基于pytorch+transformers的NER实体识别
命名实体识别作为一项基础的NLP任务，其用在信息抽取、关系抽取、图谱构建等任务中都作为基础存在。
模型构建models.py:
# -*- coding: utf-8 -*-
import torch
import torch.nn as nn
from transformers import BertModel, BertConfig
from torchcrf import CRF
import os
class Bert_BiLSTM_CRF(nn.Module): # BiLSTM加上并无多大用处,速度还慢了，可去掉LSTM层
def __init__(self, tag_to_ix, embedding_dim=768, hidden_dim=256):
super(Bert_BiLSTM_CRF, self).__init__()
self.tag_to_ix = tag_to_ix
self.tagset_size = len(tag_to_ix)
self.hidden_dim = hidden_dim
self.embedding_dim = embedding_dim
self.bert = BertModel.from_pretrained("hfl/chinese-roberta-wwm-ext")
# self.lstm = nn.LSTM(input_size=embedding_dim, hidden_size=hidden_dim//2,
# num_layers=2, bidirectional=True, batch_first=True)
self.dropout = nn.Dropout(p=0.1)
# self.linear = nn.Linear(hidden_dim, self.tagset_size)
self.linear = nn.Linear(embedding_dim, self.tagset_size)
self.crf = CRF(self.tagset_size, batch_first=True)
def _get_features(self, sentence):
with torch.no_grad():
outputs = self.bert(sentence)
# enc, _ = self.lstm(outputs.last_hidden_state)
enc = outputs.last_hidden_state
enc = self.dropout(enc)
feats = self.linear(enc)
return feats
def forward(self, sentence, tags, mask, is_test=False):
emissions = self._get_features(sentence)
if not is_test: # Training，return loss
loss=-self.crf.forward(emissions, tags, mask, reduction='mean')
return loss
else: # Testing，return decoding
decode=self.crf.decode(emissions, mask)
return decode
utils类（主要负责DataSet类如下）utils.py
# -*- coding: utf-8 -*-
import torch
from torch.utils.data import Dataset
from transformers import BertTokenizer
import pandas as pd
tokenizer = BertTokenizer.from_pretrained("hfl/chinese-roberta-wwm-ext")
ner_type = pd.read_csv("model_data/bio_type.txt") # 包含ner所有类别的csv文件
ners = ner_type["label"].tolist()
VOCAB = []
for n in ners:
VOCAB.extend(["B-" + n, "I-"+ n])
VOCAB.extend(['' , '[CLS]', '[SEP]', "O"])
tag2idx = {tag: idx for idx, tag in enumerate(VOCAB)}
idx2tag = {idx: tag for idx, tag in enumerate(VOCAB)}
MAX_LEN = 512 - 5
# MAX_LEN = 128 - 2
class NerDataset(Dataset):
''' Generate our dataset '''
def __init__(self, f_path, inference_df = None):
self.sents = []
self.tags_li = []
if inference_df is not None:
data = inference_df
else:
data = pd.read_csv(f_path)
tags = data["label"].to_list()
words = data["word"].to_list()
print("f_path is {} len_word is {} len tag is {}".format(f_path, len(words), len(tags)))
word, tag = [], []
for char, t in zip(words, tags):
if char != '。':
word.append(char)
tag.append(t)
else:
if len(word) >= MAX_LEN-2:
self.sents.append(['[CLS]'] + word[:MAX_LEN] +[char] + ['[SEP]'])
self.tags_li.append(['[CLS]'] + tag[:MAX_LEN] + [t] + ['[SEP]'])
else:
self.sents.append(['[CLS]'] + word + [char] + ['[SEP]'])
self.tags_li.append(['[CLS]'] + tag + [t] + ['[SEP]'])
word, tag = [], []
if word:
if len(word) >= MAX_LEN-2:
self.sents.append(['[CLS]'] + word[:MAX_LEN] + ['[SEP]'])
self.tags_li.append(['[CLS]'] + tag[:MAX_LEN] + ['[SEP]'])
else:
self.sents.append(['[CLS]'] + word + ['[SEP]'])
self.tags_li.append(['[CLS]'] + tag + ['[SEP]'])
word, tag = [], []
def __getitem__(self, idx):
words, tags = self.sents[idx], self.tags_li[idx]
token_ids = tokenizer.convert_tokens_to_ids(words)
laebl_ids = [tag2idx[tag] for tag in tags]
seqlen = len(laebl_ids)
return token_ids, laebl_ids, seqlen
def __len__(self):
return len(self.sents)
def PadBatch(batch):
maxlen = max([i[2] for i in batch])
token_tensors = torch.LongTensor([i[0] + [0] * (maxlen - len(i[0])) for i in batch])
label_tensors = torch.LongTensor([i[1] + [0] * (maxlen - len(i[1])) for i in batch])
mask = (token_tensors > 0)
return token_tensors, label_tensors, mask
模型在真实数据上进行train、valid、test、inference的ner_main.py脚本如下：
# -*- coding: utf-8 -*-
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils import data
import os
import warnings
import argparse
import numpy as np
from sklearn import metrics
from transformers import AdamW, get_linear_schedule_with_warmup
import pandas as pd
from models import Bert_BiLSTM_CRF
from utils import NerDataset, PadBatch, VOCAB, tokenizer, tag2idx, idx2tag
warnings.filterwarnings("ignore", category=DeprecationWarning)
device = 'cuda' if torch.cuda.is_available() else 'cpu'
def train(e, model, iterator, optimizer, scheduler, criterion, device):
model.train()
losses = 0.0
step = 0
for i, batch in enumerate(iterator):
step += 1
x, y, z = batch
x = x.to(device)
y = y.to(device)
z = z.to(device)
loss = model(x, y, z)
losses += loss.item()
loss.backward()
optimizer.step()
scheduler.step()
optimizer.zero_grad()
print("Epoch: {}, Loss:{:.4f}".format(e, losses/step))
def validate(e, model, iterator, device):
model.eval()
Y, Y_hat = [], []
losses = 0
step = 0
with torch.no_grad():
for i, batch in enumerate(iterator):
step += 1
x, y, z = batch
x = x.to(device)
y = y.to(device)
z = z.to(device)
y_hat = model(x, y, z, is_test=True)
loss = model(x, y, z)
losses += loss.item()
# Save prediction
for j in y_hat:
Y_hat.extend(j)
# Save labels
mask = (z==1)
y_orig = torch.masked_select(y, mask)
Y.append(y_orig.cpu())
Y = torch.cat(Y, dim=0).numpy()
Y_hat = np.array(Y_hat)
acc = (Y_hat == Y).mean()*100
print("Epoch: {}, Val Loss:{:.4f}, Val Acc:{:.3f}%".format(e, losses/step, acc))
return model, losses/step, acc
def test(model, iterator, device):
model.eval()
Y, Y_hat = [], []
with torch.no_grad():
for i, batch in enumerate(iterator):
x, y, z = batch
x = x.to(device)
z = z.to(device)
y_hat = model(x, y, z, is_test=True)
# Save prediction
for j in y_hat:
Y_hat.extend(j)
# Save labels
mask = (z==1).cpu()
y_orig = torch.masked_select(y, mask)
Y.append(y_orig)
Y = torch.cat(Y, dim=0).numpy()
y_true = [idx2tag[i] for i in Y]
y_pred = [idx2tag[i] for i in Y_hat]
return y_true, y_pred
if __name__=="__main__o":
ner_type = pd.read_csv("model_data/type.txt")
ners = ner_type["label"].tolist()
labels = []
for n in ners:
labels.extend(["B-" + n, "I-"+ n])
print("all type len is {}".format(len(labels)))
best_model = None
_best_val_loss = 1e18
_best_val_acc = 1e-18
parser = argparse.ArgumentParser()
parser.add_argument("--batch_size", type=int, default=256)
parser.add_argument("--lr", type=float, default=0.0005)
parser.add_argument("--n_epochs", type=int, default=40)
parser.add_argument("--trainset", type=str, default="model_data/0704_bio_train.csv")
parser.add_argument("--validset", type=str, default="model_data/0704_bio_test.csv")
parser.add_argument("--testset", type=str, default="model_data/0704_bio_test.csv")
ner = parser.parse_args()
model = Bert_BiLSTM_CRF(tag2idx).cuda()
print('Initial model Done.')
train_dataset = NerDataset(ner.trainset)
print("train data len is {}".format(len(train_dataset)))
eval_dataset = NerDataset(ner.validset)
print("validset data len is {}".format(len(eval_dataset)))
test_dataset = NerDataset(ner.testset)
print("test_dataset len is {}".format(len(test_dataset)))
print('Load Data Done.')
train_iter = data.DataLoader(dataset=train_dataset,
batch_size=ner.batch_size,
shuffle=True,
num_workers=4,
collate_fn=PadBatch)
eval_iter = data.DataLoader(dataset=eval_dataset,
batch_size=(ner.batch_size)//2,
shuffle=False,
num_workers=4,
collate_fn=PadBatch)
test_iter = data.DataLoader(dataset=test_dataset,
batch_size=(ner.batch_size)//2,
shuffle=False,
num_workers=4,
collate_fn=PadBatch)
optimizer = AdamW(model.parameters(), lr=ner.lr, eps=1e-6)
# Warmup
len_dataset = len(train_dataset)
epoch = ner.n_epochs
batch_size = ner.batch_size
total_steps = (len_dataset // batch_size) * epoch if len_dataset % batch_size == 0 else (len_dataset // batch_size + 1) * epoch
warm_up_ratio = 0.1 # Define 10% steps
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps = warm_up_ratio * total_steps, num_training_steps = total_steps)
criterion = nn.CrossEntropyLoss(ignore_index=0)
print('Start Train...,')
for epoch in range(1, ner.n_epochs+1):
train(epoch, model, train_iter, optimizer, scheduler, criterion, device)
candidate_model, loss, acc = validate(epoch, model, eval_iter, device)
if loss < _best_val_loss and acc > _best_val_acc:
best_model = candidate_model
_best_val_loss = loss
_best_val_acc = acc
print("=============================================")
y_test, y_pred = test(best_model, test_iter, device)
print(metrics.classification_report(y_test, y_pred, labels=labels, digits=3))
torch.save(best_model.state_dict(), "checkpoint/0704_ner.pt")
test_data = pd.read_csv("model_data/0704_bio_test.csv")
y_test_useful = []
y_pred_useful = []
for a, b in zip(y_test, y_pred):
if a not in ['[CLS]', '[SEP]']:
y_test_useful.append(a)
y_pred_useful.append(b)
test_data["labeled"] = y_test_useful
test_data["pred"] = y_pred_useful
test_data.to_csv("result_files/bio_test_result.csv", index=False)
其中train.csv 和 test.csv只需要满足拥有word和label列即可，sample如下：
word,label
:,O,O
右,B-空间概念,B-位置
肺,B-解剖结构,B-解剖结构
下,I-解剖结构,I-解剖结构
叶,I-解剖结构,I-解剖结构
实,B-限定语,B-限定语
质,I-限定语,I-限定语
性,I-限定语,I-限定语
结,B-异常发现,B-异常发现
节,I-异常发现,I-异常发现
占,I-异常发现,I-异常发现
位,I-异常发现,I-异常发现
",",O,O
requirements.txt如下，以上脚本基于python3.8运行无误：
torch==1.9.1
transformers==4.12.5
hanziconv==0.3.2
pandas==1.4.2
tqdm==4.64.0
scikit-learn==1.1.1
pytorch-crf==0.7.2
numpy==1.22.4