| CPC G06F 40/47 (2020.01) [G06F 40/51 (2020.01)] | 5 Claims |
|
1. A translation device comprising:
a memory configured to store a plurality of pieces of learning data in which original text for learning of a first language, a normalized sentence for learning that is a grammatically correct conversion of the original text for learning, and a translated sentence for learning that is a translation of the original text for learning into a second language different from the first language are associated with each other; and
processing circuitry configured to
perform normalized sentence learning on the plurality of pieces of learning data by combining the original text for learning and the normalized sentence for learning;
perform translated sentence learning on the plurality of pieces of learning data by combining the original text for learning and the translated sentence for learning; and
generate one normalization/translation model, which is configured to be able to output a normalized sentence of an input sentence of the first language and a translated sentence thereof into the second language on the basis of a result of the normalized sentence learning and the translated sentence learning,
wherein, on at least a part of the learning data, the translated sentence learning is performed after the normalized sentence learning,
the normalized sentence learning and the translated sentence learning are alternately performed,
on each piece of the learning data, the translated sentence learning is performed continuously after the normalized sentence learning,
the normalized sentence learning and the translated sentence learning are repeatedly performed on the plurality of pieces of learning data a plurality of times, and
the processing circuitry is further configured to
derive a loss function related to normalization and a loss function related to the generated translation for a normalization/translation model, and evaluate the normalization/translation model based on a value of each loss function, and
evaluate that the normalization/translation model is in a first state with low prediction accuracy when the normalized sentence learning and the translated sentence learning are repeatedly performed on the plurality of pieces of learning data a plurality of times, and at least one of a value of the loss function related to normalization being larger than a predetermined first threshold value and a value of the loss function related to translation being larger than a predetermined second threshold value is satisfied,
wherein the normalized sentence learning is repeatedly performed on each piece of the learning data independently, which is separate from the learning alternately performed by the normalized sentence learning and the translated sentence learning when the normalization/translation model is evaluated to be in the first state and the value of the loss function related to normalization is larger than the first threshold value, and
the translated sentence learning is repeatedly performed on each piece of the learning data independently, which is separate from the learning alternately performed by the translated sentence learning and the normalized sentence learning when the normalization/translation model is evaluated to be in the first state and the value of the loss function related to translation is larger than the second threshold value.
|