Dictionary-Based
disambiguation-based on tran
(3/
For disambiguation (for e
Step1
Count the number of times
of interest occur with transl
language corpus
Setp2
Compare the counts of the t
Step3
Choose the sense that has t
corresponding sense
d Disambiguation
nslations in a second-language
/3)
example {interest, show} )
that translations of the two senses
lations of show in the second
two different senses
the higher counts as a
Dictionary-Based
one sense per discourse, o
(1/2
Most dictionary-based a
occurrence separately.
There are constraints be
occurrences that can be
disambiguation.
One sense per discourse
The sense of a target wo
any given document.
One sense per collocatio
Nearby words provide st
the sense of a target wo
context)
Disambiguation
one sense per collocation
2)
algorithms process each
etween different
e exploited for
e
ord is highly consistent within
on
trong and consistent clues to
ord. (word sense depends on
Dictionary-Based
one sense per discourse,
(2/
The first constraint is es
The material to be disam
small documents
Or can be divided into sh
For example
Discourse initial label co
D1 living th
D1 living cla
D1 ? Alt
d Disambiguation
one sense per collocation
/2)
specially useable when
mbiguated is a collection of
hort discourses
ontext
he existence of plant and animal life
assified as either plant of animal
though bacterial and plant cells are…
Unsupervised Disa
( Schutze,1998 )
Disambiguate word sens
supporting tools such as
in the absence of labeled
Simply cluster the contex
into a number of groups
these groups without lab
The probabilistic model i
as the one used for supe
P(vj | sk) are estimated u
ambiguation (1/3)
ses without having resource to
dictionaries and thesauri and
d text.
xts of an ambiguous word
and discriminate between
beling them.
is the same Bayesian model
ervised classification, but the
using the EM algorithm.
Unsupervised Disa
EM algorithm
Initialize p(v j | sk ) Æ r
Compute likelihood l(C
IK
l(C | µ) = log∏ ∑ p(ci | sk )
i=1 k =1
While l(C | µ) is improvi
E step : hi,k = p(ci | s
ΣK p(ci
k =1
M step : Re-estimate
Σ h{ci:v j∈ci } i
p(v j | sk ) = Σ ΣK
k =1 {ci:v j∈ci }
ambiguation (2/3)
random K
∑C | µ), and P(ci ) = P(ci | sk )P(sk )
k =1
IK
∑ ∑) p(sk ) = log p(ci | sk ) p(sk )
i=1 k =1
∏ing repeat: p(ci | sk ) = p(v j | sk )
s ) v j∈ci
k
i | sk )
i,k p(sk ) = ΣiI=1hi,k
ΣkK=1ΣiI=1hi,k
h} i,k
Unsupervised Disa
KK
22
1 1
Context 1 Context 2
ambiguation (3/3)
K
2
1
Context 3
The
e End