<s>
Bitext	B-General_Concept
word	I-General_Concept
alignment	I-General_Concept
or	O
simply	O
word	B-General_Concept
alignment	I-General_Concept
is	O
the	B-Language
natural	I-Language
language	I-Language
processing	I-Language
task	O
of	O
identifying	O
translation	O
relationships	O
among	O
the	O
words	O
(	O
or	O
more	O
rarely	O
multiword	O
units	O
)	O
in	O
a	O
bitext	O
,	O
resulting	O
in	O
a	O
bipartite	O
graph	O
between	O
the	O
two	O
sides	O
of	O
the	O
bitext	O
,	O
with	O
an	O
arc	O
between	O
two	O
words	O
if	O
and	O
only	O
if	O
they	O
are	O
translations	O
of	O
one	O
another	O
.	O
</s>
<s>
Word	B-General_Concept
alignment	I-General_Concept
is	O
typically	O
done	O
after	O
sentence	O
alignment	O
has	O
already	O
identified	O
pairs	O
of	O
sentences	O
that	O
are	O
translations	O
of	O
one	O
another	O
.	O
</s>
<s>
Bitext	B-General_Concept
word	I-General_Concept
alignment	I-General_Concept
is	O
an	O
important	O
supporting	O
task	O
for	O
most	O
methods	O
of	O
statistical	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
.	O
</s>
<s>
The	O
parameters	O
of	O
statistical	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
models	O
are	O
typically	O
estimated	O
by	O
observing	O
word-aligned	O
bitexts	O
,	O
and	O
conversely	O
automatic	O
word	B-General_Concept
alignment	I-General_Concept
is	O
typically	O
done	O
by	O
choosing	O
that	O
alignment	O
which	O
best	O
fits	O
a	O
statistical	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
model	O
.	O
</s>
<s>
Circular	O
application	O
of	O
these	O
two	O
ideas	O
results	O
in	O
an	O
instance	O
of	O
the	O
expectation-maximization	B-Algorithm
algorithm	I-Algorithm
.	O
</s>
<s>
This	O
approach	O
to	O
training	O
is	O
an	O
instance	O
of	O
unsupervised	B-General_Concept
learning	I-General_Concept
,	O
in	O
that	O
the	O
system	O
is	O
not	O
given	O
examples	O
of	O
the	O
kind	O
of	O
output	O
desired	O
,	O
but	O
is	O
trying	O
to	O
find	O
values	O
for	O
the	O
unobserved	O
model	O
and	O
alignments	O
which	O
best	O
explain	O
the	O
observed	O
bitext	O
.	O
</s>
<s>
Besides	O
the	O
training	O
of	O
machine	O
translation	O
systems	O
,	O
other	O
applications	O
of	O
word	B-General_Concept
alignment	I-General_Concept
include	O
translation	O
lexicon	O
induction	O
,	O
word	O
sense	O
discovery	O
,	O
word	B-General_Concept
sense	I-General_Concept
disambiguation	I-General_Concept
and	O
the	O
cross-lingual	O
projection	O
of	O
linguistic	O
information	O
.	O
</s>
<s>
The	O
IBM	O
models	O
are	O
used	O
in	O
Statistical	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
to	O
train	O
a	O
translation	O
model	O
and	O
an	O
alignment	O
model	O
.	O
</s>
<s>
They	O
are	O
an	O
instance	O
of	O
the	O
Expectation	B-Algorithm
–	I-Algorithm
maximization	I-Algorithm
algorithm	I-Algorithm
:	O
in	O
the	O
expectation-step	O
the	O
translation	O
probabilities	O
within	O
each	O
sentence	O
are	O
computed	O
,	O
in	O
the	O
maximization	O
step	O
they	O
are	O
accumulated	O
to	O
global	O
translation	O
probabilities	O
.	O
</s>
<s>
In	O
training	O
the	O
translation	O
and	O
alignment	O
probabilities	O
can	O
be	O
obtained	O
from	O
and	O
in	O
the	O
Forward-backward	B-Algorithm
algorithm	I-Algorithm
.	O
</s>
