<s>
Brown	B-General_Concept
clustering	I-General_Concept
is	O
a	O
hard	O
hierarchical	O
agglomerative	B-Algorithm
clustering	I-Algorithm
problem	O
based	O
on	O
distributional	O
information	O
proposed	O
by	O
Peter	O
Brown	O
,	O
William	O
A	O
.	O
</s>
<s>
The	O
method	O
,	O
which	O
is	O
based	O
on	O
bigram	O
language	B-Language
models	I-Language
,	O
is	O
typically	O
applied	O
to	O
text	O
,	O
grouping	O
words	O
into	O
clusters	O
that	O
are	O
assumed	O
to	O
be	O
semantically	O
related	O
by	O
virtue	O
of	O
their	O
having	O
been	O
embedded	O
in	O
similar	O
contexts	O
.	O
</s>
<s>
In	O
natural	B-Language
language	I-Language
processing	I-Language
,	O
Brown	B-General_Concept
clustering	I-General_Concept
or	O
IBM	O
clustering	O
is	O
a	O
form	O
of	O
hierarchical	B-Algorithm
clustering	I-Algorithm
of	O
words	O
based	O
on	O
the	O
contexts	O
in	O
which	O
they	O
occur	O
,	O
proposed	O
by	O
Peter	O
Brown	O
,	O
William	O
A	O
.	O
</s>
<s>
Brown	O
,	O
Vincent	O
Della	O
Pietra	O
,	O
Peter	O
de	O
Souza	O
,	O
Jennifer	O
Lai	O
,	O
and	O
Robert	O
Mercer	O
of	O
IBM	O
in	O
the	O
context	O
of	O
language	B-Language
modeling	I-Language
.	O
</s>
<s>
The	O
intuition	O
behind	O
the	O
method	O
is	O
that	O
a	O
class-based	B-General_Concept
language	I-General_Concept
model	I-General_Concept
(	O
also	O
called	O
cluster	O
-gram	O
model	O
)	O
,	O
i.e.	O
</s>
<s>
one	O
where	O
probabilities	O
of	O
words	O
are	O
based	O
on	O
the	O
classes	O
(	O
clusters	O
)	O
of	O
previous	O
words	O
,	O
is	O
used	O
to	O
address	O
the	O
data	O
sparsity	O
problem	O
inherent	O
in	O
language	B-Language
modeling	I-Language
.	O
</s>
<s>
Brown	O
groups	O
items	O
(	O
i.e.	O
,	O
types	O
)	O
into	O
classes	O
,	O
using	O
a	O
binary	O
merging	O
criterion	O
based	O
on	O
the	O
log-probability	O
of	O
a	O
text	O
under	O
a	O
class-based	B-General_Concept
language	I-General_Concept
model	I-General_Concept
,	O
i.e.	O
</s>
<s>
is	O
a	O
greedy	B-Algorithm
heuristic	I-Algorithm
.	O
</s>
<s>
The	O
work	O
also	O
suggests	O
use	O
of	O
Brown	B-General_Concept
clusterings	I-General_Concept
as	O
a	O
simplistic	O
bigram	O
class-based	B-General_Concept
language	I-General_Concept
model	I-General_Concept
.	O
</s>
<s>
When	O
applied	O
to	O
Twitter	B-Application
data	O
,	O
for	O
example	O
,	O
Brown	B-General_Concept
clustering	I-General_Concept
assigned	O
a	O
binary	O
tree	O
path	O
to	O
each	O
word	O
in	O
unlabelled	O
tweets	O
during	O
clustering	O
.	O
</s>
<s>
Brown	B-General_Concept
clustering	I-General_Concept
has	O
also	O
been	O
explored	O
using	O
trigrams	O
.	O
</s>
<s>
Brown	B-General_Concept
clustering	I-General_Concept
as	O
proposed	O
generates	O
a	O
fixed	O
number	O
of	O
output	O
classes	O
.	O
</s>
<s>
The	O
cluster	O
memberships	O
of	O
words	O
resulting	O
from	O
Brown	B-General_Concept
clustering	I-General_Concept
can	O
be	O
used	O
as	O
features	O
in	O
a	O
variety	O
of	O
machine-learned	O
natural	B-Language
language	I-Language
processing	I-Language
tasks	O
.	O
</s>
<s>
There	O
are	O
no	O
known	O
theoretical	O
guarantees	O
on	O
the	O
greedy	B-Algorithm
heuristic	I-Algorithm
proposed	O
by	O
Brown	O
et	O
al	O
.	O
</s>
<s>
However	O
,	O
the	O
clustering	O
problem	O
can	O
be	O
framed	O
as	O
estimating	O
the	O
parameters	O
of	O
the	O
underlying	O
class-based	B-General_Concept
language	I-General_Concept
model	I-General_Concept
:	O
it	O
is	O
possible	O
to	O
develop	O
a	O
consistent	O
estimator	O
for	O
this	O
model	O
under	O
mild	O
assumptions	O
.	O
</s>
