wangqiangneu / MT-PaperReading

Record my paper reading about Machine Translation and other related works.
36 stars 2 forks source link

20-ACL-BPE-Dropout: Simple and Effective Subword Regularization #56

Open wangqiangneu opened 4 years ago

wangqiangneu commented 4 years ago

简介

针对BPE只能确定性的segmentation问题改进,提出bpe-dropout。方法很简单,首先还是学习标准的BPE,然后训练时在BPE merge时,以一定概率p(通常p=0.1)忽略本次merge,而在inference时则使用标准的BPE(等价于p=0)。相比kudo之前提出的subword regularizationbpe-dropout更简单,效果看着也不错。subword regularization需要先训练一个unigram LM做segment,再EM、viterbi生成samples,比较麻烦

有意思的点

论文信息

总结