<s>
The	O
WebCrow	B-Application
is	O
a	O
research	O
project	O
carried	O
out	O
at	O
the	O
Information	O
Engineering	O
Department	O
of	O
the	O
University	O
of	O
Siena	O
with	O
the	O
purpose	O
of	O
automatically	O
solving	O
crosswords	B-Application
.	O
</s>
<s>
The	O
scientific	O
relevance	O
of	O
the	O
project	O
can	O
be	O
understood	O
considering	O
that	O
cracking	O
crosswords	B-Application
requires	O
human-level	O
knowledge	O
.	O
</s>
<s>
Unlike	O
chess	O
and	O
related	O
games	O
and	O
there	O
is	O
no	O
closed	B-Application
world	I-Application
configuration	O
space	O
.	O
</s>
<s>
A	O
first	O
nucleus	O
of	O
technology	O
,	O
such	O
as	O
search	B-Application
engines	I-Application
,	O
information	O
retrieval	O
,	O
and	O
machine	O
learning	O
techniques	O
enable	O
computers	O
to	O
enfold	O
with	O
semantics	O
real-life	O
concepts	O
.	O
</s>
<s>
The	O
project	O
is	O
based	O
on	O
a	O
software	O
system	O
whose	O
major	O
assumption	O
is	O
to	O
attack	O
crosswords	B-Application
making	O
use	O
of	O
the	O
Web	O
as	O
its	O
primary	O
source	O
of	O
knowledge	O
.	O
</s>
<s>
WebCrow	B-Application
is	O
very	O
fast	O
and	O
often	O
thrashes	O
human	O
challengers	O
in	O
competitions	O
,	O
especially	O
on	O
multi	O
language	O
crossword	O
schemes	O
.	O
</s>
<s>
A	O
distinct	O
feature	O
of	O
the	O
WebCrow	B-Application
software	O
system	O
is	O
to	O
combine	O
properly	O
natural	B-Language
language	I-Language
processing	I-Language
(	O
NLP	B-Language
)	O
techniques	O
,	O
the	B-Application
Google	I-Application
web	B-Application
search	I-Application
engine	I-Application
,	O
and	O
constraint	B-Application
satisfaction	I-Application
algorithms	O
from	O
artificial	O
intelligence	O
to	O
acquire	O
knowledge	O
and	O
to	O
fill	O
the	O
schema	O
.	O
</s>
<s>
The	O
most	O
important	O
component	O
of	O
WebCrow	B-Application
is	O
the	O
Web	B-Application
Search	I-Application
Module	O
(	O
WSM	O
)	O
,	O
which	O
implements	O
a	O
domain	O
specific	O
web	O
based	O
question	B-Algorithm
answering	I-Algorithm
algorithm	O
.	O
</s>
<s>
The	O
way	O
WebCrow	B-Application
approaches	O
crosswords	B-Application
solving	O
is	O
quite	O
different	O
with	O
respect	O
to	O
humans	O
:	O
Whereas	O
we	O
tend	O
to	O
first	O
answer	O
clues	O
we	O
are	O
sure	O
of	O
and	O
then	O
proceed	O
filling	O
the	O
schema	O
by	O
exploiting	O
the	O
already	O
answered	O
clues	O
as	O
hints	O
,	O
WebCrow	B-Application
uses	O
two	O
clearly	O
distinct	O
stages	O
.	O
</s>
<s>
In	O
the	O
second	O
stage	O
,	O
WebCrow	B-Application
uses	O
constraint	B-Application
satisfaction	I-Application
algorithms	O
to	O
fill	O
the	O
grid	O
with	O
the	O
overall	O
most	O
likely	O
combination	O
of	O
clue	O
answers	O
.	O
</s>
<s>
In	O
order	O
to	O
interact	O
with	O
Google	B-Application
,	O
first	O
of	O
all	O
,	O
WebCrow	B-Application
needs	O
to	O
compose	O
queries	O
on	O
the	O
basis	O
of	O
the	O
given	O
clues	O
.	O
</s>
<s>
This	O
is	O
done	O
by	O
query	B-Algorithm
expansion	I-Algorithm
,	O
whose	O
purpose	O
is	O
to	O
convert	O
the	O
clue	O
into	O
a	O
query	O
expressed	O
by	O
a	O
simplified	O
and	O
more	O
appropriate	O
language	O
for	O
Google	B-Application
.	I-Application
</s>
<s>
Crosswords	B-Application
can	O
hardly	O
be	O
faced	O
by	O
using	O
encyclopedic	O
knowledge	O
only	O
,	O
since	O
many	O
clues	O
are	O
wordplays	O
or	O
are	O
otherwise	O
purposefully	O
very	O
ambiguous	O
.	O
</s>
<s>
This	O
enigmatic	O
component	O
of	O
crosswords	B-Application
is	O
faced	O
by	O
a	O
massive	O
use	O
of	O
database	O
of	O
solved	O
crosswords	B-Application
,	O
and	O
by	O
automatic	O
reasoning	O
on	O
a	O
properly	O
organized	O
knowledge	O
base	O
of	O
wired	O
rules	O
.	O
</s>
<s>
Last	O
but	O
not	O
the	O
least	O
,	O
the	O
final	O
constraint	B-Application
satisfaction	I-Application
step	O
is	O
very	O
effective	O
to	O
fill	O
the	O
correct	O
candidate	O
,	O
even	O
though	O
,	O
unlike	O
humans	O
,	O
the	O
system	O
can	O
not	O
rely	O
on	O
very	O
high	O
confidence	O
on	O
the	O
correctness	O
of	O
the	O
answer	O
.	O
</s>
<s>
The	O
outcome	O
of	O
the	O
tests	O
is	O
that	O
WebCrow	B-Application
can	O
successfully	O
compete	O
with	O
average	O
human	O
players	O
on	O
single	O
language	O
schemes	O
and	O
reaches	O
expert	O
level	O
performance	O
in	O
multi-language	O
crosswords	B-Application
.	O
</s>
<s>
However	O
,	O
WebCrow	B-Application
has	O
not	O
reached	O
expert	O
level	O
in	O
single-language	O
crosswords	B-Application
,	O
yet	O
.	O
</s>
<s>
On	O
August	O
30	O
,	O
2006	O
,	O
at	O
the	O
European	O
Conference	O
on	O
Artificial	O
Intelligence	O
(	O
ECAI2006	O
)	O
,	O
25	O
conference	O
attendees	O
and	O
53	O
internet	O
connected	O
crosswords	B-Application
lovers	O
,	O
competed	O
with	O
WebCrow	B-Application
in	O
an	O
official	O
challenge	O
organized	O
within	O
the	O
conference	O
program	O
.	O
</s>
<s>
The	O
challenge	O
consisted	O
in	O
5	O
different	O
crosswords	B-Application
(	O
2	O
in	O
Italian	O
,	O
2	O
in	O
English	O
and	O
one	O
multi-language	O
in	O
Italian	O
and	O
English	O
)	O
and	O
15	O
minutes	O
were	O
assigned	O
for	O
each	O
crossword	O
.	O
</s>
<s>
WebCrow	B-Application
ranked	O
21	O
out	O
of	O
74	O
participants	O
in	O
the	O
Italian	O
competition	O
,	O
and	O
won	O
both	O
the	O
bilingual	O
and	O
English	O
competitions	O
.	O
</s>
<s>
Several	O
competitions	O
have	O
been	O
held	O
in	O
Florence	B-Application
,	O
Italy	O
within	O
the	O
Creativity	O
Festival	O
in	O
December	O
2006	O
,	O
and	O
another	O
official	O
conference	O
competition	O
took	O
place	O
in	O
Hyderabad	O
,	O
India	O
in	O
January	O
2007	O
,	O
within	O
the	O
International	O
Conference	O
of	O
Artificial	O
Intelligence	O
,	O
where	O
it	O
ranked	O
second	O
out	O
of	O
25	O
participants	O
.	O
</s>
