本文主要介绍一个框架nlp-basictasks
nlp-basictasks是利用PyTorch深度学习框架所构建一个简单的库，旨在快速搭建模型完成一些基础的NLP任务，如分类、匹配、序列标注、语义相似度计算等。

下面利用该框架实现BERT+CRF模型做NER任务

数据集介绍

数据集来源
来源CLUE实体识别数据集

导入包

import torch,json
import sys,os
import pandas as pd
import random
import numpy as np
from nlp_basictasks.tasks import Ner
from nlp_basictasks.evaluation import nerEvaluator
from nlp_basictasks.readers.ner import InputExample

获取数据

def _create_examples(input_path,mode):
    examples = []
    with open(input_path, 'r') as f:
        idx = 0
        for line in f:
            json_d = {}
            line = json.loads(line.strip())
            text = line['text']
            label_entities = line.get('label', None)
            words = list(text)
            labels = ['O'] * len(words)
            if label_entities is not None:
                for key, value in label_entities.items():
                    for sub_name, sub_index in value.items():
                        for start_index, end_index in sub_index:
                            assert ''.join(words[start_index:end_index + 1]) == sub_name
                            if start_index == end_index:
                                labels[start_index] = 'S-' + key
                            else:
                                labels[start_index] = 'B-' + key
                                labels[start_index + 1:end_index + 1] = ['I-' + key] * (len(sub_name) - 1)
            json_d['id'] = f"{mode}_{idx}"
            json_d['context'] = " ".join(words)
            json_d['tag'] = " ".join(labels)
            json_d['raw_context'] = "".join(words)
            idx += 1
            examples.append(json_d)
    return examples
data=_create_examples('/data/nfs14/nfs/aisearch/asr/xhsun/datasets/cluener/train.json',mode='train')

构造训练集和验证集

train_examples=[]
for example in data:
    seq_in=example['context'].strip().split(' ')
    seq_out=example['tag'].strip().split(' ')
    assert len(seq_in)==len(seq_out)
    train_examples.append(InputExample(seq_in=seq_in,seq_out=seq_out))

dev_examples=train_examples[-2000:]
dev_seq_in=[]
dev_seq_out=[]
for example in dev_examples:
    dev_seq_in.append(example.seq_in)
    dev_seq_out.append(example.seq_out)
train_examples=train_examples[:-2000]

label_set=set()
for examples in data:
    label_list=examples['tag'].strip().split(' ')
    for label in label_list:
        label_set.add(label)
label2id={'[PAD]':0}
for label in label_set:
    label2id[label]=len(label2id)

print(label2id)

定义路径加载模型

model_path=''#你下载BERT模型位置，比如：'chinese-roberta-wwm/'
device='cuda'
ner_model=Ner(model_path,label2id=label2id,use_crf=True,use_bilstm=True,device=device,batch_first=True)

#可以指定是否使用CRF或者BiLSTM

训练模型

from torch.utils.data import DataLoader
batch_size=32
train_dataloader = DataLoader(train_examples, shuffle=True, batch_size=batch_size)
evaluator=nerEvaluator(label2id=label2id,seq_in=dev_seq_in,seq_out=dev_seq_out)
output_path=""#output_path是训练后保存模型的路径

ner_model.fit(train_dataloader=train_dataloader,evaluator=evaluator,epochs=5,output_path=output_path)

上图就是训练过程中各个实体的precision、recall以及f1得分

模型预测

不用100行代码即可完成NER任务，相关教程见nlp-basictasks框架做NER任务，觉得好用的话还请点个star，谢谢