0:00:05.734,0:00:08.234
Whitelaw: Hi. My name[br]is Casey Whitelaw.

0:00:08.234,0:00:09.234
I'm the Tech Lead

0:00:09.234,0:00:10.567
for the Natural Language[br]Processing Group

0:00:10.567,0:00:12.501
here in Sydney,[br]and today I'm gonna talk to you

0:00:12.501,0:00:13.901
a little bit about

0:00:13.901,0:00:16.601
some of the cool things[br]that we've added to Google Wave.

0:00:16.601,0:00:18.834
So one of the main things

0:00:18.834,0:00:21.834
that we want to stay focused on[br]in Google Wave is productivity.

0:00:21.834,0:00:23.634
We want users to be able[br]to stay productive,

0:00:23.634,0:00:25.734
whether they're reading[br]or whether they're writing.

0:00:25.734,0:00:27.601
One of the ways[br]that we've done that

0:00:27.601,0:00:29.267
is with our[br]spell correction system.

0:00:29.267,0:00:32.234
What we'd like is for users[br]just to be able to

0:00:32.234,0:00:34.534
focus on what they're typing[br]and not worry about

0:00:34.534,0:00:36.801
whether there's any mistakes[br]they've made.

0:00:36.801,0:00:39.000
We think that if people could[br]just loosen up a little bit

0:00:39.000,0:00:41.100
and, you know,[br]or maybe type 5% faster,

0:00:41.100,0:00:43.400
then that's 5% less time[br]that they spend typing.

0:00:43.400,0:00:45.734
So I'll start with an example.

0:00:45.734,0:00:47.801
It's probably the easiest way[br]to explain.

0:00:47.801,0:00:51.133
Let's say you want to meet up[br]with one of your friends.

0:00:51.133,0:00:52.300
You're having a chat.

0:00:52.300,0:00:54.534
So you write...

0:00:54.534,0:00:56.467
Let's...

0:00:56.467,0:00:58.434
met...

0:00:58.434,0:01:00.067
whoops...

0:01:00.067,0:01:03.767
tomorrow.

0:01:03.767,0:01:05.567
So here you see[br]I've made a mistake.

0:01:05.567,0:01:07.400
I've written met[br]instead of meet here.

0:01:07.400,0:01:09.501
My finger slipped on the "e."

0:01:09.501,0:01:13.534
So now, the way that we[br]implemented spelling

0:01:13.534,0:01:16.968
is we introduced an automatic[br]participant called Spelly

0:01:16.968,0:01:19.767
who works just like[br]another user

0:01:19.767,0:01:21.767
that's participating[br]on the wave with you.

0:01:21.767,0:01:24.334
So Spelly's on your wave[br]with you,

0:01:24.334,0:01:29.267
and it can see that you've[br]typed "Let's met tomorrow,"

0:01:29.267,0:01:31.367
and it's now gonna try[br]and spell-check it.

0:01:31.367,0:01:32.901
For each word...

0:01:32.901,0:01:35.534
it doesn't have any kind[br]of dictionary,

0:01:35.534,0:01:39.534
so it doesn't know whether[br]met is a well-spelled word

0:01:39.534,0:01:40.834
or a misspelling.

0:01:40.834,0:01:43.100
So to start with,[br]it comes up with a list

0:01:43.100,0:01:46.634
of possible candidate[br]corrections for this word.

0:01:46.634,0:01:50.367
So some examples of that[br]might be...

0:01:50.367,0:01:53.400
meat, the food...

0:01:53.400,0:01:56.868
or meet, the correctly[br]spelled version of this.

0:01:56.868,0:01:59.367
And you can imagine[br]lots of others.

0:01:59.367,0:02:02.200
So set or net or me--

0:02:02.200,0:02:05.367
all kinds of different words[br]that we would evaluate

0:02:05.367,0:02:09.199
to see whether they're what[br]you actually meant to type.

0:02:09.199,0:02:12.734
We've learned from the web

0:02:12.734,0:02:14.934
the kind of misspellings[br]that people make

0:02:14.934,0:02:17.200
and which things[br]are more and less likely.

0:02:17.200,0:02:18.834
So we know that,[br]for instance,

0:02:18.834,0:02:20.634
maybe slipping[br]and inserting an "A"

0:02:20.634,0:02:22.300
is relatively likely,

0:02:22.300,0:02:25.434
but misspelling[br]the very first letter

0:02:25.434,0:02:28.467
might be less likely[br]in this case.

0:02:28.467,0:02:32.667
So we've got some suggestions,[br]and the next thing that we do

0:02:32.667,0:02:34.667
is evaluate these suggestions[br]in context.

0:02:34.667,0:02:38.067
So there are other systems[br]at Google that already use

0:02:38.067,0:02:40.434
the same kind of statistical[br]language models as this,

0:02:40.434,0:02:42.467
such as the Google[br]translation system,

0:02:42.467,0:02:44.367
that essentially[br]encode information

0:02:44.367,0:02:46.133
about how language is used.

0:02:46.133,0:02:47.968
These are learned from the web

0:02:47.968,0:02:49.734
from looking at billions[br]of web pages,

0:02:49.734,0:02:51.167
so we get a really good idea

0:02:51.167,0:02:53.734
about the way that people[br]really use language in practice.

0:02:53.734,0:02:55.434
So what we would do

0:02:55.434,0:02:58.834
is look at the likelihood[br]of "Let's met tomorrow"

0:02:58.834,0:03:02.067
and "Let's meat tomorrow,"[br]less likely,

0:03:02.067,0:03:03.767
and "Let's meet tomorrow,"

0:03:03.767,0:03:06.267
which is gonna be more likely[br]than either of these.

0:03:06.267,0:03:08.067
And we combine that[br]with our error model

0:03:08.067,0:03:10.200
which tells us how likely[br]the misspellings are,

0:03:10.200,0:03:13.834
you know, without any context,[br]to get a final determination

0:03:13.834,0:03:15.901
as to what are[br]the most likely words--

0:03:15.901,0:03:19.200
most likely word[br]that you meant right here.

0:03:19.200,0:03:22.000
So in this case,[br]we would suggest meet.

0:03:22.000,0:03:25.467
Once we think[br]that a word is misspelled,

0:03:25.467,0:03:29.100
we need to get that back[br]to the Google Wave client

0:03:29.100,0:03:31.667
so that the user[br]can actually see it

0:03:31.667,0:03:34.701
and either correct it[br]automatically or manually.

0:03:34.701,0:03:36.067
Two kinds of ways

0:03:36.067,0:03:38.834
that this differs[br]from existing spelling systems.

0:03:38.834,0:03:42.000
One of them is just that[br]it's hosted.

0:03:42.000,0:03:44.100
And this means that we can do

0:03:44.100,0:03:46.033
this same kind of spelling[br]for you,

0:03:46.033,0:03:49.434
regardless of which device[br]you're connecting from.

0:03:49.434,0:03:52.767
So whether you're on your laptop[br]or your mobile or your desktop,

0:03:52.767,0:03:56.000
we can give the same[br]quality spelling, regardless.

0:03:56.000,0:03:57.901
And that applies[br]across languages too,

0:03:57.901,0:03:59.467
so, you know, we're doing this

0:03:59.467,0:04:01.267
for other alphabetic[br]languages also.

0:04:01.267,0:04:06.667
So like I said, we use large[br]statistical language models.

0:04:06.667,0:04:07.801
When I said large, you know,

0:04:07.801,0:04:09.801
we train them[br]from billions of words.

0:04:09.801,0:04:12.033
They end up being[br]many, many gigabytes.

0:04:12.033,0:04:15.501
It's pretty infeasible to run[br]these on a single machine,

0:04:15.501,0:04:17.601
which isn't such a problem[br]in a data center

0:04:17.601,0:04:19.233
where you can have[br]a set of machines

0:04:19.233,0:04:22.200
running a language model[br]and a spelling model together.

0:04:22.200,0:04:27.167
And then we can share[br]that spelling model

0:04:27.167,0:04:28.734
between many users

0:04:28.734,0:04:30.667
so that the cost per user[br]is very low.

0:04:30.667,0:04:33.534
So it's very efficient[br]for us to do this.

0:04:33.534,0:04:35.667
Once you realize[br]that you've got a system

0:04:35.667,0:04:37.801
that supports[br]collaborative editing,

0:04:37.801,0:04:39.501
that has structured data,

0:04:39.501,0:04:42.667
and that you can change[br]the user interface

0:04:42.667,0:04:45.167
by having remote participants,

0:04:45.167,0:04:47.200
then, really,[br]the sky's the limit.

0:04:47.200,0:04:49.434
I mean, there's all kinds[br]of existing

0:04:49.434,0:04:51.400
natural language technologies[br]like spell checking

0:04:51.400,0:04:53.100
or translation[br]that we can apply,

0:04:53.100,0:04:55.667
and we're seeing[br]a lot of new applications

0:04:55.667,0:04:57.667
as the way that we communicate[br]changes as well.

0:04:57.667,0:05:00.868
So, you know, really,[br]it's gonna be exciting times.