第三十六章自然语言处理-Natural Language Processing

翻译内容

Hi, I'm Carrie Anne, and welcome to Crash Course Computer Science!
(｡･∀･)ﾉﾞ嗨 我是Carrie Anne，欢迎收看计算机科学速成课

Last episode we talked about computer vision
上集我们讨论了计算机视觉 - 让电脑能看到并理解

• giving computers the ability to see and understand visual information.
  上集我们讨论了计算机视觉 - 让电脑能看到并理解

Today we're going to talk about how to give computers the ability to understand language.
今天我们讨论 怎么让计算机理解语言

You might argue they've always had this capability.
你可能会说：计算机已经有这个能力了

Back in Episodes 9 and 12,
在第9和第12集

we talked about machine language instructions,
我们聊了机器语言和更高层次的编程语言

as well as higher-level programming languages.
我们聊了机器语言和更高层次的编程语言

While these certainly meet the definition of a language,
虽然从定义来说 它们也算语言

they also tend to have small vocabularies and follow highly structured conventions.
但词汇量一般很少，而且非常结构化

Code will only compile and run if it's 100 percent free of spelling and syntactic errors.
代码只能在拼写和语法完全正确时，编译和运行

Of course, this is quite different from human languages
当然，这和人类语言完全不同 \N - 人类语言叫"自然语言"

• what are called natural languages -
  当然，这和人类语言完全不同 \N - 人类语言叫"自然语言"

containing large, diverse vocabularies,
自然语言有大量词汇

words with several different meanings,
有些词有多种含义

speakers with different accents,
不同口音

and all sorts of interesting word play.
以及各种有趣的文字游戏

People also make linguistic faux pas when writing and speaking,
人们在写作和说话时也会犯错

like slurring words together, leaving out key details so things are ambiguous,
比如单词拼在一起发音 \N 关键细节没说 导致意思模糊两可

and mispronouncing things.
以及发错音

But, for the most part, humans can roll right through these challenges.
但大部分情况下，另一方能理解

The skillful use of language is a major part of what makes us human.
人类有强大的语言能力

And for this reason,
因此，让计算机拥有语音对话的能力

the desire for computers to understand and speak our language
因此，让计算机拥有语音对话的能力

has been around since they were first conceived.
这个想法从构思计算机时就有了

This led to the creation of Natural Language Processing, or NLP,
"自然语言处理"因此诞生，简称 NLP

an interdisciplinary field combining computer science and linguistics.
结合了计算机科学和语言学的 一个跨学科领域

There's an essentially infinite number of ways to arrange words in a sentence.
单词组成句子的方式有无限种

We can't give computers a dictionary of all possible sentences
我们没法给计算机一个字典，包含所有可能句子

to help them understand what humans are blabbing on about.
让计算机理解人类在嘟囔什么

So an early and fundamental NLP problem was deconstructing sentences into bite-sized pieces,
所以 NLP 早期的一个基本问题是 \N 怎么把句子切成一块块

which could be more easily processed.
这样更容易处理

In school, you learned about nine fundamental types of English words:
上学时，老师教你 英语单词有九种基本类型：

nouns, pronouns, articles, verbs, adjectives,
名词，代词，冠词，动词，形容词

adverbs, prepositions, conjunctions, and interjections.
副词，介词，连词和感叹词

These are called parts of speech.
这叫"词性"

There are all sorts of subcategories too,
还有各种子类，比如

like singular vs. plural nouns and superlative vs. comparative adverbs,
单数名词 vs 复数名词 \N 副词最高级 vs 副词比较级

but we're not going to get into that.
但我们不会深入那些.

Knowing a word's type is definitely useful,
了解单词类型有用

but unfortunately, there are a lot words that have multiple meanings - like "rose" and "leaves",
但不幸的是，很多词有多重含义 比如 rose 和 leaves

which can be used as nouns or verbs.
可以用作名词或动词

A digital dictionary alone isn't enough to resolve this ambiguity,
仅靠字典，不能解决这种模糊问题

so computers also need to know some grammar.
所以电脑也要知道语法

For this, phrase structure rules were developed, which encapsulate the grammar of a language.
因此开发了 "短语结构规则" 来代表语法规则

For example, in English there's a rule
例如，英语中有一条规则

that says a sentence can be comprised of a noun phrase followed by a verb phrase.
句子可以由一个名词短语和一个动词短语组成

Noun phrases can be an article, like "the",
名词短语可以是冠词，如 the

followed by a noun or they can be an adjective followed by a noun.
然后一个名词，或一个形容词后面跟一个名词

And you can make rules like this for an entire language.
你可以给一门语言制定出一堆规则

Then, using these rules, it's fairly easy to construct what's called a parse tree,
用这些规则，可以做出"分析树"

which not only tags every word with a likely part of speech,
它给每个单词标了可能是什么词性

but also reveals how the sentence is constructed.
也标明了句子的结构

These smaller chunks of data allow computers to more easily access,
数据块更小 更容易处理

process and respond to information.
数据块更小 更容易处理

Equivalent processes are happening every time you do a voice search,
每次语音搜索，都有这样的流程

like: "where's the nearest pizza".
比如 "最近的披萨在哪里"

The computer can recognize that this is a "where" question,
计算机能明白这是"哪里"（where）的问题

knows you want the noun "pizza",
知道你想要名词"披萨"（pizza）

and the dimension you care about is "nearest".
而且你关心的维度是"最近的"（nearest）

The same process applies to "what is the biggest giraffe?" or "who sang thriller?"
"最大的长颈鹿是什么？"或"Thriller是谁唱的？" \N 也是这样处理

By treating language almost like lego,
把语言像乐高一样拆分，方便计算机处理

computers can be quite adept at natural language tasks.
把语言像乐高一样拆分，方便计算机处理

They can answer questions and also process commands,
计算机可以回答问题 以及处理命令

like "set an alarm for 2:20"
比如"设 2:20 的闹钟"

or "play T-Swizzle on spotify".
或"用 Spotify 播放 T-Swizzle"

But, as you've probably experienced, they fail when you start getting too fancy,
但你可能体验过，如果句子复杂一点

and they can no longer parse the sentence correctly, or capture your intent.
计算机就没法理解了

Hey Siri... me thinks the mongols doth roam too much,
嘿Siri ...... 俺觉得蒙古人走得太远了

what think ye on this most gentle mid-summer's day?
在这个最温柔的夏日的日子里，你觉得怎么样？

Siri: I'm not sure I got that.
Siri：我没明白

I should also note that phrase structure rules, and similar methods that codify language,
还有，"短语结构规则"和其他把语言结构化的方法

can be used by computers to generate natural language text.
可以用来生成句子

This works particularly well when data is stored in a web of semantic information,
数据存在语义信息网络时，这种方法特别有效

where entities are linked to one another in meaningful relationships,
实体互相连在一起

providing all the ingredients you need to craft informational sentences.
提供构造句子的所有成分

Siri: Thriller was released in 1983 and sung by Michael Jackson
Siri：Thriller 于1983年发行，由迈克尔杰克逊演唱

Google's version of this is called Knowledge Graph.
Google 版的叫"知识图谱"

At the end of 2016,
在2016年底

it contained roughly seventy billion facts about, and relationships between, different entities.
包含大概七百亿个事实，以及不同实体间的关系

These two processes, parsing and generating text,
处理, 分析, 生成文字 \N 是聊天机器人的最基本部件

are fundamental components of natural language chatbots
处理, 分析, 生成文字 \N 是聊天机器人的最基本部件

• computer programs that chat with you.
• 聊天机器人就是能和你聊天的程序

Early chatbots were primarily rule-based,
早期聊天机器人大多用的是规则.

where experts would encode hundreds of rules mapping what a user might say,
专家把用户可能会说的话，和机器人应该回复什么，\N 写成上百个规则

to how a program should reply.
专家把用户可能会说的话，和机器人应该回复什么，\N 写成上百个规则

Obviously this was unwieldy to maintain and limited the possible sophistication.
显然，这很难维护，而且对话不能太复杂.

A famous early example was ELIZA, created in the mid-1960s at MIT.
一个著名早期例子叫 Eliza\N 1960年代中期 诞生于麻省理工学院

This was a chatbot that took on the role of a therapist,
一个治疗师聊天机器人

and used basic syntactic rules to identify content in written exchanges,
它用基本句法规则 来理解用户打的文字

which it would turn around and ask the user about.
然后向用户提问

Sometimes, it felt very much like human-human communication,
有时候会感觉像和人类沟通一样

but other times it would make simple and even comical mistakes.
但有时会犯简单 甚至很搞笑的错误

Chatbots, and more advanced dialog systems,
聊天机器人和对话系统

have come a long way in the last fifty years, and can be quite convincing today!
在过去五十年发展了很多，如今可以和真人很像!

Modern approaches are based on machine learning,
如今大多用机器学习

where gigabytes of real human-to-human chats are used to train chatbots.
用上GB的真人聊天数据 来训练机器人

Today, the technology is finding use in customer service applications,
现在聊天机器人已经用于客服回答

where there's already heaps of example conversations to learn from.
客服有很多对话可以参考

People have also been getting chatbots to talk with one another,
人们也让聊天机器人互相聊天

and in a Facebook experiment, chatbots even started to evolve their own language.
在 Facebook 的一个实验里，\N 聊天机器人甚至发展出自己的语言

This experiment got a bunch of scary-sounding press,
很多新闻把这个实验 报导的很吓人

but it was just the computers crafting a simplified protocol to negotiate with one another.
但实际上只是计算机 \N 在制定简单协议来帮助沟通

It wasn't evil, it's was efficient.
这些语言不是邪恶的，而是为了效率

But what about if something is spoken
但如果听到一个句子

• how does a computer get words from the sound?
• 计算机怎么从声音中提取词汇？

That's the domain of speech recognition,
这个领域叫"语音识别"

which has been the focus of research for many decades.
这个领域已经重点研究了几十年

Bell Labs debuted the first speech recognition system in 1952,
贝尔实验室在1952年推出了第一个语音识别系统

nicknamed Audrey, the automatic digit recognizer.
绰号 Audrey，自动数字识别器

It could recognize all ten numerical digits,
如果你说得够慢，它可以识别全部十位数字

if you said them slowly enough.
如果你说得够慢，它可以识别全部十位数字

5

9

7？

The project didn't go anywhere
这个项目没有实际应用，因为手输快得多

because it was much faster to enter telephone numbers with a finger.
这个项目没有实际应用，因为手输快得多

Ten years later, at the 1962 World's Fair,
十年后，1962年的世界博览会上

IBM demonstrated a shoebox-sized machine capable of recognizing sixteen words.
IBM展示了一个鞋盒大小的机器，能识别16个单词

To boost research in the area,
为了推进"语音识别"领域的研究

DARPA kicked off an ambitious five-year funding initiative in 1971,
DARPA 在1971年启动了一项雄心勃勃的五年筹资计划

which led to the development of Harpy at Carnegie Mellon University.
之后诞生了卡内基梅隆大学的 Harpy

Harpy was the first system to recognize over a thousand words.
Harpy 是第一个可以识别1000个单词以上的系统

But, on computers of the era,
但那时的电脑

transcription was often ten or more times slower than the rate of natural speech.
语音转文字，经常比实时说话要慢十倍或以上

Fortunately, thanks to huge advances in computing performance in the 1980s and 90s,
幸运的是，1980,1990年代 计算机性能的大幅提升

continuous, real-time speech recognition became practical.
实时语音识别变得可行

There was simultaneous innovation in the algorithms for processing natural language,
同时也出现了处理自然语言的新算法

moving from hand-crafted rules,
不再是手工定规则

to machine learning techniques
而是用机器学习

that could learn automatically from existing datasets of human language.
从语言数据库中学习

Today, the speech recognition systems with the best accuracy are using deep neural networks,
如今准确度最高的语音识别系统 用深度神经网络

which we touched on in Episode 34.
我们在第34集讲过

To get a sense of how these techniques work,
为了理解原理

let's look at some speech, specifically,
我们来看一些对话声音

the acoustic signal.
我们来看一些对话声音

Let's start by looking at vowel sounds,
先看元音

like aaaaa and eeeeee.
比如 a 和 e

These are the waveforms of those two sounds, as captured by a computer's microphone.
这是两个声音的波形

As we discussed in Episode 21 - on Files and File Formats -
我们在第21集（文件格式）说过

this signal is the magnitude of displacement,
这个信号来自 麦克风内部隔膜震动的频率

of a diaphragm inside of a microphone, as sound waves cause it to oscillate.
这个信号来自 麦克风内部隔膜震动的频率

In this view of sound data, the horizontal axis is time,
在这个视图中，横轴是时间

and the vertical axis is the magnitude of displacement, or amplitude.
竖轴是隔膜移动的幅度，或者说振幅

Although we can see there are differences between the waveforms,
虽然可以看到2个波形有区别

it's not super obvious what you would point at to say,
但不能看出

"ah ha! this is definitely an eeee sound".
"啊！这个声音肯定是 e"

To really make this pop out, we need to view the data in a totally different way:
为了更容易识别，我们换个方式看：

a spectrogram.
谱图

In this view of the data, we still have time along the horizontal axis,
这里横轴还是时间

but now instead of amplitude on the vertical axis,
但竖轴不是振幅

we plot the magnitude of the different frequencies that make up each sound.
而是不同频率的振幅

The brighter the color, the louder that frequency component.
颜色越亮，那个频率的声音越大

This conversion from waveform to frequencies is done with a very cool algorithm called
这种波形到频率的转换 是用一种很酷的算法做的

a Fast Fourier Transform.
快速傅立叶变换（FFT）

If you've ever stared at a stereo system's EQ visualizer,
如果你盯过立体声系统的 EQ 可视化器

it's pretty much the same thing.
它们差不多是一回事

A spectrogram is plotting that information over time.
谱图是随着时间变化的

You might have noticed that the signals have a sort of ribbed pattern to them
你可能注意到，信号有种螺纹图案

that's all the resonances of my vocal tract.
那是我声道的回声

To make different sounds,
为了发出不同声音

I squeeze my vocal chords, mouth and tongue into different shapes,
我要把声带，嘴巴和舌头变成不同形状

which amplifies or dampens different resonances.
放大或减少不同的共振

We can see this in the signal, with areas that are brighter, and areas that are darker.
可以看到有些区域更亮，有些更暗

If we work our way up from the bottom, labeling where we see peaks in the spectrum
如果从底向上看，标出高峰

• what are called formants -
• 叫"共振峰" -

we can see the two sounds have quite different arrangements.
可以看到有很大不同

And this is true for all vowel sounds.
所有元音都是如此

It's exactly this type of information that lets computers recognize spoken vowels,
这让计算机可以识别元音

and indeed, whole words.
然后识别出整个词

Let's see a more complicated example,
让我们看一个更复杂的例子

like when I say: "she.. was.. happy"
当我说"她..很开心"的时候

We can see our "eee" sound here, and "aaa" sound here.
可以看到 e 声，和 a 声

We can also see a bunch of other distinctive sounds,
以及其它不同声音

like the "shh" sound in "she",
比如 she 中的 shh 声

the "wah" and "sss" in "was", and so on.
was 中的 wah 和 sss，等等

These sound pieces, that make up words,
这些构成单词的声音片段

are called phonemes.
叫"音素"

Speech recognition software knows what all these phonemes look like.
语音识别软件 知道这些音素

In English, there are roughly forty-four,
英语有大概44种音素

so it mostly boils down to fancy pattern matching.
所以本质上变成了音素识别

Then you have to separate words from one another,
还要把不同的词分开

figure out when sentences begin and end...
弄清句子的开始和结束点

and ultimately, you end up with speech converted into text,
最后把语音转成文字

allowing for techniques like we discussed at the beginning of the episode.
使这集视频开头里讨论的那些技术成为可能

Because people say words in slightly different ways,
因为口音和发音错误等原因

due to things like accents and mispronunciations,
人们说单词的方式略有不同

transcription accuracy is greatly improved when combined with a language model,
所以结合语言模型后，语音转文字的准确度会大大提高

which contains statistics about sequences of words.
里面有单词顺序的统计信息

For example "she was" is most likely to be followed by an adjective, like "happy".
比如："她"后面很可能跟一个形容词，\N 比如"很开心"

It's uncommon for "she was" to be followed immediately by a noun.
"她"后面很少是名词

So if the speech recognizer was unsure between, "happy" and "harpy",
如果不确定是 happy 还是 harpy，会选 happy

it'd pick "happy",
如果不确定是 happy 还是 harpy，会选 happy

since the language model would report that as a more likely choice.
因为语言模型认为可能性更高

Finally, we need to talk about Speech Synthesis,
最后, 我们来谈谈 "语音合成"

that is, giving computers the ability to output speech.
让计算机输出语音

This is very much like speech recognition, but in reverse.
它很像语音识别，不过反过来

We can take a sentence of text, and break it down into its phonetic components,
把一段文字，分解成多个声音

and then play those sounds back to back, out of a computer speaker.
然后播放这些声音

You can hear this chaining of phonemes very clearly with older speech synthesis technologies,
早期语音合成技术，可以清楚听到音素是拼在一起的

like this 1937, hand-operated machine from Bell Labs.
比如这个1937年贝尔实验室的手动操作机器

Say, "she saw me" with no expression.
不带感情的说"她看见了我"

She saw me.
她看见了我

Now say it in answer to these questions.
现在回答问题

Who saw you?
谁看见你了？

She saw me.
她看见了我

Who did she see?
她看到了谁？

She saw me.
她看见了我

Did she see you or hear you?
她看到你还是听到你说话了？

She saw me.
她看见了我

By the 1980s, this had improved a lot,
到了1980年代，技术改进了很多

but that discontinuous and awkward blending of phonemes
但音素混合依然不够好，产生明显的机器人声

still created that signature, robotic sound.
但音素混合依然不够好，产生明显的机器人声

Thriller was released in 1983 and sung by Michael Jackson.
Thriller 于1983年发行，迈克尔·杰克逊 演唱.

Today, synthesized computer voices, like Siri, Cortana and Alexa,
如今，电脑合成的声音，比如 Siri, Cortana, Alexa

have gotten much better, but they're still not quite human.
好了很多，但还不够像人

But we're soo soo close,
但我们非常非常接近了

and it's likely to be a solved problem pretty soon.
这个问题很快会被解决

Especially because we're now seeing an explosion of voice user interfaces on our phones,
现在语音界面到处都是，手机里

in our cars and homes, and maybe soon, plugged right into our ears.
汽车里，家里，也许不久之后耳机也会有.

This ubiquity is creating a positive feedback loop,
这创造一个正反馈循环

where people are using voice interaction more often,
人们用语音交互的频率会提高

which in turn, is giving companies like Google, Amazon and Microsoft
这又给了谷歌，亚马逊，微软等公司

more data to train their systems on.
更多数据来训练语音系统.

Which is enabling better accuracy,
提高准确性

which is leading to people using voice more,
准确度高了，人们更愿意用语音交互

which is enabling even better accuracy and the loop continues!
越用越好，越好越用

Many predict that speech technologies will become as common a form of interaction
很多人预测，语音交互会越来越常见

as screens, keyboards, trackpads and other physical input-output devices that we use today.
就像如今的屏幕，键盘，触控板等设备

That's particularly good news for robots,
这对机器人发展是个好消息

who don't want to have to walk around with keyboards in order to communicate with humans.
机器人就不用走来走去时 带个键盘和人类沟通

But, we'll talk more about them next week. See you then.
下周我们讲机器人. 到时见