猿创征文｜信息抽取（2）——pytorch实现Bert-BiLSTM-CRF、Bert-CRF模型进行实体抽取
文章目录
- 1 前言
- 2 数据准备
- 3 数据预处理
- 4 Bert-BiLSTM-CRF模型
- 5 Bert-CRF模型
- 6 模型训练
- 7 结果评估
- 8 训练集流水线
- 9 测试集流水线
- 10 记录遇到的一些坑
- 11 完整代码
1 前言
论文参考：
1 Neural Architectures for Named Entity Recognition
2 Attention is all you need
3 BERT: Pre-training of Deep Bidirectional Transformers for
Language Understanding
4 Bidirectional LSTM-CRF Models for Sequence Tagging
使用数据集：
https://www.datafountain.cn/competitions/529/ranking
Tips：文章可能存在一些漏洞，欢迎留言指出
2 数据准备
使用了transformers和seqeval库
安装方法：
huggingface-transformers
conda install -c huggingface transformers
seqeval
pip install seqeval -i https://pypi.tuna.tsinghua.edu.cn/simple
代码
import pandas as pd import torch from torch import optim from torch.utils.data import DataLoader from tqdm import tqdm from bert_bilstm_crf import Bert_BiLSTM_CRF, NerDataset, NerDatasetTest from bert_crf import Bert_CRF from transformers import AutoTokenizer, BertTokenizer from seqeval.metrics import f1_score # 路径 TRAIN_PATH = './dataset/train_data_public.csv' TEST_PATH = './dataset/test_public.csv' MODEL_PATH1 = './model/bert_bilstm_crf.pkl' MODEL_PATH2 = '../model/bert_crf.pkl' # 超参数 MAX_LEN = 64 BATCH_SIZE = 16 EPOCH = 5 # 预设 # 设备 DEVICE = "cuda:0" if torch.cuda.is_available() else "cpu" # tag2index tag2index = { "O": 0, # 其他 "B-BANK": 1, "I-BANK": 2, # 银行实体 "B-PRODUCT": 3, "I-PRODUCT": 4, # 产品实体 "B-COMMENTS_N": 5, "I-COMMENTS_N": 6, # 用户评论，名词 "B-COMMENTS_ADJ": 7, "I-COMMENTS_ADJ": 8 # 用户评论，形容词 } index2tag = {v: k for k, v in tag2index.items()}3 数据预处理
== 流程==
- 使用 s e r i e s . a p p l y ( l i s t ) \textcolor{red}{series.apply(list)} series.apply(list)函数将str转化为list格式
- 加载bert预训练tokenizer，使用 e n c o d e _ p l u s \textcolor{red}{encode\_plus} encode_plus函数对每一个text进行encode
- 如果是训练集，则执行如下操作：首先按照空格将每一个tag分割，并转化为索引列表，对每一个index_list，按照长度大于MAX_LEN裁剪，小于MAX_LEN填充的规则，合并为一个list，最后转化为tensor格式
代码
# 预处理 def data_preprocessing(dataset, is_train): # 数据str转化为list dataset['text_split'] = dataset['text'].apply(list) # token tokenizer = BertTokenizer.from_pretrained('bert-base-chinese') texts = dataset['text_split'].array.tolist() token_texts = [] for text in tqdm(texts): tokenized = tokenizer.encode_plus(text=text, max_length=MAX_LEN, return_token_type_ids=True, return_attention_mask=True, return_tensors='pt', padding='max_length', truncation=True) token_texts.append(tokenized) # 训练集有tag，测试集没有tag tags = None if is_train: dataset['tag'] = dataset['BIO_anno'].apply(lambda x: x.split(sep=' ')) tags = [] for tag in tqdm(dataset['tag'].array.tolist()): index_list = [0] + [tag2index[t] for t in tag] + [0] if len(index_list) MAX_LEN: # 裁剪 index_list = index_list[:MAX_LEN-1] + [0] tags.append(index_list) tags = torch.LongTensor(tags) return token_texts, tags4 Bert-BiLSTM-CRF模型
相对于Bert-CRF，中间添加了双向LSTM层。相对BiLSTM-CRF，相当于前面的word_embedding层替换为了bert预训练模型。
代码
import torch from torch import nn from torchcrf import CRF from transformers import BertModel from torch.utils.data import Dataset class Bert_BiLSTM_CRF(nn.Module): def __init__(self, tag2index): super(Bert_BiLSTM_CRF, self).__init__() self.tagset_size = len(tag2index) # bert层 self.bert = BertModel.from_pretrained('bert-base-chinese') # config = self.bert.config # lstm层 self.lstm = nn.LSTM(input_size=768, hidden_size=128, num_layers=1, batch_first=True, bidirectional=True) # dropout层 self.dropout = nn.Dropout(p=0.1) # Dense层 self.dense = nn.Linear(in_features=256, out_features=self.tagset_size) # CRF层 self.crf = CRF(num_tags=self.tagset_size) # 隐藏层 self.hidden = None # 负对数似然损失函数 def neg_log_likelihood(self, emissions, tags=None, mask=None, reduction=None): return -1 * self.crf(emissions=emissions, tags=tags, mask=mask, reduction=reduction) def forward(self, token_texts, tags): """ token_texts:{"input_size": tensor, [batch, 1, seq_len]->[batch, seq_len] "token_type_ids": tensor, [batch, 1, seq_len]->[batch, seq_len] "attention_mask": tensor [batch, 1, seq_len]->[batch, seq_len]->[seq_len, batch] } tags: [batch, seq_len]->[seq_len, batch] bert_out: [batch, seq_len, hidden_size(768)]->[seq_len, batch, hidden_size] self.hidden: [num_layers * num_directions, hidden_size(128)] out: [seq_len, batch, hidden_size * 2(256)] lstm_feats: [seq_len, batch, tagset_size] loss: tensor predictions: [batch, num] """ texts, token_type_ids, masks = token_texts['input_ids'], token_texts['token_type_ids'], token_texts['attention_mask'] texts = texts.squeeze(1) token_type_ids = token_type_ids.squeeze(1) masks = masks.squeeze(1) bert_out = self.bert(input_ids=texts, attention_mask=masks, token_type_ids=token_type_ids)[0] bert_out = bert_out.permute(1, 0, 2) # 检测设备 device = bert_out.device # 初始化隐藏层参数 self.hidden = (torch.randn(2, bert_out.size(0), 128).to(device), torch.randn(2, bert_out.size(0), 128).to(device)) out, self.hidden = self.lstm(bert_out, self.hidden) lstm_feats = self.dense(out) # 格式转换 masks = masks.permute(1, 0) masks = masks.clone().detach().bool() # masks = torch.tensor(masks, dtype=torch.uint8) # 计算损失值和预测值 if tags is not None: tags = tags.permute(1, 0) loss = self.neg_log_likelihood(lstm_feats, tags, masks, 'mean') predictions = self.crf.decode(emissions=lstm_feats, mask=masks) # [batch, 任意数] return loss, predictions else: predictions = self.crf.decode(emissions=lstm_feats, mask=masks) return predictionsDataset
class NerDataset(Dataset): def __init__(self, token_texts, tags): super(NerDataset, self).__init__() self.token_texts = token_texts self.tags = tags def __getitem__(self, index): return { "token_texts": self.token_texts[index], "tags": self.tags[index] if self.tags is not None else None, } def __len__(self): return len(self.token_texts) class NerDatasetTest(Dataset): def __init__(self, token_texts): super(NerDatasetTest, self).__init__() self.token_texts = token_texts def __getitem__(self, index): return { "token_texts": self.token_texts[index], "tags": 0 } def __len__(self): return len(self.token_texts)前向传播分析
token_texts:{
“input_size”: tensor, [batch, 1, seq_len]->[batch, seq_len]
“token_type_ids”: tensor, [batch, 1, seq_len]->[batch, seq_len]
“attention_mask”: tensor [batch, 1, seq_len]->[batch, seq_len]->[seq_len, batch]
}
tags: [batch, seq_len]->[seq_len, batch]
bert_out: [batch, seq_len, hidden_size(768)]->[seq_len, batch, hidden_size]
self.hidden: [num_layers * num_directions, hidden_size(128)]
out: [seq_len, batch, hidden_size * 2(256)]
lstm_feats: [seq_len, batch, tagset_size]
loss: tensor
predictions: [batch, num]
5 Bert-CRF模型
from torch import nn from torchcrf import CRF from transformers import BertModel class Bert_CRF(nn.Module): def __init__(self, tag2index): super(Bert_CRF, self).__init__() self.tagset_size = len(tag2index) # bert层 self.bert = BertModel.from_pretrained('bert-base-chinese') # dense层 self.dense = nn.Linear(in_features=768, out_features=self.tagset_size) # CRF层 self.crf = CRF(num_tags=self.tagset_size) # 隐藏层 self.hidden = None def neg_log_likelihood(self, emissions, tags=None, mask=None, reduction=None): return -1 * self.crf(emissions=emissions, tags=tags, mask=mask, reduction=reduction) def forward(self, token_texts, tags): """ token_texts:{"input_size": tensor, [batch, 1, seq_len]->[batch, seq_len] "token_type_ids": tensor, [batch, 1, seq_len]->[batch, seq_len] "attention_mask": tensor [batch, 1, seq_len]->[batch, seq_len]->[seq_len, batch] } tags: [batch, seq_len]->[seq_len, batch] bert_out: [batch, seq_len, hidden_size(768)]->[seq_len, batch, hidden_size] feats: [seq_len, batch, tagset_size] loss: tensor predictions: [batch, num] """ texts, token_type_ids, masks = token_texts.values() texts = texts.squeeze(1) token_type_ids = token_type_ids.squeeze(1) masks = masks.squeeze(1) bert_out = self.bert(input_ids=texts, attention_mask=masks, token_type_ids=token_type_ids)[0] bert_out = bert_out.permute(1, 0, 2) feats = self.dense(bert_out) # 格式转换 masks = masks.permute(1, 0) masks = masks.clone().detach().bool() # 计算损失之和预测值 if tags is not None: tags = tags.permute(1, 0) loss = self.neg_log_likelihood(feats, tags, masks, 'mean') predictions = self.crf.decode(emissions=feats, mask=masks) return loss, predictions else: predictions = self.crf.decode(emissions=feats, mask=masks) return predictions前向传播分析
token_texts:{
“input_size”: tensor, [batch, 1, seq_len]->[batch, seq_len]
“token_type_ids”: tensor, [batch, 1, seq_len]->[batch, seq_len]
“attention_mask”: tensor [batch, 1, seq_len]->[batch, seq_len]->[seq_len, batch]
}
tags: [batch, seq_len]->[seq_len, batch]
bert_out: [batch, seq_len, hidden_size(768)]->[seq_len, batch, hidden_size]
feats: [seq_len, batch, tagset_size]
loss: tensor
predictions: [batch, num]
6 模型训练
# 训练 def train(train_dataloader, model, optimizer, epoch): for i, batch_data in enumerate(train_dataloader): token_texts = batch_data['token_texts'].to(DEVICE) tags = batch_data['tags'].to(DEVICE) loss, predictions = model(token_texts, tags) optimizer.zero_grad() loss.backward() optimizer.step() if i % 200 == 0: micro_f1 = get_f1_score(tags, predictions) print(f'Epoch:{epoch} | i:{i} | loss:{loss.item()} | Micro_F1:{micro_f1}')7 结果评估
# 计算f1值 def get_f1_score(tags, predictions): tags = tags.to('cpu').data.numpy().tolist() temp_tags = [] final_tags = [] for index in range(BATCH_SIZE): # predictions先去掉头，再去掉尾 predictions[index].pop() length = len(predictions[index]) temp_tags.append(tags[index][1:length]) predictions[index].pop(0) # 格式转化，转化为List(str) temp_tags[index] = [index2tag[x] for x in temp_tags[index]] predictions[index] = [index2tag[x] for x in predictions[index]] final_tags.append(temp_tags[index]) f1 = f1_score(final_tags, predictions, average='micro') return f1Bert-BiLSTM-CRF
GPU_NAME:NVIDIA GeForce RTX 3060 Laptop GPU | Memory_Allocated:413399040 Epoch:0 | i:0 | loss:58.75139236450195 | Micro_F1:0.0 Epoch:0 | i:200 | loss:26.20857048034668 | Micro_F1:0.0 Epoch:0 | i:400 | loss:18.385879516601562 | Micro_F1:0.0 Epoch:1 | i:0 | loss:20.496620178222656 | Micro_F1:0.0 Epoch:1 | i:200 | loss:15.421577453613281 | Micro_F1:0.0 Epoch:1 | i:400 | loss:11.486358642578125 | Micro_F1:0.0 Epoch:2 | i:0 | loss:14.486601829528809 | Micro_F1:0.0 Epoch:2 | i:200 | loss:10.369649887084961 | Micro_F1:0.18867924528301888 Epoch:2 | i:400 | loss:8.056020736694336 | Micro_F1:0.5652173913043479 Epoch:3 | i:0 | loss:14.958343505859375 | Micro_F1:0.41025641025641024 Epoch:3 | i:200 | loss:9.968450546264648 | Micro_F1:0.380952380952381 Epoch:3 | i:400 | loss:8.947534561157227 | Micro_F1:0.5614035087719299 Epoch:4 | i:0 | loss:9.189300537109375 | Micro_F1:0.5454545454545454 Epoch:4 | i:200 | loss:8.673486709594727 | Micro_F1:0.43999999999999995 Epoch:4 | i:400 | loss:6.431578636169434 | Micro_F1:0.6250000000000001
Bert-CRF
GPU_NAME:NVIDIA GeForce RTX 3060 Laptop GPU | Memory_Allocated:409739264 Epoch:0 | i:0 | loss:57.06057357788086 | Micro_F1:0.0 Epoch:0 | i:200 | loss:12.05904483795166 | Micro_F1:0.0 Epoch:0 | i:400 | loss:13.805888175964355 | Micro_F1:0.39393939393939387 Epoch:1 | i:0 | loss:9.807424545288086 | Micro_F1:0.4905660377358491 Epoch:1 | i:200 | loss:8.098043441772461 | Micro_F1:0.509090909090909 Epoch:1 | i:400 | loss:7.059831619262695 | Micro_F1:0.611111111111111 Epoch:2 | i:0 | loss:6.629759788513184 | Micro_F1:0.6133333333333333 Epoch:2 | i:200 | loss:3.593130350112915 | Micro_F1:0.6896551724137931 Epoch:2 | i:400 | loss:6.8786163330078125 | Micro_F1:0.6666666666666666 Epoch:3 | i:0 | loss:5.009466648101807 | Micro_F1:0.6969696969696969 Epoch:3 | i:200 | loss:2.9549810886383057 | Micro_F1:0.8450704225352113 Epoch:3 | i:400 | loss:3.3801448345184326 | Micro_F1:0.868421052631579 Epoch:4 | i:0 | loss:5.864352226257324 | Micro_F1:0.626865671641791 Epoch:4 | i:200 | loss:3.308518409729004 | Micro_F1:0.7666666666666667 Epoch:4 | i:400 | loss:4.221902847290039 | Micro_F1:0.7000000000000001
分析
进行了5个epoch的训练
数据集比较小，只有7000多条数据，因此两个模型效果拟合效果相对BiLSTM+CRF模型提升不大。而添加了双向LSTM层之后，模型效果反而有所下降。
8 训练集流水线
注意学习率要设置小一点（小于1e-5），否则预测结果均为0，不收敛。
def execute(): # 加载训练集 train_dataset = pd.read_csv(TRAIN_PATH, encoding='utf8') # 数据预处理 token_texts, tags = data_preprocessing(train_dataset, is_train=True) # 数据集装载 train_dataset = NerDataset(token_texts, tags) train_dataloader = DataLoader(dataset=train_dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=4) # 构建模型 # model = Bert_BiLSTM_CRF(tag2index=tag2index).to(DEVICE) model = Bert_CRF(tag2index=tag2index).to(DEVICE) optimizer = optim.AdamW(model.parameters(), lr=1e-6) print(f"GPU_NAME:{torch.cuda.get_device_name()} | Memory_Allocated:{torch.cuda.memory_allocated()}") # 模型训练 for i in range(EPOCH): train(train_dataloader, model, optimizer, i) # 保存模型 torch.save(model.state_dict(), MODEL_PATH2)9 测试集流水线
# 测试集预测实体标签 def test(): # 加载数据集 test_dataset = pd.read_csv(TEST_PATH, encoding='utf8') # 数据预处理 token_texts, _ = data_preprocessing(test_dataset, is_train=False) # 装载测试集 dataset_test = NerDatasetTest(token_texts) test_dataloader = DataLoader(dataset=dataset_test, batch_size=BATCH_SIZE, shuffle=False, num_workers=4) # 构建模型 # model = Bert_BiLSTM_CRF(tag2index).to(DEVICE) model = Bert_CRF(tag2index).to(DEVICE) model.load_state_dict(torch.load(MODEL_PATH2)) # 模型预测 model.eval() predictions_list = [] with torch.no_grad(): for i, batch_data in enumerate(test_dataloader): token_texts = batch_data['token_texts'].to(DEVICE) predictions = model(token_texts, None) predictions_list.extend(predictions) print(len(predictions_list)) print(len(test_dataset['text'])) # 将预测结果转换为文本格式 entity_tag_list = [] index2tag = {v: k for k, v in tag2index.items()} # 反转字典 for i, (text, predictions) in enumerate(zip(test_dataset['text'], predictions_list)): # 删除首位和最后一位 predictions.pop() predictions.pop(0) text_entity_tag = [] for c, t in zip(text, predictions): if t != 0: text_entity_tag.append(c + index2tag[t]) entity_tag_list.append(" ".join(text_entity_tag)) # 合并为str并加入列表中 print(len(entity_tag_list)) result_df = pd.DataFrame(data=entity_tag_list, columns=['result']) result_df.to_csv('./data/result_df3.csv')
结果好像存在一些问题。。。
10 记录遇到的一些坑
（1）模型预测结果全为O
原因：按照之前的模型，AdamW优化器学习率设置0.001，学习率过高，导致梯度下降过程中落入了局部最低点。
解决方法：重新设置学习率为1e-6
（2）transformers的AdamW显示过期
解决方法：直接使用torch.optim的AdamW即可
（3）transformers库在ubuntu上无法使用
原因：缺少依赖
解决方法：
apt-get update apt-get install libssl1.0.0 libssl-dev
使用此代码在服务器终端上跑完后，仍会报错，原因未知，暂时用os.system()嵌入到代码中，在windows系统中无此报错。
（4）笔记本（联想R7000P2021）运行代码温度过高（最高95度）
解决方法：先用均衡模式（CPU不睿频）跑，温度只有六七十度，然后开启野兽模式跑一段时间，温度高了再切换为均衡模式。
11 完整代码
import pandas as pd import torch from torch import optim from torch.utils.data import DataLoader from tqdm import tqdm from bert_bilstm_crf import Bert_BiLSTM_CRF, NerDataset, NerDatasetTest from bert_crf import Bert_CRF from transformers import AutoTokenizer, BertTokenizer from seqeval.metrics import f1_score # 路径 TRAIN_PATH = './dataset/train_data_public.csv' TEST_PATH = './dataset/test_public.csv' MODEL_PATH1 = './model/bert_bilstm_crf.pkl' MODEL_PATH2 = '../model/bert_crf.pkl' # 超参数 MAX_LEN = 64 BATCH_SIZE = 16 EPOCH = 5 # 预设 # 设备 DEVICE = "cuda:0" if torch.cuda.is_available() else "cpu" # tag2index tag2index = { "O": 0, # 其他 "B-BANK": 1, "I-BANK": 2, # 银行实体 "B-PRODUCT": 3, "I-PRODUCT": 4, # 产品实体 "B-COMMENTS_N": 5, "I-COMMENTS_N": 6, # 用户评论，名词 "B-COMMENTS_ADJ": 7, "I-COMMENTS_ADJ": 8 # 用户评论，形容词 } index2tag = {v: k for k, v in tag2index.items()} # 训练 def train(train_dataloader, model, optimizer, epoch): for i, batch_data in enumerate(train_dataloader): token_texts = batch_data['token_texts'].to(DEVICE) tags = batch_data['tags'].to(DEVICE) loss, predictions = model(token_texts, tags) optimizer.zero_grad() loss.backward() optimizer.step() if i % 200 == 0: micro_f1 = get_f1_score(tags, predictions) print(f'Epoch:{epoch} | i:{i} | loss:{loss.item()} | Micro_F1:{micro_f1}') # 计算f1值 def get_f1_score(tags, predictions): tags = tags.to('cpu').data.numpy().tolist() temp_tags = [] final_tags = [] for index in range(BATCH_SIZE): # predictions先去掉头，再去掉尾 predictions[index].pop() length = len(predictions[index]) temp_tags.append(tags[index][1:length]) predictions[index].pop(0) # 格式转化，转化为List(str) temp_tags[index] = [index2tag[x] for x in temp_tags[index]] predictions[index] = [index2tag[x] for x in predictions[index]] final_tags.append(temp_tags[index]) f1 = f1_score(final_tags, predictions, average='micro') return f1 # 预处理 def data_preprocessing(dataset, is_train): # 数据str转化为list dataset['text_split'] = dataset['text'].apply(list) # token tokenizer = BertTokenizer.from_pretrained('bert-base-chinese') texts = dataset['text_split'].array.tolist() token_texts = [] for text in tqdm(texts): tokenized = tokenizer.encode_plus(text=text, max_length=MAX_LEN, return_token_type_ids=True, return_attention_mask=True, return_tensors='pt', padding='max_length', truncation=True) token_texts.append(tokenized) # 训练集有tag，测试集没有tag tags = None if is_train: dataset['tag'] = dataset['BIO_anno'].apply(lambda x: x.split(sep=' ')) tags = [] for tag in tqdm(dataset['tag'].array.tolist()): index_list = [0] + [tag2index[t] for t in tag] + [0] if len(index_list) MAX_LEN: # 裁剪 index_list = index_list[:MAX_LEN-1] + [0] tags.append(index_list) tags = torch.LongTensor(tags) return token_texts, tags # 执行流水线 def execute(): # 加载训练集 train_dataset = pd.read_csv(TRAIN_PATH, encoding='utf8') # 数据预处理 token_texts, tags = data_preprocessing(train_dataset, is_train=True) # 数据集装载 train_dataset = NerDataset(token_texts, tags) train_dataloader = DataLoader(dataset=train_dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=4) # 构建模型 # model = Bert_BiLSTM_CRF(tag2index=tag2index).to(DEVICE) model = Bert_CRF(tag2index=tag2index).to(DEVICE) optimizer = optim.AdamW(model.parameters(), lr=1e-6) print(f"GPU_NAME:{torch.cuda.get_device_name()} | Memory_Allocated:{torch.cuda.memory_allocated()}") # 模型训练 for i in range(EPOCH): train(train_dataloader, model, optimizer, i) # 保存模型 torch.save(model.state_dict(), MODEL_PATH2) # 测试集预测实体标签 def test(): # 加载数据集 test_dataset = pd.read_csv(TEST_PATH, encoding='utf8') # 数据预处理 token_texts, _ = data_preprocessing(test_dataset, is_train=False) # 装载测试集 dataset_test = NerDatasetTest(token_texts) test_dataloader = DataLoader(dataset=dataset_test, batch_size=BATCH_SIZE, shuffle=False, num_workers=4) # 构建模型 model = Bert_BiLSTM_CRF(tag2index).to(DEVICE) model.load_state_dict(torch.load(MODEL_PATH2)) # 模型预测 model.eval() predictions_list = [] with torch.no_grad(): for i, batch_data in enumerate(test_dataloader): token_texts = batch_data['token_texts'].to(DEVICE) predictions = model(token_texts, None) predictions_list.extend(predictions) print(len(predictions_list)) print(len(test_dataset['text'])) # 将预测结果转换为文本格式 entity_tag_list = [] index2tag = {v: k for k, v in tag2index.items()} # 反转字典 for i, (text, predictions) in enumerate(zip(test_dataset['text'], predictions_list)): # 删除首位和最后一位 predictions.pop() predictions.pop(0) text_entity_tag = [] for c, t in zip(text, predictions): if t != 0: text_entity_tag.append(c + index2tag[t]) entity_tag_list.append(" ".join(text_entity_tag)) # 合并为str并加入列表中 print(len(entity_tag_list)) result_df = pd.DataFrame(data=entity_tag_list, columns=['result']) result_df.to_csv('./data/result_df3.csv') if __name__ == '__main__': execute() test()
文章版权声明：除非注明，否则均为主机测评原创文章，转载或复制请以超链接形式并注明出处。
