doc2vec使用说明(二)gensim工具包 LabeledSentence
欢迎交流,转载请注明出处。
本文介绍gensim工具包中,带标签(一个或者多个)的文档的doc2vec 的向量表示。
应用场景: 当每个文档不仅可以由文本信息表示,还有别的其他标签信息时,比如,在商品推荐中,将每个商品看成是一个文档,我们想学习商品向量表示时,可以只使用商品的描述信息来学习商品的向量表示,但有时:商品类别等信息我们也想将其考虑进去, 最简单的方法是:当用文本信息学习到商品向量后,添加一维商品的类别信息,但只用一维来表示商品类别信息的有效性差。gensim 工具包的doc2vec提供了更加合理的方法,将商品标签(如类别)加入到商品向量的训练中,即gensim 中的LabeledSentence方法
LabeledSentence的输入文件格式:每一行为:<labels, words>, 其中labels 可以有多个,用tab 键分隔,words 用空格键分隔,eg:<id category I like my cat demon>.
输出为词典vocabuary 中每个词的向量表示,这样就可以将商品labels:id,类别的向量拼接用作商品的向量表示。
写了个例子,仅供参考(训练一定要加 min_count=1,否则词典不全,这个小问题卡了一天 Doc2Vec(sentences, size = 100, window = 5, min_count=1))
注意:下面的例子是gensim更新之前的用法,gensim更新之后,没有了labels 的属性,换为tags, 且目标向量的表示也由vacb转到docvecs 中。更新后gensim 的用法见例子2.
例子1:gensim 更新前。
# -*- coding: UTF-8 -*- import gensim, logging import os from gensim.models.doc2vec import Doc2Vec,LabeledSentence from gensim.models import Doc2Vec import gensim.models.doc2vec asin=set() category=set() class LabeledLineSentence(object): def __init__(self, filename=object): self.filename =filename def __iter__(self): with open(self.filename,'r') as infile: data=infile.readlines(); # print "length: ", len(data) for uid,line in enumerate(data): asin.add(line.split("\t")[0]) category.add(line.split("\t")[1]) yield LabeledSentence(words=line.split("\t")[2].split(), labels=[line.split("\t")[0],line.split("\t")[1]]) print 'success' logging.basicConfig(format = '%(asctime)s : %(levelname)s : %(message)s', level = logging.INFO) sentences =LabeledLineSentence('product_bpr_train.txt') model = Doc2Vec(sentences, size = 100, window = 5, min_count=1) model.save('product_bpr_model.txt') print 'success1' #for uid,line in enumerate(model.vocab): # print line print len(model.vocab) outid = file('product_bpr_id_vector.txt', 'w') outcate = file('product_bpr_cate_vector.txt', 'w') for idx, line in enumerate(model.vocab): if line in asin : outid.write(line +'\t') for idx,lv in enumerate(model[line]): outid.write(str(lv)+" ") outid.write('\n') if line in category: outcate.write(line + '\t') for idx,lv in enumerate(model[line]): outcate.write(str(lv)+" ") outcate.write('\n') outid.close() outcate.close()
例子2:gensim 更新后
# -*- coding: UTF-8 -*- import gensim, logging import os from gensim.models.doc2vec import Doc2Vec,LabeledSentence from gensim.models import Doc2Vec import gensim.models.doc2vec asin=set() category=set() class LabeledLineSentence(object): def __init__(self, filename=object): self.filename =filename def __iter__(self): with open(self.filename,'r') as infile: data=infile.readlines(); print "length: ", len(data) for uid,line in enumerate(data): print "line:",line asin.add(line.split("\t")[0]) print "asin: ",asin category.add(line.split("\t")[1]) yield LabeledSentence(words=line.split("\t")[2].split(" "), tags=[line.split("\t")[0], line.split("\t")[1]]) print 'success' logging.basicConfig(format = '%(asctime)s : %(levelname)s : %(message)s', level = logging.INFO) sentences =LabeledLineSentence('product_bpr_test_train.txt') model = Doc2Vec(sentences, size =50, window = 5, min_count=1) model.save('product_bpr_model50.txt') print 'success1' print "doc2vecs length:", len(model.docvecs) outid = file('product_bpr_id_vector50.txt', 'w') outcate = file('product_bpr_cate_vector50.txt', 'w') for id in asin: outid.write(id+"\t") for idx,lv in enumerate(model.docvecs[id]): outid.write(str(lv)+" ") outid.write("\n") for cate in category: outcate.write(cate + '\t') for idx,lv in enumerate(model.docvecs[cate]): outcate.write(str(lv)+" ") outcate.write('\n') outid.close() outcate.close()
参考:
http://rare-technologies.com/doc2vec-tutorial/
https://github.com/RaRe-Technologies/gensim/blob/develop/docs/notebooks/doc2vec-IMDB.ipynb
http://radimrehurek.com/gensim/models/doc2vec.html#blog