自然语言处理之TextRNN

TextRNN文本分类

• RNN是在自然语言处理领域非常标配的一个网络，在序列标注/命名体识别/seq2seq模型等很多场景都有应用，Recurrent Neural Network for Text Classification with Multi-Task Learning文中介绍了RNN用于分类问题的设计，下图LSTM用于网络结构原理示意图，最后一步的隐层代表着对整个句子的编码，直接接全连接层softmax输出。
  
  数据预处理
  使用torchtext库来进行文本处理，包括以下几个部分：
• 分词：torchtext使用jieba分词器作为tokenizer
• 去停用词：加载去停用词表，并在data.Field中设置 text = data.Field(sequential=True, lower=True, tokenize=tokenizer, stop_words=stop_words)
• 文本长度padding：如果需要设置文本的长度，则在data.Field中设置fix_length,否则torchtext自动将文本长度处理为最大样本长度
• 词向量转换：torchtext能自动建立word2id和id2word两个索引，并将index转换为对应词向量，如果要加载预训练词向量，在build_vocab中设置即可。

import jieba
from torchtext import data
import re
from torchtext.vocab import Vectors


def tokenizer(text): # create a tokenizer function
    regex = re.compile(r'[^\u4e00-\u9fa5aA-Za-z0-9]')
    text = regex.sub(' ', text)
    return [word for word in jieba.cut(text) if word.strip()]

# 去停用词
def get_stop_words():
    file_object = open('data/stopwords.txt')
    stop_words = []
    for line in file_object.readlines():
        line = line[:-1]
        line = line.strip()
        stop_words.append(line)
    return stop_words

def load_data(args):
    print('加载数据中...')
    stop_words = get_stop_words() # 加载停用词表
    '''
    如果需要设置文本的长度，则设置fix_length,否则torchtext自动将文本长度处理为最大样本长度
    text = data.Field(sequential=True, tokenize=tokenizer, fix_length=args.max_len, stop_words=stop_words)
    '''
    text = data.Field(sequential=True, lower=True, tokenize=tokenizer, stop_words=stop_words)
    label = data.Field(sequential=False)

    text.tokenize = tokenizer
    train, val = data.TabularDataset.splits(
            path='data/',
            skip_header=True,
            train='train.tsv',
            validation='validation.tsv',
            format='tsv',
            fields=[('index', None), ('label', label), ('text', text)],
        )

    if args.static:
        text.build_vocab(train, val, vectors=Vectors(name="data/eco_article.vector")) # 此处改为你自己的词向量
        args.embedding_dim = text.vocab.vectors.size()[-1]
        args.vectors = text.vocab.vectors

    else: text.build_vocab(train, val)

    label.build_vocab(train, val)

    train_iter, val_iter = data.Iterator.splits(
            (train, val),
            sort_key=lambda x: len(x.text),
            batch_sizes=(args.batch_size, len(val)), # 训练集设置batch_size,验证集整个集合用于测试
            device=-1
    )
    args.vocab_size = len(text.vocab)
    args.label_num = len(label.vocab)
    return train_iter, val_iter

训练

• 如果要使用预训练词向量，则data文件夹下要存放你自己的词向量
• 随机初始化Embedding进行训练

python train.py

• 随机初始化Embedding并设置是否为双向LSTM以及stack的层数

python train.py -bidirectional=True -layer-num=2

• 使用预训练词向量进行训练（词向量静态，不可调）

python train.py -static=true

• 微调预训练词向量进行训练（词向量动态，可调）

python train.py -static=true -fine-tune=true

模型代码：

import torch
import torch.nn as nn

# 循环神经网络 (many-to-one)
class TextRNN(nn.Module):
    def __init__(self, args):
        super(TextRNN, self).__init__()
        embedding_dim = args.embedding_dim
        label_num = args.label_num
        vocab_size = args.vocab_size
        self.hidden_size = args.hidden_size
        self.layer_num = args.layer_num
        self.bidirectional = args.bidirectional

        self.embedding = nn.Embedding(vocab_size, embedding_dim)
        if args.static:  # 如果使用预训练词向量，则提前加载，当不需要微调时设置freeze为True
            self.embedding = self.embedding.from_pretrained(args.vectors, freeze=not args.fine_tune)

        self.lstm = nn.LSTM(embedding_dim, # x的特征维度,即embedding_dim
                            self.hidden_size,# 隐藏层单元数
                            self.layer_num,# 层数
                            batch_first=True,# 第一个维度设为 batch, 即:(batch_size, seq_length, embedding_dim)
                            bidirectional=self.bidirectional) # 是否用双向
        self.fc = nn.Linear(self.hidden_size * 2, label_num) if self.bidirectional else nn.Linear(self.hidden_size, label_num)

    def forward(self, x):
        # 输入x的维度为(batch_size, max_len), max_len可以通过torchtext设置或自动获取为训练样本的最大长度
        x = self.embedding(x)  # 经过embedding,x的维度为(batch_size, time_step, input_size=embedding_dim)

        # 隐层初始化
        # h0维度为(num_layers*direction_num, batch_size, hidden_size)
        # c0维度为(num_layers*direction_num, batch_size, hidden_size)
        h0 = torch.zeros(self.layer_num * 2, x.size(0), self.hidden_size) if self.bidirectional else torch.zeros(self.layer_num, x.size(0), self.hidden_size)

        c0 = torch.zeros(self.layer_num * 2, x.size(0), self.hidden_size) if self.bidirectional else torch.zeros(self.layer_num, x.size(0), self.hidden_size)

        # LSTM前向传播，此时out维度为(batch_size, seq_length, hidden_size*direction_num)
        # hn,cn表示最后一个状态?维度与h0和c0一样
        out, (hn, cn) = self.lstm(x, (h0, c0))

        # 我们只需要最后一步的输出,即(batch_size, -1, output_size)
        out = self.fc(out[:, -1, :])
        return out

训练代码：

import argparse
import os
import sys
import torch
import torch.nn.functional as F
import data_processor
from model import TextRNN

parser = argparse.ArgumentParser(description='TextRNN text classifier')

parser.add_argument('-lr', type=float, default=0.01, help='学习率')
parser.add_argument('-batch-size', type=int, default=128)
parser.add_argument('-epoch', type=int, default=20)
parser.add_argument('-embedding-dim', type=int, default=128, help='词向量的维度')
parser.add_argument('-hidden_size', type=int, default=64, help='lstm中神经单元数')
parser.add_argument('-layer-num', type=int, default=1, help='lstm stack的层数')
parser.add_argument('-label-num', type=int, default=2, help='标签个数')
parser.add_argument('-bidirectional', type=bool, default=False, help='是否使用双向lstm')
parser.add_argument('-static', type=bool, default=False, help='是否使用预训练词向量')
parser.add_argument('-fine-tune', type=bool, default=True, help='预训练词向量是否要微调')
parser.add_argument('-cuda', type=bool, default=False)
parser.add_argument('-log-interval', type=int, default=1, help='经过多少iteration记录一次训练状态')
parser.add_argument('-test-interval', type=int, default=100, help='经过多少iteration对验证集进行测试')
parser.add_argument('-early-stopping', type=int, default=1000, help='早停时迭代的次数')
parser.add_argument('-save-best', type=bool, default=True, help='当得到更好的准确度是否要保存')
parser.add_argument('-save-dir', type=str, default='model_dir', help='存储训练模型位置')

args = parser.parse_args()

def train(args):
    train_iter, dev_iter = data_processor.load_data(args) # 将数据分为训练集和验证集
    print('加载数据完成')
    model = TextRNN(args)
    if args.cuda: model.cuda()
    optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
    steps = 0
    best_acc = 0
    last_step = 0
    model.train()
    for epoch in range(1, args.epoch + 1):
        for batch in train_iter:
            feature, target = batch.text, batch.label
            # t_()函数表示将(max_len, batch_size)转置为(batch_size, max_len)
            feature.data.t_(), target.data.sub_(1) # target减去1
            if args.cuda:
                feature, target = feature.cuda(), target.cuda()
            optimizer.zero_grad()
            logits = model(feature)
            loss = F.cross_entropy(logits, target)
            loss.backward()
            optimizer.step()
            steps += 1
            if steps % args.log_interval == 0:
                # torch.max(logits, 1)函数：返回每一行中最大值的那个元素，且返回其索引（返回最大元素在这一行的列索引）
                corrects = (torch.max(logits, 1)[1] == target).sum()
                train_acc = 100.0 * corrects / batch.batch_size
                sys.stdout.write(
                    '\rBatch[{}] - loss: {:.6f}  acc: {:.4f}%({}/{})'.format(steps,
                                                                             loss.item(),
                                                                             train_acc,
                                                                             corrects,
                                                                             batch.batch_size))
            if steps % args.test_interval == 0:
                dev_acc = eval(dev_iter, model, args)
                if dev_acc > best_acc:
                    best_acc = dev_acc
                    last_step = steps
                    if args.save_best:
                        print('Saving best model, acc: {:.4f}%\n'.format(best_acc))
                        save(model, args.save_dir, 'best', steps)
                else:
                    if steps - last_step >= args.early_stopping:
                        print('\nearly stop by {} steps, acc: {:.4f}%'.format(args.early_stopping, best_acc))
                        raise KeyboardInterrupt

'''
对验证集进行测试 
'''
def eval(data_iter, model, args):
    corrects, avg_loss = 0, 0
    for batch in data_iter:
        feature, target = batch.text, batch.label
        feature.data.t_(), target.data.sub_(1)
        if args.cuda:
            feature, target = feature.cuda(), target.cuda()
        logits = model(feature)
        loss = F.cross_entropy(logits, target)
        avg_loss += loss.item()
        corrects += (torch.max(logits, 1)
                     [1].view(target.size()) == target).sum()
    size = len(data_iter.dataset)
    avg_loss /= size
    accuracy = 100.0 * corrects / size
    print('\nEvaluation - loss: {:.6f}  acc: {:.4f}%({}/{}) \n'.format(avg_loss,
                                                                       accuracy,
                                                                       corrects,
                                                                       size))
    return accuracy

def save(model, save_dir, save_prefix, steps):
    if not os.path.isdir(save_dir):
        os.makedirs(save_dir)
    save_prefix = os.path.join(save_dir, save_prefix)
    save_path = '{}_steps_{}.pt'.format(save_prefix, steps)
    torch.save(model.state_dict(), save_path)

train(args)

结果：

使用单向LSTM,可以看到，前400次迭代loss几乎没有下降，接着开始快速下降，最后验证集的准确率能到91%左右（经过调参可以更高）
数据获取方式：https://download.csdn.net/download/wenweno0o/11206041