CN108932232A - A kind of illiteracy Chinese inter-translation method based on LSTM neural network - Google Patents
A kind of illiteracy Chinese inter-translation method based on LSTM neural network Download PDFInfo
- Publication number
- CN108932232A CN108932232A CN201810428619.0A CN201810428619A CN108932232A CN 108932232 A CN108932232 A CN 108932232A CN 201810428619 A CN201810428619 A CN 201810428619A CN 108932232 A CN108932232 A CN 108932232A
- Authority
- CN
- China
- Prior art keywords
- neural network
- input
- hidden layer
- word
- lstm neural
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/205—Parsing
- G06F40/211—Syntactic parsing, e.g. based on context-free grammar [CFG] or unification grammars
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- 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)
- 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
In order to change the relatively backward status of Inner Mongolia development of Mechanical Translation, the present invention has studied a kind of illiteracy Chinese inter-translation method based on LSTM neural network, compared with the existing machine translation method based on statistics, the present invention has fully considered the connection between entire sentence and context using the neural machine translation method based on LSTM first, improves the efficiency of machine translation;Secondly, improving the quality and efficiency of translation using the encoder and decoder for being substantially two-way LSTM neural network, finally, training these data as optimization algorithm using most small quantities of stochastic gradient descent algorithm, the quality of translation is further improved.
Description
Technical field
The invention belongs to machine translation mothod field, in particular to a kind of illiteracy Chinese intertranslation side based on LSTM neural network
Method.
Background technique
The Mongols is national one of the important composition of the Chinese nation 56, be the nomadic Typical Representative in grassland and
The important succession person of Grassland Culture, Mongol are then the dominant languages that Mongolians compatriot uses, and are official's languages of Mongolia
Speech, use scope is extensive, has critically important status in the world.Along with the fast development of China's economic, the Mongols with
Economy, cultural exchanges between Han nationality is more and more extensive, and the Mongols will be exchanged with Han nationality compatriot and just need to translate, and people
Work translates higher cost, this brings many inconvenience to the economic development of the Mongols.
Fortunately, with the arrival of big data era, artificial intelligence rapid development, machine translation is more and more mature, leads to
The intertranslation crossed between computer Mongol and Chinese becomes possibility.For example, a Chinese articles can be provided, machine can be certainly
It is dynamic quickly to generate a Mongolian article.In recent years, it is mutual for the values of two races culture to cover Chinese research on the machine translation
Infiltration promotes foreign trade and cultural exchanges between the foundation and promotion and Mongolia of good national relations to play very big
Effect.
In late 1940s this more than 60 years time so far, the step of development of Mechanical Translation is follow, is studied
Personnel also never stopped the research that machine translation is applied in covering Chinese translation.
The sentence element of Chinese and Mongolian sentence element different from, and have on the word order of sentence it is very big not
Together, this brings very big difficulty to Chinese machine translation is covered.
The translation system of earlier version is phrase-based machine translation, i.e. PBMT (Phrase-based Machine
Translation).The sentence of input can be divided into one group of word or phrase by PBMT, and it is individually translated.This is apparently not
Optimal Translation Strategy has ignored the connection between entire sentence and context completely.
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 LSTM neural network
Chinese inter-translation method is covered, is turned over from the pervious illiteracy Chinese machine based on statistics contacted between entire sentence and context of ignoring completely
It translates, to the neural machine translation based on LSTM contacted between entire sentence and context is fully considered, improves machine translation
Quality and efficiency, improve traditional NMT (Neural Machine Translation), change Inner Mongolia machine
The relatively backward status of translation development.
To achieve the goals above, the technical solution adopted by the present invention is that:
A kind of illiteracy Chinese inter-translation method based on LSTM neural network, using coding-decoding structure, encoder reads original language
Sentence, is encoded to the fixed vector of dimension, and decoder reads the vector, sequentially generates object language.Encoder and decoder are equal
Using LSTM neural network, in the encoder, the source language sentence of input is carried out by a two-way LSTM neural network
Coding forms context semantic vector group wherein each sentence is expressed as a context semantic vector, and the context is semantic
The coding that vector is intended to as user;In a decoder, the continuous circular flow of LSTM neural network generates in object language
Each word, while generating each word, consider input source language sentence corresponding to context semantic vector, make
The content that must be generated is consistent with the meaning of original language.
The encoder uses two-way LSTM model, the output at unidirectional LSTM neural network moment and current time
And input information before is related, and the output at two-way LSTM neural network a certain moment not only with current time and it
Preceding input information is related, also information-related with later, adequately considers contacting between entire sentence and context.
The LSTM neural network of forward directionInput x=(x is read according to the sequence of source language sentence input1..., xI), x1
Indicate the 1st word of input, xIIt indicates the i-th word of input, and calculates preceding to hidden layer state
Indicate preceding the 1st semantic vector element to hidden layer state,To the i-th semantic vector element of hidden layer state before indicating;Afterwards
To LSTM neural network read sequence (x by inputting opposite sequence with original textI..., x1), xIIndicate the i-th of input
Word, x1Indicate the 1st word of input, and to hidden layer state after calculating To hidden layer state after expression
1st semantic vector element,To the i-th semantic vector element of hidden layer state after expression, by forward and backward hidden layer state
It links together to obtain the explanation vector of each wordJ from 1 to I,It is semantic to hidden layer state before indicating
The transposition of vector,To the transposition of hidden layer state semantic vector after expression.
Decoder:
In given source language sentence x and target language { y1..., yt-1Under conditions of, decoder sequentially generates object language
Word yt, defining conditional probability in decoding layer is:
p(yt|y1,...,yt-1, X) and=g (yt-1,st,ct)
Wherein, g is activation primitive sigmoid, stIt is hidden layer state of the LSTM neural network in t moment of decoder, ctTable
Show the external input information in generating process.
Translate the guarantee of accuracy rate
In generating process above, every a Chinese sentence generates a context semantic vector, for translating a primitive
Sentence, therefore, object statement after translation without departing from text original meaning, meanwhile, the generation result of latter sentence depends on previous sentence
Word, therefore ensure that the continuity of entire sentence, ensure that the accuracy rate of translation.
Training
During training, the output vector of decoder is input to a softmax and is returned in layer, selection
Softmax returns the cross entropy of actual probability distribution and desired output that layer obtains as loss function, and loss function is:
Y is desired output, and a is reality output, and n is batch size.
In order to accelerate to train, prevent from falling into local optimum, present invention employs most small quantities of stochastic gradient descent algorithm conducts
Optimization algorithm:
Wherein,:=it is synchronized update,For learning rate;W and b ceaselessly update, when the value of w and b tends towards stability
Terminate, completes gradient descent procedures.
Compared with the existing machine translation method based on statistics, the present invention fully considered entire sentence and context it
Between connection, improve the efficiency of machine translation;Secondly, ensure that higher using the encoder and decoder for being substantially LSTM
Translation quality;Finally, improving the accuracy of translation by way of with residual error study and reducing gradient disappearance.
Detailed description of the invention
Fig. 1 is the resolution principle figure of hidden layer state.
Fig. 2 is Chinese of the invention to Mongolian translation schematic illustration.
Fig. 3 is the translation schematic illustration of Mongol of the invention to Chinese.
Specific embodiment
The embodiment that the present invention will be described in detail with reference to the accompanying drawings and examples.
Illiteracy Chinese inter-translation method based on LSTM neural network, using coding-decoding structure, encoder and decoder are all logical
Cross LSTM realization.
Wherein encoder uses two-way LSTM model, the LSTM neural network of forward directionAccording to the suitable of source language sentence input
Sequence reads input x=(x1..., xI), x1Indicate the 1st word of input, xIIndicate the i-th word of input, and before calculating to
Hidden layer state Indicate preceding the 1st semantic vector element to hidden layer state,To hidden layer shape before indicating
The i-th semantic vector element of state;Backward LSTM neural network inputs opposite sequence reading sequence by with original text
(xI..., x1), xIIndicate the i-th word of input, x1Indicate the 1st word of input, and to hidden layer state after calculating To the 1st semantic vector element of hidden layer state after expression,To the i-th semantic vector of hidden layer state after expression
Element links together forward and backward hidden layer state to obtain the explanation vector of each wordJ from 1 to
I,Indicate the preceding transposition to hidden layer state semantic vector,To the transposition of hidden layer state semantic vector after expression.
Fig. 1 is the Computing Principle of hidden layer state, which is referred to as memory block (block of memory), mainly contains three
Door (forget gate, input gate, output gate) and a memory unit (cell).The Na Tiaoshui of top in box
Horizontal line, referred to as location mode (cell state), it can control information and pass to subsequent time just as a conveyer belt.
The solution procedure of hidden layer state:
The first step:Determine that information can be by cell state first.
This decision is controlled by " forget gate " layer by sigmoid, it can be according to the output h of last momentj-1
With current input xjTo generate one 0 to 1 fjValue, to decide whether the information C for allowing last moment to acquirejBy or part it is logical
It crosses.0 indicates not allow to pass through completely, and the value between 0-1 indicates that part passes through, and 1 indicates to allow to pass through completely.
It is as follows:
fj=Sigmod (Wf*xj+Wfhj-1+bf)
Wherein:WfAnd bfIt indicates to forget the parameter looked after and guided in the neuron that door includes.
Second step:Generate the new information for needing to update.
This step includes two parts, and first part is that " input gate " layer determines which value is used to by sigmoid
It updates, ijValue be 1 when indicate do not need to update, need to update when between 0 or 0-1.Second part is tanh layers and uses next life
The candidate value of Cheng XinIt may be added in cell state as the candidate value that current layer generates,Value be 0-1
Or it is added to when 1 in cell state.The value that this two parts generates in conjunction with being updated, such as following formula:
Wherein:Wherein:WiAnd biThe parameter looked after and guided in the neuron that expression input gate includes.WCAnd bCIndicate that memory is single
The parameter looked after and guided in the neuron that member includes.
It is exactly to lose unwanted information that the first step and second step, which combine, adds the process of new information:
Third step:The output of decision model.
It is to obtain an initial output by sigmoid layers first, then uses tanh by CjValue zooms to -1 to 1
Between, then the output that obtains with sigmoid is by being multiplied, thus to hidden layer state before obtaining
Wherein:WoAnd boThe parameter looked after and guided in the neuron that expression out gate includes.
It is calculated with same method
In given source language sentence x and target language { y1..., yt-1Under conditions of, decoder sequentially generates object language
Word yt, defining conditional probability in decoding layer is:
p(yt|y1,...,yt-1, X) and=g (yt-1,st,ct)
Wherein, g is activation primitive sigmoid, stIt is hidden layer state of the LSTM neural network in t moment of decoder, ctTable
Show the external input information in generating process.
Translate the guarantee of accuracy rate
In generating process above, every a Chinese sentence generates a context semantic vector, for translating a primitive
Sentence, therefore, object statement after translation without departing from text original meaning, meanwhile, the generation result of latter sentence depends on previous sentence
Word, therefore ensure that the continuity of entire sentence, ensure that the accuracy rate of translation.
Training
During training, the vector that decoder is obtained inputs a softmax and returns in layer, obtains possible outcome
Probability distribution.The actual probability distribution for selecting softmax to obtain in desired output cross entropy as loss function.Lose letter
Number is:
Y is desired output, and a is reality output, and n is batch size.
In order to accelerate to train, prevents from falling into local optimum, the parameter w and b inside neuron are adjusted, which claims
For gradient decline, present invention employs most small quantities of stochastic gradient descent algorithms as optimization algorithm:
Wherein,:=it is synchronized update,For learning rate;W and b ceaselessly update, when the value of w and b tends towards stability
Terminate, completes gradient descent procedures.
It is two specific embodiments that the Chinese covers intertranslation below.
Embodiment 1, the Chinese translate illiteracy:
Referring to Fig. 2, firstly, the lower half portion of Fig. 2, (x1..., xt) indicate a Chinese sentence t word, user is defeated
The t word entered passes through a two-way LSTM network in order and is encoded, and forms a context semantic vector groupThen, it is encoded again by the two-way LSTM network in reverse order, forms a context semantic vector
GroupThen, the coding that these context semantic vectors are intended to as user.(Fig. 1's is upper during generation
Portion), in decoder, the continuous circular flow of LSTM neural network firstly generates the hidden layer state s at i momenti, then generate
Each of object language word yi, while generating each word, need to consider to input context language corresponding to this
Adopted vector, so that the content that this generates is consistent with the meaning of original language.
Specific translation steps are as follows:
1. the source language sentence x=(x that encoder reads input1..., xI);
2. the x read is encoded to hidden layer state using Recognition with Recurrent Neural Network by encoder, formed a context semanteme to
Amount group
3. the x read is inversely encoded to hidden layer state using Recognition with Recurrent Neural Network by encoder, method same as step 2
To hidden layer state after obtaining, to form a context semantic vector group
4. encoder links together forward and backward state to obtain the explanation vector of each word
5. the continuous circular flow of LSTM neural network of decoder generates the hidden layer state s of t momentt;
6. decoder is in given source language sentence x and target language { y1..., yt-1Under conditions of, sequentially generate target language
The word y of speecht.The coding vector for the original language to input that encoder calculates is inputed to the RNN unit of decoder, so
Afterwards, decoder can calculate probability vector according to Recognition with Recurrent Neural Network unit.I.e. for each of target language sentence word
Probability is calculated.Finally, generating object language according to the probability sampling being calculated.
Embodiment 2, illiteracy translate the Chinese:
It is translated with the Chinese and covers same method, referring to Fig. 3, the lower half portion of Fig. 3, (x1..., xt) indicate that the t of a Mongolian is a
Word, the t word that system inputs user pass through a two-way LSTM network in order and encode, and form a context language
Adopted Vector GroupsThen, it is encoded again by a two-way LSTM network in reverse order, forms about one
Literary semantic vector groupThen, the coding that these context semantic vectors are intended to as user.In the mistake of generation
In journey (top of Fig. 2), the continuous circular flow of LSTM neural network firstly generates the hidden layer state s at i momenti, then give birth to
At each of Mongolian word yi, while generating each word, need to consider to input the semanteme of context corresponding to this
Vector, so that the content that this generates is consistent with the meaning of original language.
Claims (6)
1. a kind of illiteracy Chinese inter-translation method based on LSTM neural network, using coding-decoding structure, encoder reads source language sentence
Son, is encoded to the fixed vector of dimension, and decoder reads the vector, sequentially generates object language, which is characterized in that encoder and
Decoder is all made of LSTM neural network, in the encoder, by the source language sentence of input by a two-way LSTM nerve
Network is encoded, wherein each sentence is expressed as a context semantic vector, forms context semantic vector group, it is described on
The hereafter coding that semantic vector is intended to as user;In a decoder, the continuous circular flow of LSTM neural network generates mesh
Each of poster speech word considers context language corresponding to the source language sentence of input while generating each word
Adopted vector, so that the content generated is consistent with the meaning of original language.
2. the illiteracy Chinese inter-translation method based on LSTM neural network according to claim 1, which is characterized in that the encoder makes
With two-way LSTM model, the LSTM neural network of forward directionInput x=(x is read according to the sequence of source language sentence input1...,
xI), x1Indicate the 1st word of input, xIIt indicates the i-th word of input, and calculates preceding to hidden layer state Indicate preceding the 1st semantic vector element to hidden layer state,To the i-th semantic vector element of hidden layer state before indicating;
Backward LSTM neural network inputs opposite sequence reading sequence (x by with original textI..., x1), xIIndicate the I of input
A word, x1Indicate the 1st word of input, and to hidden layer state after calculating To the 1st of hidden layer state the after expression
A semantic vector element,To the i-th semantic vector element of hidden layer state after expression, forward and backward hidden layer state is linked
The explanation vector of each word is obtained togetherJ from 1 to I,To hidden layer state semantic vector before indicating
Transposition,To the transposition of hidden layer state semantic vector after expression.
3. the illiteracy Chinese inter-translation method based on LSTM neural network according to claim 2, which is characterized in that the hidden layer state
Calculating included the following steps using block of memory (memory block) framework:
The first step:Determine that information can be by location mode (cell state) first
This decision is controlled by gate layers of forget by sigmoid, it can be according to the output h of last momentj-1With it is current
Input xjGenerate one 0 to 1 fjValue, to decide whether the information C for allowing last moment to acquirejBy or part pass through, it is as follows:
fj=Sigmod (Wf *xj+Wfhj-1+bf)
Wherein:WfAnd bfIt indicates to forget the parameter looked after and guided in the neuron that door includes;
Second step:Generate the new information for needing to update
The step includes two parts, and first part is gate layers of input and determines which value is used to update by sigmoid, and second
Part is tanh layers and is used to generate new candidate valueIt may be added to cell as the candidate value that current layer generates
In state;The value that this two parts generates is combined and is updated, such as following formula:
ij=Sigmod (Wi*xj+Wi*hj-1+bi)
Wherein:WiAnd biThe parameter looked after and guided in the neuron that expression input gate includes;WCAnd bCIndicate the mind that memory unit includes
Parameter through having been looked after and guided in member;
The first step and second step combine, that is, lose unwanted information, add the process of new information:
Third step, the output of decision model
It is to obtain an initial output by sigmoid layers first, then uses tanh by CjValue zooms between -1 to 1, then with
The output that sigmoid is obtained is by multiplication, thus to hidden layer state before obtaining
oj=Sigmod (Wo*xj+Wo*hj-1+bo)
Wherein:WoAnd boThe parameter looked after and guided in the neuron that expression out gate includes;
It is calculated with same method
4. the illiteracy Chinese inter-translation method based on LSTM neural network according to claim 1, which is characterized in that the decoder
In, in given source language sentence x and target language { y1..., yt-1Under conditions of, decoder sequentially generates the word y of object languaget,
Defining conditional probability in decoding layer is:
p(yt|y1,...,yt-1, X) and=g (yt-1,st,ct)
Wherein, g is activation primitive sigmoid, stIt is hidden layer state of the LSTM neural network in t moment of decoder, ctIndicate life
At external input information in the process, y1..., yt-1Refer to the 1st word of object language that has generated to the t-1 word.
5. the illiteracy Chinese inter-translation method based on LSTM neural network according to claim 1, which is characterized in that in the following way
Carry out model data training:
The output vector of decoder is input to a softmax to return in layer, selects softmax to return the reality that layer obtains general
As loss function, loss function is the cross entropy of rate distribution and desired output:
Y is desired output, and a is reality output, and n is batch size.
6. the illiteracy Chinese inter-translation method based on LSTM neural network according to claim 5, which is characterized in that using most in small batches with
Machine gradient descent algorithm accelerates to train as optimization algorithm, prevents from falling into local optimum, be looked after and guided by following formula:
Wherein,:=it is synchronized update,For learning rate;W and b does not stop to update, and terminates when the value of w and b tends towards stability, complete
At gradient descent procedures.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810428619.0A CN108932232A (en) | 2018-05-07 | 2018-05-07 | A kind of illiteracy Chinese inter-translation method based on LSTM neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810428619.0A CN108932232A (en) | 2018-05-07 | 2018-05-07 | A kind of illiteracy Chinese inter-translation method based on LSTM neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108932232A true CN108932232A (en) | 2018-12-04 |
Family
ID=64448397
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810428619.0A Pending CN108932232A (en) | 2018-05-07 | 2018-05-07 | A kind of illiteracy Chinese inter-translation method based on LSTM neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108932232A (en) |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109558605A (en) * | 2018-12-17 | 2019-04-02 | 北京百度网讯科技有限公司 | Method and apparatus for translating sentence |
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 |
CN109740169A (en) * | 2019-01-09 | 2019-05-10 | 北京邮电大学 | A kind of Chinese medical book interpretation method based on dictionary and seq2seq pre-training mechanism |
CN110414012A (en) * | 2019-07-29 | 2019-11-05 | 腾讯科技(深圳)有限公司 | A kind of encoder construction method and relevant device based on artificial intelligence |
CN110489766A (en) * | 2019-07-25 | 2019-11-22 | 昆明理工大学 | The Chinese-weighed based on coding conclusion-decoding gets over low-resource nerve machine translation method |
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 |
CN110717345A (en) * | 2019-10-15 | 2020-01-21 | 内蒙古工业大学 | Translation realignment recurrent neural network cross-language machine translation method |
CN112016332A (en) * | 2020-08-26 | 2020-12-01 | 华东师范大学 | Multi-modal machine translation method based on variational reasoning and multi-task learning |
CN112395892A (en) * | 2020-12-03 | 2021-02-23 | 内蒙古工业大学 | Mongolian Chinese machine translation method for realizing placeholder disambiguation based on pointer generation network |
WO2021082518A1 (en) * | 2019-11-01 | 2021-05-06 | 华为技术有限公司 | Machine translation method, machine translation model training method and device, and storage medium |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107870902A (en) * | 2016-09-26 | 2018-04-03 | 谷歌公司 | Neural machine translation system |
CN107967262A (en) * | 2017-11-02 | 2018-04-27 | 内蒙古工业大学 | A kind of neutral net covers Chinese machine translation method |
-
2018
- 2018-05-07 CN CN201810428619.0A patent/CN108932232A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107870902A (en) * | 2016-09-26 | 2018-04-03 | 谷歌公司 | Neural machine translation system |
CN107967262A (en) * | 2017-11-02 | 2018-04-27 | 内蒙古工业大学 | A kind of neutral net covers Chinese machine translation method |
Non-Patent Citations (2)
Title |
---|
COLAH: "Understanding LSTM Networks", 《COLAH.HITHUB.IO/POSTS/2015-08-UNDERSTANDING-LSTMS》 * |
WEPON_: "交叉熵代价函数", 《HTTPS://BLOG.CSDN.NET/U012162613/ARTICLE/DETAILS/44239919》 * |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109558605B (en) * | 2018-12-17 | 2022-06-10 | 北京百度网讯科技有限公司 | Method and device for translating sentences |
CN109558605A (en) * | 2018-12-17 | 2019-04-02 | 北京百度网讯科技有限公司 | Method and apparatus for translating sentence |
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 |
CN109740169A (en) * | 2019-01-09 | 2019-05-10 | 北京邮电大学 | A kind of Chinese medical book interpretation method based on dictionary and seq2seq pre-training 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 |
CN109740169B (en) * | 2019-01-09 | 2020-10-13 | 北京邮电大学 | Traditional Chinese medicine ancient book translation method based on dictionary and seq2seq pre-training mechanism |
CN110489766A (en) * | 2019-07-25 | 2019-11-22 | 昆明理工大学 | The Chinese-weighed based on coding conclusion-decoding gets over low-resource nerve machine translation method |
CN110414012A (en) * | 2019-07-29 | 2019-11-05 | 腾讯科技(深圳)有限公司 | A kind of encoder construction method and relevant device based on artificial intelligence |
CN110414012B (en) * | 2019-07-29 | 2022-12-09 | 腾讯科技(深圳)有限公司 | Artificial intelligence-based encoder construction method and related equipment |
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 |
CN110717345A (en) * | 2019-10-15 | 2020-01-21 | 内蒙古工业大学 | Translation realignment recurrent neural network cross-language machine translation method |
WO2021082518A1 (en) * | 2019-11-01 | 2021-05-06 | 华为技术有限公司 | Machine translation method, machine translation model training method and device, and storage medium |
CN112016332B (en) * | 2020-08-26 | 2021-05-07 | 华东师范大学 | Multi-modal machine translation method based on variational reasoning and multi-task learning |
CN112016332A (en) * | 2020-08-26 | 2020-12-01 | 华东师范大学 | Multi-modal machine translation method based on variational reasoning and multi-task learning |
CN112395892A (en) * | 2020-12-03 | 2021-02-23 | 内蒙古工业大学 | Mongolian Chinese machine translation method for realizing placeholder disambiguation based on pointer generation network |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108932232A (en) | A kind of illiteracy Chinese inter-translation method based on LSTM neural network | |
CN110334361B (en) | Neural machine translation method for Chinese language | |
CN110717334B (en) | Text emotion analysis method based on BERT model and double-channel attention | |
CN106126507B (en) | A kind of depth nerve interpretation method and system based on character code | |
CN107368475B (en) | Machine translation method and system based on generation of antagonistic neural network | |
Chen et al. | Neural machine translation with source dependency representation | |
CN110069790B (en) | Machine translation system and method for contrasting original text through translated text retranslation | |
CN108829684A (en) | A kind of illiteracy Chinese nerve machine translation method based on transfer learning strategy | |
CN108897740A (en) | A kind of illiteracy Chinese machine translation method based on confrontation neural network | |
CN110033008B (en) | Image description generation method based on modal transformation and text induction | |
Feng et al. | Joint extraction of entities and relations using reinforcement learning and deep learning | |
CN108153864A (en) | Method based on neural network generation text snippet | |
CN107766320A (en) | A kind of Chinese pronoun resolution method for establishing model and device | |
Tang et al. | Deep sequential fusion LSTM network for image description | |
CN110162789A (en) | A kind of vocabulary sign method and device based on the Chinese phonetic alphabet | |
CN110427616A (en) | A kind of text emotion analysis method based on deep learning | |
CN110188348A (en) | A kind of Chinese language processing model and method based on deep neural network | |
CN114398976A (en) | Machine reading understanding method based on BERT and gate control type attention enhancement network | |
CN113033189B (en) | Semantic coding method of long-short term memory network based on attention dispersion | |
CN112883722B (en) | Distributed text summarization method based on cloud data center | |
CN108388944B (en) | A kind of automatic chatting method and robot based on deep neural network | |
Zhao et al. | Synchronously improving multi-user English translation ability by using AI | |
CN110334196A (en) | Neural network Chinese charater problem based on stroke and from attention mechanism generates system | |
CN114691858B (en) | Improved UNILM digest generation method | |
CN116663578A (en) | Neural machine translation method based on strategy gradient method improvement |
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: 20181204 |