Word2vec是一群用来产生词向量的相关模型,由Google公司在2013年开放。Word2vec可以根据给定的语料库,通过优化后的训练模型快速有效地将一个词语表达成向量形式,为自然语言处理领域的应用研究提供了新的工具。
Word2vec模型为浅而双层的神经网络,网络以词表现,并且需猜测相邻位置的输入词,在word2vec中词袋模型假设下,词的顺序是不重要的。训练完成之后,word2vec模型可用来映射每个词到一个向量,可用来表示词对词之间的关系,该向量为神经网络之隐藏层。
本实验主要基于Word2vec来实现对搜狐新闻文本分类,大致步骤如下。
数据集下载地址 密码: hq5v
训练集共有24000条样本,12个分类,每个分类2000条样本。
测试集共有12000条样本,12个分类,每个分类1000条样本。
完成此步骤需要先安装gensim库,安装命令:pip install gensim
对于每一篇文章,获取文章的每一个分词在word2vec模型的相关性向量。然后把一篇文章的所有分词在word2vec模型中的相关性向量求和取平均数,即此篇文章在word2vec模型中的相关性向量。
import pandas as pd
import jieba
import time
from gensim.models import Word2Vec
import warnings
train_df = pd.read_csv('sohu_train.txt', sep='\t', header=None)
train_df.head()
for name, group in train_df.groupby(0):
print(name,len(group))
test_df = pd.read_csv('sohu_test.txt', sep='\t', header=None)
for name, group in test_df.groupby(0):
print(name, len(group))
train_df.columns = ['分类', '文章']
stopword_list = [k.strip() for k in open('stopwords.txt', encoding='utf8').readlines() if k.strip() != '']
cutWords_list = []
i = 0
startTime = time.time()
for article in train_df['文章']:
cutWords = [k for k in jieba.cut(article) if k not in stopword_list]
i += 1
if i % 1000 == 0:
print('前%d篇文章分词共花费%.2f秒' % (i, time.time() - startTime))
cutWords_list.append(cutWords)
with open('cutWords_list.txt', 'w') as file:
for cutWords in cutWords_list:
file.write(' '.join(cutWords) + '\n')
with open('cutWords_list.txt') as file:
cutWords_list = [k.split() for k in file.readlines()]
word2vec_model = Word2Vec(cutWords_list, size=100, iter=10, min_count=20)
warnings.filterwarnings('ignore')
word2vec_model.wv.most_similar('摄影')
word2vec_model.most_similar(positive=['女人', '先生'], negative=['男人'], topn=1)
word2vec_model.save('word2vec_model.w2v')
import numpy as np
import time
def getVector_v1(cutWords, word2vec_model):
count = 0
article_vector = np.zeros(word2vec_model.layer1_size)
for cutWord in cutWords:
if cutWord in word2vec_model:
article_vector += word2vec_model[cutWord]
count += 1
return article_vector / count
startTime = time.time()
vector_list = []
i = 0
for cutWords in cutWords_list[:5000]:
i += 1
if i % 1000 ==0:
print('前%d篇文章形成词向量花费%.2f秒' %(i, time.time()-startTime))
vector_list.append(getVector_v1(cutWords, word2vec_model))
X = np.array(vector_list)
import time
import numpy as np
def getVector_v3(cutWords, word2vec_model):
vector_list = [word2vec_model[k] for k in cutWords if k in word2vec_model]
cutWord_vector = np.array(vector_list).mean(axis=0)
return cutWord_vector
startTime = time.time()
vector_list = []
i = 0
for cutWords in cutWords_list[:5000]:
i += 1
if i % 1000 ==0:
print('前%d篇文章形成词向量花费%.2f秒' %(i, time.time()-startTime))
vector_list.append(getVector_v3(cutWords, word2vec_model))
X = np.array(vector_list)
调用sklearn.linear_model库的LogisticRegression方法实例化模型对象。
调用sklearn.model_selection库的train_test_split方法划分训练集和测试集。
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split
train_X, test_X, train_y, test_y = train_test_split(X, y, test_size=0.2)
logistic_model = LogisticRegression()
logistic_model.fit(train_X, train_y)
logistic_model.score(test_X, test_y)
结果为:0.7825
5.模型测试
import pandas as pd
import numpy as np
from sklearn.externals import joblib
import jieba
def getVectorMatrix(article_series):
return np.array([getVector_v3(jieba.cut(k), word2vec_model) for k in article_series])
logistic_model = joblib.load('logistic.model')
test_df = pd.read_csv('sohu_test.txt', sep='\t', header=None)
test_df.columns = ['分类', '文章']
for name, group in test_df.groupby('分类'):
featureMatrix = getVectorMatrix(group['文章'])
target = labelEncoder.transform(group['分类'])
print(name, logistic_model.score(featureMatrix, target))