Method and apparatus for training bilingual word alignment model, method and apparatus for bilingual word alignment
View Patent ↗The present invention provides method and apparatus for bilingual word alignment, method and apparatus for training bilingual word alignment model. The method for training bilingual word alignment model, comprising: training a bilingual word alignment model for a first language and a second language, using a bilingual corpus of the first and second languages; training a bilingual word alignment model for the second language and a third language, using a bilingual corpus of the second and third languages; and estimating a bilingual word alignment model for the first language and the third language, based on said bilingual word alignment model for the first and second languages and said bilingual word alignment model for the second and third languages.
1. A method for training bilingual word alignment model via a programmed computer, comprising:
training a bilingual word alignment model for a first language and a second language, using a bilingual corpus of the first and second languages;
training a bilingual word alignment model for the second language and a third language, using a bilingual corpus of the second and third languages; and
estimating a bilingual word alignment model for the first language and the third language, based on said bilingual word alignment model for the first and second languages and said bilingual word alignment model for the second and third languages, wherein
said bilingual word alignment model for the first and second languages and said bilingual word alignment model for the second and third languages respectively comprises a word translation sub-model, a position distortion sub-model, and a word fertility sub-model;
said step of estimating a bilingual word alignment model for the first language and the third language comprises:
estimating a word translation sub-model for the first and third languages, based on the word translation sub-model for the first and second languages and the word translation sub-model for the second and third languages;
estimating a position distortion sub-model for the first and third languages, based on the position distortion sub-model for the first and second languages and the position distortion sub-model for the second and third languages; and
estimating a word fertility sub-model for the first and third languages, based on the word fertility sub-model for the first and second languages and/or the word fertility sub-model for the second and third languages, the word translation sub-model for the first and second languages and/or the word translation sub-model for the second and third languages, and
said step of estimating a word translation sub-model for the first and third languages comprises:
where p CE (w c |w e ) represents the translation probability from the second language word w e to the first language word w c , p EJ (w e |w j ) represents the translation probability from the third language word w j to the second language word w e , C(w j ,w c ) represents the co-occurrence count of the first language word w c and the third language word w j , p(w c |w j ) represents the translation probability from the third language word w j , to the first language word w c ,
collecting the co-occurrence count of the first language word w c and the third language word w j , using formula
C
(
w
j
,
w
c
)
=
∑
w
e
p
EJ
(
w
e
|
w
j
)
*
p
CE
(
w
c
|
w
e
)
;
and
calculating the translation probability from the third language word w j to the first language word w c , using formula
p
(
w
c
|
w
j
)
=
C
(
w
j
,
w
c
)
∑
w
c
′
C
(
w
j
,
w
c
′
)
.
2. The method for training bilingual word alignment model according to claim 1 , wherein said step of estimating a position distortion sub-model for the first and third languages comprises:
where p EJ (k|i,l,m′) represents the probability that the i th position in the third language sentence having a length of l is corresponding to the k th position in the second language sentence having a length of m′, p CE (j|k,m′,m) represents the probability that the k th position in the second language sentence having a length of m′ is corresponding to the j th position in the first language sentence having a length of m, C(j,i,l,m) and p CJ (j|i,l,m) respectively represent the co-occurrence count and probability that the i th position in the third language sentence having a length of l is corresponding to the j th position in the first language sentence having a length of m,
collecting the co-occurrence count that the i th position in the third language sentence having a length of l is corresponding to the j th position in the first language sentence having a length of m, using formula C(j,i,l,m)=Σ k,m′ p EJ (k|i,l,m′)*p CE (j|k,m′,m); and
calculating the position distortion probability that the i th position in the third language sentence having a length of l is corresponding to the j th position in the first language sentence having a length of m, using formula
p
CJ
(
j
|
i
,
l
,
m
)
=
C
(
j
,
i
,
l
,
m
)
∑
j
′
C
(
j
′
,
i
,
l
,
m
)
.
3. The method for training bilingual word alignment model according to claim 1 , wherein said step of estimating a word fertility sub-model for the first and third languages comprises:
where p EJ (w e |w j ) represents the translation probability from the third language word w j to the second language word w e , p CE (φ i |w e ) represents the probability that the second language word w e is corresponding to φ i words of the first language, C(φ i ,w j ) and p(φ i |w j ) respectively represent the co-occurrence count and probability that the third language word w j is corresponding to φ i words of the first language,
collecting the co-occurrence count that the third language word w j is corresponding to φ i words of the first language, using formula
C
(
φ
i
,
w
j
)
=
∑
w
e
p
EJ
(
w
e
|
w
j
)
*
p
CE
(
φ
i
|
w
e
)
;
and
calculating the probability that the third language word w j is corresponding to φ i words of the first language, using formula
p
(
φ
i
|
w
j
)
=
C
(
φ
i
,
w
j
)
∑
φ
i
′
C
(
φ
i
′
,
w
j
)
.
4. A method for bilingual word alignment via a programmed computer, comprising:
obtaining a bilingual word alignment model for a first language and a third language based on the bilingual corpus of the first and second languages and the bilingual corpus of the second and third languages, by using the method for training bilingual word alignment model according to any one of claim 1 , 2 , or 3 ;
word-aligning a bilingual sentence pair of the first and third languages using said bilingual word alignment model of the first and third languages.
5. An apparatus for training bilingual word alignment model, comprising:
a first training unit configured to train a bilingual word alignment model for a first language and a second language, using a bilingual corpus of the first and second languages;
a second training unit configured to train a bilingual word alignment model for the second language and a third language, using a bilingual corpus of the second and third languages; and
a model estimating unit configured to estimate a bilingual word alignment model for the first language and the third language, based on said bilingual word alignment model for the first and second languages and said bilingual word alignment model for the second and third languages, wherein
said bilingual word alignment model for the first and second languages and said bilingual word alignment model for the second and third languages respectively comprises a word translation sub-model, a position distortion sub-model and a word fertility sub-model;
said model estimating unit comprises:
a word translation sub-model estimating unit configured to estimate a word translation sub-model for the first and third languages, based on the word translation sub-model for the first and second languages and the word translation sub-model for the second and third languages;
a position distortion sub-model estimating unit configured to estimate a position distortion sub-model for the first and third languages, based on the position distortion sub-model for the first and second languages and the position distortion sub-model for the second and third languages; and
a word fertility sub-model estimating unit configured to estimate a word fertility sub-model for the first and third languages, based on the word fertility sub-model for the first and second languages and/or the word fertility sub-model for the second and third languages, the word translation sub-model for the first and second languages and/or the word translation sub-model for the second and third languages;
where p CE (w c |w e ) represents the translation probability from the second language word w e to the first language word w c , p EJ (w e |w j ) represents the translation probability form the third language word w j to the second language word w e , C(w j ,w c ) represents the co-occurrence count of the first language word w c and the third language word w j , p(w c |w j ) represents the translation probability from the third language word w j to the first language word w c ,
said word translation sub-model estimating unit collects the co-occurrence count of the first language word w c and the third language word w j , using formula
C
(
w
j
,
w
c
)
=
∑
w
e
p
EJ
(
w
e
|
w
j
)
*
p
CE
(
w
c
|
w
e
)
,
and calculates the translation probability from the third language word w j to the first language word w c , using formula
p
(
w
c
|
w
j
)
=
C
(
w
j
,
w
c
)
∑
w
c
′
C
(
w
j
,
w
c
′
)
.
6. The apparatus for training bilingual word alignment model according to claim 5 , wherein
p EJ (k|i,l,m′) represents the probability that the i th position in the third language sentence having a length of l is corresponding to the k th position in the second language sentence having a length of m′, p CE (j|k,m′,m) represents the probability that the k th position in the second language sentence having a length of m′ is corresponding to the j th position in the first language sentence having a length of m, C(j,i,l,m) and p CJ (j|i,l,m) respectively represent the co-occurrence count and probability that the i th position in the third language sentence having a length of l is corresponding to the j th position in the first language sentence having a length of m,
said position distortion sub-model estimating unit collects the co-occurrence count that the position in the third language sentence having a length of l is corresponding to the j th position in the first language sentence having a length of m, using formula C(j,i,l,m)=Σ k,m′ p EJ (k|i,l,m′)*p CE (j|k,m′,m), and calculates the position distortion probability that the i th position in the third language sentence having a length of l is corresponding to the j th position in the first language sentence having a length of m, using formula
p
CJ
(
j
|
i
,
l
,
m
)
=
C
(
j
,
i
,
l
,
m
)
∑
j
′
C
(
j
′
,
i
,
l
,
m
)
.
7. The apparatus for training bilingual word alignment model according to claim 5 ,
wherein p EJ (w e |w j ) represents the translation probability from the third language word w j to the second language word w e , p CE (φ i |w e ) represents the probability that the second language word w e is corresponding to φ i words of the first language, C(φ i ,w j ) and p(φ i |w j ) respectively represent the co-occurrence count and probability that the third language word w j is corresponding to φ i words of the first language,
said word fertility sub-model estimating unit collects the co-occurrence count that the third language word w j is corresponding to φ i words of the first language, using formula
C
(
φ
i
,
w
j
)
=
∑
w
e
p
EJ
(
w
e
|
w
j
)
*
p
CE
(
φ
i
|
w
e
)
,
and calculates the probability that the third language word w j is corresponding to φ i words of the first language, using formula
p
(
φ
i
|
w
j
)
=
C
(
φ
i
,
w
j
)
∑
φ
i
′
C
(
φ
i
′
,
w
j
)
.
8. An apparatus for bilingual word alignment comprising:
model obtaining unit configured to obtain a bilingual word alignment model for a first language and a third language based on a the bilingual corpus of the first and second languages and the bilingual corpus of the second and third languages by the apparatus for training bilingual word alignment model according to any one of claim 5 , 6 , or 7 and;
word-alignment unit configured to word-align a bilingual sentence pair of the first and third languages using the bilingual word alignment model for the first and third languages.