JP6649124B2 - Machine translation method, machine translation device and program - Google Patents

Description

  The present invention relates to a device for translating between languages, a method for translating between languages, and a system for translating between languages.

  With the recent globalization, a machine translation device and a machine translation system that enable communication between users having different languages as their native languages are being developed. In addition, the operation of a service that provides a machine translation function has also been started, and is being used in situations such as travel conversations.

Japanese Patent No. 5653392 JP 2005-78318 A Japanese Patent No. 5097340 WO 2013/014877

Papineni, K .; Roukos, S .; , Ward, T.W. , And Zhu, W.W. J. "BLEU: a method for automatic evaluation of machine translation", Proc. pp. of the Annual Meeting of the Association of Computational Linguistics (ACL), 311-318, 2002 Natsume Takachi and Hideki Isozaki, "Verification of Automatic Evaluation of Japanese Translation Considering Scrambling Using NTCIR-9 Data", 219th Natural Language Processing Research Conference, Okayama Prefectural University, 2014 "Sequence to Sequence Learning with Neural Networks", Ilya Sustakever Papineni, and others in Advances in Neural Information Processing Systems, 27. 3104-3112, 2014 Dzmitry Bahdanau et al., "Neural Machine Translation by Jointly Learning to Align and Translate", arXiv: 1409.0473v5, ICLR 2015

  However, the above-mentioned machine translation apparatus and machine translation system required further improvement.

  One embodiment of the present invention for solving the above problem is a machine translation method in a machine translation system connected to an information output device that outputs linguistic information and performing a translation process between a first language and a second language. Receiving the translation target sentence in the first language, generating a plurality of different forward translation sentences by translating the received translation target sentence into the second language, and for each of the plurality of different forward translation sentences Generating a plurality of back-translated sentences back-translated into the first language and selecting one back-translated sentence from the plurality of back-translated sentences when the information output device is outputting the plurality of back-translated sentences; When receiving the operation to perform the forward translation, the forward translation corresponding to the one reverse translation is output.

  According to the above aspect, further improvement was realized.

FIG. 1 is a diagram illustrating an example of an overall configuration of a machine translation system according to the present embodiment. FIG. 3 is a block diagram illustrating an information display terminal configuration according to the present embodiment. FIG. 3 is a block diagram illustrating a configuration of a translation server in the present embodiment. FIG. 2 is a diagram illustrating a hardware configuration of a machine translation system according to the present embodiment. 5 is a flowchart illustrating an operation of the machine translation system according to the present embodiment. It is a flowchart in this Embodiment which shows the specific operation | movement of the translation sentence selection process. It is a flowchart in this Embodiment which shows the specific operation | movement of the reverse translation sentence selection process. It is a flowchart in this Embodiment which shows the specific operation | movement of the phrase evaluation process. It is a flowchart in this Embodiment which shows the specific operation | movement of a learning process. It is a flowchart in this Embodiment which shows the specific operation | movement of a learning process. It is an example of a general phrase table in the present embodiment. FIG. 3 is an explanatory diagram showing an outline of phrase division in the present embodiment. (A), (B), (C) is a figure which shows an example of the display screen in this Embodiment, respectively. FIG. 3 is a diagram illustrating an example of a display screen in the present embodiment.

  Each of the embodiments described below shows a specific example of the present invention. Numerical values, shapes, components, steps, order of steps, and the like shown in the following embodiments are merely examples, and do not limit the present invention. In addition, among the components in the following embodiments, components not described in the independent claims indicating the highest concept are described as arbitrary components. Further, in all the embodiments, the respective contents can be combined.

(Knowledge that led to the present invention)
In a situation where translation devices are being actively developed, it is desired that the accuracy of machine translation be perfect in order to realize smooth communication in different languages. However, in the current machine translation, it is extremely difficult to translate an arbitrary sentence without error, and the translation accuracy is enhanced by limiting the field (domain) that can be translated in a form such as travel conversation. Is far from the situation.

  Here, machine translation techniques are roughly classified into three types. In this volume, 1) Rule-based Machine Translation (RBMT), 2) Statistical Machine Translation (SMT), 3) Model Acquisition Machine Translation by Deep Neural Network (DNNMT: Deep) It will be referred to as “Neural Network Machine Translation”.

1) Rule-based machine translation (RBMT)
Rule-based machine translation (RBMT) is a system that translates based on conversion rules (translation word pairs stored as a database) constructed by hand. The database of original and translated sentences is expressed as a translation memory. There is also.

  An advantage is that, because the rules (patterns to be translated) can be strictly defined, if (a part of) the original text exists in the rules, the translation accuracy of the corresponding parts is increased, and the translation expression in the rules is increased. Are prepared in advance, it is easy to maintain consistency in the translation expression of the translation output. "

  Disadvantages include: "If there are no rules, translation accuracy will be very low, or translation will not be possible at all, and adaptability to unexpected industries and fields will be extremely low." In addition, since the rules are basically constructed and expanded manually, the cost for development is also high. Furthermore, it is necessary to add rules in order to improve translation performance, but considerable knowledge is required for rule design rules in order to construct rules on the user side and customize the translation system. However, it is not intended to be easily used by general users. For this reason, the main destination of rule-based translation (RBMT) is business translation (patent translation, etc.), and commercial translation software that uses this rule-based translation or translation memory is not available. Many.

2) Statistical machine translation (SMT)
Statistical machine translation (SMT) does not create rules like RBMT, but prepares a large number of “translation pairs in both the source language and target language” called a bilingual corpus, and uses this bilingual corpus for statistical Is a method of calculating the translation probability and generating a translated sentence. Google (registered trademark), Microsoft (registered trademark) and the like are widely used in translation engines provided on the web.

  When calculating the translation probability, this method uses a language model (generally expressing the occurrence probability between words in a monolingual language with an n-gram model) and a translation model (word or phrase between bilingual languages). In general, the correspondence between units is expressed as an alignment model, and the translated word pair relation between words at the vocabulary level is generally expressed as a vocabulary model.) Is calculated, and a parameter (probability value) of each model is obtained to realize statistical translation. It is generally called an IBM model. (If Japanese is included, a model called “rearranged model” may be added by parsing or the like).

  The advantage of this SMT is that it does not use rules like RBMT, so it is possible to avoid the problem that "the translation accuracy becomes very low when there are no rules", and a translation engine with generalization ability. Can be constructed. "

  On the other hand, the disadvantage is that it is necessary to prepare a large amount of bilingual corpora on which the probability calculation is based because it is based on stochastic expressions. In addition, while the generalization performance is improved, the translation result is merely an output having a “high stochastic score”. Therefore, there is no guarantee that the same translation accuracy can be obtained in the SMT for the case where the translation was successfully performed in the RBMT. Although the translation is relatively good as a whole / average, there are some cases where the individual translation cases are less reliable.

  Further, when performing the probability calculation, it is necessary to multiply the probability values output from the internal models (language model, translation model, etc.). At this time, in order to tune the entire SMT translation engine, machine learning is performed by weighting the probability value from each model and using the weight as a parameter. The machine evaluation value used here is literally “mechanical evaluation value”. (For example, an evaluation value called BLEU) ", it is reported that the evaluation value does not always match the subjective evaluation of the user (for example, see Non-Patent Document 2). In other words, it has a structural drawback such as "The mechanical evaluation value is high but is not linked to the user's actual evaluation".

  Also, even if we try to improve the performance on the user side or system side, even if we add a few translation corpora, the translation output is affected by the stochastic behavior after statistically processing the entire translation corpus. Another structural disadvantage is that it does not directly lead to performance improvement (performance may decrease).

3) Model-based machine translation using deep neural network (DNNMT)
Model acquisition type machine translation (DNNMT) using a deep neural network is a relatively new translation technology that applies deep learning (DL) technology to machine translation (for example, see Non-Patent Documents 3 and 4). ).

  Unlike translation methods that apply / statistically apply an input sentence or a bilingual corpus to “rules and models designed by humans” such as RBMT and SMT, the system is characterized by learning an appropriate model itself. At the present time, it is an early technology and has not yet been put into practical use as machine translation, but similar technology has already been put to practical use in Siri (registered trademark) of Apple Inc. in speech recognition.

  As an advantage, when learning of an appropriate model is successful, improvement in translation performance is expected. In particular, generalization performance exceeding SMT, and relatively long sentences that RBMT or SMT are weak (for example, 40 or more It is expected that the translation result will be output without significantly degrading the translation performance of the sentence composed of the words of

  The disadvantage is that external control of how DNNMT itself learns the model is difficult. There is no explicit "parameter" corresponding to so-called "tuning using parameters", and although there are many internal variables, it is unclear which variable is directly connected to translation performance. In other words, after determining the configuration of the neural network and the input data and teacher data, it is difficult to control how the system learns and what kind of performance it will produce, and it is unclear in principle how to improve the performance. . It is difficult to apply a general tuning method, and it is difficult to improve performance on the user side or the system side more than SMT. Even if a few sentences are added to the bilingual corpus, the translation output is affected by the learning of the neural network, and there is a structural disadvantage that what is to be output cannot be known unless the internal variables of the neural network are determined.

  The above are the features, advantages, and disadvantages when machine translation techniques are roughly classified into three types.

  On the other hand, when looking at machine translation technology from the user's point of view, if the user has insight into the target language, which is the language to be translated, even if the accuracy of the machine translation is not very high, the conversation will proceed based on the translation output result Could be possible. In other words, it can be expected that the user can judge "correct / incorrect" for the output of the machine translation, and use the useful translation part.

  However, in practice, even if the user refers to the output result of the machine translation, it is easy to imagine a case where the contents indicated by the translation result cannot be understood. For example, if the user is assumed to be Japanese, there may be cases where we want to translate Japanese into major languages (such as English), and cases where we want to translate into relatively minor languages (such as Malay or Vietnamese). There is also. For example, when the target language is English, a user who has knowledge about English may be able to use the result of the machine translation to help his or her conversation in English. On the other hand, if the target language is a minor language, the user often has no knowledge of the target language, and the user can obtain no information about whether the contents of the machine translation result are correct. Absent. That is, even if a translation of the target language is presented by machine translation, there are cases where the contents cannot be understood at all. It is said that there are thousands of languages, and it can be said that even for most users, the language without knowledge is the majority. As described above, if the meaning of the translation result of the target language cannot be understood, a translation may be presented without confirming the intention to be transmitted to the other party, and communication may be broken.

  In order to improve the accuracy of the current machine translation as much as possible, it is desired that the machine translation system itself automatically learns and improves the performance. When the machine translation system automatically learns, it is expected that the user (source language speaker) will not be burdened. It is necessary to build a machine translation system that can generate and use text. The machine translation system is required to perform automatic learning using the usage results of the user. Naturally, it is also required to minimize computer resources and development costs required for the system itself.

  That is, what is required of a machine translation system is “(requirement 1) that a user can generate and use a translated sentence without burdening the user and confirming the intention of the user in a simple manner”, “(requirement 1) 2) To be able to learn automatically using the usage results of the user (without putting a new burden on the user) "," (Requirement 3) Realization of Requirement 1 and Requirement 2 and to reduce computer resources and development cost That is the three requirements. The realization of a machine translation system that satisfies these three requirements simultaneously has become an issue.

  As a means for solving this problem, for example, as in RBMT, an example in which a source language example and a target language example are associated with each other is stored as a database or a rule / rule, and an example similar to an uttered input sentence is used. Obtain and present in both the source language and the target language (for example, Patent Document 1), calculate the distance between the input sentence and the result of reverse translation (translation from the target language to the source language) as the translation reliability, and if necessary, (For example, Patent Literature 4), and BLEU (BiLingual), a technique for evaluating the performance of a statistical machine translation (SMT) system. Evaluation Understudy (for example, Non-Patent Document 1), N back-translated sentences are obtained for an input sentence, and a comparison between the input sentence and the back-translated sentence (For example, Patent Literature 2), and machine learning of the result of determining whether a translation is effective for a target translation device when registering a bilingual phrase in a bilingual dictionary (for example, Patent Literature 3) and so on.

  Patent Literature 1 discloses a content stored in an example storage unit in which an example of a source language and an example of a target language which is a translation of the example of the source language are associated with each other in advance.

  Japanese Patent Application Laid-Open No. H11-163873 discloses, based on information stored in the above-described example storage unit, a character string of an input source language and a character string of a target language corresponding to the source language character string on a display unit. Discloses contents displayed in different display areas.

  Specifically, for the input source language character string, a plurality of similar example character strings similar to the source language character string are displayed in the same display area. Character strings of the target language corresponding to the character strings of the source language and the character strings of the plurality of similar examples are displayed in different display areas.

  The source language speaker and the target language speaker can check the example when they want to check the meaning of the input sentence. The similar sentence example selected by the target language speaker or the source language speaker is also displayed as a highlight or the like to the other speaker.

  In Patent Document 1, a character string in the source language, which is an input sentence, is acquired, a similar example in a bilingual example in a database is searched, and if the similarity is equal to or more than a threshold, it is determined that the example is a similar example. Is output as a similar example to the input sentence. At this time, when the user performs an operation of selecting the output similar example, the content of highlighting the selected example is also disclosed.

  In Patent Document 2, a forward translation sentence output from a forward machine translation unit 11 is received for an original sentence, and backward translation units A, B and C are evaluated by backward translation units 12a, 12b and 12c by an evaluation unit. The contents to be evaluated at 13 are disclosed.

  As the evaluation method at this time, the contents disclosed in Non-Patent Document 1 are generally known. In Non-Patent Document 1, after calculating the number of N-grams between the reference translation (correct translation created manually) and the translation output by machine translation, the influence of the length of the reference translation is calculated. The BLEU value is obtained by performing the added correction. BLEU is often used as an evaluation method of translation accuracy, but it is known that especially in the translation between languages having greatly different word orders such as Japanese, English, and English and Japanese, the correlation with the manual evaluation is low (for example, , Non-Patent Document 2).

  Patent Literature 2 further discloses that after performing DP matching between the back-translated sentences A, B, and C and the original, the backward-translated sentence having the maximum score and the original are output. By the evaluator comparing these, the subjective evaluation of the forward-translated sentence becomes possible.

  As described above, when performing translation in a situation where the machine translation accuracy is not perfect, the user (source language speaker) can easily and easily confirm the intention to be transmitted to the other party without burdening the user with repeated input. It was necessary for the system itself to evaluate / learn the translated utterance sentence and evaluate the user's selection as an evaluation.

  To restate what is required of a machine translation system, "(Requirement 1) A user can generate and use a translated sentence without burdening the user and confirming the intention of the user in a simple manner", "(Requirement 1) 2) To be able to learn automatically using the usage results of the user (without putting a new burden on the user) "," (Requirement 3) Realization of Requirement 1 and Requirement 2 and to reduce computer resources and development cost That is, to satisfy the three requirements at the same time.

  As means for solving this problem, the following approaches are taken in the above three types of technologies.

1) Rule-based machine translation (RBMT)
As described above, in Patent Literature 1, a similar sentence of a similar source language is output from a character string of an input source language, and a corresponding translated sentence can be selected for the output similar sentence of the source language. Output to Thus, for example, when a character string in the source language is input by voice, the effect of an input error due to a voice recognition error at the time of inputting an utterance is reduced, and the user can easily select an intention to convey.

  The paragraph number [0012] of Patent Document 1 states that “the example storage unit 105 associates an example of a source language (hereinafter also referred to as an example of a source language) with an example of a target language (hereinafter also referred to as an example of a target language). The example storage unit 105 serves as a database for performing an example search and stores pairs of the source language and the target language. This part corresponds to a rule-based translation database.

  The paragraph number [0011] of Patent Document 1 states that “the machine translation unit 103 receives the source language character string from the speech recognition unit 102 and converts the source language character string into a target language (also referred to as a second language) character string. Machine translation is performed to obtain a target language character string which is a character string of the translation result. Since specific processing of machine translation may be performed by general processing, description thereof will be omitted here. " There is no advantage or problem of using the SMT or DNNMT for the machine translation unit 103 in Patent Literature 1, and the entire speech translation system in Patent Literature 1 uses the RBMT as the RBMT by the example storage unit 105. The work is only disclosed characteristically.

  As mentioned above, the translation accuracy of RBMT is significantly reduced when there is an input sentence not described in the rules or an input of an unsupported industry or field, or translation cannot be performed at all. There is a problem that. Also in Patent Document 1, when a similar example is not described in the example storage unit 105 for an utterance input or when the field is different, there is a possibility that a translated sentence with extremely low translation accuracy may be presented, and knowledge about the target language is not obtained. If there is no user, it may lead to presenting the translation without confirming the intention transmitted to the other party (or presenting the misunderstanding of the intention transmitted to the other party), but such a problem is not disclosed and solved No measures are described.

  Furthermore, when no similar example exists in the example storage unit 105, the similar example itself is not presented, and the user has no choice but to use the input sentence (of which there are many errors) of the speech recognition result. The translation result is also only the translation result of the machine translation unit 103 for the input sentence based on the speech recognition result. In this case, the user cannot “simplely select” the “sentence according to the intention to convey appropriately”, and Patent Literature 1 does not disclose a solution to this problem.

  That is, regarding (Requirement 1), it can be said that the means for “the user to easily select (sentence)” is partially disclosed by providing a means for presenting and selecting a plurality of similar sentence search results. If there is no similar example in the example storage unit 105 corresponding to the RBMT, the result is that the option itself is lost, and there is no completely provided means for "the user to select (sentence) simply".

  Further, "confirmation of intention to be transmitted to the other party" is only limitedly solved when a search hits the content held in the example storage unit 105, and the translation result output by the machine translation unit 103 is output. Is not touched. That is, it is not mentioned whether or not the translation result of the machine translation unit 103 is in accordance with the contents of the original source language character string. The source language character string is received from the voice recognition unit 102, the source language character string is machine-translated into a character string of a target language (also referred to as a second language), and a target language character string as a character string of a translation result is obtained. Since the general process may be a general process, the description is omitted here. ", And the validity of the output of the machine translation unit 103 is not guaranteed.) Even if the content that is far from the content of the character string appears as an output of the machine translation unit 103, the user has no way of knowing it. In other words, neither the disclosure nor the solution to the above-mentioned problem of “a user without knowledge of the target language presents a translation without confirming the intention transmitted to the other party” has not been made.

  Neither (Requirement 2) nor (Requirement 3) is disclosed or suggested. Especially for requirement 3, in addition to reducing the weight of the system as a whole, in addition to the machine translation unit 103, a new example storage unit 105 corresponding to RBMT is required, which increases both computer resources and development man-hours. are doing.

  Regarding RBMT's response to "(Requirement 2) that learning can be performed automatically without imposing a new burden on the user by using the usage result of the user", an input sentence and a translated sentence as in Patent Document 3 are used. An approach is disclosed in which, when performing morphological analysis and registering a bilingual phrase in a bilingual dictionary, the result of determining whether or not the translation is effective for a target translation device is machine-learned.

  In paragraph number [0020] of Patent Document 3, "5 is a machine translation engine, for example, in rule-driven machine translation," and RBMT is exemplified as a translation system with the name of rule-driven machine translation. Discloses a method of determining the validity when registering a. In a bilingual dictionary. Here, when registering a new phrase in the rule-based bilingual dictionary, whether the registration of the phrase is valid on the system is determined by performing a morphological analysis on the previously registered bilingual pair, and then using the support vector machine. The judgment is made by referring to the judgment space identified using. That is, “registration of a new phrase in the rule base” is performed irrespective of the use of translation by the user, and no problem is disclosed or solved for automatic learning using the use result of the user.

  In general, RBMT has the disadvantage that, if there are no rules, the translation accuracy will be very low or cannot be translated at all, and the adaptability to industries and fields other than those expected is extremely low. Has not been essentially solved (for example, Patent Document 1).

  In order to improve translation performance, it is necessary to add rules, but this requires development costs. Further, in order to customize the translation system by constructing rules on the user side, considerable knowledge of rule design rules is required, so that general users cannot easily use them. In other words, as the number of rules in RBMT increases, the possibility of interference and side effects between rules increases dramatically, and adding a valid rule in one sentence example causes inconvenience in another sentence example. The phenomenon will occur frequently. "To avoid this, it is necessary to add new rules after grasping all the rules that the system has." It is difficult to add even simple rules, let alone automatic learning using. For example, in Patent Literature 3 mentioned above, among the translation rules of the RBMT, only the validity of the word translation pair is determined automatically. Not reached.

  Since the RBMT has such a fundamental disadvantage, the efficiency and simplification of “user's registration of words and examples in the RBMT database” have been disclosed as issues of the RBMT. These are for explicitly requesting the user to input and judge learning data, and that such a request should not be made to the user. With regard to “automatic learning,” no issues have been disclosed nor solved.

2) Statistical machine translation (SMT)
In SMT, the distance between the input sentence and the result of reverse translation (the input sentence is translated once into the target language and then translated again from the target language to the source language) is calculated as the translation reliability. One that presents the reverse translation result together and causes the user to rephrase or paraphrase it (for example, refer to Patent Literature 4. However, Patent Literature 4 also mentions a case where SMT is not premised. The RBMT is referred to as “grammatical rule-type translation” in paragraph number [0009] in Patent Literature 4.) As a method of evaluating the translation accuracy of the SMT, a similarity in word n-gram units is used. Automatic learning using the degree (BLEU value) (see Non-Patent Document 1), N reverse translations for a translation obtained from an input sentence are generated, and the input sentence and the reverse translation are generated. Those comparative evaluation the quality of the translation from the similarity (e.g., Patent Document 2), and the like.

  In these examples, “(Requirement 1) that a user can generate and use a translated sentence while confirming the intention to be transmitted to the user in a simple manner without putting a burden on the user” is described in these examples. A technique of presenting a translation result (Patent Literature 4) to confirm "intention transmitted to the other party" to the user is disclosed. However, regarding "the ability to generate and use a translated sentence in a simple form", the user is required to re-input or paraphrase according to the result of the reverse translation, and the problem has not been solved.

  Specifically, in the paragraph number [0013] of Patent Literature 4, "It is possible to appropriately obtain the reliability of the translation result, and to prompt the input user to re-input properly when the reliability is low. And a method for calculating and using translation reliability and a program for a translation engine. " That is, if the system determines that the reliability of the translation result is low, the user is forced not only to re-input the original text but also to change the expression of the original text until a translated text with sufficiently high reliability is output. It is necessary to keep checking the input to the translation system and the output of the translation while performing trial and error. It is also important to keep in mind that, at this time, the translation system "generates a translated sentence by what kind of internal operation", "calculates the reliability based on what criteria", and " Do you have any knowledge about "Can you get a high degree of translation?" There is no guideline, and the user is forced to input while rephrasing sentences in order to obtain a highly reliable translation result, which is a very difficult system to use from a practical point of view. . Patent Literature 4 does not propose any problem nor solve the problem, and it can be said that it is insufficient in that “a translation can be generated and used in a simple form”.

  Regarding “(Requirement 2) that learning can be performed automatically using the usage result of the user (without imposing a new burden on the user)”, “(Requirement 1) the user is not burdened, That a user can generate and use a translated sentence while easily confirming the intention to be transmitted to the other party "in both RBMT and SMT. That a translation can be generated and used while confirming the above ”has not been realized, and a method of obtaining a“ user usage result ”that satisfies (Requirement 1) has not been disclosed. Regarding the automatic learning using the “user usage result” that satisfies the condition, no problem is disclosed or no solution is presented.

  Based on the premise that "there is no method for obtaining a user's use result that satisfies (Requirement 1)", a viewpoint of an automatic evaluation and learning method of a translation system when some other learning data is given As described above, "the method of performing automatic learning based on the BLEU value, which is a method for evaluating the performance of a statistical machine translation system (see Non-Patent Document 1)", and "N reverse translation sentences for an input sentence" One that acquires and compares and evaluates the similarity between an input sentence and a back-translated sentence (for example, Patent Document 2). “When registering a bilingual phrase in a bilingual dictionary, it is necessary to determine whether the translation is valid for the target translation device. A machine learning the result of the discrimination (for example, see Patent Document 3) "and the like.

  Non-Patent Document 1 prepares a reference translation (correct data) for an input sentence in advance and compares the translation output result of the translation engine with the correct data based on a monolingual n-gram, and the comparison result is obtained. Is calculated mechanically as a value (BLEU score), and the system is tuned to increase the value. It is necessary to prepare an input sentence and a reference translation (translation sentence) as the correct answer in advance. Since the internal model of the translation engine does not change at all by this tuning but merely changes the weight, the model itself must be prepared. This method cannot be applied when learning is desired, or when correct data is not given / correct data is not uniquely determined.

  Even if the "use result of the user who satisfies the requirement 1" is obtained, the result is a result that "the user has generated and used the translated sentence while confirming the intention to be transmitted to the other party in a simple manner". Of these "results", there are no issues to be disclosed or solved regarding "whether or not the correct data", "how to uniquely make the correct data", and "how to learn the model afterwards". Not.

  Patent Document 2 discloses that the translation result is returned to the source language by a plurality of backward translators, and the similarity of the sentence between the input sentence and the plurality of backward-translated sentences is mechanically calculated. The contents for evaluating the quality of the result are disclosed. Patent Literature 2 discloses that the original translation result is evaluated by generating a plurality of reverse-translated sentences. However, the problem addressed in Patent Literature 2 is that some automatic In this respect, there is essentially no difference from the above-mentioned non-patent document 1. That is, in Patent Literature 2, some correct sentence (corresponding to a reference translation in Non-Patent Literature, equivalent to an input sentence in Patent Literature 2) and a translation result sentence (Non-Patent Literature 1 translates, and Patent Literature 2 reverses the translation A score representing a match or similarity (e.g., a score represented by the BLEU value in Non-Patent Document 1) is calculated with respect to the translated back-translation sentence), and the quality of the translation result is evaluated using the calculated value. Is disclosed.

  Note that although Patent Document 2 does not mention learning of the translation system itself, paragraph number [0048] of Patent Document 2 states that "... DP matching between three backward-translated sentences and the original sentence is performed. Since the maximum score can be the result of automatic evaluation of forward translations, it is possible to evaluate not only translations of example sentences but also all translations, reducing evaluation effort and improving reliability of evaluation. The purpose is described in paragraph [0009], "Evaluation is possible not only for translations of example sentences, but also for all translations, and the reliability of evaluation is high and labor is high. A method for evaluating a small number of machine translated sentences and a machine translation sentence evaluation device are to be constructed. Non-Patent Document 1 discloses that even when there is no correct sentence (reference translation) required in advance, Patent Document 2 discloses that evaluation can be performed by using a mechanical matching score between a back-translated sentence and an input original sentence. Thus, Patent Document 2 does not mention learning, but using this score for learning is suggested in combination with Non-Patent Document 1.

  However, even in the case where Patent Literature 2 and Non-Patent Literature 1 are combined, tuning cannot be performed if the user wants to learn the model itself, or if the matching score is incorrect or cannot be uniquely determined.

  Also, as in Non-Patent Document 1, even if “the use result of the user who satisfies the requirement 1” is obtained, the result is “the user confirms the intention to be transmitted to the other party in a simple form while confirming the translated sentence. Is generated and used. " Regarding how to judge such a result as an evaluation value, and how to perform model learning using the evaluation value, no problem is disclosed or solved.

  In other words, as for "generating and using a translated sentence while the user confirms the intention to be transmitted to the other party in a simple manner", neither the problem nor the solution is disclosed. However, as in Non-Patent Document 1, neither disclosure nor solution of the problem is disclosed.

  Regarding “(Requirement 3) Reducing computer resources and development costs while realizing requirements 1 and 2”, Non-Patent Document 1 newly calculates a BLEU value, and Patent Document 2 newly calculates a matching score. It becomes necessary, and computer resources and development costs are increasing.

3) Model-based machine translation using deep neural network (DNNMT)
Non-Patent Literature 3 and Non-Patent Literature 4 describe LSTM (Long Short Term Memory), which is a kind of deep neural net (DNN: Recursive Neural Network) and RNN, as machine translation by a neural network. Is illustrated using DNNMT. Both methods use the bilingual corpus as the correct answer data (positive and negative examples) for the input and output layers of the neural network, and directly train the intermediate layer of the neural network, so that the translation model is directly stored inside the neural network. To build. The form in which the DNN has the translation model inside depends on the type and manner of providing the learning data, the number of times of learning, the network configuration of the DNN itself, and the like. In any case, how the internal state is changed and how the translation performance changes is not disclosed (it is not clarified academically). Originally, the characteristic of a neural network is that it can learn nonlinear output. However, by using DNN, nonlinearity has dramatically increased, and a linear causal relationship between some internal parameters and output performance can be found. It is not at present.

  In other words, DNNMT is the same as RBMT or SMT in the sense that “returns some translation result for the input sentence”. However, in terms of "why the translation result was obtained", the RBMT can refer to the database describing the rules to find out why the translation result was obtained (the original rule). It can be seen that the translation result of which the maximum probability was selected from the model (the occurrence probability of each word / phrase, the alignment probability, etc.) and the language model (n-gram probability) is the translation result, whereas the DNNMT uses the rules and models. Since the neural network constructs itself equivalent to the above, there is no knowledge about the internal model or operation beyond the fact that the result output from the output layer of the neural network was a translated result sentence.

  For this reason, DNNMT is mainly focused on academic research, and has not reached the practical phase. ((Requirement 1) The translation sentence is not burdened on the user and the user confirms the intention to be transmitted to the other party in a simple manner. To be able to generate and use the same. "

  As for “(requirement 2) that learning can be automatically performed using the user's usage result (without imposing a new burden on the user)”, a provisional “user usage result that satisfies (requirement 1)” is obtained. Even if it is done, it is necessary to clarify the internal operation of "how to learn the model after that" using the result, and no problem is disclosed or solved.

  To summarize the above, the prior art has problems in the following points.

  -When similar examples or translations are not described in the bilingual database, or when the field is different, there is a possibility that a translation with extremely low translation accuracy may be presented, or translation may not be performed at all.

  -Since there is no way for the user to easily check and select the content of the translated sentence, the method of presenting a reverse-translated sentence to the input sentence or presenting the translation quality (reliability) reduces the quality and reliability of the presented content. If it is low, the user is forced to re-input, but there is no guarantee that the quality will be improved by the re-input, and the user will have to try and input the input.

  -Automatic tuning of a translation system based on a score calculated mechanically by a method such as a BLUE value has been disclosed in a prior example, but "translation is performed while the user confirms the intention transmitted to the other party in a simple manner. In the case of "generating and using a sentence", no problem is disclosed in any of evaluations based on the results and how to perform learning, and no solution is disclosed.

  Computer resources for generating similar examples and data for tuning (calculation of evaluation score, etc.) are required. Also, development and human costs for creating similar examples are required.

  Therefore, to improve the function of the machine translation system, the following improvement measures were examined.

  One aspect of the machine translation method of the machine translation system is a machine translation method in a machine translation system that connects to an information output device that outputs language information and performs a translation process between a first language and a second language. Receiving the translation target sentence of the first language, generating a plurality of different forward translation sentences by translating the received translation target sentence into the second language, and generating the first translation target sentence for each of the plurality of different forward translation sentences; Generating a plurality of back-translated sentences back-translated into a language, and selecting one back-translated sentence from the plurality of back-translated sentences when the plurality of back-translated sentences are being output in the information output device. If received, the forward-translated sentence corresponding to the one backward-translated sentence is output.

  According to the above aspect, a plurality of different forward-translated sentences generated by translating the translation target sentence of the first language into the second language are generated, and a plurality of reverse-translated sentences which are back-translated to the first language for each of the plurality of forward-translated sentences are generated. When generating and outputting a plurality of back-translated sentences on the information output device, when an operation of selecting one back-translated sentence from the plurality of back-translated sentences is received, the order corresponding to the one back-translated sentence is received. Output the translation.

  For example, a plurality of backward-translated sentences corresponding to the received translation target sentence are presented to the user, and the user is presented with a forward-translated sentence corresponding to the backward-translated sentence selected from the plurality of backward-translated sentences. Therefore, the user selects the reverse translation sentence closest to the intention of the translation target sentence inputted by himself / herself from the plurality of reverse translation sentences. When compared with a system in which only one translation sentence translated into a second language and one reverse translation sentence corresponding to the translation are presented, the reverse translation differs from the intended content of the translation target sentence. The number of situations where correction of the translation target sentence or re-input is required is reduced.

  In addition, for example, since one backward-translated sentence is selected from a plurality of backward-translated sentences by the user, the machine translation system determines the intended backward-translated sentence Among them, it is possible to obtain feedback such as which reverse-translated sentence is the most appropriate or whether it matches the user's favorite expression. Therefore, for example, when machine learning is applied to the machine translation system in the above aspect, in addition to evaluating whether or not the reverse-translated sentence presented for the input translation target sentence is appropriate, It is also possible to obtain an evaluation as to which of the back-translated sentences is appropriate among the back-translated sentences. At this time, a single translation operation in the machine translation system can provide feedback on a plurality of backward-translated sentences, so that high learning efficiency of the machine translation system can be realized.

  Further, according to the above aspect, for the machine learning in the machine translation system, the evaluation of the translation accuracy by calculating the BLEU value or the evaluation of the translation accuracy by the mechanical matching score between the input translation target sentence and the reverse translation sentence are performed. There is no need to generate learning data by selecting a backward-translated sentence by the user. Therefore, no new computer resources are required for generating the learning data, and the development cost can be reduced.

  In the above aspect, for example, the machine translation system is further connected to a voice input device that receives a voice input by a user and a text input device that receives a text input by a user, and the translation target sentence is a voice representing the translated sentence. Information, or received in the form of text information, and depending on which of the audio information or the text information received the translation target sentence, the forward-translated sentence corresponding to the one reverse-translated sentence The output form may be changed.

  According to the above aspect, the output form of the forward-translated sentence is changed depending on whether the sentence to be translated is received in the form of voice information or text information. Thereby, for example, since the output modal is determined according to the input modal, the user can freely determine the output form by changing the input form.

  In the above aspect, for example, the information output device has an audio output device and a display, and when the translation target sentence is received in the form of audio information, the information output device outputs the forward translation sentence corresponding to the one reverse translation sentence. If the sentence to be translated is output via the audio output device and the sentence to be translated is received in the form of text information, the forward translated sentence corresponding to the one reverse translated sentence may be output via the display. Good.

  According to this, since the input format and the output format correspond to each other in the same modal, the user only needs to input the translation target sentence in his / her desired output format. There is no confusion as to whether a translated sentence is output in an output form that does.

  In the above aspect, for example, the translation target sentence is received as text information indicating the translation target sentence, and a plurality of different forward-translated sentences obtained by translating the translation target sentence into the second language based on the text information are received. It may be generated.

  In the above aspect, for example, the machine translation system is further connected to a text input device that receives text input by a user, and the translation target sentence is received from the text input device as text information indicating the translation target sentence. It may be.

  In the above aspect, for example, the translation target sentence is received as voice information representing the translation target sentence, and performs voice recognition processing on the received voice information to generate text information indicating the translation target sentence, Based on the text information, a plurality of different forward-translated sentences obtained by translating the sentence to be translated into the second language may be generated.

  This allows the user to input the translation target sentence using voice. For example, there is no need to input the translation target sentence using a keyboard, a touch display, or the like, so that the user can easily input the translation target sentence.

  In the above aspect, for example, the machine translation system is further connected to a voice input device that receives a user's voice input, and the translation target sentence is received from the voice input device as voice information representing the translation target sentence. You may do it.

  In the above aspect, for example, the information output device has a display, and the plurality of reverse-translated sentences are displayed in a first area of the display, and the translation is performed in a second area different from the first area of the display. The target sentence may be displayed.

  According to the above aspect, the area to be displayed is divided between the reverse translation sentence and the translation target sentence. Since the back-translation sentence and the translation target sentence are sentences in the same language, the user can easily distinguish which is the reverse-translation sentence and which is the translation-target sentence, and is not confused.

  In the above aspect, for example, the forward-translated sentence corresponding to the one backward-translated sentence may be displayed in a third area of the display.

  As a result, each of the backward-translated sentence, the translation target sentence, and the forward-translated sentence is displayed in a different area, so that it is easy for the user to identify which sentence each is.

  In the above aspect, for example, the display direction of the forward-translated sentence corresponding to the one backward-translated sentence may be changed according to an operation on the information output device.

  Thus, for example, when the user A and the user B speak face-to-face and have a face-to-face conversation, the forward translation output in the language spoken by the user B corresponding to the translation target sentence input in the language spoken by the user A When the sentence is displayed on the information output device, if the direction of the forward-translated sentence can be changed, for example, in the reverse direction, the user A reads out the forward-translated sentence or changes the direction of the information output device itself, There is no need to inform the user B of the contents of the forward-translated sentence, and communication between different languages can be achieved by allowing two opposing users to look into the information output device from above.

  In the above aspect, for example, the display direction of the forward-translated sentence may be changed to a direction different from the display direction of the plurality of reverse-translated sentences displayed in the first area.

  In the above aspect, for example, the display direction of the forward-translated sentence may be changed to the same direction as the display direction of the plurality of reverse-translated sentences displayed in the first area.

  In the above aspect, for example, the forward-translated sentence corresponding to the one backward-translated sentence may be displayed in a different direction from the plurality of backward-translated sentences displayed in the first area.

  In the above aspect, for example, the machine translation system is a set of the forward-translated sentences obtained by translating the received sentence to be translated into the second language, and includes a forward-translated sentence group including the plurality of different forward-translated sentences. It is determined whether each of the forward-translated sentences included in the forward-translated sentence group is classified into a question sentence, an affirmative sentence, a negative sentence, or a command sentence. The translated sentence may be selected from the forward-translated sentence group based on the classified form.

  According to the above aspect, since a plurality of different forward-translated sentences are selected from the forward-translated sentence group based on the form of the sentence, for example, from the forward-translated sentence group automatically generated based on the translation target sentence, In addition, it is possible to select only the forward-translated sentence having the same form as the form of the translation target sentence, and it is possible to improve the final translation accuracy. Further, for example, a plurality of forward-translated sentences including a forward-translated sentence having a form different from the form of the translation target sentence may be selected, whereby the forward-translated sentence is generated based on the forward-translated sentence and presented to the user. Variations of multiple back-translated sentences can be increased. Therefore, for example, when applying machine learning to a machine translation system, if the user selects one backward-translated sentence from a plurality of backward-translated sentences having similar contents, even if the backward-translated sentence that is not selected is input by the user. In the case where the intention indicated by the translated target sentence is expressed, it is possible to prevent the machine translation system from learning that it is an incorrect reverse-translated sentence because it was not selected.

  In the above aspect, for example, the plurality of different forward-translated sentences may include at least two or more forward-translated sentences classified into the different forms.

  This makes it possible to increase the variations of a plurality of backward-translated sentences generated based on the forward-translated sentences and presented to the user. Therefore, for example, when applying machine learning to a machine translation system, if the user selects one backward-translated sentence from a plurality of backward-translated sentences having similar contents, even if the backward-translated sentence that is not selected is input by the user. In the case where the intention indicated by the translated target sentence is expressed, it is possible to prevent the machine translation system from learning that it is an incorrect reverse-translated sentence because it was not selected.

  In the above aspect, for example, the machine translation system generates a forward-translated sentence group that is a set of forward-translated sentences obtained by translating the received translation target sentence into the second language, and the forward-translated sentence group is A plurality of different forward-translated sentences are included, and the subject or predicate of each of the forward-translated sentences included in the forward-translated sentence group is determined. Based on the selection, it may be selected from the forward-translated sentence group.

  According to the above aspect, since a plurality of different forward-translated sentences are selected from the forward-translated sentences based on the subject or the predicate, for example, from the forward-translated sentences that are automatically generated based on the translation target sentence, Only the forward translation having the same subject or predicate as the translation target sentence can be selected, and the final translation accuracy can be improved. Further, for example, a plurality of forward-translated sentences including a forward-translated sentence having a subject or a predicate different from the translation target sentence may be selected, whereby the forward-translated sentences are generated based on the forward-translated sentences and presented to the user. Variations of multiple back-translated sentences can be increased. Therefore, for example, when applying machine learning to a machine translation system, if the user selects one backward-translated sentence from a plurality of backward-translated sentences having similar contents, even if the backward-translated sentence that is not selected is input by the user. In the case where the intention indicated by the translated target sentence is expressed, it is possible to prevent the machine translation system from learning that it is an incorrect reverse-translated sentence because it was not selected.

  In the above aspect, for example, the plurality of different forward-translated sentences may include at least two or more forward-translated sentences each determined to include a different subject or a predicate.

  This makes it possible to increase the variations of a plurality of backward-translated sentences generated based on the forward-translated sentences and presented to the user. Therefore, for example, when applying machine learning to a machine translation system, if the user selects one backward-translated sentence from a plurality of backward-translated sentences having similar contents, even if the backward-translated sentence that is not selected is input by the user. In the case where the intention indicated by the translated target sentence is expressed, it is possible to prevent the machine translation system from learning that it is an incorrect reverse-translated sentence because it was not selected.

  In the above aspect, for example, the plurality of different forward-translated sentences may be forward-translated sentences each determined to include the same subject or predicate.

  Thereby, for example, from a group of forward-translated sentences mechanically generated based on the translation-target sentence, it is possible to select only a forward-translated sentence having the same subject or predicate as the translation-target sentence, and to improve the final translation accuracy. Can be improved.

  In the above aspect, for example, the machine translation system is a set of the backward-translated sentences generated at least one or more for each of the plurality of different forward-translated sentences, and the backward-translation including the plurality of backward-translated sentences. A sentence group is generated, and for each of the back-translated sentences included in the back-translated group, an evaluation value that evaluates the similarity with the translation target sentence is calculated, and the plurality of back-translated sentences are included in the evaluation value. Based on this, it may be selected from the group of reverse-translated sentences.

  According to the above aspect, a plurality of different backward-translated sentences are selected from the backward-translated sentence group based on the similarity to the translation target sentence. Can be selected, and the final translation accuracy can be improved. In addition, for example, a plurality of backward-translated sentences including a backward-translated sentence having a low degree of similarity to the translation target sentence may be selected, thereby increasing the variations of the backward-translated sentences presented to the user. Can be. Therefore, for example, when applying machine learning to a machine translation system, if the user selects one backward-translated sentence from a plurality of backward-translated sentences having similar contents, even if the backward-translated sentence that is not selected is input by the user. In the case where the intention indicated by the translated target sentence is expressed, it is possible to prevent the machine translation system from learning that it is an incorrect reverse-translated sentence because it was not selected.

  In the above aspect, for example, the machine translation system is a set of the backward-translated sentences generated at least one or more for each of the plurality of different forward-translated sentences, and the backward-translation including the plurality of backward-translated sentences. A sentence group is generated, and for each of the back-translated sentences included in the back-translated sentence group, it is determined whether the sentence is classified into a question sentence, a positive sentence, a negative sentence, or a command sentence. May be selected from the group of back-translated sentences based on the classified form.

  According to the above aspect, since a plurality of different backward-translated sentences are selected from the backward-translated sentence group based on the form of the sentence, for example, the reverse-translated sentence group having the same form as the form of the translation target sentence is selected from the backward-translated sentence group. Only the translation can be selected, and the final translation accuracy can be improved. In addition, for example, a plurality of backward-translated sentences including a backward-translated sentence having a form different from the form of the translation target sentence may be selected, thereby increasing the variation of the backward-translated sentences presented to the user. Can be. Therefore, for example, when applying machine learning to a machine translation system, if the user selects one backward-translated sentence from a plurality of backward-translated sentences having similar contents, even if the backward-translated sentence that is not selected is input by the user. In the case where the intention indicated by the translated target sentence is expressed, it is possible to prevent the machine translation system from learning that it is an incorrect reverse-translated sentence because it was not selected.

  In the above aspect, for example, the plurality of back-translated sentences may include at least two or more of the back-translated sentences classified into the different forms.

  As a result, it is possible to increase the variations of a plurality of reverse-translated sentences presented to the user. Therefore, for example, when applying machine learning to a machine translation system, if the user selects one backward-translated sentence from a plurality of backward-translated sentences having similar contents, even if the backward-translated sentence that is not selected is input by the user. In the case where the intention indicated by the translated target sentence is expressed, it is possible to prevent the machine translation system from learning that it is an incorrect reverse-translated sentence because it was not selected.

  In the above aspect, for example, the machine translation system is a set of the backward-translated sentences generated at least one or more for each of the plurality of different forward-translated sentences, and the backward-translation including the plurality of backward-translated sentences. A sentence group is generated, and the subject or predicate of each of the back-translated sentences included in the back-translated sentence group is determined, and the plurality of back-translated sentences are based on the determined subject or predescriptor, It may be selected from a group of reverse-translated sentences.

  According to the above aspect, since a plurality of different backward-translated sentences are selected from the backward-translated sentence group based on the subject or the predicate, for example, a reverse-translated sentence having the same subject or predicate as the translation target sentence is selected from the backward-translated sentence group. Only the translation can be selected, and the final translation accuracy can be improved. Further, for example, a plurality of back-translated sentences including a back-translated sentence having a subject or a predicate different from the translation target sentence may be selected, thereby increasing the variation of the plurality of back-translated sentences presented to the user. Can be. Therefore, for example, when applying machine learning to a machine translation system, if the user selects one backward-translated sentence from a plurality of backward-translated sentences having similar contents, even if the backward-translated sentence that is not selected is input by the user. In the case where the intention indicated by the translated target sentence is expressed, it is possible to prevent the machine translation system from learning that it is an incorrect reverse-translated sentence because it was not selected.

  In the above aspect, for example, the plurality of back-translated sentences may include at least two or more back-translated sentences determined to include different subjects or predicates.

  As a result, it is possible to increase the variations of a plurality of reverse-translated sentences presented to the user. Therefore, for example, when applying machine learning to a machine translation system, if the user selects one backward-translated sentence from a plurality of backward-translated sentences having similar contents, even if the backward-translated sentence that is not selected is input by the user. In the case where the intention indicated by the translated target sentence is expressed, it is possible to prevent the machine translation system from learning that it is an incorrect reverse-translated sentence because it was not selected.

  In the above aspect, for example, the plurality of back-translated sentences may be back-translated sentences each determined to include the same subject or predicate.

  As a result, for example, only a reverse translation sentence having the same subject or predicate as the translation target sentence can be selected from the reverse translation sentence group, and the final translation accuracy can be improved.

  In the above aspect, for example, the machine translation system manages a probabilistic model referred to in the translation processing, adapts machine learning in the translation processing, and selects any one of the plurality of back-translated sentences from the plurality of back-translated sentences. May be updated based on the information indicating whether or not is selected as the one back-translated sentence to update the parameters of the probability model.

  According to the above aspect, the machine learning is performed based on information indicating which of the plurality of back-translated sentences has been selected as the one back-translated sentence, and the parameter of the probability model is To update. As a result, information indicating which one of the plurality of back-translated sentences presented for the sentence to be translated is selected is reflected in the system. It is possible to improve accuracy.

  In the above aspect, for example, the probability model includes a weight value assigned to each word or phrase used in the translation processing, and the machine translation system is a forward translation corresponding to the one reverse translation. Comparing the words or phrases included in the selected-order translated sentence with the words or phrases included in the unselected translated sentence that is a forward-translated sentence corresponding to the reverse-translated sentence other than the one reverse-translated sentence; For words or phrases included only in the translated sentence, words or phrases included only in the unselected ordered translated sentence, and words or phrases included in both the selected ordered translated sentence and the unselected ordered translated sentence Updating the weights by applying different methods of updating the weights, and updating the weights and the words or phrases corresponding to the updated weights with teacher data. May perform the machine learning is used as data.

  Thus, for example, machine learning can be performed with a difference in score between words or phrases included in the selected-order translated sentence and words or phrases not included in the selected-order translated sentence. Even in the case of a word or a phrase included in a sentence, a positive evaluation may be performed in updating the weight value. For this reason, when a partially correct translation is performed in the non-selection order translated sentence, that part can be correctly evaluated, and the evaluation result of the user can be reflected.

  Further, by machine learning, learning can be performed on the probability model while sequentially reflecting the user's selection result in units of words or phrases, and translation accuracy can be improved.

  In the above aspect, for example, the probability model includes a weight value assigned to each word or phrase used in the translation processing, and the machine translation system includes a word or phrase included in the one reverse-translated sentence, A word or a phrase included in a non-selected reverse-translated sentence that is a reverse-translated sentence other than the one reverse-translated sentence is compared, and a word or a phrase included only in the one reverse-translated sentence is compared with the non-selected reverse-translated sentence For each of the words or phrases included only in the sentence and the words or phrases included in both the one back-translated sentence and the unselected back-translated sentence, a different updating method of the weight value is applied, and the weight is applied. A value may be updated, and the machine learning may be performed using the updated weight value and the word or the phrase corresponding to the updated weight value as teacher data.

  Thus, for example, performing machine learning with a difference in score between a word or phrase included in one selected back-translated sentence and a word or phrase not included in the selected one back-translated sentence Therefore, even if the word or phrase is included in the unselected reverse-translated sentence, a positive evaluation may be performed in updating the weight value. Therefore, when a correct translation is partially performed in the non-selected reverse-translated sentence, that portion can be correctly evaluated, and the evaluation result of the user can be reflected.

  Further, by machine learning, learning can be performed on the probability model while sequentially reflecting the user's selection result in units of words or phrases, and translation accuracy can be improved.

  In the above aspect, for example, the machine translation system may include a word or a phrase included in the one back-translated sentence and a word or phrase included in a non-selected reverse-translated sentence other than the one back-translated sentence. And a word or phrase included only in the one back-translated sentence, a word or phrase included only in the unselected back-translated sentence, and the one back-translated sentence and the non-selected back-translated sentence. For each word or phrase included in both, the weight value is updated by applying a different update method of the weight value, and the updated weight value and the word corresponding to the updated weight value are updated. Alternatively, the machine learning may be performed using the phrase as teacher data.

  In the above aspect, for example, the weight value is a value as a positive example for the word corresponding to only the one back-translated sentence, and for the word corresponding to only a sentence other than the one back-translated sentence, It may be a negative value.

  Thereby, both the positive evaluation and the negative evaluation can be reflected on the weight value.

  In the above aspect, for example, the machine learning may be learning using at least one of reinforcement learning, identification learning, and neural network learning.

  Further, in another aspect, there is provided a machine translation apparatus that performs a translation process between a first language and a second language, wherein the input unit receives an input of the translation target sentence of the first language, and A translation unit that generates a forward-translated sentence translated into the second language, and a reverse-translated sentence reversely translated from the forward-translated sentence into the first language; and a translation unit corresponding to the reverse-translated sentence and the one reverse-translated sentence. An output unit that outputs a forward-translated sentence; and a user input unit that receives a user's input, wherein the translating unit generates a plurality of different forward-translated sentences for the translation target sentence, and generates the plurality of different forward-ordered sentences. The output unit generates a plurality of backward-translated sentences corresponding to each of the translated sentences, and the output unit outputs one of the backward-translated sentences from the plurality of backward-translated sentences in the user input unit when outputting the plurality of backward-translated sentences. If an input to select a back translation is received, To output the forward translation sentence corresponding to the reverse translation of.

  Further, in a second other aspect, there is provided a program for controlling an operation of a machine translation device connected to an information output device and performing a translation process between a first language and a second language, wherein Causing the computer to receive the translation target sentence in the first language, generate a plurality of different forward translated sentences obtained by translating the received translation target sentence into the second language, A plurality of back-translated sentences, each of which is back-translated into the first language, are generated, and when the plurality of back-translated sentences are displayed on the information output device, one back-translation from the plurality of back-translated sentences is generated. When an operation for selecting a sentence is received, a forward-translated sentence obtained by translating the one backward-translated sentence into the second language is output.

(Embodiment)
Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  In the following embodiments, the source language as the language before translation is described as Japanese, and the target language as the language after translation is described as English. The word pairs can be any combination of language pairs.

  The translation obtained from the translation from the source language to the target language is referred to as a forward translation, and the translation obtained from the translation from the target language to the source language is referred to as a reverse translation.

  In addition, a backward-translated sentence presented to the user is referred to as a user-provided sentence, a sentence selected by the user is referred to as a (user) selected sentence, and a reverse-translated sentence not selected is referred to as a (user) non-selected sentence.

  FIG. 1 is a diagram illustrating an example of the overall configuration of a system according to the present embodiment. It includes an information display terminal 100, a network 200, a translation server 300, a microphone 400, and a speaker 500. Examples of the information display terminal 100 include a smartphone, a tablet terminal, a dedicated display device terminal, and a personal computer (PC). Any terminal other than those listed here may be used as long as it can exchange information with the user.

  In addition, the user's input operation on the information display terminal 100 is assumed to be text input, voice input, or the like. In text input, for example, input using a touch panel or input using a keyboard can be considered. In the case of input by voice, for example, input by a microphone can be considered. In addition to this, for example, a gesture input may be used.

  When the information display terminal 100 outputs a machine translation result or the like, the result may be output via a display or the result may be output using voice.

  The information display terminal 100, the translation server 300, the microphone 400, and the speaker 500 are connected to the network 200. An example of the connection method includes a wired or wireless LAN connection, but the connection method is not limited to this as long as each component is communicably connected.

  The translation server 300 performs a machine translation process on the translation target sentence received from the information display terminal 100. For example, a character string in the source language input from the information display terminal 100 is received, and a machine translation process is performed. It also has a function of receiving machine feedback from the user on machine translation results and performing machine learning. The detailed configuration of the translation server 300 will be described later.

  Note that, for example, the information display terminal 100 and the translation server 300 may be realized integrally.

  Microphone 400 performs voice input to the machine translation system. The microphone 400 may be attached to the information display terminal 100 or may have a function of connecting to the network 200 alone. Further, when no voice input is performed to the machine translation system, the configuration of the microphone 400 is not essential.

  The speaker 500 performs audio output in the machine translation system. The speaker 500 may be attached to the information display terminal 100 or may have a function of connecting to the network 200 alone. In the case where no voice output is performed in the machine translation system, the configuration of the speaker 500 is not essential.

  The input / output modality of the machine translation system may include only one of input and output by voice and input and output by text, or may include both. When a voice input is performed from the user to the machine translation system, a voice output is performed. When a text is input by a user, the text (screen display) is output.

  FIG. 2 is a block diagram showing a configuration of information display terminal 100 in the present embodiment.

  The information display terminal 100 includes a communication unit 101, an input unit 102, an output unit 103, a control unit 104, a selected sentence detection unit 105, and a storage unit 106.

  The communication unit 101 performs communication with the translation server 300, and performs transmission of a translation target sentence input in the information display terminal 100, reception of a later-described translated sentence and a reverse-translated sentence, and the like. In addition, not only these information but also various kinds of information are transmitted and received with the translation server 300.

  The input unit 102 receives an input from a user. The input unit 102 receives inputs such as an input of a translation target sentence and a selection input of a reverse translation sentence described later. As a form of the input, a voice input or an input in a text format can be considered. When speech input is used, a speech recognition process is performed on a translation target sentence input by speech, and a character string output as a result of the speech recognition process is input to the machine translation system as an input sentence. When text input is used, input of a character string by a keyboard, a mouse, a touch panel, or the like is accepted.

  The output unit 103 outputs the translation target sentence input at the input unit 102, a plurality of reverse-translated sentences received via the communication unit 101, a translation result, and the like. The output unit 103 may be realized as a display unit that performs screen display, such as a display. For example, a touch panel display or monitor used for a smartphone, a tablet terminal, or the like is assumed. Further, it may be realized as an audio output unit that outputs audio, such as a speaker. The control unit 104 controls operations of the communication unit 101, the input unit 102, the output unit 103, the control unit 104, the selected sentence detection unit 105, and the storage unit 106.

  The selected sentence detection unit 105 detects which one of the plurality of back-translated sentences output by the output unit 103 has been selected by the user. For example, when an input to select one backward-translated sentence from a plurality of backward-translated sentences is performed in the input unit 102, user selection information indicating which backward-translated sentence is selected is output to the selected sentence detecting unit 105. Is detected. The detected user selection information is transmitted to the translation server 300 via the communication unit 101. Further, the output content of the output unit 103 may be controlled based on the user selection information. For example, when the output unit 103 is implemented by a display, control may be performed to highlight the reverse-translated sentence selected by the user or to erase the reverse-translated sentence not selected by the user from the display screen.

  Here, the backward-translated sentence selected by the user and the forward-translated sentence corresponding to the backward-translated sentence are defined as a user-selected sentence. In addition, the backward-translated sentence not selected by the user and the forward-translated sentence corresponding to the backward-translated sentence are defined as user non-selected sentences.

  The storage unit 106 performs temporary storage of information received from the translation server 300, storage of various application programs executed in the information display terminal 100, and the like.

  FIG. 3 is a block diagram showing a configuration of translation server 300 in the present embodiment. The translation server 300 includes a communication unit 210, a control unit 220, a machine translation unit 230, and a storage unit 240. Further, the machine translation unit 230 includes a forward translation unit 231, a forward translation sentence selection unit 232, a reverse translation unit 233, a reverse translation sentence selection unit 234, a selection sentence determination unit 235, a phrase division unit 236, a selection result evaluation unit 237, and learning. It has a part 238.

  The communication unit 210 performs communication with the information display terminal 100, and receives a translation target sentence input in the information display terminal 100, transmits a later-described translated sentence and a reverse translated sentence, and the like. In addition to the above information, transmission and reception of various information with the information display terminal 100 are performed.

  The control unit 220 controls various operations of the communication unit 210, the machine translation unit 230, and the storage unit 240.

  The storage unit 240 stores a phrase table that the machine translation unit 230 refers to in various types of translation processing, phrase division processing, and the like. The phrase table will be described later.

  The machine translation unit 230 executes a machine translation process on the translation target sentence received via the communication unit. The machine translation unit 230 performs machine translation by a rule-based machine translation (RBMT), a statistical machine translation (SMT), a model acquisition type machine translation using a deep neural network (DNNMT), or the like. The machine translation unit 230 evaluates the translation result and acquires a score such as an automatic evaluation score (such as BLEU) and an internal score (such as evaluation by hand).

  In addition, in order to reflect the result of the selection by the user to machine learning, a phrase table in which pairs of phrases are shown in advance as shown in FIG.

  Since rule-based machine translation (RBMT) performs translation based on conversion rules (stored pairs of translated words as a database) constructed manually, there is a possibility that the phrase table as shown in FIG. 11 is not held. is there. However, if there is a bilingual database for each phrase or word, the learning result may be reflected in the bilingual database, or a phrase table may be separately prepared.

  In the statistical machine translation (SMT), since a phrase table as shown in FIG. 11 is held in advance, it may be used.

  In machine acquisition type machine translation (DNNMT) using a deep neural network, a phrase table is often not held because a model itself is automatically constructed. Therefore, a separate phrase table may be prepared.

  Further, the object that reflects the user's selection in the learning result is not limited to this, and for example, a database that represents a pair of paraphrases between source languages may be provided. The details of the machine translation process will be described later.

  The forward translation unit 231 executes a machine translation process from the language (source language) of the translation target sentence received via the communication unit 210 to a language (target language) output as a result of translating the translation target sentence. Here, the translation from the source language to the target language is referred to as a forward translation, and the translation obtained by the forward translation is referred to as a forward translation. At this time, a plurality of forward translation sentences are generated for the translation target sentence by the forward translation process. The forward translation unit 231 performs a machine translation process with reference to the phrase table stored in the storage unit 240. The plurality of forward-translated sentences generated by the forward-translating unit 231 are referred to as a forward-translated sentence group. The forward translation unit 231 outputs the generated forward translation sentence group to the forward translation sentence selection unit 232.

  The forward-translated sentence selection unit 232 performs a forward-translated sentence selection process of selecting N forward-translated sentences from the forward-translated sentence group generated by the forward-translation unit 231. Details of this forward translation sentence selection processing will be described later. The forward-translated sentence selection unit 232 outputs the selected N forward-translated sentences to the reverse-translation unit 233.

  For each of the N forward-translated sentences selected by the forward-translated sentence selecting unit 232, the backward translating unit 233 performs machine translation from the language of the forward-translated sentence (target language) to the language of the sentence to be translated (source language). Execute the process. Here, a translation from the target language to the source language is referred to as a reverse translation, and a translation obtained by the reverse translation is referred to as a reverse translation. At this time, one or more backward-translated sentences are generated for each forward-translated sentence by the backward-translation processing. Therefore, as a result, a plurality of back-translated sentences are generated. The reverse translation unit 233 performs a machine translation process with reference to the phrase table stored in the storage unit 240. The plurality of backward-translated sentences generated by the backward-translating unit 233 are regarded as a backward-translated sentence group. The back translation unit 233 outputs the generated back translation sentence group to the back translation sentence selection unit 234.

  The backward-translated sentence selecting unit 234 performs a backward-translated sentence selection process of selecting M backward-translated sentences from the backward-translated sentence group generated by the backward-translating unit 233. This reverse translation sentence selection processing will be described later. The backward-translated sentence selection unit 234 transmits the selected M backward-translated sentences to the information display terminal 100 via the communication unit 210. On the output unit 103 of the information display terminal 100, the M reverse-translated sentences are selectively output.

  Based on the user selection information received from the information display terminal 100 via the communication unit 210, the selected sentence determination unit 235 determines which of the M reverse-translated sentences selected by the backward-translated It is determined whether a translation has been selected, and the determined information is output to the phrase division unit 236.

  The phrase division unit 236 divides the back-translated sentence into phrases or words for each of the plurality of back-translated sentences based on the plurality of back-translated sentences input from the back-translated sentence selection unit 234. Further, the forward translation corresponding to the backward translation is also divided into phrases or words. At this time, information indicating which back-translation sentence was selected from the plurality of back-translation sentences input from the selection sentence determination unit 235 may be used together. Further, a phrase table stored in the storage unit 240 may be used. Information obtained by dividing the backward-translated sentence and the forward-translated sentence into units of phrases or words, and user selection information are output to the selection result evaluation unit 237.

  For phrase division, a phrase table indicating correspondence between phrases or words between bilingual languages as shown in Statistical Machine Translation (SMT) is often used, but a fixed phrase table is always used. It is not necessary and may be similar. When a phrase table is used in machine translation, division may be performed using the phrase table. Alternatively, a separately prepared phrase table or the like may be used, or if there is a bilingual dictionary or the like, it may be used.

  The selection result evaluation unit 237 evaluates the forward-translated sentence and the backward-translated sentence based on the information input from the phrase dividing unit 236. At this time, different evaluations may be performed for the user-selected sentence and the user-unselected sentence based on the user selection information. A detailed evaluation method will be described later. The selection result evaluation unit 237 outputs evaluation information obtained by evaluating the forward-translated sentence and the backward-translated sentence to the learning unit 238.

  The learning unit 238 performs machine learning in the machine translation process by updating the phrase table stored in the storage unit 240 based on the evaluation information input from the selection result evaluation unit 237. That is, the evaluation information is reflected in the phrase table. The target of machine learning may be a phrase table referred to by the forward translator 231 or a phrase table referred to by the reverse translator 233. It is not always necessary to reflect the evaluation information in the phrase table. For example, the result may be reflected in a paraphrase dictionary or a dictionary of words, and machine learning in the machine translation process may be performed. A detailed method of reflecting the evaluation information on the phrase table will be described later.

  FIG. 4 is a diagram illustrating a hardware configuration of a computer that realizes functions of each unit of the information display terminal by a program. The computer 1000 includes an input device 1001 such as an input button and a touch pad, an output device 1002 such as a display and a speaker, a CPU (Central Processing Unit) 1003, a ROM (Read Only Memory) 1004, a RAM (Random Access Memory) 1005, and the like. Prepare. The computer 1000 reads information from a recording device such as a hard disk device, a storage device 1006 such as an SSD (Solid State Drive), a DVD-ROM (Digital Versatile Disk Read Only Memory), or a USB (Universal Serial Bus) memory. 1007, a transmission / reception device 1008 for performing communication via a network may be provided. The above-described units are connected by a bus 1009.

  Then, the reading device 1007 reads the program from a recording medium on which the program for realizing the function of each unit is recorded, and stores the program in the storage device 1006. Alternatively, the transmission / reception device 1008 communicates with a server device connected to the network, and causes the storage device 1006 to store a program downloaded from the server device for realizing the function of each unit.

  Then, the CPU 1003 copies the program stored in the storage device 1006 to the RAM 1005, and sequentially reads out and executes the instructions included in the program from the RAM 1005, thereby realizing the functions of the respective units. When the program is executed, information obtained by the various processes described in the embodiments is stored in the RAM 1005 or the storage device 1006, and is appropriately used.

  FIG. 5 is a flowchart illustrating the operation of the machine translation system according to the present embodiment. For simplicity, in this flowchart, the input unit 102 and the output unit 103 of the information display terminal 100 are realized by a touch panel display. In the information display terminal 100, it is assumed that the input of the user and the output of the translation result and the like are performed via the touch panel display. For example, the input unit 102 and the output unit 103 such as a keyboard and a display are independent. It may be. Further, input / output by voice may be performed. For the sake of explanation, there are some places where the explanation is made with the source language (native language) being Japanese and the target language being English. However, these are merely examples, and the source language and the target language may be in any combination.

  First, in step S401, a translation target sentence input by the user is acquired. In step S402, a machine translation process is performed on the translation target sentence. The machine translation process here is a forward translation process for translating a translation target sentence in the source language into a character string (forward translation sentence) in the target language. At this time, for example, statistical machine translation (SMT) scores the likelihood of translation from a translation model and a language model. The translation model is a statistical model that defines the likelihood of the translated word, and the language model is a statistical model that defines the likelihood of the arrangement of words in the output language. From these two models, a plurality of forward-translated sentences are generated by scoring the certainty as translation and outputting the translation results in the order of the scores. A group of translated sentences.

  The specific processing of the machine translation according to the present embodiment is a general machine translation processing, and a description thereof will be omitted.

  In step S403, N forward-translated sentences are selected from the forward-translated sentence group based on a predetermined criterion. For example, an evaluation score may be given to each of a plurality of forward-translated sentences included in the group of forward-translated sentences, and N items may be selected from those having the highest evaluation scores. Alternatively, N items may be selected at random regardless of the evaluation score. And so on. In addition, in consideration of the contents indicated by a plurality of forward-translated sentences included in the forward-translated sentence group, the selected N forward-translated sentences may be selected so as not to have the same meaning. In addition, when a plurality of forward-translated sentences included in the forward-translated sentence group include only a small number of forward-translated sentences having different purposes, a process of adding forward-translated sentences having different purposes may be performed as necessary. A detailed method of selecting a forward-translated sentence will be described later with reference to FIG.

  If it is determined that it is necessary to increase the number of forward-translated sentences after the processing of step S403 is performed (YES in S404), the process returns to step S402 and performs the forward-translation processing again. At this time, in order to obtain a forward-translated sentence that is different from the already-acquired forward-translated sentence, one having a lower score than the previous one is selected. Alternatively, another translation index may be used (for example, rank-based intuitive bilingual evaluation score). Further, if a database of paraphrases of the source language is held, a similar sentence can be created by applying them to the input sentence, and the forward translation process can be executed again. As a result, a sentence having a different surface but having the same meaning as the input sentence is input, so that a different forward-translated sentence can be obtained.

  When the forward-translated sentence selection processing is executed in step S403 after the forward-translated processing is executed again, the forward-translated sentence may be selected based on a different criterion from the previous one, or the forward-translated sentence may be selected based on the same criteria. You may choose.

  If it is determined that it is not necessary to increase the number of forward-translated sentences after the processing in step S403 is performed (No in S404), the flow proceeds to the reverse translation processing in step S405. In step S405, a reverse translation process for performing reverse translation on the N forward-translated sentences obtained in step S403 is executed. Assuming that the translation from the source language to the target language is a forward translation, the reverse translation is a reverse translation from the target language to the source language. For each of the N forward-translated sentences, a reverse translation process for generating an arbitrary backward-translated sentence is performed. The reverse translation process of generating an arbitrary reverse translated sentence is a reverse translation process of generating a one-to-one corresponding reverse translated sentence for each of the N forward translated sentences, and a reverse translation process in the N forward translated sentences. Means a reverse translation process in which there is a forward translation sentence not performed, a reverse translation process of generating a plurality of reverse translations for one forward translation, and the like. By this back translation process, a plurality of back translation sentences are generated. The plurality of back-translated sentences are referred to as a back-translated sentence group for convenience of explanation.

  In addition, the criteria for outputting the reverse-translated sentence may be determined based on some system criteria, or may be determined by the user. The standard of the system here means that the score of a forward-translated sentence is calculated using an evaluation such as BLEU or manual evaluation, and for example, a forward-translated sentence having a low score is not subjected to a reverse translation process (a certain forward-translated sentence). For example, for a forward-translated sentence having a high score, an arbitrary number of backward-translated sentences are obtained (for a given forward-translated sentence, a plurality of backward-translated sentences are generated). When the user determines the number of backward-translated sentences, for example, the number of backward-translated sentences to be generated for one forward-translated sentence may be set. However, the present invention is not limited to this.

  Step S406 is a backward-translated sentence selection process of selecting M backward-translated sentences from the backward-translated sentence group obtained in step S405. In the reverse translation sentence selection processing, processing that is substantially the same as the forward translation sentence selection processing in step S403 is performed. A detailed selection method will be described later with reference to FIGS.

  If it is determined that it is necessary to increase the number of backward-translated sentences after the backward-translated sentence selection process of step S406 is performed (YES in S407), the process returns to step S402 and performs the forward translation process again. When the forward-translated sentence selection processing is executed in step S403 after the forward-translated processing is executed again, the forward-translated sentence may be selected based on a different criterion from the previous one, or the forward-translated sentence may be selected based on the same criteria. You may choose.

  If it is determined that there is no need to increase the number of back-translated sentences after the back-translation process of step S406 (No in S407), the M backward-translated sentences selected from the back-translated sentence group are sent to the user presented sentence. Is output from the information display terminal 100 in the next step.

  In step S408, M reverse-translated sentences are transmitted to the information display terminal 100 and displayed on the touch panel display.

  In step S409, the selected sentence detection unit 105 detects whether one of the M back-translated sentences displayed on the touch panel display of the information display terminal 100 has been selected.

  When it is determined that there is no selection of a reverse translation sentence for a certain period, the machine translation system returns to the initial state and accepts the input sentence of the user (No in S409). At this time, the display screen of the touch panel display is reset.

  Also, when the user performs some kind of reset operation, the machine translation system similarly returns to the initial state and accepts the user's input.

  When the selected sentence detection unit 105 detects that any of the backward-translated sentences has been selected (Yes in S409), user selection information indicating which of the backward-translated sentences has been selected is transmitted to the translation server 300. You.

  In step S411, the selected sentence determination unit 235 obtains M reverse-translated sentences from the backward-translated sentence selecting unit 234, and based on the user selection information received from the information display terminal 100, the information display terminal 100 It is determined which of the back-translated sentences is selected. The following steps will be described with the selected backward-translated sentence among the M backward-translated sentences as the selected translated sentence and the unselected backward-translated sentence as the unselected backward-translated sentence. The selected sentence determination unit 235 outputs the selected backward-translated sentence and the unselected backward-translated sentence to the phrase division unit 236.

  In step S412, a phrase division process is performed on the selected backward-translated sentence and the unselected backward-translated sentence, and the forward-translated sentence corresponding to each of the selected backward-translated sentence and the unselected backward-translated sentence.

  Phrase division is to divide a target sentence into phrases or words that are shorter units. A specific example of the phrase division processing will be described later.

  FIG. 11 is an example of a general phrase table in the present embodiment. The phrase table is a table that represents a correspondence between phrases and words between the source language and the target language.

  In FIG. 11, when the source language is Japanese and the target language is English, from the left, the Japanese phrase, the English phrase, and the English-to-Japanese translation probability of the phrase (the probability that the English phrase is translated into a Japanese phrase) ), The product of the translation probabilities of words in the English-Japanese direction (the product of the translation probabilities of each word in the phrase when English is translated into Japanese), the Japanese-English translation probability of the phrase (the Japanese phrase is (Probability of being translated into a phrase), and the product of the translation probabilities of words in the Japanese and English directions (the product of the translation probabilities of each word in the phrase when Japanese is translated into English). However, the phrase table does not necessarily need to include all of this information, and is not limited to this notation. Since this phrase table includes translation probabilities, it is also called a probability model in a broad sense.

  For example, in the phrase table shown in FIG. 11, the probability that the phrase PH1 is translated into the phrase PH1 is 0.38, the probability that the word SD3 is translated into the word SD1, and the probability that the word SD4 is translated into the word SD4 are: The product is 0.04, the probability that the phrase PH1 is translated into the phrase PH2 is 0.05, and the product of the probability that the word SD1 is translated into the word SD3 and the probability that the word SD2 is translated into the word SD4 is 0.02. It represents that.

  Using such a phrase table, a phrase division process is performed on the selected backward-translated sentence, the unselected backward-translated sentence, and the forward-translated sentence corresponding to each of these backward-translated sentences.

  FIG. 12 is an explanatory diagram illustrating an outline of phrase division.

  FIG. 12 shows reverse-translated sentences RS10, RS20, and RS30 expressed in the source language, and forward-translated sentences TS10, TS20, and TS30 in the target language corresponding to each of these backward-translated sentences. When the phrase division processing is performed on the back-translation sentence of the source language, for example, when the back-translation sentence RS10 is divided into phrases, the back-translation sentence RS10 is divided into three phrases of phrases PH11, PH12, and PH12. When a phrase division process is performed on a forward-translated sentence in the target language, for example, when the forward-translated sentence TS10 is phrase-divided, the forward-translated sentence TS10 is divided into three phrases (words) of phrases PH14, PH15, and PH16.

  The phrase into which the target sentence to be subjected to the phrase division processing is divided into phrases depends on the character strings of the source language and the target language described in the phrase table, and thus may not be uniquely determined.

  Step S413 is a phrase evaluation process of evaluating a score for each phrase output by the phrase division process according to a predetermined standard. Details of the phrase evaluation processing will be described later.

  In step S414, the forward-translated sentence corresponding to the selected reverse-translated sentence determined in S411 is transmitted to the information display terminal 100, and is displayed on the touch panel display as a translation result. At this time, the selected reverse translation sentence displayed on the touch panel display may be highlighted. In addition, the display of the unselected reverse-translated sentence may be deleted on the touch-panel display, and if it is possible to clearly indicate that the displayed translation result is a forward-translated sentence corresponding to the selected reverse-translated sentence selected by the user, Display may be performed.

  Since a series of processes in step S414 and steps S412 to S414 can be operated in parallel, step S414 may be performed before the phrase dividing process in step S412 and after the learning process in step S415. It may be executed at the timing of.

  In step S415, machine learning such as reinforcement learning, identification learning, and neural network learning is performed based on the score for each phrase obtained in step S413. Details of this processing will be described with reference to FIGS.

  FIG. 6 is a flowchart showing a specific operation of the translation sentence selection process in the present embodiment.

  Referring to FIG. 6, specific processes of the forward translation sentence selection process executed by forward translation sentence selection unit 232 in step S403 and the reverse translation sentence selection process executed by reverse translation sentence selection unit 234 in step S406 will be described. explain. For convenience of explanation, the two processes of the forward translation sentence selection process and the backward translation sentence selection process are collectively referred to as a translation sentence selection process.

  The forward-translated sentence selection unit 232 performs a forward-translated sentence selection process of selecting N forward-translated sentences from the forward-translated sentence group generated by the forward-translation unit 231, and the backward-translated sentence selection unit 234 performs reverse-translation sentence selection. A backward-translated sentence selection process of selecting M backward-translated sentences from the backward-translated sentence group generated by the unit 233 is executed. N forward-translated sentences are selected based on the evaluation scores of the plurality of forward-translated sentences included in the forward-translated sentence group, and based on the evaluation scores of the plurality of backward-translated sentences included in the backward-translated sentence group. Thus, M backward-translated sentences are selected.

  The following description applies to both forward-translation sentence selection processing and reverse-translation sentence selection processing, so it is described as translation-sentence selection processing, and forward-translated sentences and reverse-translated sentences are collectively referred to as translated sentences. I do. The forward-translated sentences and the backward-translated sentences are referred to as translated sentences. Further, in practice, N forward-translated sentences are selected and M backward-translated sentences are selected, but in the following description, they are described as N without distinction.

  In step S501, Nk translation sentences (1 <= N, 0 <= k <= N) having a high evaluation score are selected from the translation sentence group. Here, as an example of the evaluation score, BLEU, which is often used as a technique for evaluating translation accuracy, can be mentioned. Other methods for evaluating translation accuracy include WER (Word Error Rate), METEOR (Metric for Evaluation of Translation with Explicit Ordering), RIBES (Rank-based Intuitive), and RIBES (Rank-based Integrative Evaluation). Any of these may be used, and other methods may be used without being limited thereto.

  The remaining k sentences are selected from among the translated sentences not selected in step S501 in the translated sentence group (S502).

  In order to select k translations, a translation having an evaluation score within a predetermined threshold may be extracted, and k translations may be randomly selected from the extracted translations. Alternatively, translated sentences having an evaluation score within a predetermined threshold value may be extracted, and among the extracted translated sentences, k sentences may be selected in ascending order of score. When selecting N−k translations, a translation with a high evaluation score is selected. However, when selecting k translations, a translation with a high evaluation score is not necessarily selected. When only translations with a high evaluation score, which are automatically given according to a specific evaluation criterion, are selected, it is highly likely that all the selected translations are sentences having similar contents. In consideration of presenting a plurality of back-translated sentences to the user on the information display terminal 100 and allowing the user to select one back-translated sentence therefrom, it is possible to present a plurality of back-translated sentences selected from a somewhat different viewpoint. preferable.

  If only similar reverse-translated sentences are presented and the user selects them, there is a concern that a high learning effect cannot be obtained in a machine learning process for a machine translation system described later. As a correct example (correct answer) to the machine translation system, when a machine learning process is performed using the selected backward-translated sentence selected by the user and the unselected backward-translated sentence not selected by the user as teacher data. Because the selected reverse-translated sentence learned and the non-selected reverse-translated sentence learned as a negative example (incorrect answer) are similar sentences, there is a remarkable difference between the positive and negative examples for the machine translation system. I can't do it, and I can't expect the effects of learning. Therefore, as described in the present embodiment, it is preferable to select a plurality of reverse-translated sentences from a somewhat different viewpoint.

  Also, if only similar back-translated sentences are presented, the user needs to select one back-translated sentence in consideration of the small differences between the back-translated sentences, and thus intuitively selects one back-translated sentence. There is also a concern that the selection of the back translation may take time. By presenting a plurality of reverse-translated sentences selected from different viewpoints as in the present embodiment, the user can intuitively select the translation content intended by the user from among them.

  In addition, since the backward-translated sentence is generated from the forward-translated sentence, it may be preferable to select a plurality of forward-translated sentences with a somewhat different idea in the stage of selecting the forward-translated sentence as well.

  Further, as a different method for selecting k translations, k translations may be selected based on past user selection information. For example, for each translated sentence, the number of times selected by the user in the past is stored (for a backward-translated sentence, the number of direct selections is stored; for a forward-translated sentence, the number of times the corresponding reverse-translated sentence is selected is stored). Then, k translations may be selected in order from the one with the highest number of times stored in the translation sentence group. Also, instead of selecting based on such a direct number of times, based on the user's past use history of the machine translation system, the tendency of the translated sentence that the user can easily select for the translation target sentence is analyzed. Alternatively, k translation sentences may be selected based on the analyzed tendency. Note that if a selection criterion different from the selection criterion for selecting Nk translations is used, the selection method for selecting k translations is not limited to these.

  Further, in the method of selecting the k translated sentences, a process of excluding the translated sentences having the same meaning as the Nk translated sentences may be performed. In addition, a process for excluding a translated sentence having the same meaning from among the k translated sentences may be performed. Alternatively, if the k translations do not include a translation having a purpose different from the N−k translations, or if the number of translations is small, the process adds a translation having a different purpose. May be performed. In addition, regarding the reverse translation sentence selection processing, a process such as selecting a reverse translation sentence having the same meaning as that of the translation target sentence may be performed.

  For example, the translation sentence group counts the number of translation sentences in the respective forms of question sentence, affirmative sentence, negative sentence, and command sentence (the number of different sentences). Perform forward translation or reverse translation.

  There is also a method of performing a syntax analysis on a translated sentence. The parsing is performed on each of the plurality of translations included in the translation sentence group to determine what the word indicating the subject is, and the number of different subjects indicating the number of types of the subject is less than or equal to the threshold. If so, the forward translation process or the reverse translation process is performed again. At this time, a different number may be calculated not for the subject but for the predicate. Further, both of them may be used.

  Here, the number of different sentences, the number of different subjects, the number of different verbs or objects, and the like are used as evaluation scores, and the forward translation sentence selection processing and / or the reverse translation sentence selection are performed so that the predetermined number of differences is included. Processing may be performed.

  When executing the forward-translated sentence selection processing, for example, an evaluation criterion of “the number of different sentences is set to 2 or more” is provided, and the forward-translated sentence can be selected using the number of different sentences as an evaluation score. For each of the plurality of forward-translated sentences included in the forward-translated sentence group, analysis is sequentially performed using syntactic analysis and / or semantic analysis. It is also possible to score whether or not the type has appeared as a different number of sentences and create l small sets including two or more types.

  In addition, for example, when executing the reverse translation sentence selection processing, an evaluation criterion of “the number of different subjects may be set to a predetermined number or less” may be provided, and a reverse translation sentence may be selected using the number of different subjects as an evaluation score.

  The forward-translated sentence selection process or the backward-translated sentence selection process is not limited to these examples, and an arbitrary forward-translated sentence selection or reverse-translated sentence selection method may be used.

  Further, in the reverse translation sentence selection process, a comparison evaluation with the translation target sentence may be performed, and a value obtained by the comparison evaluation may be used as the evaluation score. For example, a syntax analysis is performed on the translation target sentence and a plurality of back translation sentences in the group of back translation sentences, and the similarity between the translation target sentence and each of the plurality of back translation sentences is determined. May be used to select k reverse-translated sentences.

  The present invention is not limited to these examples, and any evaluation score may be used in combination. Further, N may be set in advance as a predetermined number, and scores may be combined and used so that the number of differences included therein becomes a desired number.

  The above-described method of selecting a translation is merely an example, and the present invention is not limited thereto.

  When k = 0, all N words are selected in order from the translation with the highest evaluation score. When k = N, all of the N translations are selected by a selection method other than the translations having the highest evaluation score.

  The standard of the system here means that the score of a forward-translated sentence is calculated using an evaluation such as BLEU or manual evaluation, and for example, a forward-translated sentence having a low score is not subjected to a reverse translation process (a certain forward-translated sentence). For example, for a forward-translated sentence having a high score, an arbitrary number of backward-translated sentences are obtained (for a given forward-translated sentence, a plurality of backward-translated sentences are generated). When the user determines the number of backward-translated sentences, for example, the number of backward-translated sentences to be generated for one forward-translated sentence may be set. However, the present invention is not limited to this.

  FIG. 7 is a flowchart showing a specific operation of the reverse translation sentence selection process in the present embodiment.

  In the description of the reverse translation process (step S405) with reference to FIG. And good things. Here, a process of selecting one backward-translated sentence from a plurality of backward-translated sentences obtained from one forward-translated sentence will be described.

  The following process is performed on all N forward-translated sentences selected by the forward-translated sentence selection process.

  In step S601, the backward-translated sentence generated for the forward-translated sentence A among the N forward-translated sentences is extracted.

  In step S602, the number of extracted backward-translated sentences is determined. When no reverse-translated sentence has been generated for the forward-translated sentence A, that is, when no reverse-translated sentence has been generated for the forward-translated sentence A (0 in step S602), Since the translation A is not presented to the user, it is deleted (step S603).

  Next, when there is one backward-translated sentence for the forward-translated sentence A (one in step S602), the backward-translated sentence is determined as the backward-translated sentence corresponding to the forward-translated sentence A (step S604).

  Finally, when two or more backward-translated sentences have been generated for the forward-translated sentence A (two or more at step S602), one of the optimal backward-translated sentences corresponding to the forward-translated sentence A is selected. Are determined (step S605). The determination method is selected, for example, by using a method of referring to an automatic evaluation score or a manual evaluation score. These are repeated for all N forward-translated sentences.

  Finally, among the reverse-translated sentences obtained by these processes, a process of removing a reverse-translated sentence of the same meaning, or adding a translated sentence having a different meaning when there are few reverse-translated sentences of different purposes, etc. May be performed as needed (step S606). The processing here is the same as the processing described in the description of FIG.

  In the above description, one backward-translated sentence is selected for one forward-translated sentence in step S605 of FIG. However, a plurality of backward-translated sentences may be selected for one forward-translated sentence. In this case, among the plurality of backward-translated sentences presented to the user, there exists a backward-translated sentence whose corresponding forward-translated sentence is the same.

  FIG. 8 is a flowchart showing a specific operation of the phrase evaluation processing in the present embodiment.

  In the initial state of the flow, it is assumed that a selected reverse translation sentence, a non-selection reverse translation, and a forward translation corresponding to each of the selection reverse translation and the non-selection reverse translation are divided into phrases in advance. I do. It is also assumed that a phrase table corresponding to a phrase obtained by dividing the backward-translated sentence and the forward-translated sentence into phrases is similarly acquired. This phrase evaluation process is performed for each of the backward-translated sentence and the forward-translated sentence. For convenience of explanation, a phrase division process for the backward-translated sentence will be described as an example.

  In step S701, the phrases included in these selected reverse-translated sentences and non-selected reverse-translated sentences are compared for each reverse-translated sentence, and the presence or absence of a phrase existing only in the selected reverse-translated sentence is confirmed. I do.

  If there is a phrase that exists only in the selected reverse-translated sentence (Yes in step S701), the user-selected score is added to the phrase that exists only in the selected reverse-translated sentence (step S702). This user selection score is a score that makes a phrase appearing only in the selected sentence a good one. Finally, in the processing of FIG. 9, the forward translation sentence (if the target of the phrase evaluation processing is a forward translation sentence, the translation target sentence Is reflected in the Japanese-English translation probability or the English-Japanese translation probability of the phrase table together with the corresponding phrase of ()). At this time, any method may be used as a method for adding scores. For example, a method of uniformly adding points to the corresponding phrase, adding points depending on the length of the corresponding phrase, and the like are assumed. When the score addition is completed, the flow proceeds to step S703. If there is no phrase that exists only in the selected reverse translation (No in step S701), the flow proceeds to step S703 without performing any special processing.

  Similarly, in step S703, the phrases included in the selected reverse-translated sentences and the non-selected reverse-translated sentences are compared for each reverse-translated sentence, and the phrases existing only in the non-selected reverse-translated sentences are compared. Check if there is any.

  If there is a phrase that exists only in the unselected reverse-translated sentence (Yes in step S703), the user-selected score is deducted for a phrase that exists only in the unselected reverse-translated sentence (step S704). Any method may be used as the score deduction method at this time. For example, a method of uniformly deducting points for the corresponding phrase, a method of deducting points depending on the length of the corresponding phrase, and the like are assumed.

  In steps S702 and S704, adding or subtracting points from the user-selected score is not necessarily essential. That is, in the flowchart of FIG. 8, it is assumed that no score addition or deduction is performed for a phrase existing only in the selected reverse-translated sentence, and a user-selected score is deducted for a phrase existing only in the non-selected reverse-translated sentence. Is also good. Alternatively, the user-selected score may be added to the phrase that exists only in the selected reverse-translated sentence, and the user-selected score may not be added or deducted to the phrase that exists only in the non-selected reverse-translated sentence.

  In addition, scores may be added to phrases that are present in the selected reverse-translated sentence and some non-selected reverse-translated sentences. In this case, for example, a value considering the user selection score of the phrase included only in the selected reverse translation and the user selection score of the phrase included only in the non-selected reverse translation (for example, a phrase included only in the selected reverse translation) It is conceivable to add the user selection score of the phrase and the average value of the user selection scores of the phrases included only in the unselected reverse-translated sentences. The method of adding these scores is an example, and the present invention is not limited thereto.

  Hereinafter, a specific example will be described with reference to FIG. For example, a case where three back-translated sentences are presented to the information display terminal 100 and the user selects one of the back-translated sentences will be described. At this time, it is assumed that the three reverse-translated sentences presented to the information display terminal 100 are reverse-translated sentences RS10, RS20, and RS30, and the reverse-translated sentence selected by the user is the backward-translated sentence RS10. For convenience of description, a reverse-translated sentence selected by the user is defined as a selected reverse-translated sentence, and a reverse-translated sentence not selected by the user is defined as a non-selected reverse-translated sentence. In the initial state of the flow, the backward-translated sentence RS10, which is the selected backward-translated sentence, is divided into phrases PH11, PH12, and PH13. Similarly, for the unselected reverse-translated sentence, the reverse-translated sentence RS20 is divided into phrases PH21, PH22, and PH23, and the reverse-translated sentence RS30 is divided into phrases PH31, PH32, and PH33.

  In step S701, it is confirmed whether or not there is a phrase that exists only in the backward-translated sentence RS10 that is the selected backward-translated sentence among the backward-translated sentences RS10, RS20, and RS30. Then, since the phrase PH12 is included only in the reverse-translated sentence RS10, which is the selected reverse-translated sentence, the user selection score of the phrase PH12 is added to “+1”.

  Similarly, in S702, it is confirmed whether or not there is a phrase that exists only in the backward-translated sentences RS20 and RS30 that are the unselected backward-translated sentences. Then, since the phrase PH22 (PH32) and the phrase PH31 are included only in the unselected reverse-translated sentences, the user-selected scores of the phrases PH22 (PH32) and PH31 are deducted to "-1".

  Here, for the phrases PH11 (PH21) and PH13 (PH23, PH33) included in both the selected reverse-translated sentence and the non-selected reverse-translated sentence, no points are added or subtracted from the user-selected score.

  By the above-described processing, the final amount of addition or deduction of the user-selected score for each phrase is as follows. "± 0" for phrase PH11 (PH21), "-1" for phrase PH31, "-1" for phrase PH22 (PH32), "+1" for phrase PH12, phrase For PH13 (PH23, PH33), an additional point or a deducted amount such as “± 0” is obtained. Since the amount of points added and the amount of points deducted here are just examples, points may be added or deducted in an order larger than this, or points may be added or deducted in an order smaller than this.

  The phrase evaluation process for the forward-translated sentence that is the target language will be described below using a specific example with reference to FIG. The forward-translated sentences RS10, TS20, and TS30 correspond to the backward-translated sentences RS10, RS20, and RS30, respectively. In the initial state of the flow, the forward-translated sentence TS10 is divided into phrases PH14, PH15, and PH16. The forward-translated sentence TS20 is divided into phrases PH24, PH25 and PH26, and the TS30 is divided into phrases PH34, PH35 and PH36.

  In S701, it is confirmed whether or not there is a phrase that exists only in the forward-translated sentence 1 corresponding to the selected reverse-translated sentence 1. Then, since the phrase PH16 is included only in the forward-translated sentence TS10, the user-selected score of the phrase PH16 is added to "+1".

  Similarly, in S702, it is confirmed whether there is a phrase that exists only in one of the forward-translated sentences TS20 and TS30 corresponding to the backward-translated sentences RS20 and RS30 that are the unselected backward-translated sentences. Then, since the phrases PH26 (PH36) and the phrase PH34 are included only in the forward-translated sentences TS20 or TS30, the user-selected scores of the phrases PH26 (PH36) and PH34 are respectively deducted to "-1".

  Phrase PH24 and phrase PH15 (PH25, PH35) included in forward-translated sentence TS10 corresponding to the selected reverse-translated sentence and included in either forward-translated sentence TS20 or forward-translated sentence TS30 corresponding to the unselected reverse-translated sentence. Regarding, no point is added or subtracted from the user-selected score.

  By the above-described processing, the final score increment or deduction for each phrase is “± 0” for the phrase PH24 (PH34), “−1” for the phrase PH34, and the phrase PH26 (PH36). For the phrase PH16, "+1" for the phrase PH16, and "± 0" for the phrase PH15 (PH25, PH35). Since the point addition amount and the point deduction amount of the user selection score here are examples, the point addition or the deduction may be performed in an order larger than this, or may be performed in the order smaller than this.

  FIG. 9 is a flowchart showing a specific operation of the learning process in the present embodiment.

  In step S801, a phrase pair of a phrase included in the forward translation corresponding to the selected reverse translation and a phrase included in the selected reverse translation, or a phrase included in the translation target sentence corresponding to the selected reverse translation, A phrase pair with a phrase included in the forward translation corresponding to the selected reverse translation is acquired. Phrase pairs are two phrases that have correspondence (having the same meaning) between the source language and the target language during machine translation. Further, the user selection score obtained in the processing of FIG. 8 is also obtained at the same time.

  In this phrase pair, if the user selection score for the value of the phrase table referred to when translated from the source language to the target language is defined using the example shown in FIG. 12, for example, the phrase PH31 → phrase PH34: −1 / Phrase PH22 (PH32) → Phrase PH26 (PH35): -1 / Phrase PH11 (PH21) → Phrase PH14 (PH24): 0 / Phrase PH13 (PH23, PH33) → Phrase PH15 (PH25, PH35): 0 / Phrase PH12 → Phrase PH16: It becomes like +1.

  Also, in the phrase pair, defining a user selection score for the value of the phrase table referred to when the target language is translated from the target language, for example, the phrase PH34 → phrase PH31: −1 / phrase PH26 (PH36) → phrase PH22 (PH32): -1 / Phrase PH14 (PH24) → Phrase PH11 (PH21): 0 / Phrase PH15 (PH25, PH35) → Phrase PH13 (PH23, PH33): 0 / Phrase PH16 → Phrase PH12: +1 Become.

  In step S802, the above-mentioned user-selected score is reflected in the English-Japanese translation probability or the Japanese-English translation probability of the phrase table stored in the storage unit 240. Also, the user selection score may be given a slope or weight, for example, by multiplying the user selection score by a certain value and then reflecting the result in the phrase table.

  Using these, the machine translation unit 230 and the reverse translation unit 233 perform machine learning such as reinforcement learning, identification learning, and neural network learning.

  In conventional machine translation, the probability values in a phrase table as shown in FIG. 11 are tuned based on a bilingual corpus (data obtained by collecting pairs of sentences that are translated into each other between two different languages). However, there has never been a method of performing machine learning with a difference between scores included in and not included in a user-selected sentence, and the machine translation system according to the present disclosure is more user-friendly. Evaluation results can be reflected.

  Furthermore, by machine learning, a translation model or a language model can be trained on a translation model or a language model created from a bilingual corpus prepared in advance while sequentially mixing user selection results in phrase units. Therefore, accuracy can be improved.

  Furthermore, by performing machine learning, optimal parameters are selected based on the data. As a result, the selection result of a person (user) is reflected on the translation system, so that a translation system that is easy for a person to use can be constructed.

  In addition to performing such machine learning, a new corpus can be generated from the obtained phrases and used as a bilingual corpus for translation engine learning.

  FIG. 10 is a flowchart illustrating a specific process of the learning unit 238 according to the present embodiment. The contents of the flowchart shown in FIG. 10 will be described using the backward-translated sentence and the forward-translated sentence shown in FIG.

  In step S901, a phrase pair of a user-selected sentence and its forward translation is acquired.

  For example, it is assumed that the user selects the backward-translated sentence RS10 while the backward-translated sentences RS10, RS20, and RS30 are displayed in the user-provided sentence display area 1102 (when the backward-translated sentence RS10 is a user-selected sentence). explain. Since the reverse translation sentence RS10 is selected by the user, phrase pairs such as the phrase PH11 and the phrase PH14, the phrase PH12 and the phrase PH16, and the phrase PH13 and the phrase PH15 are acquired.

  In step S902, a phrase pair in the input sentence and its forward-translated sentence used in the machine translation by the machine translation unit 230 is acquired.

  For example, when the content indicated by the reverse translation sentence RS30 is an input sentence, a phrase pair such as the phrases PH31 and PH34, the phrases PH32 and PH36, and the phrases PH33 and PH35 is acquired.

  In step S903, a phrase having the same target language character string as the phrase acquired in the input sentence and the user-selected sentence is acquired. For example, assume that the phrase pairs in the user selection sentence are as follows. The phrase pairs in the user-selected sentence are the phrases PH11 and PH14, the phrases PH12 and PH16, and the phrases PH13 and PH15. On the other hand, it is assumed that the phrase pair of the input sentence is as follows. The phrase pairs of the input sentence are the phrases PH31 and PH34, the phrases PH32 and PH36, and the phrases PH41 and PH42.

  At this time, in the phrase of the user-selected sentence and the phrase of the input sentence, the source language phrase PH33 and the phrase PH41 have the same meaning but have different expressions. Further, the phrases PH33 and PH41 are corresponding phrases between the phrase of the user selection sentence and the phrase of the input sentence.

  Finally, it is checked whether the target language is the same and the source language is different, and the different language is stored as a paraphrase (paraphrase) (Yes in S904, S905). That is, since the source language is different even if the target language is the same, these can be regarded as paraphrasing in the source language.

  For example, the phrases PH33 and PH41 can be regarded as paraphrases, and these are held as paraphrases between source languages.

  The paraphrase can be referred to when performing machine translation in the machine translation unit 230, or referred to as a paraphrase in the source language before performing the translation.

  FIG. 13 is a diagram illustrating an example of a display screen according to the present embodiment.

  For example, as shown in FIG. 13A, when an input of an original sentence to be translated is received from a user, the contents of the input original sentence OS1 are displayed in the input sentence display area 1101.

  Next, as shown in FIG. 13B, a reverse translation result for the translated sentence obtained by translating the original sentence OS1 is displayed in the user-presented sentence display area 1102.

  As an example, a mode in which three back-translated sentences are output will be described here. As the reverse translation result, for example, a reverse translation sentence RS1, a reverse translation sentence RS2, and a reverse translation sentence RS3 are output and displayed in the user presentation sentence display area 1102. At this time, the reverse-translated sentences RS1 to RS3 displayed in the user-provided sentence display area 1102 in FIG. 13B are similar sentences having the same meaning in the source language. Due to the characteristics of the reverse translation process, these are expected to be similar sentences having the same meaning, but may be implemented to output sentences having different meanings.

  Next, as shown in FIG. 13C, the user checks the reverse translation result displayed in the user-provided sentence display area 1102, and selects the reverse-translated sentence closest to the input content intended by the user. . Here, for example, the back-translated sentence RS1 is selected for the original sentence OS1.

  When the user selects a backward-translated sentence, a translated sentence corresponding to the selected backward-translated sentence is displayed in the translation result display area 1103. Here, a translated sentence TS1, which is a translated sentence corresponding to the reverse-translated sentence RS1, is displayed.

  Note that the screen display is not limited to the layout shown in FIGS. 13A, 13B, and 13C. Various buttons may be arranged as necessary. For example, when an operation on a button is performed after inputting an original text to be translated, a translation process may be performed. The reverse translation may be displayed in the user-provided sentence display area 1102 by performing an operation on the button. Also, the arrangement, displayed contents, and orientation of the input sentence display area 1101, the user presentation sentence display area 1102, and the translation result display area 1103 are not limited to the above-described contents.

  FIG. 14 is a diagram illustrating an example of a display screen according to the present embodiment.

  The display is partially different from FIG. Here, the direction of the sentence displayed in translation result display area 1201 is different from the direction of the sentence displayed in input sentence display area 1202 and user presented sentence display area 1203. This assumes a situation in which two users (a source language speaker and a target language speaker) communicate with each other across an information display terminal. That is, the sentences displayed in the input sentence display area 1202 and the user presented sentence display area 1203 are displayed in a direction suitable for the source language speaker, and the sentences displayed in the translation result display area 1201 are displayed in the target language speech. Is displayed in an orientation suitable for the person. This allows the source language speaker to read out the translated sentence for the input sentence to the target language speaker or change the orientation of the information display terminal so that the target language speaker can easily check the translated sentence. Since there is no need to perform such communication, smooth communication between users who speak different languages via an information display terminal or the like is possible. The direction of the translation result display area 1201 can be changed by an arbitrary operation by the user. Further, similarly to FIG. 13, the arrangement of each area, the displayed content, and the orientation are not limited to these.

  As described above, the translation method according to one embodiment of the present invention has been described based on the embodiments, but the present invention is not limited to these embodiments. Unless departing from the gist of the present invention, various modifications conceivable by those skilled in the art are applied to this embodiment, or forms configured by combining components in different embodiments are also included in the scope of the present invention. .

  For example, in the above description, one backward-translated sentence is selected by the user from a plurality of backward-translated sentences presented to the information display terminal 100, but a plurality of backward-translated sentences may be selected. . For example, one is selected as a user's selection and a forward-translated sentence is presented, but other non-user-selected sentences may be evaluated and the result may be reflected in the system as a learning result. As a method of evaluation here, for example, the user ranks the user non-selected sentences in the order of good or bad, selects a non-selected sentence that is as permissible as the user selected sentence, The evaluation is performed by a method in which the user selects an unacceptable non-selected sentence. By performing these operations, it is possible to evaluate sentences that have not been selected, and by reflecting these in the system, it is possible to learn the system.

  Further, in the above description, a user-provided sentence or a translation of a result is output as a text with respect to an input sentence as shown in FIG. 13, but these are presented with text and voice or only voice. It does not matter. In this case, the user may select one of the user-provided sentences from the user-provided sentences via the microphone.

  The machine translation method according to the present invention is useful in a machine translation system that connects to an information output device that outputs linguistic information and performs translation processing between a first language and a second language.

REFERENCE SIGNS LIST 100 Information display terminal 101 Communication unit 102 Input unit 103 Output unit 104 Control unit 105 Selected sentence detection unit 106 Storage unit 200 Network 210 Communication unit 220 Control unit 230 Machine translation unit 231 Forward translation unit 232 Forward translation selection unit 233 Reverse translation unit 234 Reverse translation sentence selection unit 235 Selected sentence judgment unit 236 Phrase division unit 237 Selection result evaluation unit 238 Learning unit 240 Storage unit 300 Translation server 400 Microphone 500 Speaker 1000 Computer 1001 Input device 1002 Output device 1003 CPU
1004 ROM
1005 RAM
1006 storage device 1007 reading device 1008 transmitting / receiving device 1009 bus

Claims (16)

A machine translation method in a machine translation system connected to an information output device that outputs language information and performing a translation process between a first language and a second language,
Receiving the translation target sentence in the first language,
Generating a plurality of different forward-translated sentences obtained by translating the received sentence to be translated into the second language,
For each of the plurality of different forward-translated sentences, generate a plurality of back-translated sentences back-translated into the first language,
If the operation of selecting one reverse-translated sentence from the plurality of backward-translated sentences is received while outputting the plurality of backward-translated sentences in the information output device, the information corresponding to the one backward-translated sentence is received. Output a forward translation,
Generating a forward-translated sentence group that is a set of forward-translated sentences obtained by translating the received sentence to be translated into the second language, and that includes the plurality of different forward-translated sentences;
For each of the forward-translated sentences included in the forward-translated sentence group, it is determined whether the sentence is classified into a question sentence, an affirmative sentence, a negative sentence, or a command sentence,
The plurality of different forward-translated sentences are determined by selecting a forward-translated sentence having the same form as the classified form from the forward-translated sentence group,
Machine translation method. The machine translation system is further connected to a voice input device that receives a voice input by the user, and a text input device that receives a text input by the user,
The translation target sentence is received in the form of audio information or text information representing the translation,
The form of output of the forward-translated sentence corresponding to the one backward-translated sentence is changed according to which form of the voice information or the text information has been received.
The machine translation method according to claim 1. The information output device has an audio output device and a display,
If the translation target sentence is received in the form of audio information, the forward translation corresponding to the one reverse translation is output via the audio output device,
If the translation target sentence is received in the form of text information, the forward translation corresponding to the one reverse translation is output via the display.
3. The machine translation method according to claim 2. The machine translation system is further connected to a text input device that receives text input by a user,
The translation target sentence is received from the text input device as text information indicating the translation target sentence,
Based on the text information, generating a plurality of different forward-translated sentences that translated the sentence to be translated into the second language,
The machine translation method according to claim 1. The machine translation system is further connected to a voice input device that receives a user's voice input,
The translation target sentence is received from the voice input device as audio information representing the translation target sentence,
Performing voice recognition processing on the received voice information to generate text information indicating the translation target sentence,
Based on the text information, generating a plurality of different forward-translated sentences that translated the sentence to be translated into the second language,
The machine translation method according to claim 1. The information output device has a display,
The plurality of reverse-translated sentences are displayed in a first area of the display,
The translation target sentence is displayed in a second area different from the first area of the display,
The machine translation method according to claim 1. The forward-translated sentence corresponding to the one backward-translated sentence is displayed in a third area of the display.
The machine translation method according to claim 6. The forward-translated sentence corresponding to the one backward-translated sentence is displayed in a different direction from the plurality of backward-translated sentences displayed in the first area.
The machine translation method according to claim 7.   The machine translation system,
  A set of the backward-translated sentences generated at least one or more for each of the plurality of different forward-translated sentences, and a backward-translated sentence group including the plurality of backward-translated sentences is generated,
  For each of the backward-translated sentences included in the backward-translated group, calculate an evaluation value that evaluates the degree of similarity with the translation target sentence,
  The plurality of back-translated sentences are selected from the group of back-translated sentences based on the evaluation value.
  The machine translation method according to claim 1. A machine translation method in a machine translation system connected to an information output device that outputs language information and performing a translation process between a first language and a second language,
Receiving the translation target sentence in the first language,
Generating a plurality of different forward-translated sentences obtained by translating the received sentence to be translated into the second language,
For each of the plurality of different forward-translated sentences, generate a plurality of back-translated sentences back-translated into the first language,
If the operation of selecting one reverse-translated sentence from the plurality of backward-translated sentences is received while outputting the plurality of backward-translated sentences in the information output device, the information corresponding to the one backward-translated sentence is received. Output a forward translation,
Generating a forward-translated sentence group that is a set of the forward-translated sentences obtained by translating the received sentence to be translated into the second language;
The forward-translated sentence group includes the plurality of different forward-translated sentences,
Determine the subject or predicate of each forward-translated sentence included in the forward-translated sentence group,
The plurality of different forward-translated sentences are determined by selecting a forward-translated sentence having the same subject or predicate as the determined subject or pre-described word from the forward-translated sentence group ,
The machine translation method. A machine translation method in a machine translation system connected to an information output device that outputs language information and performing a translation process between a first language and a second language,
Receiving the translation target sentence in the first language,
Generating a plurality of different forward-translated sentences obtained by translating the received sentence to be translated into the second language,
For each of the plurality of different forward-translated sentences, generate a plurality of back-translated sentences back-translated into the first language,
If the operation of selecting one reverse-translated sentence from the plurality of backward-translated sentences is received while outputting the plurality of backward-translated sentences in the information output device, the information corresponding to the one backward-translated sentence is received. Output a forward translation,
A set of the backward-translated sentences generated at least one or more for each of the plurality of different forward-translated sentences, and a backward-translated sentence group including the plurality of backward-translated sentences is generated,
For each of the back-translated sentences included in the back-translated sentence group, determine whether the sentence is classified into a question sentence, a positive sentence, a negative sentence, or a command sentence,
Wherein the plurality of reverse translation, from the reverse translation Bungun, is selected by selecting the inverse translation is the same form as classified above embodiment,
The machine translation method. A machine translation method in a machine translation system connected to an information output device that outputs language information and performing a translation process between a first language and a second language,
Receiving the translation target sentence in the first language,
Generating a plurality of different forward-translated sentences obtained by translating the received sentence to be translated into the second language,
For each of the plurality of different forward-translated sentences, generate a plurality of back-translated sentences back-translated into the first language,
If the operation of selecting one reverse-translated sentence from the plurality of backward-translated sentences is received while outputting the plurality of backward-translated sentences in the information output device, the information corresponding to the one backward-translated sentence is received. Output a forward translation,
A set of the backward-translated sentences generated at least one or more for each of the plurality of different forward-translated sentences, and a backward-translated sentence group including the plurality of backward-translated sentences is generated,
Determine the subject or predicate of each of the back-translated sentences included in the back-translated sentence group,
The plurality of backward-translated sentences are selected from the backward-translated sentence group by selecting a backward-translated sentence having the same subject or predicate as the determined subject or pre- determined word,
The machine translation method. The machine translation system manages a probability model referred to in the translation processing,
The probability model includes a weight value given to each word or phrase used in the translation process,
The machine translation system comprises:
A word or a phrase included in the selected-order translated sentence that is a forward-translated sentence corresponding to the one backward-translated sentence, and a non-selected translated sentence that is a forward-translated sentence corresponding to a reverse-translated sentence other than the one backward-translated sentence. Compares with the contained word or phrase,
A word or phrase included only in the selected order translation,
A word or phrase included only in the non-selection order translation,
For the words or phrases included in both the selected-order translated sentence and the non-selected-order translated sentence, the weight values are updated by applying different weight value updating methods, and the updated weight values are updated. Performing the machine learning using the word or the phrase corresponding to the updated weight value as teacher data;
The machine translation method according to claim 12 . The machine translation system manages a probability model referred to in the translation processing,
The probability model includes a weight value given to each word or phrase used in the translation process,
The machine translation system comprises:
Compare the word or phrase included in the one back-translated sentence and the word or phrase included in a non-selected reverse-translated sentence that is a back-translated sentence other than the one back-translated sentence,
A word or phrase that is included only in said one back translation,
A word or phrase included only in the unselected reverse translation,
For words or phrases included in both the one back translation and the unselected back translation, update the weight value by applying a different update method of the weight value,
Performing the machine learning using the updated weight value and the word or the phrase corresponding to the updated weight value as teacher data,
The machine translation method according to claim 12 . A machine translation device that performs a translation process between a first language and a second language,
An input unit configured to receive an input of a sentence to be translated in the first language;
A translation unit that generates a forward-translated sentence obtained by translating the translation target sentence into the second language, and a reverse-translated sentence obtained by reverse-translating the forward-translated sentence into the first language;
An output unit that outputs the backward-translated sentence and the forward-translated sentence corresponding to the one backward-translated sentence;
And a user input unit for receiving a user's input. The translation unit generates a plurality of different forward-translated sentences for the translation target sentence, and a plurality of reverse translations corresponding to each of the plurality of different forward-translated sentences. Generate a statement,
The output unit, when outputting the plurality of back-translated sentences, when the user input unit receives an input for selecting one back-translated sentence from the plurality of back-translated sentences, and it outputs the order translation sentence corresponding to the translation,
The translation unit,
Generating a forward-translated sentence group that is a set of forward-translated sentences obtained by translating the input target sentence into the second language, and that includes the plurality of different forward-translated sentences;
For each of the forward-translated sentences included in the forward-translated sentence group, it is determined whether the sentence is classified into a question sentence, an affirmative sentence, a negative sentence, or a command sentence,
The plurality of different forward-translated sentences are determined by selecting a forward-translated sentence having the same form as the classified form from the forward-translated sentence group,
Machine translation device. A program for controlling an operation of a machine translation device connected to an information output device and performing a translation process between a first language and a second language,
For the computer of the machine translation device,
Receiving the translation target sentence in the first language,
Generating a plurality of different forward-translated sentences by translating the received sentence to be translated into the second language;
Generating a plurality of back-translated sentences back-translated into the first language for each of the plurality of different forward-translated sentences;
In the information output device, when an operation of selecting one backward-translated sentence from the plurality of backward-translated sentences is displayed while displaying the plurality of backward-translated sentences, the one backward-translated sentence is displayed in the second backward-translated sentence. To output the translated text translated into two languages ,
Generating a forward-translated sentence group that is a set of the forward-translated sentences obtained by translating the received translation target sentence into the second language and includes the plurality of different forward-translated sentences;
For each of the forward-translated sentences included in the forward-translated sentence group, a question sentence, an affirmative sentence, a negative sentence, a command sentence is determined,
The plurality of different forward-translated sentences are determined by selecting a forward-translated sentence having the same form as the classified form from the forward-translated sentence group,
program.

Images