<s>
Co-training	B-Algorithm
is	O
a	O
machine	O
learning	O
algorithm	O
used	O
when	O
there	O
are	O
only	O
small	O
amounts	O
of	O
labeled	O
data	O
and	O
large	O
amounts	O
of	O
unlabeled	O
data	O
.	O
</s>
<s>
One	O
of	O
its	O
uses	O
is	O
in	O
text	B-Algorithm
mining	I-Algorithm
for	O
search	B-Application
engines	I-Application
.	O
</s>
<s>
Co-training	B-Algorithm
is	O
a	O
semi-supervised	B-General_Concept
learning	I-General_Concept
technique	O
that	O
requires	O
two	O
views	O
of	O
the	O
data	O
.	O
</s>
<s>
Co-training	B-Algorithm
first	O
learns	O
a	O
separate	O
classifier	O
for	O
each	O
view	O
using	O
any	O
labeled	O
examples	O
.	O
</s>
<s>
The	O
original	O
co-training	B-Algorithm
paper	O
described	O
experiments	O
using	O
co-training	B-Algorithm
to	O
classify	O
web	O
pages	O
into	O
"	O
academic	O
course	O
home	O
page	O
"	O
or	O
not	O
;	O
the	O
classifier	O
correctly	O
categorized	O
95%	O
of	O
788	O
web	O
pages	O
with	O
only	O
12	O
labeled	O
web	O
pages	O
as	O
examples	O
.	O
</s>
<s>
The	O
paper	O
has	O
been	O
cited	O
over	O
1000	O
times	O
,	O
and	O
received	O
the	O
10	O
years	O
Best	O
Paper	O
Award	O
at	O
the	O
25th	O
International	B-General_Concept
Conference	I-General_Concept
on	I-General_Concept
Machine	I-General_Concept
Learning	I-General_Concept
(	O
ICML	B-General_Concept
2008	O
)	O
,	O
a	O
renowned	O
computer	B-General_Concept
science	I-General_Concept
conference	O
.	O
</s>
<s>
Krogel	O
and	O
Scheffer	O
showed	O
in	O
2004	O
that	O
co-training	B-Algorithm
is	O
only	O
beneficial	O
if	O
the	O
data	O
sets	O
are	O
independent	O
;	O
that	O
is	O
,	O
if	O
one	O
of	O
the	O
classifiers	O
correctly	O
labels	O
a	O
data	O
point	O
that	O
the	O
other	O
classifier	O
previously	O
misclassified	O
.	O
</s>
<s>
Co-training	B-Algorithm
has	O
been	O
used	O
to	O
classify	O
web	O
pages	O
using	O
the	O
text	O
on	O
the	O
page	O
as	O
one	O
view	O
and	O
the	O
anchor	O
text	O
of	O
hyperlinks	O
on	O
other	O
pages	O
that	O
point	O
to	O
the	O
page	O
as	O
the	O
other	O
view	O
.	O
</s>
<s>
Co-training	B-Algorithm
can	O
work	O
on	O
"	O
unlabeled	O
"	O
text	O
that	O
has	O
not	O
already	O
been	O
classified	O
or	O
tagged	B-General_Concept
,	O
which	O
is	O
typical	O
for	O
the	O
text	O
appearing	O
on	O
web	O
pages	O
and	O
in	O
emails	O
.	O
</s>
<s>
The	O
co-training	B-Algorithm
models	O
utilize	O
both	O
classifiers	O
to	O
determine	O
the	O
likelihood	O
that	O
a	O
page	O
will	O
contain	O
data	O
relevant	O
to	O
the	O
search	O
criteria.	O
"	O
</s>
<s>
Text	O
on	O
websites	O
can	O
judge	O
the	O
relevance	O
of	O
link	O
classifiers	O
,	O
hence	O
the	O
term	O
"	O
co-training	B-Algorithm
"	O
.	O
</s>
<s>
Mitchell	O
claims	O
that	O
other	O
search	O
algorithms	O
are	O
86%	O
accurate	O
,	O
whereas	O
co-training	B-Algorithm
is	O
96%	O
accurate	O
.	O
</s>
<s>
Co-training	B-Algorithm
was	O
used	O
on	O
FlipDog.com,	O
a	O
job	O
search	O
site	O
,	O
and	O
by	O
the	O
U.S.	O
Department	O
of	O
Labor	O
,	O
for	O
a	O
directory	O
of	O
continuing	O
and	O
distance	O
education	O
.	O
</s>
<s>
It	O
has	O
been	O
used	O
in	O
many	O
other	O
applications	O
,	O
including	O
statistical	B-General_Concept
parsing	I-General_Concept
and	O
visual	O
detection	O
.	O
</s>
