CMeKG代码解读(以项目为导向从零开始学习知识图谱)（五）
书接上文：CMeKG代码解读(以项目为导向从零开始学习知识图谱)（四）_chen_nnn的博客-CSDN博客作者从零开始学习和知识图谱有关技术和内容，而本文的核心内容是对CMeKG的python代码进行学习和解读，供大家讨论参考共同进步。CMeKG（Chinese Medical Knowledge Graph）是利用自然语言处理与文本挖掘技术，基于大规模医学文本数据，以人机结合的方式研发的中文医学知识图谱。https://blog.csdn.net/chen_nnn/article/details/122861121
新的一个python文件，继续加油！！！
目录
medical_ner.py
medical_ner类：
from_input():
from_text():
spilt_entity_input():
predict_sentence():
predict_file():
结语
medical_ner.py
import codecs
import torch
from torch.autograd import Variable
from torch.utils.data import TensorDataset
from torch.utils.data import DataLoader
import json
from utils import load_vocab
from ner_constant import *
from model_ner import BERT_LSTM_CRF
import os从命名方式易知，该文件是用来进行医学实体识别的。
medical_ner类：
class medical_ner(object):
def __init__(self):
self.NEWPATH = '/Users/yangyf/workplace/model/medical_ner/model.pkl'
self.vocab = load_vocab('/Users/yangyf/workplace/model/medical_ner/vocab.txt')
self.vocab_reverse = {v: k for k, v in self.vocab.items()}
self.model = BERT_LSTM_CRF('/Users/yangyf/workplace/model/medical_ner', tagset_size, 768, 200, 2,
dropout_ratio=0.5, dropout1=0.5, use_cuda=use_cuda)
if use_cuda:
self.model.to(device)和前面的医学文本分词类似，首先是加载命名实体识别的字典，然后使用BERT_LSTM_CRF模型进行训练，并且询问cuda是否准备好，准备则在cuda上运行。
from_input():
def from_input(self, input_str):
raw_text = []
textid = []
textmask = []
textlength = []
text = ['[CLS]'] + [x for x in input_str] + ['[SEP]']
raw_text.append(text)
cur_len = len(text)
# raw_textid = [self.vocab[x] for x in text] + [0] * (max_length - cur_len)
raw_textid = [self.vocab[x] for x in text if self.vocab.__contains__(x)] + [0] * (max_length - cur_len)
textid.append(raw_textid)
raw_textmask = [1] * cur_len + [0] * (max_length - cur_len)
textmask.append(raw_textmask)
textlength.append([cur_len])
textid = torch.LongTensor(textid)
textmask = torch.LongTensor(textmask)
textlength = torch.LongTensor(textlength)
return raw_text, textid, textmask, textlength和医学文本分词的单行输入的情况都一样，前期处理都类似。
from_text():
def from_txt(self, input_path):
raw_text = []
textid = []
textmask = []
textlength = []
with open(input_path, 'r', encoding='utf-8') as f:
for line in f.readlines():
if len(line.strip())==0:
continue
if len(line) > 448:
line = line[:448]
temptext = ['[CLS]'] + [x for x in line[:-1]] + ['[SEP]']
cur_len = len(temptext)
raw_text.append(temptext)
tempid = [self.vocab[x] for x in temptext[:cur_len]] + [0] * (max_length - cur_len)
textid.append(tempid)
textmask.append([1] * cur_len + [0] * (max_length - cur_len))
textlength.append([cur_len])
textid = torch.LongTensor(textid)
textmask = torch.LongTensor(textmask)
textlength = torch.LongTensor(textlength)
return raw_text, textid, textmask, textlength同样和医学文本分词的多行输入处理函数一样，只不过在此函数中能处理的文本长度更长。（我怀疑是不是敲错了，一位1和4还是蛮近的）
spilt_entity_input():
def split_entity_input(self,label_seq):
entity_mark = dict()
entity_pointer = None
for index, label in enumerate(label_seq):
#print(f"before: {label_seq}")
if label.split('-')[-1]=='B':
category = label.split('-')[0]
entity_pointer = (index, category)
entity_mark.setdefault(entity_pointer, [label])
elif label.split('-')[-1]=='M':
if entity_pointer is None: continue
if entity_pointer[1] != label.split('-')[0]: continue
entity_mark[entity_pointer].append(label)
elif label.split('-')[-1]=='E':
if entity_pointer is None: continue
if entity_pointer[1] != label.split('-')[0]: continue
entity_mark[entity_pointer].append(label)
else:
entity_pointer = None
# print(entity_mark)
return entity_mark对输入进函数的文本序列首先进行enumerate标号处理，然后根据标签判断是否是‘B’、‘M’和‘E’中的哪一种（B stands for 'beginning' (signifies beginning of an NE)、M stands for 'middle' (signifies that the word is inside an NE)、E stands for 'end' (signifies that the word is the end of an NE)）。如果判断标签里面设定的是B，则记录这个开始标签的类别并放到实体指针当中，然后按照指针对应标签内容的方式存储在字典内；如果判断为M，则再次判断是否有已经存在的实体指针，如果没有实体指针，则继续循环，如果有实体指针，再判定实体指针的类别是否和该标签的内的类别一致，如果一致则按照实体指针存储在entity_mark字典中，不一致继续循环遍历；如果判断为E，和判断为M的操作相同，最终返回entity_mark字典。而从函数中我们也可以大概推断一下label_seq里面的存储方式，最后一个元素用来表示文本的内容处于实体内什么样的位置，第一个元素存储了该实体的类别。
predict_sentence():
def predict_sentence(self, sentence):
tag_dic = {"d": "疾病", "b": "身体", "s": "症状", "p": "医疗程序", "e": "医疗设备", "y": "药物", "k": "科室",
"m": "微生物类", "i": "医学检验项目"}
if sentence == '':
print("输入为空！请重新输入")
return
if len(sentence) > 448:
print("输入句子过长，请输入小于148的长度字符！")
sentence = sentence[:448]
raw_text, test_ids, test_masks, test_lengths = self.from_input(sentence)
test_dataset = TensorDataset(test_ids, test_masks, test_lengths)
test_loader = DataLoader(test_dataset, shuffle=False, batch_size=1)
self.model.load_state_dict(torch.load(self.NEWPATH, map_location=device))
self.model.eval()
for i, dev_batch in enumerate(test_loader):
sentence, masks, lengths = dev_batch
batch_raw_text = raw_text[i]
sentence, masks, lengths = Variable(sentence), Variable(masks), Variable(lengths)
if use_cuda:
sentence = sentence.to(device)
masks = masks.to(device)
predict_tags = self.model(sentence, masks)
predict_tags.tolist()
predict_tags = [i2l_dic[t.item()] for t in predict_tags[0]]
predict_tags = predict_tags[:len(batch_raw_text)]
pred = predict_tags[1:-1]
raw_text = batch_raw_text[1:-1]
entity_mark = self.split_entity_input(pred)
entity_list = {}
if entity_mark is not None:
for item, ent in entity_mark.items():
# print(item, ent)
entity = ''
index, tag = item[0], item[1]
len_entity = len(ent)
for i in range(index, index + len_entity):
entity = entity + raw_text[i]
entity_list[tag_dic[tag]] = entity
# print(entity_list)
return entity_list该函数的前半部分是将上一篇文章当中的predict_sentence（）和recover_to_text（）两个函数结合起来，因此前半部分作用也都一样。然后是对预测出来的pred列表放到上一个spilt_entity_input()函数中进行切分和组合处理，如果最后切分出来的字典非空则针对其中的内容进行提取处理。从该函数中也可以看出，上一个函数返回的字典中所包含的内容形式，包含有实体在文本中的位置，实体的类型，还有实体本身。
predict_file():
def predict_file(self, input_file, output_file):
tag_dic = {"d": "疾病", "b": "身体", "s": "症状", "p": "医疗程序", "e": "医疗设备", "y": "药物", "k": "科室",
"m": "微生物类", "i": "医学检验项目"}
raw_text, test_ids, test_masks, test_lengths = self.from_txt(input_file)
test_dataset = TensorDataset(test_ids, test_masks, test_lengths)
test_loader = DataLoader(test_dataset, shuffle=False, batch_size=1)
self.model.load_state_dict(torch.load(self.NEWPATH, map_location=device))
self.model.eval()
op_file = codecs.open(output_file, 'w', 'utf-8')
for i, dev_batch in enumerate(test_loader):
sentence, masks, lengths = dev_batch
batch_raw_text = raw_text[i]
sentence, masks, lengths = Variable(sentence), Variable(masks), Variable(lengths)
if use_cuda:
sentence = sentence.to(device)
masks = masks.to(device)
predict_tags = self.model(sentence, masks)
predict_tags.tolist()
predict_tags = self.model(sentence, masks)
predict_tags.tolist()
predict_tags = [i2l_dic[t.item()] for t in predict_tags[0]]
predict_tags = predict_tags[:len(batch_raw_text)]
pred = predict_tags[1:-1]
raw_text = batch_raw_text[1:-1]
entity_mark = self.split_entity_input(pred)
entity_list = {}
if entity_mark is not None:
for item, ent in entity_mark.items():
entity = ''
index, tag = item[0], item[1]
len_entity = len(ent)
for i in range(index, index + len_entity):
entity = entity + raw_text[i]
entity_list[tag_dic[tag]] = entity
op_file.write("".join(raw_text))
op_file.write("\n")
op_file.write(json.dumps(entity_list, ensure_ascii=False))
op_file.write("\n")
op_file.close()
print('处理完成！')
print("结果保存至 {}".format(output_file))此时是处理的内容是多行输入的情况，和上一篇文章中的多行处理类似，同时处理思路也与本篇文章中的单行处理思路相同，函数结束部分添加了部分文档操作的内容，方便使用。
结语
有前面的医学文本分词做铺垫，医疗实体识别的难度降低了不少，但是关键的是处理这两项任务的模型还没有阅读，这才是本次工程的重点任务。如果您看到这里的话，希望您能留下一个免费的赞赞，这对我的帮助很大，对我本人同样也有着激励作用。