CN110442880A - A kind of interpretation method, device and the storage medium of machine translation translation - Google Patents
A kind of interpretation method, device and the storage medium of machine translation translation Download PDFInfo
- Publication number
- CN110442880A CN110442880A CN201910721252.6A CN201910721252A CN110442880A CN 110442880 A CN110442880 A CN 110442880A CN 201910721252 A CN201910721252 A CN 201910721252A CN 110442880 A CN110442880 A CN 110442880A
- Authority
- CN
- China
- Prior art keywords
- translation
- beam search
- evaluation function
- term
- word
- 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
Landscapes
- Machine Translation (AREA)
Abstract
The invention discloses interpretation method, device and the storage mediums of a kind of machine translation translation, comprising: receives source statement to be translated;Word segmentation processing is carried out to the source statement;Obtain the part of speech of each word in the participle;According to term vector model, the part of speech is incorporated in term vector corresponding to word, fused term vector sequence is obtained;By the term vector sequence inputting into coder-decoder model, encoding and decoding result is obtained;For the encoding and decoding as a result, carrying out evaluation of result based on beam search evaluation function, wherein the beam search evaluation function includes in the penalty term based on length vs and the penalty term for repeating to detect;Translation is obtained according to the evaluation result.Using the embodiment of the present invention, improve duplicated in translation segment and omit source statement the problem of, it is applied widely, with strong points, translation translation quality it is higher.
Description
Technical field
The present invention relates to machine translation translation technical field of improvement more particularly to a kind of translation sides of machine translation translation
Method, device and storage medium.
Background technique
Language is a kind of most important carrier of the mankind's usually information interchange, it has the development of entire society very heavy
The influence wanted, the method for Machine automated translation have become a current urgent demand.Realize oneself of different language
Dynamicization is translated by huge application controls.
Currently, rule-based machine translation method needs the linguist of profession to formulate a large amount of rule, cost of labor
Height, poor expandability.Machine translation method based on intermediate language needs to formulate a set of general intermediate language, and difficulty is too high,
And robustness is low.Although the machine translation method cost of labor based on statistics is lower, scalability is improved, translation quality
It is still poor.Machine translation method neural network based is current state-of-the-art machine translation method, but for translating translation
Quality still have an improved space.
Summary of the invention
The purpose of the present invention is to provide interpretation method, device and the storage mediums of a kind of machine translation translation, it is intended to solve
Certainly existing Machine Translation Model generates the poor problem of translation quality.
To achieve the goals above, the present invention provides a kind of interpretation method of machine translation translation, which comprises
Receive source statement to be translated;
Word segmentation processing is carried out to the source statement;
Obtain the part of speech of each word in the participle;
According to term vector model, the part of speech is incorporated in term vector corresponding to word, fused term vector is obtained
Sequence;
By the term vector sequence inputting into coder-decoder model, encoding and decoding result is obtained;
For the encoding and decoding as a result, carrying out evaluation of result based on beam search evaluation function, wherein the beam search
Evaluation function includes in the penalty term based on length vs and the penalty term for repeating to detect;
Translation is obtained according to the evaluation result.
Further, the beam search evaluation function embodies are as follows:
S (Y, X)=log (P (Y | X))+d (x)+l (x)
Wherein, x (Y, X) is beam search evaluation function, and log (P (Y/X)) is the probability function that Y occurs when X occurs, d
It (x) is based on the penalty term for repeating detection, l (x) is the punishment based on length vs, and P is distribution function;
The penalty term based on lenth ratio is added in beam search evaluation function, is turned over for solving translation appearance part leakage
The problem of;
It is added in beam search evaluation function based on the penalty term for repeating detection, duplicates content for solving translation
The problem of.
Further, the specific formula expression for repeating detection penalty term d (x) are as follows:
Wherein, c is the index where current translation of words, and δ, which attaches most importance to, rechecks the range of survey, and ε is penalty coefficient, and y is candidate
Matrix corresponding to translation, yc-j, yc-i-jRespectively repeat two matrixes of detection, i, j, to traverse variable.
Further, described to be directed to the encoding and decoding as a result, carrying out the step of evaluation of result based on beam search evaluation function
Suddenly, comprising:
The length of the source statement and the lenth ratio of target translation;
The lenth ratio is fitted by linear regression, obtains Cumulative Distribution Function;
When occurring end of the sentence label and common word in the candidate word of beam search simultaneously, translation is over general
Rate FX(x) the probability 1-F not terminated with translationX(x) it is added separately to evaluation function l (x)=θ FX(x), not_EOS, l (x)=θ
(1-FX(x)), EOS, wherein EOS is end of the sentence label, and θ is parameter;
When candidate word is end of the sentence label mark, also untranslated good probability is multiplied by penalty factor as penalty term;
And when candidate word is not end of the sentence label mark, the probability for completing translation is multiplied by penalty factor as penalty term;
The obtained penalty term based on lenth ratio is added in the evaluation function of beam search;
Evaluation of result is carried out based on beam search evaluation function.
Further, in the coder-decoder model, encoder section and decoder section use bidirectional circulating
Neural network.
Further, it is described by the term vector sequence inputting into coder-decoder model, obtain encoding and decoding result
The step of, comprising:
By the term vector sequence inputting into coder-decoder model;
Term vector sequence is converted the vector that forms a complete sentence by the deep learning frame based on coding decoder;
It is being based on decoder, sentence vector is converted into term vector sequence.
In addition, described device includes processor and passes through logical the invention also discloses a kind of machine translation translation device
The memory that letter bus is connected to the processor;Wherein,
The memory, for storing the interpretive program of machine translation translation;
The processor, for executing the interpretive program of the machine translation translation, to realize described in any item machines
Translate the translation steps of translation.
And a kind of computer storage medium, the computer storage medium is stored with one or more program, described
One or more program can be executed by one or more processor, so that the execution of one or more of processors is any
The translation steps of machine translation translation described in.
Using a kind of interpretation method, device and the storage medium of machine translation translation provided in an embodiment of the present invention, having
While building vector in effect ground establishes the semantic association between different terms, to the meaning comprising them under different parts of speech, and
And have modified beam search evaluation function improve duplicated in translation segment and omit source statement the problem of, the scope of application
Extensively, with strong points, translation translation quality it is higher.
Detailed description of the invention
Fig. 1 is a kind of flow diagram of the embodiment of the present invention.
Fig. 2 is a kind of structural schematic diagram of the embodiment of the present invention.
Fig. 3 is another structural schematic diagram of the embodiment of the present invention.
Fig. 4 is the penalty term algorithm description schematic diagram of the repetition detection of the embodiment of the present invention.
Fig. 5 is the penalty term algorithm description schematic diagram of the lenth ratio of the embodiment of the present invention.
Fig. 6 is the english translation effect diagram of the embodiment of the present invention.
Specific embodiment
Illustrate embodiments of the present invention below by way of specific specific example, those skilled in the art can be by this specification
Other advantages and efficacy of the present invention can be easily understood for disclosed content.The present invention can also pass through in addition different specific realities
The mode of applying is embodied or practiced, the various details in this specification can also based on different viewpoints and application, without departing from
Various modifications or alterations are carried out under spirit of the invention.
Language model (language Model) is a simple, unified, abstract formal system, the objective thing of language
It is real to be automatically processed by computer after the description of language model, thus language model is for the letter of natural language
Breath processing is of great significance, and the researchs such as part-of-speech tagging, syntactic analysis, speech recognition are all played an important role.
In machine translation problem, the original language sentence of input and the object language translation of output can be seen as sequence,
Therefore machine translation can be regarded to sequence as to sequence problem.The method of the solution sequence of mainstream to sequence is encoder solution at present
Original language statement coding is sentence vector by code device model, encoder, and sentence vector is decoded again and obtains target language by decoder
Say translation.
It should be noted that usually using Recognition with Recurrent Neural Network RNN as encoder and decoder.Recognition with Recurrent Neural Network
(RNN, RecurrentNeural Network) is a kind of neural network structure of classics, it includes the neural unit of circulation,
Therefore serializability data be can handle and allow the persistence of data information.RNN can be by current input and input before
It is trained and is exported together as parameter.Bidirectional circulating neural network (Bi-RNN, Bidirectional
Recurrent Neural Network), it is based on a kind of improved network structure of Recognition with Recurrent Neural Network.In some tasks,
The input of network is not only related with past input, also has certain association with subsequent input.Therefore in addition to inputting positive sequence
Outside, input Inverse order sequence is also needed.And bidirectional circulating neural network is made of two layers of Recognition with Recurrent Neural Network, it is supported simultaneously
Positive sequence and Inverse order sequence are inputted, the performance of network is effectively improved.
One group of Language Modeling and feature learning skill in term vector (WordVector) embedded natural language processing (NLP)
The general designation of art, wherein the word or expression from vocabulary is mapped to the vector of real number.In concept, it is related to from each
Mathematics insertion of the one-dimensional space of word to the vector row space with more low dimensional.Skip-gram model is in nerve net
When network train language model, for generating the distributed model structure indicated of word, Skip-gram model is by current word
Term vector predicts the context of this word as input.Beam search (Beam Search) beam search, which is that one kind is heuristic, to be searched
Rope algorithm explores figure by most promising node in extension finite aggregate.
Beam search is the optimization of best-first search, it is possible to reduce its memory requirements.Best search is pre- according to attempting
Surveying partial solution and total solution (dbjective state) has how close some heuristic orders to arrange all parts
The graph search of solution (state).But in beam search, the best local solution of only predetermined quantity is just left time
Choosing.
Please refer to Fig. 1-6.It should be noted that only the invention is illustrated in a schematic way for diagram provided in the present embodiment
Basic conception, only shown in schema then with related component in the present invention rather than component count, shape when according to actual implementation
Shape and size are drawn, when actual implementation kenel, quantity and the ratio of each component can arbitrarily change for one kind, and its component cloth
Office's kenel may also be increasingly complex.
As Fig. 1 present invention provides a kind of interpretation method of machine translation translation, which comprises
S110 receives source statement to be translated.
S120 carries out word segmentation processing to source statement;
It is understood that carrying out word segmentation processing to the every sentence received in source statement to be translated.
S130 obtains the part of speech of each word in the participle;
It should be noted that being segmented to every sentence in source statement, then using part-of-speech tagging tool to each
Word carries out part-of-speech tagging, obtains the part of speech of each word, and the abbreviation symbol of corresponding part of speech is obtained by inquiry part of speech abbreviations table
Number.Finally former word is attached to obtain word/part of speech character string and be replaced with corresponding part of speech dummy suffix notation by " _ " symbol
For the former word in source statement.
The part of speech is incorporated in term vector corresponding to word according to term vector model, obtains fused word by S140
Sequence vector;
One group of Language Modeling and feature learning skill in term vector (WordVector) embedded natural language processing (NLP)
The general designation of art, wherein the word or expression from vocabulary is mapped to the vector of real number.In concept, it is related to from each
Mathematics insertion of the one-dimensional space of word to the vector row space with more low dimensional.
In a kind of implementation of the present invention, word all in source statement/part of speech character string is obtained to step S120, S130
Dictionary is counted and constructed, then each word/part of speech character string in dictionary is indexed and is saved.Then by sentence
In word/part of speech character string index value sequence inputting for being converted into index value, and every sentence being represented to skip-gram mould
It is trained in type, obtains the trained term vector for having merged part of speech feature, obtain fused term vector sequence.
Illustratively, as shown in Fig. 2, input be w (t) export after skip-gram model training respectively w (t-2),
w(t-1)、w(t+2)、w(t+1)。
S150 obtains encoding and decoding result by the term vector sequence inputting into coder-decoder model.
It should be noted that trained term vector is replaced by term vector sequence inputting coder-decoder model
Sentence in corpus is converted term vector sequence by the word for each sentence that primitive material is concentrated.Again using term vector sequence as
Input is sent into coder-decoder model, and encoding and decoding result is obtained.Coder-decoder model structure is as shown in Figure 3.
S160, for the encoding and decoding as a result, carrying out evaluation of result based on beam search evaluation function, wherein the wave
Beam search evaluation function includes in the penalty term based on length vs and the penalty term for repeating to detect;
It is understood that beam search is a kind of heuristic search algorithm, by being most hopeful in extension finite aggregate
Node explore figure.Beam search is the optimization of best-first search, it is possible to reduce its memory requirements.Best search is root
There are how close some heuristic orders to arrange according to predicted portions solution and total solution (dbjective state) is attempted
The graph search of all partial solutions (state).But in beam search, the only best local solution ability quilt of predetermined quantity
It is left candidate.The embodiment of the present invention improves the evaluation function of beam search, is added based on the penalty term for repeating detection
With penalty term of the addition based on lenth ratio.
S170 obtains translation according to the evaluation result;
By coder-decoder model and beam search, final translation is obtained.
In a kind of implementation of the invention, the beam search evaluation function is embodied are as follows:
S (Y, X)=log (P (Y | X))+d (x)+l (x)
Wherein, x (Y, X) is beam search evaluation function, and log (P (Y/X)) is the probability function that Y occurs when X occurs, d
It (x) is based on the penalty term for repeating detection, l (x) is the punishment based on length vs, and P is distribution function;
The penalty term based on lenth ratio is added in beam search evaluation function, is turned over for solving translation appearance part leakage
The problem of;
It is added in beam search evaluation function based on the penalty term for repeating detection, duplicates content for solving translation
The problem of.
It should be noted that the embodiment of the present invention improves the evaluation function of beam search, it is added and is based on length ratio
The penalty term of value and the penalty term detected based on repetition.Lenth ratio penalty term is too long or too short for machine translation translation length
The problem of, penalty term is obtained by the ratio of the length and translation length that count source statement, and for beam search to candidate word
Evaluation function in, repeat to detect penalty term and be compared by the way that translation is divided into different size of segment, and will be weighed
The position of compound word between position to be translated at a distance from take into account, finally by the penalty term acquired be used in beam search to time
It selects in the evaluation function of word.The problem of duplicating segment in translation and omitting source statement is improved, it is applied widely, be directed to
Property it is strong, translation translation quality is higher.
Further, the specific formula expression for repeating detection penalty term d (x) are as follows:
Wherein, c is the index where current translation of words, and δ, which attaches most importance to, rechecks the range of survey, and ε is penalty coefficient, and y is candidate
Matrix corresponding to translation, yc-j, yc-i-jRespectively repeat two matrixes of detection, i, j, to traverse variable.
As Fig. 4 passes through using the parameter δ and ε in the candidate sentence and formula of entire beam search as the input of algorithm
The segment for dividing multiple and different sizes is compared, and calculates separately respective penalty term, is finally weighted cumulative.Fig. 5 is herein
The value F for the cumulative distribution function that current candidate word, current length are calculatedX(x) parameter θ and in formula is as algorithm
Input, obtain whether candidate word is EOS by vector operation first, if being be 1, be otherwise 0.Then pass through the shape of dot product
Formula obtains the value of l (x) in formula.
In a kind of implementation of the invention, it is described be directed to the encoding and decoding as a result, based on beam search evaluation function into
The step of row evaluation of result, comprising:
The length of the source statement and the lenth ratio of target translation;
The lenth ratio is fitted by linear regression, obtains Cumulative Distribution Function;
When occurring end of the sentence label and common word in the candidate word of beam search simultaneously, translation is over general
Rate FX(x) the probability 1-F not terminated with translationX(x) it is added separately to evaluation function l (x)=θ FX(x), not_EOS, l (x)=θ
(1-FX(x)), EOS, wherein EOS is end of the sentence label, and θ is parameter;
When candidate word is end of the sentence label mark, also untranslated good probability is multiplied by penalty factor as penalty term;
And when candidate word is not end of the sentence label mark, the probability for completing translation is multiplied by penalty factor as penalty term;
The obtained penalty term based on lenth ratio is added in the evaluation function of beam search;
Evaluation of result is carried out based on beam search evaluation function.
It is understood that first respectively count source statement length and target translation length, and calculate source statement with
Then the lenth ratio of target translation is fitted to obtain its Cumulative Distribution Function F by the lenth ratio that linear regression obtainsX
(x)=P (X < x), wherein
That is the lenth ratio of the length of target translation and source statement.When there is EOS (sentence simultaneously in the candidate word of beam search
End label) and when common word, by the probability FX (x) that translation the is over and probability 1-F that translation does not terminate alsoX(x) divide
It is not added to their evaluation function l (x)=θ FX(x), not_EOS, l (x)=θ (1-FX(x)), EOS.When candidate word is EOS
When mark, also untranslated good probability is multiplied by penalty factor as penalty term, and when candidate word is not EOS mark, it will be complete
Penalty factor is multiplied by as penalty term at the probability of translation.The obtained penalty term based on lenth ratio is finally added to wave beam
In the evaluation function of search, such as Fig. 5.
By coder-decoder model and beam search, final optimal translation is obtained, as shown in Figure 6.
It should be noted that coding decoder is deep learning frame in conjunction with beam search evaluation function, obtain final
Optimal translation, solve the problems, such as to duplicate segment and omit source statement, it is applied widely, with strong points, translate
Literary quality is higher.
Further, in the coder-decoder model, encoder section and decoder section use bidirectional circulating
Neural network.
It should be noted that bidirectional circulating neural network (Bi-RNN, Bidirectional Recurrent Neural
It Network), is based on a kind of improved network structure of Recognition with Recurrent Neural Network.In some tasks, the input of network not only with mistake
The input gone is related, also has certain association with subsequent input.Therefore it other than inputting positive sequence, also needs to input reverse sequence
Column.And bidirectional circulating neural network is made of two layers of Recognition with Recurrent Neural Network, it supports while inputting positive sequence and inverse
To sequence, the performance of network is effectively improved.
In a kind of implementation of the present invention, it is described by the term vector sequence inputting into coder-decoder model, obtain
The step of obtaining encoding and decoding result, comprising:
By the term vector sequence inputting into coder-decoder model;
Term vector sequence is converted the vector that forms a complete sentence by the deep learning frame based on coding decoder;
It is being based on decoder, sentence vector is converted into term vector sequence.
It is understood that coder-decoder is deep learning frame, encoder is for term vector sequence to be converted into
Sentence vector, decoder are used to sentence vector being converted into term vector sequence.
The present invention also provides a kind of translating equipment of machine translation translation, described device includes processor and passes through
The memory that communication bus is connected to the processor;Wherein,
The memory, for storing the interpretive program of machine translation translation;
The processor, for executing the interpretive program of the machine translation translation, to realize described in any item machines
Translate the translation steps of translation.
The present invention also provides a kind of computer storage mediums to be stored with one or more program, one or more
A program can be executed by one or more processor, so that one or more of processors execute described in any item machines
The translation steps of device translation translation.
The above-described embodiments merely illustrate the principles and effects of the present invention, and is not intended to limit the present invention.It is any ripe
The personage for knowing this technology all without departing from the spirit and scope of the present invention, carries out modifications and changes to above-described embodiment.Cause
This, institute is complete without departing from the spirit and technical ideas disclosed in the present invention by those of ordinary skill in the art such as
At all equivalent modifications or change, should be covered by the claims of the present invention.
Claims (8)
1. a kind of interpretation method of machine translation translation, which is characterized in that the described method includes:
Receive source statement to be translated;
Word segmentation processing is carried out to the source statement;
Obtain the part of speech of each word in the participle;
According to term vector model, the part of speech is incorporated in term vector corresponding to word, fused term vector sequence is obtained;
By the term vector sequence inputting into coder-decoder model, encoding and decoding result is obtained;
For the encoding and decoding as a result, carrying out evaluation of result based on beam search evaluation function, wherein the beam search evaluation
Function includes in the penalty term based on length vs and the penalty term for repeating to detect;
Translation is obtained according to the evaluation result.
2. a kind of interpretation method of machine translation translation according to claim 1, which is characterized in that the beam search is commented
Valence function embodies are as follows:
S (Y, X)=log (P (Y | X))+d (x)+l (x)
Wherein, s (Y, X) is wave beam evaluation function, and x (Y, X) is beam search evaluation function, and log (P (Y/X)) is when X occurs
The probability function that Y occurs, d (x) are based on the penalty term for repeating detection, and l (x) is the punishment based on length vs, and P is distribution letter
Number;
In beam search evaluation function be added the penalty term based on lenth ratio, for solve translation occur part leakage turn over ask
Topic;
It is added in beam search evaluation function based on the penalty term for repeating detection, duplicates asking for content for solving translation
Topic.
3. a kind of interpretation method of machine translation translation according to claim 2, which is characterized in that the repetition detection is punished
The specific formula of a d (x) is penalized to express are as follows:
Wherein, c is the index where current translation of words, and δ, which attaches most importance to, rechecks the range of survey, and ε is penalty coefficient, and y is candidate translation
Corresponding matrix, yc-j, yc-i-jRespectively repeat two matrixes of detection, i, j, to traverse variable.
4. a kind of interpretation method of machine translation translation according to claim 2 or 3, which is characterized in that described to be directed to institute
Encoding and decoding are stated as a result, the step of carrying out evaluation of result based on beam search evaluation function, comprising:
The length of the source statement and the lenth ratio of target translation;
The lenth ratio is fitted by linear regression, obtains Cumulative Distribution Function;
When occurring end of the sentence label and common word in the candidate word of beam search simultaneously, probability F that translation is overX
(x) the probability 1-F not terminated with translationX(x) it is added separately to evaluation function l (x)=θ FX(x), not_EOS, l (x)=θ (1-
FX(x)), EOS, wherein EOS is end of the sentence label, and θ is parameter;
When candidate word is end of the sentence label mark, also untranslated good probability is multiplied by penalty factor as penalty term;
And when candidate word is not end of the sentence label mark, the probability for completing translation is multiplied by penalty factor as penalty term;
The obtained penalty term based on lenth ratio is added in the evaluation function of beam search;
Evaluation of result is carried out based on beam search evaluation function.
5. a kind of interpretation method of machine translation translation according to claim 1, which is characterized in that the encoder decoding
In device model, encoder section and decoder section use bidirectional circulating neural network.
6. a kind of interpretation method of machine translation translation according to claim 1, which is characterized in that it is described by institute's predicate to
The step of sequence inputting is measured into coder-decoder model, obtains encoding and decoding result, comprising:
By the term vector sequence inputting into coder-decoder model;
Term vector sequence is converted the vector that forms a complete sentence by the deep learning frame based on coding decoder;
It is being based on decoder, sentence vector is converted into term vector sequence.
7. a kind of translating equipment of machine translation translation, which is characterized in that described device includes processor and passes through communication always
The memory that line is connected to the processor;Wherein,
The memory, for storing the interpretive program of machine translation translation;
The processor, for executing the interpretive program of the machine translation translation, to realize as any in claim 1 to 6
The translation steps of machine translation translation described in.
8. a kind of computer storage medium, which is characterized in that the computer storage medium is stored with one or more program,
One or more of programs can be executed by one or more processor, so that one or more of processors execute
Such as the translation steps of machine translation translation described in any one of claims 1 to 6.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910721252.6A CN110442880B (en) | 2019-08-06 | 2019-08-06 | Translation method, device and storage medium for machine translation |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910721252.6A CN110442880B (en) | 2019-08-06 | 2019-08-06 | Translation method, device and storage medium for machine translation |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110442880A true CN110442880A (en) | 2019-11-12 |
CN110442880B CN110442880B (en) | 2022-09-30 |
Family
ID=68433418
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910721252.6A Active CN110442880B (en) | 2019-08-06 | 2019-08-06 | Translation method, device and storage medium for machine translation |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110442880B (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112541364A (en) * | 2020-12-03 | 2021-03-23 | 昆明理工大学 | Chinese-transcendental neural machine translation method fusing multilevel language feature knowledge |
CN112632996A (en) * | 2020-12-08 | 2021-04-09 | 浙江大学 | Entity relation triple extraction method based on comparative learning |
CN113191165A (en) * | 2021-07-01 | 2021-07-30 | 南京新一代人工智能研究院有限公司 | Method for avoiding duplication of machine translation fragments |
CN113435215A (en) * | 2021-06-22 | 2021-09-24 | 北京捷通华声科技股份有限公司 | Machine translation method and device |
CN113836950A (en) * | 2021-09-22 | 2021-12-24 | 广州华多网络科技有限公司 | Commodity title text translation method and device, equipment and medium thereof |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2018058046A1 (en) * | 2016-09-26 | 2018-03-29 | Google Llc | Neural machine translation systems |
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 |
-
2019
- 2019-08-06 CN CN201910721252.6A patent/CN110442880B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2018058046A1 (en) * | 2016-09-26 | 2018-03-29 | Google Llc | Neural machine translation systems |
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 (1)
Title |
---|
樊文婷等: "融合先验信息的蒙汉神经网络机器翻译模型", 《中文信息学报》 * |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112541364A (en) * | 2020-12-03 | 2021-03-23 | 昆明理工大学 | Chinese-transcendental neural machine translation method fusing multilevel language feature knowledge |
CN112632996A (en) * | 2020-12-08 | 2021-04-09 | 浙江大学 | Entity relation triple extraction method based on comparative learning |
CN113435215A (en) * | 2021-06-22 | 2021-09-24 | 北京捷通华声科技股份有限公司 | Machine translation method and device |
CN113191165A (en) * | 2021-07-01 | 2021-07-30 | 南京新一代人工智能研究院有限公司 | Method for avoiding duplication of machine translation fragments |
CN113191165B (en) * | 2021-07-01 | 2021-09-24 | 南京新一代人工智能研究院有限公司 | Method for avoiding duplication of machine translation fragments |
CN113836950A (en) * | 2021-09-22 | 2021-12-24 | 广州华多网络科技有限公司 | Commodity title text translation method and device, equipment and medium thereof |
CN113836950B (en) * | 2021-09-22 | 2024-04-02 | 广州华多网络科技有限公司 | Commodity title text translation method and device, equipment and medium thereof |
Also Published As
Publication number | Publication date |
---|---|
CN110442880B (en) | 2022-09-30 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110442880A (en) | A kind of interpretation method, device and the storage medium of machine translation translation | |
Munkhdalai et al. | Neural semantic encoders | |
CN107967262B (en) | A kind of neural network illiteracy Chinese machine translation method | |
CN106484682B (en) | Machine translation method, device and electronic equipment based on statistics | |
Yin et al. | Neural enquirer: Learning to query tables with natural language | |
CN104331449B (en) | Query statement and determination method, device, terminal and the server of webpage similarity | |
CN111191002B (en) | Neural code searching method and device based on hierarchical embedding | |
CN106202153A (en) | The spelling error correction method of a kind of ES search engine and system | |
CN106484681A (en) | A kind of method generating candidate's translation, device and electronic equipment | |
CN110287482B (en) | Semi-automatic participle corpus labeling training device | |
CN115048944B (en) | Open domain dialogue reply method and system based on theme enhancement | |
CN113641830B (en) | Model pre-training method, device, electronic equipment and storage medium | |
CN114528898A (en) | Scene graph modification based on natural language commands | |
CN104572631A (en) | Training method and system for language model | |
CN112818091A (en) | Object query method, device, medium and equipment based on keyword extraction | |
CN113971394A (en) | Text repeat rewriting system | |
Li et al. | Auto completion of user interface layout design using transformer-based tree decoders | |
CN111125323A (en) | Chat corpus labeling method and device, electronic equipment and storage medium | |
CN109992785A (en) | Content calculation method, device and equipment based on machine learning | |
Park et al. | Natural language generation using dependency tree decoding for spoken dialog systems | |
CN111325015B (en) | Document duplicate checking method and system based on semantic analysis | |
Wang | Short Sequence Chinese‐English Machine Translation Based on Generative Adversarial Networks of Emotion | |
CN116432637A (en) | Multi-granularity extraction-generation hybrid abstract method based on reinforcement learning | |
CN115757694A (en) | Recruitment industry text recall method, system, device and medium | |
Wang et al. | Knowledge base question answering system based on knowledge graph representation learning |
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 |
Inventor after: Lin Xinyue Inventor after: Liu Jin Inventor after: Song Junjie Inventor before: Lin Xinyue Inventor before: Liu Jin |
|
CB03 | Change of inventor or designer information | ||
GR01 | Patent grant | ||
GR01 | Patent grant |