<s>
The	O
knowledge	O
acquisition	O
bottleneck	O
is	O
perhaps	O
the	O
major	O
impediment	O
to	O
solving	O
the	O
word	B-General_Concept
sense	I-General_Concept
disambiguation	I-General_Concept
(	O
WSD	O
)	O
problem	O
.	O
</s>
<s>
Unsupervised	B-General_Concept
learning	I-General_Concept
methods	O
rely	O
on	O
knowledge	O
about	O
word	O
senses	O
,	O
which	O
is	O
barely	O
formulated	O
in	O
dictionaries	O
and	O
lexical	O
databases	O
.	O
</s>
<s>
Supervised	B-General_Concept
learning	I-General_Concept
methods	O
depend	O
heavily	O
on	O
the	O
existence	O
of	O
manually	O
annotated	O
examples	O
for	O
every	O
word	O
sense	O
,	O
a	O
requisite	O
that	O
can	O
be	O
met	O
only	O
for	O
a	O
handful	O
of	O
words	O
for	O
testing	O
purposes	O
,	O
as	O
it	O
is	O
done	O
in	O
the	O
Senseval	B-General_Concept
exercises	O
.	O
</s>
<s>
WSD	O
has	O
been	O
traditionally	O
understood	O
as	O
an	O
intermediate	O
language	O
engineering	O
technology	O
which	O
could	O
improve	O
applications	O
such	O
as	O
information	B-Library
retrieval	I-Library
(	O
IR	O
)	O
.	O
</s>
<s>
In	O
this	O
case	O
,	O
however	O
,	O
the	O
reverse	O
is	O
also	O
true	O
:	O
Web	B-Application
search	I-Application
engines	I-Application
implement	O
simple	O
and	O
robust	O
IR	O
techniques	O
that	O
can	O
be	O
successfully	O
used	O
when	O
mining	O
the	O
Web	O
for	O
information	O
to	O
be	O
employed	O
in	O
WSD	O
.	O
</s>
<s>
The	O
most	O
direct	O
way	O
of	O
using	O
the	O
Web	O
(	O
and	O
other	O
corpora	O
)	O
to	O
enhance	O
WSD	O
performance	O
is	O
the	O
automatic	B-General_Concept
acquisition	I-General_Concept
of	I-General_Concept
sense-tagged	I-General_Concept
corpora	I-General_Concept
,	O
the	O
fundamental	O
resource	O
to	O
feed	O
supervised	B-General_Concept
WSD	O
algorithms	O
.	O
</s>
<s>
acquisition	O
by	O
direct	O
Web	B-Application
searching	I-Application
(	O
searches	O
for	O
monosemous	O
synonyms	O
,	O
hypernyms	O
,	O
hyponyms	O
,	O
parsed	O
gloss	O
 '	O
words	O
,	O
etc	O
.	O
</s>
<s>
to	O
mine	O
the	O
web	O
for	O
word	B-General_Concept
sense	I-General_Concept
disambiguation	I-General_Concept
.	O
</s>
<s>
This	O
is	O
the	O
case	O
of	O
the	O
monosemous	O
relatives	O
plus	O
bootstrapping	O
with	O
Semcor	O
seeds	O
technique	O
and	O
the	O
examples	O
taken	O
from	O
the	O
ODP	O
Web	B-Application
directories	I-Application
.	O
</s>
<s>
It	O
has	O
been	O
shown	O
that	O
a	O
mainstream	O
supervised	B-General_Concept
learning	I-General_Concept
technique	O
trained	O
exclusively	O
with	O
web	O
data	O
can	O
obtain	O
better	O
results	O
than	O
all	O
unsupervised	O
WSD	O
systems	O
which	O
participated	O
at	O
Senseval-2	O
.	O
</s>
<s>
Web	O
examples	O
made	O
a	O
significant	O
contribution	O
to	O
the	O
best	O
Senseval-2	O
English	O
all-words	O
system	O
.	O
</s>
<s>
High	O
precision	O
in	O
the	O
retrieved	O
examples	O
(	O
i.e.	O
,	O
correct	O
sense	O
assignments	O
for	O
the	O
examples	O
)	O
does	O
not	O
necessarily	O
lead	O
to	O
good	O
supervised	B-General_Concept
WSD	O
results	O
(	O
i.e.	O
,	O
the	O
examples	O
are	O
possibly	O
not	O
useful	O
for	O
training	O
)	O
.	O
</s>
<s>
The	O
most	O
complete	O
evaluation	O
of	O
Web	O
examples	O
for	O
supervised	B-General_Concept
WSD	O
indicates	O
that	O
learning	O
with	O
Web	O
data	O
improves	O
over	O
unsupervised	O
techniques	O
,	O
but	O
the	O
results	O
are	O
nevertheless	O
far	O
from	O
those	O
obtained	O
with	O
hand-tagged	O
data	O
,	O
and	O
do	O
not	O
even	O
beat	O
the	O
most-frequent-sense	O
baseline	O
.	O
</s>
<s>
Results	O
with	O
Web	O
data	O
seem	O
to	O
be	O
very	O
sensitive	O
to	O
small	O
differences	O
in	O
the	O
learning	O
algorithm	O
,	O
to	O
when	O
the	O
corpus	O
was	O
extracted	O
(	O
search	B-Application
engines	I-Application
change	O
continuously	O
)	O
,	O
and	O
on	O
small	O
heuristic	O
issues	O
(	O
e.g.	O
,	O
differences	O
in	O
filters	O
to	O
discard	O
part	O
of	O
the	O
retrieved	O
examples	O
)	O
.	O
</s>
<s>
It	O
is	O
unclear	O
whether	O
this	O
is	O
simply	O
a	O
problem	O
of	O
Web	O
data	O
,	O
or	O
an	O
intrinsic	O
problem	O
of	O
supervised	B-General_Concept
learning	I-General_Concept
techniques	O
,	O
or	O
just	O
a	O
problem	O
of	O
how	O
WSD	O
systems	O
are	O
evaluated	O
(	O
indeed	O
,	O
testing	O
with	O
rather	O
small	O
Senseval	B-General_Concept
data	O
may	O
overemphasize	O
sense	O
distributions	O
compared	O
to	O
sense	O
distributions	O
obtained	O
from	O
the	O
full	O
Web	O
as	O
corpus	O
)	O
.	O
</s>
<s>
In	O
any	O
case	O
,	O
Web	O
data	O
has	O
an	O
intrinsic	O
bias	O
,	O
because	O
queries	B-Library
to	O
search	B-Application
engines	I-Application
directly	O
constrain	O
the	O
context	O
of	O
the	O
examples	O
retrieved	O
.	O
</s>
<s>
There	O
are	O
approaches	O
that	O
alleviate	O
this	O
problem	O
,	O
such	O
as	O
using	O
several	O
different	O
seeds/queries	O
per	O
sense	O
or	O
assigning	O
senses	O
to	O
Web	B-Application
directories	I-Application
and	O
then	O
scanning	O
directories	O
for	O
examples	O
;	O
but	O
this	O
problem	O
is	O
nevertheless	O
far	O
from	O
being	O
solved	O
.	O
</s>
<s>
The	O
Web	O
as	O
a	O
social	O
network	O
has	O
been	O
successfully	O
used	O
for	O
cooperative	O
annotation	O
of	O
a	O
corpus	O
(	O
OMWE	O
,	O
Open	O
Mind	O
Word	O
Expert	O
project	O
)	O
,	O
which	O
has	O
already	O
been	O
used	O
in	O
three	O
Senseval-3	O
tasks	O
(	O
English	O
,	O
Romanian	O
and	O
Multilingual	O
)	O
.	O
</s>
<s>
The	O
Web	O
has	O
been	O
used	O
to	O
enrich	O
WordNet	O
senses	O
with	O
domain	O
information	O
:	O
topic	O
signatures	O
and	O
Web	B-Application
directories	I-Application
,	O
which	O
have	O
in	O
turn	O
been	O
successfully	O
used	O
for	O
WSD	O
.	O
</s>
<s>
Also	O
,	O
some	O
research	O
benefited	O
from	O
the	O
semantic	O
information	O
that	O
the	O
Wikipedia	O
maintains	O
on	O
its	O
disambiguation	B-General_Concept
pages	O
.	O
</s>
