CN108897740A - A kind of illiteracy Chinese machine translation method based on confrontation neural network - Google Patents
A kind of illiteracy Chinese machine translation method based on confrontation neural network Download PDFInfo
- Publication number
- CN108897740A CN108897740A CN201810428132.2A CN201810428132A CN108897740A CN 108897740 A CN108897740 A CN 108897740A CN 201810428132 A CN201810428132 A CN 201810428132A CN 108897740 A CN108897740 A CN 108897740A
- Authority
- CN
- China
- Prior art keywords
- network
- translation
- machine translation
- training
- confrontation
- 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.)
- Pending
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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Machine Translation (AREA)
Abstract
A kind of illiteracy Chinese interpretation method based on confrontation neural network introduces the differentiation network D of one with generation network G confrontation on the basis of the generation network G of former machine translation;It is described differentiate network D mainly to generate network G output realize two-value classification, judge the translation of object language, if it is from training Parallel Corpus, then return value be 1, if from generate network G machine translation as a result, if return value be 0.When the probability distribution of truthful data (can not be difficult to) calculate (for example original language parallel corpora data are less), the training mechanism fought by generator and arbiter can make generator go to approach the probability distribution for being difficult to calculate.
Description
Technical field
The invention belongs to field of computer technology, in particular to a kind of illiteracy Chinese machine translation side based on confrontation neural network
Method.
Background technique
Machine translation (Machine Translation, abbreviation MT) is that one in natural language processing field is earliest
Branch is studied, it is a kind of natural language to be transformed into the another kind nature with identical meaning using computer (machine)
The process of language.Machine translation is a research direction of artificial intelligence, has highly important scientific research value and practical value.
Along with the rapid development of globalization process deepened constantly with internet, machine translation mothod at home and abroad politics, economy, society
Meeting, cultural exchanges etc. play increasingly important role.
Currently, machine translation method neural network based is that machine translation field obtains the optimal method of translation effect.
It mainly uses " coding-decoding " structure, is made of two parts of encoder and decoder, the two is all made of Recognition with Recurrent Neural Network
(Recurrent Neural Network, RNN) and long short-term memory (Long Short-Term Memory, LSTM) network knot
Structure.The process of translation includes:Firstly, the source language sentence of input is converted into a term vector sequence as circulation by encoder
The input of neural network, encoder can export the intensive vector an of regular length, referred to as context vector.Then, decoder
Using context vector as input, a Softmax classifier is combined using another Recognition with Recurrent Neural Network, exports target language
The term vector sequence of speech.Finally, term vector is mapped one by one using dictionary becomes object language word, entire translation process is completed.
However, the bilingual parallel corporas for needing largely manually to mark using existing RNN and LSTM neural network model
Library.But existing realistic problem is, it is the artificial heavy workload for marking parallel corpora, at high cost, and lack the people of high quality
Work marks Parallel Corpus.And these deficiencies will largely influence last translation quality.This is existing nerve net
The bottleneck problem that network Machine Translation Model faces.And especially for some rare foreign languages (such as Mongol) for, these problems are more
It is prominent, can be used for training the Parallel Corpus data of neural network model less, be thus difficult to train and construct high quality,
High performance machine translation system.
Summary of the invention
In order to overcome the disadvantages of the above prior art, the purpose of the present invention is to provide a kind of based on confrontation neural network
Cover Chinese machine translation method.
To achieve the goals above, the technical solution adopted by the present invention is that:
It is a kind of to be drawn on the basis of the generation network G of former machine translation based on the illiteracy Chinese interpretation method for fighting neural network
Enter the differentiation network D of one with generation network G confrontation;It is described to differentiate that network D mainly realizes two-value point to the output for generating network G
Class judges the translation of object language, and if it is the Parallel Corpus from training, then return value is 1, if from net is generated
Network G machine translation as a result, then return value be 0.
The differentiation network D uses multilayer perceptron BP network model, and the method for the two-value classification is:
Using the form of amendment linear unit (RELU):
Wherein, x makes a living into network G to the input signal for differentiating network D;α is adjustable constant, and α answers very little, can such as set
For α=0.001;
Multilayer perceptron feedforward neural network carries out information propagation by following formula:
a(l)=fl(W(l)·a(l-1)+b(l))
Wherein, a(l)Indicate the output of l layers of neuron, flIndicate the activation primitive of l layers of neuron, W(l)Indicate that l-1 layers are arrived l
The weight matrix of layer, b(l)Indicate l-1 layers to l layers of biasing, wherein activation primitive uses sigmoid activation primitive:
When multilayer perceptron BP network model carries out two-value classification, using input layer vector X as the defeated of first layer
Enter a(0)It substitutes into f (Wa+b), calculates output a(l)As entire function output vector Y, the larger dimension institute of numerical value in Y is selected
The classification of representative indicates that translation is derived from Parallel Corpus, is also derived from generation network G as classification results.
The generation network G uses convolutional neural networks (CNN), is made of encoder and decoder two parts;The volume
Code device and decoder are all the depth CNN of multilayer, obtain short distance Dependency Specification using CNN convolution kernel, and by increasing CNN
Depth obtains remote Dependency Specification, and every layer decoder is equipped with an attention mechanism.
Described to differentiate network D by training with the confrontation type for generating network G, synchronous raising generates network G and generates target language
The ability of speech and raising differentiate the ability that network D judges translation source;In confrontation type training process, differentiate network D for sentencing
Disconnected translation out is derived from Parallel Corpus data, or generates the result of network G machine translation;
Differentiate that the process of network D study makes a living into network G and differentiates the competition process between network D, specifically includes:
One is taken from authentic specimen and the sample generated by generation network G at random, allows and differentiates that network D goes to judge whether
It is true;
By the mechanism of Machine Learning of competitive mode, makes to generate network G and differentiate that the performance of network D is constantly promoted;When entire
Network reaches Nash Equilibrium state, i.e., when two network parameters are stablized, training is completed;It is turned at this point, generating the machine that network G generates
It translates as a result, having been able to out-trick and differentiates network D, it is made to think translation from parallel corpora;At this point, generation network G model is
It can be used as the Machine Translation Model of output.
The process of confrontation type training can be regarded as following optimization task:
V (D, G)=EX~Pdata (x)[log D(x)]+EZ~Pz (z)[log(1-D(G(z)))]
The optimization task is about the cost function (Value Function) for differentiating network D and generation network G, wherein x
Indicate the data in Parallel Corpus;Z indicates that input generates the source language data of network;E indicates the expectation of the event;G (z) table
Show and generates the translation data that network G generates;D (x) is to differentiate that network D judges whether x derives from the probability of Parallel Corpus;D(G
(z)) it indicates to differentiate that network D judges to generate the probability whether translation that network G generates derives from Parallel Corpus.In training process
Fixed party updates the parameter of another network, alternating iteration, so that the mistake of other side maximizes, finally, generates network G energy
The distribution of sample data is estimated, model is generated and implicitly defines a probability distribution P (g), data are really distributed as P
(data), there are optimal solutions when P (g)=P (data), that is, reach Nash Equilibrium, generate network G at this time and have restored instruction
Practice the distribution of data, the accuracy rate of discrimination model is equal to 50%.
By that will generate network G and differentiate that network D carries out series connection training and completely generates confrontation network to be formed;It will generate
The input vector of encoder and the output vector of decoder are attached in network G, are passed to as input and are differentiated network D;Together
When, the output result 0 or 1 for differentiating network D is fed back into generation network G.
The Parallel Corpus of the training be by source text and its it is parallel it is corresponding translate Chinese language originally constitute it is bilingual/multi-lingual
Corpus participates in the training for generating network D.
Compared with prior art, the beneficial effects of the invention are as follows:
When the probability distribution of truthful data (can not be difficult to) calculate (for example original language parallel corpora data are less),
Traditional neural network model not can be used directly.But confrontation neural network still can be used, and pass through generator and differentiation
The training mechanism of device confrontation, can make generator go to approach the probability distribution for being difficult to calculate.
Detailed description of the invention
Fig. 1 is the flow chart of the illiteracy Chinese machine translation method the present invention is based on confrontation neural network.
Fig. 2 is convolutional neural networks (CNN) neural network and differentiation network structure.
Fig. 3 is encoder using convolutional neural networks schematic diagram.
Fig. 4 is decoder using convolutional neural networks schematic diagram.
Specific embodiment
The embodiment that the present invention will be described in detail with reference to the accompanying drawings and examples.
As shown in Figure 1, the present invention is a kind of illiteracy Chinese machine translation method based on confrontation neural network, in former machine translation
Generation network G on the basis of, introduce one with generate network G fight differentiation network D;By the side for realizing two-value classification
Method judges the translation of object language, and if it is the Parallel Corpus from training, then return value is 1, if from net is generated
Network G machine translation as a result, then return value be 0.
Differentiate that network D uses multilayer perceptron BP network model, two-value classification method is:
Using the form of amendment linear unit (RELU):
RELU is selected, comparing tanh in gradient decline has faster convergence rate, and computing cost is smaller.
Entire multilayer perceptron feedforward neural network carries out information transmitting by following formula:
z(l)=W(l)·a(l-1)+b(l) (1)
a(l)=fl(z(l)) (2)
(1) and (2) is merged:
a(l)=fl(W(l)·a(l-1)+b(l)) (3)
Wherein, z(l)It indicates the input of l layers of neuron, indicates a(l)Indicate the output of l layers of neuron, flIndicate l layers of nerve
The activation primitive of member, W(l)Indicate l-1 layers to l layers of weight matrix, b(l)Indicate l-1 layers to l layers of biasing.Wherein activation primitive
flIt is all made of sigmoid activation primitive:
When multilayer perceptron BP network model carries out two-value classification, using input layer vector X as the defeated of first layer
Enter a(0)It substitutes into f (w, a, b), calculates output a(l)As entire function output vector, then using the output vector as
The input vector x of softmax classifier is inputted.
Further, the softmax classifier functions form is as follows:
Wherein, i indicates some class categories, (θ1,θ2,…,θk) it is corresponding classifier parameters of all categories, k is classification
Sum.P (y=i | x;θ) indicate that the x of input belongs to the probability of classification i.
Further, for applying in softmax classifier of the invention, i ∈ (1,2), k=2.If i=1 indicates to come
Derived from the classification of Parallel Corpus, i=2 is indicated from the classification for generating network.Its functional form is as follows:
Final output vector Y shape formula isWherein, P (G) indicates translation from the probability for generating network G, P (C)
It indicates that translation derives from the probability of Parallel Corpus, selects in Y that classification representated by the larger dimension of numerical value is as classification results, i.e.,
If P (G) is larger, P (G) is selected, then differentiates that network-feedback value is 0;Otherwise selection P (C), then value of feedback is 1.
It generates network G and uses convolutional neural networks (CNN), be made of encoder and decoder two parts;Encoder, decoding
Device two parts utilize CNN convolution kernel to obtain short distance Dependency Specification, and obtain remote rely on by increasing CNN depth and believe
Breath, therefore encoder and decoder are all the depth CNN of multilayer, every layer decoder is equipped with an attention mechanism.
Network D is differentiated by training with the confrontation type for generating network G, and synchronous raising generates network G and generates object language
Ability and differentiation network D judge the ability in translation source;In confrontation type training process, differentiate network D for judging translation
Parallel Corpus data are derived from, or former machine translation generates the result of network G machine translation;
Differentiate that the process of network D study makes a living into network G and differentiates the competition process between network D;It specifically includes:
One is taken from authentic specimen and the sample generated by generation network at random, allows and differentiates that network D goes to judge whether
It is true;
By the mechanism of Machine Learning of competitive mode, makes to generate network G and differentiate that the performance of network D is constantly promoted;When entire
Network reaches Nash Equilibrium state, i.e., when two network parameters are stablized, training is completed;It is turned at this point, generating the machine that network G generates
It translates as a result, having been able to out-trick and differentiates network D, it is made to think translation from parallel corpora;At this point, generating the model of network G
It can be used as the Machine Translation Model of output.
It generates network G and differentiation network D carries out series connection training and fights network to form complete generate;Network G will be generated
The input vector of middle encoder and the output vector of decoder are attached, and are passed to as input and are differentiated network D;Meanwhile it will
Differentiate that the output result 0 or 1 of network D feeds back to generation network G.
Confrontation type training section training process of the present invention regards following optimization task as:
V (D, G)=EX~Pdata (x)[log(D(x))]+EZ~Pz (z)[log(1-D(G(z)))]
The optimization task can regard the process optimized to two tasks as:
1. the optimization of couple D:
2. the optimization of couple G:
Wherein, x indicates the data in Parallel Corpus;Z indicates that input generates the source language data of network;G (z) indicates life
The translation data generated at network G;D (x) is to differentiate that network D judges whether x derives from the probability of Parallel Corpus;D(G(z))
It indicates to differentiate that network D judges to generate the probability whether translation that network G generates derives from Parallel Corpus.
It can be seen that the optimization for D, if differentiating, the ability of network D is stronger, D (x) Ying Yue great,
D (G (x)) is smaller, and V (D, G) will increase at this time, therefore is optimized for tending to its maximum value maxD for D.
For generating network G, the purpose of the network is to generate the translation for being enough to confuse and differentiating network, that is, is generated more accurate
Translation, therefore D (G (z)) value should increase as far as possible, V (D, G) value will reduce at this time, therefore be optimized for tending to its minimum value for G
minG。
Below with reference to embodiment, invention is further explained.
As shown in Fig. 2, being made of in generating network decoder and encoder two parts:Affiliated encoder uses convolution
Neural network (Convolutional Neural Networks), the encoder is first converted to the source language sentence of input
One term vector sequence carries out characteristic processing by convolutional layer, activation primitive, pond layer and full articulamentum.Then the decoding
Device utilizes another convolutional neural networks, using term vector after the processing of encoder output as input, through convolutional neural networks
Decode term vector sequence, final output translation.Then differentiation network is passed to using the output vector of decoder as input, together
When, the output result for differentiating network (0 or 1) is fed back into living network.
Wherein, convolutional layer carries out feature extraction to term vector;Pond layer carries out Feature Compression and extracts main feature and reduce
Computation complexity;Full articulamentum connects all features.
For example, being translated into object language Mongolian with " tomorrow will rain " for source language sentence.It is first that sentence inputting is raw
At network, encoder is initially entered, encoder uses convolutional neural networks.Encoder encodes sentence, by source language sentence
A term vector sequence is converted to, feature extraction then is carried out to the term vector sequence by convolutional layer, pond layer carries out feature
After feature is attached by compression, full articulamentum, then send decoder to, as shown in Figure 3;Decoder equally uses convolutional Neural
Network exports the term vector sequence of object language by decoder processes, and term vector is then mapped as mesh one by one by dictionary
Poster words are finally completed translation process, outputAs shown in Figure 4;And it will solution
The output vector of code device passes to differentiation network as input, differentiates that 0 or 1 are fed back to generation net by two-value classification by network
Network illustrates to differentiate that network thinks that the translation from Parallel Corpus, has generated the training objective of network if value of feedback is 1
Through reaching;If value of feedback is 0, illustrate to differentiate that network is thought to generate the translation of network training from network is generated to translation
Model is not yet completed, and need to continue to train.
Claims (7)
1. a kind of illiteracy Chinese interpretation method based on confrontation neural network, which is characterized in that in the generation network G of former machine translation
On the basis of, introduce the differentiation network D of one with generation network G confrontation;The output for differentiating network D mainly to network G is generated
It realizes two-value classification, judges the translation of object language, if it is the Parallel Corpus from training, then return value is 1, if coming
Derived from generation network G machine translation as a result, then return value is 0.
2. the illiteracy Chinese machine translation method according to claim 1 based on confrontation neural network, which is characterized in that the differentiation
Network D uses multilayer perceptron BP network model, and the method for the two-value classification is:
Using the form of amendment linear unit (RELU):
Wherein, x makes a living into network G to the input signal for differentiating network D;α is adjustable constant;
Multilayer perceptron feedforward neural network carries out information propagation by following formula:
a(l)=fl(W(l)·a(l-1)+b(l))
Wherein, a(l)Indicate the output of l layers of neuron, flIndicate the activation primitive of l layers of neuron, W(l)Indicate l-1 layers to l layers
Weight matrix, b(l)Indicate l-1 layers to l layers of biasing, wherein activation primitive uses sigmoid activation primitive:
When multilayer perceptron BP network model carries out two-value classification, using input layer vector X as the input a of first layer(0)
It substitutes into f (Wa+b), calculates output a(l)As entire function output vector Y, select in Y representated by the larger dimension of numerical value
Classification indicate that translation is derived from Parallel Corpus, be also derived from generation network G as classification results.
3. the illiteracy Chinese machine translation method according to claim 1 based on confrontation neural network, which is characterized in that the generation
Network G uses convolutional neural networks (CNN), is made of encoder and decoder two parts;The encoder and decoder are all
The depth CNN of multilayer obtains short distance Dependency Specification using CNN convolution kernel, and is obtained at a distance by increasing CNN depth
Dependency Specification, every layer decoder are equipped with an attention mechanism.
4. the illiteracy Chinese machine translation method according to claim 1 based on confrontation neural network, which is characterized in that the differentiation
Network D with the confrontation type for generating network G by training, and synchronous raising generation network G generates the ability of object language and raising is sentenced
Other network D judges the ability in translation source;In confrontation type training process, differentiate network D for judging that translation is derived from
Parallel Corpus data, or generate the result of network G machine translation;
Differentiate that the process of network D study makes a living into network G and differentiates the competition process between network D, specifically includes:
One is taken from authentic specimen and the sample generated by generation network G at random, allows and differentiates that network D goes to judge whether it is
Very;
By the mechanism of Machine Learning of competitive mode, makes to generate network G and differentiate that the performance of network D is constantly promoted;Work as whole network
Reach Nash Equilibrium state, i.e., when two network parameters are stablized, training is completed;At this point, generating the machine translation knot that network G generates
Fruit, having been able to out-trick differentiates network D, it is made to think translation from parallel corpora;At this point, generating network G model can make
For the Machine Translation Model of output.
5. the illiteracy Chinese machine translation method according to claim 4 based on confrontation neural network, which is characterized in that by confrontation type
Trained process regards following optimization task as:
V (D, G)=EX~Pdata (x)[log D(x)]+EZ~Pz (z)[log(1-D(G(z)))]
The optimization task is about the cost function (Value Function) for differentiating network D and generation network G, wherein x expression
Data in Parallel Corpus;Z indicates that input generates the source language data of network;E indicates the expectation of the event;G (z) indicates life
The translation data generated at network G;D (x) is to differentiate that network D judges whether x derives from the probability of Parallel Corpus;D(G(z))
It indicates to differentiate that network D judges to generate the probability whether translation that network G generates derives from Parallel Corpus.It is fixed in training process
One side updates the parameter of another network, alternating iteration, so that the mistake of other side maximizes, finally, generating network G can estimate
The distribution of sample data out generates model and implicitly defines a probability distribution P (g), and data are really distributed as P (data),
There are optimal solutions when P (g)=P (data), that is, reach Nash Equilibrium, generate network G at this time and have restored training data
Distribution, the accuracy rate of discrimination model is equal to 50%.
6. the illiteracy Chinese machine translation method according to claim 4 based on confrontation neural network, which is characterized in that generate network
G and differentiation network D carries out series connection training and fights network to form complete generate;By generate network G in encoder input to
Amount and the output vector of decoder are attached, and are passed to as input and are differentiated network D;Meanwhile the output knot that will differentiate network D
Fruit 0 or 1 feeds back to generation network G.
7. the illiteracy Chinese machine translation method according to claim 1 based on confrontation neural network, which is characterized in that the training
Parallel Corpus be to participate in generating by source text and its parallel corresponding bilingual/multi-lingual corpus translating Chinese language and originally constituting
The training of network D.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810428132.2A CN108897740A (en) | 2018-05-07 | 2018-05-07 | A kind of illiteracy Chinese machine translation method based on confrontation neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810428132.2A CN108897740A (en) | 2018-05-07 | 2018-05-07 | A kind of illiteracy Chinese machine translation method based on confrontation neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108897740A true CN108897740A (en) | 2018-11-27 |
Family
ID=64342627
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810428132.2A Pending CN108897740A (en) | 2018-05-07 | 2018-05-07 | A kind of illiteracy Chinese machine translation method based on confrontation neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108897740A (en) |
Cited By (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109543827A (en) * | 2018-12-02 | 2019-03-29 | 清华大学 | Production fights network equipment and training method |
CN109670036A (en) * | 2018-12-17 | 2019-04-23 | 广州大学 | A kind of news comment automatic generation method and its device |
CN109670178A (en) * | 2018-12-20 | 2019-04-23 | 龙马智芯(珠海横琴)科技有限公司 | Sentence-level bilingual alignment method and device, computer readable storage medium |
CN109887047A (en) * | 2018-12-28 | 2019-06-14 | 浙江工业大学 | A kind of signal-image interpretation method based on production confrontation network |
CN109993678A (en) * | 2019-03-26 | 2019-07-09 | 南京联创北斗技术应用研究院有限公司 | It is a kind of to fight the robust steganography method for generating network based on depth |
CN110288609A (en) * | 2019-05-30 | 2019-09-27 | 南京师范大学 | A kind of multi-modal whole-heartedly dirty image partition method of attention mechanism guidance |
CN110309512A (en) * | 2019-07-05 | 2019-10-08 | 北京邮电大学 | A kind of Chinese grammer error correction method thereof based on generation confrontation network |
CN110334361A (en) * | 2019-07-12 | 2019-10-15 | 电子科技大学 | A kind of neural machine translation method towards rare foreign languages language |
CN110472238A (en) * | 2019-07-25 | 2019-11-19 | 昆明理工大学 | Text snippet method based on level interaction attention |
CN110598221A (en) * | 2019-08-29 | 2019-12-20 | 内蒙古工业大学 | Method for improving translation quality of Mongolian Chinese by constructing Mongolian Chinese parallel corpus by using generated confrontation network |
CN111046178A (en) * | 2019-11-29 | 2020-04-21 | 北京邮电大学 | Text sequence generation method and system |
CN111414770A (en) * | 2020-02-24 | 2020-07-14 | 内蒙古工业大学 | Semi-supervised Mongolian neural machine translation method based on collaborative training |
CN111859995A (en) * | 2020-06-16 | 2020-10-30 | 北京百度网讯科技有限公司 | Training method and device of machine translation model, electronic equipment and storage medium |
CN112100908A (en) * | 2020-08-31 | 2020-12-18 | 西安工程大学 | Garment design method for generating confrontation network based on multi-condition deep convolution |
CN112489158A (en) * | 2021-01-13 | 2021-03-12 | 河北大学 | Enhancement method for low-dose PET image by using cGAN-based adaptive network |
CN113657129A (en) * | 2021-09-06 | 2021-11-16 | 内蒙古工业大学 | Mongolian Chinese neural machine translation method combining generation of confrontation network and near-end optimization strategy |
CN115660037A (en) * | 2022-10-26 | 2023-01-31 | 广东东方思维科技有限公司 | Method for distinguishing through deep neural network model |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107368475A (en) * | 2017-07-18 | 2017-11-21 | 中译语通科技(北京)有限公司 | A kind of machine translation method and system based on generation confrontation neutral net |
CN107464210A (en) * | 2017-07-06 | 2017-12-12 | 浙江工业大学 | A kind of image Style Transfer method based on production confrontation network |
-
2018
- 2018-05-07 CN CN201810428132.2A patent/CN108897740A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107464210A (en) * | 2017-07-06 | 2017-12-12 | 浙江工业大学 | A kind of image Style Transfer method based on production confrontation network |
CN107368475A (en) * | 2017-07-18 | 2017-11-21 | 中译语通科技(北京)有限公司 | A kind of machine translation method and system based on generation confrontation neutral net |
Non-Patent Citations (3)
Title |
---|
LIJUN WU 等: "Adversarial Neural Machine Translation", 《HTTPS://ARXIV.ORG/ABS/1704.06933V3》 * |
洪洋 等: "深度卷积对抗生成网络综述", 《第18届中国系统仿真技术及其应用学术年会论文集》 * |
蒋昂波 等: "ReLU激活函数优化研究", 《传感器与微系统》 * |
Cited By (29)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109543827B (en) * | 2018-12-02 | 2020-12-29 | 清华大学 | Generating type confrontation network device and training method |
CN109543827A (en) * | 2018-12-02 | 2019-03-29 | 清华大学 | Production fights network equipment and training method |
US11574199B2 (en) | 2018-12-02 | 2023-02-07 | Tsinghua University | Generative adversarial network device and training method thereof |
CN109670036A (en) * | 2018-12-17 | 2019-04-23 | 广州大学 | A kind of news comment automatic generation method and its device |
CN109670178B (en) * | 2018-12-20 | 2019-10-08 | 龙马智芯(珠海横琴)科技有限公司 | Sentence-level bilingual alignment method and device, computer readable storage medium |
CN109670178A (en) * | 2018-12-20 | 2019-04-23 | 龙马智芯(珠海横琴)科技有限公司 | Sentence-level bilingual alignment method and device, computer readable storage medium |
CN109887047B (en) * | 2018-12-28 | 2023-04-07 | 浙江工业大学 | Signal-image translation method based on generation type countermeasure network |
CN109887047A (en) * | 2018-12-28 | 2019-06-14 | 浙江工业大学 | A kind of signal-image interpretation method based on production confrontation network |
CN109993678A (en) * | 2019-03-26 | 2019-07-09 | 南京联创北斗技术应用研究院有限公司 | It is a kind of to fight the robust steganography method for generating network based on depth |
CN109993678B (en) * | 2019-03-26 | 2020-04-07 | 南京联创北斗技术应用研究院有限公司 | Robust information hiding method based on deep confrontation generation network |
CN110288609A (en) * | 2019-05-30 | 2019-09-27 | 南京师范大学 | A kind of multi-modal whole-heartedly dirty image partition method of attention mechanism guidance |
CN110288609B (en) * | 2019-05-30 | 2021-06-08 | 南京师范大学 | Multi-modal whole-heart image segmentation method guided by attention mechanism |
CN110309512A (en) * | 2019-07-05 | 2019-10-08 | 北京邮电大学 | A kind of Chinese grammer error correction method thereof based on generation confrontation network |
CN110334361B (en) * | 2019-07-12 | 2022-11-22 | 电子科技大学 | Neural machine translation method for Chinese language |
CN110334361A (en) * | 2019-07-12 | 2019-10-15 | 电子科技大学 | A kind of neural machine translation method towards rare foreign languages language |
CN110472238A (en) * | 2019-07-25 | 2019-11-19 | 昆明理工大学 | Text snippet method based on level interaction attention |
CN110472238B (en) * | 2019-07-25 | 2022-11-18 | 昆明理工大学 | Text summarization method based on hierarchical interaction attention |
CN110598221A (en) * | 2019-08-29 | 2019-12-20 | 内蒙古工业大学 | Method for improving translation quality of Mongolian Chinese by constructing Mongolian Chinese parallel corpus by using generated confrontation network |
CN111046178A (en) * | 2019-11-29 | 2020-04-21 | 北京邮电大学 | Text sequence generation method and system |
CN111046178B (en) * | 2019-11-29 | 2023-06-20 | 北京邮电大学 | Text sequence generation method and system |
CN111414770A (en) * | 2020-02-24 | 2020-07-14 | 内蒙古工业大学 | Semi-supervised Mongolian neural machine translation method based on collaborative training |
CN111859995B (en) * | 2020-06-16 | 2024-01-23 | 北京百度网讯科技有限公司 | Training method and device of machine translation model, electronic equipment and storage medium |
CN111859995A (en) * | 2020-06-16 | 2020-10-30 | 北京百度网讯科技有限公司 | Training method and device of machine translation model, electronic equipment and storage medium |
CN112100908B (en) * | 2020-08-31 | 2024-03-22 | 西安工程大学 | Clothing design method for generating countermeasure network based on multi-condition deep convolution |
CN112100908A (en) * | 2020-08-31 | 2020-12-18 | 西安工程大学 | Garment design method for generating confrontation network based on multi-condition deep convolution |
CN112489158A (en) * | 2021-01-13 | 2021-03-12 | 河北大学 | Enhancement method for low-dose PET image by using cGAN-based adaptive network |
CN112489158B (en) * | 2021-01-13 | 2023-05-12 | 河北大学 | Enhancement method for low-dose PET image by adopting adaptive network based on cGAN |
CN113657129A (en) * | 2021-09-06 | 2021-11-16 | 内蒙古工业大学 | Mongolian Chinese neural machine translation method combining generation of confrontation network and near-end optimization strategy |
CN115660037A (en) * | 2022-10-26 | 2023-01-31 | 广东东方思维科技有限公司 | Method for distinguishing through deep neural network model |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108897740A (en) | A kind of illiteracy Chinese machine translation method based on confrontation neural network | |
CN107368475B (en) | Machine translation method and system based on generation of antagonistic neural network | |
CN109783657B (en) | Multi-step self-attention cross-media retrieval method and system based on limited text space | |
Herzig et al. | Neural semantic parsing over multiple knowledge-bases | |
CN108829684A (en) | A kind of illiteracy Chinese nerve machine translation method based on transfer learning strategy | |
CN109766427B (en) | Intelligent question-answering method based on collaborative attention for virtual learning environment | |
CN108416065A (en) | Image based on level neural network-sentence description generates system and method | |
CN108932232A (en) | A kind of illiteracy Chinese inter-translation method based on LSTM neural network | |
CN110489554B (en) | Attribute-level emotion classification method based on location-aware mutual attention network model | |
CN110427616A (en) | A kind of text emotion analysis method based on deep learning | |
Zhang et al. | Learning sentiment-inherent word embedding for word-level and sentence-level sentiment analysis | |
CN112527966A (en) | Network text emotion analysis method based on Bi-GRU neural network and self-attention mechanism | |
CN111782788A (en) | Automatic emotion reply generation method for open domain dialogue system | |
Luo et al. | Hierarchical transfer learning architecture for low-resource neural machine translation | |
CN116663578A (en) | Neural machine translation method based on strategy gradient method improvement | |
CN113657125B (en) | Mongolian non-autoregressive machine translation method based on knowledge graph | |
CN111563148A (en) | Dialog generation method based on phrase diversity | |
Wang et al. | Information-enhanced hierarchical self-attention network for multiturn dialog generation | |
Maslennikova | ELMo Word Representations For News Protection. | |
CN114328866A (en) | Strong anthropomorphic intelligent dialogue robot with smooth and accurate response | |
CN115408515A (en) | Automatic classification method for online collaborative discussion interactive text | |
Wang | Short Sequence Chinese‐English Machine Translation Based on Generative Adversarial Networks of Emotion | |
Wang et al. | Emotional conversation generation with bilingual interactive decoding | |
CN116976361A (en) | RLHF-based self-adaptive machine translation method and storage medium | |
CN116258147A (en) | Multimode comment emotion analysis method and system based on heterogram convolution |
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 | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181127 |