def get_radical(word):
radical = Radical(RunOption.Radical)
rs = [radical.trans_ch(ele) for ele in word]
str = ''
for r in rs:
if r is not None:
str += r
return str
def process_text(idx, split_method=None, split_name='train'):
'''
读取文本 切割 然后打上标记 并提取边界、词性、
偏旁部首、拼音等文本特征'''
data = {} #一个很大的字典，包含很多内容
# ////////////////获取句子///////////////////////////
if split_method is None:
with open(f'{train_dir}/{idx}.txt', 'r', encoding='utf-8') as f:
texts = f.readlines()
else:
with open(f'{train_dir}/{idx}.txt', 'r', encoding='utf-8') as f:
texts = f.read() #读取整篇文章
texts = split_method(texts)
data['word'] = texts
# /////////////////获取标签/////////////////////////////
tag_list = ['O' for s in texts for x in s] # 双重循环
# print(tag_list)
tag = pd.read_csv(f'{train_dir}/{idx}.ann', header=None, sep='\t')
for i in range(tag.shape[0]):
tag_item = tag.iloc[i][1].split(' ')
cls, start, end = tag_item[0], int(tag_item[1]), int(tag_item[-1])
tag_list[start] = 'B-' + cls #tag_list的下标开始改动
for j in range(start + 1, end):
tag_list[j] = 'I-' + cls
# return tag_list# 只是弄好了一个全文章的标记，但是我们要做成一句话形式的标记匹配
assert len([x for s in texts for x in s]) == len(tag_list) # 保证两个序列长度一致
# text_list =''
# for t in texts:
# text_list += t
# textes = []
# tags = []
# start = 0
# end = 0
# max = len(tag_list)
# for s in texts:
# l = len(s)
# end += l
# tags.append(tag_list[start:end])
# start +=l
# data['label'] = tags # 做好标签
# # print(tags,texts) #做好了标签与文本的对应关系
#///////////////提取词性和词边界特征/////////////////
word_bounds = ['M' for item in tag_list] #保存词语的分词边界
word_flags = []
for text in texts:
for word, flag in psg.cut(text): # word 中国，flag：ns
if len(word) == 1: #说明是单个词
start = len(word_flags)
word_bounds[start] = 'S'
word_flags.append(flag)
else:
start = len(word_flags)
word_bounds[start] = 'B'
word_flags += [flag] * len(word)
end = len(word_flags) - 1
word_bounds[end] = 'E'
#////////统一截断///////////////
bounds = []
flags = []
tags = []
start = 0
end = 0
for s in texts:
l = len(s)
end += l
bounds.append(word_bounds[start:end])
flags.append(word_flags[start:end])
tags.append(tag_list[start:end])
start += l
data['bound'] = bounds
data['flag'] = flags
data['label'] = tags
# return texts[0], tags[0], bounds[0], flags[0]#此处已经完成了以上四个特征的输出
# /////////////////获取拼音特征和偏旁部首/////////////////////
radical = Radical(RunOption.Radical)
pinyin = Radical(RunOption.Pinyin)
#对于没有偏旁部首的字标上UNK
data['radical'] = [[
radical.trans_ch(x) if radical.trans_ch(x) is not None else 'UNK'
for x in s
] for s in texts]
data['pinyin'] = [[
pinyin.trans_ch(x) if pinyin.trans_ch(x) is not None else 'UNK'
for x in s
] for s in texts]
# return texts[0], tags[0], bounds[0], flags[0], data['radical'][0], data['pinyin'][0]
#数据的几个特征都对应上了。
#//////////////////////////存储数据///////////////////////////////////
num_samples = len(texts)
num_col = len(data.keys())
dataset = []
for i in range(num_samples):
records = list(zip(*[list(v[i]) for v in data.values()]))
#这个符号是解压的意思
dataset += records + [['sep'] * num_col] # 一句话结束后，要用sep进行隔开。
dataset = dataset[:-1] #最后一个sep不要隔开
dataset = pd.DataFrame(dataset, columns=data.keys())
save_path = f'data/{split_name}/{idx}.csv'
def clean_word(w): #现在可以去掉空格，标记已经打好了。
if w == '\n':
return 'LB'
if w in [' ', '\t', '\u2003']:
return 'SPACE'
if w.isdigit():
return 'NUM' #将数字变成统一的数字
return w
dataset['word'] = dataset['word'].apply(clean_word)
dataset.to_csv(save_path, index=False, encoding='utf-8')
from cnradical import Radical, RunOption radical = Radical(RunOption.Radical) pinyin = Radical(RunOption.Pinyin) text = '你好，今天早上吃饭了吗？Eastmount' radical_out = [radical.trans_ch(ele) for ele in text] pinyin_out = [pinyin.trans_ch(ele) for ele in text] print(radical_out) print(pinyin_out) radical_out = radical.trans_str(text) pinyin_out = pinyin.trans_str(text) print(radical_out) print(pinyin_out)
def input_from_line_with_feature(line):
"""
此函数将单一输入句子进行实体识别，构造为具体如下形式
[[[raw_text]], [[word]], [[bound]], [[flag]], [[label]], [[radical]], [[pinyin]]]
这里多一列，到时候输入为[1:]
:param line:输入的单一句子
:param char_to_id:词典转索引
:return:
"""
with open(f'datas/prepare_data/dict.pkl', 'rb') as f:
map_dict = pickle.load(f)
def item2id(data, w2i):
return [w2i[x] if x in w2i else w2i['UNK'] for x in data]
inputs = list()
feature_names = ['word', 'bound', 'flag', 'radical', 'pinyin', 'label']
line = full_to_half(line)
line = replace_html(line)
chars = [[char for char in line]]
# 获取标签，先全部打上O
tag_list = ['O' for _ in line]
# 提取词性和词边界特征
word_bounds = ['M' for _ in tag_list] # 保存每个词的边界
word_flags = [] # 保存词性
# 遍历带词性的切分
for word, flag in psg.cut(line):
# 单个词的时候
if len(word) == 1:
start = len(word_flags)
word_bounds[start] = 'S'
word_flags.append(flag)
else:
start = len(word_flags)
word_bounds[start] = 'B'
word_flags += [flag] * len(word)
# 这里end需要-1
end = len(word_flags) - 1
word_bounds[end] = 'E'
bounds = [word_bounds]
flags = [word_flags]
# 由于是测试将label置为空
targets = [[]]
# 获取偏旁和拼音特征
radical = Radical(RunOption.Radical)
pinyin = Radical(RunOption.Pinyin)
# 这里循环迭代去获取，None的去填充
radicals = [[
radical.trans_ch(x) if radical.trans_ch(x) is not None else 'UNK'
for x in line
]]
pinyins = [[
pinyin.trans_ch(x) if pinyin.trans_ch(x) is not None else 'UNK'
for x in line
]]
inputs.append(chars)
inputs.append(bounds)
inputs.append(flags)
inputs.append(radicals)
inputs.append(pinyins)
inputs.append(targets)
# 开始循环转化为数字索引
id_inputs = [[line]]
for i, feature in enumerate(feature_names):
id_inputs.append([item2id(inputs[i][0], map_dict[feature][2])])
return id_inputs[0][0], id_inputs[1:]
def process_text(idx, split_method=None, split_name='train'):
"""
功能: 读取文本并切割,接着打上标记及提取词边界、词性、偏旁部首、拼音等特征
param idx: 文件的名字 不含扩展名
param split_method: 切割文本方法
param split_name: 存储数据集 默认训练集, 还有测试集
return
"""
#定义字典 保存所有字的标记、边界、词性、偏旁部首、拼音等特征
data = {}
#--------------------------------------------------------------------
# 获取句子
#--------------------------------------------------------------------
if split_method is None:
#未给文本分割函数 -> 读取文件
with open(f'data/{train_dir}/{idx}.txt',
encoding='utf8') as f: #f表示文件路径
texts = f.readlines()
else:
#给出文本分割函数 -> 按函数分割
with open(f'data/{train_dir}/{idx}.txt', encoding='utf8') as f:
outfile = f'data/train_data_pro/{idx}_pro.txt'
print(outfile)
texts = f.read()
texts = split_method(texts, outfile)
#提取句子
data['word'] = texts
print(texts)
#--------------------------------------------------------------------
# 获取标签(实体类别、起始位置)
#--------------------------------------------------------------------
#初始时将所有汉字标记为O
tag_list = ['O' for s in texts for x in s] #双层循环遍历每句话中的汉字
#读取ANN文件获取每个实体的类型、起始位置和结束位置
tag = pd.read_csv(f'data/{train_dir}/{idx}.ann', header=None,
sep='\t') #Pandas读取 分隔符为tab键
#0 T1 Disease 1845 1850 1型糖尿病
for i in range(tag.shape[0]): #tag.shape[0]为行数
tag_item = tag.iloc[i][1].split(' ') #每一行的第二列 空格分割
#print(tag_item)
#存在某些实体包括两段位置区间 仅获取起始位置和结束位置
cls, start, end = tag_item[0], int(tag_item[1]), int(tag_item[-1])
#print(cls,start,end)
#对tag_list进行修改
tag_list[start] = 'B-' + cls
for j in range(start + 1, end):
tag_list[j] = 'I-' + cls
#断言 两个长度不一致报错
assert len([x for s in texts for x in s]) == len(tag_list)
#print(len([x for s in texts for x in s]))
#print(len(tag_list))
#--------------------------------------------------------------------
# 分割后句子匹配标签
#--------------------------------------------------------------------
tags = []
start = 0
end = 0
#遍历文本
for s in texts:
length = len(s)
end += length
tags.append(tag_list[start:end])
start += length
print(len(tags))
#标签数据存储至字典中
data['label'] = tags
#--------------------------------------------------------------------
# 提取词性和词边界
#--------------------------------------------------------------------
#初始标记为M
word_bounds = ['M' for item in tag_list] #边界 M表示中间
word_flags = [] #词性
#分词
for text in texts:
#带词性的结巴分词
for word, flag in psg.cut(text):
if len(word) == 1: #1个长度词
start = len(word_flags)
word_bounds[start] = 'S' #单个字
word_flags.append(flag)
else:
start = len(word_flags)
word_bounds[start] = 'B' #开始边界
word_flags += [flag] * len(word) #保证词性和字一一对应
end = len(word_flags) - 1
word_bounds[end] = 'E' #结束边界
#存储
bounds = []
flags = []
start = 0
end = 0
for s in texts:
length = len(s)
end += length
bounds.append(word_bounds[start:end])
flags.append(word_flags[start:end])
start += length
data['bound'] = bounds
data['flag'] = flags
#--------------------------------------------------------------------
# 获取拼音和偏旁特征
#--------------------------------------------------------------------
radical = Radical(RunOption.Radical) #提取偏旁部首
pinyin = Radical(RunOption.Pinyin) #提取拼音
#提取拼音和偏旁 None用特殊符号替代UNK
radical_out = [[
radical.trans_ch(x) if radical.trans_ch(x) is not None else 'UNK'
for x in s
] for s in texts]
pinyin_out = [[
pinyin.trans_ch(x) if pinyin.trans_ch(x) is not None else 'UNK'
for x in s
] for s in texts]
#赋值
data['radical'] = radical_out
data['pinyin'] = pinyin_out
#--------------------------------------------------------------------
# 存储数据
#--------------------------------------------------------------------
#获取样本数量
num_samples = len(texts) #行数
num_col = len(data.keys()) #列数 字典自定义类别数 6
print(num_samples)
print(num_col)
dataset = []
for i in range(num_samples):
records = list(zip(*[list(v[i]) for v in data.values()])) #压缩
dataset += records + [['sep'] * num_col] #每处理一句话sep分割
#records = list(zip(*[list(v[0]) for v in data.values()]))
#for r in records:
# print(r)
#最后一行sep删除
dataset = dataset[:-1]
#转换成dataframe 增加表头
dataset = pd.DataFrame(dataset, columns=data.keys())
#保存文件 测试集 训练集
save_path = f'data/prepare/{split_name}/{idx}.csv'
dataset.to_csv(save_path, index=False, encoding='utf-8')
#--------------------------------------------------------------------
# 处理换行符 w表示一个字
#--------------------------------------------------------------------
def clean_word(w):
if w == '\n':
return 'LB'
if w in [' ', '\t', '\u2003']: #中文空格\u2003
return 'SPACE'
if w.isdigit(): #将所有数字转换为一种符号 数字训练会造成干扰
return 'NUM'
return w
#对dataframe应用函数
dataset['word'] = dataset['word'].apply(clean_word)
#存储数据
dataset.to_csv(save_path, index=False, encoding='utf-8')
def process_text(idx, split_method=None, split_name='train'): # 用来接收切分好的句子
"""
读取文本---切割---然后打上标记---并提取词边界、词性、部首、拼音、等文本特征
:param idx: 文件的名字,不含扩展名
:param split_method: 切割文本的函数---来自上一个写好的程序
split_name: 最终保存的文件夹名字
:return:
"""
data = {}
#------------------------------切分句子----------------------------------------
if split_method is None: # 如果未将切分函数传给split_method
with open(f'../datas/{train_dir}/{idx}.txt', 'r',
encoding='utf8') as f:
# 打开路径的新写法,读取不同的文件
texts = f.readlines() # 按行,做简单的切割
else:
with open(f'../datas/{train_dir}/{idx}.txt', 'r',
encoding='utf8') as f:
texts = f.read() # 读整篇文章
texts = split_method(texts) # 调用上次的切分函数
data['word'] = texts
# 从one中提取切分好的句子,然后作为一个值赋给data字典里['word']这个键
#------------------------------依据.ann人工标注,首先给 每一个字都打上标签----------------------------------------
# 一、 .txt,对每一个字标记为'O'
tag_list = ['O' for s in texts
for x in s] # s遍历了text里的每一句话,再遍历每一句话里的每一个字---双重循环
# return tag_list # (检查站: main-1用于查看给所有字标记'O')
tag = pd.read_csv(f'../datas/{train_dir}/{idx}.ann', header=None,
sep='\t') # 读取对应的.ann文件
# 二、 获取.ann人工标注中的类别和边界
for i in range(tag.shape[0]): # 这里是做一个文件里的第一行,即一行一行读
tag_item = tag.iloc[i][1].split(' ')
# 取每一行的第二列,获取的实体类别以及起始位置,eg:Disease 18445 1850
# print(tag_item) # (检查站)
cls, start, end = tag_item[0], int(tag_item[1]), int(
tag_item[-1]) # 字符串转换成整数
# 分别抽取出实体类别,起始位置,终止位置,分别保存在cls,start,end里
# 三、 按照人工标注给已经标注为'O'的根据句子长度,标注B,I标签
tag_list[start] = 'B-' + cls # 每一句话中,给起始位置打B+类别名
for j in range(start + 1, end): # 起始位置之后...到最后的结束为止
tag_list[j] = 'I-' + cls # 后面的位置打I+类别名
assert len([x for s in texts for x in s]) == len(tag_list)
# 保证两个序列长度一致,即需要对切分好的每一句话里面的每一个字的长度==打完标记后的每一个字的长度
# texts是切分好的句子,s遍历这些切分好的句子,x再遍历每一个句子的每一个字,
# text_list是整篇文章根据人工标注打的bi标签,对起始位置和结束位置标记好的实体结果
# return tag_list # (检查站:可断点查看)*
#-----------------------------提取词性和词边界特征----------------------------------------
word_bounds = ['M' for item in tag_list
] # 定义词边界,全部标记为M,text_list是对起始位置和结束位置标记好的实体
word_flags = [] # 用来保存每次切好的词,并给下一次切根据长度提供起始位置
for text in texts:
for word, flag in psg.cut(text): # 对每一句话进行带词性切词
if len(word) == 1: # 单个字(词)时
start = len(word_flags) # 确定起始位置,起始位0
word_bounds[start] = 'S' # 对单个的词标注修改为S
word_flags.append(flag) # 把标注好的词添加进word_flags
else: # 当不是单个词时
start = len(word_flags) # 确定起始位置
word_bounds[start] = 'B' # 每个词的第一个字符标注位B
word_flags += [flag] * len(word) # 将这个词的每个字都加上词性标记
end = len(word_flags) - 1 # 确定结束位置
word_bounds[end] = 'E' # 将最后一个字标注位E
#---------统一做,本来放在上面,现在和共用一个循环-------------------------------------------------
# 需要对text_list同样进行切分,切成texts一样长度的句子,从一整篇文章变成一句话一句话的样子
tags = []
bounds = []
flags = []
start = 0
end = 0
for s in texts: # 一句话一句话遍历整个.txt文本的句子
l = len(s) # 计算每个切分好的句子的长度
end += l # 结束位置
bounds.append(
word_bounds[start:end]) # 按照切分好句子的长度对打完标签的txst_list标签进行截取
flags.append(word_flags[start:end]) # 按照切分好句子的长度对打完标签的txst_list标签进行截取
tags.append(tag_list[start:end])
start += l # 定位下一句话的起始位置
data['bound'] = bounds # 词边界特征
data['flag'] = flags # 词性特征
data['label'] = tags
# -----------------------------提取拼音特征----------------------------------------
radical = Radical(RunOption.Radical) # 提取偏旁部首
pinyin = Radical(RunOption.Pinyin) # 提取拼音
# 通过库提取偏旁特征,对没有偏旁的标注位 UNK
data['radical'] = [[
radical.trans_ch(x) if radical.trans_ch(x) is not None else 'UNK'
for x in s
] for s in texts] # 列表推导式
# 提取拼音特征,对没有拼音的标注PAD
data['pinyin'] = [[
pinyin.trans_ch(x) if pinyin.trans_ch(x) is not None else 'UNK'
for x in s
] for s in texts] # 列表推导式
# return texts[1],tags[1],bounds[1],flags[1],data['radical'][1],data['pinyin'][1] # (检查站)
# return (len(data['label'])) #(检查站)
#----------------------------------合并和存储切分好的文件---------------------------------------
num_samples = len(texts) # 有多少句话
num_col = len(data.keys()) # 统计有多少列
dataset = []
for i in range(num_samples):
records = list(
zip(*[list(v[i])
for v in data.values()])) # 将6个列表分别拆分成一个字一个列表,*=解压,无*=压缩
dataset += records + [['sep'] * num_col] # 每存完一个句子需要一行sep进行隔离,sep为元组格式
dataset = dataset[:-1] # 最后一行sep不要
dataset = pd.DataFrame(dataset, columns=data.keys()) # 转换成数据框
save_path = f'../data/prepare/{split_name}/{idx}.csv'
#print(dataset) # (检查站)
def clean_word(w): # 由于已经标注完成,可以清理换行,空格等符号
if w == '\n': # 换行
return 'LB'
if w in [' ', '\t', '\u2003']: # 空格,TAB,中文空格
return 'SPACE'
if w.isdigit(): # 命名实体识别不关心数字是多少,所以把数字变成一个类
return 'num'
return w
dataset['word'] = dataset['word'].apply(clean_word) # 存储之前进行清洗
dataset.to_csv(save_path, index=False, encoding='utf-8')
def process_text(file_id, split_method=None, split_name='train_small'):
"""
本函数作用：用于读取文本、切割、打标记 并提取词边界、词性、偏旁部首和拼音等文本特征。
:param file_name: 文件的名字，不含扩展名（.txt 和 .ann）
:param split_method: 切割文本的方法（是一个函数）
:param split_name: 最终保存的文件夹的名字
:return:
"""
data = {} # 用来存所有的字，标记，边界，偏旁部首，拼音等。
# ****************************************获取句子********************************************
if split_method == None: # 如果没有切割文本的方法，就一行当做一个句子处理。
with open(f'{train_dir}/{file_id}.txt', encoding='utf-8') as f:
texts = f.readlines()
else:
with open(f'{train_dir}/{file_id}.txt', encoding='utf-8') as f:
texts = f.read()
texts = split_method(texts)
# print(texts)
data['word'] = texts
# ****************************************获取标签*********************************************
tag_list = ['O' for sent in texts for word in sent]
tag = pd.read_csv(f'{train_dir}/{file_id}.ann', header=None, sep='\t')
for row in range(tag.shape[0]):
tag_item = tag.iloc[row][1].split(
' ') # 获取实体和实体在文本中的起始位置。 iloc 主要是通过行号获取行数据
entity_cls, tag_begin, tag_end = tag_item[0], int(tag_item[1]), int(
tag_item[-1])
tag_list[tag_begin] = 'B-' + entity_cls # 起始位置写上'B-'
for i in range(tag_begin + 1, tag_end): # 同一个实体类的后面写上'I-'
tag_list[i] = 'I-' + entity_cls
assert len([word for sent in texts
for word in sent]) == len(tag_list) # 保证每个字都对应一个标签。
tag_split_list = []
sen_b = 0
sen_e = 0
for s in texts:
leng = len(s)
sen_e += leng
tag_split_list.append(tag_list[sen_b:sen_e])
sen_b += leng
data['label'] = tag_split_list
#**************************************提取词边界和词性特征*************************************
word_bounds = ['M' for s_tag in tag_list
] # 保存词的边界(B, E, M, S)，刚初始化时，为每一个字都打上'M'的标记。
word_flags = [] # 保存每个字的词性特征。
for sent in texts:
for word, flag in psg.cut(
sent): # 中国,ns 成人,n 2,m 型,k 糖尿病,n HBA1C,eng
if len(word) == 1:
start = len(word_flags)
word_bounds[start] = 'S'
word_flags.append(flag)
else:
start = len(word_flags)
word_bounds[start] = 'B'
word_flags += [flag] * len(word)
end = len(word_flags) - 1
word_bounds[end] = 'E'
bounds = []
flags = []
sen_b = 0
sen_e = 0
for s in texts:
leng = len(s)
sen_e += leng
bounds.append(word_bounds[sen_b:sen_e])
flags.append(word_flags[sen_b:sen_e])
sen_b += leng
data['bound'] = bounds
data['flag'] = flags
# return texts[0], tag_split_list[0], bounds[0], flags[0]
# **************************************获取拼音特征*************************************
radical = Radical(RunOption.Radical)
pinyin = Radical(RunOption.Pinyin)
# 提取每个字的偏旁部首.对于没有偏旁部首的字标上PAD
data['radical'] = radical_out = [[
radical.trans_ch(x) if radical.trans_ch(x) is not None else 'UNK'
for x in s
] for s in texts]
# 提取每个字的拼音.对于没有拼音的字标上PAD
data['pinyin'] = pinyin_out = [[
pinyin.trans_ch(x) if pinyin.trans_ch(x) is not None else 'UNK'
for x in s
] for s in texts]
# return texts[0], tag_split_list[0], bounds[0], flags[0], data['radical'][0], data['pinyin'][0]
# ****************************************存储数据****************************************
num_samples = len(texts) # 获取总共有多少句话
num_col = len(data.keys()) # 获取特征项的个数（列数）
dataset = []
for i in range(num_samples):
records = list(zip(*[list(v[i]) for v in data.values()]))
dataset += records + [['sep'] * num_col] # 每存完一句话，需要一行sep进行隔离。
dataset = dataset[:-1] # 最后一行的分隔符sep不要
dataset = pd.DataFrame(dataset,
columns=data.keys()) # 转换成dataframe（一个表格型的数据结构）
# f-string用大括号 {} 表示被替换字段，其中直接填入替换内容：
save_path = f'./wt_pytorch_Medical_BiLSTM_CRF_NER/data/data_preparation/{split_name}/{file_id}.csv'
def clean_word(word):
if word == '\n':
return 'LB'
if word in [' ', '\t', '\u2003']:
return 'SPACE'
# 把所有数字变成一种符号（命名实体识别的任务不关心是什么数字，只知道是数字就可以了）
# 提高泛化能力。
if word.isdigit():
return 'num'
return word
dataset['word'] = dataset['word'].apply(clean_word)
# dataset.to_csv(save_path, index=False, encoding='utf-8')
dataset.to_csv(save_path, index=False, encoding='utf-8')
def process_text(idx, target_dir='train', split_method=None):
assert target_dir in ['train', 'test'], "数据只分训练集测试集"
data = {}
# ------------------- get word ---------------------------------------------
if not split_method:
with open(f'../data/{target_dir}/{idx}.txt', 'r',
encoding='utf-8') as f:
texts = f.readlines() # 按行分割
else:
with open(f'../data/{target_dir}/{idx}.txt', 'r',
encoding='utf-8') as f:
texts = f.read()
texts = split_method(texts) # 自定义分割方法
data['word'] = texts
# ------------------- get label ----------------------------------------------
tag_list = ['PAD' for s in texts for x in s]
if target_dir == 'train':
tag = pd.read_csv(f'../data/{target_dir}/{idx}.ann',
header=None,
sep='\t')
for i in range(tag.shape[0]):
tag_item = tag.iloc[i][1].split(' ')
cls, start, end = tag_item[0], int(tag_item[1]), int(tag_item[-1])
tag_list[start] = 'B-' + cls
for j in range(start + 1, end):
tag_list[j] = 'I-' + cls
assert len([x for s in texts for x in s]) == len(tag_list)
tags = []
start = 0
end = 0
for item in texts:
l = len(item)
end += l
tags.append(tag_list[start:end])
start += l
data['label'] = tags
# ----------------------------- 词性词边界--------------------------------------------
word_bounds = ['I' for s in texts for x in s]
word_flags = []
# 增加词性与词范围
for text in texts:
for word, flag in jieba.posseg.cut(text):
if len(word) == 1:
start = len(word_flags)
word_bounds[start] = 'S'
word_flags.append(flag)
else:
start = len(word_flags)
word_bounds[start] = 'B'
word_flags += [flag] * len(word)
end = len(word_flags) - 1
word_bounds[end] = 'E'
data['bound'] = []
data['pos_tag'] = []
start = 0
end = 0
for item in texts:
l = len(item)
end += l
data['pos_tag'].append(word_flags[start:end])
data['bound'].append(word_bounds[start:end])
start += l
assert len(word_bounds) == len([x for s in texts for x in s])
assert len(word_flags) == len(word_bounds)
# -------------------------------- 部首 --------------------------------
from cnradical import Radical, RunOption
radical = Radical(RunOption.Radical)
pinyin = Radical(RunOption.Pinyin)
radical_out = [[
radical.trans_ch(ele) if radical.trans_ch(ele) is not None else 'PAD'
for ele in text
] for text in texts]
data['pinyin'] = [[
pinyin.trans_ch(ele) if pinyin.trans_ch(ele) is not None else 'PAD'
for ele in text
] for text in texts]
data['radical'] = radical_out
# --------------------------------------------------------------------------
num_samples = len(texts)
num_col = len(data.keys())
train_file = f'../data/working/{target_dir}/{idx}.csv'
dataset = []
for i in range(num_samples):
records = list(zip(*[list(v[i]) for v in data.values()]))
dataset += records + [['sep'] * num_col]
dataset = dataset[:-1]
dataset = pd.DataFrame(dataset, columns=data.keys())
def clean_word(w):
if w == '\n':
return 'LB'
elif w in [' ', '\t', '\u2003']:
return 'SPACE'
elif w.isdigit():
return 'num'
else:
return w
dataset['word'] = dataset['word'].apply(clean_word)
dataset.to_csv(train_file, sep='\t', index=False, encoding='utf8')
def process_text(idx, split_method=None):
"""
功能: 读取文本并切割,接着打上标记及提取词边界、词性、偏旁部首、拼音等特征
param idx: 文件的名字 不含扩展名
param split_method: 切割文本方法
return
"""
#定义字典 保存所有字的标记、边界、词性、偏旁部首、拼音等特征
data = {}
#--------------------------------------------------------------------
#获取句子
#--------------------------------------------------------------------
if split_method is None:
#未给文本分割函数 -> 读取文件
with open(f'data/{train_dir}/{idx}.txt',
encoding='utf8') as f: #f表示文件路径
texts = f.readlines()
else:
#给出文本分割函数 -> 按函数分割
with open(f'data/{train_dir}/{idx}.txt', encoding='utf8') as f:
outfile = f'data/train_data_pro/{idx}_pro.txt'
print(outfile)
texts = f.read()
texts = split_method(texts, outfile)
#提取句子
data['word'] = texts
print(texts)
#--------------------------------------------------------------------
# 获取标签
#--------------------------------------------------------------------
#初始时将所有汉字标记为O
tag_list = ['O' for s in texts for x in s] #双层循环遍历每句话中的汉字
#读取ANN文件获取每个实体的类型、起始位置和结束位置
tag = pd.read_csv(f'data/{train_dir}/{idx}.ann', header=None,
sep='\t') #Pandas读取 分隔符为tab键
#0 T1 Disease 1845 1850 1型糖尿病
for i in range(tag.shape[0]): #tag.shape[0]为行数
tag_item = tag.iloc[i][1].split(' ') #每一行的第二列 空格分割
#print(tag_item)
#存在某些实体包括两段位置区间 仅获取起始位置和结束位置
cls, start, end = tag_item[0], int(tag_item[1]), int(tag_item[-1])
#print(cls,start,end)
#对tag_list进行修改
tag_list[start] = 'B-' + cls
for j in range(start + 1, end):
tag_list[j] = 'I-' + cls
#断言 两个长度不一致报错
assert len([x for s in texts for x in s]) == len(tag_list)
#print(len([x for s in texts for x in s]))
#print(len(tag_list))
#--------------------------------------------------------------------
# 分割后句子匹配标签
#--------------------------------------------------------------------
tags = []
start = 0
end = 0
#遍历文本
for s in texts:
length = len(s)
end += length
tags.append(tag_list[start:end])
start += length
print(len(tags))
#标签数据存储至字典中
data['label'] = tags
#--------------------------------------------------------------------
# 提取词性和词边界
#--------------------------------------------------------------------
#初始标记为M
word_bounds = ['M' for item in tag_list] #边界 M表示中间
word_flags = [] #词性
#分词
for text in texts:
#带词性的结巴分词
for word, flag in psg.cut(text):
if len(word) == 1: #1个长度词
start = len(word_flags)
word_bounds[start] = 'S' #单个字
word_flags.append(flag)
else:
start = len(word_flags)
word_bounds[start] = 'B' #开始边界
word_flags += [flag] * len(word) #保证词性和字一一对应
end = len(word_flags) - 1
word_bounds[end] = 'E' #结束边界
#存储
bounds = []
flags = []
start = 0
end = 0
for s in texts:
length = len(s)
end += length
bounds.append(word_bounds[start:end])
flags.append(word_flags[start:end])
start += length
data['bound'] = bounds
data['flag'] = flags
#--------------------------------------------------------------------
# 获取拼音特征
#--------------------------------------------------------------------
radical = Radical(RunOption.Radical) #提取偏旁部首
pinyin = Radical(RunOption.Pinyin) #提取拼音
#提取拼音和偏旁 None用特殊符号替代
radical_out = [[
radical.trans_ch(x) if radical.trans_ch(x) is not None else 'PAD'
for x in s
] for s in texts]
pinyin_out = [[
pinyin.trans_ch(x) if pinyin.trans_ch(x) is not None else 'PAD'
for x in s
] for s in texts]
#赋值
data['radical'] = radical_out
data['pinyin'] = pinyin_out
#return texts, tags, bounds, flags
return texts[0], tags[0], bounds[0], flags[0], radical_out[0], pinyin_out[
0]
def process_text(idx, split_method=None, split_name='train'):
"""
读取文本，切割，然后打上标记并提取词边界、词性、偏旁部首、拼音等文本特征
:param idx:文件的名字 不含扩展名
:param split_method:切割文本的方法
:param split_name:判断是保存训练集文件还是测试集
:return:
"""
data = {}
# 获取句子
if split_method is None:
with open(f'./datas/{train_dir}/{idx}.txt', 'r', encoding='utf-8') as f:
texts = f.readlines()
else:
with open(f'./datas/{train_dir}/{idx}.txt', 'r', encoding='utf-8') as f:
texts = f.read()
texts = split_text(texts)
data['word'] = texts
# 获取标签，先全部打上O
tag_list = ['O' for s in texts for _ in s]
# 读取对应的ann文件
tag = pd.read_csv(f'./datas/{train_dir}/{idx}.ann', header=None, sep='\t')
for i in range(tag.shape[0]):
# 获取实体类别以及起始位置
tag_item = tag.iloc[i][1].split(' ')
# 开始打标签
cls, start, end = tag_item[0], int(tag_item[1]), int(tag_item[-1])
# 起始实体打上B
tag_list[start] = 'B-' + cls
# 其他的打上I
for j in range(start + 1, end):
tag_list[j] = 'I-' + cls
# 做检查长度是否相等
assert len([x for s in texts for x in s]) == len(tag_list)
# 提取词性和词边界特征
word_bounds = ['M' for _ in tag_list] # 保存每个词的边界
word_flags = [] # 保存词性
for text in texts:
# 遍历带词性的切分
for word, flag in psg.cut(text):
# 单个词的时候
if len(word) == 1:
start = len(word_flags)
word_bounds[start] = 'S'
word_flags.append(flag)
else:
start = len(word_flags)
word_bounds[start] = 'B'
word_flags += [flag] * len(word)
# 这里end需要-1
end = len(word_flags) - 1
word_bounds[end] = 'E'
# 这里保存词性，统一截断
bounds = []
flags = []
tags = []
start = 0
end = 0
for s in texts:
ldx = len(s)
end += ldx
# 分句子显示
bounds.append(word_bounds[start:end])
flags.append(word_flags[start:end])
tags.append(tag_list[start:end])
start += ldx
data['bound'] = bounds
data['flag'] = flags
data['label'] = tags
# 获取偏旁和拼音特征
radical = Radical(RunOption.Radical)
pinyin = Radical(RunOption.Pinyin)
# 这里循环迭代去获取，None的去填充
data['radical'] = [[radical.trans_ch(x) if radical.trans_ch(x) is not None else 'UNK' for x in s] for s in texts]
data['pinyin'] = [[pinyin.trans_ch(x) if pinyin.trans_ch(x) is not None else 'UNK' for x in s] for s in texts]
# 存储数据
num_samples = len(texts)
num_col = len(data.keys())
dataset = []
# 获取形如('中', 'B', 'ns', 'O', '丨', 'zhōng'), ('国', 'E', 'ns', 'O', '囗', 'guó')
for i in range(num_samples):
recoders = list(zip(*[list(v[i]) for v in data.values()])) # *是解压的意思
# 需要加入隔离符号对其隔离
dataset += recoders + [['sep'] * num_col]
# 最后一个不要
dataset = dataset[:-1]
# 转换成dataframe
dataset = pd.DataFrame(dataset, columns=data.keys())
# csv存储路径
save_path = f'datas/prepare_data/{split_name}/{idx}.csv'
# 现在开始可以处理换行符了
def clean_word(w):
if w == '\n':
return 'LB'
if w in [' ', '\t', '\u2003']:
return 'SPACE'
# 对所有的数字要统一处理
if w.isdigit():
return 'num'
return w
dataset['word'] = dataset['word'].apply(clean_word)
dataset.to_csv(save_path, index=False, encoding='utf-8')