开端
本人有幸加入哈工大赛尔NLP实验室。慢慢学习与NLP有关的知识。这篇论文是要求阅读项目中的一篇。本文也不算翻译,只是总结每一部分的重点。我希望能够通过这种写博客的方式巩固自己的知识,也和大家一起分享。
综述
这篇论文介绍的是在word2vec上预训练词向量使用卷积神经网络完成句子级别的分类任务。简单的CNN使用静态的词向量可以达到多个好结果在多任务上。如果进一步允许学习词向量能够获得更好的表现。
介绍
介绍部分介绍了深度学习模型在计算机视觉和语音识别获得了很好的成果。许多深度学习方面的研究都使用了从语言模型获得的词向量表示。 词向量把单词从稀疏的独热码向量投影到一个低维的紧密表示。在这个表示中相似的词向量的距离相近。 卷积神经网络使用卷积滤波器获得局部特征。一开始使用在计算机视觉领域,逐渐在NLP领域获取了成果。 在目前的工作中,我们训练一个CNN一层卷积在词向量上。一开始我们的词向量静止。尽管模型很简单,还是取得了不错的成果。
模型
xi∈Rk表示第i个单词由一个k维向量表示。长度为n的句子表示为 x1:n = x1 ⊕ x2 ⊕ . . . ⊕ xn ⊕表示向量的连接。 xi:i+j 表示对向量i…i+j连接 ci = f(w · xi:i+h?1 + b)表示了一个特征。其实就是把i…i+h-1的向量首尾连接,与w做向量乘法,得到一个值。这样做其是和卷积等价的。h是窗口大小。w∈R(h*k)是一个卷积核。 对每个窗口计算特征,把他们连到一起c = [c1, c2, . . . , cn?h+1]得到一个特征图。c ∈ Rn?h+1 然后使用最大池化c? = max{c},为了捕获最明显的特征。 上面过程是使用一个的卷积核。模型中使用多个卷积核,来获取多个特征图。这些来自多卷积核的特征最后连接到一起放入softmax分类器中。 模型还实验了多输入,意思是一个向量静态一个向量动态然后两个结果相加。
正则化
在这个论文中,正则化使用了l2正则化和dropout层。 dropout用来阻止过拟合。就是让神经元以一定概率失活。 就是在倒数第二层中训练中 z = [?c1, . . . , c?m] 不使用y = w · z + b 而是y = w · (z ? r) + b, r是一个向量取值从0,1按伯努利分布随机采样按dropout_rate也就是论文中的p为等于1的概率。 然后测试的时候,按照p来缩放向量而不dropout。
训练结果
训练结果我就不详细写了。
最后附上一份自己的代码
没有用训练好的词向量,就是学习架构用的,最后F1值也没有达到论文水平,72左右。
import torch
import torch.nn as nn
import torch.nn.functional as F
class MyModel(nn.Module):
def __init__(self, embeddings, n_features, n_classes, batch_size, dropout_prob=0.5):
super().__init__()
self.embeddings = embeddings
self.n_features = n_features
self.n_classes = n_classes
self.dropout_prob = dropout_prob
self.batch_size = batch_size
self.embed_size = embeddings.shape[1]
self.my_embeddings = nn.Embedding(embeddings.shape[0], self.embed_size)
self.my_embeddings.weight = nn.Parameter(torch.tensor(embeddings))
self.fc1_1 = nn.Conv1d(in_channels=self.embed_size,out_channels=100,kernel_size = 1)
self.fc1_2 = nn.Conv1d(in_channels=self.embed_size, out_channels=100, kernel_size=4)
self.fc1_3 = nn.Conv1d(in_channels=self.embed_size, out_channels=100, kernel_size=5)
self.fc1_1_2 = nn.ReLU()
self.fc1_2_2 = nn.ReLU()
self.fc1_3_2 = nn.ReLU()
self.fc3_1 = nn.Linear(in_features = 300, out_features =100)
self.fc3_2 = nn.ReLU()
self.fc3_3 = nn.Dropout(0.5)
self.fc4_1 = nn.Linear(in_features=100, out_features=50)
self.fc4_2 = nn.ReLU()
self.fc4_3 = nn.Dropout(0.5)
self.fc5= nn.Linear(in_features=50,out_features=2)
self.softmax = nn.Softmax(dim=1)
def forward(self,x):
x = x.long()
x = self.my_embeddings(x)
x = x.permute([0,2,1]).to(torch.float32)
x1 = self.fc1_1(x)
x2 = self.fc1_2(x)
x3 = self.fc1_3(x)
x1 = F.relu(self.fc1_1_2(F.max_pool1d(x1, kernel_size = x1.shape[2])))
x2 = F.relu(self.fc1_2_2(F.max_pool1d(x2, kernel_size=x2.shape[2])))
x3 = F.relu(self.fc1_3_2(F.max_pool1d(x3, kernel_size=x3.shape[2])))
x = torch.cat((x3, x2, x1),1).squeeze(2)
x = self.fc3_1(x)
x = self.fc3_2(x)
x = self.fc3_3(x)
x = self.fc4_1(x)
x = self.fc4_2(x)
x = self.fc4_3(x)
x = self.fc5(x)
return self.softmax(x)
from torch.utils.data.dataloader import DataLoader
import model
import dataget
import torch.optim as optim
from sklearn.metrics import f1_score, confusion_matrix, accuracy_score, classification_report, precision_recall_fscore_support
import matplotlib.pyplot as plt
mydataset = dataget.MyDataSet()
mymodel = model.MyModel(batch_size=64,dropout_prob=0.5,embeddings=mydataset.word2vec,n_classes=2,n_features=mydataset.maxlen)
mydatasetloader = DataLoader(dataset=mydataset,batch_size=64)
mytestdatasetloader = DataLoader(dataset=dataget.MytestSet(mydataset),batch_size=64)
loss_function = torch.nn.CrossEntropyLoss()
optimzer = optim.Adam(mymodel.parameters(),lr=0.0001)
history_loss = list()
for epoches in range(100):
tr_loss = 0
nb_tr_steps = 0
train_logits_list = []
train_labels_list = []
tv_loss = 0
nb_tv_steps = 0
valid_logits_list = []
valid_labels_list = []
mymodel.train()
for data,mask,labels in tqdm(mydatasetloader):
predict = mymodel(data)
labels = labels.long()
print(labels.shape)
print(predict.shape)
loss = loss_function(predict, labels)
loss.backward()
predict = predict.view(-1, 2).detach().cpu().numpy()
labels = labels.view(-1).to('cpu').numpy()
train_logits_list += [int(x) for x in np.argmax(predict, axis=1)]
train_labels_list += [int(x) for x in labels]
tr_loss = tr_loss+loss.item()
nb_tr_steps = nb_tr_steps +1
optimzer.step()
optimzer.zero_grad()
train_loss = tr_loss
train_accuracy = metrics.accuracy_score(train_labels_list, train_logits_list)
train_w_f1 = metrics.f1_score(train_labels_list, train_logits_list, average='weighted')
mymodel.eval()
for data,mask,labels in tqdm(mytestdatasetloader):
predict = mymodel(data)
labels = labels.long()
loss = loss_function(predict, labels)
tv_loss = tv_loss+loss.item()
nb_tv_steps = nb_tv_steps +1
predict = predict.view(-1, 2).detach().cpu().numpy()
labels = labels.view(-1).to('cpu').numpy()
valid_logits_list += [int(x) for x in np.argmax(predict, axis=1)]
valid_labels_list += [int(x) for x in labels]
valid_loss = tv_loss
valid_accuracy = metrics.accuracy_score(valid_labels_list, valid_logits_list)
valid_w_f1 = metrics.f1_score(valid_labels_list, valid_logits_list, average='weighted')
history_loss.append(valid_loss)
print('\nEpoch %d, train_loss=%.5f, train_acc=%.2f, train_w_f1=%.2f,valid_loss=%.5f, valid_acc=%.2f, valid_w_f1=%.2f'
%(epoches, train_loss, train_accuracy * 100, train_w_f1 * 100,valid_loss, valid_accuracy * 100, valid_w_f1 * 100))
fig=plt.figure(num=1,figsize=(4,4))
plt.subplot(111)
plt.plot(np.arange(0,len(history_loss)),history_loss)
plt.show()
import random
import chardet
import torch
import pickle
import numpy as np
from torch.utils.data.dataset import Dataset
from PIL import Image
import numpy as np
import torch
from tqdm import tqdm
def get_word2id(data_paths):
print('加载数据集')
wordid = {'PAD': '0'}
for path in data_paths:
with open(path,encoding="Windows-1252") as f:
for line in tqdm(f):
words = line.strip().split()
for word in words:
if word not in wordid.keys():
wordid[word] = len(wordid)
print(wordid)
return wordid
def get_word2vec(word2id):
word2vec = np.array(np.random.uniform(-1., 1., [len(word2id) + 1, 50]))
return word2vec
def get_corpus(word2id):
print('loading corpus...')
contents = list()
labels = list()
masks = list()
maxlen = 0
with open('E:\\python_project\\dataset\\rt-polaritydata\\rt-polarity.neg', encoding="Windows-1252") as f:
for line in tqdm(f):
words = line.strip().split()
if not words: continue
content = [word2id[word] for word in words]
if len(content) >maxlen:maxlen = len(content)
label = 0
contents.append(content)
labels.append(label)
with open('E:\\python_project\\dataset\\rt-polaritydata\\rt-polarity.pos', encoding="Windows-1252") as f:
for line in tqdm(f):
words = line.strip().split()
if not words: continue
content = [word2id[word] for word in words]
if len(content) > maxlen: maxlen = len(content)
label = 1
contents.append(content)
labels.append(label)
for index,line in enumerate(contents):
addline = [0 for i in range(maxlen - len(line))]
mask = [1 for i in range(len(line))]
line = line + addline
mask = mask + addline
masks.append(mask)
contents[index] = line
cc = list(zip(contents,labels))
random.shuffle(cc)
contents[:], labels[:] = zip(*cc)
return contents,masks,labels,maxlen
class MyDataSet(Dataset):
def __init__(self):
datapath = ['E:\\python_project\\dataset\\rt-polaritydata\\rt-polarity.pos',
'E:\\python_project\\dataset\\rt-polaritydata\\rt-polarity.neg']
self.word2id = get_word2id(data_paths=datapath)
self.corpus = get_corpus(self.word2id)
self.word2vec = get_word2vec(self.word2id)
self.maxlen = self.corpus[3]
def __len__(self):
return 8000
def __getitem__(self, index):
'''
:param index:
:return: 编码,掩码,标签
'''
return np.array(self.corpus[0][index]), np.array(self.corpus[1][index]),np.array(self.corpus[2][index])
class MytestSet(Dataset):
def __init__(self,MyDataSet):
self.mydataset = MyDataSet
def __len__(self):
return 2050
def __getitem__(self, index):
'''
:param index:
:return: 编码,掩码,标签
'''
return np.array(self.mydataset.corpus[0][index+8000]), np.array(self.mydataset.corpus[1][index+8000]), np.array(self.mydataset.corpus[2][index+8000])
|