<s>
Google	B-Algorithm
Neural	I-Algorithm
Machine	I-Algorithm
Translation	I-Algorithm
(	O
GNMT	B-Algorithm
)	O
is	O
a	O
neural	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
(	O
NMT	O
)	O
system	O
developed	O
by	O
Google	O
and	O
introduced	O
in	O
November	O
2016	O
,	O
that	O
uses	O
an	O
artificial	B-Architecture
neural	I-Architecture
network	I-Architecture
to	O
increase	O
fluency	O
and	O
accuracy	O
in	O
Google	B-Application
Translate	I-Application
.	O
</s>
<s>
GNMT	B-Algorithm
improves	O
on	O
the	O
quality	O
of	O
translation	O
by	O
applying	O
an	O
example-based	B-General_Concept
(	O
EBMT	B-General_Concept
)	O
machine	B-Application
translation	I-Application
method	O
in	O
which	O
the	O
system	O
"	O
learns	O
from	O
millions	O
of	O
examples	O
"	O
.	O
</s>
<s>
GNMT	B-Algorithm
's	O
proposed	O
architecture	O
of	O
system	O
learning	O
was	O
first	O
tested	O
on	O
over	O
a	O
hundred	O
languages	O
supported	O
by	O
Google	B-Application
Translate	I-Application
.	O
</s>
<s>
GNMT	B-Algorithm
attempts	O
to	O
translate	O
whole	O
sentences	O
at	O
a	O
time	O
,	O
rather	O
than	O
just	O
piece	O
by	O
piece	O
.	O
</s>
<s>
The	O
GNMT	B-Algorithm
network	O
can	O
undertake	O
interlingual	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
by	O
encoding	O
the	O
semantics	O
of	O
the	O
sentence	O
,	O
rather	O
than	O
by	O
memorizing	O
phrase-to-phrase	O
translations	O
.	O
</s>
<s>
The	O
Google	B-Application
Brain	I-Application
project	O
was	O
established	O
in	O
2011	O
in	O
the	O
"	O
secretive	O
Google	O
X	O
research	O
lab	O
"	O
by	O
Google	O
Fellow	O
Jeff	O
Dean	O
,	O
Google	O
Researcher	O
Greg	O
Corrado	O
,	O
and	O
Stanford	O
University	O
Computer	O
Science	O
professor	O
Andrew	O
Ng	O
.	O
</s>
<s>
In	O
November	O
2016	O
,	O
Google	B-Algorithm
Neural	I-Algorithm
Machine	I-Algorithm
Translation	I-Algorithm
system	O
(	O
GNMT	B-Algorithm
)	O
was	O
introduced	O
.	O
</s>
<s>
Since	O
then	O
,	O
Google	B-Application
Translate	I-Application
began	O
using	O
neural	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
(	O
NMT	O
)	O
in	O
preference	O
to	O
its	O
previous	O
statistical	B-General_Concept
methods	I-General_Concept
(	O
SMT	O
)	O
which	O
had	O
been	O
used	O
since	O
October	O
2007	O
,	O
with	O
its	O
proprietary	O
,	O
in-house	O
SMT	O
technology	O
.	O
</s>
<s>
Google	B-Application
Translate	I-Application
's	O
NMT	O
system	O
uses	O
a	O
large	O
artificial	B-Architecture
neural	I-Architecture
network	I-Architecture
capable	O
of	O
deep	B-Algorithm
learning	I-Algorithm
.	O
</s>
<s>
By	O
using	O
millions	O
of	O
examples	O
,	O
GNMT	B-Algorithm
improves	O
the	O
quality	O
of	O
translation	O
,	O
using	O
broader	O
context	O
to	O
deduce	O
the	O
most	O
relevant	O
translation	O
.	O
</s>
<s>
GNMT	B-Algorithm
's	O
proposed	O
architecture	O
of	O
system	O
learning	O
was	O
first	O
tested	O
on	O
over	O
a	O
hundred	O
languages	O
supported	O
by	O
Google	B-Application
Translate	I-Application
.	O
</s>
<s>
GNMT	B-Algorithm
did	O
not	O
create	O
its	O
own	O
universal	O
interlingua	O
but	O
rather	O
aimed	O
at	O
finding	O
the	O
commonality	O
between	O
many	O
languages	O
using	O
insights	O
from	O
psychology	O
and	O
linguistics	O
.	O
</s>
<s>
The	O
new	O
translation	B-Application
engine	I-Application
was	O
first	O
enabled	O
for	O
eight	O
languages	O
:	O
to	O
and	O
from	O
English	O
and	O
French	O
,	O
German	O
,	O
Spanish	O
,	O
Portuguese	O
,	O
Chinese	O
,	O
Japanese	O
,	O
Korean	O
and	O
Turkish	O
in	O
November	O
2016	O
.	O
</s>
<s>
Support	O
for	O
Hebrew	O
and	O
Arabic	O
was	O
also	O
added	O
with	O
help	O
from	O
the	O
Google	B-Application
Translate	I-Application
Community	O
in	O
the	O
same	O
month	O
.	O
</s>
<s>
The	O
GNMT	B-Algorithm
system	O
is	O
said	O
to	O
represent	O
an	O
improvement	O
over	O
the	O
former	O
Google	B-Application
Translate	I-Application
in	O
that	O
it	O
will	O
be	O
able	O
to	O
handle	O
"	O
zero-shot	O
translation	O
"	O
,	O
that	O
is	O
it	O
directly	O
translates	O
one	O
language	O
into	O
another	O
(	O
for	O
example	O
,	O
Japanese	O
to	O
Korean	O
)	O
.	O
</s>
<s>
Google	B-Application
Translate	I-Application
previously	O
first	O
translated	O
the	O
source	O
language	O
into	O
English	O
and	O
then	O
translated	O
the	O
English	O
into	O
the	O
target	O
language	O
rather	O
than	O
translating	O
directly	O
from	O
one	O
language	O
to	O
another	O
.	O
</s>
<s>
A	O
July	O
2019	O
study	O
in	O
Annals	O
of	O
Internal	O
Medicine	O
found	O
that	O
"	O
Google	B-Application
Translate	I-Application
is	O
a	O
viable	O
,	O
accurate	O
tool	O
for	O
translating	O
non	O
–	O
English-language	O
trials	O
"	O
.	O
</s>
<s>
Since	O
many	O
medical	O
studies	O
are	O
excluded	O
from	O
systematic	O
reviews	O
because	O
the	O
reviewers	O
do	O
not	O
understand	O
the	O
language	O
,	O
GNMT	B-Algorithm
has	O
the	O
potential	O
to	O
reduce	O
bias	O
and	O
improve	O
accuracy	O
in	O
such	O
reviews	O
.	O
</s>
<s>
As	O
of	O
December	O
2021	O
,	O
all	O
of	O
the	O
languages	O
of	O
Google	B-Application
Translate	I-Application
support	O
GNMT	B-Algorithm
,	O
with	O
Latin	O
being	O
the	O
most	O
recent	O
addition	O
.	O
</s>
