Parallel processing-based translation method and apparatus

Patent metadata

Field	Value
Publication number	US-10902216-B2
Application number	US-201715450333-A
Country	US
Kind code	B2
Filing date	Mar 6, 2017
Priority date	Aug 10, 2016
Publication date	Jan 26, 2021
Grant date	Jan 26, 2021

Abstract

A translation method and apparatus may respectively perform or include: using one or more processors, plural different translation processes, in parallel, for a source sentence in a first language, including encoding, to generate respective feature vectors, the source sentence in each of two or more translation processes of the plural translation processes or the source sentence and a variation of the source sentence in respective translation processes of the plural translation processes, and decoding each of the respective feature vectors to generate respective plural candidate sentences in a second language; and selecting a final sentence in the second language from the respective plural candidate sentences in the second language.

First claim

What is claimed is:

1. A processor implemented translation method comprising:
performing, using one or more processors, plural different translation processes, in parallel, for a source sentence in a first language, including
generating first respective feature vectors by encoding the source sentence in each of two or more translation processes of the plural different translation processes, and generating second respective feature vectors by encoding the source sentence and a variation of the source sentence in at least one translation process of the plural different translation processes, and
generating respective plural candidate sentences in a second language by decoding each of the first respective feature vectors and the second respective feature vectors; and
selecting a final sentence in the second language from the respective plural candidate sentences in the second language,
wherein the generating of the respective candidate sentences includes generating a preset number of candidate sentences in each translation process using respective m-best beam search algorithms, with the m-best beam search algorithm being a beam search algorithm with less complexity than an n-best beam search algorithm, with n being greater than m.

2. The method of claim 1 ,
wherein the plural different translation processes are respectively implemented through different graphic processor units (GPUs), each of the GPUs including a translation encoder and a translation decoder, and
wherein the source sentence is input to at least two of the GPUs or the source sentence is input to at least one of the GPUs and the variation of the source sentence is input to another at least one of the GPUs, to perform the different translation processes.

3. The method of claim 1 , wherein the plural different translation processes are respectively implemented through the one or more processors, of a translation system, that each include one or more translation encoders of plural encoders of the translation system and one or more translation decoders of plural decoders of the translation system,
where the translation method further includes inputting, in parallel, the source sentence to at least two of the plural encoders.

4. The method of claim 3 , further comprising:
inputting the variation of the source sentence to at least one of the plural encoders, varying at least one of the plural encoders, and/or varying at least one of the plural decoders, for the performing of the plural different translation processes.

5. The method of claim 4 , further comprising:
generating the variation of the source sentence by changing a word order of the source sentence based on information associated with the first language and/or replacing, with a synonym, a word included in the source sentence based on the information associated with the first language; and
inputting the variation of the source sentence to the at least one of the plural encoders.

6. The method of claim 4 , further comprising:
performing the varying of the at least one of the plural encoders, including changing the at least one encoder by respectively applying noise to a corresponding parameter value or a corresponding activation function to be respectively applied to the at least one encoder; or
performing the varying of the at least one of the plural decoders, including changing the at least one decoder by respectively applying noise to a corresponding parameter value or a corresponding activation function to be respectively applied to the at least one decoder.

7. The method of claim 3 , further comprising:
varying or changing an encoder in at least one of the one or more processors and/or varying or changing a decoder in at least one of the one or more processors, for the performing of the plural different translation processes.

8. The method of claim 7 , further comprising at least a correspondingly one of:
performing the varying or changing of the encoder by replacing the encoder with another encoder having been trained by at least one of a different initial training value, training set, or training sequence than an initial training value, training set, or training sequence that was used to train the encoder; and
performing the varying or changing of the decoder by replacing the decoder with another decoder having been trained by at least one of a different initial training value, training set, or training sequence than an initial training value, training set, or training sequence that was used to train the decoder.

9. The method of claim 1 , wherein the selecting of the final sentence in the second language comprises:
calculating respective scores of multiple candidate sentences, of the respective plural candidate sentences, in the second language; and
selecting, as the final sentence in the second language, one of the multiple candidate sentences that has a highest score among the multiple candidate sentences in the second language.

10. The method of claim 9 , wherein the calculating of the respective scores of the multiple candidate sentences in the second language comprises:
calculating respective scores of only a single candidate sentence in the second language from each of the translation processes.

11. The method of claim 1 , wherein the selecting of the final sentence in the second language comprises:
calculating scores corresponding to each of corresponding candidate sentences in the second language in each of the translation processes using respective rescoring models;
recalculating scores corresponding to one or more of the corresponding candidate sentences from each of the translation processes using statistics of scores corresponding to each of the one or more of the corresponding candidate sentences; and
selecting, as the final sentence in the second language, one of the one or more of the corresponding candidate sentences that has a highest recalculated score among the recalculated scores.

12. A non-transitory computer-readable storage medium storing instructions that, when executed by the one or more processors, cause the one or more processors to perform the method of claim 1 .

13. A translation apparatus comprising:
at least one processor, of one or more processors, configured to perform plural different translation processes, in parallel, for a source sentence in a first language through
generation of first respective feature vectors by an encoding of the source sentence in each of two or more translation processes of the plural different translation processes, and generation of second respective feature vectors by an encoding of the source sentence and a variation of the source sentence in at least one translation process of the plural different translation processes, and
generation of respective plural candidate sentences in a second language by a decoding of each of the first respective feature vectors and the second respective feature vectors; and
an output processor, of the one or more processors, configured to select a final sentence in the second language from the respective plural candidate sentences in the second language,
wherein the generating of the respective candidate sentences includes generating a preset number of candidate sentences in each translation process using respective beam search algorithms,
wherein the respective beam search algorithms are m-best beam search algorithms with less complexity than an n-best beam search algorithm, as n being greater than m.

14. The translation apparatus of claim 13 , wherein each of the one or more processors includes at least one encoder to perform a corresponding encoding of the source sentence to generate

Assignees

• Samsung Electronics Co Ltd

Inventors

• Lee Hodong
• Park Youngki
• NA Hwidong
• Mun Minyoung
• SONG Inchul

Classifications

• G06N3/045

  Combinations of networks · CPC title

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title

• G06N3/08

  Learning methods · CPC title

• G06N3/09

  Supervised learning · CPC title