Machine translation method, machine translation system, program, and non-transitory computer-readable storage medium

Abstract

To enhance accuracy of machine translation, to obtain a translated document with high translation accuracy, and to faithfully translate a document to be translated.SOLUTION: A document to be translated is translated by using a neural network to generate a first translated document. A modification-target word or phrase is determined from words and phrases contained in the document to be translated on the basis of an analysis result for the first translated document. The modification-target word or phrase is replaced with a high frequency word in learning data used for learning in the neural network to modify the document to be translated. And the modified document to be translated is translated by using the neural network to generate a second translated document.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-transitory computer-readable storage medium.

[0002] Note that one embodiment of the present invention is not limited to the above technical field. Examples of the technical field of one embodiment of the present invention include a semiconductor device, a display device, a light-emitting device, a power storage device, a memory device, an electronic device, a lighting device, an input device (e.g., a touch sensor), an input / output device (e.g., a touch panel), a driving method thereof, or a manufacturing method thereof. [Background technology]

[0003] Research and development of machine translation, which uses a computer to translate one natural language into another, is being actively carried out. Machine translation includes rule-based machine translation, which performs translation based on rules, statistical machine translation, which performs translation using a language model and a translation model, and neural machine translation, which performs translation using a neural network.

[0004] In statistical machine translation, learning is carried out through multiple steps using various tools. For this reason, training data must be prepared for each tool. On the other hand, neural machine translation uses a single neural network for learning, so the only training data required is a bilingual corpus that pairs documents to be translated with their translations. This makes it easier to obtain training data, and reduces the effort required to artificially create training data. Furthermore, neural machine translation tends to achieve higher translation accuracy than rule-based machine translation and statistical machine translation. For these reasons, efforts are underway to put neural machine translation into practical use.

[0005] However, machine translation is still not perfect and mistranslations occur. Patent Document 1 discloses a translation method aimed at preventing the occurrence of mistranslations. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] JP 2019-20950 A Summary of the Invention [Problem to be solved by the invention]

[0007] An object of one embodiment of the present invention is to improve the accuracy of machine translation.An object of one embodiment of the present invention is to obtain a translated document with high translation accuracy by machine translation.An object of one embodiment of the present invention is to obtain a translation result that is faithful to the original document by using machine translation.

[0008] An object of one embodiment of the present invention is to provide a highly accurate machine translation system, and an object of one embodiment of the present invention is to provide a highly accurate machine translation method.

[0009] An object of one embodiment of the present invention is to provide a machine translation system capable of faithfully translating a document to be translated.An object of one embodiment of the present invention is to provide a machine translation method capable of faithfully translating a document to be translated.

[0010] Note that the description of these problems does not preclude the existence of other problems. One embodiment of the present invention does not necessarily have to solve all of these problems. Problems other than these can be extracted from the description of the specification, drawings, and claims. [Means for solving the problem]

[0011] One embodiment of the present invention is a machine translation method that generates a first translated document by translating a document to be translated using a neural network, determines words to be corrected from among words contained in the document to be translated based on the analysis results of the first translated document, corrects the document to be translated by replacing the words to be corrected with high-frequency words in the training data used to train the neural network, and generates a second translated document by translating the corrected document to be translated using the neural network.

[0012] In the step of determining the phrases to be corrected, it is preferable to determine a sentence to be corrected from the first translation document, and to determine the phrases to be corrected from among phrases included in a sentence in the document to be translated that corresponds to the sentence to be corrected.

[0013] One aspect of the present invention is a machine translation method that uses a neural network to translate a document to be translated, generates a translated document, analyzes the translated document to obtain a score based on the translation accuracy of the translated document, and repeats the steps of correcting the document to be translated, generating a translated document based on the correction, and obtaining a score for the translated document based on the correction until a termination condition is met. Correcting the document to be translated includes a step of determining a correction target phrase from among phrases contained in the document to be translated based on the analysis result of the translated document generated before this step, and a step of correcting the document to be translated by replacing the correction target phrase with a high-frequency word in the learning data used to train the neural network. Generating a translated document based on the correction includes a step of translating the document to be translated that reflects the latest correction using a neural network, thereby generating a translated document based on the latest correction.

[0014] In the step of determining the words and phrases to be corrected, it is preferable to determine a sentence to be corrected from within the translation document, and to determine the words and phrases to be corrected from among words and phrases contained in a sentence in the document to be translated that corresponds to the sentence to be corrected.

[0015] The high frequency words are preferably selected from among synonyms and similar words of the phrase to be amended.

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

[0017] One embodiment of the present invention is a machine translation system having a processing unit, which has the following functions: generating a first translated document by translating a document to be translated using a neural network; determining words to be corrected from among words contained in the document to be translated based on the analysis results of the first translated document; correcting the document to be translated by replacing the words to be corrected with high-frequency words in the training data used to train the neural network; and generating a second translated document by translating the corrected document to be translated using the neural network.

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

[0019] The program may be provided to the computer by various types of temporary computer-readable storage media. Examples of the temporary computer-readable storage media include electric signals, optical signals, and electromagnetic waves. The temporary computer-readable storage media can provide the program to the computer via a wired communication path such as an electric wire or an optical fiber, or via a wireless communication path.

[0020] One aspect of the present invention is a program that causes a processor to execute the steps of: generating a first translated document by translating a document to be translated using a neural network; determining words to be corrected from among words contained in the document to be translated based on the analysis result of the first translated document; correcting the document to be translated by replacing the words to be corrected with high-frequency words in the training data used to train the neural network; and generating a second translated document by translating the corrected document to be translated using the neural network. Another aspect of the present invention is a non-transitory computer-readable storage medium having the program stored therein.

[0021] As the non-transitory computer-readable storage medium, various types of tangible storage media can be used. Examples of the non-transitory computer-readable storage medium include volatile memory such as RAM (Random Access Memory) and non-volatile memory such as ROM (Read Only Memory). Other examples include recording media drives such as hard disk drives (HDD) and solid state drives (SSD), magneto-optical disks, CD-ROMs, CD-Rs, etc. Effect of the Invention

[0022] According to one embodiment of the present invention, the accuracy of machine translation can be improved. According to one embodiment of the present invention, a translated document with high translation accuracy can be obtained. According to one embodiment of the present invention, a translation result that is faithful to the original document can be obtained by using machine translation.

[0023] According to one aspect of the present invention, a highly accurate machine translation system and a highly accurate machine translation method are provided.

[0024] According to one aspect of the present invention, it is possible to provide a machine translation system capable of faithfully translating a document to be translated. According to one aspect of the present invention, it is possible to provide a machine translation method capable of faithfully translating a document to be translated.

[0025] Note that the description of these effects does not preclude the existence of other effects. One embodiment of the present invention does not necessarily have all of these effects. Effects other than these can be extracted from the description in the specification, drawings, and claims. [Brief description of the drawings]

[0026] [Figure 1] FIG. 1 is a flow diagram showing an example of a machine translation method. [Diagram 2]FIG. 2 is a schematic diagram showing an example of a machine translation method. [Diagram 3] FIG. 3 is a flow diagram illustrating an example of a machine translation method. [Figure 4] FIG. 4 is a block diagram showing an example of a machine translation system. [Diagram 5] FIG. 5 is a block diagram illustrating an example of a machine translation system. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0027] The embodiments will be described in detail with reference to the drawings. However, the present invention is not limited to the following description, and it will be easily understood by those skilled in the art that the modes and details of the present invention can be modified in various ways without departing from the spirit and scope of the present invention. Therefore, the present invention should not be interpreted as being limited to the description of the embodiments shown below.

[0028] In the configuration of the invention described below, the same parts or parts having similar functions are denoted by the same reference numerals in different drawings, and the repeated explanations are omitted. In addition, when referring to similar functions, the same hatch pattern may be used and no particular reference numeral may be used.

[0029] In addition, for ease of understanding, the position, size, range, etc. of each component shown in the drawings may not represent the actual position, size, range, etc. Therefore, the disclosed invention is not necessarily limited to the position, size, range, etc. disclosed in the drawings.

[0030] (Embodiment 1) In this embodiment, a machine translation method according to one embodiment of the present invention will be described with reference to FIGS.

[0031] In a machine translation method according to an embodiment of the present invention, artificial intelligence (AI) is used when translating a document to be translated to generate a translated document. Specifically, in the machine translation method according to an embodiment of the present invention, an artificial neural network (ANN, hereinafter also simply referred to as a neural network) is used. The neural network is realized by a circuit (hardware) or a program (software).

[0032] In this specification, the term "neural network" refers to a general model that imitates the neural circuit network of a living organism, determines the connection strength between neurons through learning, and has a problem-solving ability. A neural network has an input layer, an intermediate layer (hidden layer), and an output layer.

[0033] In this specification and the like, when discussing neural networks, determining the connection strengths (also called weighting coefficients) between neurons from existing information may be referred to as "learning".

[0034] In this specification and the like, constructing a neural network using the connection weights obtained by learning and deriving a new conclusion from it may be called "inference." In one aspect of the present invention, translating a document to be translated using a neural network corresponds to inference.

[0035] In this specification and the like, translation using a neural network may be referred to as "neural machine translation."

[0036] Here, neural machine translation can sometimes result in mistranslations.

[0037] To train a neural network, training data (also called a training corpus) containing a large number of sets of documents to be translated and their translations is used. If the document to be translated contains words or phrases that are not included in the training data, neural machine translation may not be able to translate it properly.

[0038] Furthermore, in a neural network, the higher the dimensions of the input layer and the output layer, the greater the amount of calculation. Therefore, the vocabulary of each language (language of the document to be translated and language of the translated document) used during learning is limited to a certain number. Specifically, the vocabulary used during learning is limited to words that are frequently used in the learning data (hereinafter also referred to as high-frequency words), and words that are not often used in the learning data (hereinafter also referred to as low-frequency words) are replaced with special symbols as unknown words. Even when words replaced as unknown words in this way are included in the document to be translated, it is difficult for neural machine translation to translate them appropriately.

[0039] Furthermore, even if a sentence uses words selected as high-frequency words in the training data, mistranslations are likely to occur in neural machine translation if the word order or combination is unusual.

[0040] There are also errors that are unique to neural machine translation, such as repeated use of the same word (or phrase) or use of words (or phrases) that do not correspond to words (or phrases) in the original document.

[0041] One method for reducing mistranslations in a translation document is to correct the document to be translated into expressions that fit the trained model of a neural network before performing neural machine translation. However, if the document to be translated is corrected before performing neural machine translation, there is a risk that the document to be translated will be over-corrected, and the translation result may not be faithful to the document to be translated.

[0042] In a machine translation method according to one embodiment of the present invention, a first translated document is generated by translating a document to be translated using a neural network. Next, based on the analysis result of the first translated document, a phrase to be corrected is determined from among the phrases contained in the document to be translated. Next, the phrase to be corrected is replaced with a high-frequency word in the training data used to train the neural network, thereby correcting the document to be translated. Then, a second translated document is generated by translating the corrected document to be translated using the neural network.

[0043] In this machine translation method, neural machine translation is performed once and the translation result is analyzed. Then, based on the analysis result, words to be corrected in the to-be-translated document are determined. Therefore, excessive correction of the to-be-translated document can be prevented, and a translated document that is faithful to the to-be-translated document can be generated. When correcting words in the to-be-translated document, the words are replaced with high-frequency words in the training data used to train the neural network. This can improve the translation accuracy in neural machine translation.

[0044] Alternatively, in one embodiment of the machine translation method of the present invention, a to-be-translated document is translated using a neural network to generate a translated document, and an error check is performed on the translated document to obtain a score based on the translation accuracy of the translated document. The to-be-translated document is then corrected, a translated document is generated based on the corrections, and a score is obtained for the translated document based on the corrections, and this is repeated until a termination condition is satisfied.

[0045] The mistranslation in the translation document is not limited to one sentence, and there may be multiple sentences including mistranslations. If many parts of the to-be-translated document are corrected at once, it becomes difficult to determine which parts are effective and which parts are not. In addition, there may be multiple correction candidates for the to-be-translated document, and there may be multiple correction candidates. Therefore, multiple to-be-translated documents each having different corrections may be created, and the multiple to-be-translated documents may be translated to generate multiple translated documents based on the respective corrections. For example, the to-be-translated document may be repeatedly corrected until the above-mentioned score reaches a reference value. Alternatively, for example, a predetermined number of multiple to-be-translated documents each having different corrections and translation documents corresponding to the multiple to-be-translated documents may be created, and the translation document with the highest score may be output as the translation result.

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

[0047] [Step A1: Obtain the translation document SD1] First, the to-be-translated document SD1 is obtained. The to-be-translated document SD1 may contain one or more sentences.

[0048] For example, the document SD1 to be translated is supplied from the outside as text data, voice data, or image data. When voice data or image data is supplied, text data is created based on that data.

[0049] The document to be translated is not particularly limited, and may be, for example, a document related to intellectual property. Specific examples of documents related to intellectual property include specifications, claims, and abstracts used in patent applications. Furthermore, documents related to intellectual property include publications such as patent documents (unexamined patent publications, patent gazettes, etc.), utility model publications, design publications, and papers. Publications published in countries around the world, not limited to publications published domestically, may be used as documents related to intellectual property.

[0050] In addition, the document to be translated may be an e-mail, a book, a newspaper, a paper, a report, a column, or any other written work that contains text. The document to be translated may also be a specification, a medical document, or the like.

[0051] Furthermore, there is no particular limitation on the language of the document to be translated. For example, documents written in Japanese, English, Chinese, Korean, etc. can be used.

[0052] [Step A2: Neural machine translation is performed on the to-be-translated document SD1 to generate a translated document TD1] Next, as shown in FIG. 2(A), the processing unit 103 uses a neural network NN to translate the to-be-translated document SD1, thereby generating a translated document TD1.

[0053] As the neural network NN, for example, a recurrent neural network (RNN) that can handle variable-length data such as text data can be used.

[0054] There are no particular limitations on the model or mechanism used in neural machine translation. For example, a sequence-to-sequence model, a transformer model, an attention mechanism, etc. can be used.

[0055] The language of the translation document is not particularly limited as long as it is different from the language of the document to be translated, and for example, Japanese, English, Chinese, Korean, etc. can be used.

[0056] [Step A3: Determine the correction target sentence CTS from the translated document TD1] Next, the translated document TD1 is analyzed to determine sentences that may be mistranslated (sentences to be corrected CTS). Figure 2(B) shows an example in which sentences that may be mistranslated and are included in the translated document TD1 are determined as sentences to be corrected CTS.

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

[0058] In analyzing the translated document TD1, it is preferable to detect at least one of the following: sentences in which the same word (or phrase) is repeated, sentences that contain words (or phrases) that are not included in the translated document SD1, sentences that contain grammatical errors, and sentences that contain special symbols.

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

[0060] For example, a word dictionary prepared in advance may be used to compare nouns contained in the to-be-translated document SD1 with nouns contained in the translated document TD1.

[0061] In addition, the translation document TD1 may be back-translated into the language of the original document SD1, and the sentences in the original document SD1 may be compared.

[0062] To detect sentences that may be mistranslated, the probability value of the output result of the neural machine translation model may be analyzed. In neural machine translation, the word with the highest probability in the translation language is output, taking into account which words in the translated sentence are most closely related (weight coefficients in the middle layer). If the probability of the word is low, the sentence containing the word can be determined to be possibly mistranslated.

[0063] In order to detect sentences that may be mistranslated, information on past translation results stored on a server, etc. may be used. For example, information on the strength of the word association between the original sentence and the translated sentence (weight coefficients in the intermediate layer of a neural network NN) may be stored, and if the association with words different from the past is stronger, it may be determined that there is a possibility of mistranslation.

[0064] The number of sentences to be revised may be one or more. When there are multiple sentences to be revised, for example, steps A4 and A5 may be performed for each sentence, and then the process may proceed to step A6. Alternatively, steps A4 to A6 may be performed for a certain sentence, and then the process may return to step A3 or A4 to repeat the same process for the other sentences.

[0065] [Step A4: Determine the correction target phrase from among the phrases contained in the sentence S1 in the translation document SD1 that corresponds to the correction target sentence CTS] Next, a phrase (a phrase to be corrected) that is predicted to be the cause of the mistranslation is determined from among the phrases contained in the document SD1 to be translated. FIG. 2(B) shows an example in which the document SD1 to be translated has a sentence S1 that corresponds to the sentence to be corrected CTS, and a phrase to be corrected is determined from among the phrases contained in the sentence S1. The phrase to be corrected may be either a word or a phrase.

[0066] Among the words contained in sentence S1, words that are predicted to be the cause of the mistranslation include, for example, words that are used relatively infrequently among the words used when training the neural network NN (i.e., words that are not replaced with special symbols as unknown words). Therefore, by referring to information on the frequency of use of words in the training data used to train the neural network NN, it is possible to determine that the least frequently used word or a phrase that includes the least frequently used word is the word to be corrected.

[0067] The frequency of use of a word can be obtained by, for example, calculating the Term Frequency (TF) value of the word. In addition, the Term Frequency Inverse Document Frequency (TF-IDF) value of the word may be used to determine the words to be corrected.

[0068] To detect words that are predicted to cause mistranslations, the weight coefficients of the intermediate layers in the neural network NN may be analyzed. In neural machine translation, by referring to the weight coefficients of the intermediate layers, it is possible to know which words in the translated text are closely related to the output words. For example, words that are weakly related to any words in the translated text can be determined to be words to be corrected.

[0069] The method for determining the correction target phrase is not limited to the above, and various methods can be used.

[0070] [Step A5: Replace the correction target phrase with a high-frequency word in the training data, and correct the translated document SD1] Next, the to-be-translated document SD1 is corrected by replacing the to-be-corrected phrase with a high-frequency word in the training data used to train the neural network NN, to generate a to-be-translated document SD2. In the following, the to-be-translated document SD2 after correction shown in FIG. 2(B) has a sentence S1' in which the to-be-corrected phrase in sentence S1 has been replaced with another phrase.

[0071] By replacing the words to be corrected with high-frequency words in the training data used to train the neural network NN, the translation accuracy in neural machine translation can be improved.

[0072] However, it is undesirable that as a result of correcting the to-be-translated document SD1 to reduce mistranslations, a translation faithful to the content of the to-be-translated document SD1 before the correction can no longer be performed. Therefore, it is preferable that the words to be corrected are replaced with synonyms or similar words of the words to be corrected. This makes it possible to prevent the content of the to-be-translated document SD1 from changing significantly even if the to-be-translated document SD1 is corrected. Therefore, it is possible to perform a translation faithful to the content of the to-be-translated document SD1 before the correction, while reducing mistranslations.

[0073] Specifically, it is preferable to extract multiple synonyms and similar words of the word to be corrected, select from among them a word that is frequently used in the training data used to train the neural network NN, and replace the word to be corrected with that word.

[0074] The synonyms and the like of the correction target phrase can be obtained, for example, by using a dictionary prepared in advance. Also, the synonyms and the like of the correction target phrase may be extracted based on the similarity or distance of the distributed representation vectors of the words.

[0075] High frequency words in the training data used to train the neural network NN can be selected based on information about the frequency of use of words in the training data.

[0076] The frequency of use of a word can be calculated, for example, by calculating the TF value of the word, or the TF-IDF value of the word may be used.

[0077] In addition, the candidate for the word (or phrase) to replace the correction target phrase may be one or more. For example, the most frequently used word in the learning data may be the candidate for the replacement word, or multiple frequently used words may be the candidates for the replacement word. When there are multiple candidates for the replacement word, multiple translation documents may be created in which the correction target phrase is replaced with different words, and multiple translation documents based on each correction may be created in step A6. In this case, as described above in step A3, if a score based on the translation accuracy can be calculated by analyzing the translation document, it is easy to determine which of the multiple translation documents is the most preferable translation result, which is preferable.

[0078] When the phrase to be corrected includes a plurality of words, each word may be replaced with another word, or only some of the words may be replaced with another word.

[0079] [Step A6: Neural machine translation is performed on the corrected translation document SD2 to generate a translated document TD2] Next, as shown in FIG. 2(C), the corrected to-be-translated document SD2 is translated using the neural network NN of the processing unit 103 to generate a translated document TD2.

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

[0081] For example, the translated document TD2 is output to the outside as text data, voice data, or image data. When voice data or image data is output, the voice data or image data is created based on the text data generated in step A6.

[0082] In the machine translation method of the present embodiment, an example in which the processing is performed automatically by a machine has been described, but some processing may be performed by a human being. For example, the specification of the sentence to be corrected in step A3 and the specification of the word or phrase to be corrected in step A4 may be performed by a human being.

[0083] Furthermore, the data used in the machine translation method of this embodiment and the generated data may be appropriately stored in a server, memory, etc. For example, at least one of the to-be-translated document acquired in step A1, the translation document generated in steps A2 and A6, and the corrected to-be-translated document generated in step A5 may be stored in a server, memory, etc.

[0084] In addition, when retraining the neural network, one or both of the set of the original document SD1 and the translation document TD1 and the set of the original document SD2 and the translation document TD2 may be added to the training data. This can improve the translation accuracy in neural machine translation.

[0085] <Machine translation method example 2> A flowchart of the machine translation method is shown in Fig. 3. As shown in Fig. 3, the machine translation method according to one embodiment of the present invention has nine steps, namely, step A11 to step A19. Note that detailed explanations of the same steps as those in step A1 to step A7 above will be omitted.

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

[0087] [Step A12: Neural machine translation is performed on the to-be-translated document to generate a translated document] Next, similarly to step A2, the to-be-translated document is translated using a neural network to generate a translated document.

[0088] [Step A13: Check for errors] Next, an error check is performed. In the error check, the translated document generated prior to this step (step A12 or step A18 described below) is analyzed, and a score based on the translation accuracy of the translated document is calculated. The score serves as an evaluation index for the translation accuracy. Therefore, when error correction is performed multiple times and multiple correction proposals (translated documents) are obtained, the score can be used to determine which translated document has the highest translation accuracy.

[0089] [Step A14: Determine whether the error correction termination condition is met] Next, it is determined whether the error correction end condition is met. If so, the process proceeds to step A19. If not, the process proceeds to step A15.

[0090] The condition for ending the error correction is not particularly limited. For example, the error correction may be ended when the score calculated in step A13 reaches a reference value (mistranslations are sufficiently reduced). Alternatively, the error correction may be ended when the error correction is performed a predetermined number of times.

[0091] [Step A15: Determine the sentence to be corrected] Next, similarly to step A3, sentences that may be mistranslated (sentences to be corrected) are determined from the translation document. The sentences to be corrected may be determined using the results of the error check performed in step A13.

[0092] [Step A16: Determine the correction target phrase from among the phrases contained in the sentence in the translation document corresponding to the correction target phrase] Next, similarly to step A4, a phrase that is predicted to be the cause of the mistranslation (a phrase to be corrected) is determined from among the phrases contained in the document to be translated.

[0093] [Step A17: Replace the correction target phrase with a high-frequency word in the training data, and correct the translated document] Next, similarly to step A5, the translation document is corrected by replacing the words to be corrected with high frequency words in the training data used for training the neural network.

[0094] [Step A18: Neural machine translation is performed on the corrected translation document to generate a translated document] Next, in the same manner as in step A6, the corrected to-be-translated document is translated using a neural network to generate a translated document. After step A18, the process proceeds to step A13.

[0095] [Step A19: Output the translated document] If the error correction termination condition is met in step A14, the translated document is output. For example, if the error correction is terminated because the score calculated in step A13 reaches a reference value, it is preferable to output the translated document whose score has reached the reference value. Also, if the error correction is terminated after a predetermined number of error corrections, it is preferable to compare the scores of a predetermined number of translated documents and output the translated document with the highest translation accuracy. The number of translated documents to be output may be one or more.

[0096] As described above, in the machine translation method of this embodiment, neural machine translation is performed once, and the words to be corrected in the to-be-translated document are determined based on the translation result. Therefore, excessive correction of the to-be-translated document can be prevented, and a translated document that is faithful to the to-be-translated document can be generated. Furthermore, when correcting a word in the to-be-translated document, the word is replaced with a high-frequency word in the training data used to train the neural network. This can improve the translation accuracy in neural machine translation.

[0097] This embodiment mode can be combined with other embodiment modes as appropriate. In addition, in the case where a plurality of configuration examples are shown in one embodiment mode in this specification, the configuration examples can be combined as appropriate.

[0098] (Embodiment 2) In this embodiment, a machine translation system according to one embodiment of the present invention will be described with reference to FIGS.

[0099] The machine translation system of this embodiment can translate one natural language into another natural language by using the machine translation method shown in embodiment 1. Therefore, it is possible to translate a document to be translated faithfully, and the translation accuracy is high.

[0100] <Machine translation system configuration example 1> 4 shows a block diagram of the machine translation system 100. In the drawings attached to this specification, the components are classified by function and shown as independent blocks in the block diagram, but in reality, it is difficult to completely separate the components by function, and one component may be involved in multiple functions. Also, one function may be involved in multiple components, and for example, the processing performed by the processing unit 103 may be executed by different servers depending on the processing.

[0101] The machine translation system 100 includes at least a processing unit 103. The machine translation system 100 shown in FIG.

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

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

[0104] Methods for inputting the document to be translated include, for example, key input using a keyboard or touch panel, voice input using a microphone, reading from a recording medium, image input using a scanner or camera, and acquisition via communication.

[0105] It is preferable that the machine translation system 100 has a function of converting voice data into text data. For example, the processing unit 103 may have this function. Alternatively, the machine translation system 100 may further have a voice conversion unit having this function.

[0106] The machine translation system 100 preferably has an optical character recognition (OCR) function. This makes it possible to recognize characters contained in image data and create text data. For example, the processing unit 103 may have this function. Alternatively, the machine translation system 100 may further have a character recognition unit having this function.

[0107] [Transmission path 102] The transmission path 102 has a function of transmitting various data. Data can be transmitted and received between the input unit 101, the processing unit 103, the storage unit 105, the database 107, and the output unit 109 via the transmission path 102. For example, data such as a document to be translated and a translated document is transmitted and received via the transmission path 102.

[0108] [Processing section 103] The processing unit 103 has a function of performing calculations using data supplied from the input unit 101, the storage unit 105, the database 107, etc. The processing unit 103 can supply the calculation results to the storage unit 105, the database 107, the output unit 109, etc.

[0109] The processing unit 103 has a function of performing neural machine translation to generate a translated document. For example, the processing unit 103 can translate a document to be translated that is input to the input unit 101 or a document to be translated that is corrected by the processing unit 103.

[0110] The processing unit 103 has a function of determining a sentence to be corrected from the translation document. The processing unit 103 also has a function of determining a phrase to be corrected from among phrases included in a sentence in the to-be-translated document corresponding to the sentence to be corrected. The processing unit 103 has a function of replacing the phrase to be corrected with a high-frequency word in the learning data, correcting the to-be-translated document, and generating a corrected to-be-translated document.

[0111] The processing unit 103 preferably uses a transistor having a metal oxide in a channel formation region. Since the off-state current of the transistor is extremely low, the transistor can be used as a switch for holding charge (data) flowing into a capacitor element functioning as a memory element, thereby ensuring a long data holding period. By using this characteristic in at least one of the register and the cache memory of the processing unit 103, the processing unit 103 can be operated only when necessary, and the processing unit 103 can be turned off in other cases by saving information of the immediately preceding process in the memory element. In other words, normally-off computing is possible, and the power consumption of the machine translation system can be reduced.

[0112] Note that in this specification and the like, a transistor in which an oxide semiconductor or a metal oxide is used for a channel formation region is referred to as an oxide semiconductor transistor or an OS transistor. The channel formation region of an OS transistor preferably contains a metal oxide.

[0113] In this specification and the like, the term "metal oxide" refers to an oxide of a metal in a broad sense. Metal oxides are classified into oxide insulators, oxide conductors (including transparent oxide conductors), oxide semiconductors (also referred to as oxide semiconductors or simply OSs), and the like. For example, when a metal oxide is used in a semiconductor layer of a transistor, the metal oxide may be referred to as an oxide semiconductor. In other words, when a metal oxide has at least one of an amplifying function, a rectifying function, and a switching function, the metal oxide can be referred to as a metal oxide semiconductor, or an OS for short.

[0114] The metal oxide in the channel formation region preferably contains indium (In). When the metal oxide in the channel formation region contains indium, the carrier mobility (electron mobility) of the OS transistor is increased. The metal oxide in the channel formation region is preferably an oxide semiconductor containing an element M. The element M is preferably aluminum (Al), gallium (Ga), or tin (Sn). Other elements that can be used for the element M include boron (B), silicon (Si), titanium (Ti), iron (Fe), nickel (Ni), germanium (Ge), yttrium (Y), zirconium (Zr), molybdenum (Mo), lanthanum (La), cerium (Ce), neodymium (Nd), hafnium (Hf), tantalum (Ta), and tungsten (W). However, the element M may be a combination of a plurality of the above elements. The element M is, for example, an element having a high bond energy with oxygen. For example, it is an element that has a higher bond energy with oxygen than indium. In addition, the metal oxide in the channel formation region preferably contains zinc (Zn). Metal oxides containing zinc may be easily crystallized.

[0115] The metal oxide in the channel formation region is not limited to a metal oxide containing indium. The semiconductor layer may be a metal oxide containing zinc but not indium, such as zinc tin oxide or gallium tin oxide, a metal oxide containing gallium, or a metal oxide containing tin.

[0116] The processing section 103 may be formed using a transistor including silicon in a channel formation region.

[0117] In addition, the processing portion 103 preferably includes a combination of a transistor including an oxide semiconductor in a channel formation region and a transistor including silicon in a channel formation region.

[0118] The processing unit 103 includes, for example, an arithmetic circuit or a central processing unit (CPU: Central Processing Unit).

[0119] The processing unit 103 may have a microprocessor such as a DSP (Digital Signal Processor) or a GPU (Graphics Processing Unit). The microprocessor may be realized by a PLD (Programmable Logic Device) such as an FPGA (Field Programmable Gate Array) or an FPAA (Field Programmable Analog Array). The processing unit 103 can perform various data processing and program control by interpreting and executing instructions from various programs using the processor. Programs that can be executed by the processor are stored in at least one of a memory area of ​​the processor and the storage unit 105.

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

[0121] The RAM may be, for example, a dynamic random access memory (DRAM) or a static random access memory (SRAM), and a virtual memory space is allocated and used as a working space for the processing unit 103. The operating system, application programs, program modules, program data, lookup tables, and the like stored in the storage unit 105 are loaded into the RAM for execution. The data, programs, and program modules loaded into the RAM are each directly accessed and operated by the processing unit 103.

[0122] ROM can store BIOS (Basic Input / Output System) and firmware that do not require rewriting. Examples of ROM include mask ROM, OTPROM (One Time Programmable Read Only Memory), EPROM (Erasable Programmable Read Only Memory), etc. Examples of EPROM include UV-EPROM (Ultra-Violet Erasable Programmable Read Only Memory), which allows the erasure of stored data by exposure to ultraviolet light, EEPROM (Electrically Erasable Programmable Read Only Memory), flash memory, etc.

[0123] [Storage section 105] The storage unit 105 has a function of storing a program executed by the processing unit 103. The storage unit 105 may also have a function of storing, for example, the calculation results generated by the processing unit 103 (translated document, error check results, a list of correction candidate sentences, a list of words and phrases to be corrected, the corrected translated document, etc.) and data input to the input unit 101 (translated document, etc.).

[0124] The storage unit 105 has at least one of a volatile memory and a non-volatile memory. The storage unit 105 may have a volatile memory such as a DRAM or an SRAM. The storage unit 105 may have a non-volatile memory such as a ReRAM (Resistive Random Access Memory, also called a resistance change type memory), a PRAM (Phase change Random Access Memory), a FeRAM (Ferroelectric Random Access Memory), an MRAM (Magnetoresistive Random Access Memory, also called a magnetoresistive type memory), or a flash memory. The storage unit 105 may also have a recording media drive such as a hard disk drive (HDD) and a solid state drive (SSD).

[0125] [Database 107] The database 107 has at least a function of storing information on the frequency of use of words in the learning data used to learn the neural network NN. The database 107 may also have a function of storing, for example, the calculation results (translated document, error check results, a list of correction candidate sentences, a list of words to be corrected, the corrected translated document, etc.) generated by the processing unit 103, and data (translated document, etc.) input to the input unit 101. The storage unit 105 and the database 107 may not be separated from each other. For example, the machine translation system may have a storage unit having the functions of both the storage unit 105 and the database 107.

[0126] The memories included in the processing unit 103, the storage unit 105, and the database 107 can each be considered as an example of a non-transitory computer-readable storage medium.

[0127] [Output section 109] The output unit 109 has a function of supplying data to the outside of the machine translation system 100. For example, it can supply the result of the calculation in the processing unit 103 to the outside. For example, it can supply a translated document corresponding to an input document to be translated to the outside.

[0128] The machine translation system 100 may have a function of generating either or both of audio data and image data using the text data of a translation document.

[0129] <Machine translation system configuration example 2> 5 shows a block diagram of a machine translation system 150. The machine translation system 150 includes a server 151 and a terminal 152 (such as a personal computer).

[0130] The server 151 includes a communication unit 161a, a transmission path 162, a processing unit 163a, and a database 167. Although not shown in FIG. 5, the server 151 may further include a storage unit, an input / output unit, and the like.

[0131] The terminal 152 includes a communication unit 161b, a transmission path 168, a processing unit 163b, a storage unit 165, and an input / output unit 169. Although not shown in FIG. 5, the terminal 152 may further include a database and the like.

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

[0133] The document to be translated received by the communication unit 161a is stored in a database 167 or a storage unit (not shown) via the transmission path 162. Alternatively, the document to be translated may be directly supplied from the communication unit 161a to the processing unit 163a.

[0134] The generation of a translation document, the determination of a sentence to be corrected, the determination of a phrase to be corrected, and the correction of a document to be translated, which have been described in the first embodiment, each require high processing power. The processing unit 163a of the server 151 has a higher processing power than the processing unit 163b of the terminal 152. Therefore, it is preferable that each of these processes is performed by the processing unit 163a.

[0135] Then, a translated document is generated by processing unit 163a. ​​The translated document is stored in database 167 or a storage unit (not shown) via transmission path 162. Alternatively, the translated document may be directly supplied from processing unit 163a to communication unit 161a. Thereafter, the translated document is output from server 151 to terminal 152. The translated document is transmitted from communication unit 161a to communication unit 161b.

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

[0137] [Transmission path 162 and transmission path 168] The transmission paths 162 and 168 have a function of transmitting data. Data can be transmitted and received between the communication unit 161a, the processing unit 163a, and the database 167 via the transmission path 162. Data can be transmitted and received between the communication unit 161b, the processing unit 163b, the storage unit 165, and the input / output unit 169 via the transmission path 168.

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

[0139] [Storage section 165] The storage unit 165 has a function of storing a program executed by the processing unit 163b. The storage unit 165 also has a function of storing the calculation result generated by the processing unit 163b, the data input to the communication unit 161b, 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 to learn the neural network NN. The database 167 may also have a function of storing the calculation results generated by the processing unit 163a and the data input to the communication unit 161a. Alternatively, 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 Unit 161a and Communication Unit 161b] Using the communication units 161a and 161b, data can be transmitted and received between the server 151 and the terminal 152. A hub, a router, a modem, or the like can be used as the communication units 161a and 161b. Data can be transmitted and received using either a wired connection or wirelessly (for example, radio waves, infrared rays, or the like).

[0142] This embodiment mode can be combined with other embodiment modes as appropriate. [Explanation of symbols]

[0143] CTS Correction Target Sentence S1 sentence TD1 Translation Documents TD2 Translation Documents SD1 Translated document SD2 Translated documents 100 Machine Translation Systems 101 Input section 102 Transmission Line 103 Processing section 105 Storage section 107 Database 109 Output section 150 Machine Translation System 151 Server 152 terminals 161a Communications Department 161b Communications Department 162 Transmission Line 163a Processing section 163b Processing section 165 Storage section 167 Database 168 Transmission Line 169 Input / output section

Claims

1. A processing unit of the machine translation system translates the to-be-translated document using a neural network to generate a first translated document; A processing unit of the machine translation system determines a sentence suspected of being mistranslated from the first translated document; A processing unit of the machine translation system determines a correction target phrase from among phrases included in the translation target document corresponding to the sentence suspected of being mistranslated, A processing unit of the machine translation system corrects the translation target phrase by replacing the correction target phrase with a high frequency word in training data used for training the neural network, A processing unit of the machine translation system translates the corrected to-be-translated document using the neural network to generate a second translated document; The sentence suspected of being mistranslated is at least one of a sentence in which the same word is repeated, a sentence containing a word not contained in the translated document, a sentence containing a grammatical error, and a sentence containing a special symbol; The correction target phrase is a least frequently used word or a phrase including a least frequently used word in training data used to train the neural network, A machine translation method, wherein the high frequency words are selected from among synonyms and similar words of the correction target phrase.

2. A processing unit of the machine translation system translates the to-be-translated document using a neural network to generate a translated document; A processing unit of the machine translation system analyzes the translated document to obtain a score based on the translation accuracy of the translated document; a processing unit of the machine translation system repeats correcting the to-be-translated document, generating a translated document based on the correction, and obtaining the score of the translated document based on the correction, until a termination condition is satisfied; The modification of the translated document includes: A process in which a processing unit of the machine translation system determines a sentence that is suspected to be mistranslated from the translation document, and a process in the machine translation system determines a correction target phrase from among phrases included in the translation document that correspond to the sentence that is suspected to be mistranslated; and correcting the document to be translated by a processing unit of the machine translation system by replacing the correction target phrase with a high frequency word in training data used for training the neural network, The generation of the translation document based on the correction includes: a step of a processor of the machine translation system translating the to-be-translated document reflecting the latest revisions using the neural network to generate a translation document based on the latest revisions; The sentence suspected of being mistranslated is at least one of a sentence in which the same word is repeated, a sentence containing a word not contained in the translated document, a sentence containing a grammatical error, and a sentence containing a special symbol; The correction target phrase is a least frequently used word or a phrase including a least frequently used word in training data used to train the neural network, A machine translation method, wherein the high frequency words are selected from among synonyms and similar words of the correction target phrase.

3. A processing unit is provided. The processing unit includes: A function of generating a first translated document by translating the to-be-translated document using a neural network; determining a sentence from the first translation document that is suspected to be a mistranslation; A function of determining a phrase to be corrected from among phrases contained in the translation document corresponding to the sentence suspected to be mistranslated; A function of correcting the translation document by replacing the correction target phrase with a high frequency word in training data used for training the neural network; and a function of translating the corrected to-be-translated document using the neural network to generate a second translated document, The sentence suspected of being mistranslated is at least one of a sentence in which the same word is repeated, a sentence containing a word not contained in the translated document, a sentence containing a grammatical error, and a sentence containing a special symbol; The correction target phrase is a least frequently used word or a phrase including a least frequently used word in training data used to train the neural network, A machine translation system, wherein the high frequency words are selected from among synonyms and similar words of the correction target phrase.

4. generating a first translated document by translating the to-be-translated document using a neural network; determining a sentence from the first translation document that is suspected to be mistranslated; determining a correction target phrase from among phrases included in the translation document corresponding to the sentence suspected to be mistranslated; correcting the to-be-translated document by replacing the correction target phrase with a high-frequency word in training data used to train the neural network; and generating a second translated document by translating the corrected to-be-translated document using the neural network; The sentence suspected of being mistranslated is at least one of a sentence in which the same word is repeated, a sentence containing a word not contained in the translated document, a sentence containing a grammatical error, and a sentence containing a special symbol; The correction target phrase is a least frequently used word or a phrase including a least frequently used word in training data used to train the neural network, The high frequency words are selected from among synonyms and anaphoric words of the correction target phrase.

5. A non-transitory computer-readable storage medium having the program according to claim 4 stored thereon.

Images