JP6965951B2 - Training methods, devices and storage media for neural machine translation models - Google Patents

Description

The present invention relates to the field of neural machine translation technology in natural language processing (NLP), and specifically to training methods, devices and storage media for neural machine translation models.

Neural Machine Translation (NMT) refers to a machine translation method that uses a neural network directly to perform end-to-end translation modeling. Unlike traditional methods of perfecting certain modules in statistical machine translation using deep running techniques, neural machine translation uses a simple and intuitive method to accomplish the translation task. First, a neural network called an encoder is used to encode the source language sentence into a dense vector, and then a neural network called a decoder is used to decode the target language sentence from the vector. The neural network model is commonly referred to as an "encoder-decoder" structure.

Conventional techniques often use a bilingual mutual translation quality evaluation (BLEU) algorithm to evaluate machine translation quality. The design concept of the BLEU algorithm is consistent with the idea of judging the quality of machine translation. That is, the closer the machine translation result is to the result of professional artificial translation, the better the translation quality. N-gram is a statistical language model, which can display one sentence as a word string composed of N consecutive words. By calculating the probability of a sentence using the combination information between adjacent words in the context, it is judged whether or not the logic of this one sentence is appropriate. The BLEU algorithm uses N-gram matching rules. Thereby, the occupancy rate similar to N grams in the predicted translation and the reference translation can be calculated, and the evaluation index of the machine translation quality can be obtained.

Currently, common NMT models include the series-series (seq2seq) model, the convolution series-series (convS2S) model, and the transformer model. These prior arts improve machine translation performance by improving the neural machine model itself. Therefore, further improving the translation performance of conventional neural machine translation and realizing more accurate translation between the source language and the target language is a technical problem to be solved immediately in this technical field. ..

In view of the above technical problems, the embodiment of the present invention provides a training method, an apparatus and a storage medium for a neural machine translation model, and improves the translation performance of the neural machine translation.

In order to solve the above technical problem, an embodiment of the present invention is a method for training a neural machine translation model, which is a step of calculating the frequency of occurrence of N grams in a target sentence corpus, and is a step of calculating the frequency of occurrence of the target sentence corpus. Contains a plurality of target sentences, N is 2 or more, a step and a high frequency N gram whose appearance frequency is higher than a predetermined threshold are selected from the N gram, and are present in the target sentence by a predetermined delimiter. A step of synthesizing the high-frequency N-grams into one integrated word to obtain an updated target sentence corpus, and a step of training a neural machine translation model using the source sentence corpus and the updated target sentence corpus. And, a training method characterized by including.

Preferably, after training the neural machine translation model, the trained neural machine translation model is used to translate the sentence to be translated to obtain the predicted sentence, and the integrated word is present in the predicted sentence. In the case, the step of outputting the predicted sentence after resetting the integrated word existing in the predicted sentence to N separate words, and the predicted sentence when the integrated word does not exist in the predicted sentence. Further includes a step of outputting as it is.

Preferably, the step of resetting the integrated word present in the predicted sentence to N separate words is the integrated word present in the predicted sentence based on a predetermined delimiter in the integrated word. Includes a step of dividing into and retrieving N words.

Preferably, the step of training the neural machine translation model using the source sentence corpus and the updated target sentence corpus uses a parallel corpus consisting of the source sentence in the source sentence corpus and the target sentence corresponding to the source sentence. , The integrated word present in the target sentence, including the step of training the neural machine translation model, is not divided in the training.

Preferably, the N grams are 2 grams, 3 grams or 4 grams.

Preferably, the neural machine translation model is a seq2seq model, a convS2S model, or a transformer model.

Further, an embodiment of the present invention is a training device for a neural machine translation model, which is a frequency calculation unit for calculating the appearance frequency of N grams in a target sentence corpus, and the target sentence corpus includes a plurality of target sentences. , N is 2 or more, the frequency calculation unit and the high frequency N gram whose appearance frequency is higher than a predetermined threshold are selected from the N gram, and the high frequency N gram existing in the target sentence by a predetermined delimiter. A word set unit that synthesizes into one integrated word and obtains an updated target sentence corpus, and a model training unit that trains a neural machine translation model using the source sentence corpus and the updated target sentence corpus. , Including training equipment.

Preferably, the training device uses a neural machine translation model trained by the model training unit to translate a sentence to be translated to obtain a predicted sentence, and the integrated word is present in the predicted sentence. After resetting the integrated word existing in the predicted sentence to N separate words, the predicted sentence is output, and if the integrated word does not exist in the predicted sentence, the predicted sentence is output as it is. Also includes a translation unit, which does.

Preferably, in the training device, the translation unit divides the integrated word existing in the predicted sentence based on a predetermined delimiter in the integrated word, and acquires N words. ..

Preferably, in the training apparatus, the model training unit trains a neural machine translation model using a parallel corpus consisting of a source sentence in the source sentence corpus and a target sentence corresponding to the source sentence, and is present in the target sentence. The united word to be used is not divided in training.

Preferably, in the training device, the N grams are 2 grams, 3 grams or 4 grams.

Preferably, in the training apparatus, the neural machine translation model is a seq2seq model, a convS2S model, or a transformer model.

Further, an embodiment of the present invention provides a training device for a neural machine translation model, which includes a memory, a processor, and a computer program stored in the memory, and when the computer program is executed by the processor, the above Provided is an apparatus characterized in that steps in a method of training a neural machine translation model are realized.

Finally, an embodiment of the present invention is a computer-readable storage medium in which a computer program is stored, and when the computer program is executed by a processor, the steps of the training method of the neural machine translation model are realized. Provided is a storage medium characterized by being used.

Compared with the prior art, the training method, apparatus and storage medium of the neural machine translation model provided by the embodiment of the present invention have a higher frequency of occurrence in the target sentence corpus than a predetermined threshold in the training process of the neural machine translation model. By training the high-frequency N-grams as indivisible words, the actual translation is performed using the trained neural machine translation model, and the high-frequency N-grams contained in the obtained predicted sentence are divided into N words. This will give you a translation result that contains more N grams of correctly translated words. As a result, the scoring result of the neural machine translation is improved, and the machine translation quality is improved.

In order to more clearly explain the technical solutions of the embodiments of the present invention, the drawings necessary for the description of the embodiments of the present invention will be briefly described below, and as is clear, the drawings in the following description are the present invention. It is only a few examples of the invention, and for those skilled in the art, it is possible to obtain other drawings based on these drawings without any creative effort.
FIG. 1 is a flowchart showing a training method of a neural machine translation model according to an embodiment of the present invention. FIG. 2 is another flowchart showing a training method of the neural machine translation model according to the embodiment of the present invention. FIG. 3 is a diagram schematically showing the structure of the training device of the neural machine translation model according to the embodiment of the present invention. FIG. 4 is a diagram schematically showing another structure of the training device of the neural machine translation model according to the embodiment of the present invention. FIG. 5 is a diagram schematically showing still another structure of the training device of the neural machine translation model according to the embodiment of the present invention.

In order to clarify the technical problems, technical solutions and advantages to be solved by the present invention, the following will be described in detail with reference to the drawings and specific examples. In the following description, specific details, such as specific configurations and components, are provided to aid in a complete understanding of embodiments of the present invention. Therefore, it will be apparent to those skilled in the art that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the invention. In addition, the description of known functions and configurations will be omitted for the sake of brevity.

"One Example" or "One Example" as referred to throughout the specification means that a particular feature, structure or property associated with an example is included in at least one embodiment of the present invention. It should be understood to do. Therefore, the terms "in one embodiment" or "in one embodiment" that appear throughout the specification do not necessarily refer to the same embodiment. Moreover, these particular features, structures and properties can optionally be incorporated into one or more embodiments in any suitable manner.

In various embodiments of the present invention, it should be understood that the magnitude of each process number below does not mean before or after the order of execution. The execution order of each process is determined by its function or its own logic, and does not limit the execution process according to the embodiment of the present invention.

FIG. 1 is a flowchart showing a training method of a neural machine translation model provided by an embodiment of the present invention. The training method of the neural machine translation model can improve the translation performance of the neural machine translation model obtained by training. Specifically, the neural machine translation model is a series-series (seq2seq) model, a convolution series-series (convS2S) model, or a transformer model. Of course, the embodiments of the present invention can still apply other types of neural machine translation models. The present invention does not limit this in detail. As shown in FIG. 1, the training method of the neural machine translation model provided by the embodiment of the present invention includes the following.

In step 101, the frequency of occurrence of N grams in the target sentence corpus is calculated, the target sentence corpus contains a plurality of target sentences, and the N is 2 or more.

In the process of training a neural machine translation model, the training corpus generally includes a source sentence corpus and a target sentence corpus. A source sentence corpus contains source sentences in multiple source languages, and a target sentence corpus contains target sentences in multiple target languages. Each source sentence has a target sentence corresponding to the corresponding source sentence, and both form one parallel corpus. In the embodiment of the present invention, in step 101 above, the frequency of occurrence of various N-grams in the target sentence corpus is calculated. For example, suppose a corpus of target sentences contains 1 million target sentences, and one N-gram appears 20,000 times in total in these target sentences, and the frequency of occurrence of the N-gram is 2/100 = 0.02. As a matter of course, the frequency of appearance can be statistic according to the number of occurrences, and in this case, the frequency of appearance of the N-gram is 200,000 times.

Here, you can refer to the related description of the prior art for the concept of N-grams. Generally, the N-gram may be N consecutive words or N consecutive words and punctuation marks in the corresponding sentence. These words and punctuation may be continuous in the sentence, and detailed description is omitted here in order to save space. Preferably, N is an integer greater than or equal to 2 and can take a value of, for example, 2, 3 or 4. Of course, it may be another larger number. As a preferred embodiment, the BLEU algorithm generally uses 4 grams to evaluate machine translation performance, so in the examples of the present invention, N grams is preferably 4 grams. For example, if the target language is English, the target sentence is "it is said that it will rain tomorrow", and Ngram is "it is said that", "is said that it", and "said that". "it will" ... "it will rain tomorrow" and so on. In the embodiment of the present invention, for "it is said that", the frequency of occurrence of the 4 grams in all target sentences of the target sentence corpus is calculated.

In step 102, one of the high-frequency N-grams present in the target sentence is selected from the N-grams with a high-frequency N-gram whose appearance frequency is higher than a preset threshold value, and via a predetermined delimiter. Obtain an updated target sentence corpus by constructing it as a united word.

Here, the embodiment of the present invention determines whether or not N-gram is a high-frequency N-gram based on whether or not the appearance frequency of N-gram is higher than a preset threshold value. A high frequency N-gram means that the N-gram appears frequently in the target sentence corpus. For this reason, the N-grams are often used together. In consideration of the above factors, the embodiment of the present invention constructs a high-frequency N-gram as one indivisible integrated word. The indivisible means prohibiting the integrated word from being divided into smaller subwords during the training process of the neural machine translation model.

In order to facilitate the identification of the integrated word consisting of the high frequency N gram during model training, in the embodiment of the present invention, each word in the high frequency N gram is concatenated by using a predetermined delimiter. , Form one united word. For example, "@_" can be used as a delimiter to connect each word in the Ngram. Taking the above "it is said that" as an example, the above delimiter gives the unified word "it @ _is @ _said @ _that". Through the above processing, the embodiment of the present invention realizes the update of the target sentence corpus by using the high frequency N grams existing in each target sentence in the target sentence corpus as an integrated word. Of course, if a high frequency N-gram is not present in a target sentence, the above process does not need to be performed.

Note that the word division algorithm adopted in the subsequent training process of the neural machine translation model may use a specific delimiter. The predetermined delimiter in step 102 needs to be distinguished from the delimiter used in the word division algorithm, i.e., a different delimiter is used. For example, taking the Byte Pair Encoder (BPE) algorithm as an example, if "@@" is used as a delimiter in BPE and the subsequent training model adopts the BPE algorithm, a predetermined delimiter is used in step 102. The symbol must use a delimiter different from "@@".

In step 103, the source sentence corpus and the updated target sentence corpus are used to train the neural machine translation model.

In step 103 above, an embodiment of the invention utilizes an updated target sentence corpus and an original source sentence corpus to train a neural machine translation model and trained final neural for translation from source language to target language. Get a machine translation model.

In the above training process, the embodiment of the present invention trains a neural machine translation model using a parallel corpus consisting of a source sentence in the source sentence corpus and a target sentence corresponding to the source sentence. Among them, when the target sentence contains an integrated word composed of high-frequency N-grams, the division of the integrated word is prohibited in the training process. That is, in the training process, if the integrated word is present in the target sentence, the integrated word is not further divided.

The embodiment of the present invention can be applied to a neural machine translation model. For example, in step 103, the neural machine translation model is a sequence-series (seq2seq) model, a convolution sequence-series (convS2S) model, or a transformer model. Other types of neural machine translation models can be applied, but the present invention does not limit this in detail.

According to the above steps, in the embodiment of the present invention, the neural machine translation model is trained using the integrated word composed of the high-frequency N-grams, and the high-frequency N-grams are trained as one in the training process. Guaranteed. The neural machine translation model trained in this way improves the scoring result of the neural machine translation and improves the machine translation quality by obtaining the translation result containing many high-frequency N-grams at the time of actual translation.

In the method for training the neural machine translation model according to the embodiment of the present invention, a neural machine translation model having higher translation performance can be obtained by the above steps 101 to 103. Subsequent translation from the source language to the target language can be performed using the trained neural machine translation model.

FIG. 2 shows a training method of a neural machine translation model provided by an embodiment of the present invention. Following step 103, the following steps are further included.

In step 104, the trained neural machine translation model is used to translate the translation sentence to be translated to obtain the predicted sentence.

In step 105, when the integrated word is present in the predicted sentence, the predicted sentence is output after resetting the integrated word existing in the predicted sentence to N words to be separated.

In step 106, if the integrated word does not exist in the predicted sentence, the predicted sentence is output as it is.

In step 104 above, translation is performed using the neural machine translation model obtained in step 103, and the translated predicted sentence may include an integrated word composed of high frequency N grams. For this reason, an embodiment of the present invention is in step 105 to further divide the integrated words that may be present therein. Specifically, based on a predetermined delimiter in the integrated word, a division point between adjacent words is determined, and the integrated word existing in the predicted sentence is further divided into N You get a word. Of course, if the integrated word does not exist in the predicted sentence, the predicted sentence is output as it is in step 106.

Through the above steps, the embodiment of the present invention realizes a translation application by a trained neural machine translation model. The pre-trained neural machine translation model improves the score result of neural machine translation and improves the machine translation quality by obtaining the translation result containing more N grams at the time of actual translation.

Based on the above method, the examples of the present invention further provide an apparatus for carrying out the above method. As shown in FIG. 3, the training device 300 for the neural machine translation model according to the embodiment of the present invention includes the following units.

The frequency calculation unit 301 calculates the frequency of occurrence of N grams in the target sentence corpus, and the target sentence corpus includes a plurality of target sentences, and the N is 2 or more.

The word set unit 302 selects a high frequency N gram from the N gram whose frequency of occurrence is higher than a preset threshold value, and uses a predetermined delimiter to select the high frequency N gram existing in the target sentence. Obtain an updated target sentence corpus by constructing it as a single integrated word.

The model training unit 303 trains the neural machine translation model using the source sentence corpus and the updated target sentence corpus.

With the above unit, the neural machine translation model training device 300 according to the embodiment of the present invention trains the neural machine translation model by using the integrated word composed of high frequency N grams, and is high in the training process. Frequency N grams are guaranteed to train as one. The neural machine translation model trained in this way improves the scoring result of the neural machine translation and improves the machine translation quality by obtaining the translation result containing many high-frequency N-grams at the time of actual translation.

Preferably, the model training unit 303 trains a neural machine translation model using a parallel corpus consisting of a source sentence in the source sentence corpus and a target sentence corresponding to the source sentence, in which the model training unit 303 is present in the target sentence. The division of the integrated word is prohibited in the training process.

Preferably, the N grams are 2 grams or 3 grams or 4 grams.

Preferably, the neural machine translation model is a series-series (seq2seq) model or a convolution series-series (convS2S) model or a transformer model.

Preferably, as shown in FIG. 4, the training device 300 of the neural machine translation model further translates the translation sentence to be translated by using the neural machine translation model trained by the model training unit 303. Obtain the predicted sentence; if the integrated word is present in the predicted sentence, after resetting the integrated word existing in the predicted sentence to N words that separate the integrated word, the predicted sentence is output; the predicted sentence is output. Includes a translation unit 304 that outputs the predicted sentence as it is when the integrated word does not exist in.

Preferably, the translation unit 304 divides the integrated word existing in the predicted sentence based on a predetermined delimiter in the integrated word to obtain N words.

FIG. 5 is a block diagram showing an example of the hardware configuration of the training device of the neural machine translation model according to the embodiment of the present invention. As shown in FIG. 5, the training device 500 of the neural machine translation model includes a processor 502 and a memory 504 in which computer program commands are stored. When the computer program command is executed by the processor 502, the frequency of occurrence of N grams in the target sentence corpus is calculated, the target sentence corpus contains a plurality of target sentences, and the N is 2 or more; the N grams. Select a high-frequency N-gram whose appearance frequency is higher than a preset threshold, and construct the high-frequency N-gram existing in the target sentence as one integrated word via a predetermined delimiter. Obtains an updated target sentence corpus; the source sentence corpus and the updated target sentence corpus are used to perform steps to train a neural machine translation model.

Further, as shown in FIG. 5, the training device 500 of the neural machine translation model further includes a network interface 501, an input device 503, a hard disk 505, and a display device 506.

Each of the above interfaces is connected to each device via a bus architecture. A bus architecture is a bus and bridge that can contain any number of interconnects. Specifically, one or more central processing units (CPUs) represented by the processor 502 and various circuits of one or more memories represented by the memory 504 are connected. Also, from the bus architecture, various other circuits such as external devices, regulators and power management circuits are connected. In this way, these devices are communicably connected by the bus architecture. In addition to the data bus, the bus architecture includes a power bus, a control bus, and a status signal bus. These are known techniques in the field of the present invention, and detailed description thereof will be omitted in the text.

The network interface 501 is an interface that is connected to a network (for example, the Internet or a LAN), collects a source sentence corpus and a target sentence corpus from the network, and stores them in the hard disk 505.

The input device 503 is a means for receiving various commands input from the user, transmitting the commands to the processor 502, and executing the commands. Further, the input device 503 includes a keyboard, a clicking means (for example, a mouse, a trackball, a touch board) and the like.

The display device 506 is a means for displaying the result of executing the command by the processor 502. For example, the progress of model training and the translation result of the sentence to be translated are displayed.

The memory 504 is a memory that stores programs and data necessary for executing the operating system (OS), and data such as intermediate results in the calculation process of the processor 502.

The memory 504 according to the embodiment of the present invention is a volatile memory or a non-volatile memory, or a memory including both volatile and non-volatile. Among them, non-volatile memory is ROM, PROM, EPROM, EEPROM, flash memory. Volatile memory is RAM and is used as an external cache. However, the memory 504 used in the devices and methods described herein is not limited to these memories and may be other suitable type of memory.

In some embodiments, the memory 504 stores an executable module or data configuration or submodules or extension modules thereof, an operating system (OS) 5041 and an application program (APP) 5042.

Among them, the operating system 5041 includes various system programs such as a framework layer, a core library layer, and a driving layer, and is used to realize various mission-critical tasks and hardware-based tasks. The application program 442 includes various application programs such as a web browser (Browser), and is for realizing various application operations. A program that executes the method according to this embodiment is included in the application program 5042.

The method according to the embodiment of the present invention is applied to or realized by the processor 402. Processor 502 is an integrated circuit board capable of processing signals. Each step of the above method is realized by an integrated logic circuit or software-type command which is hardware in the processor 502. The processor 502 is a general purpose processor, digital signal processing device (DSP), dedicated integrated circuit (ASIC), off-the-shelf programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware component. Each method, step and logic box disclosed in the examples of the present invention can be realized or implemented. The general-purpose processor may be a microprocessor or any general processor. Each step of the method according to the embodiment of the present invention may be realized by being executed by a decoder which is hardware, or may be realized by a combination of hardware and software which can go to the decoder. Software modules are stored in storage media mature in the art, such as random memory, flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, registers. From memory 504, which includes a storage medium in which the software is stored, processor 502 reads information and implements the steps of the above method in accordance with the hardware.

The embodiments described above are realized by hardware, software, firmware, middleware, microcode, or a combination thereof. Among them, regarding the realization of hardware, the processing unit is one or more dedicated integrated circuits (ASIC), digital signal processing processor (DSP), digital signal processing device (DSPD), programmable logic circuit (PLD), field. It is implemented by programmable gate arrays (FPGAs), general purpose processors, controllers, microcontrollers, microprocessors, other electronic units that perform the functions of the present invention, or a combination thereof.

Further, regarding the realization of software, the above technology is realized by a module (for example, a process, a function, etc.) that realizes the functions described above. The software code is stored in memory and executed by the processor. The memory is realized inside or outside the processor.

Specifically, when the computer program is executed by the processor 502, after training the neural machine translation model, the trained neural machine translation model is used to translate the translation sentence to be translated to obtain the predicted sentence. When the integrated word is present in the predicted sentence, the integrated word existing in the predicted sentence is reset to N words to be separated, and then the predicted sentence is output; the integrated word is output in the predicted sentence. If the word does not exist, the step of outputting the predicted sentence as it is is included.

Specifically, when the computer program is executed by the processor 502, N pieces are divided into the integrated words existing in the predicted sentence based on a predetermined delimiter in the integrated word. Includes steps to get the word.

Specifically, when the computer program is executed by processor 502, a neural machine translation model is trained using a parallel corpus consisting of a source sentence in the source sentence corpus and a target sentence corresponding to the source sentence, in which the neural machine translation model is trained. , The division of the integrated word present in the target sentence is prohibited in the training process.

Preferably, the N grams are 2 grams or 3 grams or 4 grams.

Preferably, the neural machine translation model is a series-series (seq2seq) model or a convolution series-series (convS2S) model or a transformer model.

Those skilled in the art of the present invention will appreciate that the unit and algorithm steps of each of the examples described above are realized in electronic hardware, or in combination with computer software and electronic hardware. Is easily conceived. Performing these functions in either hardware or software depends on the specific application and design constraints of the invention. Those skilled in the art can achieve the above functions in a manner according to a particular application, but should not go beyond the scope of the present invention.

Also, for convenience and brevity for explanation, it will be apparent to those skilled in the art that the specific work processes of the above systems, devices and units can be referred to in the corresponding processes in the above examples. A detailed description will be omitted.

In the embodiments provided in the present application, it should be understood that the disclosed devices and methods may be implemented in other ways. For example, the device embodiment described above is a schematic one. For example, the division of the unit is only a logical function division, and it is possible to have another division method when actually realizing the unit, for example, a plurality of divisions. Units or components can be combined or integrated, or some features can be ignored or not implemented. Also, the coupling or direct coupling or communication connection between the displayed or discussed may be an indirect coupling or communication connection of some interface, device or unit, may be electrical, mechanical or other. It may be in the form of.

The units described as separate parts may or may not be physically separate. The parts displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed among a plurality of network units. Some or all of the units can be selected according to the actual needs to realize the solution of the embodiment of the present invention.

Each functional unit according to the embodiment of the present invention may be integrated into one processing unit, physically alone, or two or more as one unit.

The function is realized in the form of a software functional unit and can be stored in a computer-readable storage medium when sold or used as an independent product. In this case, the technical plan of the present invention is essentially, or a part that contributes to the prior art or a part of the technical plan is expressed in the form of a software product. The computer software product is stored in a storage medium and causes a computer device (personal computer, server, network device, etc.) to perform all or part of the steps of the method according to each embodiment of the present invention. Includes directives. The above-mentioned storage medium includes various media such as a USB memory, a removable disk, a ROM, a RAM, a magnetic disk, or an optical disk, which can store a program code.

As described above, it is only a specific embodiment of the present invention, and the scope of protection of the present invention is not limited to this, and can be easily changed or replaced within the technical scope disclosed by those skilled in the art. And all should be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims (11)

A training method for neural machine translation models
A step of calculating the frequency of occurrence of N grams in a target sentence corpus, wherein the target sentence corpus contains a plurality of target sentences and N is 2 or more.
A high-frequency N-gram having an appearance frequency higher than a predetermined threshold was selected from the N-gram, and the high-frequency N-gram existing in the target sentence was synthesized into one integrated word by a predetermined delimiter and updated. Steps to get the target sentence corpus and
Steps to train a neural machine translation model using the source sentence corpus and the updated target sentence corpus,
Using a trained neural machine translation model, the steps to translate the sentence to be translated and obtain the predicted sentence,
When the integrated word is present in the predicted sentence, the step of outputting the predicted sentence after resetting the integrated word existing in the predicted sentence to N separate words.
A training method comprising a step of outputting the predicted sentence as it is when the integrated word does not exist in the predicted sentence. The step of resetting the integrated word existing in the predicted sentence to N separate words is
Based on a predetermined delimiter in said integral words, the performed division to the integrated word present in the predicted sentence, the step of obtaining N pieces of word, in claim 1, characterized in that it comprises Described training method. The steps to train a neural machine translation model using the source sentence corpus and the updated target sentence corpus are:
Including a step of training a neural machine translation model using a parallel corpus consisting of a source sentence in the source sentence corpus and a target sentence corresponding to the source sentence.
The training method according to claim 1, wherein the integrated word existing in the target sentence is not divided in the training. The training method according to any one of claims 1 to 3 , wherein the N gram is 2 gram, 3 gram or 4 gram. The training method according to claim 4 , wherein the neural machine translation model is a seq2seq model, a convS2S model, or a transformer model. A training device for neural machine translation models
A frequency calculation unit that calculates the frequency of occurrence of N grams in a target sentence corpus, wherein the target sentence corpus contains a plurality of target sentences and N is 2 or more.
A high-frequency N-gram having an appearance frequency higher than a predetermined threshold was selected from the N-gram, and the high-frequency N-gram existing in the target sentence was synthesized into one integrated word by a predetermined delimiter and updated. With a word set unit to get the target sentence corpus,
A model training unit that trains neural machine translation models using the source sentence corpus and the updated target sentence corpus,
Using the neural machine translation model trained by the model training unit, the sentence to be translated is translated to obtain the predicted sentence, and when the integrated word is present in the predicted sentence, the said sentence existing in the predicted sentence. Training including a translation unit that outputs the predicted sentence after resetting the integrated word to N separate words, and outputs the predicted sentence as it is when the integrated word does not exist in the predicted sentence. Device. Said translation unit, based on a predetermined delimiter in said integral words, the performed division to the integrated word present in the predicted sentence, claim and acquires N pieces of word 6 The training device described in. The model training unit trains a neural machine translation model using a parallel corpus consisting of a source sentence in the source sentence corpus and a target sentence corresponding to the source sentence.
The training device according to claim 6 , wherein the integrated word present in the target sentence is not divided in training. The training device according to any one of claims 6 to 8 , wherein the N gram is 2 gram, 3 gram or 4 gram. The training device according to claim 9 , wherein the neural machine translation model is a seq2seq model, a convS2S model, or a transformer model. The step of the method for training a neural machine translation model according to any one of claims 1 to 5 , when the computer program is stored in a computer-readable storage medium and the computer program is executed by the processor. A storage medium characterized in that

Images