BriefGPT.xyz
Apr, 2017
神经机器翻译的后期最小贝叶斯风险解码
Later-stage Minimum Bayes-Risk Decoding for Neural Machine Translation
HTML
PDF
Raphael Shu, Hideki Nakayama
TL;DR
本研究提出使用最小贝叶斯风险编解码技术(MBR)来消除基于beam search算法的模型自信度导致的输出错误，并使用GPU批量计算贝叶斯风险值以加速解码，并在机器翻译中验证后期MBR解码的优越性。
Abstract
For extended periods of time,
sequence generation models
rely on
beam search algorithm
to generate output sequence. However, the correctness of beam search degrades when the a model is over-confident about a subo
→