本系列为完成复旦大学qxp老师的 nlp-beginner 系列任务的记录,邱老师不愧是NLP行业巨擘,该项目作为入门有非常好的作用。github项目链接如下。本文为该系列的第一次任务。项目链接附于文末。
任务目标
实现基于logistic/softmax regression的文本分类
数据集:Classify the sentiment of sentences from the Rotten Tomatoes dataset
数据集说明:
Submissions are evaluated on classification accuracy (the percent of labels that are predicted correctly) for every parsed phrase. The sentiment labels are:
0 - negative
1 - somewhat negative
2 - neutral
3 - somewhat positive
4 - positive
项目实现
数据读取
读入数据,该数据集共有4列,最后一列为情感倾向,共有5种类别
df_train = pd.read_csv('../input/sentiment-analysis-on-movie-reviews/train.tsv.zip', sep='\t')
df_test = pd.read_csv('../input/sentiment-analysis-on-movie-reviews/test.tsv.zip', sep='\t')
数据预处理
第一步需要对数据进行处理,经过本步操作完成分词
def clean_sentences(df):
reviews = []
for sent in tqdm(df['Phrase']):
# 去除其他字符
text = re.sub('[^a-zA-Z]', ' ', sent)
# 分词
words = word_tokenize(text.lower())
# 去除停用词
new_words = [char for char in words if char.lower() not in stopwords.words('english')]
# 将单词恢复成原型
lem = WordNetLemmatizer()
lem_words = [lem.lemmatize(i) for i in new_words]
reviews.append(lem_words)
return reviews
train_sentences = clean_sentences(df_train)
test_sentences = clean_sentences(df_test)
同时对于目标进行one-hot编码
train_target = torch.zeros(df_train["Sentiment"].shape[0],5)
train_target.scatter_(dim = 1, value = 1 , index = torch.tensor(df_train["Sentiment"]).unsqueeze(1))
词向量化
首先统计出所有单词建立词表,然后通过词表将句子向量化,选取最长的句子长度作为向量的维度
vocab = set()
len_max = 0
for sent in tqdm(train_sentences):
vocab.update(sent)
if len(sent) >len_max:
len_max = len(sent)
print(len(vocab),len_max)
word2index_list = {word : i for (i,word) in enumerate(vocab) }
def coding_sentence(sentences):
sentences_coding = torch.zeros(len(sentences),len_max)
for i ,line in enumerate(sentences):
for j,word in enumerate(line):
index = word2index_list[word]
sentences_coding[i][j] = index
for j in range(len(line),len_max):
sentences_coding[i][j] = 0
return sentences_coding
train_sentence = coding_sentence(train_sentences)
n-gram
phrases = df_train["Phrase"]
data = phrases[:1000]
dict_words = dict()
phrase_list = []
for line in data:
line = re.sub(r'[^a-zA-Z0-9]+'," ",line)
line = line.lower().split()
phrase_list.append(line)
data =phrase_list
for d in range(1,dimension+1):
for words in data:
for i in range(len(words) -d + 1):
temp = words[i:i+d]
temp = "_".join(temp)
if temp not in dict_words:
dict_words[temp] = len(dict_words)
length = len(dict_words)
train_matrix = torch.zeros(len(data),length)
for d in range(1,dimension+1):
for i, words in enumerate(data):
for j in range(len(words)-d+1):
temp = words[j:j+d];
temp = "_".join(temp)
train_matrix[i][dict_words[temp]]=1
train_matrix
数据集划分
对数据集进行划分,20%设为验证集
n_samples = train_sentence.shape[0]
n_val = int(0.2 * n_samples)
shuffled_indices = torch.randperm(n_samples)
train_indices = shuffled_indices[:-n_val]
val_indices = shuffled_indices[-n_val:]
train_indices, val_indices # 划分结果是随机的
X_train = train_sentence[train_indices]
Y_train = train_target[train_indices]
X_val = train_sentence[val_indices]
Y_val = train_target[val_indices]
训练模型
def training_loop(n_epochs , optimizer , model , loss_fn , X_train , Y_train , X_val , Y_val):
for epoch in range(1,n_epochs+1):
soft_max = torch.nn.Softmax()
train_pre = model(X_train)
train_loss = loss_fn(train_pre , Y_train)
val_pre = model(X_val)
val_loss = loss_fn(val_pre,Y_val)
optimizer.zero_grad()
train_loss.backward()
optimizer.step()
if epoch == 1 or epoch % 500 ==0 :
print("Epoch %d loss is %.4f val_loss is %.4f"%(epoch,train_loss,val_loss))
参数如下:
class LogistRegression(nn.Module):
def __init__(self):
super().__init__()
self.fc = nn.Linear(30, 5)
def forward(self, x):
out = self.fc(x)
out = torch.sigmoid(out)
return out
loss_fn = nn.MSELoss()
model = LogistRegression()
optimizer = torch.optim.SGD(model.parameters(),lr =1e-2)