System and method for predicting an optimal machine translation system for a user based on an updated user profile
A system and method predict an optimal machine translation system for a first of a set of users. The method includes, for each of the users, providing a respective user profile which includes rankings for at least some machine translation systems from a set of machine translation systems. The user profile of the first user is updated, based on the user profiles of at least a subset of the other users. The updating includes generating at least one missing ranking. An optimal translation system for the first user from the set of machine translation systems is predicted, based on the updated user profile computed for the first user.
1. A method for predicting an optimal machine translation system for a user comprising:
for each of a set of users, providing a respective user profile which includes rankings for at least some machine translation systems from a set of machine translation systems, the set of users including a first user and a plurality of other users, wherein the rankings are pairwise rankings for pairs of machine translation systems;
updating the user profile of the first user based on the user profiles of at least a subset of the other users, the updating including generating at least one missing ranking, the updating including:
for each of a subset of the other users whose user profiles include a pairwise ranking for a selected pair of the machine translation systems, computing a similarity between the first user's user profile and the respective other user's user profile;
identifying, based on the computed similarities, a set of the other users as nearest neighbors to the first user, the identifying comprising applying at least one criterion for inclusion of the other users in the nearest neighbors; and
computing a pairwise ranking for the pair of the machine translation systems as a function of the pairwise rankings of the nearest neighbors for the selected pair of machine translation systems, wherein when a number of the nearest neighbors does not meet a threshold number of nearest neighbors, the pairwise ranking for the pair of the machine translation systems is computed as a function of the pairwise rankings of at least one other of the users in addition to the nearest neighbors;
predicting an optimal machine translation system for the first user from the set of machine translation systems based on the pairwise rankings for the pairs of machine translation systems in the updated user profile computed for the first user for translation of source text in a source language to target text in a target language;
outputting a machine translation of source text in the target language for the first user, based on the prediction; and
wherein at least one of the updating and the predicting of the optimal translation system is performed with a processor.
2. The method of claim 1 , wherein the pairwise ranking for the pair of the machine translation systems is computed as a function of weighted pairwise rankings of the nearest neighbors for the selected pair of machine translation systems, wherein the weights are based on the computed similarities.
3. The method of claim 1 , wherein the providing comprises generating the pairwise rankings from user rankings of machine translations of source text strings performed by a plurality of the machine translation systems.
4. The method of claim 1 , further comprising, for a second selected pair of machine translation systems, repeating the identifying of a set of nearest neighbors and the computing of a function of the pairwise rankings of the nearest neighbors and wherein the predicting of the optimal machine translation system is also based on a pairwise ranking computed for the second selected pair of machine translation systems.
5. The method of claim 1 , wherein the updating comprises decomposing a matrix of the user profiles into a plurality of latent factor matrices and constructing an updated matrix of user profiles based on the latent factor matrices.
6. The method of claim 1 , further comprising translating source text from a source language into target text in a target language with the predicted machine translation system.
7. The method of claim 1 , wherein each machine translation system in the set of machine translation systems is configured for translating source text from a same source language into target text in a same target language as the other machine translation systems in the set of machine translation systems.
8. The method of claim 1 , further comprising outputting information based on the predicted optimal machine translation system.
9. A computer program product comprising a non-transitory recording medium storing instructions, which when executed on a computer, causes the computer to perform the method of claim 1 .
10. A system comprising memory which stores instructions for performing the method of claim 1 and a processor in communication with the memory for executing the instructions.
11. A method for predicting an optimal machine translation system for a user comprising:
for each of a set of users, providing a respective user profile which includes rankings for at least some machine translation systems from a set of machine translation systems, the set of users including a first user and a plurality of other users, the rankings being pairwise rankings for pairs of the machine translation systems in the set of machine translation systems, each user profile comprising a vector of elements, one element for each of the machine translation pairs that the user has evaluated, each element (i, j) of the vector being assigned a value:
p
u
(
i
,
j
)
=
w
u
(
i
,
j
)
-
l
u
(
i
,
j
)
w
u
(
i
,
j
)
+
l
u
(
i
,
j
)
(
1
)
where w u (I,j) and I u (i,j) are the number of wins and losses of system S i vs. system s j , as judged by user u;
updating the user profile of the first user based on the user profiles of at least a subset of the other users, the updating including generating at least one missing ranking;
predicting an optimal translation system for the first user from the set of machine translation systems based on the updated user profile computed for the first user,
wherein at least one of the updating and the predicting of the optimal translation system is performed with a processor.
12. The method of claim 11 , wherein the updating comprises:
for each of a subset of the other users whose user profiles include a pairwise ranking for a selected pair of the machine translation systems, computing a similarity between the first user's user profile and the respective other user's user profile;
identifying, based on the computed similarities, a set of the other users as nearest neighbors to the first user;
computing a pairwise ranking for the pair of the machine translation systems as a function of the pairwise rankings of the nearest neighbors for the selected pair of machine translation systems.
13. The method of claim 12 , wherein the computing of the similarity comprises computing a cosine similarity between the user profiles.
14. The method of claim 12 , wherein the identifying of the set of the other users as nearest neighbors to the first user comprises applying at least one criterion for inclusion of the other users in the nearest neighbors.
15. The method of claim 14 , wherein the at least one criterion includes a criterion selected from:
the other user meeting a threshold on the computed similarity;
the other user having at least a threshold quantity of pairwise rankings for pairs of machine translation systems in common with the first user; and
combinations thereof.
16. The method of claim 14 , wherein when a number of the nearest neighbors does not meet a threshold number of nearest neighbors, the pairwise ranking for the pair of the machine translation systems is computed as a function of the pairwise rankings of at least one other of the users in addition to the nearest neighbors.
17. The method of claim 12 , wherein the updating of the user profile comprises computing a function ƒ for the first user u and a MT system pair (s i , s j ):
f
ctp
(
u
)
(
i
,
j
)
=
(
Σ
u
′
∈
MSU
(
u
)
p
u
′
(
i
,
j
)
*
sim
(
u
,
u
′
)
Σ
u
′
∈
MSU
(
u
)
sim
(
u
,
u
′
)
)
,
or
f
ctp
(
u
)
(
i
,
j
)
=
sign
(
Σ
u
′
∈
MSU
(
u
)
p
u
′
(
i
,
j
)
*
sim
(
u
,
u
′
)
)
,
ƒ ctp (u) (i,j) =sign(Σ u′ϵMSU(u) P u′ (i,j) *sim(u, u′)),
where MSU(u) is the set of nearest neighbors u′ of u;
each P u′ (i,j) is a respective ranking of user u′ ϵ MSU(u) for (s i , s j ); and
sim(u, u′) is the similarity computed between the two users.
18. A system for predicting an optimal machine translation system for a user comprising:
memory which stores a user profile for each of a set of users, each user profile including pairwise rankings for at least some pairs of machine translation systems drawn from a set of associated machine translation systems;
a similarity computation component which computes a similarity between a user profile of a first of the set of users and a user profile of another other of the set of users, for each of a subset of the users whose user profiles include a pairwise ranking for a selected pair of the machine translation systems;
a nearest neighbor identification component which identifies, based on the computed similarities, a set of the other users as nearest neighbors to the first user;
a prediction component which predicts an optimal machine translation system for the first user from the set of machine translation systems, based on a pairwise ranking computed for the selected pair of machine translation systems, the pairwise ranking for the pair of the machine translation systems being computed as a function of the pairwise rankings of the nearest neighbors for the selected pair of machine translation systems, wherein when a number of the nearest neighbors does not meet a threshold number of nearest neighbors, the pairwise ranking for the pair of the machine translation systems is computed as a function of the pairwise rankings of at least one other of the users in addition to the nearest neighbors; and
a processor which implements the similarity computation component, nearest neighbor identification component, and prediction component.
19. The system of claim 18 , further comprising at least one of:
a user profile generator which generates the user profiles, the user profile generator being implemented by a processor; and
a translation generator which generates a translation using the predicted machine translation system, the translation generator being implemented by a processor.
20. A translation method comprising:
providing a set of machine translation systems for a selected language pair;
for each of a set of users, providing a respective user profile which includes pairwise rankings for at least some pairs of the machine translation systems in the set of machine translation systems, the set of users including a first user and a plurality of other users, each user profile comprising a vector of elements, one element for each of the machine translation pairs that the user has evaluated, each element (i, j) of the vector being assigned a value:
p
u
(
i
,
j
)
=
w
u
(
i
,
j
)
-
l
u
(
i
,
j
)
w
u
(
i
,
j
)
+
l
u
(
i
,
j
)
(
1
)
where w u (i,j) and I u (i,j) are the number of wins and losses of system s i vs. system s j , as judged by user u;
for each of at least one pair of the machine translation systems, identifying a set of nearest neighbors to the first user based on a computed similarity between the first user's user profile and a respective other user's user profile;
predicting an optimal machine translation system for the first user from the set of machine translation systems based on the profiles of the nearest neighbors; and
translating source language text into target language text using the optimal translation system,
wherein at least one of the identifying of the set of nearest neighbors, the predicting of the optimal translation system, and the translating of the source language text is performed with a processor.