The method and relevant apparatus that a kind of method of translation, target information determine
The method and relevant apparatus that a kind of method of translation, target information determine
 CN 107,368,476 A
 Filed: 07/25/2017
 Published: 11/21/2017
 Est. Priority Date: 07/25/2017
 Status: Active Application
First Claim
1. A kind of 1. method of translation, it is characterised in that methods described is applied to neural network machine translation NMT systems, the sideMethod includes：
 Coded treatment is carried out to pending text message using encoder, to obtain source vector representation sequence, wherein, it is described to treatProcessing text message belongs to first language；
According to source context vector corresponding to first moment of source vector representation retrieval, wherein, in the sourceHereafter vector is used to represent pending source content；
Determine that the first translation is vectorial and/or second turns over according to the source vector representation sequence and the source context vectorVector is translated, wherein, described first translates what is be not translated into the moment of the amount instruction first inherent source vector representation sequenceSource content, second translates in the source being translated into the moment of amount instruction second in the source vector representation sequenceHold, second moment is a moment adjacent before first moment；
It is vectorial to the described first translation and/or second translation is vectorial and the source context vector is entered using decoderRow decoding process, to obtain the target information at the first moment, wherein, the target information belongs to second language.
Chinese PRB Reexamination
Abstract
The invention discloses a kind of method that target information determines, including：Coded treatment is carried out to pending text message, to obtain source vector representation sequence；According to source context vector corresponding to source vector representation the first moment of retrieval, source context vector is used to represent pending source content；First translation vector and/or the second translation vector are determined according to source vector representation sequence and source context vector, first translates the source content not being translated into the first moment of amount instruction inherence source vector representation sequence, and second translates the source content being translated into the second moment of amount instruction inherence source vector representation sequence；Decoding process is carried out to the first translation vector and/or the second translation vector and source context vector, to obtain the target information at the first moment.The present invention also provides the method and target information determining device of a kind of translation.The present invention can reduce the model training difficulty of decoder, improve the translation effect of translation system.

18 Citations
Translation method and device, storage medium and electronic equipment  
Patent #
CN 111,597,829 A
Filed 05/19/2020

Current Assignee

Neural network machine interpretation method, model, electric terminal and storage medium  
Patent #
CN 110,472,255 A
Filed 08/20/2019

Current Assignee

Machine translation method and device  
Patent #
CN 109,858,045 A
Filed 02/01/2019

Current Assignee

A kind of chapter grade text interpretation method and device  
Patent #
CN 110,489,761 A
Filed 05/15/2018

Current Assignee

TEXT TRANSLATION METHOD AND DEVICE, AND STORAGE MEDIUM  
Patent #
WO2020108400A1
Filed 11/22/2019

Current Assignee

INFORMATION PROCESSING METHOD AND APPARATUS, AND STORAGE MEDIUM  
Patent #
WO2020103721A1
Filed 11/11/2019

Current Assignee

A kind of message prompt method, device and terminal device  
Patent #
CN 109,274,814 A
Filed 08/20/2018

Current Assignee

A kind of automatic questionanswering method, device and storage medium  
Patent #
CN 108,363,763 A
Filed 02/05/2018

Current Assignee

Machine translation method and device  
Patent #
CN 109,446,534 A
Filed 09/21/2018

Current Assignee

CHAPTERLEVEL TEXT TRANSLATION METHOD AND DEVICE  
Patent #
WO2019218809A1
Filed 04/10/2019

Current Assignee

INFORMATION TRANSLATION METHOD AND DEVICE, AND STORAGE MEDIUM AND ELECTRONIC DEVICE  
Patent #
WO2019161753A1
Filed 02/15/2019

Current Assignee

MACHINE TRANSLATION METHOD AND DEVICE, AND COMPUTERREADABLE STORAGE MEDIUM  
Patent #
WO2019154210A1
Filed 01/30/2019

Current Assignee

TRANSLATION METHOD, TARGET INFORMATION DETERMINING METHOD AND RELATED DEVICE, AND STORAGE MEDIUM  
Patent #
WO2019019916A1
Filed 07/11/2018

Current Assignee

Translation method and translation device based on neural network model  
Patent #
CN 105,068,998 A
Filed 07/29/2015

Current Assignee

TRANSLATION APPARATUS, LEARNING APPARATUS, TRANSLATION METHOD, AND STORAGE MEDIUM  
Patent #
US 20160179790A1
Filed 05/23/2014

Current Assignee
N/A

A kind of based on charactercoded degree of depth nerve interpretation method and system  
Patent #
CN 106,126,507 A
Filed 06/22/2016

Current Assignee

Short text emotion factor extraction method and device based on deep learning  
Patent #
CN 106,372,058 A
Filed 08/29/2016

Current Assignee

Neural machine translation systems with rare word processing  
Patent #
CN 106,663,092 A
Filed 10/23/2015

Current Assignee

15 Claims

1. A kind of 1. method of translation, it is characterised in that methods described is applied to neural network machine translation NMT systems, the sideMethod includes：

Coded treatment is carried out to pending text message using encoder, to obtain source vector representation sequence, wherein, it is described to treatProcessing text message belongs to first language； According to source context vector corresponding to first moment of source vector representation retrieval, wherein, in the sourceHereafter vector is used to represent pending source content； Determine that the first translation is vectorial and/or second turns over according to the source vector representation sequence and the source context vectorVector is translated, wherein, described first translates what is be not translated into the moment of the amount instruction first inherent source vector representation sequenceSource content, second translates in the source being translated into the moment of amount instruction second in the source vector representation sequenceHold, second moment is a moment adjacent before first moment； It is vectorial to the described first translation and/or second translation is vectorial and the source context vector is entered using decoderRow decoding process, to obtain the target information at the first moment, wherein, the target information belongs to second language.


2. according to the method for claim 1, it is characterised in that described that pending text message is compiled using encoderCode processing, to obtain source vector representation sequence, including：

The pending text message is inputted to the encoder； Coded treatment is carried out to the pending text message using the encoder； The source vector representation sequence is obtained according to the result of coded treatment, wherein, it is each in the source vector representation sequenceIndividual source vector belongs to the first language.


3. method according to claim 1 or 2, it is characterised in that described vectorial to the described first translation using decoderAnd/or second translation is vectorial and the source context vector carries out decoding process, to obtain the target at the first momentInformation, including：

Described first translation is vectorial and/or second translation is vectorial and the source context vector is inputted to the solutionCode device； Using the decoder is vectorial to the described first translation and/or second translation is vectorial and the source context toAmount carries out decoding process； The translation content of the pending text message is obtained according to the result of decoding process, wherein, the translation content is instituteState the target information at the first moment.


4. a kind of method that target information determines, it is characterised in that including：

Coded treatment is carried out to pending text message, to obtain source vector representation sequence； According to source context vector corresponding to first moment of source vector representation retrieval, wherein, in the sourceHereafter vector is used to represent pending source content； Determine that the first translation is vectorial and/or second turns over according to the source vector representation sequence and the source context vectorVector is translated, wherein, described first translates what is be not translated into the moment of the amount instruction first inherent source vector representation sequenceSource content, second translates in the source being translated into the moment of amount instruction second in the source vector representation sequenceHold, second moment is a moment adjacent before first moment； Vectorial to the described first translation and/or described second translation is vectorial and the source context vector is carried out at decodingReason, to obtain the target information at the first moment.


5. according to the method for claim 4, it is characterised in that described according to the source vector representation sequence and describedSource context vector determines the first translation vector, including：

The 3rd translation vector according to corresponding to the second moment described in the source vector representation retrieval； Source context vector described in the described 3rd translation vector sum is handled using default neural network model, to obtainThe first translation vector.


6. according to the method for claim 4, it is characterised in that described according to the source vector representation sequence and describedSource context vector determines that the first translation vector sum second translates vector, including：

The 3rd translation vector according to corresponding to the second moment described in the source vector representation retrieval； Source context vector described in the described 3rd translation vector sum is handled using default neural network model, to obtainThe first translation vector； The position occurred according to the source context vector in the source vector representation sequence, obtain second translationVector, wherein, the second translation vector is used to update the 4th translation vector corresponding to first moment, and the described 4th turns overIt is after being handled using the default neural network model source context vector described in the described second translation vector sum to translate vectorObtain.


7. according to the method for claim 4, it is characterised in that described according to the source vector representation sequence and describedSource context vector determines the second translation vector, including：
The position occurred according to the source context vector in the source vector representation sequence, obtain second translationVector, wherein, the second translation vector is used to generate the 4th translation vector corresponding to first moment, and the described 4th turns overIt is after being handled using the default neural network model source context vector described in the described second translation vector sum to translate vectorObtain.

8. the method according to claim 5 or 6, it is characterised in that it is described using default neural network model to described theSource context vector described in three translation vector sums is handled, to obtain the first translation vector, including：
The source context vector is subtracted from the described 3rd translation vector using gating cycle unit GRU, it is described to obtainFirst translation vector.

9. the method according to claim 5 or 6, it is characterised in that it is described using default neural network model to described theSource context vector described in three translation vector sums is handled, to obtain the first translation vector, including：

Source context vector described in the described 3rd translation vector sum is handled using GRU, to obtain intermediate vector； The intermediate vector is entered into row interpolation with the described 3rd translation vector to merge, to obtain the first translation vector.


10. according to the method for claim 4, it is characterised in that described according to the source vector representation retrieval theSource context vector corresponding to one moment, including：

According to the decoder states at second moment, the second translation vector, the 3rd translation be vectorial and the source toAmount represents the vector of source content in sequence, determines the alignment probability of source content； According to the alignment probability of the source content and the semantic vector of the source content, determine that first moment is correspondingThe source context vector.


11. according to the method for claim 4, it is characterised in that described to translate vectorial and/or described second to described firstTranslate vectorial and described source context vector and carry out decoding process, it is described with before obtaining the target information at the first momentMethod also includes：

According to the decoder states at second moment, the target information at second moment, the source context vector, instituteState that the first translation is vectorial and the second translation vector, determine the decoder states at first moment； Described vectorial to the described first translation and/or described second translation is vectorial and the source context vector decodesProcessing, to obtain the target information at the first moment, including： Decoder states, the source context vector to first moment, first translation are vectorial and/or described theTwo translation vectors carry out decoding process, to obtain the target information at first moment.


12. the method according to claim 10 or 11, it is characterised in that methods described also includes：

Vectorial and described 3rd translation vector is translated according to described first, obtains the first index desired value, wherein, described firstIndex desired value is used to represent consistent disposition semantic between following translation vector change and the target information at first momentCondition； Vectorial and described 4th translation vector is translated according to described second, obtains the second index desired value, wherein, described secondIndex desired value is used to represent to translate consistent disposition semantic between vector change and the target information at first moment in the pastCondition； Training objective is determined according to the first index desired value and the second index desired value, wherein, the training meshMark for building default neural network model.


13. A kind of 13. target information determining device, it is characterised in that including：

Coding module, for carrying out coded treatment to pending text message, to obtain source vector representation sequence； First acquisition module, during for encoding the obtained source vector representation retrieval the first according to the coding moduleSource context vector corresponding to quarter, wherein, the source context vector is used to represent pending source content； First determining module, for the source vector representation sequence and described for encoding to obtain according to the coding moduleThe source context vector that one acquisition module obtains determines the first translation vector and/or the second translation vector, wherein, it is describedFirst translates the source content not being translated into the moment of the amount instruction first inherent source vector representation sequence, the second translationThe source content being translated into the moment of the amount instruction second inherent source vector representation sequence, second moment beAn adjacent moment before first moment； Decoder module, for first determining module is determined it is described first translation it is vectorial and/or described second translate toAmount and the source context vector carry out decoding process, to obtain the target information at the first moment.


14. A kind of 14. target information determining device, it is characterised in that including：
 Memory, processor and bus system；
Wherein, the memory is used for storage program； The processor is used to perform the program in the memory, comprises the following steps： Coded treatment is carried out to pending text message, to obtain source vector representation sequence； According to source context vector corresponding to first moment of source vector representation retrieval, wherein, in the sourceHereafter vector is used to represent pending source content； Determine that the first translation is vectorial and/or second turns over according to the source vector representation sequence and the source context vectorVector is translated, wherein, described first translates what is be not translated into the moment of the amount instruction first inherent source vector representation sequenceSource content, second translates in the source being translated into the moment of amount instruction second in the source vector representation sequenceHold, second moment is a moment adjacent before first moment； Vectorial to the described first translation and/or described second translation is vectorial and the source context vector is carried out at decodingReason, to obtain the target information at the first moment； The bus system is used to connect the memory and the processor, so that the memory and the processorCommunicated.
 Memory, processor and bus system；

15. a kind of computerreadable recording medium, including instruction, when run on a computer so that computer performs such asMethod described in claim 412.
Specification(s)