数据增强是一种有效的防止过拟合并且可以加大训练集的有效方式，NLP中的数据增强方法有回译，同义词替换，随机插入，随机交换，随机删除，对抗训练这几种方式，本篇文章简单介绍一下这几种方式。
1.回译
#经过测试，这个翻译的包翻译的时间是最短的from pygtrans import Translatedef backTran(wordss):client = Translate()text1 = client.translate(wordss)text2 = client.translate(text1.translatedText, target='en')return text2.translatedText
使用样例
words="hello world"#传入一个英文的字符串就行backTranWord=backTran(words)
2.同义词替换
######################################################################### 同义词替换# 替换一个语句中的n个单词为其同义词########################################################################from nltk.corpus import stopwords#引入停用词，因为对停用词进行数据增强相当于没有增强from nltk.corpus import wordnet as wn#引入同义词import randomstop_words=stopwords.words('english')for w in ['!',',','.','?','-s','-ly','</s>','s']:stop_words.add(w)#这里传入的words是一个列表,#eg:"hello world".split(" ") or ["hello","world"]def synonym_replacement(words, n):new_words = words.copy()random_word_list = list(set([word for word in words if word not in stop_words]))random.shuffle(random_word_list)num_replaced = 0 for random_word in random_word_list:synonyms = get_synonyms(random_word)if len(synonyms) >= 1:synonym = random.choice(synonyms) new_words = [synonym if word == random_word else word for word in new_words] num_replaced += 1if num_replaced >= n: breaksentence = ' '.join(new_words)new_words = sentence.split(' ')return " ".join(new_words)#获取同义词def get_synonyms(word):nearbyWordSet=wn.synsets(word)return nearbyWordSet[0].lemma_names()
使用样例
words="hello world"lenW=len(words)/4#随机替换1/4的词语newWrods=synonym_replacement(words.split(" "),lenW)
3.随机插入
######################################################################### 随机插入# 随机在语句中插入n个词#########################################################################这里传入的words是一个列表,#eg:"hello world".split(" ") or ["hello","world"]def random_insertion(words, n):new_words = words.copy()for _ in range(n):add_word(new_words)return " ".join(new_words)#插入单词，这里插入随机挑选的单词的同义词def add_word(new_words):synonyms = []counter = 0 while len(synonyms) < 1:random_word = new_words[random.randint(0, len(new_words)-1)]synonyms = get_synonyms(random_word)counter += 1if counter >= 10:returnrandom_synonym = random.choice(synonyms)random_idx = random.randint(0, len(new_words)-1)new_words.insert(random_idx, random_synonym)
使用样例
words="hello world"lenW=len(words)/4#随机替换1/4的词语newWrods=random_insertion(words.split(" "),lenW)
4.随机交换
######################################################################### 随机交换# 随机交换几次#########################################################################这里传入的words是一个列表,#eg:"hello world".split(" ") or ["hello","world"]def random_swap(words, n):new_words = words.copy()for _ in range(n):new_words = swap_word(new_words)return " ".join(new_words)def swap_word(new_words):random_idx_1 = random.randint(0, len(new_words)-1)random_idx_2 = random_idx_1counter = 0while random_idx_2 == random_idx_1:random_idx_2 = random.randint(0, len(new_words)-1)counter += 1if counter > 3:return new_wordsnew_words[random_idx_1], new_words[random_idx_2] = new_words[random_idx_2], new_words[random_idx_1] return new_words
words="hello world"lenW=len(words)/4#随机替换1/4的词语newWrods=random_swap(words.split(" "),lenW)
5.随机删除
######################################################################### 随机删除# 以概率p删除语句中的词#########################################################################这里传入的words是一个列表,#eg:"hello world".split(" ") or ["hello","world"]def random_deletion(words, p):if len(words) == 1:return wordsnew_words = []for word in words:r = random.uniform(0, 1)if r > p:new_words.append(word)if len(new_words) == 0:rand_int = random.randint(0, len(words)-1)return [words[rand_int]]return " ".join(new_words)
words="hello world"#随机替换1/4的词语newWrods=random_deletion(words.split(" "),1/4)
6.对抗学习进行数据增强
对抗学习对数据进行增强的的原理是对embedding好的词向量上进行扰动，从而达到生成新样本的效果，这个过程是在训练过程中进行的。本次比赛中使用了两种扰动方式，Fast Gradient Method(FGM)
和Projected Gradient Descent(PGD)，下面给出使用方法
1） FGM
#定义import torchclass FGM(object):def __init__(self, model):super(FGM, self).__init__()self.model = modelself.backup = {}def attack(self, epsilon=1., emb_name='emb.'):# emb_name这个参数要换成你模型中embedding的参数名for name, param in self.model.named_parameters():if param.requires_grad and emb_name in name:self.backup[name] = param.data.clone()norm = torch.norm(param.grad)if norm != 0 and not torch.isnan(norm):r_at = epsilon * param.grad / normparam.data.add_(r_at)def restore(self, emb_name='emb.'):# emb_name这个参数要换成你模型中embedding的参数名for name, param in self.model.named_parameters():if param.requires_grad and emb_name in name: assert name in self.backupparam.data = self.backup[name]self.backup = {}# 使用fgm = FGM(model)for batch_input, batch_label in data:# 正常训练loss = model(batch_input, batch_label)loss.backward() # 反向传播，得到正常的grad# 对抗训练fgm.attack() # 在embedding上添加对抗扰动loss_adv = model(batch_input, batch_label)loss_adv.backward() # 反向传播，并在正常的grad基础上，累加对抗训练的梯度fgm.restore() # 恢复embedding参数# 梯度下降，更新参数optimizer.step()model.zero_grad()
PGD
#定义import torchclass PGD():def __init__(self, model):self.model = modelself.emb_backup = {}self.grad_backup = {}def attack(self, epsilon=1., alpha=0.3, emb_name='emb.', is_first_attack=False):# emb_name这个参数要换成你模型中embedding的参数名for name, param in self.model.named_parameters():if param.requires_grad and emb_name in name:if is_first_attack:self.emb_backup[name] = param.data.clone()norm = torch.norm(param.grad)if norm != 0 and not torch.isnan(norm):r_at = alpha * param.grad / normparam.data.add_(r_at)param.data = self.project(name, param.data, epsilon)def restore(self, emb_name='emb.'):# emb_name这个参数要换成你模型中embedding的参数名for name, param in self.model.named_parameters():if param.requires_grad and emb_name in name: assert name in self.emb_backupparam.data = self.emb_backup[name]self.emb_backup = {}def project(self, param_name, param_data, epsilon):r = param_data - self.emb_backup[param_name]if torch.norm(r) > epsilon:r = epsilon * r / torch.norm(r)return param_data + rdef backup_grad(self):for name, param in self.model.named_parameters():if param.requires_grad:self.grad_backup[name] = param.graddef restore_grad(self):for name, param in self.model.named_parameters():if param.requires_grad:param.grad = self.grad_backup[name]#使用pgd = PGD(model)K = 3for batch_input, batch_label in data:# 正常训练loss = model(batch_input, batch_label)loss.backward() # 反向传播，得到正常的gradpgd.backup_grad()# 对抗训练for t in range(K):pgd.attack(is_first_attack=(t==0)) # 在embedding上添加对抗扰动, first attack时备份param.dataif t != K-1:model.zero_grad()else:pgd.restore_grad()loss_adv = model(batch_input, batch_label)loss_adv.backward() # 反向传播，并在正常的grad基础上，累加对抗训练的梯度pgd.restore() # 恢复embedding参数# 梯度下降，更新参数optimizer.step()model.zero_grad()
