“Google published a paper that explains exactly what they are doing and provides metrics as well [1,2].”
Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation
8 层 LSTM 加注意力的谷歌神经机器翻译系统
论文介绍 GNMT:编码器和解码器各 8 层的深度 LSTM,带注意力和残差连接;注意力把解码器底层连到编码器顶层以提高并行度,推理时用低精度算术加速,并把词拆成 wordpiece 子词单元来处理罕见词,束搜索使用长度归一化和覆盖惩罚。在 WMT'14 英法和英德上取得有竞争力的结果,在简单句的人工并排评测中,比谷歌的短语式生产系统平均减少 60% 的翻译错误。
评论者把它当作「谷歌用神经翻译取代统计翻译」的出处,并借它讨论 BLEU 评测、用强化学习直接优化 BLEU,以及开源复现与谷歌模型的差异。
HN 书评
推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及“This result is on an independently set up evaluation framework, compared to Google's which used their own framework.”
“Also there are metrics e.g. in the report released yesterday you can find BLEU scores on WMT 2014 tasks, and the clear conclusion is that Google is way ahead.”
“A practical example would be Google's recently announced Machine Translation System. It uses RL to directly optimize BLEU scores on translated sentences.”
“The code implements a pretty close model to what Google Translate has published.”
“See the recent NY Times feature [1] and the arxiv paper [2].”
看全部 7 条 →
“starting with Google introducing Neural Machine Translation[1] to replace their statistical model”
- Google's Neural Machine Translation: Bridging the Gap Between Humans and Machines ↗turing · 2016-09-27 · 2 分
- Google’s Neural Machine Translation System ↗fitzwatermellow · 2016-09-27 · 236 分