读论文《Distributed Representations of Words and Phrases and their Compositionality》
2017-08-24 00:35
615 查看
读论文《Distributed Representations of Words and Phrases and their Compositionality》
introduce
这篇论文是对《Efficient Estimation of Word Representations in Vector Space》的补充,介绍了使用Skip-gram模型和Hierarchical Softmax训练模式的训练方法。并补充了Negative Sampling的训练模式替代Negative Sampling,获得更快的训练效果。本文还提出了对高频词进行二次抽样的方法,以及衡量短语的方法,学习短语的表示。
method
Skip-gram的目标是最大化下面的概率分布其中, c是训练上下文的大小。比如c是2,就预测上文2个词和下文2个词。
p(wt+j|wt)是表示wt条件下出现wt+j的概率。
如果采用正常的softmax,p(wt+j|wt)这么算:
其中W是词表的大小,而一般大语料的词表都是非常大的,所以对全部输出使用softmax的训练难度很大,所以Skip-gram采用了用哈夫曼树结构构建的Hierarchical Softmax,最终的概率密度函数如下:
本文中还介绍了Negative Sampling的方法,它以一个正样本和几个随机采取的负样本作为优化目标
其中文章对高频词进行二次抽样的函数如下
作者以构造的得分函数评价连续的词是否为短语的可能性,当大于阖值时就作为短语。最后以短语中各单词的词嵌入之和作为该短语的词嵌入(文中没找到,但表格5中是用加的)
my view
1.负采样是否适用于其他模型2.怎么表示短语的词嵌入,各单词之和?那句子,语段呢?
本文地址:http://blog.csdn.net/qq_31456593/article/details/77515156
相关文章推荐
- 翻译:Distributed Representations of Words and Phrases and their Compositionality
- 《Distributed Representations of Words and Phrases and their Compositionality》笔记
- Distributed Representations of Words and Phrasesand their Compositionality
- Distributed Representations of Words and Phrasesand their Compositionality
- [NLP论文阅读]Distributed Representations of Sentences and Documents
- [论文泛读] Accurate and efficient stochastic reliability analysis of composite services using their compact Markov reward model representations (SCC, 2007)
- NLP-文献-Distributed Representations of Sentences and Documents
- 【每周论文】Time, Clocks, and Ordering of Events in a Distributed System
- 【论文阅读】Distributed Representation of Sentence and Document
- Words and Their Stories: Some of Our Listeners' Favorite Sayings
- 论文阅读:Joint Learning of Single-image and Cross-image Representations for Person Re-identification
- 【deep learning学习笔记】Distributed Representations of Sentences and Documents
- 1607.CVPR-Joint Learning of Single-image and Cross-image Representations for Person ReID 论文笔记
- 《Document Classification by Inversion of Distributed Language Representations》分享
- 自然语言处理入门以及TensorFlow官网教程Vector Representations of words简介
- A survey of task allocation and load balance in distributed system阅读笔记
- Microservices and the First Law of Distributed Objects
- 经典论文翻译导读之《Large-scale Incremental Processing Using Distributed Transactions and Notifications》
- 【论文阅读】Author2Vec: Learning Author Representations by Combining Content and Link Information
- 综述论文:图像标注中的BoW表示 Bag-of-Words Representation in Image Annotation: A Review