BriefGPT.xyz
Apr, 2020
极化-VAE:基于相似度的离散化表现学习用于文本生成
Polarized-VAE: Proximity Based Disentangled Representation Learning for Text Generation
HTML
PDF
Vikash Balasubramanian, Ivan Kobyzev, Hareesh Bahuleyan, Ilya Shapiro, Olga Vechtomova
TL;DR
本文提出了polarized-VAE方法,它通过反映数据点之间在属性相似性方面的相似性度量,可以解开潜在空间中选定的属性的不可避免联系,相对于其他属性的联系降至最低,展现出更好的表现,并适用于其他属性分离任务。
Abstract
Learning
disentangled representations
of real world data is a challenging open problem. Most previous methods have focused on either fully
supervised approaches
which use attribute labels or
→