Machine translation system

Abstract

To improve accuracy of machine translation, obtain a translated document with high accuracy, and faithfully translate a translation target document.SOLUTION: A system includes the steps of: generating a first translated document by translating a translation target document using a neural network; determining a phrase to be corrected among phrases in the translation target document on the basis of an analysis result of the first translated document; revising the translation target document by replacing the phrase to be corrected with a high-frequency word in training data used to train the neural network;.and generating a second translated document by translating the revised translation target document using a neural network.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] One aspect of the present invention relates to a machine translation method, a machine translation system, a program, and a non-temporary computer-readable storage medium. Note that one aspect of the present invention is not limited to the above technical field. Examples of the technical field of one aspect of the present invention include semiconductor devices, display devices, light-emitting devices, power storage devices, storage devices, electronic devices, lighting devices,

[0002] input devices (e.g., touch sensors, etc.), input / output devices (e.g., touch panels, etc.), their driving methods, or their manufacturing methods.

Background Art

[0003] Research and development of machine translation, which translates a certain natural language into another natural language using a computer, is actively underway. Examples of machine translation include rule-based machine translation that performs translation based on rules, statistical machine translation that performs translation using language models and translation models, and neural machine translation that performs translation using a neural network.

[0004] In statistical machine translation, learning is performed through multiple steps using various tools. Therefore, it is necessary to prepare learning data according to each tool. On the other hand, neural machine translation performs learning with a single neural network, so the necessary learning data is only a parallel corpus of the source document and the translated document. Therefore, learning data is easily obtained, and the labor of artificially creating learning data can be reduced. And neural machine translation tends to achieve higher translation accuracy than rule-based machine translation and statistical machine translation. Therefore, the practical application of neural machine translation is progressing.

[0005] However, machine translation is not yet perfect, and mistranslations occur. Patent Document 1 describes the occurrence of mistranslations. Translation methods aimed at preventing such errors have been disclosed. [Prior art documents] [Patent Documents]

[0006] [Patent Document 1] Japanese Patent Publication No. 2019-20950 [Overview of the project] [Problems that the invention aims to solve]

[0007] One aspect of the present invention aims to improve the accuracy of machine translation. One of the challenges of this invention is to obtain highly accurate translated documents through machine translation. One of the objectives of this approach is to obtain a translation result that is faithful to the original document using machine translation. .

[0008] One aspect of the present invention aims to provide a highly accurate machine translation system. One aspect of the invention aims to provide a highly accurate machine translation method.

[0009] One aspect of the present invention is a machine translation system that can faithfully translate a document to be translated. One of the objectives is to provide a solution. One aspect of the present invention provides a solution that faithfully translates the document to be translated. One of the objectives is to provide a machine translation method that can perform the following tasks.

[0010] Furthermore, the description of these problems does not preclude the existence of other problems. One aspect of the present invention is It is not necessarily required to resolve all of these issues. Specifications, drawings, invoices. From the description of the items, it is possible to extract problems other than these.

Means for Solving the Problems

[0011] One aspect of the present invention is to translate a source document using a neural network to generate a first translation document, determine a target sentence to be corrected from the sentences included in the source document based on the analysis result of the first translation document, and replace the target sentence to be corrected with a high-frequency word in the training data used for the learning of the neural network to correct the source document, and then translate the corrected source document using the neural network to generate a second translation document. This is a machine translation method.

[0012]

[0013]

[0014]

[0015]

[0016]

[0017]

[0018]

[0019]

[0020]

[0021]

[0022]

[0023]

[0024]

[0025]

[0026] [Figure 1] [Figure 2] [Figure 3] [Figure 4] [Figure 5]

[0027]

[0028]

[0029]

[0030]

[0031] [[ID=15G]]

[0032]

[0033]

[0034]

[0035]

[0036]

[0037]

[0038] [[ID=T88]]

[0039] By translating the translated document with the latest corrections reflected using a neural network, a translated document based on the latest corrections is generated. This includes the step of generating a translated document.

[0014] In the step of determining the sentence to be corrected, from within the translated document, the sentence to be corrected is determined, and preferably, the word or phrase to be corrected is determined from among the words or phrases included in the sentence in the translated document corresponding to the sentence to be corrected. This is preferred.

[0015] High-frequency words are preferably selected from among synonyms and similar words of the word or phrase to be corrected.

[0016] One aspect of the present invention is a document search system having a function of performing at least any one of the above machine translation methods.

[0017] One aspect of the present invention is a machine translation system having a processing unit, and the processing unit has a function of generating a first translated document by translating a translated document using a neural network, a function of determining a word or phrase to be corrected from among the words or phrases included in the translated document based on the analysis result of the first translated document, a function of correcting the translated document by replacing the word or phrase to be corrected with a high-frequency word in the learning data used for the learning of the neural network, and a function of generating a second translated document by translating the corrected translated document using a neural network.

[0018] One aspect of the present invention is a program having a function of causing a processor to execute at least any one of the above machine translation methods. One aspect of the present invention is a non-temporary computer-readable storage medium in which the program is stored.

[0019] The program is stored on a computer using various types of temporary computer-readable storage media. It may be supplied to. Examples of temporary computer-readable storage media include electrical signals and optical signals. There are numbers and electromagnetic waves. Temporary computer-readable storage media include electric wires and optical fibers, etc. The program can be supplied to the computer via a wired communication channel or a wireless communication channel.

[0020] One aspect of the present invention involves translating a document to be translated using a neural network, The steps include generating a translated document and, based on the analysis results of the first translated document, including the contents of the document to be translated. The steps involve determining which words to modify from among the words that are generated, and then using neural networks to process the words to be modified. By replacing the words in the training data used to train the network with high-frequency words, the translated text The process involves correcting the document and then translating the corrected document using a neural network. This involves a program that causes the processor to execute the steps of generating a second translated document. One aspect of the present invention is a non-temporary computer-readable memory in which the program is stored. It is a medium.

[0021] As non-temporary computer-readable storage media, various types of tangible storage media can be used. It is possible. As a non-temporary computer-readable storage medium, for example, RAM (Random Access Memory) Volatile memory such as DOM Access Memory, ROM (Read Only) Examples of non-volatile memory include hard disk drives. (Hard Disk Drive: HDD) and Solid State Drive (Soli d State Drive (SSD) and other recording media drives, magneto-optical disks, C Examples include D-ROMs and CD-Rs. [Effects of the Invention]

[0022] According to one aspect of the present invention, the accuracy of machine translation can be improved. According to one aspect of the present invention, A highly accurate translated document can be obtained. According to one aspect of the present invention, machine translation is used This allows you to obtain translation results that are faithful to the original document.

[0023] According to one aspect of the present invention, a highly accurate machine translation system can be provided. Therefore, we can provide a highly accurate machine translation method.

[0024] According to one aspect of the present invention, a machine translation system that can faithfully translate a document to be translated is It is possible to provide a translation. According to one aspect of the present invention, it is possible to translate the document to be translated faithfully. We can provide a machine translation method.

[0025] Furthermore, the description of these effects does not preclude the existence of other effects. One aspect of the present invention is It is not necessarily required to have all of these effects. It is possible to extract effects other than those listed above. [Brief explanation of the drawing]

[0026] [Figure 1] Figure 1 is a flowchart illustrating an example of a machine translation method. [Figure 2] Figure 2 is a schematic diagram illustrating an example of a machine translation method. [Figure 3] Figure 3 is a flowchart illustrating an example of a machine translation method. [Figure 4] Figure 4 is a block diagram showing an example of a machine translation system. [Figure 5] Figure 5 is a block diagram showing an example of a machine translation system. [Modes for carrying out the invention]

[0027] Embodiments will be described in detail with reference to the drawings. However, the present invention is not limited to the following description. Without departing from the spirit and scope of the present invention, its form and details may be modified in various ways. It will be easily understood by those skilled in the art to obtain this. Therefore, the present invention is as shown in the embodiments below. The interpretation is not limited to the content stated herein.

[0028] In the configuration of the invention described below, the same part or part having a similar function is used. The same symbol is used consistently across different drawings, and explanations of its repetition are omitted. When referring to a function, the same hatch pattern may be used, and a specific symbol may not be assigned.

[0029] Furthermore, the position, size, and extent of each component shown in the drawings are, for the sake of ease of understanding, actually The location, size, and range may not be described. Therefore, the disclosed invention is not always Furthermore, it is not limited to the location, size, scope, etc., disclosed in the drawings.

[0030] (Embodiment 1) In this embodiment, a machine translation method according to one aspect of the present invention will be explained using Figures 1 to 3. .

[0031] In one aspect of the present invention, a machine translation method is used when translating a document to be translated and generating a translated document. We will use artificial intelligence (AI). Specifically In one aspect of the present invention, a machine translation method is used, which involves an artificial neural network (ANN:Art Physical Neural Network, hereafter simply referred to as Neural Network (Also noted) uses a neural network. A neural network is a circuit (hardware) or program. This is achieved through (software).

[0032] In this specification, a neural network is defined as a network that mimics the neural network of a living organism and learns This refers to a general set of models that determine the strength of connections between neurons and give them problem-solving abilities. A neural network has an input layer, an intermediate layer (hidden layer), and an output layer.

[0033] In this specification, when describing neural networks, we will use existing information to... The process of determining the strength of the connection between neurons (also called weight coefficients) is called "learning." There are cases where this is the case.

[0034] In this specification, etc., neural networks are constructed using connection strengths obtained through learning. The process of constructing something and deriving new conclusions from it is sometimes called "inference." In one aspect of this invention... In this context, translating a document using a neural network is equivalent to inference. It can be said that...

[0035] In this specification and elsewhere, translation using neural networks is referred to as "neural machine translation." It is sometimes called that.

[0036] In this case, neural machine translation can sometimes produce mistranslations.

[0037] Training a neural network requires a large number of sets of documents to be translated and documents to be translated. Training data (also called a training corpus, etc.) is used. If the document being translated contains words or phrases that are not suitable for neural machine translation, the machine translation will translate them appropriately. There are some things I can't do.

[0038] Furthermore, in neural networks, the higher the dimensionality of the input and output layers, the more complex they become. The computational complexity becomes enormous. Therefore, each language used during training (the language of the document to be translated, the translated text) The vocabulary of the written language is limited to a certain number. Specifically, the vocabulary used during learning is limited to learning Limited to words that are frequently used in the data (hereinafter also referred to as high-frequency words), training data Words that are not used very often (hereinafter also called low-frequency words) are treated as unknown words and specially marked. It is replaced with a number. Words that have been replaced in this way as unknown words are included in the translated document. Even in such cases, neural machine translation may have difficulty translating it accurately.

[0039] Furthermore, even in sentences using words selected as high-frequency words in the training data, the words When the arrangement or combination of words is unusual, mistranslations are more likely to occur in neural machine translation.

[0040] Furthermore, there are errors that are characteristic of neural machine translation. The translated document may contain errors such as the repetition of the same word (or phrase), or An error occurs when a word (or phrase) in the translated document does not correspond to a word (or phrase). - and other problems are likely to occur.

[0041] As a method to reduce mistranslations in translated documents, the document to be translated is analyzed by a neural network. One approach is to modify the expression to match the trained model before performing neural machine translation. However, if you modify the document to be translated before performing neural machine translation, the document to be translated will be affected. There is a risk of over-correcting, which could result in a translation that is not faithful to the original document. be.

[0042] Therefore, in one aspect of the present invention, a machine translation method is used, first, by using a neural network By translating the translated document, the first translated document is generated. Next, the analysis results of the first translated document are... Based on the results, the words to be corrected are determined from among the words and phrases contained in the document to be translated. Next, the corrections are made. The target words are placed in the high-frequency words of the training data used to train the neural network. The translated document is modified by making changes. Then, after the modification, a neural network is used. By translating the first document to be translated, a second translated document is generated.

[0043] This machine translation method performs neural machine translation once and then analyzes the translation result. Based on the analysis results, the words to be modified in the document to be translated are determined. This can suppress excessive revisions to the document and generate a translated document that is faithful to the original document. It can be done. Also, when correcting the words in the document to be translated, the words in question can be processed by a neural network. Replace the high-frequency words in the training data used for learning the neural machine. This can improve the accuracy of translations.

[0044] Alternatively, in a machine translation method according to one aspect of the present invention, a neural network is used to translate the text to be translated. By translating the book, a translated document is generated, and by performing error checking on the translated document, the translation A score is obtained based on the translation accuracy of the document, and the translated document is revised until the completion conditions are met. The process involves generating a translated document based on the revisions, obtaining a score for the translated document based on the revisions, and repeating this process. return.

[0045] A mistranslation in a translated document is not limited to just one sentence; there may be multiple sentences containing mistranslations. If you modify many parts of the document being translated, it becomes difficult to determine which modifications are effective and which parts are not. It becomes difficult to determine whether a revision is undesirable. Furthermore, there may be more than one candidate for revision in the document being translated. In some cases, multiple correction candidates exist. Therefore, different corrections are applied to each. By creating multiple documents to be translated and translating each of those documents, Multiple translated documents may be generated based on this correction. For example, if the above score reaches a baseline value You may repeatedly revise the document to be translated until you achieve the desired result. Alternatively, for example, each revision may be different. Multiple documents to be translated that have been corrected, and the translated documents corresponding to those multiple documents to be translated, You can create multiple translations and output the one with the highest score as the translation result.

[0046] <Example of machine translation method 1> Figure 1 shows a flowchart of the machine translation method. As shown in Figure 1, one embodiment of the present invention The machine translation method has seven steps, from step A1 to step A7. Figure 2(A) shows... Figure 2(B) shows a schematic diagram of step A2, and Figure 2(B) shows schematic diagrams of steps A3, A4, and A5. A schematic diagram of step A6 is shown in 2(C).

[0047] [Step A1: Obtain the document to be translated, SD1] First, obtain the document to be translated, SD1. Even if the document to be translated, SD1 contains only one sentence, multiple sentences may be included. It can also be a number.

[0048] For example, the document to be translated, SD1, can be text data, audio data, or image data. It is supplied from an external source. If audio data or image data is supplied, then that data will be used. Next, create text data.

[0049] There are no particular limitations on the documents to be translated; for example, documents relating to intellectual property. Documents related to property include, specifically, the specification used in the patent application, the claims, and Examples include abstracts. Furthermore, documents related to intellectual property include patent documents (published patent documents). Examples include publications such as patent gazettes, utility model gazettes, design gazettes, and academic papers. This includes not only publications issued within Japan, but also publications issued in countries around the world, as intellectual property documents. It can be used as such.

[0050] In addition, the documents to be translated include emails, books, newspapers, academic papers, reports, columns, and You may also use various other copyrighted works, including other texts. Furthermore, the documents to be translated may include specifications, medical documents, etc. Medical documents and other similar materials may be used.

[0051] Furthermore, there are no particular limitations on the language of the document to be translated; for example, Japanese, English, Chinese, Korean. Documents written in Japanese or other languages ​​can be used.

[0052] [Step A2: Perform neural machine translation on document SD1 to generate translated document TD1] ] Next, as shown in Figure 2(A), the neural network NN of the processing unit 103 is used Then, by translating the document to be translated, SD1 is generated, resulting in the translated document TD1.

[0053] A neural network (NN) can handle variable-length data, such as text data. Recurrent neural networks can do this. You can use network(RNN).

[0054] The models and mechanisms used in neural machine translation are not particularly limited. For example, Sequ Sequence-to-Sequence model, Transformer model, Atten A tion mechanism or similar can be used.

[0055] The language of the translated document is not particularly limited as long as it is a different language from the language of the document being translated. For example, Japanese Japanese, English, Chinese, Korean, and other languages ​​can be used.

[0056] [Step A3: Determine the sentence to be corrected (CTS) from the translated document TD1] Next, we analyze the translated document TD1 to identify potentially mistranslated sentences (sentences to be corrected, CTS). In Figure 2(B), the sentence to be corrected (CTS) is a possible mistranslation contained in the translated document TD1. This shows an example where gender was determined.

[0057] Step A3 can be considered an error check of the entire translated document TD1. It is preferable that a score can be calculated that serves as an evaluation index for the translation accuracy of the translated document TD1.

[0058] In the analysis of the translated document TD1, sentences with the same word (or phrase) repeated, and the translated text. Sentences containing words (or phrases) not included in document SD1, or sentences with grammatical errors. It is preferable to detect at least one of the following: a sentence, or a sentence containing special characters. It seems so.

[0059] One or more types of processing can be used to detect sentences that may be mistranslated. For example By using various natural language processing techniques, it is possible to detect sentences that may be mistranslated.

[0060] For example, using a pre-prepared word dictionary, the nouns contained in the document to be translated (SD1) and the translated document You may also compare it with the nouns included in TD1.

[0061] Furthermore, the result of back-translating the translated document TD1 into the language of the document to be translated SD1, and the language of the document to be translated SD1 You may compare the texts between them.

[0062] To detect potentially mistranslated sentences, the probability of the output result of the neural machine translation model is... The values ​​can also be analyzed. In neural machine translation, the question is which words in the translated text are most strongly related. Considering the weight coefficients of the intermediate layer, the most probable word in the translation language is output. If the probability of a word is low, it can be determined that a sentence containing that word may be a mistranslation. ru.

[0063] In order to detect sentences that may be mistranslated, past translation results stored on servers, etc. Information may be used. For example, information on the strength of the relationship between words in the translated text and the translated text ( The weight coefficients of the hidden layer in a neural network (NN) are saved, and different from the past. If the word is strongly associated with something else, it can be determined that there is a possibility of mistranslation.

[0064] Note that there may be one or more sentences to be corrected. If there are multiple sentences to be corrected, For example, after performing steps A4 and A5 for each sentence, step A6 You can proceed to step A4. Alternatively, for a given sentence, perform steps A4 through A6. Then, return to step A3 or step A4 and repeat the same process for the other sentences. You may do so.

[0065] [Step A4: Words contained in sentence S1 in the translated document SD1 that correspond to the sentence CTS to be corrected] [Determine which words or phrases to revise from within the sentence.] Next, from among the words and phrases contained in the document to be translated, SD1, we identify the words and phrases that are predicted to be the cause of the mistranslation. Determine the words / phrases to be corrected. In Figure 2(B), the translated document SD1 is corrected in the sentence CTS. An example that has a corresponding sentence S1 and determines the words to be modified from among the words contained in that sentence S1. This indicates that the words or phrases to be corrected may be single words or phrases.

[0066] Among the words and phrases included in sentence S1, the words and phrases that are predicted to be the cause of the mistranslation are, for example, ni Words used during training of a neural network (NN) (i.e., words that are treated as unknown words with special symbols) Among the words that have not been replaced, some are those that are used relatively infrequently. Information on the frequency of word usage in the training data used to train the neural network (NN). By referring to this, you can find the least frequently used word, or a phrase containing the least frequently used word. The terms can be identified as words that need to be corrected.

[0067] Word frequency can be calculated, for example, by determining the word's TF (Term Frequency) value. This can be determined by the TF-IDF (Ter m Frequency (Inverse Document Frequency) value You may also use [this].

[0068] To detect words and phrases that are predicted to be the cause of mistranslations, a neural network (NN) is used. The weight coefficients of the hidden layer may also be analyzed. In neural machine translation, the weight coefficients of the hidden layer By referring to this, you can see which words in the translated text the output word is most closely related to. For example, a word that has little relation to any other word in the text being translated can be identified as a word to be modified. Cut.

[0069] Furthermore, the method for determining the words to be corrected is not limited to the above, and various methods can be used. .

[0070] [Step A5: Replace the words to be corrected with high-frequency words in the training data and translate the document into a new document.] [Modify SD1] Next, the words to be corrected are used in the training data used to train the neural network NN. By replacing high-frequency words, the translated document SD1 is modified, and the translated document SD2 is created. In the following, the revised translated document SD2 shown in Figure 2(B) has the same words as the words to be revised in sentence S1. It has sentence S1' which is replaced with the phrase.

[0071] The words to be corrected were identified as high-frequency words in the training data used to train the neural network (NN). By replacing words with other words, the accuracy of translation in neural machine translation can be improved.

[0072] However, as a result of modifying the translated document SD1 to reduce mistranslations, the original translated document S It is undesirable to lose the ability to provide a translation that is faithful to the content of D1. Therefore, the phrases to be corrected are: It is preferable that the phrase to be modified be replaced with a synonym or related term. Even if you modify the document to be translated, the content of the document to be translated will change significantly. This can be prevented. Therefore, a translation that is faithful to the content of the original document to be translated, SD1, can be performed. Furthermore, it can help reduce mistranslations.

[0073] Specifically, multiple synonyms and related words for the word to be corrected are extracted, and from among them, neural networks are selected. Select the words that appear most frequently in the training data used to train the network neural network, and then select the words that appear most frequently in the training data. Therefore, it is preferable to replace the phrase to be modified.

[0074] Synonyms and related terms for the words to be corrected can be obtained, for example, by using a dictionary prepared in advance. This is possible. Also, based on the similarity or proximity of the word's distributed representation vectors, You may also extract synonyms and related terms for the phrase to be modified.

[0075] High-frequency words in the training data used to train a neural network (NN) are... Selection can be made based on information about the frequency of word usage in the database.

[0076] The frequency of use of a word can be determined, for example, by calculating the word's TF value. Also, You may also use the TF-IDF value of the word.

[0077] Note that there may be one or more candidates for the word (or phrase) to replace the word or phrase to be corrected. This may also be the case. For example, the most frequently used word in the training data may be used to replace the word It can be used as a candidate word, or as a candidate word to replace multiple frequently used words. Good. If there are multiple candidate words to replace, each will be a different word to replace the phrase to be modified. Create multiple revised documents to be translated, and in step A6, create multiple documents based on the respective revisions. You may create a translated document. In this case, as described above in step A3, the translation of the document By analyzing the data, a score based on translation accuracy can be calculated, allowing us to determine which of multiple translated documents is the best. It is preferable because it makes it easier to determine whether the translation result is satisfactory.

[0078] If the phrase to be corrected contains multiple words, each word may be replaced with another word. You may replace only the word in the section with another word.

[0079] [Step A6: Perform neural machine translation on the corrected document to be translated, SD2, and then translate the translated document TD2. [Generate] Next, as shown in Figure 2(C), the neural network NN of the processing unit 103 is used Then, by translating the revised document to be translated, the translated document TD2 is generated.

[0080] [Step A7: Output the translated document TD2] Then, output the translated document TD2.

[0081] For example, a translated document TD2 can be output as text data, audio data, or image data. Output to the section. If outputting audio data or image data, generate it in step A6. Based on the text data, audio or image data is created.

[0082] In this embodiment, the machine translation method described is an example in which processing is performed automatically by a machine. However, some human processing may be involved. For example, the items to be corrected in step A3. The specification of sentences, the specification of words to be corrected in step A4, etc., may be done by a person.

[0083] Furthermore, the data used in the machine translation method of this embodiment and the generated data may be appropriately shared. It may be saved in a computer or memory, etc. For example, the document to be translated obtained in step A1, step The translated documents generated in Step A2 and Step A6, and the corrected documents generated in Step A5. At least one of the translated documents may be stored on a server, in memory, or elsewhere.

[0084] Furthermore, when retraining the neural network, the document to be translated (SD1) and the translated document (TD1) are used. The set of documents, and either or both of the sets of the document to be translated (SD2) and the translated document (TD2), This can be added to the training data. This will improve the translation accuracy in neural machine translation. It is possible to measure this.

[0085] <Example of machine translation method 2> Figure 3 shows a flowchart of the machine translation method. As shown in Figure 3, one embodiment of the present invention The machine translation method has nine steps, from step A11 to step A19. For parts similar to steps A1 to A7, detailed explanations will be omitted.

[0086] [Step A11: Obtain the document to be translated] First, as in step A1, obtain the document to be translated.

[0087] [Step A12: Perform neural machine translation on the document to be translated and generate the translated document] Next, similar to step A2, the document to be translated is translated using a neural network. This generates the translated document.

[0088] [Step A13: Perform error checking] Next, we perform an error check. In the error check, we perform an error check in the step before this step (Step A12). Alternatively, the translated document generated in step A18) described below is analyzed to determine the translation accuracy of the translated document. A score is calculated based on this. This score serves as an evaluation index for translation accuracy. Therefore, - When multiple revisions are made and multiple revised versions (translated documents) are obtained, the score is used to assess translation accuracy. This allows you to determine which translated documents have a high quality rating.

[0089] [Step A14: Determine if the error correction completion conditions are met.] Next, determine if the error correction completion conditions are met. If so, proceed to step A19. If the conditions are not met, proceed to step A15.

[0090] The conditions for ending error correction are not particularly limited. For example, if the score calculated in step A13 is Error correction may be terminated when the threshold value is reached (the number of mistranslations has been sufficiently reduced). Alternatively, error correction may be terminated after a predetermined number of error corrections have been performed.

[0091] [Step A15: Determine the sentence to be corrected] Next, similar to step A3, identify sentences in the translated document that may contain mistranslations (sentences to be corrected). Determine the corrected sentences using the results of the error check in step A13. You may do so.

[0092] [Step A16: From among the words and phrases contained in the sentence in the translated document that corresponds to the sentence to be corrected, [Determine the correct target phrase] Next, similar to step A4, we identify the words and phrases in the document to be translated that are the cause of the mistranslation. Determine the predicted words (words to be corrected).

[0093] [Step A17: Replace the words to be corrected with high-frequency words in the training data, and translate the sentence [Revise the document] Next, similar to step A5, the words to be modified were used to train the neural network. The translated document is modified by replacing high-frequency words in the training data with those that are frequently used in the training data.

[0094] [Step A18: Perform neural machine translation on the modified document to be translated and generate the translated document.] Next, similar to step A6, we use a neural network to process the modified document to be translated. The translated document is generated by translating. After step A18, proceed to step A13.

[0095] [Step A19: Output the translated document] If the error correction completion conditions are met in step A14, the translated document will be output. If the score calculated in step A13 reaches the target value, error correction is terminated. It is preferable to output a translated document in which the score reaches the standard value. Also, error correction If the error correction is completed after performing the process a predetermined number of times, the scores of a predetermined number of translated documents will be compared. It is preferable to output the translated document with the highest translation accuracy. The number of translated documents to output is... It may be one or more.

[0096] As described above, in the machine translation method of this embodiment, neural machine translation is performed once, and Based on the translation results, the words and phrases to be modified in the translated document are determined. This can suppress excessive revisions to the document and generate a translated document that is faithful to the original document. It can be done. Also, when correcting the words in the document to be translated, the words in question can be processed by a neural network. Replace the high-frequency words in the training data used for learning the neural machine. This can improve the accuracy of translations.

[0097] This embodiment can be appropriately combined with other embodiments. Furthermore, this specification Furthermore, if multiple configuration examples are shown within a single embodiment, the configuration examples may be combined as appropriate. It is possible to do so.

[0098] (Embodiment 2) In this embodiment, a machine translation system according to one aspect of the present invention will be described using Figures 4 and 5. I will reveal it.

[0099] The machine translation system of this embodiment uses the machine translation method shown in Embodiment 1 to perform a certain self It is possible to translate one natural language into another natural language. Therefore, it is possible to be faithful to the document being translated. It can translate, and the translation accuracy is high.

[0100] <Example of a machine translation system configuration 1> Figure 4 shows a block diagram of the machine translation system 100. Note that the drawings attached to this specification are... This classifies the components by function and shows them as independent blocks in a block diagram. However, it is difficult to completely separate the actual components by function, and one component is complex. It may also be related to the function of numbers. Furthermore, one function may be related to multiple components. For example, the processing performed by the processing unit 103 is executed on different servers depending on the processing. Sometimes.

[0101] The machine translation system 100 has at least a processing unit 103. The machine translation system shown in Figure 4 The stem 100 further includes an input unit 101, a transmission line 102, a storage unit 105, and a database 1 It has 07 and an output unit 109.

[0102] [Input section 101] The input unit 101 receives the document to be translated from outside the machine translation system 100. The document to be translated, supplied to 101, is transmitted via the transmission path 102 to the processing unit 103 and the storage unit 105. , or supplied to database 107.

[0103] The document to be translated is input as, for example, text data, audio data, or image data. ru.

[0104] Input methods for the document to be translated include, for example, key input using a keyboard or touch panel. Force, voice input using a microphone, reading from recording media, scanner, camera, etc. Examples include image input and data acquisition using communication.

[0105] The machine translation system 100 has the function of converting speech data into text data. Preferably, the processing unit 103 may have the function. Alternatively, the machine translation system The Tem 100 may further have an audio conversion unit having the said function.

[0106] The machine translation system 100 preferably has an optical character recognition (OCR) function. This allows for the recognition of characters contained in image data and the creation of text data. For example, the processing unit 103 may have the function. Or, the machine translation system 100 However, it may also have a character recognition unit that has the said function.

[0107] [Transmission path 102] The transmission line 102 has the function of transmitting various data. Input unit 101, processing unit 103, Data transmission and reception between the memory unit 105, the database 107, and the output unit 109 is performed via transmission path 1 This can be done via 02. For example, data such as the document to be translated and the translated document can be transmitted through the transmission channel It is transmitted and received via 102.

[0108] [Processing step 103] The processing unit 103 receives data supplied from the input unit 101, storage unit 105, database 107, etc. It has the function of performing calculations using data. The processing unit 103 stores the calculation results in the storage unit 105 This can be supplied to the database 107, output unit 109, etc.

[0109] The processing unit 103 has the function of performing neural machine translation and generating translated documents. For example, input The document to be translated, input to the power unit 101, or the document to be translated modified by the processing unit 103, It can be translated.

[0110] The processing unit 103 has the function of determining the sentences to be corrected from the translated document. 103 is a selection of words and phrases from the text in the translated document that correspond to the sentence to be corrected. It has the function of determining the phrase. The processing unit 103 determines the phrase to be corrected based on the high frequency in the training data. It has the function of replacing words with grammatical units, modifying the document to be translated, and generating the modified document to be translated.

[0111] The processing unit 103 uses a transistor having a metal oxide in the channel formation region. Preferred. Because the transistor has an extremely low off-current, the transistor can be used as a memory element. It is used as a switch to hold the charge (data) that flows into a capacitive element that functions as such. This ensures that data can be retained for a long period of time. By using it in at least one of the registers and cache memory of 103, The processing unit 103 is operated only when necessary, and in other cases, the information of the previous processing is stored in the memory element. By putting it into standby mode, the processing unit 103 can be turned off. In other words, normally This enables power-saving computing, which can lead to lower power consumption in machine translation systems. ru.

[0112] In this specification, etc., when an oxide semiconductor or metal oxide is used in the channel formation region. Transistors are called Oxide Semiconductor transistors, or OS transistors. It is called a transistor. The channel formation region of an OS transistor may contain a metal oxide. preferable.

[0113] In this specification, metal oxide refers to metals in a broad sense. It is an oxide. Metal oxides are oxide insulators and oxide conductors (including transparent oxide conductors). Oxide semiconductors (also called OS) They are classified into the following categories. For example, when a metal oxide is used in the semiconductor layer of a transistor, the metal Oxides are sometimes referred to as oxide semiconductors. In other words, metal oxides have amplification and rectification effects. , and if it has at least one switching action, the metal oxide is a metal oxide Semiconductors (metal oxide semiconductors), abbreviated as OS It is possible.

[0114] The metal oxide in the channel-forming region preferably contains indium (In). If the metal oxide in the Nell-forming region is an indium-containing metal oxide, the OS Transis The carrier mobility (electron mobility) of the ion becomes higher. Also, the metallic acid present in the channel-forming region The oxide is preferably an oxide semiconductor containing element M. Element M is aluminum (Al). It is preferable that it be gallium (Ga) or tin (Sn). Other applicable elements M The elements include boron (B), silicon (Si), titanium (Ti), iron (Fe), and nickel. Kel (Ni), Germanium (Ge), Yttrium (Y), Zirconium (Zr), Mo Ribdenum (Mo), Lanthanum (La), Cerium (Ce), Neodymium (Nd), Hafniu Examples include fluorine (Hf), tantalum (Ta), and tungsten (W). However, as element M... In some cases, it is acceptable to combine multiple of the aforementioned elements. Element M, for example, can be combined with oxygen. It is an element with high bonding energy. For example, its bonding energy with oxygen is higher than that of indium. It is an element. Furthermore, the metal oxides that the channel-forming region contains include zinc (Zn). This is preferable. Zinc-containing metal oxides may be prone to crystallization.

[0115] The metal oxides present in the channel-forming regions are not limited to indium-containing metal oxides. The semiconductor layer is made of materials such as zinc tin oxide and gallium tin oxide, which do not contain indium. These included metal oxides containing zinc, metal oxides containing gallium, and metal oxides containing tin. That's fine.

[0116] Furthermore, the processing unit 103 may use a transistor that includes silicon in its channel formation region. stomach.

[0117] Furthermore, the processing unit 103 includes a transistor containing an oxide semiconductor in the channel formation region, and a channel It is preferable to use a transistor containing silicon in the Nel-forming region in combination with the other transistor.

[0118] The processing unit 103 is, for example, an arithmetic circuit or a central processing unit (CPU). It has an processing unit, etc.

[0119] The processing unit 103 includes a DSP (Digital Signal Processor) and a GP (Ground Processing Unit). It has a microprocessor such as U (Graphics Processing Unit). It is acceptable to do so. Microprocessors are FPGAs (Field Programmable Arrays). Field Programmable Array), FPAA (Field Programmable A Programmable Logic Dev (PLD) such as a rectangular array The configuration may be implemented by ice. The processing unit 103 is determined by the processor. By interpreting and executing instructions from various programs, various data processing and programmatic processes are performed. It is possible to perform the action. The programs that can be executed by the processor are those that the processor possesses. It is stored in at least one of the memory area and the storage unit 105.

[0120] The processing unit 103 may have main memory. The main memory may be volatile memory such as RAM. It has at least one of Mori and non-volatile memory such as ROM.

[0121] For example, RAM can be DRAM (Dynamic Random Access Memory). mory), SRAM (Static Random Access Memory), etc. This is used, and a memory space is virtually allocated and used as the workspace for the processing unit 103. The operating system and application programs stored in the memory unit 105. Program modules, program data, and lookup tables are used for execution. These are then loaded into RAM. These data, programs, and programs loaded into RAM are then loaded into RAM. Each program module is directly accessed and operated by the processing unit 103.

[0122] The ROM contains BIOS (Basic Input / Output) which does not require rewriting. It can store the System and firmware, etc. As for ROM, SCRROM, OTPROM (One Time Programmable Read) Only Memory), EPROM (Erasable Programmable Examples include Read Only Memory. EPROMs include ultraviolet light UV-EPROM (Ultra-Violet) enables the erasure of stored data by irradiation. Erasable Programmable Read Only Memory), EEPROM (Electrically Erasable Programmability) Examples include Read Only Memory (e) and flash memory.

[0123] [Storage section 105] The memory unit 105 has the function of storing the program to be executed by the processing unit 103. The memory unit 105 contains, for example, the calculation results (translated document, error check results) generated by the processing unit 103. The results include a list of suggested sentences for correction, a list of words to be corrected, the corrected translated document, etc., and the input section. It may also have a function to store data entered into 101 (such as the document to be translated).

[0124] The storage unit 105 has at least one of volatile memory and non-volatile memory. The unit 105 may have, for example, volatile memory such as DRAM or SRAM. Part 105 is, for example, ReRAM (Resistive Random Access). Memory (also called resistive random-access memory), PRAM (Phase change R) andom Access Memory), FeRAM (Ferroelectric Random Access Memory), MRAM (Magnetoresis (also known as magnetically resistive random access memory) Alternatively, it may have non-volatile memory such as flash memory. Also, storage unit 105 These include hard disk drives (HDDs) and solid-state drives. Recording media such as state drives (Solid State Drive: SSD) It's okay to have live performances.

[0125] [Database 107] Database 107 contains at least the training data used to train the neural network NN. It has a function to store information on the frequency of word usage in the database. For example, the calculation results (translation document, error check results, correction candidates) generated by the processing unit 103. (List of supplementary texts, list of words to be corrected, corrected translated document, etc.), and input into input unit 101 It may also have a function to store translated data (such as translated documents). Databases 05 and 107 do not need to be separated from each other. For example, a machine translation system The system has a storage unit that has the functions of both a storage unit 105 and a database 107. It's fine if you do that.

[0126] Furthermore, the memory of the processing unit 103, the storage unit 105, and the database 107 is Therefore, it can be considered an example of a non-temporary computer-readable storage medium.

[0127] [Output section 109] The output unit 109 has the function of supplying data to the outside of the machine translation system 100. If so, the calculation results in the processing unit 103 can be supplied to an external source. For example, the input The translated document corresponding to the document to be translated can be supplied externally.

[0128] Furthermore, the machine translation system 100 uses the text data of the translated document to process the audio data and It may have the function to generate one or both of the image data.

[0129] <Example of a machine translation system configuration 2> Figure 5 shows a block diagram of the machine translation system 150. The machine translation system 150 is a service It has a base station 151 and a terminal 152 (such as a personal computer).

[0130] Server 151 includes a communication unit 161a, a transmission line 162, a processing unit 163a, and Database 1 It has 67. Although not shown in Figure 5, the server 151 also has a storage unit, an input / output unit, etc. It may have.

[0131] Terminal 152 includes a communication unit 161b, a transmission line 168, a processing unit 163b, a storage unit 165, and input It has an output unit 169. Although not shown in Figure 5, terminal 152 further has a database It may have any of the following.

[0132] The user of the machine translation system 150 inputs the document to be translated from terminal 152 to server 151. To be translated. The document to be translated is transmitted from communication unit 161b to communication unit 161a.

[0133] The document to be translated received by the communication unit 161a is transmitted via the transmission path 162 to Database 167. Alternatively, the document to be translated is stored in the memory unit (not shown). It may also be supplied directly to the processing unit 163a.

[0134] The generation of the translated document, determination of the sentence to be corrected, determination of the words and phrases to be corrected, and as described in Embodiment 1 Furthermore, correcting the translated document requires high processing power. Server 151 has Processing unit 163a has higher processing power than processing unit 163b, which is located in terminal 152. These processes are preferably carried out by the processing unit 163a.

[0135] Then, the processing unit 163a generates the translated document. The translated document is transmitted via the transmission line 162. The data is then stored in database 167 or a storage unit (not shown). Alternatively, the translated document is Alternatively, the processing unit 163a may supply the data directly to the communication unit 161a. After that, the server 15 From 1, the translated document is output to terminal 152. The translated document is sent from communication unit 161a to communication unit It will be sent to 161b.

[0136] [I / O section 169] Data is supplied to the input / output unit 169 from outside the machine translation system 150. 169 has the function of supplying data to the machine translation system 150. The input and output sections may be separate, as in translation system 100.

[0137] [Transmission lines 162 and 168] Transmission lines 162 and 168 have the function of transmitting data. Communication unit 161a, processing Data transmission and reception between the data processing unit 163a and the database 167 is via the transmission path 162. This can be done. Communication unit 161b, processing unit 163b, storage unit 165, and input / output unit 16 Data transmission and reception between points 9 can be performed via the transmission line 168.

[0138] [Processing Unit 163a and Processing Unit 163b] The processing unit 163a receives data supplied from the communication unit 161a and the database 167, etc. It has the function of performing calculations using the following. The processing unit 163b has the communication unit 161b, the storage unit 165, It also has the function of performing calculations using data supplied from the input / output unit 169, etc. Sections 163a and 163b can refer to the description of the processing unit 103. Processing unit 163a is Therefore, it is preferable that the processing capacity is higher than that of processing unit 163b.

[0139] [Storage section 165] The memory unit 165 has the function of storing the program to be executed by the processing unit 163b. The storage unit 165 stores the calculation results generated by the processing unit 163b and the data input to the communication unit 161b. has a function of storing the data input to the input / output unit 169 and the like.

[0140] [Database 167] The database 167 has at least a function of storing information on the frequency of use of words in the learning data used for learning the neural network NN. In addition, the database 167 may have a function of storing the calculation results generated by the processing unit 163a and the data input to the communication unit 161a. Or, the server 151 may have a storage unit separate from the database 167, and the storage unit may have a function of storing the calculation results generated by the processing unit 163a and the data input to the communication unit 161a.

[0141] [Communication units 161a and 161b] Using the communication units 161a and 161b, data can be transmitted and received between the server 151 and the terminal 152. As the communication units 161a and 161b, a hub, a router, a modem, etc. can be used. For data transmission and reception, wired or wireless (for example, radio waves, infrared rays, etc.) can be used.

[0142] This embodiment can be appropriately combined with other embodiments.

Explanation of Signs

[0143] CTS correction target sentence S1 sentence TD1 translation document TD2 translation document SD1 source document to be translated SD2 source document to be translated 100 Machine translation system 101 Input unit 102 Transmission path 103 Processing unit ​​​​​​​105 Storage section 107 Databases 109 Output section 150 Machine Translation Systems 151 Servers 152 terminals 161a Communications Department 161b Communications Department 162 transmission lines 163a Processing Unit 163b Processing Unit 165 Storage section 167 Databases 168 transmission lines 169 Input / output section

Claims

[Claim 1] Having a processing unit, The aforementioned processing unit, The system has a function to generate a first translated document by translating the document to be translated using a neural network, A function to identify sentences suspected of being mistranslated from the aforementioned first translated document, A function to determine the phrase to be corrected from among the phrases contained in the translated document corresponding to the sentence suspected of being mistranslated, The function modifies the document to be translated by replacing the aforementioned words to be modified with high-frequency words in the training data used to train the neural network, The system has a function to generate a second translated document by translating the modified document to be translated using the aforementioned neural network, The sentences suspected of being mistranslated are at least one of the following: sentences in which the same word is repeated, sentences in which a word not included in the translated document is included, sentences that are grammatically incorrect, and sentences that contain special characters. The aforementioned words to be corrected are words that have weak connections to any of the words in the translated text, as detected by analyzing the weight coefficients of the intermediate layer of the neural network. The machine translation system selects the aforementioned high-frequency words from among synonyms and related words of the aforementioned words to be modified.

Images