基于TextCNN、LSTM与Transformer模型的疫情微博情绪分类

任务概述

微博情绪分类任务旨在识别微博中蕴含的情绪，输入是一条微博，输出是该微博所蕴含的情绪类别。在本次任务中，我们将微博按照其蕴含的情绪分为以下六个类别之一：积极、愤怒、悲伤、恐惧、惊奇和无情绪。

• 数据集来源
  本数据集(疫情微博数据集)内的微博内容是在疫情期间使用相关关键字筛选获得的疫情微博，其内容与新冠疫情相关
• 数据集标签
  每条微博被标注为以下六个类别之一: neural (无情绪)、angry (愤怒)、sad (悲伤)、surprise (惊奇)。
• 数据集规模
  疫情微博训练数据集包括6,606条微博，测试数据集包含5,000条微博。
• 数据集形式
  数据集为json格式，包含三个字段：数据编号（id），文本（content），情绪标签（label）。
  示例： {“id”: 11, “content”: “武汉加油！中国加油！安徽加油！”, “label”: “happy”}
• 下载
  链接：https://pan.baidu.com/s/13_czouycHR8mK0pHzuH7gw
  提取码：t81p

实验设计

文件框架

简单易懂，不多逼逼

main函数

import torch
import torch.nn as nn
import torch.optim as optim
import pickle as pkl
from src.models.textCNN import TextCNN
from src.models.textRNN import TextRNN
from src.models.Transformer import Transformer
from src.Config import Config
from src.get_data import get_data
from src.train import train

if __name__ == '__main__':
    config = Config()
    batch_size = config.batch_size
    learning_rate = config.learning_rate

    train_dataloader, test_dataloader, n_vocab = get_data(batch_size)
    config.n_vocab = n_vocab

    # model = TextCNN(config).to(Config.device)
    model = TextRNN(config).to(Config.device)
    # model = Transformer(config).to(Config.device)

    # 导入word2vec训练出来的预训练词向量
    id_vec = open(Config.id_vec_path, 'rb')
    id_vec = pkl.load(id_vec)
    id_vec = torch.tensor(list(id_vec.values())).to(Config.device)
    if config.embedding_pretrained:
        model.embedding = nn.Embedding.from_pretrained(id_vec)

    loss = nn.CrossEntropyLoss().to(Config.device)
    optimizer = optim.Adam(params=model.parameters(), lr=learning_rate)

    train(model, loss, optimizer, train_dataloader, test_dataloader, Config.epoches, Config.device)

配置信息

import torch


class Config():
    train_data_path = '../data/virus_train.txt'
    test_data_path = '../data/virus_eval_labeled.txt'
    vocab_path = '../data/vocab.pkl'
    split_word_all_path = '../data/split_word_all.txt'
    model_file_name_path = '../data/vec_model.txt'
    id_vec_path = '../data/id_vec.pkl'

    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

    word_level = True   # 按照字级别进行分词

    embedding_pretrained = False   # 是否使用预训练的词向量

    label_fields = {'neural': 0, 'happy': 1, 'angry': 2, 'sad': 3, 'fear': 4, 'surprise': 5}
    all_seq_len = 64  # 句子长度，长剪短补

    batch_size = 128
    learning_rate = 0.0001
    epoches = 50
    dropout = 0.5
    num_classes = 6

    embed_dim = 300
    n_vocab = 0

数据准备

import re
import os
import json
import jieba
import pickle as pkl
import numpy as np
import gensim.models.word2vec as w2v
import torch
from src.Config import Config
import torch.utils.data as Data

train_data_path = Config.train_data_path
test_data_path = Config.test_data_path
vocab_path = Config.vocab_path

label_fields = Config.label_fields
all_seq_len = Config.all_seq_len

UNK, PAD = '<UNK>', '<PAD>'  # 未知字，padding符号

# 构造字典
def build_vocab(content_list, tokenizer):
    file_split_word = open(Config.split_word_all_path, 'w', encoding='utf-8')
    vocab_dic = {}
    for content in content_list:
        word_lines = []
        for word in tokenizer(content):
            vocab_dic[word] = vocab_dic.get(word, 0) + 1
            word_lines.append(word)

        str = " ".join(word_lines) + "
"
        file_split_word.write(str)

    file_split_word.close()
    vocab_dic.update({UNK: len(vocab_dic), PAD: len(vocab_dic) + 1})
    vocab_dic = {word_count: idx for idx, word_count in enumerate(vocab_dic)}
    return vocab_dic

def build_id_vec(vocab_dic, model):
    model.wv.add_vector(UNK, np.zeros(300))
    model.wv.add_vector(PAD, np.ones(300))
    id2vec = {}
    for word in vocab_dic.keys():
        id = vocab_dic.get(word, vocab_dic.get(UNK))
        vec = model.wv.get_vector(word)
        id2vec.update({id: vec})
    return id2vec

# 预训练词向量
def train_vec():
    model_file_name = Config.model_file_name_path
    sentences = w2v.LineSentence(Config.split_word_all_path)
    model = w2v.Word2Vec(sentences, vector_size=300, window=20, min_count=0)
    model.save(model_file_name)

# 读入数据与数据清洗
def load_data(root):
    content_list = []
    content_token_list = []
    label_list = []
    # 不同分词器
    if Config.word_level:
        tokenizer = lambda x: [y for y in x]
    else:
        tokenizer = lambda x: jieba.cut(x, cut_all=False)

    file = open(root, 'r', encoding='utf-8')

    datas = json.load(file)
	# 多种数据清洗方式
    # pattern = re.compile(r'[^u4e00-u9fa5|，|。|！|？|[|]]')
    pattern = re.compile(r'[^u4e00-u9fa5|，|。|！|？]')
    # pattern = re.compile(r'[^u4e00-u9fa5|，|。]')       # seq_len=32 CNN：67%-68%  RNN：61%-62%  Transformer：63-64%
    # pattern = re.compile(r'[^u4e00-u9fa5|，|。|！]')       # CNN：65%-66%
    for data in datas:
        content_after_clean = re.sub(pattern, '', data['content'])
        content_list.append(content_after_clean)
        label_list.append(label_fields[data['label']])

    if os.path.exists(vocab_path):
        vocab = pkl.load(open(vocab_path, 'rb'))
    else:
        vocab = build_vocab(content_list, tokenizer)
        pkl.dump(vocab, open(vocab_path, 'wb'))
        if Config.embedding_pretrained:
            train_vec()
            model = w2v.Word2Vec.load(Config.model_file_name_path)
            id_vec = build_id_vec(vocab, model)
            pkl.dump(id_vec, open(Config.id_vec_path, 'wb'))

    for content in content_list:
        word_line = []
        token = list(tokenizer(content))
        seq_len = len(token)
        if seq_len < all_seq_len:
            token.extend([PAD] * (all_seq_len - seq_len))
        else:
            token = token[:all_seq_len]

        for word in token:
            word_line.append(vocab.get(word, vocab.get(UNK)))

        content_token_list.append(word_line)

    n_vocab = len(vocab)

    return content_token_list, label_list, n_vocab

# 将数据映射为Dataset
class WeiBboDataset(Data.Dataset):
    def __init__(self, content_token_list, label_list):
        super(WeiBboDataset, self).__init__()
        self.content_token_list = content_token_list
        self.label_list = label_list

    def __getitem__(self, index):
        label = float(self.label_list[index])
        return torch.tensor(self.content_token_list[index]), torch.tensor(label)

    def __len__(self):
        return len(self.label_list)

# 核心函数
def get_data(batch_size):
    train_content_token_list, train_label_list, n_vocab = load_data(train_data_path)
    test_content_token_list, test_label_list, _ = load_data(test_data_path)

    train_dataset = WeiBboDataset(train_content_token_list, train_label_list)
    test_dataset = WeiBboDataset(test_content_token_list, test_label_list)

    train_dataloader = Data.DataLoader(dataset=train_dataset, batch_size=batch_size, shuffle=True)
    test_dataloader = Data.DataLoader(dataset=test_dataset, batch_size=batch_size, shuffle=False)
    return train_dataloader, test_dataloader, n_vocab


if __name__ == '__main__':
    get_data(32)

模型搭建

TextCNN

import torch.nn as nn
import torch
import torch.nn.functional as F


class TextCNN(nn.Module):
    def __init__(self, Config):
        super(TextCNN, self).__init__()

        self.filter_sizes = (2, 3, 4)  # 卷积核尺寸
        self.num_filters = 64  # 卷积核数量(channels数)

        self.embedding = nn.Embedding(Config.n_vocab, Config.embed_dim)
        self.convs = nn.ModuleList(
            [nn.Conv2d(1, self.num_filters, (k, Config.embed_dim)) for k in self.filter_sizes])
        self.dropout = nn.Dropout(Config.dropout)
        self.fc = nn.Linear(self.num_filters * len(self.filter_sizes), Config.num_classes)

    def conv_and_pool(self, x, conv):
        x = F.relu(conv(x))
        x = x.squeeze(3)
        x = F.max_pool1d(x, x.size(2)).squeeze(2)
        return x

    def forward(self, x):
        out = self.embedding(x)
        out = out.unsqueeze(1)
        out = torch.cat([self.conv_and_pool(out, conv) for conv in self.convs], 1)
        out = self.dropout(out)
        out = self.fc(out)
        return out

LSTM

import os
import torch
import torch.nn as nn
import numpy as np


class TextRNN(nn.Module):
    def __init__(self, Config):
        super(TextRNN, self).__init__()
        self.hidden_size = 128  # lstm隐藏层
        self.num_layers = 2  # lstm层数
        self.embedding = nn.Embedding(Config.n_vocab, Config.embed_dim)
        self.lstm = nn.LSTM(Config.embed_dim, self.hidden_size, self.num_layers,
                            bidirectional=True, batch_first=True, dropout=Config.dropout)
        self.fc = nn.Linear(self.hidden_size * 2, Config.num_classes)

    def forward(self, x):
        out = self.embedding(x)  # [batch_size, seq_len, embeding]=[128, 32, 300]
        out, _ = self.lstm(out)
        out = self.fc(out[:, -1, :])  # 句子最后时刻的 hidden state
        return out

Transformer

import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import copy


class Transformer(nn.Module):
    def __init__(self, Config):
        super(Transformer, self).__init__()

        self.hidden = 1024
        self.last_hidden = 512
        self.num_head = 5
        self.num_encoder = 2
        self.dim_model = 300

        self.embedding = nn.Embedding(Config.n_vocab, Config.embed_dim)

        self.postion_embedding = Positional_Encoding(Config.embed_dim, Config.all_seq_len, Config.dropout, Config.device)
        self.encoder = Encoder(self.dim_model, self.num_head, self.hidden, Config.dropout)
        self.encoders = nn.ModuleList([
            copy.deepcopy(self.encoder)
            # Encoder(config.dim_model, config.num_head, config.hidden, config.dropout)
            for _ in range(self.num_encoder)])

        self.fc1 = nn.Linear(Config.all_seq_len * self.dim_model, Config.num_classes)
        # self.fc2 = nn.Linear(config.last_hidden, config.num_classes)
        # self.fc1 = nn.Linear(config.dim_model, config.num_classes)

    def forward(self, x):
        out = self.embedding(x)
        out = self.postion_embedding(out)
        for encoder in self.encoders:
            out = encoder(out)
        out = out.view(out.size(0), -1)
        # out = torch.mean(out, 1)
        out = self.fc1(out)
        return out


class Encoder(nn.Module):
    def __init__(self, dim_model, num_head, hidden, dropout):
        super(Encoder, self).__init__()
        self.attention = Multi_Head_Attention(dim_model, num_head, dropout)
        self.feed_forward = Position_wise_Feed_Forward(dim_model, hidden, dropout)

    def forward(self, x):
        out = self.attention(x)
        out = self.feed_forward(out)
        return out


class Positional_Encoding(nn.Module):
    def __init__(self, embed, pad_size, dropout, device):
        super(Positional_Encoding, self).__init__()
        self.device = device
        self.pe = torch.tensor([[pos / (10000.0 ** (i // 2 * 2.0 / embed)) for i in range(embed)] for pos in range(pad_size)])
        self.pe[:, 0::2] = np.sin(self.pe[:, 0::2])
        self.pe[:, 1::2] = np.cos(self.pe[:, 1::2])
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        out = x + nn.Parameter(self.pe, requires_grad=False).to(self.device)
        out = self.dropout(out)
        return out


class Scaled_Dot_Product_Attention(nn.Module):
    '''Scaled Dot-Product Attention '''
    def __init__(self):
        super(Scaled_Dot_Product_Attention, self).__init__()

    def forward(self, Q, K, V, scale=None):
        '''
        Args:
            Q: [batch_size, len_Q, dim_Q]
            K: [batch_size, len_K, dim_K]
            V: [batch_size, len_V, dim_V]
            scale: 缩放因子 论文为根号dim_K
        Return:
            self-attention后的张量，以及attention张量
        '''
        attention = torch.matmul(Q, K.permute(0, 2, 1))
        if scale:
            attention = attention * scale
        # if mask:  # TODO change this
        #     attention = attention.masked_fill_(mask == 0, -1e9)
        attention = F.softmax(attention, dim=-1)
        context = torch.matmul(attention, V)
        return context


class Multi_Head_Attention(nn.Module):
    def __init__(self, dim_model, num_head, dropout=0.0):
        super(Multi_Head_Attention, self).__init__()
        self.num_head = num_head
        assert dim_model % num_head == 0
        self.dim_head = dim_model // self.num_head
        self.fc_Q = nn.Linear(dim_model, num_head * self.dim_head)
        self.fc_K = nn.Linear(dim_model, num_head * self.dim_head)
        self.fc_V = nn.Linear(dim_model, num_head * self.dim_head)
        self.attention = Scaled_Dot_Product_Attention()
        self.fc = nn.Linear(num_head * self.dim_head, dim_model)
        self.dropout = nn.Dropout(dropout)
        self.layer_norm = nn.LayerNorm(dim_model)

    def forward(self, x):
        batch_size = x.size(0)
        Q = self.fc_Q(x)
        K = self.fc_K(x)
        V = self.fc_V(x)
        Q = Q.view(batch_size * self.num_head, -1, self.dim_head)
        K = K.view(batch_size * self.num_head, -1, self.dim_head)
        V = V.view(batch_size * self.num_head, -1, self.dim_head)
        # if mask:  # TODO
        #     mask = mask.repeat(self.num_head, 1, 1)  # TODO change this
        scale = K.size(-1) ** -0.5  # 缩放因子
        context = self.attention(Q, K, V, scale)

        context = context.view(batch_size, -1, self.dim_head * self.num_head)
        out = self.fc(context)
        out = self.dropout(out)
        out = out + x  # 残差连接
        out = self.layer_norm(out)
        return out


class Position_wise_Feed_Forward(nn.Module):
    def __init__(self, dim_model, hidden, dropout=0.0):
        super(Position_wise_Feed_Forward, self).__init__()
        self.fc1 = nn.Linear(dim_model, hidden)
        self.fc2 = nn.Linear(hidden, dim_model)
        self.dropout = nn.Dropout(dropout)
        self.layer_norm = nn.LayerNorm(dim_model)

    def forward(self, x):
        out = self.fc1(x)
        out = F.relu(out)
        out = self.fc2(out)
        out = self.dropout(out)
        out = out + x  # 残差连接
        out = self.layer_norm(out)
        return out

训练和测试

import os
import torch
import torch.nn as nn
from torch.autograd import Variable
from utils.draw_loss_pic import draw_loss_pic

os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"


def train(net, loss, optimizer, train_loader, test_loader, epoches, device):
    train_loss = []
    train_acc = []
    test_loss = []
    test_acc = []
    for epoch in range(epoches):
        net.train()
        total_loss = 0.0
        correct = 0
        sample_num = 0
        for batch_idx, (data, target) in enumerate(train_loader):
            data = data.to(device).long()
            target = target.to(device).long()
            optimizer.zero_grad()
            output = net(data)
            ls = loss(output, target)
            ls.backward()
            optimizer.step()
            total_loss += ls.item()
            sample_num += len(target)
            max_output = output.data.max(1, keepdim=True)[1].view_as(target)
            correct += (max_output == target).sum()

        print('epoch %d, train_loss %f, train_acc: %f' % (epoch + 1, total_loss/sample_num, float(correct.data.item()) / sample_num))
        train_loss.append(total_loss/sample_num)
        train_acc.append(float(correct.data.item()) / sample_num)

        test_ls, test_accury = test(net, test_loader, device, loss)
        test_loss.append(test_ls)
        test_acc.append(test_accury)

    draw_loss_pic(train_loss, test_loss, "loss")
    draw_loss_pic(train_acc, test_acc, "acc")


def test(net, test_loader, device, loss):
    net.eval()
    total_loss = 0.0
    correct = 0
    sample_num = 0
    for batch_idx, (data, target) in enumerate(test_loader):
        data = data.to(device)
        target = target.to(device).long()
        output = net(data)
        ls = loss(output, target)
        total_loss += ls.item()
        sample_num += len(target)
        max_output = output.data.max(1, keepdim=True)[1].view_as(target)
        correct += (max_output == target).sum()

    print('test_loss %f, test_acc: %f' % (
        total_loss / sample_num, float(correct.data.item()) / sample_num))
    return total_loss / sample_num, float(correct.data.item()) / sample_num

运行结果

在分词器为按字进行分词、句子长度为64、batch_size为128、learning_rate为0.0001、数据清洗方式为“保留中文、逗号、句号、感叹号、问号、emoji（带中括号）”，训练模型为CNN的情况下，损失曲线和准确率曲线如下图所示：

参考

https://blog.csdn.net/cui_yonghua/article/details/121094116

https://ask.csdn.net/questions/672138?ops_request_misc&request_id&biz_id=106&utm_term=re%E5%BA%93%E5%8C%B9%E9%85%8D%E4%B8%AD%E6%8B%AC%E5%8F%B7&utm_medium=distribute.pc_search_result.none-task-ask-2~ask~sobaiduweb~default-2-672138.pc_ask&spm=1018.2226.3001.4187

https://blog.csdn.net/Littewood/article/details/123393736?ops_request_misc=%257B%2522request%255Fid%2522%253A%2522167258426316800180660412%2522%252C%2522scm%2522%253A%252220140713.130102334..%2522%257D&request_id=167258426316800180660412&biz_id=0&utm_medium=distribute.pc_search_result.none-task-blog-2~all~sobaiduend~default-2-123393736-null-null.142^v68^control,201^v4^add_ask,213^v2^t3_control2&utm_term=lambda%E7%94%A8%E6%B3%95&spm=1018.2226.3001.4449

https://blog.csdn.net/benzhujie1245com/article/details/117173090