<s>
Statistical	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
(	O
SMT	O
)	O
is	O
a	O
machine	B-Application
translation	I-Application
paradigm	O
where	O
translations	O
are	O
generated	O
on	O
the	O
basis	O
of	O
statistical	O
models	O
whose	O
parameters	O
are	O
derived	O
from	O
the	O
analysis	O
of	O
bilingual	O
text	O
corpora	O
.	O
</s>
<s>
The	O
statistical	O
approach	O
contrasts	O
with	O
the	O
rule-based	B-General_Concept
approaches	I-General_Concept
to	I-General_Concept
machine	I-General_Concept
translation	I-General_Concept
as	O
well	O
as	O
with	O
example-based	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
,	O
and	O
has	O
more	O
recently	O
been	O
superseded	O
by	O
neural	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
in	O
many	O
applications	O
(	O
see	O
this	O
article	O
's	O
final	O
section	O
)	O
.	O
</s>
<s>
The	O
first	O
ideas	O
of	O
statistical	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
were	O
introduced	O
by	O
Warren	O
Weaver	O
in	O
1949	O
,	O
including	O
the	O
ideas	O
of	O
applying	O
Claude	O
Shannon	O
's	O
information	O
theory	O
.	O
</s>
<s>
Statistical	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
was	O
re-introduced	O
in	O
the	O
late	O
1980s	O
and	O
early	O
1990s	O
by	O
researchers	O
at	O
IBM	O
's	O
Thomas	O
J	O
.	O
Watson	O
Research	O
Center	O
and	O
has	O
contributed	O
to	O
the	O
significant	O
resurgence	O
in	O
interest	O
in	O
machine	B-Application
translation	I-Application
in	O
recent	O
years	O
.	O
</s>
<s>
Before	O
the	O
introduction	O
of	O
neural	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
,	O
it	O
was	O
by	O
far	O
the	O
most	O
widely	O
studied	O
machine	B-Application
translation	I-Application
method	O
.	O
</s>
<s>
The	O
idea	O
behind	O
statistical	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
comes	O
from	O
information	O
theory	O
.	O
</s>
<s>
One	O
approach	O
which	O
lends	O
itself	O
well	O
to	O
computer	O
implementation	O
is	O
to	O
apply	O
Bayes	O
Theorem	O
,	O
that	O
is	O
,	O
where	O
the	O
translation	O
model	O
is	O
the	O
probability	O
that	O
the	O
source	O
string	O
is	O
the	O
translation	O
of	O
the	O
target	O
string	O
,	O
and	O
the	O
language	B-Language
model	I-Language
is	O
the	O
probability	O
of	O
seeing	O
that	O
target	O
language	O
string	O
.	O
</s>
<s>
Performing	O
the	O
search	O
efficiently	O
is	O
the	O
work	O
of	O
a	O
machine	B-Application
translation	I-Application
decoder	O
that	O
uses	O
the	O
foreign	O
string	O
,	O
heuristics	O
and	O
other	O
methods	O
to	O
limit	O
the	O
search	O
space	O
and	O
at	O
the	O
same	O
time	O
keeping	O
acceptable	O
quality	O
.	O
</s>
<s>
This	O
trade-off	O
between	O
quality	O
and	O
time	O
usage	O
can	O
also	O
be	O
found	O
in	O
speech	B-Application
recognition	I-Application
.	O
</s>
<s>
As	O
the	O
translation	B-Application
systems	I-Application
are	O
not	O
able	O
to	O
store	O
all	O
native	O
strings	O
and	O
their	O
translations	O
,	O
a	O
document	O
is	O
typically	O
translated	O
sentence	O
by	O
sentence	O
,	O
but	O
even	O
this	O
is	O
not	O
enough	O
.	O
</s>
<s>
Language	B-Language
models	I-Language
are	O
typically	O
approximated	O
by	O
smoothed	O
n-gram	O
models	O
,	O
and	O
similar	O
approaches	O
have	O
been	O
applied	O
to	O
translation	O
models	O
,	O
but	O
there	O
is	O
additional	O
complexity	O
due	O
to	O
different	O
sentence	O
lengths	O
and	O
word	O
orders	O
in	O
the	O
languages	O
.	O
</s>
<s>
The	O
statistical	B-General_Concept
translation	I-General_Concept
models	O
were	O
initially	O
word	O
based	O
(	O
Models	O
1-5	O
from	O
IBM	O
Hidden	O
Markov	O
model	O
from	O
Stephan	O
Vogel	O
and	O
Model	O
6	O
from	O
Franz-Joseph	O
Och	O
)	O
,	O
but	O
significant	O
advances	O
were	O
made	O
with	O
the	O
introduction	O
of	O
phrase	O
based	O
models	O
.	O
</s>
<s>
Later	O
work	O
incorporated	O
syntax	B-Application
or	O
quasi-syntactic	O
structures	O
.	O
</s>
<s>
The	O
most	O
frequently	O
cited	O
benefits	O
of	O
statistical	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
over	O
rule-based	O
approach	O
are	O
:	O
</s>
<s>
Rule-based	O
translation	B-Application
systems	I-Application
require	O
the	O
manual	O
development	O
of	O
linguistic	O
rules	O
,	O
which	O
can	O
be	O
costly	O
,	O
and	O
which	O
often	O
do	O
not	O
generalize	O
to	O
other	O
languages	O
.	O
</s>
<s>
Statistical	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
usually	O
works	O
less	O
well	O
for	O
language	O
pairs	O
with	O
significantly	O
different	O
word	O
order	O
.	O
</s>
<s>
Word-based	O
translation	B-Application
systems	I-Application
can	O
relatively	O
simply	O
be	O
made	O
to	O
cope	O
with	O
high	O
fertility	O
,	O
such	O
that	O
they	O
could	O
map	O
a	O
single	O
word	O
to	O
multiple	O
words	O
,	O
but	O
not	O
the	O
other	O
way	O
about	O
.	O
</s>
<s>
An	O
example	O
of	O
a	O
word-based	O
translation	B-Application
system	I-Application
is	O
the	O
freely	O
available	O
GIZA++	O
package	O
(	O
GPLed	O
)	O
,	O
which	O
includes	O
the	O
training	O
program	O
for	O
IBM	O
models	O
and	O
HMM	O
model	O
and	O
Model	O
6	O
.	O
</s>
<s>
The	O
alignments	O
are	O
used	O
to	O
extract	O
phrases	O
or	O
deduce	O
syntax	B-Application
rules	O
.	O
</s>
<s>
The	O
second	O
model	O
is	O
trained	O
using	O
the	O
expectation	B-Algorithm
maximization	I-Algorithm
algorithm	I-Algorithm
,	O
similarly	O
to	O
the	O
word-based	O
IBM	B-General_Concept
model	I-General_Concept
.	O
</s>
<s>
Syntax-based	O
translation	O
is	O
based	O
on	O
the	O
idea	O
of	O
translating	O
syntactic	O
units	O
,	O
rather	O
than	O
single	O
words	O
or	O
strings	O
of	O
words	O
(	O
as	O
in	O
phrase-based	O
MT	O
)	O
,	O
i.e.	O
</s>
<s>
The	O
idea	O
of	O
syntax-based	O
translation	O
is	O
quite	O
old	O
in	O
MT	O
,	O
though	O
its	O
statistical	O
counterpart	O
did	O
not	O
take	O
off	O
until	O
the	O
advent	O
of	O
strong	O
stochastic	B-General_Concept
parsers	I-General_Concept
in	O
the	O
1990s	O
.	O
</s>
<s>
Examples	O
of	O
this	O
approach	O
include	O
DOP-based	O
MT	O
and	O
,	O
more	O
recently	O
,	O
synchronous	B-General_Concept
context-free	I-General_Concept
grammars	I-General_Concept
.	O
</s>
<s>
Hierarchical	O
phrase-based	O
translation	O
combines	O
the	O
strengths	O
of	O
phrase-based	O
and	O
syntax-based	O
translation	O
.	O
</s>
<s>
It	O
uses	O
synchronous	B-General_Concept
context-free	I-General_Concept
grammar	I-General_Concept
rules	O
,	O
but	O
the	O
grammars	O
may	O
be	O
constructed	O
by	O
an	O
extension	O
of	O
methods	O
for	O
phrase-based	O
translation	O
without	O
reference	O
to	O
linguistically	O
motivated	O
syntactic	O
constituents	O
.	O
</s>
<s>
A	O
language	B-Language
model	I-Language
is	O
an	O
essential	O
component	O
of	O
any	O
statistical	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
system	O
,	O
which	O
aids	O
in	O
making	O
the	O
translation	O
as	O
fluent	O
as	O
possible	O
.	O
</s>
<s>
A	O
good	O
language	B-Language
model	I-Language
will	O
for	O
example	O
assign	O
a	O
higher	O
probability	O
to	O
the	O
sentence	O
"	O
the	O
house	O
is	O
small	O
"	O
than	O
to	O
"	O
small	O
the	O
is	O
house	O
"	O
.	O
</s>
<s>
Other	O
than	O
word	O
order	O
,	O
language	B-Language
models	I-Language
may	O
also	O
help	O
with	O
word	O
choice	O
:	O
if	O
a	O
foreign	O
word	O
has	O
multiple	O
possible	O
translations	O
,	O
these	O
functions	O
may	O
give	O
better	O
probabilities	O
for	O
certain	O
translations	O
in	O
specific	O
contexts	O
in	O
the	O
target	O
language	O
.	O
</s>
<s>
Problems	O
that	O
statistical	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
have	O
to	O
deal	O
with	O
include	O
:	O
</s>
<s>
Sentence	O
aligning	O
can	O
be	O
performed	O
through	O
the	O
Gale-Church	B-General_Concept
alignment	I-General_Concept
algorithm	I-General_Concept
.	O
</s>
<s>
Sentence	O
alignment	O
is	O
usually	O
either	O
provided	O
by	O
the	O
corpus	O
or	O
obtained	O
by	O
aforementioned	O
Gale-Church	B-General_Concept
alignment	I-General_Concept
algorithm	I-General_Concept
.	O
</s>
<s>
Solutions	O
are	O
the	O
IBM-Models	B-General_Concept
or	O
the	O
HMM-approach	O
.	O
</s>
<s>
In	O
speech	B-Application
recognition	I-Application
,	O
the	O
speech	O
signal	O
and	O
the	O
corresponding	O
textual	O
representation	O
can	O
be	O
mapped	O
to	O
each	O
other	O
in	O
blocks	O
in	O
order	O
.	O
</s>
<s>
For	O
SMT	O
,	O
the	O
machine	B-Application
translator	I-Application
can	O
only	O
manage	O
small	O
sequences	O
of	O
words	O
,	O
and	O
word	O
order	O
has	O
to	O
be	O
thought	O
of	O
by	O
the	O
program	O
designer	O
.	O
</s>
<s>
Different	O
location	O
changes	O
can	O
be	O
ranked	O
with	O
the	O
help	O
of	O
the	O
language	B-Language
model	I-Language
and	O
the	O
best	O
can	O
be	O
selected	O
.	O
</s>
<s>
Recently	O
,	O
Skype	B-Application
voice	O
communicator	O
started	O
testing	O
speech	O
translation	O
.	O
</s>
<s>
However	O
,	O
machine	B-Application
translation	I-Application
is	O
following	O
technological	O
trends	O
in	O
speech	O
at	O
a	O
slower	O
rate	O
than	O
speech	B-Application
recognition	I-Application
.	O
</s>
<s>
In	O
fact	O
,	O
some	O
ideas	O
from	O
speech	B-Application
recognition	I-Application
research	O
have	O
been	O
adopted	O
by	O
statistical	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
.	O
</s>
<s>
The	O
rapid	O
increase	O
in	O
the	O
computing	O
power	O
of	O
tablets	B-Device
and	O
smartphones	B-Application
,	O
combined	O
with	O
the	O
wide	O
availability	O
of	O
high-speed	O
mobile	O
Internet	O
access	O
,	O
makes	O
it	O
possible	O
for	O
them	O
to	O
run	O
machine	B-Application
translation	I-Application
systems	I-Application
.	O
</s>
<s>
For	O
example	O
,	O
Apple	O
’s	O
iOS	B-Operating_System
8	I-Operating_System
allows	O
users	O
to	O
dictate	O
text	O
messages	O
.	O
</s>
<s>
A	O
built-in	O
ASR	B-Application
system	O
recognizes	O
the	O
speech	O
and	O
the	O
recognition	O
results	O
are	O
edited	O
by	O
an	O
online	O
system	O
.	O
</s>
<s>
Recent	O
years	O
have	O
seen	O
a	O
growing	O
interest	O
in	O
combining	O
speech	B-Application
recognition	I-Application
,	O
machine	B-Application
translation	I-Application
and	O
speech	B-Application
synthesis	I-Application
.	O
</s>
<s>
To	O
achieve	O
speech-to-speech	O
translation	O
,	O
n-best	O
lists	O
are	O
passed	O
from	O
the	O
ASR	B-Application
to	O
the	O
statistical	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
system	O
.	O
</s>
