Original subtitles

Alright, so we're not yet to the point where we can tell intentionally funny stories.

What can we do with language?

Well I mentioned Siri. Siri may not be the best at telling bedtime stories

but Siri does some amazing things, and the pieces that make that up are

actually used in many places in industry. There's automatic speech recognition,

where you go from speech to text. There's text-to-speech synthesis which

is an easier problem where you go from the text to the speech. And then there's dialogue systems

that integrate all this together, linguistic analysis.

Let me show you what a

speech recognition system looks like

just kind of when you point it at the TV. So this is not customized to a

specific speaker, this is not over some great microphone like how your phones

have really sophisticated microphones these days.

This is just plugged straight into the TV as essentially automatic transcription. Let's see

how well it does, and in particular watch the errors.

So, what's interesting about this. First of all,

is it good? Is it bad?

It does a lot of stuff.

It does a lot of things right. It makes some mistakes.

The linguistics are of multiple kinds, so for example, here

"The classmates said their final goodbyes".

That's like good buys like Best Buy. right. That is exactly the sounds that

the reporter said.

The failing here in this case was not in the acoustic modeling which tries to

connect

the wave forms to the underlying linguistic sounds.

Here the failing is there multiple things that sound the same.

You gotta figure which one the reporter could possibly mean in the context.

This is a sad story right. Somebody died, people are not going shopping, right, and

we know this is humans, but the system does not and so in this case this

isn't a problem the language model. there are other cases here where the problem is

more in the acoustics and putting all the stuff together in some probabilistic framework

that lets you reconcile it all,

that's a big part of how speech recognition works.

We'll have more discussion on that later.

We can do more with language than just manipulate the signal from speech to text.

This is actually my research area. We can do things like question answering.

We talked a little about Watson and we'll have a lot more

later in the course about Watson.

So Watson is basically a question answering system.

Like, yeah, there's this layer of remembering to phrase it as a question 'cause

you're on Jeopardy and making sure you wager the right amount on the

Daily Double and that kinda stuff but to a first approximation a question comes in,

Watson kind of has to

dig through a lot of information like you know largely Wikipedia

and connect up some answer to the question,

so that you know how to respond. Basically a question answering system,

although an amazingly cool demonstration of a very good one.

Another thing we can do is machine translation. How many of you have used

a tool like Google translate?

So, you know, again, C-3PO. How good is machine translation?

Well, depends on the language pair. I mean, if I'm looking at a page, say in Chinese,

and I don't speak any Chinese, the machine translation's pretty good because I was kind

of starting with nothing. But if I actually speak the language maybe

I'm better off reading reading it in its natural form. You can see some of these problems if

you do round trip from say, English to Chinese and back, and you can see how

good what comes back--actually that's a good way to make an unintentionally funny story.

What else can we do: things like web search, really are about a lot of things.

It has something to do with the words but also kind of click stream information,

and kind of a local search and things like that. And so there's a lot that goes into

web search. A big part of that is the language.

Text classification, spam filtering. Again, spam filtering is a case where it's

part language, part not language. We'll talk more about spam filtering later

--and so on. These are the kinds of things you can do in the domain of natural language.

We're no longer trying so hard to tell stories funny or otherwise.

We're trying to build things like this.

And there has been a lot of traction. There's a lot of stuff we can build.

We're not yet to C-3PO, but

we actually can now translate Russian,

which we couldn't do in the fifties even though they thought would be able to do

it by the sixties. But now, today, we can.

It only took

something like twelve times longer than they thought it would.