没有合适的资源?快使用搜索试试~ 我知道了~
资源推荐
资源详情
资源评论
NEURAL MACHINE TRANSLATION
A DISSERTATION
SUBMITTED TO THE DEPARTMENT OF COMPUTER SCIENCE
AND THE COMMITTEE ON GRADUATE STUDIES
OF STANFORD UNIVERSITY
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
Minh-Thang Luong
December 2016
c
Copyright by Minh-Thang Luong 2017
All Righ ts Reserved
ii
I certify that I have read this dissertation and that, in my opinion, it is fully
adequate in scope and quality as a dissertation for the degree of Doctor of
Philosophy.
(Christopher D. Manning) Principal Adviser
I certify that I have read this dissertation and that, in my opinion, it is fully
adequate in scope and quality as a dissertation for the degree of Doctor of
Philosophy.
(Dan Jurafsky)
I certify that I have read this dissertation and that, in my opinion, it is fully
adequate in scope and quality as a dissertation for the degree of Doctor of
Philosophy.
(Andrew Ng)
I certify that I have read this dissertation and that, in my opinion, it is fully
adequate in scope and quality as a dissertation for the degree of Doctor of
Philosophy.
(Quoc V. Le)
iii
Abstract
Being able to communicate seamlessly across the entire repertoire of human languages is,
to me, an ultimately rewarding goal for an intelligent system. Despite great progress in
the field of Statistical Machine Translation (SMT) over the past two decades, translati o n
quality has not yet satisfied users; at the same time, SMT systems have become increasing
complex wit h many different components built separately, rendering it extremely difficult
to make further advancement. Recently, Neural Machin e Translation (NMT) emerges as
a promising solution to the problem of machine translation. At i ts core, NMT consists of
a single deep neural network with millions of neurons that learn t o directly map source
sentences to target sentences. NMT is powerful because it is an end-to-end deep-learning
framework that is significantly better th an SMT in capturing long-range dependencies in
sentences and generalizing well to unseen texts.
This dissertation presents all of th e essence of Neural Machine Translation (NMT),
through which I discuss how I have pushed the limits of NMT, making it applicable to a
wide variety of languages with state-of-the-art performance. My contributions in clude ad-
dressing the rare word problem with copy mechanisms, improving the attention mechanism
to better select local contexts in the source sentence, and translating at the character level
with a hybrid architecture. Towards the future of NMT, I discuss how to utilize data from a
wide variety of tasks such as p arsi ng, image caption generation, and unsupervised learning
to improve translati on; as well as how to compress NM T models for mobile devices. I
conclude with h ow my work influences subsequent research as well as provide an in-depth
coverage on the existing research landscape, highlight potential research directions, and
speculate on future elements needed to further advance NMT.
iv
Acknowledgements
As I was writing these lines of acknowledgement s, it struck me t hat 5 years have already
passed, blazingly fast but truly rewarding. At the beginning of my PhD, I thought it would
be a long, painful multi-year period of my li fe (yes, there had been ups and downs!), but
by the time I was close to finishing, my PhD duration became surprising short as I kept
wondering: will I ever have a chance to live again in such a stimulating, inspiring, and
caring environm ent with little p ressure and full freedom for academic and personal devel-
opment that Stanford and the CS department had offered? All of my achievements will not
be possible without the help of m any people.
First and foremost, I would like to extend my deepest appreciation to my advisor, Chris
Manning, whom I canno t find any poi nt to complain about. Chris has given me so much in-
spiration for my academic career and provided me countless support for almost any matters
including personal ones. I learned from him how to build up my academic “brand”, when
to say no to distractors, what to give to others, and many, many more great things. Without
him, it will be uncertain if I can make it all the way through my PhD. To get a sense of
Chris’s hard work, I invite readers to visit th is link https://github.com/lmthang/
thesis/issues which documents all of his edits to my thesis, page by page.
I would also like to thank other members in my thesis committee, Dan Jurafsky, Andrew
Ng, Quoc Le, and Christopher Potts for providing feedback to my t hesis and making sure I
can graduate. I am grateful to Quoc Le, Ilya Sutskever, Oriol Vinyals, and other colleagues
at Google Brain who have inspired me to do research in deep learning. It was a fortunate
turn of m y life when interni n g at Google Brain in 2014 and 2015 during which I h ad
learned how to train my first deep neural network in Neural Machine Translation! Quoc
Le deserves special menti on for being my mai n hosts during t h es e two internships and for
v
剩余155页未读,继续阅读
资源评论
zhuf14
- 粉丝: 16
- 资源: 57
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
最新资源
- jaudiotagger音频元数据
- 基于bert4keras 和tf2的多标签文本分类源代码+数据集
- 学生信息表.html
- C#WPF音乐、视频播放器
- YINTR24373-Next-Generation_DRAM_2024-Focus_on_HBM_and_3D_DRAM.pdf
- ME2320D-VB一款N-Channel沟道SOT23的MOSFET晶体管参数介绍与应用说明
- java swing + mysql 自动售货机系统
- ME2319-VB一款P-Channel沟道SOT23的MOSFET晶体管参数介绍与应用说明
- 基于asp.net的宠物管理系统源码+项目说明(高分项目).zip
- 05.11+77777777777777777777
资源上传下载、课程学习等过程中有任何疑问或建议,欢迎提出宝贵意见哦~我们会及时处理!
点击此处反馈
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功