<s>
Word-sense	B-General_Concept
disambiguation	I-General_Concept
(	O
WSD	O
)	O
is	O
the	O
process	O
of	O
identifying	O
which	O
sense	O
of	O
a	O
word	O
is	O
meant	O
in	O
a	O
sentence	O
or	O
other	O
segment	O
of	O
context	O
.	O
</s>
<s>
In	O
computational	O
linguistics	O
,	O
it	O
is	O
an	O
open	O
problem	O
that	O
affects	O
other	O
computer-related	O
writing	O
,	O
such	O
as	O
discourse	O
,	O
improving	O
relevance	O
of	O
search	B-Application
engines	I-Application
,	O
anaphora	O
resolution	O
,	O
coherence	O
,	O
and	O
inference	O
.	O
</s>
<s>
Given	O
that	O
natural	O
language	O
requires	O
reflection	O
of	O
neurological	O
reality	O
,	O
as	O
shaped	O
by	O
the	O
abilities	O
provided	O
by	O
the	O
brain	O
's	O
neural	B-General_Concept
networks	I-General_Concept
,	O
computer	O
science	O
has	O
had	O
a	O
long-term	O
challenge	O
in	O
developing	O
the	O
ability	O
in	O
computers	O
to	O
do	O
natural	B-Language
language	I-Language
processing	I-Language
and	O
machine	O
learning	O
.	O
</s>
<s>
Many	O
techniques	O
have	O
been	O
researched	O
,	O
including	O
dictionary-based	O
methods	O
that	O
use	O
the	O
knowledge	O
encoded	O
in	O
lexical	B-Application
resources	I-Application
,	O
supervised	B-General_Concept
machine	I-General_Concept
learning	I-General_Concept
methods	O
in	O
which	O
a	O
classifier	B-General_Concept
is	O
trained	O
for	O
each	O
distinct	O
word	O
on	O
a	O
corpus	O
of	O
manually	O
sense-annotated	O
examples	O
,	O
and	O
completely	O
unsupervised	B-General_Concept
methods	I-General_Concept
that	O
cluster	O
occurrences	O
of	O
words	O
,	O
thereby	O
inducing	O
word	O
senses	O
.	O
</s>
<s>
Among	O
these	O
,	O
supervised	B-General_Concept
learning	I-General_Concept
approaches	O
have	O
been	O
the	O
most	O
successful	O
algorithms	O
to	O
date	O
.	O
</s>
<s>
On	O
finer-grained	O
sense	O
distinctions	O
,	O
top	O
accuracies	O
from	O
59.1	O
%	O
to	O
69.0	O
%	O
have	O
been	O
reported	O
in	O
evaluation	O
exercises	O
(	O
SemEval-2007	O
,	O
Senseval-2	O
)	O
,	O
where	O
the	O
baseline	O
accuracy	O
of	O
the	O
simplest	O
possible	O
algorithm	O
of	O
always	O
choosing	O
the	O
most	O
frequent	O
sense	O
was	O
51.4	O
%	O
and	O
57%	O
,	O
respectively	O
.	O
</s>
<s>
Disambiguation	B-General_Concept
requires	O
two	O
strict	O
inputs	O
:	O
a	O
dictionary	B-Operating_System
to	O
specify	O
the	O
senses	O
which	O
are	O
to	O
be	O
disambiguated	O
and	O
a	O
corpus	O
of	O
language	O
data	O
to	O
be	O
disambiguated	O
(	O
in	O
some	O
methods	O
,	O
a	O
training	O
corpus	O
of	O
language	O
examples	O
is	O
also	O
required	O
)	O
.	O
</s>
<s>
WSD	O
task	O
has	O
two	O
variants	O
:	O
"	O
lexical	O
sample	O
"	O
(	O
disambiguating	O
the	O
occurrences	O
of	O
a	O
small	O
sample	O
of	O
target	O
words	O
which	O
were	O
previously	O
selected	O
)	O
and	O
"	O
all	O
words	O
"	O
task	O
(	O
disambiguation	B-General_Concept
of	O
all	O
the	O
words	O
in	O
a	O
running	O
text	O
)	O
.	O
</s>
<s>
Later	O
,	O
Bar-Hillel	O
(	O
1960	O
)	O
argued	O
that	O
WSD	O
could	O
not	O
be	O
solved	O
by	O
"	O
electronic	O
computer	O
"	O
because	O
of	O
the	O
need	O
in	O
general	O
to	O
model	O
all	O
world	B-General_Concept
knowledge	I-General_Concept
.	O
</s>
<s>
In	O
the	O
1970s	O
,	O
WSD	O
was	O
a	O
subtask	O
of	O
semantic	O
interpretation	O
systems	O
developed	O
within	O
the	O
field	O
of	O
artificial	B-Application
intelligence	I-Application
,	O
starting	O
with	O
Wilks	O
 '	O
preference	O
semantics	O
.	O
</s>
<s>
By	O
the	O
1980s	O
large-scale	O
lexical	B-Application
resources	I-Application
,	O
such	O
as	O
the	O
Oxford	O
Advanced	O
Learner	O
's	O
Dictionary	B-Operating_System
of	O
Current	O
English	O
(	O
OALD	O
)	O
,	O
became	O
available	O
:	O
hand-coding	O
was	O
replaced	O
with	O
knowledge	O
automatically	O
extracted	O
from	O
these	O
resources	O
,	O
but	O
disambiguation	B-General_Concept
was	O
still	O
knowledge-based	O
or	O
dictionary-based	O
.	O
</s>
<s>
In	O
the	O
1990s	O
,	O
the	O
statistical	O
revolution	O
advanced	O
computational	O
linguistics	O
,	O
and	O
WSD	O
became	O
a	O
paradigm	O
problem	O
on	O
which	O
to	O
apply	O
supervised	B-General_Concept
machine	I-General_Concept
learning	I-General_Concept
techniques	O
.	O
</s>
<s>
The	O
2000s	O
saw	O
supervised	B-General_Concept
techniques	O
reach	O
a	O
plateau	O
in	O
accuracy	O
,	O
and	O
so	O
attention	O
has	O
shifted	O
to	O
coarser-grained	O
senses	O
,	O
domain	B-General_Concept
adaptation	I-General_Concept
,	O
semi-supervised	B-General_Concept
and	O
unsupervised	O
corpus-based	O
systems	O
,	O
combinations	O
of	O
different	O
methods	O
,	O
and	O
the	O
return	O
of	O
knowledge-based	O
systems	O
via	O
graph-based	O
methods	O
.	O
</s>
<s>
Still	O
,	O
supervised	B-General_Concept
systems	O
continue	O
to	O
perform	O
best	O
.	O
</s>
<s>
One	O
problem	O
with	O
word	B-General_Concept
sense	I-General_Concept
disambiguation	I-General_Concept
is	O
deciding	O
what	O
the	O
senses	O
are	O
,	O
as	O
different	O
dictionaries	B-Operating_System
and	O
thesauruses	O
will	O
provide	O
different	O
divisions	O
of	O
words	O
into	O
senses	O
.	O
</s>
<s>
Some	O
researchers	O
have	O
suggested	O
choosing	O
a	O
particular	O
dictionary	B-Operating_System
,	O
and	O
using	O
its	O
set	O
of	O
senses	O
to	O
deal	O
with	O
this	O
issue	O
.	O
</s>
<s>
Most	O
research	O
in	O
the	O
field	O
of	O
WSD	O
is	O
performed	O
by	O
using	O
WordNet	B-General_Concept
as	O
a	O
reference	O
sense	O
inventory	O
for	O
English	O
.	O
</s>
<s>
WordNet	B-General_Concept
is	O
a	O
computational	O
lexicon	B-Application
that	O
encodes	O
concepts	O
as	O
synonym	B-General_Concept
sets	I-General_Concept
(	O
e.g.	O
</s>
<s>
Other	O
resources	O
used	O
for	O
disambiguation	B-General_Concept
purposes	O
include	O
Roget	O
's	O
Thesaurus	O
and	O
Wikipedia	O
.	O
</s>
<s>
More	O
recently	O
,	O
BabelNet	B-Application
,	O
a	O
multilingual	O
encyclopedic	O
dictionary	B-Operating_System
,	O
has	O
been	O
used	O
for	O
multilingual	O
WSD	O
.	O
</s>
<s>
in	O
the	O
Senseval/SemEval	O
competitions	O
parts	O
of	O
speech	O
are	O
provided	O
as	O
input	O
for	O
the	O
text	O
to	O
disambiguate	B-General_Concept
)	O
.	O
</s>
<s>
The	O
success	O
rate	O
for	O
part-of-speech	O
tagging	O
algorithms	O
is	O
at	O
present	O
much	O
higher	O
than	O
that	O
for	O
WSD	O
,	O
state-of-the	O
art	O
being	O
around	O
96%	O
accuracy	O
or	O
better	O
,	O
as	O
compared	O
to	O
less	O
than	O
75%	O
accuracy	O
in	O
word	B-General_Concept
sense	I-General_Concept
disambiguation	I-General_Concept
with	O
supervised	B-General_Concept
learning	I-General_Concept
.	O
</s>
<s>
Some	O
AI	B-Application
researchers	O
like	O
Douglas	O
Lenat	O
argue	O
that	O
one	O
cannot	O
parse	O
meanings	O
from	O
words	O
without	O
some	O
form	O
of	O
common	O
sense	O
ontology	B-Language
.	O
</s>
<s>
Moreover	O
,	O
sometimes	O
the	O
common	O
sense	O
is	O
needed	O
to	O
disambiguate	B-General_Concept
such	O
words	O
like	O
pronouns	O
in	O
case	O
of	O
having	O
anaphoras	O
or	O
cataphoras	O
in	O
the	O
text	O
.	O
</s>
<s>
In	O
information	B-Library
retrieval	I-Library
,	O
a	O
sense	O
inventory	O
is	O
not	O
necessarily	O
required	O
,	O
because	O
it	O
is	O
enough	O
to	O
know	O
that	O
a	O
word	O
is	O
used	O
in	O
the	O
same	O
sense	O
in	O
the	O
query	B-Library
and	O
a	O
retrieved	O
document	O
;	O
what	O
sense	O
that	O
is	O
,	O
is	O
unimportant	O
.	O
</s>
<s>
For	O
example	O
,	O
in	O
Senseval-2	O
,	O
which	O
used	O
fine-grained	O
sense	O
distinctions	O
,	O
human	O
annotators	O
agreed	O
in	O
only	O
85%	O
of	O
word	O
occurrences	O
.	O
</s>
<s>
In	O
2009	O
,	O
a	O
task	O
–	O
named	O
lexical	B-General_Concept
substitution	I-General_Concept
–	O
was	O
proposed	O
as	O
a	O
possible	O
solution	O
to	O
the	O
sense	O
discreteness	O
problem	O
.	O
</s>
<s>
The	O
task	O
consists	O
of	O
providing	O
a	O
substitute	O
for	O
a	O
word	O
in	O
context	O
that	O
preserves	O
the	O
meaning	O
of	O
the	O
original	O
word	O
(	O
potentially	O
,	O
substitutes	O
can	O
be	O
chosen	O
from	O
the	O
full	O
lexicon	B-Application
of	O
the	O
target	O
language	O
,	O
thus	O
overcoming	O
discreteness	O
)	O
.	O
</s>
<s>
Deep	O
approaches	O
presume	O
access	O
to	O
a	O
comprehensive	O
body	O
of	O
world	B-General_Concept
knowledge	I-General_Concept
.	O
</s>
<s>
Additionally	O
due	O
to	O
the	O
long	O
tradition	O
in	O
computational	O
linguistics	O
,	O
of	O
trying	O
such	O
approaches	O
in	O
terms	O
of	O
coded	O
knowledge	O
and	O
in	O
some	O
cases	O
,	O
it	O
can	O
be	O
hard	O
to	O
distinguish	O
between	O
knowledge	O
involved	O
in	O
linguistic	O
or	O
world	B-General_Concept
knowledge	I-General_Concept
.	O
</s>
<s>
This	O
approach	O
,	O
while	O
theoretically	O
not	O
as	O
powerful	O
as	O
deep	O
approaches	O
,	O
gives	O
superior	O
results	O
in	O
practice	O
,	O
due	O
to	O
the	O
computer	O
's	O
limited	O
world	B-General_Concept
knowledge	I-General_Concept
.	O
</s>
<s>
Dictionary	B-Operating_System
-	O
and	O
knowledge-based	O
methods	O
:	O
These	O
rely	O
primarily	O
on	O
dictionaries	B-Operating_System
,	O
thesauri	O
,	O
and	O
lexical	O
knowledge	O
bases	O
,	O
without	O
using	O
any	O
corpus	O
evidence	O
.	O
</s>
<s>
Semi-supervised	B-General_Concept
or	I-General_Concept
minimally	I-General_Concept
supervised	I-General_Concept
methods	I-General_Concept
:	O
These	O
make	O
use	O
of	O
a	O
secondary	O
source	O
of	O
knowledge	O
such	O
as	O
a	O
small	O
annotated	O
corpus	O
as	O
seed	O
data	O
in	O
a	O
bootstrapping	B-Application
process	O
,	O
or	O
a	O
word-aligned	O
bilingual	O
corpus	O
.	O
</s>
<s>
Supervised	B-General_Concept
methods	I-General_Concept
:	O
These	O
make	O
use	O
of	O
sense-annotated	O
corpora	O
to	O
train	O
from	O
.	O
</s>
<s>
Unsupervised	B-General_Concept
methods	I-General_Concept
:	O
These	O
eschew	O
(	O
almost	O
)	O
completely	O
external	O
information	O
and	O
work	O
directly	O
from	O
raw	O
unannotated	O
corpora	O
.	O
</s>
<s>
These	O
methods	O
are	O
also	O
known	O
under	O
the	O
name	O
of	O
word	B-General_Concept
sense	I-General_Concept
discrimination	I-General_Concept
.	O
</s>
<s>
Two	O
shallow	O
approaches	O
used	O
to	O
train	O
and	O
then	O
disambiguate	B-General_Concept
are	O
Naïve	B-General_Concept
Bayes	I-General_Concept
classifiers	I-General_Concept
and	O
decision	B-Algorithm
trees	I-Algorithm
.	O
</s>
<s>
In	O
recent	O
research	O
,	O
kernel-based	B-Algorithm
methods	I-Algorithm
such	O
as	O
support	B-Algorithm
vector	I-Algorithm
machines	I-Algorithm
have	O
shown	O
superior	O
performance	O
in	O
supervised	B-General_Concept
learning	I-General_Concept
.	O
</s>
<s>
The	O
Lesk	B-General_Concept
algorithm	I-General_Concept
is	O
the	O
seminal	O
dictionary-based	O
method	O
.	O
</s>
<s>
Two	O
(	O
or	O
more	O
)	O
words	O
are	O
disambiguated	O
by	O
finding	O
the	O
pair	O
of	O
dictionary	B-Operating_System
senses	O
with	O
the	O
greatest	O
word	O
overlap	O
in	O
their	O
dictionary	B-Operating_System
definitions	O
.	O
</s>
<s>
For	O
example	O
,	O
when	O
disambiguating	O
the	O
words	O
in	O
"	O
pine	O
cone	O
"	O
,	O
the	O
definitions	O
of	O
the	O
appropriate	O
senses	O
both	O
include	O
the	O
words	O
evergreen	O
and	O
tree	O
(	O
at	O
least	O
in	O
one	O
dictionary	B-Operating_System
)	O
.	O
</s>
<s>
An	O
alternative	O
to	O
the	O
use	O
of	O
the	O
definitions	O
is	O
to	O
consider	O
general	O
word-sense	O
relatedness	O
and	O
to	O
compute	O
the	O
semantic	O
similarity	O
of	O
each	O
pair	O
of	O
word	O
senses	O
based	O
on	O
a	O
given	O
lexical	O
knowledge	O
base	O
such	O
as	O
WordNet	B-General_Concept
.	O
</s>
<s>
Graph-based	O
methods	O
reminiscent	O
of	O
spreading	B-Algorithm
activation	I-Algorithm
research	O
of	O
the	O
early	O
days	O
of	O
AI	B-Application
research	I-Application
have	O
been	O
applied	O
with	O
some	O
success	O
.	O
</s>
<s>
More	O
complex	O
graph-based	O
approaches	O
have	O
been	O
shown	O
to	O
perform	O
almost	O
as	O
well	O
as	O
supervised	B-General_Concept
methods	I-General_Concept
or	O
even	O
outperforming	O
them	O
on	O
specific	O
domains	O
.	O
</s>
<s>
Also	O
,	O
automatically	O
transferring	O
knowledge	O
in	O
the	O
form	O
of	O
semantic	O
relations	O
from	O
Wikipedia	O
to	O
WordNet	B-General_Concept
has	O
been	O
shown	O
to	O
boost	O
simple	O
knowledge-based	O
methods	O
,	O
enabling	O
them	O
to	O
rival	O
the	O
best	O
supervised	B-General_Concept
systems	O
and	O
even	O
outperform	O
them	O
in	O
a	O
domain-specific	O
setting	O
.	O
</s>
<s>
The	O
use	O
of	O
selectional	O
preferences	O
(	O
or	O
selectional	O
restrictions	O
)	O
is	O
also	O
useful	O
,	O
for	O
example	O
,	O
knowing	O
that	O
one	O
typically	O
cooks	O
food	O
,	O
one	O
can	O
disambiguate	B-General_Concept
the	O
word	O
bass	O
in	O
"	O
I	O
am	O
cooking	O
basses	O
"	O
(	O
i.e.	O
,	O
it	O
's	O
not	O
a	O
musical	O
instrument	O
)	O
.	O
</s>
<s>
Supervised	B-General_Concept
methods	I-General_Concept
are	O
based	O
on	O
the	O
assumption	O
that	O
the	O
context	O
can	O
provide	O
enough	O
evidence	O
on	O
its	O
own	O
to	O
disambiguate	B-General_Concept
words	O
(	O
hence	O
,	O
common	O
sense	O
and	O
reasoning	O
are	O
deemed	O
unnecessary	O
)	O
.	O
</s>
<s>
Probably	O
every	O
machine	O
learning	O
algorithm	O
going	O
has	O
been	O
applied	O
to	O
WSD	O
,	O
including	O
associated	O
techniques	O
such	O
as	O
feature	B-General_Concept
selection	I-General_Concept
,	O
parameter	O
optimization	O
,	O
and	O
ensemble	B-Algorithm
learning	I-Algorithm
.	O
</s>
<s>
Support	B-Algorithm
Vector	I-Algorithm
Machines	I-Algorithm
and	O
memory-based	B-General_Concept
learning	I-General_Concept
have	O
been	O
shown	O
to	O
be	O
the	O
most	O
successful	O
approaches	O
,	O
to	O
date	O
,	O
probably	O
because	O
they	O
can	O
cope	O
with	O
the	O
high-dimensionality	O
of	O
the	O
feature	O
space	O
.	O
</s>
<s>
However	O
,	O
these	O
supervised	B-General_Concept
methods	I-General_Concept
are	O
subject	O
to	O
a	O
new	O
knowledge	O
acquisition	O
bottleneck	O
since	O
they	O
rely	O
on	O
substantial	O
amounts	O
of	O
manually	O
sense-tagged	O
corpora	O
for	O
training	O
,	O
which	O
are	O
laborious	O
and	O
expensive	O
to	O
create	O
.	O
</s>
<s>
Because	O
of	O
the	O
lack	O
of	O
training	O
data	O
,	O
many	O
word	B-General_Concept
sense	I-General_Concept
disambiguation	I-General_Concept
algorithms	O
use	O
semi-supervised	B-General_Concept
learning	I-General_Concept
,	O
which	O
allows	O
both	O
labeled	O
and	O
unlabeled	O
data	O
.	O
</s>
<s>
It	O
uses	O
the	O
‘	O
One	O
sense	O
per	O
collocation’	O
and	O
the	O
‘	O
One	O
sense	O
per	O
discourse’	O
properties	O
of	O
human	O
languages	O
for	O
word	B-General_Concept
sense	I-General_Concept
disambiguation	I-General_Concept
.	O
</s>
<s>
From	O
observation	O
,	O
words	O
tend	O
to	O
exhibit	O
only	O
one	O
sense	O
in	O
most	O
given	O
discourse	O
and	O
in	O
a	O
given	O
collocation	B-General_Concept
.	O
</s>
<s>
The	O
bootstrapping	B-Application
approach	O
starts	O
from	O
a	O
small	O
amount	O
of	O
seed	O
data	O
for	O
each	O
word	O
:	O
either	O
manually	O
tagged	O
training	O
examples	O
or	O
a	O
small	O
number	O
of	O
surefire	O
decision	O
rules	O
(	O
e.g.	O
,	O
'	O
play	O
 '	O
in	O
the	O
context	O
of	O
'	O
bass	O
 '	O
almost	O
always	O
indicates	O
the	O
musical	O
instrument	O
)	O
.	O
</s>
<s>
The	O
seeds	O
are	O
used	O
to	O
train	O
an	O
initial	O
classifier	B-General_Concept
,	O
using	O
any	O
supervised	B-General_Concept
method	O
.	O
</s>
<s>
This	O
classifier	B-General_Concept
is	O
then	O
used	O
on	O
the	O
untagged	O
portion	O
of	O
the	O
corpus	O
to	O
extract	O
a	O
larger	O
training	O
set	O
,	O
in	O
which	O
only	O
the	O
most	O
confident	O
classifications	O
are	O
included	O
.	O
</s>
<s>
The	O
process	O
repeats	O
,	O
each	O
new	O
classifier	B-General_Concept
being	O
trained	O
on	O
a	O
successively	O
larger	O
training	O
corpus	O
,	O
until	O
the	O
whole	O
corpus	O
is	O
consumed	O
,	O
or	O
until	O
a	O
given	O
maximum	O
number	O
of	O
iterations	O
is	O
reached	O
.	O
</s>
<s>
Other	O
semi-supervised	B-General_Concept
techniques	O
use	O
large	O
quantities	O
of	O
untagged	O
corpora	O
to	O
provide	O
co-occurrence	O
information	O
that	O
supplements	O
the	O
tagged	O
corpora	O
.	O
</s>
<s>
These	O
techniques	O
have	O
the	O
potential	O
to	O
help	O
in	O
the	O
adaptation	O
of	O
supervised	B-General_Concept
models	O
to	O
different	O
domains	O
.	O
</s>
<s>
Word-aligned	O
bilingual	O
corpora	O
have	O
been	O
used	O
to	O
infer	O
cross-lingual	O
sense	O
distinctions	O
,	O
a	O
kind	O
of	O
semi-supervised	B-General_Concept
system	O
.	O
</s>
<s>
Unsupervised	B-General_Concept
learning	I-General_Concept
is	O
the	O
greatest	O
challenge	O
for	O
WSD	O
researchers	O
.	O
</s>
<s>
The	O
underlying	O
assumption	O
is	O
that	O
similar	O
senses	O
occur	O
in	O
similar	O
contexts	O
,	O
and	O
thus	O
senses	O
can	O
be	O
induced	B-General_Concept
from	O
text	O
by	O
clustering	B-Algorithm
word	O
occurrences	O
using	O
some	O
measure	O
of	O
similarity	O
of	O
context	O
,	O
a	O
task	O
referred	O
to	O
as	O
word	B-General_Concept
sense	I-General_Concept
induction	I-General_Concept
or	O
discrimination	O
.	O
</s>
<s>
Then	O
,	O
new	O
occurrences	O
of	O
the	O
word	O
can	O
be	O
classified	O
into	O
the	O
closest	O
induced	B-General_Concept
clusters/senses	O
.	O
</s>
<s>
Performance	O
has	O
been	O
lower	O
than	O
for	O
the	O
other	O
methods	O
described	O
above	O
,	O
but	O
comparisons	O
are	O
difficult	O
since	O
senses	O
induced	B-General_Concept
must	O
be	O
mapped	O
to	O
a	O
known	O
dictionary	B-Operating_System
of	O
word	O
senses	O
.	O
</s>
<s>
If	O
a	O
mapping	B-Algorithm
to	O
a	O
set	O
of	O
dictionary	B-Operating_System
senses	O
is	O
not	O
desired	O
,	O
cluster-based	O
evaluations	O
(	O
including	O
measures	O
of	O
entropy	O
and	O
purity	O
)	O
can	O
be	O
performed	O
.	O
</s>
<s>
Alternatively	O
,	O
word	B-General_Concept
sense	I-General_Concept
induction	I-General_Concept
methods	O
can	O
be	O
tested	O
and	O
compared	O
within	O
an	O
application	O
.	O
</s>
<s>
For	O
instance	O
,	O
it	O
has	O
been	O
shown	O
that	O
word	B-General_Concept
sense	I-General_Concept
induction	I-General_Concept
improves	O
Web	B-Application
search	I-Application
result	O
clustering	B-Algorithm
by	O
increasing	O
the	O
quality	O
of	O
result	O
clusters	O
and	O
the	O
degree	O
diversification	O
of	O
result	O
lists	O
.	O
</s>
<s>
It	O
is	O
hoped	O
that	O
unsupervised	B-General_Concept
learning	I-General_Concept
will	O
overcome	O
the	O
knowledge	O
acquisition	O
bottleneck	O
because	O
they	O
are	O
not	O
dependent	O
on	O
manual	O
effort	O
.	O
</s>
<s>
Representing	O
words	O
considering	O
their	O
context	O
through	O
fixed	O
size	O
dense	O
vectors	O
(	O
word	B-General_Concept
embeddings	I-General_Concept
)	O
has	O
become	O
one	O
of	O
the	O
most	O
fundamental	O
blocks	O
in	O
several	O
NLP	B-Language
systems	O
.	O
</s>
<s>
Even	O
though	O
most	O
of	O
traditional	O
word	B-General_Concept
embedding	I-General_Concept
techniques	O
conflate	O
words	O
with	O
multiple	O
meanings	O
into	O
a	O
single	O
vector	O
representation	O
,	O
they	O
still	O
can	O
be	O
used	O
to	O
improve	O
WSD	O
.	O
</s>
<s>
A	O
simple	O
approach	O
to	O
employ	O
pre-computed	O
word	B-General_Concept
embeddings	I-General_Concept
to	O
represent	O
word	O
senses	O
is	O
to	O
compute	O
the	O
centroids	O
of	O
sense	O
clusters	O
.	O
</s>
<s>
In	O
addition	O
to	O
word	B-General_Concept
embeddings	I-General_Concept
techniques	O
,	O
lexical	O
databases	O
(	O
e.g.	O
,	O
WordNet	B-General_Concept
,	O
ConceptNet	O
,	O
BabelNet	B-Application
)	O
can	O
also	O
assist	O
unsupervised	O
systems	O
in	O
mapping	B-Algorithm
words	O
and	O
their	O
senses	O
as	O
dictionaries	B-Operating_System
.	O
</s>
<s>
Some	O
techniques	O
that	O
combine	O
lexical	O
databases	O
and	O
word	B-General_Concept
embeddings	I-General_Concept
are	O
presented	O
in	O
AutoExtend	O
and	O
Most	O
Suitable	O
Sense	O
Annotation	O
(	O
MSSA	O
)	O
.	O
</s>
<s>
synsets	B-General_Concept
in	O
WordNet	B-General_Concept
)	O
objects	O
as	O
nodes	O
and	O
the	O
relationship	O
between	O
nodes	O
as	O
edges	O
.	O
</s>
<s>
In	O
MSSA	O
,	O
an	O
unsupervised	O
disambiguation	B-General_Concept
system	O
uses	O
the	O
similarity	O
between	O
word	O
senses	O
in	O
a	O
fixed	O
context	O
window	O
to	O
select	O
the	O
most	O
suitable	O
word	O
sense	O
using	O
a	O
pre-trained	O
word	B-General_Concept
embedding	I-General_Concept
model	O
and	O
WordNet	B-General_Concept
.	O
</s>
<s>
For	O
each	O
context	O
window	O
,	O
MSSA	O
calculates	O
the	O
centroid	O
of	O
each	O
word	O
sense	O
definition	O
by	O
averaging	O
the	O
word	B-General_Concept
vectors	I-General_Concept
of	O
its	O
words	O
in	O
WordNet	B-General_Concept
's	O
glosses	O
(	O
i.e.	O
,	O
short	O
defining	O
gloss	O
and	O
one	O
or	O
more	O
usage	O
example	O
)	O
using	O
a	O
pre-trained	O
word	B-General_Concept
embeddings	I-General_Concept
model	O
.	O
</s>
<s>
After	O
all	O
words	O
are	O
annotated	O
and	O
disambiguated	O
,	O
they	O
can	O
be	O
used	O
as	O
a	O
training	O
corpus	O
in	O
any	O
standard	O
word	B-General_Concept
embedding	I-General_Concept
technique	O
.	O
</s>
<s>
In	O
its	O
improved	O
version	O
,	O
MSSA	O
can	O
make	O
use	O
of	O
word	O
sense	O
embeddings	O
to	O
repeat	O
its	O
disambiguation	B-General_Concept
process	O
iteratively	O
.	O
</s>
<s>
Domain-driven	O
disambiguation	B-General_Concept
;	O
</s>
<s>
Hindi	O
:	O
Lack	O
of	O
lexical	B-Application
resources	I-Application
in	O
Hindi	O
have	O
hindered	O
the	O
performance	O
of	O
supervised	B-General_Concept
models	O
of	O
WSD	O
,	O
while	O
the	O
unsupervised	O
models	O
suffer	O
due	O
to	O
extensive	O
morphology	O
.	O
</s>
<s>
The	O
creation	O
of	O
the	O
has	O
paved	O
way	O
for	O
several	O
Supervised	B-General_Concept
methods	I-General_Concept
which	O
have	O
been	O
proven	O
to	O
produce	O
a	O
higher	O
accuracy	O
in	O
disambiguating	O
nouns	O
.	O
</s>
<s>
Unsupervised	B-General_Concept
methods	I-General_Concept
rely	O
on	O
knowledge	O
about	O
word	O
senses	O
,	O
which	O
is	O
only	O
sparsely	O
formulated	O
in	O
dictionaries	B-Operating_System
and	O
lexical	O
databases	O
.	O
</s>
<s>
Supervised	B-General_Concept
methods	I-General_Concept
depend	O
crucially	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
so	O
far	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
successfully	O
mine	O
the	O
Web	O
for	O
information	O
to	O
use	O
in	O
WSD	O
.	O
</s>
<s>
The	O
historic	O
lack	O
of	O
training	O
data	O
has	O
provoked	O
the	O
appearance	O
of	O
some	O
new	O
algorithms	O
and	O
techniques	O
,	O
as	O
described	O
in	O
Automatic	B-General_Concept
acquisition	I-General_Concept
of	I-General_Concept
sense-tagged	I-General_Concept
corpora	I-General_Concept
.	O
</s>
<s>
They	O
can	O
vary	O
from	O
corpora	O
of	O
texts	O
,	O
either	O
unlabeled	O
or	O
annotated	O
with	O
word	O
senses	O
,	O
to	O
machine-readable	O
dictionaries	B-Operating_System
,	O
thesauri	O
,	O
glossaries	O
,	O
ontologies	B-Language
,	O
etc	O
.	O
</s>
<s>
Before	O
the	O
organization	O
of	O
specific	O
evaluation	O
campaigns	O
most	O
systems	O
were	O
assessed	O
on	O
in-house	O
,	O
often	O
small-scale	O
,	O
data	B-General_Concept
sets	I-General_Concept
.	O
</s>
<s>
In	O
order	O
to	O
define	O
common	O
evaluation	O
datasets	B-General_Concept
and	O
procedures	O
,	O
public	O
evaluation	O
campaigns	O
have	O
been	O
organized	O
.	O
</s>
<s>
Senseval	B-General_Concept
(	O
now	O
renamed	O
SemEval	B-General_Concept
)	O
is	O
an	O
international	O
word	B-General_Concept
sense	I-General_Concept
disambiguation	I-General_Concept
competition	O
,	O
held	O
every	O
three	O
years	O
since	O
1998	O
:	O
(	O
1998	O
)	O
,	O
(	O
2001	O
)	O
,	O
(	O
2004	O
)	O
,	O
and	O
its	O
successor	O
,	O
(	O
2007	O
)	O
.	O
</s>
<s>
The	O
objective	O
of	O
the	O
competition	O
is	O
to	O
organize	O
different	O
lectures	O
,	O
preparing	O
and	O
hand-annotating	O
corpus	O
for	O
testing	O
systems	O
,	O
perform	O
a	O
comparative	O
evaluation	O
of	O
WSD	O
systems	O
in	O
several	O
kinds	O
of	O
tasks	O
,	O
including	O
all-words	O
and	O
lexical	O
sample	O
WSD	O
for	O
different	O
languages	O
,	O
and	O
,	O
more	O
recently	O
,	O
new	O
tasks	O
such	O
as	O
semantic	O
role	O
labeling	O
,	O
gloss	O
WSD	O
,	O
lexical	B-General_Concept
substitution	I-General_Concept
,	O
etc	O
.	O
</s>
<s>
The	O
systems	O
submitted	O
for	O
evaluation	O
to	O
these	O
competitions	O
usually	O
integrate	O
different	O
techniques	O
and	O
often	O
combine	O
supervised	B-General_Concept
and	O
knowledge-based	O
methods	O
(	O
especially	O
for	O
avoiding	O
bad	O
performance	O
in	O
lack	O
of	O
training	O
examples	O
)	O
.	O
</s>
<s>
As	O
technology	O
evolves	O
,	O
the	O
Word	B-General_Concept
Sense	I-General_Concept
Disambiguation	I-General_Concept
(	O
WSD	O
)	O
tasks	O
grows	O
in	O
different	O
flavors	O
towards	O
various	O
research	O
directions	O
and	O
for	O
more	O
languages	O
:	O
</s>
<s>
Classic	B-General_Concept
monolingual	I-General_Concept
WSD	I-General_Concept
evaluation	O
tasks	O
use	O
WordNet	B-General_Concept
as	O
the	O
sense	O
inventory	O
and	O
are	O
largely	O
based	O
on	O
supervised/semi	O
-supervised	O
classification	O
with	O
the	O
manually	O
sense	O
annotated	O
corpora	O
:	O
</s>
<s>
Classic	O
English	O
WSD	O
uses	O
the	O
Princeton	B-General_Concept
WordNet	I-General_Concept
as	O
it	O
sense	O
inventory	O
and	O
the	O
primary	O
classification	O
input	O
is	O
normally	O
based	O
on	O
the	O
corpus	O
.	O
</s>
<s>
Classical	O
WSD	O
for	O
other	O
languages	O
uses	O
their	O
respective	O
WordNet	B-General_Concept
as	O
sense	O
inventories	O
and	O
sense	O
annotated	O
corpora	O
tagged	O
in	O
their	O
respective	O
languages	O
.	O
</s>
<s>
Multilingual	O
WSD	O
evaluation	O
tasks	O
focused	O
on	O
WSD	O
across	O
2	O
or	O
more	O
languages	O
simultaneously	O
,	O
using	O
their	O
respective	O
WordNets	B-General_Concept
as	O
its	O
sense	O
inventories	O
or	O
BabelNet	B-Application
as	O
multilingual	O
sense	O
inventory	O
.	O
</s>
<s>
It	O
evolved	O
from	O
the	O
Translation	O
WSD	O
evaluation	O
tasks	O
that	O
took	O
place	O
in	O
Senseval-2	O
.	O
</s>
<s>
Word	B-General_Concept
Sense	I-General_Concept
Induction	I-General_Concept
and	I-General_Concept
Disambiguation	I-General_Concept
task	I-General_Concept
is	O
a	O
combined	O
task	O
evaluation	O
where	O
the	O
sense	O
inventory	O
is	O
first	O
induced	B-General_Concept
from	O
a	O
fixed	O
training	O
set	O
data	O
,	O
consisting	O
of	O
polysemous	O
words	O
and	O
the	O
sentence	O
that	O
they	O
occurred	O
in	O
,	O
then	O
WSD	O
is	O
performed	O
on	O
a	O
different	O
testing	O
data	B-General_Concept
set	I-General_Concept
.	O
</s>
