<s>
A	O
constrained	B-General_Concept
conditional	I-General_Concept
model	I-General_Concept
(	O
CCM	O
)	O
is	O
a	O
machine	O
learning	O
and	O
inference	O
framework	O
that	O
augments	O
the	O
learning	O
of	O
conditional	O
(	O
probabilistic	O
or	O
discriminative	O
)	O
models	O
with	O
declarative	O
constraints	O
.	O
</s>
<s>
Models	O
of	O
this	O
kind	O
have	O
recently	O
attracted	O
much	O
attention	O
within	O
the	B-Language
natural	I-Language
language	I-Language
processing	I-Language
(	O
NLP	B-Language
)	O
community	O
.	O
</s>
<s>
Formulating	O
problems	O
as	O
constrained	B-Application
optimization	I-Application
problems	I-Application
over	O
the	O
output	O
of	O
learned	O
models	O
has	O
several	O
advantages	O
.	O
</s>
<s>
Using	O
this	O
declarative	O
framework	O
frees	O
the	O
developer	O
from	O
low	O
level	O
feature	B-General_Concept
engineering	I-General_Concept
while	O
capturing	O
the	O
problem	O
's	O
domain-specific	O
properties	O
and	O
guarantying	O
exact	O
inference	O
.	O
</s>
<s>
Making	O
decisions	O
in	O
many	O
domains	O
(	O
such	O
as	O
natural	B-Language
language	I-Language
processing	I-Language
and	O
computer	O
vision	O
problems	O
)	O
often	O
involves	O
assigning	O
values	O
to	O
sets	O
of	O
interdependent	O
variables	O
where	O
the	O
expressive	O
dependency	O
structure	O
can	O
influence	O
,	O
or	O
even	O
dictate	O
,	O
what	O
assignments	O
are	O
possible	O
.	O
</s>
<s>
These	O
settings	O
are	O
applicable	O
not	O
only	O
to	O
Structured	B-General_Concept
Learning	I-General_Concept
problems	O
such	O
as	O
semantic	O
role	O
labeling	O
,	O
but	O
also	O
for	O
cases	O
that	O
require	O
making	O
use	O
of	O
multiple	O
pre-learned	O
components	O
,	O
such	O
as	O
summarization	O
,	O
textual	O
entailment	O
and	O
question	O
answering	O
.	O
</s>
<s>
In	O
all	O
these	O
cases	O
,	O
it	O
is	O
natural	O
to	O
formulate	O
the	O
decision	O
problem	O
as	O
a	O
constrained	B-Application
optimization	I-Application
problem	I-Application
,	O
with	O
an	O
objective	O
function	O
that	O
is	O
composed	O
of	O
learned	O
models	O
,	O
subject	O
to	O
domain	O
-	O
or	O
problem-specific	O
constraints	O
.	O
</s>
<s>
Constrained	B-General_Concept
conditional	I-General_Concept
models	I-General_Concept
form	O
a	O
learning	O
and	O
inference	O
framework	O
that	O
augments	O
the	O
learning	O
of	O
conditional	O
(	O
probabilistic	O
or	O
discriminative	O
)	O
models	O
with	O
declarative	O
constraints	O
(	O
written	O
,	O
for	O
example	O
,	O
using	O
a	O
first-order	O
representation	O
)	O
as	O
a	O
way	O
to	O
support	O
decisions	O
in	O
an	O
expressive	O
output	O
space	O
while	O
maintaining	O
modularity	O
and	O
tractability	O
of	O
training	O
and	O
inference	O
.	O
</s>
<s>
In	O
most	O
applications	O
of	O
this	O
framework	O
in	O
NLP	B-Language
,	O
following	O
,	O
Integer	O
Linear	O
Programming	O
(	O
ILP	O
)	O
was	O
used	O
as	O
the	O
inference	O
framework	O
,	O
although	O
other	O
algorithms	O
can	O
be	O
used	O
for	O
that	O
purpose	O
.	O
</s>
<s>
Constraints	O
assigned	O
an	O
infinite	O
penalty	O
are	O
known	O
as	O
hard	B-Application
constraints	I-Application
,	O
and	O
represent	O
unfeasible	O
assignments	O
to	O
the	O
optimization	O
problem	O
.	O
</s>
<s>
Identifying	O
the	O
correct	O
(	O
or	O
optimal	O
)	O
learning	O
representation	O
is	O
viewed	O
as	O
a	O
structured	B-General_Concept
prediction	I-General_Concept
process	O
and	O
therefore	O
modeled	O
as	O
a	O
CCM	O
.	O
</s>
<s>
The	O
advantages	O
of	O
the	O
CCM	O
declarative	O
formulation	O
and	O
the	O
availability	O
of	O
off-the-shelf	O
solvers	O
have	O
led	O
to	O
a	O
large	O
variety	O
of	O
natural	B-Language
language	I-Language
processing	I-Language
tasks	O
being	O
formulated	O
within	O
the	O
framework	O
,	O
including	O
semantic	O
role	O
labeling	O
,	O
syntactic	O
parsing	O
,	O
coreference	O
resolution	O
,	O
summarization	O
,	O
transliteration	O
,	O
natural	B-General_Concept
language	I-General_Concept
generation	I-General_Concept
and	O
joint	O
information	O
extraction	O
.	O
</s>
<s>
Although	O
theoretically	O
solving	O
an	O
Integer	O
Linear	O
Program	O
is	O
exponential	O
in	O
the	O
size	O
of	O
the	O
decision	O
problem	O
,	O
in	O
practice	O
using	O
state-of-the-art	O
solvers	O
and	O
approximate	B-General_Concept
inference	I-General_Concept
techniques	O
large	O
scale	O
problems	O
can	O
be	O
solved	O
efficiently	O
.	O
</s>
<s>
The	O
key	O
advantage	O
of	O
using	O
an	O
ILP	O
solver	O
for	O
solving	O
the	O
optimization	O
problem	O
defined	O
by	O
a	O
constrained	B-General_Concept
conditional	I-General_Concept
model	I-General_Concept
is	O
the	O
declarative	O
formulation	O
used	O
as	O
input	O
for	O
the	O
ILP	O
solver	O
,	O
consisting	O
of	O
a	O
linear	O
objective	O
function	O
and	O
a	O
set	O
of	O
linear	O
constraints	O
.	O
</s>
