US 8,265,923 B2Granted

Statistical machine translation employing efficient parameter training

AssigneeXerox Corporation

InventorsSamidh Chatterjee, Nicola Cancedda

Life of this patent

Estimating the term…

Abstract

A statistical machine translation (SMT) system employs a conditional translation probability conditioned on the source language content. A model parameters optimization engine is configured to optimize values of parameters of the conditional translation probability using a translation pool comprising candidate aligned translations for source language sentences having reference translations. The model parameters optimization engine adds candidate aligned translations to the translation pool by sampling available candidate aligned translations in accordance with the conditional translation probability.

Claims11

Dependent claims

A method comprising:

translating source language content in a source natural language to a target natural language using statistical machine translation (SMT) employing a conditional translation probability conditioned on the source language content;
and
optimizing values of parameters of the conditional translation probability by an iterative optimization process operating on a translation pool, the optimizing including adding candidate aligned translations to the translation pool by sampling available candidate aligned translations for a source language sentence in accordance with the conditional translation probability;
wherein the conditional translation probability is quantitatively equivalent to:
P
⁡
(
e
,
a
|
f
)
=
exp
⁡
(
∑
k
=
1
K
⁢
λ
k
⁢
h
k
⁡
(
e
,
a
,
f
)
)
∑
(
e
′
,
a
′
)
∈
L
⁢
exp
⁡
(
∑
k
=
1
K
⁢
λ
k
⁢
h
k
⁡
(
e
′
,
a
′
,
f
)
)
where f denotes the source language sentence, L denotes the set of available candidate aligned translations, (e, a) denotes a candidate aligned translation for which the conditional probability is computed, hk(. . .), k=1, . . . , K denotes a set of K feature functions, and λk, k=1, . . . , K denotes the parameters of the conditional translation probability;
wherein the sampling comprises:
selecting a sampled candidate aligned translation by traversing a translation lattice representing the available candidate aligned translations for the source language sentence from its root node to its final node wherein each transition from a current node to a next node is selected based on conditional translation probabilities
σ
⁡
(
e
i
)
=
∑
k
=
1
K
⁢
λ
k
⁢
h
k
⁡
(
e
i
)
of the available edges ei leading away from the current node;
wherein the SMT and the optimizing are implemented by a SMT system embodied by at least one digital processor.

98 more elements

A non-transitory storage medium storing instructions executable on a digital processor to perform a method including (i) translating source language content in a source natural language to a target natural language based on a conditional translation probability conditioned on the source language content and (ii) tuning the conditional translation probability using a translation pool,

wherein the tuning includes:
selecting candidate aligned translations for a source language sentence having a reference translation by sampling available candidate aligned translations for the source language sentence in accordance with the conditional translation probability conditioned on the source language sentence;
wherein the conditional translation probability for a translation (e,a) of source language content f is monotonically increasing with
exp
⁡
(
∑
k
=
1
K
⁢
λ
k
⁢
h
k
⁡
(
e
,
a
,
f
)
)
where hk(. . .), k=1, . . . , K denotes a set of K feature functions and λk, k=1, . . . , K denotes tuning parameters of the conditional translation probability;
wherein the selecting comprises:
constructing a translation lattice representing the available candidate aligned translations for the source language sentence;
and
sampling the available candidate aligned translations by sampling paths through the translation lattice from its root node to its final node wherein each transition from a current node to a next node is selected based on conditional translation probabilities
σ
⁡
(
e
i
)
=
∑
k
=
1
K
⁢
λ
k
⁢
h
k
⁡
(
e
i
)
of the available edges ei leading away from the current node, and
adding the selected candidate aligned translations to the translation pool.

53 more elements

An apparatus comprising:

a statistical machine translation (SMT) system employing a conditional translation probability conditioned on the source language content;
and
a model parameters optimization engine configured to optimize values of parameters of the conditional translation probability using a translation pool comprising candidate aligned translations for source language sentences having reference translations, the model parameters optimization engine adding candidate aligned translations to the translation pool by sampling available candidate aligned translations in accordance with the conditional translation probability;
wherein the conditional translation probability is quantitatively equivalent to:
P
⁡
(
e
,
a
|
f
)
=
exp
⁡
(
∑
k
=
1
K
⁢
λ
k
⁢
h
k
⁡
(
e
,
a
,
f
)
)
∑
(
e
′
,
a
′
)
∈
L
⁢
exp
⁡
(
∑
k
=
1
K
⁢
λ
k
⁢
h
k
⁡
(
e
′
,
a
′
,
f
)
)
where f denotes the source language sentence, L denotes the set of available candidate aligned translations, (e, a) denotes a candidate aligned translation for which the conditional probability is computed, hk(. . .), k=1, . . . , K denotes a set of K feature functions, and λk, k=1, . . . , K denotes the parameters of the conditional translation probability;
wherein the sampling comprises:
selecting a sampled candidate aligned translation by traversing a translation lattice representing the available candidate aligned translations for the source language sentence from its root node to its final node wherein each transition from a current node to a next node is selected based on conditional translation probabilities
σ
⁡
(
e
i
)
=
∑
k
=
1
K
⁢
λ
k
⁢
h
k
⁡
(
e
i
)
of the available edges ei leading away from the current node;
wherein the SMT system and the model parameters optimization engine are embodied by one or more digital processors.

98 more elements

Dependencies are inferred from the claim text.

patentgrep

Loading patent…

Abstract