seg in segment:
        if seg.word == '<':
            flag = 1
            pre = 0
            continue
        elif seg.word == '>':
            flag = 0
            pre = 0
            continue

        if flag == 0:
            while start < len(pieces) and getType(pieces[start]) != 'OTHER':
                stack.append(getType(pieces[start]))
                start += 1
            while start < len(pieces) and getType(pieces[start][1:]) != 'OTHER':
                stack.pop()
                start += 1
            while start < len(pieces) and getType(pieces[start]) != 'OTHER':
                stack.append(getType(pieces[start]))
                start += 1
            index = pieces[start].find(seg.word, pre)
            pre = index + 1
            if len(stack) == 0:
                type = 'O'
                if start < len(pieces) and index + len(seg.word) == len(pieces[start]):
                    start += 1
            else:
                if start < len(pieces):
                    if index == 0 and len(seg.word) == len(pieces[start]):
                        type = 'S-' + stack[-1]
                        start += 1
                    elif index == 0 and len(seg.word) != len(pieces[start]):
                        type = 'B-' + stack[-1]
                    elif index != -1 and len(pieces[start]) - index == len(seg.word):
                        if start + 1 == len(pieces) or getType(pieces[start + 1]) == 'OTHER':
                            type = 'E-' + stack[-1]
                        else:
                            type = 'I-' + stack[-1]
                        start += 1
                    elif index != -1:
                        type = 'I-' + stack[-1]

            features.append([seg.word, seg.flag, type])
            # print '%s, %s, %s' % (seg.word, seg.flag, type)
    return features


def go(text, i):
    while i < len(text):
        if text[i] == '>':
            break
        else:
            i += 1
    return i


# 将标注过的ner数据集,提取出实体
def getNamedEntity(word, ner):
    ans = []
    cur = ''
    for i, tag in enumerate(ner):
        if 'B' == tag.split('-')[0]:
            cur += word[i]
        elif 'I' == tag.split('-')[0]:
            cur += word[i]
        elif 'E' == tag.split('-')[0]:
            cur += word[i]
            ans.append(cur)
            cur = ''
        elif 'S' == tag.split('-')[0]:
            if len(cur) == 0:
                ans.append(word[i])
            else:
                cur += word[i]
    return ans


if __name__ == '__main__':
    # fw = file('test1.txt', 'w')
    # for i in range(100, 101):
    #     filename = 'data/病史特点-' + str(i) + '.txt'
    #     answer = text2nerformat(readFileUTF8(filename))
    #     for [word, pos, ner] in answer:
    #         fw.write(word + '\t' + pos + '\t' + ner + '\n')
    #     print 'file ' + str(i) + ' has already finished!'
    # fw.flush()
    # fw.close()
    fr = codecs.open('test1.txt', 'r', 'utf-8')
    data = []
    for line in fr:
        fields = line.strip().split('\t')
        if len(fields) == 3:
            data.append(fields)
    word = [w for w, tag, label in data]
    ner = [label for w, tag, label in data]
    ans = getNamedEntity(word, ner)
    for a in ans:
        print (a)

(1)python模块之codecs: 自然语言编码转换:

 如果我们处理的文件里的字符编码是其他类型的呢?这个读取进行做处理也需要特 殊的处理的。codecs也提供了方法.

学习笔记——基于条件随机场的医疗电子病例的命名实体识别_第1张图片

 参考文章: python模块之codecs: 自然语言编码转换_Mingz技术博客-CSDN博客

在使用NLPIR分词的时候,对输入文档的编码格式是有严格要求的,在函数初始化的时候可以设置输入源文档的编码格式。

但是源文档的编码可能一会儿是utf-8一会儿是gbk,这就要求统一一下格式,不能格式一乱就报错了,

参考文章:   https://blog.csdn.net/chixujohnny/article/details/51782826

(2)open

关于open相关内容请参考博主这篇文章

参考链接:https://blog.csdn.net/weixin_51130521/article/details/119614510

(3)line.strip()

line.strip()会把'\\n'(空行)替换为''

参考链接:https://blog.csdn.net/u010565244/article/details/19193635

(4)enumerate

enumerate()使用

  • 如果对一个列表,既要遍历索引又要遍历元素时,首先可以这样写:

  • 上述方法有些累赘,利用enumerate()会更加直接和优美:

  • list1 = ["这", "是", "一个", "测试"]
    for index, item in enumerate(list1):
        print index, item
    >>>
    0 这
    1 是
    2 一个
    3 测试
  • enumerate还可以接收第二个参数,用于指定索引起始值,如:
  • list1 = ["这", "是", "一个", "测试"]
    for index, item in enumerate(list1, 1):
        print index, item
    >>>
    1 这
    2 是
    3 一个
    4 测试

    参考链接:https://blog.csdn.net/churximi/article/details/51648388

(5)res[englishtag[i]] = []

学习笔记——基于条件随机场的医疗电子病例的命名实体识别_第2张图片

(6)re.compile

正则表达式re.compile()

compile()与findall()一起使用,返回一个列表。

eg:

import re


def main():
    content = 'Hello, I am Jerry, from Chongqing, a montain city, nice to meet you……'
    regex = re.compile('\w*o\w*')
    x = regex.findall(content)
    print(x)


if __name__ == '__main__':
    main()
# ['Hello', 'from', 'Chongqing', 'montain', 'to', 'you']

参考链接:https://blog.csdn.net/Darkman_EX/article/details/80973656

Python中正则表达式re.S的作用

import re
a = """sdfkhellolsdlfsdfiooefo:
877898989worldafdsf"""
b = re.findall('hello(.*?)world',a)
c = re.findall('hello(.*?)world',a,re.S)
print ('b is ' , b)
print ('c is ' , c)
 
 
# 输出结果:
# b is  []
# c is  ['lsdlfsdfiooefo:\n877898989']

注意:只有三单引或者三双引号的情况下,可以直接回车(\n)换行写。其他双引号,单引号写法不同。这里不做其他解释。

在字符串a中,包含换行符\n,在这种情况下:

如果不使用re.S参数,则只在每一行内进行匹配,如果一行没有,就换下一行重新开始。

使用re.S参数以后,正则表达式会将这个字符串作为一个整体,在整体中进行匹配

原文链接:https://blog.csdn.net/weixin_42781180/article/details/81302806

re.compile与sub

import re 
eliminate = re.compile('[!"#$%&\'()*+,-./:;<=>?@[\\]^_`{|}~\s]+')
text = eliminate.sub("@", "你!好!") 
text
 
#  输出:'你@好@'
 
text = eliminate.sub("", "你!好!") 
text
# 输出:'你好'

如图所示:compile与sub组合替换掉已经匹配的字符串

 2,训练模型。通过crf_unit.py,训练CRF模型,目前CRF中的特征包括上下两个词语及其词性,分词和词性标注调用[jieba](https://github.com/fxsjy/jieba)


import codecs
import pycrfsuite
import string
import zhon.hanzi as zh
from sklearn.model_selection import ShuffleSplit, cross_val_score
from sklearn.neighbors import KNeighborsClassifier

import reader
from sklearn.metrics import classification_report
from sklearn.preprocessing import LabelBinarizer



# 获取数据
def readData(filename):
    fr = codecs.open(filename, 'r', 'utf-8')
    data = []
    for line in fr:
        fields = line.strip().split('\t')
        if len(fields) == 3:
            data.append(fields)
    return data


train = readData('train.txt')
test = readData('test.txt')


# 判断是否为标点符号
# punctuation
def ispunctuation(word):
    punctuation = string.punctuation + zh.punctuation
    if punctuation.find(word) != -1:
        return True
    else:
        return False


# 特征定义
def word2features(sent, i):
    """返回特征列表"""
    word = sent[i][0]
    postag = sent[i][1]
    features = [
        'bias',
        'word=' + word,
        'word_tag=' + postag,
    ]
    if i > 0:
        features.append('word[-1]=' + sent[i-1][0])
        features.append('word[-1]_tag=' + sent[i-1][1])
        if i > 1:
            features.append('word[-2]=' + sent[i-2][0])
            features.append('word[-2, -1]=' + sent[i-2][0] + sent[i-1][0])
            features.append('word[-2]_tag=' + sent[i-2][1])
    if i < len(sent) - 1:
        features.append('word[1]=' + sent[i+1][0])
        features.append('word[1]_tag=' + sent[i+1][1])
        if i < len(sent) - 2:
            features.append('word[2]=' + sent[i+2][0])
            features.append('word[1, 2]=' + sent[i+1][0] + sent[i+2][0])
            features.append('word[2]_tag=' + sent[i+2][1])
    return features


def sent2feature(sent):
    return [word2features(sent, i) for i in range(len(sent))]


def sent2label(sent):
    return [label for word, tag, label in sent]


def sent2word(sent):
    return [word for word, tag, label in sent]


X_train = sent2feature(train)
y_train = sent2label(train)

X_test = sent2feature(test)
y_test = sent2label(test)

# 训练模型
model = pycrfsuite.Trainer(verbose=True)
model.append(X_train, y_train)
model.set_params({
    'c1': 1.0,  # coefficient for L1 penalty
    'c2': 1e-3,  # coefficient for L2 penalty
    'max_iterations': 100,  # stop earlier
    # include transitions that are possible, but not observed
    'feature.possible_transitions': True,
    'feature.minfreq': 3
})

model.train('./medical.crfsuite')


# 预测数据
tagger = pycrfsuite.Tagger()
tagger.open('./medical.crfsuite')

# 一份测试数据集
print (' '.join(sent2word(readData('test1.txt'))))
predicted = tagger.tag(sent2feature(readData('test1.txt')))
correct = sent2label(readData('test1.txt'))

# 预测结果对比
print ('Predicted: ', ' '.join(predicted))
print ('Correct: ', ' '.join(correct))

# 预测准确率
num = 0
for i, tag in enumerate(predicted):
    if tag == correct[i]:
        num += 1
print( 'accuracy: ', num * 1.0 / len(predicted))


# 实体抽取结果
ans = reader.getNamedEntity(sent2word(readData('test1.txt')), predicted)
for a in ans:
    print (a)

(1)import zhon.hanzi as zh

见博主这篇文章:https://blog.csdn.net/weixin_51130521/article/details/119729670

(2)re.sub

该函数主要用于替换字符串中的匹配项

本文链接:https://blog.csdn.net/jackandsnow/article/details/103885422

(3)Python中的 .join()用法

Python中的 .join() 函数经常被大家使用到,之前面试的时候也被问到过,在这里记录一下:

这个函数展开来写应该是str.join(item),join函数是一个字符串操作函数

str表示字符串(字符),item表示一个成员,注意括号里必须只能有一个成员,比如','.join('a','b')这种写法是行不通的

举个例子:

','.join('abc')

上面代码的含义是“将字符串abc中的每个成员以字符','分隔开再拼接成一个字符串”,输出结果为:

'a,b,c'

join里放列表、元组、字典也是可以的

';'.join([a,b,c])
>>  'a;b;c'

(4) sklearn.preprocessing.LabelBinarizer()的用法

对于标称型数据来说,preprocessing.LabelBinarizer是一个很好用的工具。比如可以把yes和no转化为0和1,或是把incident和normal转化为0和1。当然,对于两类以上的标签也是适用的。这里举一个简单的例子,说明将标签二值化以及其逆过程。

from sklearn import preprocessing
feature = [[0,1], [1,1], [0,0], [1,0]]
label= ['yes', 'no', 'yes', 'no']
lb = preprocessing.LabelBinarizer() #构建一个转换对象
Y = lb.fit_transform(label)
re_label = lb.inverse_transform(Y)
print(Y)
print(re_label)

 结果

[[1]
 [0]
 [1]
 [0]]
['yes' 'no' 'yes' 'no']

3,评估模型。调用crf_unit.py中的bio_classification_report方法,评估模型。


'''''
# 评估模型
用这个会报错  Recall and F-score are ill-defined and being set to 0.0 in samples with 
no true labels. Use `zero_division` parameter to control this behavior.  
 _warn_prf(average, modifier, msg_start, len(result))
def bio_classification_report(y_true, y_pred):
    """
    Classification report for a l ist of BIOSE-encoded sequences.
    It computes token-level metrics and discards 'O' labels.
    :param y_true:
    :param y_pred:
    :return:
    """
    lb = LabelBinarizer()
    y_true_combined = lb.fit_transform(y_true)
    y_pred_combined = lb.transform(y_pred)

    tagset = set(lb.classes_) - {'O'}
    tagset = sorted(tagset, key=lambda tag: tag.split('-', 1)[::-1])
    class_indices = {
        cls: idx for idx, cls in enumerate(lb.classes_)
    }

    return classification_report(
        y_true_combined,
        y_pred_combined,
        labels=[class_indices[cls] for cls in tagset],
        target_names=tagset
    )


y_pred = list(tagger.tag(X_test))
print (bio_classification_report(y_test, y_pred))
'''''
def knn(self,X_train,X_test,Y_train,Y_test):

   #implementación del algoritmo
   knn = KNeighborsClassifier(n_neighbors=3).fit(X_train,Y_train)
   #10XV
   cv = ShuffleSplit(n_splits=10, test_size=0.3, random_state=0)
   puntajes = sum(cross_val_score(knn, X_test, Y_test,
                                        cv=cv,scoring='f1_weighted'))/10

   print(puntajes)


参考链接

immense8342/medical_ner_crfsuite: 基于条件随机场的医疗电子病例的命名实体识别 (github.com)

你可能感兴趣的:(python,自然语言处理,机器学习,sklearn,scikit-learn)