WEBVTT

00:00:05.734 --> 00:00:08.234
Whitelaw: Hi. My name
is Casey Whitelaw.

00:00:08.234 --> 00:00:09.234
I'm the Tech Lead

00:00:09.234 --> 00:00:10.567
for the Natural Language
Processing Group

00:00:10.567 --> 00:00:12.501
here in Sydney,
and today I'm gonna talk to you

00:00:12.501 --> 00:00:13.901
a little bit about

00:00:13.901 --> 00:00:16.601
some of the cool things
that we've added to Google Wave.

00:00:16.601 --> 00:00:18.834
So one of the main things

00:00:18.834 --> 00:00:21.834
that we want to stay focused on
in Google Wave is productivity.

00:00:21.834 --> 00:00:23.634
We want users to be able
to stay productive,

00:00:23.634 --> 00:00:25.734
whether they're reading
or whether they're writing.

00:00:25.734 --> 00:00:27.601
One of the ways
that we've done that

00:00:27.601 --> 00:00:29.267
is with our
spell correction system.

00:00:29.267 --> 00:00:32.234
What we'd like is for users
just to be able to

00:00:32.234 --> 00:00:34.534
focus on what they're typing
and not worry about

00:00:34.534 --> 00:00:36.801
whether there's any mistakes
they've made.

00:00:36.801 --> 00:00:39.000
We think that if people could
just loosen up a little bit

00:00:39.000 --> 00:00:41.100
and, you know,
or maybe type 5% faster,

00:00:41.100 --> 00:00:43.400
then that's 5% less time
that they spend typing.

00:00:43.400 --> 00:00:45.734
So I'll start with an example.

00:00:45.734 --> 00:00:47.801
It's probably the easiest way
to explain.

00:00:47.801 --> 00:00:51.133
Let's say you want to meet up
with one of your friends.

00:00:51.133 --> 00:00:52.300
You're having a chat.

00:00:52.300 --> 00:00:54.534
So you write...

00:00:54.534 --> 00:00:56.467
Let's...

00:00:56.467 --> 00:00:58.434
met...

00:00:58.434 --> 00:01:00.067
whoops...

00:01:00.067 --> 00:01:03.767
tomorrow.

00:01:03.767 --> 00:01:05.567
So here you see
I've made a mistake.

00:01:05.567 --> 00:01:07.400
I've written met
instead of meet here.

00:01:07.400 --> 00:01:09.501
My finger slipped on the "e."

00:01:09.501 --> 00:01:13.534
So now, the way that we
implemented spelling

00:01:13.534 --> 00:01:16.968
is we introduced an automatic
participant called Spelly

00:01:16.968 --> 00:01:19.767
who works just like
another user

00:01:19.767 --> 00:01:21.767
that's participating
on the wave with you.

00:01:21.767 --> 00:01:24.334
So Spelly's on your wave
with you,

00:01:24.334 --> 00:01:29.267
and it can see that you've
typed "Let's met tomorrow,"

00:01:29.267 --> 00:01:31.367
and it's now gonna try
and spell-check it.

00:01:31.367 --> 00:01:32.901
For each word...

00:01:32.901 --> 00:01:35.534
it doesn't have any kind
of dictionary,

00:01:35.534 --> 00:01:39.534
so it doesn't know whether
met is a well-spelled word

00:01:39.534 --> 00:01:40.834
or a misspelling.

00:01:40.834 --> 00:01:43.100
So to start with,
it comes up with a list

00:01:43.100 --> 00:01:46.634
of possible candidate
corrections for this word.

00:01:46.634 --> 00:01:50.367
So some examples of that
might be...

00:01:50.367 --> 00:01:53.400
meat, the food...

00:01:53.400 --> 00:01:56.868
or meet, the correctly
spelled version of this.

00:01:56.868 --> 00:01:59.367
And you can imagine
lots of others.

00:01:59.367 --> 00:02:02.200
So set or net or me--

00:02:02.200 --> 00:02:05.367
all kinds of different words
that we would evaluate

00:02:05.367 --> 00:02:09.199
to see whether they're what
you actually meant to type.

00:02:09.199 --> 00:02:12.734
We've learned from the web

00:02:12.734 --> 00:02:14.934
the kind of misspellings
that people make

00:02:14.934 --> 00:02:17.200
and which things
are more and less likely.

00:02:17.200 --> 00:02:18.834
So we know that,
for instance,

00:02:18.834 --> 00:02:20.634
maybe slipping
and inserting an "A"

00:02:20.634 --> 00:02:22.300
is relatively likely,

00:02:22.300 --> 00:02:25.434
but misspelling
the very first letter

00:02:25.434 --> 00:02:28.467
might be less likely
in this case.

00:02:28.467 --> 00:02:32.667
So we've got some suggestions,
and the next thing that we do

00:02:32.667 --> 00:02:34.667
is evaluate these suggestions
in context.

00:02:34.667 --> 00:02:38.067
So there are other systems
at Google that already use

00:02:38.067 --> 00:02:40.434
the same kind of statistical
language models as this,

00:02:40.434 --> 00:02:42.467
such as the Google
translation system,

00:02:42.467 --> 00:02:44.367
that essentially
encode information

00:02:44.367 --> 00:02:46.133
about how language is used.

00:02:46.133 --> 00:02:47.968
These are learned from the web

00:02:47.968 --> 00:02:49.734
from looking at billions
of web pages,

00:02:49.734 --> 00:02:51.167
so we get a really good idea

00:02:51.167 --> 00:02:53.734
about the way that people
really use language in practice.

00:02:53.734 --> 00:02:55.434
So what we would do

00:02:55.434 --> 00:02:58.834
is look at the likelihood
of "Let's met tomorrow"

00:02:58.834 --> 00:03:02.067
and "Let's meat tomorrow,"
less likely,

00:03:02.067 --> 00:03:03.767
and "Let's meet tomorrow,"

00:03:03.767 --> 00:03:06.267
which is gonna be more likely
than either of these.

00:03:06.267 --> 00:03:08.067
And we combine that
with our error model

00:03:08.067 --> 00:03:10.200
which tells us how likely
the misspellings are,

00:03:10.200 --> 00:03:13.834
you know, without any context,
to get a final determination

00:03:13.834 --> 00:03:15.901
as to what are
the most likely words--

00:03:15.901 --> 00:03:19.200
most likely word
that you meant right here.

00:03:19.200 --> 00:03:22.000
So in this case,
we would suggest meet.

00:03:22.000 --> 00:03:25.467
Once we think
that a word is misspelled,

00:03:25.467 --> 00:03:29.100
we need to get that back
to the Google Wave client

00:03:29.100 --> 00:03:31.667
so that the user
can actually see it

00:03:31.667 --> 00:03:34.701
and either correct it
automatically or manually.

00:03:34.701 --> 00:03:36.067
Two kinds of ways

00:03:36.067 --> 00:03:38.834
that this differs
from existing spelling systems.

00:03:38.834 --> 00:03:42.000
One of them is just that
it's hosted.

00:03:42.000 --> 00:03:44.100
And this means that we can do

00:03:44.100 --> 00:03:46.033
this same kind of spelling
for you,

00:03:46.033 --> 00:03:49.434
regardless of which device
you're connecting from.

00:03:49.434 --> 00:03:52.767
So whether you're on your laptop
or your mobile or your desktop,

00:03:52.767 --> 00:03:56.000
we can give the same
quality spelling, regardless.

00:03:56.000 --> 00:03:57.901
And that applies
across languages too,

00:03:57.901 --> 00:03:59.467
so, you know, we're doing this

00:03:59.467 --> 00:04:01.267
for other alphabetic
languages also.

00:04:01.267 --> 00:04:06.667
So like I said, we use large
statistical language models.

00:04:06.667 --> 00:04:07.801
When I said large, you know,

00:04:07.801 --> 00:04:09.801
we train them
from billions of words.

00:04:09.801 --> 00:04:12.033
They end up being
many, many gigabytes.

00:04:12.033 --> 00:04:15.501
It's pretty infeasible to run
these on a single machine,

00:04:15.501 --> 00:04:17.601
which isn't such a problem
in a data center

00:04:17.601 --> 00:04:19.233
where you can have
a set of machines

00:04:19.233 --> 00:04:22.200
running a language model
and a spelling model together.

00:04:22.200 --> 00:04:27.167
And then we can share
that spelling model

00:04:27.167 --> 00:04:28.734
between many users

00:04:28.734 --> 00:04:30.667
so that the cost per user
is very low.

00:04:30.667 --> 00:04:33.534
So it's very efficient
for us to do this.

00:04:33.534 --> 00:04:35.667
Once you realize
that you've got a system

00:04:35.667 --> 00:04:37.801
that supports
collaborative editing,

00:04:37.801 --> 00:04:39.501
that has structured data,

00:04:39.501 --> 00:04:42.667
and that you can change
the user interface

00:04:42.667 --> 00:04:45.167
by having remote participants,

00:04:45.167 --> 00:04:47.200
then, really,
the sky's the limit.

00:04:47.200 --> 00:04:49.434
I mean, there's all kinds
of existing

00:04:49.434 --> 00:04:51.400
natural language technologies
like spell checking

00:04:51.400 --> 00:04:53.100
or translation
that we can apply,

00:04:53.100 --> 00:04:55.667
and we're seeing
a lot of new applications

00:04:55.667 --> 00:04:57.667
as the way that we communicate
changes as well.

00:04:57.667 --> 00:05:00.868
So, you know, really,
it's gonna be exciting times.