CN108920472A - A kind of emerging system and method for the machine translation system based on deep learning - Google Patents
A kind of emerging system and method for the machine translation system based on deep learning Download PDFInfo
- Publication number
- CN108920472A CN108920472A CN201810722720.7A CN201810722720A CN108920472A CN 108920472 A CN108920472 A CN 108920472A CN 201810722720 A CN201810722720 A CN 201810722720A CN 108920472 A CN108920472 A CN 108920472A
- Authority
- CN
- China
- Prior art keywords
- translation
- module
- machine translation
- word
- vector
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/40—Processing or translation of natural language
- G06F40/58—Use of machine translation, e.g. for multi-lingual retrieval, for server-side translation for client devices or for real-time translation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biomedical Technology (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Machine Translation (AREA)
Abstract
The present invention relates to the emerging systems and method of a kind of machine translation system based on deep learning, belong to vocabulary integration technology field.The emerging system includes input module, coding module, decoder module and output module；The fusion method includes input step, coding step, decoding step and output step；The emerging system and method have the characteristics that improve machine translation system fusion performance indicator.
Description
Technical field
The present invention relates to the emerging systems and method of a kind of machine translation system based on deep learning, belong to vocabulary fusion
Technical field.
Background technique
Machine translation system fusion is that the translation of different machine translation systems is higher from one quality of Combination nova generation
A kind of method of translation.Vocabulary grade system globe area therein is that the most popular now method and performance are very good.In order into one
It walks lifting system and merges performance, deep neural network model is introduced in machine translation translation control fusion.However nerve net
Network model usually requires a large amount of data and is trained, and could be more than classical Fusion Model in performance.It is turned in actual machine
It in the task of translation fusion, is influenced by training data scale, directly is usually unable to be managed using the training of small-scale data
The performance thought is unable to satisfy user's real time translation demand.
Summary of the invention
The present invention is bad in order to solve in existing lexical system integration technology vocabulary fusion performance, is unable to satisfy for real-time
The problem of translation demand, proposes the emerging system and method for a kind of machine translation system based on deep learning.
A kind of emerging system of the machine translation system based on deep learning, the technical solution taken are as follows：
The emerging system of the machine translation system includes：
For inputting the input module of the n machine translation translation of 1 source language sentence and the source language sentence,
In, n is the integer greater than 2；
For the source language sentence and n machine translation translation to be encoded into the coding module of vector；
Vector for exporting the coding module is merged the fusing and decoding module for being decoded into sentence again,
In, the sentence being decoded into again that merges is translation after merging；
For carrying out the model parameter training module of parameter learning to the fusing and decoding module；The model training module
Model foundation mould including translation model of the building one from original language to object language on the basis of original encoding-decoded model
Block；
For exporting the output module of translation after the fusion.
Further, the coding module includes：
For the source language sentence to be encoded into the source of the sequence of a vector using bidirectional circulating neural network structure
Language sentence encoding submodule；
For the attention mechanism module by introducing source language sentence corresponding to the machine translation translation to described
The machine translation text encoding submodule that machine translation translation is encoded；
Include in the machine translation text encoding submodule：
For inputting the input module of machine translation sentence and original language sequence vector；,
For exporting the output module of the sequence vector of machine translation；
For utilizing condition door recirculating network unit (cGRU：Conditional Gated Recurrent Unit) it calculates
The vector calculation module of each vector in the sequence vector.Attention mechanism (attention mechanism) is turned in machine
The essence translated is that the word alignment relationship between source language and the target language is set up in the distribution of weight.
Wherein, the attention mechanism module of the source language sentence includes：
Coding input module (the h of source language sentencei)；
The weight calculation module of the coding of source language sentence (calculates each αi,j)；
Attention vector generation module (cj)；
Further, the bidirectional circulating neural network structure uses GRU (Gated Recurrent Unit) structure.
Further, the fusing and decoding module includes the cGRUmutil-att (Conditional of attention more than one
Gated Recurrent Unit with multi-attention) unit, wherein
CGRU unit is cascade two GRU units.And cGRUmutil-att unit is permitted on the basis of cGRU unit
Many notice that force vector is connected together as the input of unit.Further, the search space of the fusing and decoding module
Restrictive condition it is as follows：
The vocabulary of final translation is the union for merging translation candidate vocabulary, that is, the word for decoding generation must be in fusion translation
The word occurred in candidate；
The original word order that fusion translation is kept when decoding requires successively to select the word in translation candidate from left to right
The candidate of word as final translation；
The word that some word is aligned if being used with it cannot be used again, i.e., the word of all alignment can only be most
Occur in whole translation candidate primary.
A method of the machine translation system fusion based on deep learning, the technical solution taken are as follows：
The fusion method includes：
For inputting the input step of the n machine translation translation of 1 source language sentence and the source language sentence,
In, n is the integer greater than 2；
For the source language sentence and n machine translation translation to be encoded into the coding step of vector；
Vector for exporting the coding step is merged the fusing and decoding step for being decoded into sentence again,
In, the sentence being decoded into again that merges is translation after merging；
For carrying out the model parameter training step of parameter learning to the fusing and decoding step；The model training step
Model foundation step including translation model of the building one from original language to object language on the basis of original encoding-decoded model
Suddenly；
For exporting the output step of translation after the fusion.
Further, the coding step includes：
For the source language sentence to be encoded into the source of the sequence of a vector using bidirectional circulating neural network structure
Language sentence coding sub-step；
For the attention mechanism step by introducing source language sentence corresponding to the machine translation translation to described
The machine translation text coding sub-step that machine translation translation is encoded；
Include in the machine translation text coding sub-step：
For inputting the input step of machine translation sentence and original language sequence vector；,
For exporting the output step of the sequence vector of machine translation；
For utilizing condition door recirculating network unit (cGRU：Conditional Gated Recurrent Unit) it calculates
The vector calculation step of each vector in the original language sequence vector.
Wherein, the attention mechanism step of the source language sentence includes：
Coding input step (the h of source language sentencei)；
The weight calculation step of the coding of source language sentence (calculates each αi,j)；
Pay attention to force vector generation step (cj)；Further, the bidirectional circulating neural network structure uses GRU structure.
Further, the fusing and decoding step includes the cGRUmutil-att unit of attention more than one
Further, the restrictive condition of the search space of the fusing and decoding step is as follows：
The vocabulary of final translation is the union for merging translation candidate vocabulary, that is, the word for decoding generation must be in fusion translation
The word occurred in candidate；
The original word order that fusion translation is kept when decoding requires successively to select the word in translation candidate from left to right
The candidate of word as final translation；
The word that some word is aligned if being used with it cannot be used again, i.e., the word of all alignment can only be most
Occur in whole translation candidate primary.
Beneficial effect of the present invention：
The emerging system and method for a kind of machine translation system based on deep learning proposed by the present invention, by using double
To Recognition with Recurrent Neural Network structure, and the design of structure and processing method to coding module and fusing and decoding module, make machine
Translation system fusion needs not move through a large amount of data training, and can make its BLEU-4 score performance is more than classical Fusion Model,
And then it obtains machine translation system fusion in performance and is greatly promoted.Also, since the system and method needs on a small quantity
It is that it small-scale data training pattern can be used directly in practical applications can be obtained higher word that data, which can be trained,
It converges and merges performance, make machine translation system fusion that can carry out docking collocation use, Jin Eryou with other small-scale training patterns
Effect improves the practicability of machine translation system fusion.
The emerging system and method for a kind of machine translation system based on deep learning proposed by the present invention are in practical application
In without obtaining a large amount of machine translation to the system for participating in fusion in advance vocabulary fusion can be realized, can be effectively reduced acquisition
The time that translation expends is translated, the efficiency of vocabulary fusion is largely improved, can effectively meet client's real time translation
Demand.
Detailed description of the invention
Fig. 1 is the system structure diagram of the system of the machine translation system fusion of the present invention based on deep learning.
Fig. 2 is the structural schematic diagram of GRU structure of the present invention.
Fig. 3 is the structural schematic diagram of attention mechanism module structure of the present invention.
Fig. 4 is the structural schematic diagram of more attention mechanism module structures of the present invention.
Fig. 5 is the structural schematic diagram of model training module of the present invention.
Specific embodiment
The present invention will be further described combined with specific embodiments below, but the present invention should not be limited by the examples.
Embodiment 1：
A kind of emerging system of the machine translation system based on deep learning, using coding-classical in deep learning
Decoded model frame, as shown in Figure 1, the system of machine translation system fusion includes：
For inputting the input module of the n machine translation translation of 1 source language sentence and the source language sentence,
In, n is the integer greater than 2；
For the source language sentence and n machine translation translation to be encoded into the coding module of vector；
Vector for exporting the coding module is merged the fusing and decoding module for being decoded into sentence again,
In, the sentence being decoded into again that merges is translation after merging；
For carrying out the model parameter training module of parameter learning to the fusing and decoding module；The model training module
Model building module including translation model of the building one from original language to object language on the basis of master mould；
For exporting the output module of translation after the fusion.
In Fig. 1, S is source language sentence, T(1),T(2),…T(n)It is the n machine translation translation of source language sentence S, H respectively
It is the coding of source language sentence, Q(1),Q(2),…Q(n)It is the coding of n machine translation translation respectively, E is translating for final output
Text.System gives a source language sentence S, and corresponding n machine translation result T (1), T (2), T (n), fusion
An optimal translation E is found in model trial in translation search space.
Wherein, the coding module includes：
For the source language sentence to be encoded into the source of the sequence of a vector using bidirectional circulating neural network structure
Language sentence encoding submodule；The bidirectional circulating neural network structure uses GRU structure, and specific structure is as shown in Figure 2.Scheming
In 2, h is the state of GRU unit, and h (underscore) is the transitory state of GRU unit, and z is resetting door, and r is resetting door.Substantially
On, the two gate vectors determine which information finally can be as the output of gating cycle unit.The two door control mechanisms
It is characterized in that, they can save the information in long-term sequence, and will not remove or because with prediction not phase at any time
It closes and removes.
For the attention mechanism module by introducing source language sentence corresponding to the machine translation translation to described
The machine translation text encoding submodule that machine translation translation is encoded；Wherein, attention mechanism (attention
It mechanism be) in the essence in machine translation is that the word alignment between source language and the target language is set up in the distribution of weight
Relationship.Due to the usual quality of machine translation and imperfect, the information for being included is not fully accurate, and therefore, the present embodiment introduces
The attention mechanism module of source language sentence can supplement the information lost in machine translation translation with mistake.
Include in the machine translation text encoding submodule：
For inputting the input module of machine translation sentence and original language sequence vector；
For exporting the output module of the sequence vector of machine translation；
For utilizing condition door recirculating network unit (cGRU：Conditional Gated Recurrent Unit) it calculates
The vector calculation module of each vector in the sequence vector.
Wherein, due to machine translation system generate translation it is general and imperfect, expressed by information it is not accurate enough, because
This such as encodes machine translation still according to the above-mentioned method encoded to original language, then in an encoding process
It can include certain noise.And the expressed information in source language sentence is typically considered accurately, therefore to machine
During translation is encoded, the attention of the source language sentence corresponding to it is introduced, it in this way can be more accurately to machine
Translation is encoded, and coding accuracy is effectively improved.The attention mechanism module structure is as shown in figure 3, the source language sentence
Son attention mechanism module include：
Coding input module (the h of source language sentencei)；
The weight calculation module of the coding of source language sentence (calculates each αi,j)；
Attention vector generation module (cj)；
The fusing and decoding module uses a kind of cGRUmutil-att unit of more attentions, the cGRUmutil-att
Unit, which allows to input multiple context vectors and these vectors link together, constitutes a new context vector.Input is
The sequence vector of the sequence vector of source language sentence and its multiple machine translations, final output are fused translations, specific
Structure it is as shown in Figure 4.CGRU unit is cascade two GRU units.And cGRUmutil-att unit is in cGRU unit
On the basis of allow multiple attention force vectors to be connected together as the inputs of unit.
Advantage using the cGRUmutil-att unit of more attentions is, can be by can letter in machine translation translation
Breath is united, and influences the generation of final translation, is turned over so that the translation ultimately generated can combine a machine
Information in translation.CGRUmutil-att can effectively calculate in machine translation the position of each word and final simultaneously
The alignment relation of the position of the word of translation.
The vocabulary in fixed size theoretically is used when decoding based on softmax, that is to say, that is allowed to generate to participate in and be melted
Close the word except translation vocabulary.And in practice, using participation when classical more translations fusion task decodes under normal circumstances
Word in the translation of fusion is as dynamic vocabulary.Therefore, it is used in the present embodiment and decoded search space is limited.
The restrictive condition of the search space of the fusing and decoding module is as follows：
The vocabulary of final translation is the union for merging translation candidate vocabulary, that is, the word for decoding generation must be in fusion translation
The word occurred in candidate；
The original word order that fusion translation is kept when decoding requires successively to select the word in translation candidate from left to right
The candidate of word as final translation；
The word that some word is aligned if being used with it cannot be used again, i.e., the word of all alignment can only be most
Occur in whole translation candidate primary.
The advantage of above-mentioned restrictive condition is on the one hand to can solve the unregistered word problem in decoding, because even translation
There are unregistered words, can also find corresponding word in the former candidate translation for participating in fusion.On the other hand, nerve can be alleviated
The problem of excessive translation and translation deficiency in network decoding, effectively improves the accuracy and translation efficiency of translation.
A kind of emerging system of the machine translation system based on deep learning proposed in the present embodiment inputs machine translation
The sequence vector of sentence and its original language exports as the sequence vector of machine translation.Utilize the attention of source language sentence
(Attention) mechanism solves original language Dependence Problem, and with condition gating cycle network unit (Conditional
Gated Recurrent Unit, cGRU) carry out each vector in the sequence of calculation, influence of the noise to coding is effectively eliminated, is had
Effect improves coding accuracy.A kind of emerging system for machine translation system based on deep learning that the present embodiment proposes solves
The problem of unregistered word in decoding influences translation accuracy, effectively improves the accuracy of translation.Alleviate neural network solution again
The problem of excessive translation and translation deficiency in code, further improves the accuracy and translation efficiency of translation process, effectively mentions
The performance of high emerging system.
Embodiment 2
A kind of fusion method of the machine translation system based on deep learning, the technical solution taken are as follows：
The fusion method includes：
For inputting the input step of the n machine translation translation of 1 source language sentence and the source language sentence,
In, n is the integer greater than 2；
For the source language sentence and n machine translation translation to be encoded into the coding step of vector；
Vector for exporting the coding step is merged the fusing and decoding step for being decoded into sentence again,
In, the sentence being decoded into again that merges is translation after merging；
For carrying out the model parameter training step of parameter learning to the fusing and decoding step；The model training step
Model foundation step including translation model of the building one from original language to object language on the basis of master mould；
For exporting the output step of translation after the fusion.
The coding step includes：
For the source language sentence to be encoded into the source of the sequence of a vector using bidirectional circulating neural network structure
Language sentence coding sub-step；The bidirectional circulating neural network structure uses GRU structure.
For the attention mechanism step by introducing source language sentence corresponding to the machine translation translation to described
The machine translation text coding sub-step that machine translation translation is encoded；
Include in the machine translation text coding sub-step：
For inputting the input step of machine translation sentence and original language sequence vector；
For exporting the output step of the sequence vector of machine translation；
For utilizing condition door recirculating network unit (cGRU：Conditional Gated Recurrent Unit) it calculates
The vector calculation step of each vector in the sequence vector.
Wherein, the attention mechanism step of the source language sentence includes：
Coding input step (the h of source language sentencei)；
The weight calculation step of the coding of source language sentence (calculates each αi,j)；
Pay attention to force vector generation step (cj)；
The restrictive condition of the search space of the fusing and decoding step is as follows：
The vocabulary of final translation is the union for merging translation candidate vocabulary, that is, the word for decoding generation must be in fusion translation
The word occurred in candidate；
The original word order that fusion translation is kept when decoding requires successively to select the word in translation candidate from left to right
The candidate of word as final translation；
The word that some word is aligned if being used with it cannot be used again, i.e., the word of all alignment can only be most
Occur in whole translation candidate primary.
The experimental verification process of the emerging system of machine translation system of the present invention based on deep learning and method is such as
Under：
For the validity of verification method, the selection of our data needs while meeting following condition：1) first
The data that stage uses are the disclosed bilingual datas that can be obtained, and scale is large enough to train neural network translation mould
Type, and there is good performance；2) second stage uses disclosed training data, and scale is smaller, because in practical application
Scene can not obtain large-scale machines translation result in real time.3) training data of second stage and final test collection will include source language
Speech, the translation of reference translation and multiple black box machine translation systems.Therefore, the training set of model first stage we use
For the Czech of the Europarl_v7 of WMT publication to the bilingual data collection conduct of English, which includes about 645K bilingual sentences
Machine translation translation Fusion training collection right, that the training set of model second stage is provided using WMT2011 translation fusion task,
The translation knot of 8 black box machine translation systems of 1003 sentence pair of bilingual data and each sentence pair including Czech to English
Fruit.The test set that test set is provided using WMT2011 translation fusion task comprising the bilingual data 2000 of Czech to English
Sentence pair and each sentence pair merge the training set of task corresponding 8 black box machine translation system with front WMT2011 translation
The translation result of system, without loss of generality, we have randomly selected the translation result of 4 translation systems, the translation of this four systems
BLEU-4 score, as shown in table 1.
The translation BLEU score of each machine translation system of table 1
| System identifier | BLEU-4 |
| System A | 19.65 |
| System B | 21.58 |
| System C | 22.00 |
| System D | 23.53 |
Specific experiment result is as follows：
We have selected the classical system globe area model (MEMT) based on ballot as our baseline model, should
Method repeatedly does well in performance in multiple language tasks of the system globe area of WMT.The BLEU of each system globe area method
Score is as shown in table 2
The BLEU score of each system globe area method of table 2
| Fusion method title | BLEU-4 |
| System top performance | 23.53 |
| Baseline emerging system MEMT performance | 24.07 |
| Context of methods： | 24.95 |
| Context of methods (does not use decoding constraint) | 24.74 |
| Context of methods (without using original language, does not use decoding constraint) | 24.32 |
It can be concluded that, the performance of system globe area model neural network based is significantly better than classical from experimental result
The performance of baseline Fusion Model MEMT, reason essentially consist in the method based on depth integration can more effectively retain it is to be fused
The information of machine translation, these information influence each other in neural network, can more effectively merge to translation.It merges simultaneously
Method after original language information relative to no fusion original language information improves 0.42 BLEU score.This shows source language
Speech still is able to bring more effective informations in last fusing and decoding, to help the decoding merged.Decoding constraint is used
Model can be made further to improve 0.21 BLEU score.This shows the feature in classical model, and still some are largely effective
Information, do not excavated by depth integration model, thus could after fusion further lift scheme performance.
Although the present invention has been disclosed in the preferred embodiment as above, it is not intended to limit the invention, any to be familiar with this
The people of technology can do various changes and modification, therefore protection of the invention without departing from the spirit and scope of the present invention
Range should subject to the definition of the claims.
Claims (10)
1. a kind of emerging system of the machine translation system based on deep learning, which is characterized in that the emerging system includes：
For inputting the input module of the n machine translation translation of 1 source language sentence and the source language sentence, wherein n is
Integer greater than 2；
For the source language sentence and n machine translation translation to be encoded into the coding module of vector；
Vector for exporting the coding module is merged the fusing and decoding module for being decoded into sentence again, wherein institute
Stating the sentence that fusion is decoded into again is translation after merging；
For carrying out the model parameter training module of parameter learning to the fusing and decoding module；The model training module includes
The model building module of translation model of the building one from original language to object language on the basis of original encoding-decoded model；
For exporting the output module of translation after the fusion.
2. the emerging system of the machine translation system based on deep learning according to claim 1, which is characterized in that the volume
Code module include：
For the source language sentence to be encoded into the original language of the sequence of a vector using bidirectional circulating neural network structure
Sentence encoding submodule；
For the attention mechanism module by introducing source language sentence corresponding to the machine translation translation to the machine
The machine translation text encoding submodule that translation translation is encoded；
Wherein：
Include in the machine translation text encoding submodule：
For inputting the input module of machine translation sentence and original language sequence vector；
For exporting the output module of the sequence vector of machine translation；
For calculating the vector calculation module of each vector in the sequence vector using condition door recirculating network unit.
The attention mechanism module of the source language sentence includes：
The coding input module of source language sentence；
The weight calculation module of the coding of source language sentence；
Attention vector generation module.
3. the emerging system of the machine translation system based on deep learning according to claim 2, which is characterized in that described double
GRU structure is used to Recognition with Recurrent Neural Network structure.
4. the emerging system of the machine translation system based on deep learning according to claim 1, which is characterized in that described to melt
Close the cGRUmutil-att unit that decoder module includes attention more than one.
5. the emerging system of the machine translation system based on deep learning according to claim 1, which is characterized in that described to melt
The restrictive condition for closing the search space of decoder module is as follows：
The vocabulary of final translation is the union for merging translation candidate vocabulary, that is, it must be candidate in fusion translation for decoding the word of generation
The word of middle appearance；
The original word order that fusion translation is kept when decoding, that is, require successively to select from left to right word in translation candidate as
The candidate of the word of final translation；
The word that some word is aligned if being used with it cannot be used again, i.e., the word of all alignment can only translated finally
Occur in literary candidate primary.
6. a kind of fusion method of the machine translation system based on deep learning, which is characterized in that the method includes：
For inputting the input step of the n machine translation translation of 1 source language sentence and the source language sentence, wherein n is
Integer greater than 2；
For the source language sentence and n machine translation translation to be encoded into the coding step of vector；
Vector for exporting the coding step is merged the fusing and decoding step for being decoded into sentence again, wherein institute
Stating the sentence that fusion is decoded into again is translation after merging；
For carrying out the model parameter training step of parameter learning to the fusing and decoding step；The model training step includes
The model foundation step of translation model of the building one from original language to object language on the basis of master mould；
For exporting the output step of translation after the fusion.
7. the fusion method of machine translation system according to claim 1, which is characterized in that the coding step includes：
For the source language sentence to be encoded into the original language of the sequence of a vector using bidirectional circulating neural network structure
Sentence coding sub-step；
For the attention mechanism step by introducing source language sentence corresponding to the machine translation translation to the machine
The machine translation text coding sub-step that translation translation is encoded；
Wherein：
Include in the machine translation text coding sub-step：
For inputting the input step of machine translation sentence and original language sequence vector；
For exporting the output step of the sequence vector of machine translation；
For calculating the vector calculation step of each vector in the sequence vector using condition door recirculating network unit.
The attention mechanism step of the source language sentence includes：
The coding input step of source language sentence；
The weight calculation step of the coding of source language sentence；
Pay attention to force vector generation step.
8. the fusion method of machine translation system according to claim 2, which is characterized in that the bidirectional circulating neural network
Structure uses GRU structure.
9. the fusion method of machine translation system according to claim 1, which is characterized in that the fusing and decoding step includes
The cGRUmutil-att unit of attention more than one.
10. the fusion method of machine translation system according to claim 1, which is characterized in that the fusing and decoding step
The restrictive condition of search space is as follows：
The vocabulary of final translation is the union for merging translation candidate vocabulary, that is, it must be candidate in fusion translation for decoding the word of generation
The word of middle appearance；
The original word order that fusion translation is kept when decoding, that is, require successively to select from left to right word in translation candidate as
The candidate of the word of final translation；
The word that some word is aligned if being used with it cannot be used again, i.e., the word of all alignment can only translated finally
Occur in literary candidate primary.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810722720.7A CN108920472B (en) | 2018-07-04 | 2018-07-04 | Fusion system and method of machine translation system based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810722720.7A CN108920472B (en) | 2018-07-04 | 2018-07-04 | Fusion system and method of machine translation system based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108920472A true CN108920472A (en) | 2018-11-30 |
| CN108920472B CN108920472B (en) | 2020-01-10 |
Family
ID=64425489
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810722720.7A Active CN108920472B (en) | 2018-07-04 | 2018-07-04 | Fusion system and method of machine translation system based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108920472B (en) |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109740168A (en) * | 2019-01-09 | 2019-05-10 | 北京邮电大学 | A kind of classic of TCM ancient Chinese prose interpretation method based on knowledge of TCM map and attention mechanism |
| CN110046359A (en) * | 2019-04-16 | 2019-07-23 | 苏州大学 | Neural machine translation method based on sample guidance |
| CN110175338A (en) * | 2019-05-31 | 2019-08-27 | 北京金山数字娱乐科技有限公司 | A kind of data processing method and device |
| CN110263348A (en) * | 2019-03-06 | 2019-09-20 | 腾讯科技（深圳）有限公司 | Interpretation method, device, computer equipment and storage medium |
| CN110263353A (en) * | 2019-06-25 | 2019-09-20 | 北京金山数字娱乐科技有限公司 | A kind of machine translation method and device |
| CN110472253A (en) * | 2019-08-15 | 2019-11-19 | 哈尔滨工业大学 | A kind of Sentence-level mechanical translation quality estimation model training method based on combination grain |
| CN111274827A (en) * | 2020-01-20 | 2020-06-12 | 南京新一代人工智能研究院有限公司 | Suffix translation method based on multi-target learning of word bag |
| CN111652004A (en) * | 2020-05-09 | 2020-09-11 | 清华大学 | Fusion method and device for machine translation system |
| CN112183080A (en) * | 2020-10-20 | 2021-01-05 | 新疆大学 | Uyghur-Chinese machine translation system based on word and morpheme mixed model |
| US11947925B2 (en) | 2020-05-21 | 2024-04-02 | International Business Machines Corporation | Adaptive language translation using context features |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8326598B1 (en) * | 2007-03-26 | 2012-12-04 | Google Inc. | Consensus translations from multiple machine translation systems |
| CN103646019A (en) * | 2013-12-31 | 2014-03-19 | 哈尔滨理工大学 | Method and device for fusing multiple machine translation systems |
| US9201871B2 (en) * | 2010-06-11 | 2015-12-01 | Microsoft Technology Licensing, Llc | Joint optimization for machine translation system combination |
| CN103605644B (en) * | 2013-12-02 | 2017-02-01 | 哈尔滨工业大学 | Pivot language translation method and device based on similarity matching |
| CN106407184A (en) * | 2015-07-30 | 2017-02-15 | 阿里巴巴集团控股有限公司 | Decoding method used for statistical machine translation, and statistical machine translation method and apparatus |
| CN104462072B (en) * | 2014-11-21 | 2017-09-26 | 中国科学院自动化研究所 | The input method and device of computer-oriented supplementary translation |
| CN107590135A (en) * | 2016-07-07 | 2018-01-16 | 三星电子株式会社 | Automatic translating method, equipment and system |
-
2018
- 2018-07-04 CN CN201810722720.7A patent/CN108920472B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8326598B1 (en) * | 2007-03-26 | 2012-12-04 | Google Inc. | Consensus translations from multiple machine translation systems |
| US9201871B2 (en) * | 2010-06-11 | 2015-12-01 | Microsoft Technology Licensing, Llc | Joint optimization for machine translation system combination |
| CN103605644B (en) * | 2013-12-02 | 2017-02-01 | 哈尔滨工业大学 | Pivot language translation method and device based on similarity matching |
| CN103646019A (en) * | 2013-12-31 | 2014-03-19 | 哈尔滨理工大学 | Method and device for fusing multiple machine translation systems |
| CN104462072B (en) * | 2014-11-21 | 2017-09-26 | 中国科学院自动化研究所 | The input method and device of computer-oriented supplementary translation |
| CN106407184A (en) * | 2015-07-30 | 2017-02-15 | 阿里巴巴集团控股有限公司 | Decoding method used for statistical machine translation, and statistical machine translation method and apparatus |
| CN107590135A (en) * | 2016-07-07 | 2018-01-16 | 三星电子株式会社 | Automatic translating method, equipment and system |
Non-Patent Citations (1)
| Title |
|---|
| 李茂西，等；: "机器翻译系统融合技术综述", 《中文信息学报》 * |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109740168A (en) * | 2019-01-09 | 2019-05-10 | 北京邮电大学 | A kind of classic of TCM ancient Chinese prose interpretation method based on knowledge of TCM map and attention mechanism |
| CN109740168B (en) * | 2019-01-09 | 2020-10-13 | 北京邮电大学 | Traditional Chinese medicine classical book and ancient sentence translation method based on traditional Chinese medicine knowledge graph and attention mechanism |
| CN110263348A (en) * | 2019-03-06 | 2019-09-20 | 腾讯科技（深圳）有限公司 | Interpretation method, device, computer equipment and storage medium |
| CN110046359A (en) * | 2019-04-16 | 2019-07-23 | 苏州大学 | Neural machine translation method based on sample guidance |
| CN110175338A (en) * | 2019-05-31 | 2019-08-27 | 北京金山数字娱乐科技有限公司 | A kind of data processing method and device |
| CN110175338B (en) * | 2019-05-31 | 2023-09-26 | 北京金山数字娱乐科技有限公司 | Data processing method and device |
| CN110263353A (en) * | 2019-06-25 | 2019-09-20 | 北京金山数字娱乐科技有限公司 | A kind of machine translation method and device |
| CN110263353B (en) * | 2019-06-25 | 2023-10-13 | 北京金山数字娱乐科技有限公司 | Machine translation method and device |
| CN110472253B (en) * | 2019-08-15 | 2022-10-25 | 哈尔滨工业大学 | Sentence-level machine translation quality estimation model training method based on mixed granularity |
| CN110472253A (en) * | 2019-08-15 | 2019-11-19 | 哈尔滨工业大学 | A kind of Sentence-level mechanical translation quality estimation model training method based on combination grain |
| CN111274827A (en) * | 2020-01-20 | 2020-06-12 | 南京新一代人工智能研究院有限公司 | Suffix translation method based on multi-target learning of word bag |
| CN111652004A (en) * | 2020-05-09 | 2020-09-11 | 清华大学 | Fusion method and device for machine translation system |
| US11947925B2 (en) | 2020-05-21 | 2024-04-02 | International Business Machines Corporation | Adaptive language translation using context features |
| CN112183080A (en) * | 2020-10-20 | 2021-01-05 | 新疆大学 | Uyghur-Chinese machine translation system based on word and morpheme mixed model |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108920472B (en) | 2020-01-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108920472A (en) | A kind of emerging system and method for the machine translation system based on deep learning | |
| Zheng et al. | A pre-training based personalized dialogue generation model with persona-sparse data | |
| Chen et al. | Two-stream network for sign language recognition and translation | |
| CN105843801B (en) | The structure system of more translation Parallel Corpus | |
| CN110334361A (en) | A kind of neural machine translation method towards rare foreign languages language | |
| CN107967262A (en) | A kind of neutral net covers Chinese machine translation method | |
| CN108897740A (en) | A kind of illiteracy Chinese machine translation method based on confrontation neural network | |
| CN110688861B (en) | Multi-feature fusion sentence-level translation quality estimation method | |
| CN110309287A (en) | The retrieval type of modeling dialog round information chats dialogue scoring method | |
| CN110851575B (en) | Dialogue generating system and dialogue realizing method | |
| CN108932232A (en) | A kind of illiteracy Chinese inter-translation method based on LSTM neural network | |
| CN112257465B (en) | Multi-mode machine translation data enhancement method based on image description generation | |
| Dutta Chowdhury et al. | Multimodal neural machine translation for low-resource language pairs using synthetic data | |
| Zheng et al. | Enhancing neural sign language translation by highlighting the facial expression information | |
| CN108984539A (en) | The neural machine translation method of translation information based on simulation future time instance | |
| Wang et al. | A text-guided generation and refinement model for image captioning | |
| CN112507733A (en) | Dependency graph network-based Hanyue neural machine translation method | |
| CN114691858B (en) | Improved UNILM digest generation method | |
| CN113657125B (en) | Mongolian non-autoregressive machine translation method based on knowledge graph | |
| Xue et al. | Lipformer: learning to lipread unseen speakers based on visual-landmark transformers | |
| CN116721176B (en) | Text-to-face image generation method and device based on CLIP supervision | |
| CN112926344A (en) | Word vector replacement data enhancement-based machine translation model training method and device, electronic equipment and storage medium | |
| Shi et al. | Adding Visual Information to Improve Multimodal Machine Translation for Low‐Resource Language | |
| Wu et al. | Enhancing pre-trained models with text structure knowledge for question generation | |
| Liu et al. | A novel domain adaption approach for neural machine translation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| CB03 | Change of inventor or designer information | ||
| CB03 | Change of inventor or designer information |
Inventor after: Yang Muyun Inventor after: Zhu Junguo Inventor after: Zhao Tiejun Inventor after: Zhu Conghui Inventor after: Cao Hailong Inventor after: Xu Bing Inventor after: Zheng Dequan Inventor before: Yang Muyun Inventor before: Zhu Junguo Inventor before: Zhao Tiejun Inventor before: Zhu Conghui Inventor before: Cao Hailong Inventor before: Xu Bing Inventor before: Zheng Dequan |
|
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| TR01 | Transfer of patent right | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20210125 Address after: Building 9, accelerator, 14955 Zhongyuan Avenue, Songbei District, Harbin City, Heilongjiang Province Patentee after: INDUSTRIAL TECHNOLOGY Research Institute OF HEILONGJIANG PROVINCE Address before: 150001 No. 92 West straight street, Nangang District, Heilongjiang, Harbin Patentee before: HARBIN INSTITUTE OF TECHNOLOGY |
|
| TR01 | Transfer of patent right | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20230315 Address after: 150027 Room 412, Unit 1, No. 14955, Zhongyuan Avenue, Building 9, Innovation and Entrepreneurship Plaza, Science and Technology Innovation City, Harbin Hi tech Industrial Development Zone, Heilongjiang Province Patentee after: Heilongjiang Industrial Technology Research Institute Asset Management Co.,Ltd. Address before: Building 9, accelerator, 14955 Zhongyuan Avenue, Songbei District, Harbin City, Heilongjiang Province Patentee before: INDUSTRIAL TECHNOLOGY Research Institute OF HEILONGJIANG PROVINCE |