ADM-201 dump PMP dumps pdf SSCP exam materials CBAP exam sample questions

电脑如何与人进行交谈? – 译学馆
未登录,请登录后再发表信息
最新评论 (0)
播放视频

电脑如何与人进行交谈?

Behind the Mic: The Science of Talking with Computers

这个短视频是关于……
[音乐]
[music]
[法语]
[SPEAKING IN FRENCH]
[各种声音]
[VOICES SPEAKING]
[笑声]
[LAUGHTER]
[电脑的声音]
[COMPUTER NOISES]
与电脑进行交谈
杰弗里:我们来到这个世界
GEOFFREY HINTON: Wecome into this world
天生就拥有学习
with the innateabilities to learn
与其他有情感人进行交流
to interact withother sentient beings.
假设你必须通过写短信与他人交流
Suppose you had to interact with other people by writing little messages to them.
那会是十分痛苦的
It’d be a real pain.
与其同样的是我们和电脑交流的方法
And that’s how weinteract with computers.
这比起和他们聊天更轻松
It’s much easierjust to talk to them.
如果电脑能够理解
It’s just so mucheasier if the computers
我们所说的话,那将更加轻松
could understandwhat we’re saying.
要做到这样 你需要完善的语音识别系统
And for that, you need reallygood speech recognition. NARRATOR:
解说员:第一个语音识别系统由贝尔实验室于1952年研发
The first speech recognition system was developed by Bell Laboratories in 1952.
它只能识别一个人说出来的若干单词
It could only recognizenumbers spoken by one person.
在20世纪70年代 卡内基·梅隆研制出
In the 1970s, CarnegieMellon came out
Harpy系统 它能识别超过1000个单词
with the Harpy system, whichwas able to recognize over 1,000
并且识别出来自同一个单词的
words and could recognizedifferent pronunciations
不同发音
of the same word.
电脑男性声音:番茄
MALE COMPUTER VOICE: Tomato.
电脑女性声音:番茄
FEMALE COMPUTER VOICE: Tomato. NARRATOR:
解说员:语音识别在80年代继续引入了
Speech recognition continued in the’80s with the introduction
隐马尔科夫模型 使用更加
of the hidden Markov model, which used a more
数学化的方法来分析声波
mathematical approachto analyzing sound waves
使我们得到了现今的很多突破
and led to many of thebreakthroughs we have today.
杰夫:你录入未经加工的声波
JEFF DEAN: You’re taking invery raw audio waveforms.
男声:像是你从手机的麦克风上获得的
MALE SPEAKER: Like you getfrom a microphone on your phone
或其他的声波
or whatever.
男性电脑声音:奶酪汉堡
MALE COMPUTERVOICE: Cheeseburger.
弗朗索瓦丝:我们把它剁成小块
FRANCOISE BEAUFAYS: Wechop it into small pieces
然后它试着识别出
and it tries toidentify which phoneme
那段语音的最后一部分说的是哪个音位
was spoken in thatlast piece of speech.
杰弗里:因此一个音位是
GEOFFREY HINTON: So a phoneme is a kind
一种表达性语言的最小单位
of primitive unit for expressing words.
杰夫:[发出音位的声音]
JEFF DEAN:[SOUNDING OUT PHONEMES]
然后它会把那些音位连起来
JEFF DEAN: And then it willwant to stitch those together
变成最有可能的单词比如帕洛阿尔托
into likely wordslike Palo Alto.
雷:现在的语音识别很擅长
RAY KURZWEIL: Speech recognition today is quite good
识别你说的话
at transcribing what you’ve said.
男声:托皮卡天气怎么样?
MALE SPEAKER: What’s theweather like in Topeka?
罗伯托:你也可以谈论旅行
ROBERTO PIERACCINI: Youcan talk about travels.
或者可以谈论你的联系人
You can talk aboutyour contacts.
雷:比如 我能在哪儿买披萨?
RAY KURZWEIL: Likewhere can I get pizza? PHONE:
电话:这是披萨店名单
Here are thelistings for pizza.
雷:艾菲尔铁塔有多高?
RAY KURZWEIL: How tallis the Eiffel Tower? PHONE:
电话:艾菲尔铁塔……
The Eiffel Tower is–
弗朗索瓦丝:我们的提高显而易见
FRANCOISE BEAUFAYS: We’vemade tremendous improvements
日新月异
very quickly.
男声:谁是美国第21任总统?
MALE SPEAKER: Who is the 21stPresident of the United States? PHONE:
电话:切斯特阿瑟是第21任
Chester A.Arthur was the 21st–
男声:好的 谷歌
MALE SPEAKER: OK, Google.
他来自哪里?
Where’s he from?
雷:几年前 要与电脑交流
RAY KURZWEIL: Years ago,you had to be an engineer
你必须是一个工程师
to interact with computers.
现在 每个人都可以做到
I mean, today,everybody can interact.
罗伯托:然而 还有一件事
ROBERTO PIERACCINI: One thing, though,
电脑的理解力依然处在初期阶段
that is still in the infancy is the understanding.
杰弗里:我们需要更复杂的语言理解模型
GEOFFREY HINTON: We need a farmore sophisticated language
以期能够理解句子的意思
understanding modelthat understands what the sentence means.
要达到这个程度还有很远的路要走
And we’re still a verylong way from having that.
爱丽森:我们使用语言的能力是
ALISON GOPNIK: Our ability to use language is one
我们创造文明的一大助力
of the things that helps us have culture.
它帮助我们延续传统
It’s one of the things thathelps us pass on traditions
代代相续
from one generation to another.
了解语言系统如何工作
Figuring out about howthe system of language works,
即使那看起来是个很简单的问题
even though that seemslike a really easy problem,
其实也非常困难
it turns out to be one that’s very hard
但是每个婴儿在两岁时
but that every baby has cracked by the time
都会咿呀学语
they’re two years old.
女童:这有两个L
FEMALE CHILD: There’s two L’s.
女声:这有两个L
FEMALE SPEAKER: There’s two L’s.

Yeah.
E-L-L-I 然后
E-L-L-I and then–
女童:E
FEMALE CHILD: E.
女声:E
FEMALE SPEAKER: E.
罗伯托:语言非常复杂和令人费解
ROBERTO PIERACCINI:Language is extremely complex and sophisticated.
比尔:从语义学角度来说
BILL BYRNE: From the semantics
雷:反语
RAY KURZWEIL: Irony–
弗朗索瓦丝:重音男声:面部表情
FRANCOISE BEAUFAYS: Strong accents — MALE SPEAKER:facial expressions
雷:人类情绪
RAY KURZWEIL: Human emotion because
都是我们沟通的一部分
that’s part of how we communicate.
比尔:幽默
BILL BYRNE: Humor.
雷:难道我要小心别冒犯了恐龙吗?
RAY KURZWEIL: Do I have to be careful not to offend the dinosaur?
比尔:语言有太多不同层面
BILL BYRNE: Language hasso many different layers,
这也正是为什么它是个复杂的问题
and that’s why it’s sucha difficult problem.
杰弗里:现在 人类智力 人类大脑的算力
GEOFFREY HINTON: At present, thehuman brain, and the learning
比语言理解力这种能力
algorithms in the humanbrain, are far, far better
要好太多
at things likelanguage understanding.
然而人脑还是更善于模式识别
And they’re still a lotbetter at pattern recognition.
比尔:因此
BILL BYRNE: So
我们能否完全复制大脑对语言的理解方式
whether or not we replicate exactly what the brain does to understand language
与理解发音的模式还是个问题
and to understand speech,is still a question.
杰弗里:很多年来
GEOFFREY HINTON: Formany, many years,
我们相信神经网络要比
we believed that neuralnetworks should work better
主要通过表格查找的现存无声科技
than the dumb existingtechnology that’s
发挥着更好的作用
basically just table lookup.
在2009年 我的两个学生
And then in 2009,two of my students,
做的更好 我做的不多
with a little input fromme, got it working better.
一开始只是做的好一点
And the first time it justworked a little bit better,
但很明显
but then it was obvious
他们可以工作的更为出色并有所成就
that this could be developed to something that worked much better.
大脑有大量的神经元
The brain has thesegazillions of neurons
能同时进行多项运算
all computing in parallel.
并且大脑里所有的知识都包含在
And all of the knowledge in the brain is
神经元之间的连接强度上
in the strength of the connection between neurons.
我的意思是神经系统
What I mean by neuralnet is something
是指能用传统的电脑进行模仿的事物
that’s simulated on aconventional computer,
但是只是概略地提取信息
but is designed to workin very, very roughly
就像大脑的工作方式一样
the same way as the brain.
所以直到最近 人们仍然
So until quite recently,people got features
通过手工工程得到事物的特征
by hand engineering.
他们关注于声波 做傅里叶分析
They looked at sound waves,and they did Fourier analysis.
尝试计算出
And they tried to figure out,
我们应该把哪一部分培养成识别系统
what features should we feed to the pattern recognition system?
通过了解自己的特征来了解神经网络
And the thing about neural networks is they learn their own features.
尤其是他们能够学习
And in particular,they can learn features
这些信息的特点
and then they can learn features
通过不断地学习这些特点
of features and then they can learn features of features of features.
语音识别就会有巨大的提高
And that’s led toa huge improvement in speech recognition.
杰夫:同时你也可以用它们做语言识别任务
JEFF DEAN: But you can also use them for language understanding tasks.
你这样的做法
And the way you dothis is you represent
是在高维空间中表达语言
words in very highdimensional spaces.
杰弗里:我们现在可以进行处理
GEOFFREY HINTON: We cannow deal with analogies
用一连串数字表示一个词语的类比
where a word is representedas a list of numbers.
比如
So for example,
如果我给出100个数字
if I take the list of 100 numbers that
这些数字代表巴黎 然后减掉法国
represents Paris andI subtract from France
再加上意大利 然后看一下我得到的数字
France and I add to itItaly, and if I look
最有可能的是
at the numbers I’vegot, the closest thing
这一串数字代表罗马
is the list of numbersthat represents Rome.
所以 第一次用神经元转换这些数字时
So by first converting wordsinto these numbers using
你是可以推断出结果的
a neural net, you can actuallydo this analogical reasoning.
我认为5年内
I predict that in the next five years,
这些神经系统就会
it will become clear that these new big deep neural networks
利用新算法学习
with the new learningalgorithms are going to give us
使我们更好地理解语言
much better languageunderstanding.
爱丽森:开始的时候
ALISON GOPNIK: When we started out,
我们认为像象棋和数学或者逻辑这样的事情
we thought that things like chess or mathematics
它们都是
or logic, those were goingto be the things that
很难模仿的
were really hard.
其实没那么难
They’re not that hard.
我是说 我们可以利用机器
I mean, we can end up witha machine that actually
它们实际上能把象棋下的和大师一个水平
can do chess as well as agrandmaster can play chess.
我们的思考方式
The things that we thought were
对于电脑来说太过简单了
going to be easy for a computer system,
但像理解语言这一类事
like understandinglanguage, those things
被证明对电脑非常难
have turned out tobe incredibly hard.
比尔:我简直不能想象“我们做到了”
BILL BYRNE: I can’t evenimagine the”we’ve done it”
在这个非常时刻 因为
moment quite yet, just because there are so
还有许多没解决的棘手问题
many pieces of this puzzle that are unsolved,
从科学的角度来看
both from a science point of view,
也可以从科技实现的角度来看
as well as from a technicalimplementation point of view.
有很多的未知数
There’s a lot of unknowns.
爱丽森:那些是很大的变革
ALISON GOPNIK: Those arethe great revolutions.
这不仅仅是
They’re not just when we
我们用已知的知识研究出的成果
fiddle a little with what we already know,
而是我们发现了一些全新的不可预测的事物
but when wediscover something completely new and unexpected.
杰夫:我想一旦电脑
JEFF DEAN: I think once you kind
能达到人们正常表现的范围
of are in the ballpark of human normal performance, that
那会是非常了不起的
will be pretty remarkable.
[播放音乐]
[MUSIC PLAYING]

发表评论

译制信息
视频概述

在科技发展日新月异的今天,我们开始关注于与电脑的交流。试图通过语言这种更为简单直接的方式下达指令,那么电脑又是如何做到接受指令,完成任务的呢?

听录译者

收集自网络

翻译译者

NALIEZM

审核员

审核员 D

视频来源

https://www.youtube.com/watch?v=yxxRAHVtafI

相关推荐