Translation device, control program for translation device, and translation evaluation method using translation device
The translation device encrypts and preprocesses text to ensure accurate and confidential translation by using double encryption and bilingual dictionaries, addressing neural machine translation's accuracy and security issues.
Patent Information
- Application Number
- JP2019145271
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2019-08-07
- Publication Date
- 2024-11-06
- Estimated Expiration
- 2039-08-07
AI Technical Summary
Neural machine translation systems face challenges in accurately translating information, particularly in documents requiring strict accuracy, such as laws and patents, due to issues like mistranslations, omissions, handling of text decorations and styles, and the need for large amounts of parallel sentences and computational resources. Additionally, there are concerns about confidentiality and security when using external translation services over the internet.
A translation device that encrypts sensitive text using a first encryption method, translates it through external resources, and decrypts it using a second encryption method, ensuring confidentiality and accuracy by replacing words with codes and adjusting sentence structure to prevent leaks and mistranslations.
Maintains confidentiality of translated documents by double encryption, reduces mistranslations and computational costs, and ensures accurate translation by preprocessing text to simplify sentence structure and using bilingual dictionaries.
Smart Images

Figure 0007578255000001 
Figure 0007578255000002 
Figure 0007578255000003
Abstract
Description
[Technical field]
[0001] The present invention relates to a translation device, a control program for the translation device, and a translation evaluation method using the translation device. [Background technology]
[0002] Research into the technologies that form the basis of machine translation (MT), such as morphological analysis and dependency analysis, began in the 1950s. In the 1980s, with the improvement of computer performance, rule-based machine translation technology became somewhat practical.
[0003] In the 1990s, the method of statistical machine translation (STM) was developed. This involves statistically processing the correspondence between words and sentence structures between different languages, and selecting words and translations with a high probability of being translated from translation candidates. In statistical machine translation, a translation model determines the words that correspond to the original words and the word order of the words according to probability. Furthermore, a language model determines the correct order of the translated words according to probability. In statistical machine translation, a large amount of bilingual data (training data) is required to build a translation engine (translation model and language model). The translation engine is tuned appropriately by comparing the translations made using the translation engine with the correct translations.
[0004] In statistical machine translation, in order to correctly determine the word order, a technique is known in which a sentence in the source language is subjected to syntactic analysis and the word order of the words in the source language is converted in advance into the word order of the target language using dependency relationships.
[0005] In the 2010s, neural machine translation (NMT), which uses deep learning with neural networks, appeared. It has been rapidly put into practical use since around 2015, and is replacing statistical machine translation. Neural machine translation generally does not perform pre-processing of the source text to be translated (such as changing the word order), but translates the text directly (end to end) using a neural network.
[0006] In neural machine translation, words are treated as real-valued vectors, and as distributed representations of several hundred dimensions. In other words, the meaning and syntactic information of words are expressed by vectors (word embedding). Flexible translation is possible because the semantic relationship between one source word and its translation can be inferred to some extent from the semantic relationship between other source words and their translations.
[0007] In machine translation, the length of the input data (length of the original text) is not constant, and previously input data must be used. For this reason, recurrent neural networks (RNNs) are generally used in neural machine translation. Words (or characters) of the original text are input sequentially into the recurrent neural network, and when a code indicating the end of the sentence (EOS: End of String) is output, the translation of the sentence is considered complete.
[0008] In neural machine translation, the weighting coefficients of each neuron (node) become translation knowledge, so there is no need for a translation table (a phrase table or other table that associates original words with translated words). Neural machine translation does not translate by replacing and rearranging the constituent words of the original text, as in statistical machine translation, but rather creates a new translation from the input text so that it is consistent with the learned language model. [Prior art documents] [Patent documents]
[0009] [Patent Document 1] JP 2018-120584 A Summary of the Invention [Problem to be solved by the invention]
[0010] Compared to rule-based and statistical machine translation, neural machine translation has the advantages of fluent translation and the ability to flexibly translate even untrained sentences (through word embedding, etc.).
[0011] However, because neural machine translation is not a technology that translates by replacing the source text, as in rule-based machine translation and statistical machine translation, it is difficult to translate the information contained in the input text exactly (there is a possibility of missing parts in the translation).In addition, there are cases where duplicated parts of the translation are output.
[0012] Furthermore, neural machine translation cannot use the technique used in statistical machine translation, in which the word order of the source text is converted into the word order of the target language in advance, which makes it difficult to use the training data that has been used up until now (dictionary data, word order data, etc.).
[0013] Furthermore, while neural machine translation allows for flexible translation through word embedding, it can also result in translations with completely different meanings being created by selecting words with completely different meanings (especially for words that occur infrequently, proper nouns that occur infrequently, etc.). In addition, for long or complex sentences, translations that do not make sense can sometimes be created.
[0014] Furthermore, unlike rule-based machine translation and statistical machine translation, neural machine translation generally has the problem that it is difficult for humans to understand the translation process, and it is difficult to predict whether an accurate translation will be output. This makes neural machine translation prone to mistranslations, and the problem is that mistranslations are likely to be provided as the final result.
[0015] For everyday conversations where a few mistranslations or omissions are acceptable as long as the gist can be understood, or for documents conveying less important information, the fluency and flexibility of the generated translation may be more important than the accuracy of the content. For such documents, neural machine translation is extremely useful.
[0016] On the other hand, for documents that require strict accuracy, such as laws and patents, mistranslations and missing translations can be fatal. For this reason, even when using neural machine translation, it is necessary to spend a considerable amount of time and effort checking for mistranslations and missing translations. Furthermore, insufficient checking can lead to fatal mistranslations and missing translations, which can cause financial damage.
[0017] Another issue with neural machine translation is that it can be difficult to handle characters with special styles, such as superscripts, subscripts, italics, bold, and underlines. ~ In some cases, superscripts are enclosed in tags to indicate that they are superscripts, and in other cases, superscripts are expressed directly in character codes such as Unicode. If these are mixed and used for machine learning, or if the character decorations and styles in the document at the time of translation are handled differently from those used during training, incorrect translation results may be output.
[0018] Superscripts and subscripts are used to write chemical symbols, chemical formulas, mathematical expressions, variables, etc., but in neural machine translation, due to the effect of word embedding, chemical symbols, chemical formulas, mathematical expressions, and variables that may appear similar at first glance but are completely different may be output as translation results. Because errors in chemical symbols, chemical formulas, mathematical expressions, and variables are difficult to spot at a glance, they often remain as mistranslations in the translation deliverables.
[0019] Furthermore, in neural machine translation, there is a problem that the same word in the original text is translated into multiple different words. For example, the word "address" in the original document may be translated as "address", "addressee", or "addressee" depending on the place where it appears, even though it has the same meaning. Also, for example, the word "display unit" in the original document may be translated as "display unit", "display portion", or "display means" depending on the place where it appears, even though it has the same meaning. Such fluctuations in spelling and output of words with different meanings are not desirable, especially in the translation of legal documents and patent documents, which require strictness.
[0020] Furthermore, neural machine translation requires the preparation of a huge amount of bilingual text for training, and because of the large amount of calculation required for training, it requires a lot of time and computer resources (such as GPUs).
[0021] To reduce the cost of neural machine translation, it is also possible to perform neural machine translation using trained external computers (Internet-connected sites, servers, etc. operated by other companies or by other people) that are not dedicated to the company. The company in 1 provides a trained neural machine translation system on the Internet, and multiple users (an unspecified number of users) use the system on the Internet. Note that "unspecified number of users" means that the site (server) is open to the public and can be accessed from anywhere.
[0022] When using such a system, users of the system send document data via the Internet to a trained neural machine translation computer operated by another company, and receive the results of the machine translation performed on that server via the Internet. Sending and receiving can be done using an API (Application Programming Interface) provided by the system provider, or can be done using general communication protocols such as HTTP and HTTPS.
[0023] Transmission and reception can also be done via a web browser, which is a user interface. In this case, the user pastes the sentence (text) to be translated into a form displayed on the web browser using copy and paste or other methods, and then sends it to an external neural machine translation computer. The translation results, which are the response from the neural machine translation computer, are displayed on the web browser. The data exchange format for this type of communication is JSON (JavaScript Object Notation), etc.
[0024] In a shared neural machine translation system like the one described above, system users do not need to prepare huge volumes of bilingual texts for training, nor do they need to spend a lot of time on training or maintenance. In addition, system users do not need to own computer resources (servers, GPUs, etc.) to build the system. In other words, users can conveniently use the neural machine translation service while reducing the time required for training the neural network, the hardware costs, and the maintenance costs and effort to almost zero.
[0025] On the other hand, when translating highly confidential documents (such as legal documents such as contracts, especially patent documents before publication, confidential internal documents, research and development documents, etc.), it is necessary to avoid sending these documents to external parties via the Internet.
[0026] Even if the communication path is encrypted using HTTPS or similar, the transmitted document will naturally be converted into plain text at the destination. Therefore, if you use a service provided by another company, you should ultimately avoid sending highly confidential documents over the Internet. Ensuring security by encrypting the communication path and the security risk of passing confidential information to outside parties (other companies, etc.) are two separate issues and need to be addressed separately and independently.
[0027] In this regard, it may be possible to conclude a confidentiality agreement with the recipient of the document (a company that provides neural machine translation services over the Internet). However, in an online society, information is easily leaked, and the leak is difficult to detect, making it unfeasible and dangerous to conceal information through a confidentiality agreement. Furthermore, there is no guarantee that such a confidentiality agreement will be fulfilled. Furthermore, it is difficult to regularly supervise and manage whether the confidentiality agreement is being fulfilled on an ongoing basis.
[0028] Today, computers and smartphones have become commonplace tools, and the Internet and web browsers have become necessities of life that anyone can easily use. However, due to a lack of understanding of the Internet system, communication protocols, and information transmission routes, there are a constant stream of cases where highly confidential documents are sent carelessly to online machine translation services. In addition, the risk of information leaks when using online machine translation services is often overlooked.
[0029] In translation and dictionary services via the web, users input the sentences and words they want to translate from a web browser or similar (this can be done by typing, voice input, copy and paste, etc.). The input sentences and words are sent to a server owned (or managed) by the provider of the translation or dictionary service, and stored together with information about the sender.
[0030] Providers of translation and dictionary services can establish usage policies (terms of use) for the accumulated information, but some do not. Some terms of use include a clause stating, "When a user uploads, provides, or transmits content, the user grants the service provider (and any third parties working with the service provider) a worldwide license to use, store, copy, modify, create derivative works from, transmit (to the public), publish, publicly display, and distribute that content."
[0031] There have been many cases where highly confidential documents have been leaked to the outside (especially overseas) when translators at translation companies, or freelance translators who have been commissioned by translation companies, carelessly enter highly confidential documents (or parts of them or words) into a computer or smartphone web browser. There is also the problem that such leaks can continue for years without anyone noticing - the translator himself, the translator's manager, the translation company, or the party ordering the translation.
[0032] In particular, leaking highly confidential documents such as contracts, research and development documents, and unreleased patents to the outside (especially overseas) must be absolutely avoided not only from the perspective of protecting the interests of an individual company, but also from the perspective of protecting the national interest.
[0033] For example, the following is stated in the press release issued by the Information-Technology Promotion Agency, Japan on February 20, 2015 entitled "Caution: Beware of unintended information leaks of information entered into cloud services" (quoted from https: / / www.ipa.go.jp / about / press / 20150220.html):
[0034] Following the revelation that text entered into an online translation service was published online as is, the IPA (Information-technology Promotion Agency, Japan, Chairman: Kazumasa Fujie) has decided to issue a new warning regarding the mindset required for cloud service use, in order to raise users' awareness of the issue. In recent years, various cloud services have been developed, and the use of services aimed at individuals, as well as businesses, is on the rise. For example, there are services that allow you to store photos, documents, and other data on the Internet and use them anytime, anywhere, and that allow easy translation. However, while cloud services have rapidly spread due to their convenience, there have been repeated reports of unintended information leaks caused by users using the services without properly understanding the content and risks of the services(*1).
[0035] In 2013, a problem occurred in which users of Google Groups, a service provided by Google, were not properly aware of the settings for the scope of information disclosure, resulting in communications being viewable by people other than those involved. In addition, in an IPA survey (*2), 31.3% of respondents answered that they were "not concerned at all" or "not very concerned" about "information entered into the browser and search history being collected by the company that provides the browser," indicating a need to improve users' awareness and knowledge of cloud services.
[0036] On the other hand, service providers are also required to provide easy-to-understand explanations of the content of the service and how the information will be handled by the service provider in order to prevent information leaks due to insufficient explanations or lack of awareness on the part of users.
[0037] In addition, the following items are stated in the "February 2014 Call" by the Information-Technology Promotion Agency, Japan (quoted from https: / / www.ipa.go.jp / security / txt / 2014 / 02outline.html):
[0038] "Are you leaking information to the outside without knowing it?" ~ Key points when using cloud services ~ (omitted)
[0039] While cloud services are convenient to use, it is unavoidable to hand over some information to the service provider.
[0040] The three cloud services in Table 1 below are services that we use casually in our daily work. In particular, "online translation services" allow you to easily translate text without installing translation software on your computer, but you can also say that you are handing over the original text to be translated to the company.
[0041] If you handle confidential information in your work, please be aware of this and use it with caution.
[0042] Table 1: How to use cloud services that may pose risks such as information leakage
[0043] Online Translation Service
[0044] Overview When you copy and paste the document you want to translate onto a web page, the translation program on the service provider's side will automatically translate it.
[0045] Risks of use
[0046] The copied and pasted contents of the original document will be sent to an external party, and if the contents are classified as confidential information, you will unknowingly be violating company regulations.
[0047] Examples of anticipated damage
[0048] An employee was negotiating a contract with an overseas company using encrypted emails. The decrypted emails were written in English, so they were translated into Japanese using an online translation service. This resulted in the contract information being leaked. Although the company had taken the trouble to use encrypted emails to prevent information leaks, the information was leaked via another route, the online translation service.
[0049] Although the above warnings have been issued, security awareness among users of external (cloud, etc.) translation systems remains low due to their convenience. [Means for solving the problem]
[0050] An object of the present invention is to provide a translation device that allows a user to easily correct a translated text.
[0051] In order to achieve the above-mentioned object, according to one aspect of the present invention, in a translation device that corrects a result of translating a sentence written in a first language into a sentence written in a second language, the sentence written in the first language includes a plurality of character strings, and the translation device includes a determination means that determines whether the sentence written in the second language includes a character string corresponding to a character string included in the first language, and a correction means that corrects the sentence written in the second language based on a determination result of the determination means.
[0052] Preferably, the text written in the second language is a text obtained by machine translating a text written in the first language.
[0053] Preferably, the character string is either a character string indicating a word, a character string indicating a word followed immediately by a reference symbol, or a cryptographic character string, or a character string containing either a character string indicating a word, a character string indicating a word followed immediately by a reference symbol, or a cryptographic character string.
[0054] Preferably, the translation device acquires a sentence written in the second language which is a machine translation result of a sentence written in the first language in which a character string indicating a certain word, or a character string including a character string indicating a certain word, is replaced with a corresponding other character string, and the determination means determines whether the replaced character string is included in the sentence written in the second language.
[0055] Preferably, the determining means makes a determination for each unit constituting the document, and the correcting means corrects each unit constituting the document.
[0056] Preferably, the determining means performs a determination on the entire document, and the correcting means performs a correction on the entire document.
[0057] Preferably, the determination result includes at least one of character strings not included in the sentence written in the second language and the number of character strings not included in the sentence written in the second language.
[0058] Preferably, the determination means determines whether a character string corresponding to a character string contained in a sentence written in the second language is contained in a sentence written in the first language.
[0059] According to another aspect of the present invention, in a control program for a translation device that corrects a result of translating a sentence written in a first language into a sentence written in a second language, the translation device includes a computer, and the sentence written in the first language includes a plurality of character strings, and the control program for the translation device causes the computer to execute a determination step of determining whether the sentence written in the second language includes a character string corresponding to a character string included in the first language, and a correction step of correcting the sentence written in the second language based on a determination result of the determination step.
[0060] According to yet another aspect of the present invention, in a translation evaluation method using a translation device that corrects a result of translating a sentence written in a first language into a sentence written in a second language, the translation device includes a computer, and the sentence written in the first language includes a plurality of character strings, and the translation evaluation method includes a determination step of determining whether the sentence written in the second language includes a character string corresponding to a character string included in the first language, and a correction step of correcting the sentence written in the second language based on a result of the determination step. Effect of the Invention
[0061] According to the present invention, at least one of the above problems can be solved. [Brief description of the drawings]
[0062] [Figure 1] 1 is a block diagram showing a configuration of a translation system according to a first embodiment of the present invention. [Diagram 2] 1 is a functional block diagram showing a configuration of a translation system according to a first embodiment of the present invention. [Diagram 3] FIG. 2 is a diagram showing a specific example of a data structure stored in a database 207. [Figure 4] 3 is a flowchart showing a Japanese-English translation process of a computer program included in the translation system in the first embodiment of the present invention. [Diagram 5] 5 is a flowchart continuing from FIG. 4. [Figure 6] 4 is a flowchart showing a word and cipher registration process of a computer program included in the translation system in the first embodiment of the present invention. [Figure 7] 13 is a flowchart showing a word and cipher registration process of a computer program included in a translation system in accordance with a second embodiment of the present invention. [Figure 8] 13 is a flowchart showing a Japanese-English translation process of a computer program included in a translation system in the third embodiment of the present invention. [Figure 9] 13 is a flowchart showing a Japanese-English translation process of a computer program included in a translation system in the fourth embodiment of the present invention. [Figure 10] FIG. 13 is a diagram showing a specific example of a data structure stored in a first database in database 207 in the translation system in the fifth embodiment of the present invention. [Figure 11] FIG. 13 is a diagram showing a specific example of a data structure stored in a second database in database 207 in the translation system in the fifth embodiment of the present invention. [Figure 12] 13 is a flowchart showing a Japanese-English translation process of a computer program included in a translation system in the fifth embodiment of the present invention. [Figure 13] 13 is a flowchart continuing from FIG. 12. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0063] [First embodiment]
[0064] FIG. 1 is a block diagram showing the configuration of a translation system according to a first embodiment of the present invention.
[0065] With reference to the figure, the translation system comprises a computer 100 operated by a user (system user), a LAN (Local Area Network) 300, the Internet 400, other company's computer resources 200, and the company's own computer resources 500.
[0066] The computer 100 operated by the user is composed of a CPU 101, a communication unit 103, an I / O 105, a ROM 107, a RAM 109, a storage device 111, an audio input / output unit 113, a display 115, a graphics unit 117, a keyboard 119, and a mouse 121.
[0067] A computer 100 is connected to a LAN 300, and the LAN 300 is connected to the Internet 400. The LAN 300 is connected to the company's own computer resources 500, and the Internet 400 is connected to other company's computer resources 200.
[0068] Here, computer resources refer to all or part of servers, personal computers, storage devices (storage within a computer, NAS (Network Attached Storage), etc.), and information and communication paths (networks, load balancers, switches, routers, etc.), as well as the CPUs, memory, storage devices (hard disks, optical or magnetic storage devices, SSDs and other semiconductor devices), ROM, motherboards, input devices such as keyboards, mice, and microphones, output devices such as displays and speakers, and software such as computer programs and data that are read from storage devices and generally run on memory.
[0069] The company's own computer resources 500 are computer resources owned by the user (or the organization to which the user belongs). The organization may be an organization consisting of one user, or an organization consisting of multiple personnel.
[0070] The company's own computer resources 500 and the computer 100 are not directly connected to the Internet 400, but are connected to the Internet 400 via a router (not shown) of the LAN 300. In addition, by providing a firewall function in the router, the company's own computer resources 500, and the computer 100, access to the company's own computer resources 500 and the computer 100 from the Internet 400 side is prohibited. This prevents the company's own computer resources 500 and the computer 100 from being operated from the outside via the Internet 400, and also prohibits external access to data stored in the company's own computer resources 500 and the computer 100.
[0071] The other company's computer resource 200 is a computer resource that is not owned by the user (or the organization to which the user belongs). The user can temporarily borrow the computing power and storage of the other company's computer resource 200 by accessing it from the computer 100. The form of use may be either paid or free of charge.
[0072] The original data to be translated, which is a highly confidential document, is written in HTML or plain text and is recorded in the company's computer resources 500 or computer 100. If the original data to be translated is an image, it is converted into an HTML document or a text document by OCR (Optical Character Recognition / Reader) software recorded in the company's computer resources 500 or computer 100, and is then processed.
[0073] The communication protocol between the company's own computer resource 500 and computer 100 and the other company's computer resource 200 is HTTPS (Hypertext Transfer Protocol Secure) or the like. That is, HTTP communication is performed on secure encryption (a data encryption method using a private key and a public key) provided by the SSL (Secure Sockets Layer) / TLS (Transport Layer Security) protocol, so the contents of communication between the company's own computer resource 500 and computer 100 and the other company's computer resource 200 can be kept confidential, and the contents of communication can be prevented from leaking to third parties. Encryption communication other than HTTPS may be used as long as it can prevent the contents of communication from leaking to third parties.
[0074] Naturally, communication contents encrypted by HTTPS or the like are decrypted by the other company's computer resource 200. For this reason, even if a communication protocol such as HTTPS is used, it is not desirable to transmit the plaintext of a document that should be kept secret within the user or the organization to which the user belongs to the other company's computer resource 200. In addition, the operator of the other company's computer resource 200 may have an agreement in the terms of use that reserves the right to disclose, use, etc., of the transmitted data. In such a case, it is technically possible to transmit the plaintext of a document that should be kept secret within the user or the organization to which the user belongs to the other company's computer resource 200, but is commercially impossible.
[0075] The user does not know in detail how the communication contents restored to plain text by the other company's computer resource 200 are actually stored and used by that other company. In addition, the user generally does not know the configuration and connections of the other company's computer resource 200, and therefore cannot know the specific handling of the received data within that other company.
[0076] In reality, the leakage of confidential information occurs frequently and is difficult to prevent. Therefore, even if a confidentiality agreement is concluded with the operator of the other company's computer resource 200, it is still undesirable to transmit the plaintext of a document that should be kept confidential to the other company's computer resource 200.
[0077] That is, even if encrypted communication such as HTTPS is used, the text to be translated, which is highly confidential, or any part of it, should not be transmitted to the external computer resource 200 of another company.
[0078] Therefore, in this embodiment, computer 100 encrypts part of the highly confidential text to be translated by replacing it with a string of characters that is difficult to decipher (hereinafter, this will be referred to as "first encryption"), and then, when sending the text to external other company's computer resource 200, further encrypts it using a protocol such as HTTPS (SSL encrypted communication) (hereinafter, this will be referred to as "second encryption").
[0079] In the other company's computer resource 200, the communication contents encrypted by the second encryption are restored to plain text by a protocol such as HTTPS (this corresponds to the "second encryption", and is hereinafter referred to as "second decryption"). However, the first encryption is not decrypted, and the document that remains encrypted by the first encryption is translated by a translation program in the other company's computer resource 200. The translation (text document or HTML document) created by this translation is encrypted by a protocol such as HTTPS (i.e., encrypted by the "second encryption" method) and sent to the computer 100 via the Internet 400. The computer 100 restores the received text to plain text by a protocol such as HTTPS, which is the "second decryption" method. In this state, the first encryption has not yet been decrypted.
[0080] The computer 100 performs "first decryption", which is a decryption process for the "first encryption", on the received translated data. It also performs subsequent processes as necessary. This allows the user of the computer 100 to obtain the complete translated text.
[0081] Even if communication data is leaked to a party other than the company or other companies shown in Figure 1, it has been doubly encrypted using the first encryption and the second encryption, which prevents information from being leaked to a third party other than the parties involved in the communication.
[0082] Also, the other companies shown in Figure 1 can obtain information that has been decrypted for the second encryption, but cannot decrypt the first encryption. This prevents information from leaking from the company shown in Figure 1 to the other companies shown in Figure 1. In other words, the other companies shown in Figure 1 (people who manage the computer resource 200) can only get the text that has been first encrypted. Furthermore, the translated text also remains first encrypted. This means that the other companies shown in Figure 1 cannot obtain either the complete unencrypted data before translation, or the complete unencrypted translated data.
[0083] The first encryption must be of a level that allows the other company's computer resource 200 to perform machine translation of the encrypted text, and must be strong enough that no one other than the company can understand the contents.
[0084] FIG. 2 is a functional block diagram showing a configuration of a translation system according to the first embodiment of the present invention.
[0085] 1 stores a plurality of computer-readable executable programs which are executed in sequence by CPU 101 to cause computer 100 to operate as a translation device. The programs are read from storage device 111 and expanded on RAM 109. The programs are executed on RAM 109 by CPU 101.
[0086] By executing the program, as shown in FIG. 2, the computer 100 functions as a control unit 201, a communication unit 103, a search and replace unit 203, a document editing and word registration unit 205, a database 207, a memory 209, a storage device 111, a display / output unit 211, and an input unit 213.
[0087] The control unit 201 is a functional block that performs various controls on the entire translation device. The communication unit 103 is a functional block for communicating with in-house and external computer resources. The search and replace unit 203 is a functional block that searches for specific elements (words, phrases, etc.) in at least one of the document to be translated and the machine-translated document, and replaces the searched elements with corresponding character strings or codes. The document editing and word registration unit 205 is a functional block that edits at least one of the document to be translated and the machine-translated document, and registers dictionaries for word conversion.
[0088] The database 207 is a database in which a word dictionary is mainly registered. The database 207 may be implemented using database server software, or may be implemented by recording a simple text file such as a CSV file or a table in a spreadsheet program (such as EXCEL), searching the file, and reading out the corresponding data.
[0089] The memory 209 is a working area for temporarily storing data. The storage device 111 is a non-volatile storage device, and stores data before translation, data after translation, data to be recorded in a database, data in the memory 209, etc. The display / output unit 211 is a user interface such as a display and a speaker. The input unit 213 is a user interface such as a mouse, a keyboard, a microphone, etc.
[0090] FIG. 3 is a diagram showing a specific example of a data structure stored in the database 207. As shown in FIG.
[0091] Here, the database 207 is made up of a table having numbers (record numbers, serial numbers of data), words, passwords, bilingual words, and the like as column names (column names, field names).
[0092] As the words, nouns are registered, but verbs, adverbs, adjectives, etc. may also be registered. In addition, nouns include compound nouns consisting of multiple nouns (such as "electric vehicle" which means electricity + automobile).
[0093] The number column records a serial number for each record. The word column registers the words in the original text to be translated. The code column registers the code corresponding to that word. The bilingual word column registers the translated word that corresponds to that word (and code).
[0094] For example, in Fig. 3, translation data used for translating between Japanese and English is registered. That is, Japanese words, their corresponding English words (translated words), and their corresponding codes are registered. By using the table in Fig. 3, it is possible to convert a Japanese word into its corresponding code, a Japanese word into its corresponding English word, a code into its corresponding Japanese word, a code into its corresponding English word, an English word into its corresponding code, or an English word into its corresponding Japanese word.
[0095] Furthermore, translation between three or more languages is possible by adding columns and registering bilingual words in other languages in the table in Figure 3. For example, by adding a column of bilingual Chinese words, translation between Japanese and English, English and Japanese, Japanese and Chinese, Chinese and Japanese, English and Chinese and English (as well as conversion between words in each language and codes) becomes possible.
[0096] For example, in the record numbered "0," the database records in association the original word "denki jidosha" (electric vehicle), the encrypted character string "AAA," and the bilingual word "electric vehicle."
[0097] Such a database may be created in advance by the user, or may be created during or outside of the translation process by repeating the process shown in the flowchart of Figure 6 or 7. Also, a company may create the database and provide it to the user.
[0098] By using such a database, it is possible to convert words into codes, convert words into their translations, and also convert codes into their translations.It is also possible to convert translations back into the original words (reverse translation).
[0099] For words and translation words, one basic word (such as "car," "means," "part," "vehicle," "device," or "unit") may be registered in one record, or a compound word (such as "electric vehicle," "signal input means," "display unit," "electric vehicle," "signal input unit," or "display unit") may be registered in one record.
[0100] The code consists of three letters of the alphabet and functions as a virtual word during machine translation. The code starts with AAA and expresses the record number (number) written in decimal as a 26-hexadecimal number, with the letters A to Z corresponding to the numbers 0 to 25.
[0101] In other words, the last digit of the code is the 0th power of 26, the second last digit of the code is the 1st power of 26, and the third last digit (the most significant digit) of the code is the 2nd power of 26.
[0102] For example, if the "number" column contains a 0, it will be encrypted with "AAA" (A is the base 26 code for zero), and if the "number" column contains a 1, it will be encrypted with "AAB" (A is the base 26 code for zero, and B is the base 26 code for one). Similarly, if the "number" column contains a 3501, it will be encrypted with "FER" (base 26 code for 3501).
[0103] That is, the most significant digit of the cipher (the last three digits) is assigned an alphabet that corresponds to the integer part of the value obtained by dividing the number by 26 to the power of 2 (i.e. 676). The middle digit of the cipher (the last two digits) is assigned an alphabet that corresponds to the integer part of the value obtained by dividing the remainder of the above division by 26 to the power of 1 (i.e. 26). The least significant digit of the cipher (the last one) is assigned an alphabet that corresponds to the remainder of that division. This is equivalent to converting from decimal to base 26.
[0104] The number of digits in the code is not limited to three. Also, among the three-digit codes, those that have meaning as words (e.g., CPU, RAM, USB, NOX, etc.) may not be registered together with the corresponding number as reserved words in order to prevent misunderstandings and mistranslations.
[0105] In this embodiment, the record number is converted into a hexadecimal number to create the code, but the code may be determined randomly or according to other rules.
[0106] The word and the translation word may be registered as the same thing. For example, a chemical formula or an abbreviation (HCl, CPU, etc.) may be registered as the same word and translation word. In this case, the same word is registered in both Japanese and English. Furthermore, the word (Japanese) may be full-width characters, and the translation word (English) may be half-width characters.
[0107] Also, a word may be registered in which a part of the original word is written in the translation language, such as "information display unit" and its corresponding translation is "information display unit."
[0108] The code may be expressed in 26-hexadecimal notation, following the usual hexadecimal notation (using 0-9 and A-F), using 0-9 and A-P. Lowercase alphabetic characters and symbols (!"#$%&, etc.) may also be used. The code may be a random character string that does not overlap among multiple code strings. It is desirable that the characters that make up the string consist of alphabetic characters, symbols, or a combination of alphabetic characters and symbols. Pictograms such as emoticons may also be used as the code.
[0109] As in this embodiment, if the code is AAA, AAB, AAC, . . ., ZZY, ZZZ, and each digit is a 26-hexadecimal number ranging from A to Z, and is expressed as a three-digit (three-letter) alphabet, then 26*26*26=17576 words can be registered in the database. If this is insufficient, the number of digits in the code can be increased to register records (pairs of words, code, and translated words) in the database in the number equal to the power of 26.
[0110] As in this embodiment, by expressing the code in 26-hexadecimal notation with A to Z being 0 to 25 respectively, there is an advantage that it is not clear at a glance what the word represents. Also, in neural machine translation, a character string of alphabets or symbols is recognized as a single word, so there is an advantage that the code can be machine translated while remaining as a single word in the original text.
[0111] In other words, when translating using neural machine translation, strings of letters or symbols are not translated and are output as is (however, there are exceptions).
[0112] For example, when the original translation text "electric vehicle 100 includes engine 101" is substituted with the table in FIG. 3, the encrypted text "AAA 100 includes AAB 101" is generated. This text is translated by neural machine translation into the text "AAA 100 includes AAB 101." After the neural machine translation is performed, the text in the table in FIG. 3 is converted into the translation of the original word to obtain the translated text "electric vehicle 100 includes engine 101."
[0113] It is possible for a code word to be converted to a plural form, for example, the original text "multiple AAA" could be translated to "a plural of AAAs". In this case, too, the translation "books" can be obtained by replacing the code word "AAA" in "AAAs" with the corresponding English word (for example, "book") by adding an "s" to the end.
[0114] However, this method results in misspellings of words such as "bus," "leaf," and "city" (words that do not become correctly pluralized by simply adding an "s"). For this reason, it is necessary to run a conventional spell-check routine after translation to ensure that these words are spelled correctly. It is also possible to prepare a table that maps "buss" to "buses," "leafs" to "leaves," and "city" to "cities" (a table that maps misspellings to correct spellings), and then perform a bulk conversion after machine translation.
[0115] Also, depending on the learning results of the neural network, the code in the input source text may differ from the code in the corresponding translation (a type of mistranslation).
[0116] To prevent this, when performing neural machine translation of the original text, it is a good idea to surround the encrypted part with characters that indicate that it is a special string (inserting special characters before and after the encrypted part). For example, by inserting space characters before and after the encrypted part and sending it to the neural network, such as "AAA 100 contains AAB 101," it is possible to prevent the encrypted part from being converted into another code (a type of mistranslation).
[0117] In addition to characters indicating spaces, it is also effective to surround code parts in the document to be translated with brackets (""). Similarly, it is also effective to surround them with quotation characters ('', "", "", '', '', etc.), parentheses, double parentheses, double brackets, square brackets, double square brackets, curly brackets, tortoiseshell brackets, double tortoiseshell brackets, angle brackets, double angle brackets, guillemets, or corner brackets. This allows the enclosed parts to be indicated as parts with a special meaning (code) that is different from the other parts, and then neural machine translation can be performed on them.
[0118] Characters inserted (added) by these processes are deleted in post-processing after machine translation. For example, if a code is enclosed in brackets and machine translated as "AAA" 100, it will be translated as "AAA" 100. The double quotation marks are not in the original text, but were added during the above process, so they are deleted after machine translation to become AAA 100. Deletion is possible using regular expressions.
[0119] If the encryption uses characters other than the alphabet or symbols, such as kanji or hiragana, there is a problem that they are translated into English and cannot be decrypted (for example, the character string "asai" assigned as the encryption string may be translated into English as "shallow," or a meaningless character string such as "aai" may be translated into English as "Aai"). Therefore, it is preferable to use a character string of the alphabet or symbols for the encryption.
[0120] Furthermore, when using the table in Figure 3, the first letter of a sentence in the translated English text may end up in lowercase (in English, the first letter of a sentence should start with a capital letter). In this case, it is a good idea to capitalize the first letter of the sentence during the spell check later. For example, you could use a regular expression to search for the alphabet at the beginning of a sentence (or the letter after a period) and make it all uppercase.
[0121] The data in the database of FIG. 3 may be registered by purchasing or obtaining it from a vendor, or may be registered by the user.
[0122] Fig. 4 is a flowchart showing the Japanese-English translation process of the computer program included in the translation system in the first embodiment of the present invention. Fig. 5 is a flowchart following Fig. 4.
[0123] The process shown in this flowchart is executed by reading a computer program recorded in storage device 111 onto RAM 109 and sequentially executing it by CPU 101. The program may be stored in an executable format of compiled machine language, may be in a format in which source code is sequentially executed by an interpreter, or may be in a format written in an intermediate language.
[0124] Here, word processing software (such as Microsoft's WORD or the open source OpenOffice) is used, and the program in the flowchart of Figure 4 is executed using its macro (such as VBA: Visual Basic For Applications). The database may be a dedicated database or a CSV file, but here the database is formed using spreadsheet software such as Microsoft's EXCEL, and the database is accessed using the VBA of the word processing software. The word processing software and the type of program language used are not limited in the practice of the invention.
[0125] 4, in step S101, the user opens a document file in which a sentence to be translated is recorded using a word processing software (the document file is expanded from the auxiliary storage device to the RAM). A spreadsheet software program is also opened.
[0126] The document file may be a text file written in plain text format, a word processing software document file, or an HTML file. It may also be a file in which characters included in an image are converted to text by OCR. An image file may be input, and a file may be created in which the character portion is converted to text data by the OCR function in step S101. As long as a sentence can be written in the file, the file format is not limited to a specific one. The document may be received from another computer using a file transfer protocol or email software, or may be input from a storage device such as a USB memory. It may also be input by the keyboard 119 or a microphone. The document may be downloaded from the Internet.
[0127] Let's say we are translating a patent specification about an electric vehicle. The document file contains the following:
[0128] "
[0023]
[0129] The electric vehicle 100 comprises an engine 101 and a display unit 102, and the engine 101 is connected to a signal input means 103 and the display unit 102. When the input of the signal input means 103 is high, the display unit 102 displays a warning.' The translation process in this embodiment will be explained using the translation of this sentence as an example. Note that
[0023] is the paragraph number in the document.
[0130] In step S103, the document to be translated is pre-processed in order to:
[0131] (1) Clarify the units of text (text divisions) to be sent to the other company's computer resource 200 at one time.
[0132] (2) In order to reduce mistranslations, the length of one sentence processed at one time in the other company's computer resource 200 is shortened.
[0133] (3) Clarify the parts of a sentence that have different meanings.
[0134] For example, in this embodiment, the unit of text sent to the other company's computer resource 200 at one time is from the beginning of the sentence to the period (.). Assuming that the paragraph number is also a unit of text, in preprocessing, the paragraph number portion is searched using regular expressions and a period is added after it (in the above example, a period is added after
[0023] ).
[0135] In addition, compound sentences (sentences consisting of two or more coordinate clauses (a group of words including a combination of a subject and a predicate)) are less likely to result in mistranslations if they are broken down into shorter sentences. For this reason, in preprocessing, compound sentences are broken down into two or more sentences (in the above example, the sentence "The electric vehicle 100 has an engine 101 and a display unit 102, and the engine 101 is connected to signal input means 103 and the display unit 102" is converted into two sentences: "The electric vehicle 100 has an engine 101 and a display unit 102. The engine 101 is connected to signal input means 103 and the display unit 102.").
[0136] A complex sentence consisting of an adverbial clause and a main clause indicating a condition, time, cause, reason, etc., can be translated more accurately by processing it as one sentence rather than converting it into two sentences and processing them, because the adverbial clause and the main clause are related to each other. On the other hand, there is a trade-off in that a longer sentence is more likely to cause mistranslation. In this embodiment, the adverbial clause and the main clause are treated as one processing unit, but by inserting a line feed code between them, machine translation is performed assuming that the two are separate clauses (machine translation is performed after inserting a line feed code between the clauses that make up the complex sentence). This makes it possible to make the translation more accurate. (In the above example, a line feed code is inserted after "If it is high" in the sentence "If the input of the signal input means 103 is high, the display unit 102 displays a warning.").
[0137] Similar processing may be performed not only on adverbial clauses, but also on adverbial phrases, adjective clauses, and adjective phrases.
[0138] By the process in step S103,
[0139] "
[0023]
[0140] The electric vehicle 100 includes an engine 101 and a display unit 102. The engine 101 is connected to a signal input means 103 and the display unit 102. When the input of the signal input means 103 is high, the display unit 102 displays a warning.
[0141] "
[0023] .
[0142] The electric vehicle 100 includes an engine 101 and a display unit 102. The engine 101 is connected to a signal input means 103 and the display unit 102. When the input of the signal input means 103 is high,
[0143] The display unit 102 displays a warning." (A line feed code is inserted after "If yes.").
[0144] In step S105, the words in the sentence to be translated after the process in step S103 are converted into bilingual words in a one-to-one manner using the database in FIG.
[0145] This process reads records in the database in Figure 3 in order, starting with record number 0, performs a full-text search of the document to be translated using the words in those records as target words, and replaces the words found with the translated words in the same record in Figure 3. Note that it is also possible to replace the words with the code written in that record here, but this would reduce human readability during later checks, so it is preferable to replace them with translated words that are understandable to humans (translators).
[0146] In the example of FIG. 3, first, "electric vehicle" in record number 0 is set as the search word in the document to be translated, and "electric vehicle" is set as the word to replace it (word after replacement). The CPU searches for the word "electric vehicle" (target word) from the beginning of the document to be translated, and if it exists, it is replaced with "electric vehicle". Note that before the replacement, a dialog box such as "Do you want to replace XX with △△?" may be displayed to the user, and whether or not to replace the word may be determined based on the user's YES / NO input. Also, full-text search and batch replacement (replacement of all target words in the document) may be performed mechanically without obtaining the user's consent. When the processing of record number 0 is completed, it is determined whether there is a next record, and the same processing is performed up to the last record.
[0147] It is desirable to insert half-width or full-width spaces (blank symbols) before and after replacement words in order to improve readability and to make it clear that they are word boundaries during machine translation. Replacement words may also be enclosed in brackets ("") to make it clear that they are definite words. They may also be enclosed in quotation characters ('', "", "", '', '', etc.), parentheses, double parentheses, double brackets, square brackets, double square brackets, curly brackets, tortoiseshell brackets, double tortoiseshell brackets, angle brackets, double angle brackets, guillemets, or corner brackets, or other symbols. In this way, they may be enclosed in symbols that indicate that the enclosed part is different from the other parts.
[0148] If no data has yet been registered in the database (FIG. 3) (such as at the beginning of system implementation), the process of step S105 in FIG. 4 is skipped (not executed, and the sentence to be translated remains unchanged).
[0149] If the data in Figure 3 is registered in the database,
[0150] "
[0023] .
[0151] The electric vehicle 100 includes an engine 101 and a display unit 102. The engine 101 is connected to a signal input means 103 and the display unit 102. When the input of the signal input means 103 is high,
[0152] The display unit 102 displays a warning.
[0153] "
[0023] .
[0154] The electric vehicle 100 includes an engine 101 and a display 102. The engine 101 is connected to a signal input unit 103 and a display 102. When the input of the signal input unit 103 is high,
[0155] display 102 will display a warning." (Here, to improve readability, half-width spaces (blank symbols) are inserted before and after the replaced words.)
[0156] In step S107, the user visually checks the document created in the processes up to step S105, and if necessary, edits the sentences, registers new words and passwords, and converts the newly registered words all at once within the document.
[0157] This is done in order to:
[0158] (1) Change any long sentences that remain into shorter sentences.
[0159] (2) Correct any words that have been erroneously converted in the batch conversion in step S105.
[0160] (3) Words to be replaced with their corresponding translations and words to be encrypted are converted in bulk and registered in a database.
[0161] That is, the user visually checks the document created in the process up to step S105, and converts any long sentences that remain into shorter ones. For example, a compound sentence consisting of two or more sentences is divided by a period to create two or more sentences with the same meaning.
[0162] The user also corrects any words that were incorrectly converted in the batch conversion in step S105. There are cases where the same word should be translated into a different word due to differences in technical fields, etc. In such cases, the user converts the word using batch conversion. Also, if necessary, the user updates the database to the correct word, or prepares multiple databases to be used depending on the technical field, etc.
[0163] For words that were not converted in step S105 because the words were not registered in the database (particularly the words before the reference symbols), the user registers new words, translation words, and codes in step S107, and performs bulk conversion of the newly registered words within the document.
[0164] 6 is a flow chart showing a word and cipher registration process of a computer program included in the translation system in the first embodiment of the present invention. This process is executed in response to a user's input in step S107 of FIG.
[0165] With reference to Fig. 6, it is assumed that the user is performing input and editing processing of a sentence in a document in step S203. If the user considers it necessary to register a new word in the database, the user selects that word in the document in step S205. This is done by moving the cursor to the beginning (or end) of the word to be registered while the document is being displayed in a word processing software, and pressing the directional keys to the end (or beginning) of the word to be registered while pressing the shift key, thereby highlighting the word to be registered (or distinguishing it from the rest by changing the color or adding an underline, etc.). A word may also be selected by dragging the word with the mouse.
[0166] When a shortcut key for word registration is pressed (or a display indicating word registration is selected from the displayed menu), a dialog box is displayed in step S207. The dialog box has a field for entering the translation of the selected word. The user determines the translation of the selected word by entering the translation of the selected word in this field. Also, candidates for the translation of the selected word may be displayed in the dialog box by a database different from that in FIG. 3 or a learned neural network, and the translation of the selected word may be determined by accepting a selection from the user.
[0167] Note that the word selected in step S205 may be a word partially converted in step S105 of FIG. 4. For example, when a translation of "display unit" for "display unit" is recorded in the database, "information display unit" in the document is converted to "information display unit" by the process of step S105. The user may select "information" and register the word "information" as a translation in the database, or may select "information display unit" and register the word "information display unit" as a translation in the database.
[0168] In step S209, a full-text search is performed for the same word as the word selected in step S205, and in step S211, it is replaced with the translation determined in step S207. Also in this replacement, similar to step S105 of FIG. 4, a dialog box such as "Replace ○○ with △△?" may be displayed to the user before replacement, and it may be determined whether to replace the word based on the user's YES / NO input. Also, without obtaining the user's consent, full-text search and batch replacement (replacing all target words in the document) may be performed mechanically.
[0169] In step S213, the bottom line of the database (FIG. 3) where data is registered is searched for. The line just below this is set as the registration line for the current word, and in step S215 a password is created from this record number. In step S217, new numbers, words, passwords, bilingual words, etc. are registered in the line just below the bottom line of the database. This updates the database, which is a translation asset.
[0170] After the process in step S217, the process returns to step S203. If no word is registered in step S205, the process returns to step S203.
[0171] The event for starting the word registration process in step S205 may be pressing a shortcut key for word registration, selecting a display indicating word registration from a displayed menu, or voice input (such as "word registration"), but it is preferable to use a shortcut key (so that words can be registered without using a mouse or voice input). It is preferable that the shortcut key is assigned to pressing both the control key and a specific key, for example, so as not to interfere with character input.
[0172] 4 is completed, a check is made in step S109 to see whether there are any parts in the document generated in the processes up to step S107 that are difficult to machine translate. Specifically, this is as follows:
[0173] (1) If the length of a sentence (from the beginning of the sentence to the period) is longer than a specified length, the sentence is deemed difficult to machine translate and a warning is output.
[0174] (2) If the distance from the beginning of a sentence or a line break code to the next line break code or a period is longer than a specified length, the sentence is deemed difficult to machine translate and a warning is output.
[0175] (3) If there is a word that appears a predetermined number of times or more and has not been converted (registered in the database) in the process of step S105 or S107, a warning is output to request that it be converted.
[0176] (4) If there is a word that appears before the reference code and that has not been converted (registered in the database) in the process of step S105 or S107, a warning is output to request that it be converted.
[0177] (5) If the subject and predicate in a sentence do not correspond, a warning is output.
[0178] (6) If a sentence does not contain one subject and one predicate, a warning is issued.
[0179] If the check fails in step S109, the user continues the process in step S107. Note that the process in step S109 may be omitted.
[0180] In step S111, the document obtained by the processes up to step S109 is subjected to a process of converting English words into codes all at once.
[0181] This process is similar to the process in step S105, and involves reading records in the database in FIG. 3 in order starting from record number 0, performing a full-text search of the document to be translated using the bilingual words in those records as target words, and replacing the words found with the code listed in the same record in FIG. 3.
[0182] This means:
[0183] "
[0023] .
[0184] The electric vehicle 100 includes an engine 101 and a display 102. The engine 101 is connected to a signal input unit 103 and a display 102. When the input of the signal input unit 103 is high,
[0185] display 102 displays a warning.
[0186] "
[0023] .
[0187] The AAA 100 includes an AAB 101 and an FER 102. The AAB 101 is connected to an AAC 103 and an FER 102. When the input of the AAC 103 is high,
[0188] FER 102 displays a warning.'
[0189] In this way, words are replaced with the code written in the record, generating a document that is incomprehensible to humans (first encryption).
[0190] When reading records in the database of FIG. 3 in order from record number 0, it is preferable to sort the data in the process of step S111 in order of length of translated words (or in order of number of words included).
[0191] For example, "display unit" and "information display unit" may be registered as bilingual words in a database. When the word "information display unit" is found in a sentence, only the "display unit" portion is searched for and encrypted first, and "information" remains unencrypted. Therefore, by searching in order of length of bilingual words (in order of containing more words), it is possible to encrypt longer words first. Search terms may be picked up from the database in order of length of bilingual words (in order of containing more words) without sorting the data in the database.
[0192] In step S113, a search may be performed to see if there are any words (particularly nouns) before the reference code that cannot be converted into a code. In other words, if there are any words before the reference code that cannot be converted into a code, in step S115, the user is prompted to register them in the database, as in step S109. The processing in steps S113 and S115 ensures that the words before the reference code are encrypted. Note that the processing in steps S113 and S115 may be omitted.
[0193] In step S117, the document data created by the word processing software in the process up to step S115 is converted to text data (data such as UTF-8). Although it may be converted to HTML data, it is preferable to use text data in order to reduce the amount of data to be communicated and the amount of data to be translated.
[0194] 5, one sentence to be processed (a processing unit to be machine translated at one time) is identified, starting from the top of the data obtained in step S117. This is a process in which the data from the top to the next period is treated as the first processed data, and the data up to the next period is treated as the second processed data, and so on until the end of the sentence (the end of the nth processed data). The end of the nth processed data (end of the data) is the period or the last character of the data.
[0195] "
[0023] .
[0196] The AAA 100 includes an AAB 101 and an FER 102. The AAB 101 is connected to an AAC 103 and an FER 102. When the input of the AAC 103 is high,
[0197] FER 102 displays a warning.
[0198] The first sentence would be "
[0023] ."
[0199] The second sentence would be "(line break code) AAA 100 has AAB 101 and FER 102."
[0200] The third sentence would be "AAB 101 is connected to AAC 103 and FER 102."
[0201] The fourth sentence is: "If the AAC 103 input is high, (line feed code)
[0202] FER 102 displays a warning." These are processed in order from the first sentence, and are processed in step S121. Note that the second and fourth sentences are also processed as a single sentence, but they contain a line feed code.
[0203] In step S121, it is determined whether the sentence to be processed does not require translation. In the above example, the first sentence is the paragraph number "
[0023] .", which contains only symbols and numbers. In this way, if the sentence is only composed of symbols, numbers, and alphabets, there is no need to translate it, so it is added directly to a variable (queue) to which the translation result is assigned. Alternatively, it may be added after processing such as converting full-width characters to half-width characters. In other words, if the answer is YES in step S121, the sentence is processed as necessary in step S123 and added to the end of the queue. This makes it possible to reduce communication costs and translation costs. It also makes it possible to prevent mistranslations from occurring.
[0204] If the answer is NO in step S121, then in step S125 the sentence is sent by HTTPS communication to the external other company's computer resource 200. The encryption performed in the HTTPS communication is the second encryption.
[0205] In step S127, the translation result of that one sentence is received from the external other company's computer resource 200. This communication is also encrypted using HTTPS, and is decrypted by the computer 100 (second decryption). The translation result is added to the end of the queue.
[0206] When the translation is added to the end of the queue, a space (or two spaces) may be inserted after any period. In other words, received documents may be joined with a space between them. This is to improve the readability of the translated text.
[0207] In step S129, it is determined whether the processing has been completed up to the last sentence (the nth sentence), and if the result is NO, the processing from step S121 (processing of the next sentence) is repeated. This completes the processing from the first sentence to the nth sentence.
[0208] If the answer is YES in step S129, the queue contains the translation result. For example, in the above example,
[0209] "
[0023] .
[0210] The AAA 100 includes an AAB 101 and an FER 102. The AAB 101 is connected to an AAC 103 and an FER 102. When the input of the AAC 103 is high,
[0211] FER 102 displays a warning.
[0212] "
[0023] .
[0213] AAA 100 includes AAB 101 and FER 102. AAB 101 is connected to AAC 103 and FER 102. When an input of AAC 103 is high,
[0214] FER 102 displays warning." and this is included in the queue. In this way, the translation is also encrypted, so even if someone outside of your company obtains the translation, they cannot know its contents.
[0215] In step S131, the code in the queue is converted into an English word.
[0216] This process is similar to the process in step S105, and involves reading records in the database in Figure 3 in order starting from record number 0, performing a full-text search of the document to be translated using the ciphertext described in that record as the target word, and replacing the discovered ciphertext with the bilingual word described in the same record in Figure 3.
[0217] This allows the above
[0218] "
[0023] .
[0219] AAA 100 includes AAB 101 and FER 102. AAB 101 is connected to AAC 103 and FER 102. When an input of AAC 103 is high,
[0220] FER 102 displays warning.
[0221] The sentence:
[0222] "
[0023] .
[0223] electric vehicle 100 includes engine 101 and display 102. engine 101 is connected to signal input unit 103 and display 102. When an input of signal input unit 103 is high,
[0224] display 102 displays warning.”
[0225] By performing this conversion in-house, confidentiality of the document is guaranteed.
[0226] In this conversion, it is advisable to set a full-text search flag so that the code is not matched to a part of the word. For example, this prevents the word "ABS" in ABSTRACT from being mistakenly converted as a code. If you are using Microsoft's VBA, this process can be achieved by setting the MatchWholeWord property to True and performing a full-text search.
[0227] In addition, in order to prevent words with meanings such as CPU and RAM from being translated as code words, it is desirable to prevent words with such meanings from being registered in the database as code words (to be reserved words).
[0228] In step S133, post-processing is performed, which specifically includes the following processing.
[0229] (1) The process of deleting unnecessary spaces. For example, when two or more spaces are inserted in a position other than after a period, the process of combining them into one. Also, the process of deleting spaces between numbers and letters in a reference code consisting of letters and numbers (such as 100A).
[0230] (2) A process to delete line feed codes other than those following a comma or a period. For example, in the above example, the process to delete the line feed code following "When an input of signal input unit 103 is high,"
[0231] (3) Converting incorrect plurals to the correct form, for example, converting boxes to boxes.
[0232] (4) The process of changing the word "said" used in the translation of "the aforementioned" to the more general word "the."
[0233] (5) When special codes, such as enclosing a code in brackets, are added during translation, these codes are deleted. For example, if a code such as AAA100 is machine translated with brackets added, such as "AAA" 100, double quotation marks that were not in the original, such as "AAA" 100, are added, and this is deleted.
[0234] (6) Correcting articles. The first time a word appears, it should be preceded by an indefinite article, and the second or subsequent times, a definite article should be added.
[0235] (7) A process for deleting full-width periods. (This is a process for deleting periods such as the one in
[0023] above.)
[0236] In step S135, the user checks the post-processed document and corrects any errors.
[0237] In step S137, any errors in words or syntax are corrected using a spell checker in the word processing software, and any errors in plurals are also corrected at the same time. The user corrects and edits the errors, and the final translation is completed.
[0238] In this way, machine translation can be performed using external computer resources while maintaining the confidentiality of both the original text and the translation.
[0239] In addition, by registering a word containing superscript, subscript, italic, bold, or underlined characters in the database, it is converted into a code and machine translated. The code can later be converted back into a word containing superscript, subscript, italic, bold, or underlined characters, which prevents the word containing superscript, subscript, italic, bold, or underlined characters from being mistranslated by machine translation. For example, the word may be registered in the database as "SiO2", its corresponding code as "ABC", and its translation as "SiO2", or it may be registered as in record 3504 in Figure 3. If the word before translation and the translated word are the same, only the word and the code may be registered in the database, and the translation may not be registered. The word may be registered in full-width and the translation as half-width (or vice versa) in the database.
[0240] In addition, if the document has low confidentiality, the encryption in steps S111 to S115 and the decryption in step S131 are not performed, and the document that has been partially converted into bilingual words may be machine translated. In this case, the first encryption and the first decryption are not performed, and only the second encryption and the second decryption are performed. This method is effective when it is not necessary to keep the contents of the translation secret from other companies in FIG. 1. This method can prevent the words in the document after machine translation from fluctuating. In this method, the bilingual words converted by the batch conversion may be machine translated after adding brackets or double quotes to surround them. This allows the machine translation to be performed in a state where the words are indicated as words that do not need to be translated, thereby reducing the fluctuating words and mistranslations in the document after machine translation.
[0241] There are multiple translations for "electric vehicle," including "electric vehicle," "electrical vehicle," "electric-powered vehicle," "electronic vehicle," "battery car," and "battery vehicle." When machine translation is performed using a neural network, it is not clear which translation will be used depending on the context, and the translated word expression will vary. Also, the appropriate translation often differs depending on the field of the document to be translated. In this embodiment, words are converted into unified translations in advance before machine translation using a neural network. This eliminates variation in expression.
[0242] If the password is in full-width characters and machine translated from Japanese to English, the password will be in half-width characters after machine translation. Therefore, in the conversion in step S131, the entire text must be converted taking this into account. If the password is in half-width characters from the beginning, this problem will not occur.
[0243] [Second embodiment]
[0244] FIG. 7 is a flowchart showing a word and encryption registration process of a computer program included in the translation system in the second embodiment of the present invention.
[0245] The process shown in FIG. 6 is not limited to the process within the translation system of FIG. 1, but can be performed separately and independently as a word registration process (database creation process).
[0246] That is, after opening a document in a word processing software in step S201 in FIG. 7, it is also possible to perform the processes in and after step S203 (the same as the processes in and after step S203 in FIG. 6).
[0247] In the first and second embodiments, if translation is performed without using a code (when translating a document that does not need to be kept secret), there is no need to register a code in the database. In this case, the process of step S215 is unnecessary.
[0248] In addition, although the password is registered in the database, it is also possible to use a password that is uniquely determined from the record number, and to calculate the password from the record number by computer 100 without recording the password in the database.
[0249] The processing shown in such a flowchart is executed by reading a computer program recorded in storage device 111 onto RAM 109 and sequentially executing it by CPU 101. The program may be stored in an executable format of compiled machine language, may be in a format in which source code is sequentially executed by an interpreter, or may be in a format written in an intermediate language.
[0250] The programs in the flowcharts of Figures 6 and 7 may be executed using a word processing software (such as Microsoft's WORD or the open source OpenOffice) and its macros (such as VBA: Visual Basic For Applications). The database may be a dedicated database or a CSV file, but in this example, the database is formed using a spreadsheet software such as Microsoft's EXCEL, and the database is accessed using the VBA of the word processing software.
[0251] 7, in step S201, the user opens a document file in which a sentence to be translated is recorded using a word processing software (the document file is expanded from the auxiliary storage device to the RAM). A spreadsheet software is also opened.
[0252] The document file may be a text file written in plain text format, a word processing software document file, or an HTML file. It may also be a file in which characters included in an image are converted to text by OCR. An image file may be input, and a file may be created in which the character portion is converted to text data by the OCR function in step S201. As long as a sentence can be written in the file, the file format is not limited to a specific one. The document may be received from an external computer using a file transfer protocol or email software, or may be input from a storage device such as a USB memory. It may also be input by the keyboard 119 or a microphone. The document may be downloaded from the Internet.
[0253] Let's say we are translating a patent specification about an electric vehicle. The document file contains the following:
[0254] Assume that the following sentence was written: "The electric vehicle 100 comprises an engine 101 and a display unit 102, and the engine 101 is connected to a signal input means 103 and the display unit 102. When the input of the signal input means 103 is high, the display unit 102 displays a warning."
[0255] 4, the user checks the contents of the opened document and corrects them if necessary, and also adds (inputs) text if necessary.
[0256] In step S205 in FIG. 4, the user selects a word to be registered in the database by visually inspecting the text. This is done by moving the cursor to the beginning or end of the word and highlighting the target word (or displaying the text and / or background in a different color from the rest of the text) by pressing SHIFT+directional keys or dragging the mouse. Words are mainly registered for words that precede a reference symbol, but words without reference symbols may also be registered. A reference symbol is a symbol used to refer to symbols (such as numbers or letters used with leader lines to indicate components) written in a drawing. For example, the "100" in "electric vehicle 100" is the reference symbol, and the word that precedes the reference symbol means "electric vehicle."
[0257] The selection of words may be made by human judgment (key or mouse input) as described above, but it may also be possible to automatically detect and select a word that exists before a reference symbol, for example. Moreover, it is also possible to automatically detect and select a word regardless of the presence or absence of a reference symbol.
[0258] If a registration button (or a shortcut key for registration) displayed on the screen is pressed while a word is selected, then in step S207 CPU 101 accepts input of a translation of the word selected in step S205. This may be done by displaying a dialog box on the screen and accepting input from the user via a keyboard or voice, or by using a UI (user interface) that displays multiple word candidates from a dictionary and allows the user to select a translation from among them. This determines the correspondence between the word before translation and the word after translation.
[0259] In step S209, the CPU 101 performs a full-text search of the document to be translated, searches for the pre-translation word determined in the process up to step S207, and converts it to the corresponding translated word in step S211. The full-text search may be performed without user confirmation, or a dialog box such as "Replace with (translated word)? (YES / NO)" may be displayed on the screen each time a word is searched, and the user may select whether to replace the word. As a result, all words in the entire text, including the word selected in step S205, are converted to the translated word determined in step S207 at once. For example, if "electric vehicle" is associated with the word "electric vehicle", the word "electric vehicle" in the document is converted to "electric vehicle" at once.
[0260] In step S213, CPU 101 searches database 207 for the bottom line where information is registered (the bottom line in FIG. 3), and obtains the number (serial number) of the next blank line.
[0261] In step S215, the acquired number is converted to hexadecimal and a code for the target word is created. In step S217, the code is recorded as one record in the line below the bottom line of registered information in the database, together with the word before and after translation. The information registered in the database becomes an asset as a translation dictionary for future use.
[0262] After the process of step S217, the process from step S203 is repeated. If the registration button is not pressed in step S205, the process from step S203 is repeated.
[0263] Once the full text has been processed, the user saves the document, which is the source text with the partial translation, and which may then be subjected to neural machine translation.
[0264] The words registered in the database in the processes of Figures 6 and 7 are mainly words written before the reference symbols. For example, if the original text to be translated is "The electric vehicle 100 has an engine 101 and a display unit 102, and the engine 101 is connected to a signal input means 103 and the display unit 102. When the input of the signal input means 103 is high, the display unit 102 displays a warning," then "electric vehicle," "engine," "display unit," and "signal input means" are registered.
[0265] Words such as "input" and "warning" that do not have a reference code may also be registered in the database 207 in FIG. 4 together with the password and the translation word.
[0266] Furthermore, at least some of the words accompanied by the reference signs may be excluded from registration, i.e. at least some of "electric vehicle", "engine", "display unit", and "signal input means" may not be registered.
[0267] After registration in step S213, the words in the source text may be sorted in order of length, with longer words closer to number 0 (higher) and shorter words closer to higher numbers (lower).
[0268] Through the above processing, the sentence "The electric vehicle 100 includes an engine 101 and a display unit 102. The engine 101 is connected to a signal input means 103 and the display unit 102. When the input of the signal input means 103 is high, the display unit 102 displays a warning." is changed to:
[0269] This is converted into the sentence "The electric vehicle 100 is equipped with an engine 101 and a display 102. The engine 101 is connected to a signal input unit 103 and a display 102. When the input of the signal input unit 103 is high, the display 102 displays a warning." The user checks this sentence and determines whether there are any errors. If there are any errors, the user edits the sentence on a word processing software.
[0270] [Third embodiment]
[0271] FIG. 8 is a flowchart showing a Japanese-English translation process of a computer program included in a translation system in the third embodiment of the present invention.
[0272] The process shown in this flowchart is executed by reading a computer program recorded in storage device 111 onto RAM 109 and sequentially executing it by CPU 101. The program may be stored in an executable format of compiled machine language, may be in a format in which source code is sequentially executed by an interpreter, or may be in a format written in an intermediate language.
[0273] With reference to the figure, in step S101, CPU 101 inputs a document (sentence) to be translated in accordance with a user's instruction. The document input here is assumed to be the sentence after conversion has been performed by the process in FIG. 7. Here, an example will be described in which the sentence "Electric vehicle 100 is equipped with engine 101 and display 102, and engine 101 is connected to signal input unit 103 and display 102. When the input to signal input unit 103 is high, display 102 displays a warning." is processed.
[0274] In step S103, CPU 101 searches for parts of the input document that are difficult to machine translate. For example, when translating from Japanese, the document is broken down into multiple sentences by dividing it with a comma such as "." (a period if translating from English), and it is determined whether each sentence is within a predetermined length (i.e., whether the length of each sentence is equal to or less than a predetermined threshold) (the threshold is selected to be, for example, a value of about 80 to 120 characters depending on the performance of other company's computer resource 200).
[0275] If a sentence is long, machine translation is likely to be inaccurate, so the length of the sentence is checked in advance and long sentences are deemed untranslatable. For long sentences, for example, if the sentence is a compound sentence, a sentence such as "A is B, and C is D" is converted into multiple sentences such as "A is B. C is D," and a conversion / editing process is performed to shorten each sentence. This conversion may be performed automatically, or the user may be shown sentences that are determined to be long and prompted to manually correct them. In addition, when performing automatic conversion, it is desirable to show the user the sentence after automatic conversion (and the sentence before conversion, if necessary) and have the user check whether the conversion has been performed correctly.
[0276] To check for difficult parts of machine translation, it is possible to check whether a sentence has a subject (for example, whether it contains a phrase equivalent to "XX is" or "XX is"), whether the subject and predicate in the sentence correspond, etc. To perform machine translation more accurately and without errors, it is also possible to set the condition for a sentence to be translatable, for example, that there is only one subject and only one predicate in the sentence.
[0277] These checks may be performed by matching sentence patterns using regular expressions. Checks may also be performed using morphological analysis. For example, a trained recurrent neural network (RNN) may be used to input the sentence to be checked and output a sentence that is not difficult to machine translate, or to output a signal indicating whether or not there is a problem and prompt the user to make corrections if there is a problem.
[0278] In step S105, a check is made to see if the sentence is machine translatable, and if there is a difficult part, the process in step S103 is repeated.
[0279] If it is determined in step S105 that machine translation is possible, preprocessing of the sentence is performed in step S107. This is a process to facilitate machine translation so that there are no mistakes in the machine translation. For example, a compound sentence has a structure of "When A is B, C does D.", but in order to prevent translation mistakes, a marking code is embedded in the delimiting part of the compound sentence (immediately after "when," in the above sentence). The code is a line feed code, a space, a delimiting character, etc., and a code that does not affect the translation and makes the delimiting part easy to understand for a machine is selected. In addition, for example, a compound sentence having a structure such as "A is B, C is D.", the marking code may be embedded in the delimiting part of the compound sentence (immediately after "and," in the above sentence). The process in step S107 reduces the unit of sentences processed in the machine translation, preventing the selection of an incorrect subject-predicate combination in the machine translation and other mistranslations.
[0280] By the processing in steps S103 to S107, for example,
[0281] The sentence "The electric vehicle 100 includes an engine 101 and a display 102. The engine 101 is connected to a signal input unit 103 and a display 102. When the input of the signal input unit 103 is high, the display 102 displays a warning." is
[0282] The electric vehicle 100 includes an engine 101 and a display 102. The engine 101 is connected to a signal input unit 103 and a display 102. When the input of the signal input unit 103 is high, (line feed code)
[0283] Display 102 shows a warning."
[0284] (The text is converted into a single sentence in step S103, and line feed codes are inserted at the boundaries of sentences that form a compound sentence in step S107).
[0285] In step S109, the database in the storage device 111 is referenced to convert words (particularly nouns) that appear before the reference signs in the document to be translated into code (first encryption). The database compares the original text, the encrypted character string, and the translated character string.
[0286] In step S109, the words before the reference code are converted into a code in a lump.
[0287] The electric vehicle 100 includes an engine 101 and a display 102. The engine 101 is connected to a signal input unit 103 and a display 102. When the input of the signal input unit 103 is high, (line feed code)
[0288] display 102 displays a warning.
[0289] "AAA100 includes AAB101 and FER102. AAB101 is connected to AAC103 and FER102. When the input of AAC103 is high, (line feed code)
[0290] FER102 will display a warning."
[0291] is replaced with the ciphertext of
[0292] In such encrypted text, hexadecimal numbers such as "AAA" are processed as words, so that it can be machine translated by an external third-party computer resource 200 without revealing the true meaning to the outside.
[0293] The process in step S109 is performed as follows.
[0294] (1) Search from the beginning of the document for sentence components consisting of combinations of words and reference symbols.
[0295] (2) The words of the searched components are searched for in a database.
[0296] (3) The corresponding code is obtained from the database.
[0297] (4) The words of the components found in (2) above are replaced with the code obtained in (3) above.
[0298] (5) Search for the next occurrence in the sentence of a sentence component consisting of a word and a reference symbol combination.
[0299] (6) Repeat the process from (1) above until all components have been processed. If all components have been processed, end the process.
[0300] Alternatively, the process in step S109 may be performed as follows.
[0301] (1) Read the first word registered in the database.
[0302] (2) Search for all the words that have been read out, starting from the beginning of the document, and replace them with the corresponding passwords registered in the database (full-text replacement).
[0303] (3) The next word registered in the database is read, and the process in (2) above is repeated. When all words registered in the database have been processed, the process ends.
[0304] In the process from step S119 onwards, the data before machine translation obtained in the process up to step S109 is transmitted, starting from the first sentence, via HTTPS to the other company's computer resource 200 which has completed neural network training.
[0305] More specifically, the first sentence of the data before translation (e.g., the sentence up to the period ".") is transmitted to the external other company's computer resource 200 by HTTPS in step S119. This transmission is encrypted (second encryption). In response to the request, the external other company's computer resource 200 translates the received data into the target language and transmits it to the computer 100.
[0306] In step S121, the computer 100 obtains translated data of one sentence from the external other company's computer resource 200 as a response via HTTPS.
[0307] In step S123, it is determined whether translation has been completed up to the last sentence, and if NO, processing of the next sentence is performed in step S119. If translation has been completed up to the last sentence, the process proceeds to step S125.
[0308] In step S125, post-processing of the translated document is performed. Here, the translations obtained in steps S119 to S123 are compiled into one document. In addition, since the translations obtained by the first encryption in steps S109 to S113 contain codes, the codes are converted into corresponding translations in a batch. This is performed, for example, as follows.
[0309] (1) Read the first cipher (e.g., "AAA") registered in the database.
[0310] (2) The retrieved ciphertext is searched from the beginning of the document and replaced with the corresponding translation from the database (full-text replacement).
[0311] (3) The next cipher registered in the database is read, and the process from (2) above is repeated. When all ciphers registered in the database have been processed, the process ends.
[0312] In step S127, a spelling check is performed, including checking plural words, etc. The user performs a final check of the translation, and corrects or edits it if there is a problem.
[0313] By performing the process in step S127,
[0314] "AAA 100 comprises AAB 101 and FER 102. AAB 101 is connected to AAC 103 and FER 102. When an input of AAC 103 is high,(line break code)
[0315] FER 102 displays alert.
[0316] From the text,
[0317] "electric vehicle 100 comprises engine 101 and display 102. engine 101 is connected to signal input unit 103 and display 102. When an input of signal input unit 103 is high,(line break code)
[0318] display 102 displays alert.”
[0319] The following sentence is obtained.
[0320] In steps S125 and S127, the same processes as in steps S135 and S137 in FIG. 5 are performed, thereby making it possible to obtain the same effects as in the first embodiment.
[0321] [Fourth embodiment]
[0322] FIG. 9 is a flowchart showing a Japanese-English translation process of a computer program included in a translation system in the fourth embodiment of the present invention.
[0323] The process in this flowchart is the same as that in FIG. 8, but with the addition of steps S111 and S113.
[0324] After the process in step S109, in step S111, it is determined whether there is a word preceding the reference sign that could not be converted because it is not registered in the database. This may be determined by using a regular expression to determine whether there is a sentence component consisting of a combination of a word and a reference sign, or the user may visually check.
[0325] If the answer is YES in step S111, the process proceeds to step S113. In step S113, the word that could not be converted in step S111, its code, and its translation are registered in the database as a set. For example, the word that could not be converted may be highlighted by highlighting it in black and white, and a dialog box including the words "Please enter a translation for XX" and a word input field may be displayed to accept input from the user. For example, the word "denki jiyuu (electric vehicle)" in the document may be highlighted, and a dialog box including the words "Please enter a translation for 'denki jiyuu (electric vehicle)'" and a word input field may be displayed to accept input of "electric vehicle" from the user. The word, its code, and its translation are registered in the database as a set. It is preferable that the code is generated automatically.
[0326] The process in step S113 is repeated until step S111 becomes NO.
[0327] If the result in step S111 is NO, the process proceeds to step S119.
[0328] [others]
[0329] The above-described multiple embodiments and the elements contained therein (part of the configurations, part of the processes) can be combined or replaced to create new embodiments.
[0330] When performing full-text batch conversion (processing to convert words into translated words or translated words into codes), it is desirable to read search target words from the database in order from the longest word. For example, if there are two records in the database, one for converting "detection means" to "detection means" and the other for converting "information detection means" to "information detection means", the longer word "information detection means" is converted into "information detection means" in a batch process, and then the shorter word "detection means" is converted into "detection means" in a batch process. This prevents only the "detection means" part of "information detection means" in the document from being converted, and improves the efficiency of translation. Similarly, when converting words into codes, by reading the words from the database in order from the longest word, it is possible to prevent only a part of a set of words with a series of meanings from being converted into codes.
[0331] When words are read out in ascending order of record number and converted all at once, the records may be sorted by the number of characters of the registered words.
[0332] In the case of a patent specification, there is a section "Explanation of symbols" at the end, which generally explains important reference symbols and the components (words) they represent. Only the words written in the explanation of symbols section may be encrypted. Before machine translation, a warning may be issued if the words written in the explanation of symbols section are not encrypted in the document.
[0333] In the case of a patent specification, the mode for carrying out the invention, a brief description of the drawings, the effects of the invention, the means for solving the problems, the document name (specification), the name of the invention, the technical field, the background art, the prior art documents, the patent documents, the outline of the invention, the explanation of the symbols, the document name (claims), the document name (abstract), etc. may each be regarded as a word and encrypted. This makes it difficult to tell that the text to be translated is a patent specification.
[0334] In documents such as patent specifications that provide explanations with reference to reference numbers in drawings, the noun word preceding the reference number in the document has a particularly important meaning. By replacing at least a part of the word preceding the reference number with a code (or a translation), the present invention can be highly effective. Of course, at least a part of the words other than the word preceding the reference number in the document may be replaced with a code (or a translation). Also, it is possible to replace at least a part of the words other than the word preceding the reference number in the document with a code (or a translation), without replacing the word preceding the reference number in the document with a code (or a translation).
[0335] Of the sets of noun words and reference signs that appear in a document (sets consisting of a noun word and a reference sign that immediately follows it), it is desirable to register only the noun word portion in the database as a target for batch conversion, and not include the reference sign in the batch conversion. In this way, the same word with different reference signs can be processed using the database (the database can be used more effectively when translating other documents in the future).
[0336] The invention is also applicable to documents that do not contain reference signs, in which case at least some of the words in the document are replaced with a code word (or a translation) and then machine translated.
[0337] The batch conversion process using the database in step S105 in FIG. 4 and step S109 in FIG. 8 may be performed multiple times in succession. For example, if a user records a record of the conversion from "detection means" to "detection means" in advance in the database, "information detection means" in the translation document may be converted to "information detection means". This is a case where a part of a word is converted. In this case, the user selects "information detection means" as a word and records "information detection means" as a new record in the database as its translation. When the next document is translated using that database, if the word "information detection means" exists in the document, the word is first converted to "information detection means" by batch conversion using the database. If the search term "information detection means" appears after that in the database search order, "information detection means" is converted to "information detection means" without any problems. However, if "information detection means" is before "detection means" in the search order in the database, "information detection means" will not be converted to "information detection means" by performing batch conversion process using the database only once.
[0338] Therefore, the batch conversion process using the database is repeated twice (or more) to perform batch conversion again. As a result, even if the original text contains a part of a translated word, such as "information detection means," it is converted to the correct word, such as "information detection means."
[0339] In addition, although the embodiments have been described as examples of machine translation using a neural network, statistical machine translation, phrase-based or rule-based machine translation, artificial intelligence machine translation, deep learning machine translation, etc. may also be used, and the present invention is not limited to machine translation using a neural network.
[0340] The method of transmitting and receiving information to the external other company's computer resource 200 may be via a WEB page. Information may be transmitted and received by an automated method of browser operation. Information may also be transmitted and received using an API provided by a company that operates the external other company's computer resource.
[0341] The post-processing may include removing unnecessary spaces in the text, standardizing words, converting uppercase to lowercase in the text, spell checking, and correcting plural forms of words.
[0342] [Effects of the embodiment]
[0343] According to the above embodiment, both the original text and the translated text are encrypted in two stages, the first encryption and the second encryption. Both encryption stages ensure that the document remains confidential from third parties. Furthermore, the first encryption stage ensures that the document remains confidential from translation service providers. This makes it possible to prevent information from leaking to the outside while still using an external translation service.
[0344] The confidentiality of information can be ensured even if the computer resource that transmits the information is used by an unspecified number of people (or multiple people, other people). Note that "use by an unspecified number of people" means that the site (server) is open to the public and can be accessed from anywhere.
[0345] Furthermore, the conventional neural machine translation has the problem that it is not possible to translate the information contained in the input sentence exactly and without omissions (there is a possibility of missing parts in the translation). Also, there are cases where duplicated parts of the translation are output. In this embodiment, the original text can be compressed (made shorter and simpler) by encryption. Also, the sentence structure can be simplified (made shorter and simpler). This has the effect that even when neural machine translation is adopted, the information contained in the input sentence can be translated exactly and without omissions in the translation, and there is also the effect that duplicated parts of the translation are less likely to be output.
[0346] Furthermore, it has the advantage that previously used learning data (such as bilingual dictionary data and word order data) can be easily used as data to be registered in the database.
[0347] Furthermore, while conventional neural machine translation allows for flexible translation through word embedding, it can also result in the creation of translations with completely different meanings (particularly for words that occur infrequently, proper nouns that occur infrequently, etc.) by selecting completely different words (sometimes meaningless character strings). In addition, for long or complex sentences, it can also result in translations that are incomprehensible. In this embodiment, words are converted into parallel words or codes in advance, all at once, and then machine translated (preferably with quotation marks or the like added). This prevents machine translation from translating words in the original text into parallel words with completely different meanings.
[0348] [Fifth embodiment]
[0349] The hardware configuration of the translation system in the fifth embodiment of the present invention is the same as that in the first to fourth embodiments, and therefore description thereof will not be repeated here. In the fifth embodiment, a sentence to be translated, written in a first language, includes a set of a word that is a noun and a reference sign that follows immediately thereafter. The translation device includes a replacement unit that replaces the set of the word that is a noun and the reference sign that follows immediately thereafter with a character string that corresponds to the set. The sentence that has been replaced by the replacement unit is transmitted to an external computer via the Internet. The translation device receives the translated sentence written in the second language from the computer via the Internet.
[0350] The differences between the fifth embodiment and the above embodiments will be described below.
[0351] FIG. 10 is a diagram showing a specific example of a data structure stored in the first database of the database 207 in the translation system according to the fifth embodiment of the present invention.
[0352] The data registered in this first database does not include a column of codes, as compared with the data in FIG. 3. Words in a first language and their corresponding words in a second language are recorded in association (paired). The data in the first database may be registered by the same registration process as described in the first to fourth embodiments, or past dictionary data or external dictionary data may be registered in the first database. Furthermore, corresponding words in a third to Nth language may be registered to enable translation into multiple languages.
[0353] FIG. 11 is a diagram showing a specific example of a data structure stored in the second database in the database 207 in the translation system according to the fifth embodiment of the present invention.
[0354] The data registered in the second database is created from the data recorded in the first database and a document to be translated, which includes a plurality of sets each consisting of a noun word and a reference code immediately following it. The second database includes columns for numbers, words, translated words, and codes. Of course, it is also possible to enable translation into a plurality of languages by registering corresponding words in a third to an Nth language. In the fifth embodiment, the second database serves as a code table (codebook) used when performing the first encryption and the first decryption.
[0355] In creating the second database, first, the contents of the first database are copied (see lines numbered 0 to 3504), and a password for identifying each word is recorded. The password is the same as that in the first to fourth embodiments.
[0356] Next, a set of a noun word and a reference code immediately following it, which are included in a sentence written in the first language to be translated, is detected. In this detection, a part in which one or more nouns are consecutive and a part in which one or more reference codes such as numbers or letters are consecutively followed may be detected as a set by word segmentation and morphological analysis of the sentence, or a regular expression may be used for the detection. In addition, a word registered in the first database of FIG. 10 may be searched for in the document to be translated, and if a reference code is followed by the word, the word and the reference code may be detected as a set. With regard to a set of a noun word and a reference code immediately following it, the word may be a word in the first language or a word in the second language (a bilingual word).
[0357] In particular, if the text to be translated is written in a first language, a set consisting of a word in the first language and a reference symbol immediately following it is detected. If the text to be translated is written in a first language, but some words have already been replaced with words in a second language, a set consisting of a word in the second language and a reference symbol immediately following it is detected.
[0358] In this example, it is assumed that the sentence to be translated contains a set consisting of the word "electric vehicle 100" and the reference code immediately following it. "Electric vehicle" in the set is replaced with "electric vehicle" using the data in the first database in FIG. 10 (or the second database in FIG. 11). This creates a correspondence between "electric vehicle 100" and "electric vehicle 100," and a new code is assigned to this correspondence, which is then newly registered in the second database in FIG. 11 (see line numbered 4012).
[0359] In this case, the word "electric car 100" may be registered as a word in the first language, or ""AAA" 100" may be registered in which the "electric car" portion is encrypted with the row data numbered 0. In FIG. 11, ""AAA" 100" is registered.
[0360] If you enclose the encrypted portion in special symbols such as double quotation marks, it will be easier to process later, as it will be clear that the portion is encrypted. In other words, when converting to encryption, you can enclose the encryption in special symbols such as double quotation marks and insert it into a document, or when registering the encryption in the database in Figure 11, you can enclose the encryption in special symbols such as double quotation marks and register it. Of course, it is not necessary to enclose it in special symbols.
[0361] Similarly, sets of other words and reference symbols contained in the sentence to be translated are sequentially registered in the second database. Note that in the document to be translated, there are cases where the same word is given different reference symbols (for example, "gear 501a" and "gear 501b"), in which case, each is given a different code and registered in a different line (the translated words are "gear 501a" and "gear 501b", respectively, and each is given a different code).
[0362] By using the second database, as in the first to fourth embodiments, two-way substitution is possible between words, translation words, and codes, and also between words with reference signs, translation words with reference signs, and codes.
[0363] Furthermore, by holding the first database as a database of translation knowledge (a database for word conversion and word translation) and using the second database as a database for translating documents to be translated (a so-called custom-made database), it becomes possible to associate a set of word + reference code with a code or a set of bilingual word + reference code, even for multiple documents in which different reference codes are assigned to the same word.
[0364] In this embodiment, a set of words and reference symbols can be replaced with a code for translation. This allows the sentence to be translated (sent to the server) to be shortened (reducing the amount of data), and translation costs can be reduced. In addition, because the sentence to be machine translated is shorter, mistranslations are less likely to occur.
[0365] In translation, the document to be translated is encrypted using the second database, and if the word in the set in the document to be translated is in the first language, the word in the "word" column of the data in Figure 11 is converted to the encryption for that row.
[0366] In this case, it is possible that a set of word + reference code exists in the document to be translated, and that the word appears without a reference code. In order to shorten the encrypted text, it is desirable to first replace the set of word + reference code with a code (replacement using lines after number 4012 in FIG. 11), and then replace the word with a code (replacement using lines from number 0 to before number 4012 in FIG. 11). This allows words with a reference code to be encrypted as a set of word + reference code, and words without a reference code to be encrypted as just the word.
[0367] Furthermore, if a code + reference code is registered in the word column as shown in Figure 11, the code may be decoded by the database, converted into a word + reference code, and then converted into a code.
[0368] Also, if the words in the set in the document to be translated are replaced with the second language, the words written in the column of the "translated words" in the data in FIG. 11 are converted to the code in that row. At this time, it is possible that a set of translated words + reference signs exists in the document to be translated, and the translated words appear without the reference signs. In order to shorten the text after encryption, it is preferable to first replace the set of translated words + reference signs with the code (replacement using the lines after number 4012 in FIG. 11), and then replace the translated words with the code (replacement using the lines after number 4012 in FIG. 11). In this way, the translated words with the reference signs can be encrypted as a set of translated words + reference signs, and the translated words without the reference signs can be encrypted as the translated words alone.
[0369] If a bilingual word + reference code is registered in the bilingual word column as shown in Figure 11, it is searched for in the text and converted into a code. A bilingual word may also be partially encrypted (e.g., "AAA" 100) and registered in the database.
[0370] If a set of word + reference code is registered as a code + reference code in the second database as shown in FIG. 11, the second database is managed by dividing it into a word-code association part (numbers 0 to 3504 in FIG. 11) and a word + reference code and code association part (numbers 4012 and after in FIG. 11). First, a step of converting a word into a code using the word-code association part is executed, and then a step of converting a word into a code using the word + reference code and code association part is executed. As a result, a word with a reference code can be encrypted as a set of word + reference code, and a word without a reference code can be encrypted only as a bilingual word. Such processing can be appropriately changed according to the algorithm. For example, the second database may be processed as separate databases for the word-code association part (numbers 0 to 3504 in FIG. 11) and the word + reference code and code association part (numbers 4012 and after in FIG. 11).
[0371] When encrypting words or translations, it is desirable to encrypt longer words or translations (words with more characters). For example, when both "input means" (corresponding cipher JKL) and "signal input means" (corresponding cipher YSH) are recorded in the database with different ciphers, if the "input means" with fewer characters is processed first, the "signal input means" included in the sentence to be processed will be converted to "signal JKL", which is undesirable from the viewpoints of shortening sentences and reducing mistranslations.
[0372] By prioritizing encryption of longer words (with more characters) and bilingual words, the "signal input means" included in the sentence to be processed is correctly converted into the "YSH" code, making it possible to shorten the sentence and reduce mistranslations.
[0373] In this way, by using the second database in Fig. 11, a set of words and reference symbols contained in the text to be translated written in the first language is replaced with the corresponding code, and the words contained in the text to be translated written in the first language are replaced with the corresponding code (first encryption). The document that has undergone the first encryption is sent to an external server, where it is processed (translated) while maintaining the confidentiality of the document.
[0374] For documents in the second language that have been translated by an external server, since they contain codes, they are decrypted using the second database (code table, codebook) in Fig. 11. That is, the codes in the text are replaced with the corresponding data of the translated words in Fig. 11.
[0375] For example, suppose you are translating a patent specification about an electric vehicle. The document file contains the following:
[0376] "
[0023]
[0377] The electric vehicle 100 includes an engine 101 and a display unit 102, and the engine 101 is connected to a signal input means 103 and the display unit 102. When the input of the signal input means 103 is high, the display unit 102 displays a warning. In this embodiment, the sentence after the first encryption is as follows:
[0378] "
[0023]
[0379] "HSD" has "HSE" and "HEF", and "HSE" is connected to "HEG" and "HEF". When the input of "HEG" is high, "HEF" displays a warning."
[0380] (For simplicity of explanation, it is assumed that no pre-processing is performed.) Since the reference symbols can be included in the encryption as a single unit in this way, the sentences can be shorter than those in the first to fourth embodiments, and the possibility of mistranslation is reduced. Also, words without reference symbols can be encrypted.
[0381] Furthermore, when this text is translated by an external computer resource, the result is
[0382] "
[0023] .
[0383] "HSD" includes "HSE" and "HEF". "HSE" is connected to "HEG" and "HEF". When an input of "HEG" is high, "HEF" displays warning." By decoding this using the second database, the following sentence can be obtained as in the first to fourth embodiments:
[0384] "
[0023] .
[0385] electric vehicle 100 includes engine 101 and display 102. engine 101 is connected to signal input unit 103 and display 102. When an input of signal input unit 103 is high, display 102 displays warning.
[0386] (For simplicity, we will assume that no post-processing is performed.)
[0387] 12 and 13 are flow charts showing the Japanese-English translation process of a computer program included in a translation system in the fifth embodiment of the present invention.
[0388] The differences between this flowchart and the flowcharts of FIGS. 4 and 5 will now be described.
[0389] In step S107, the selected word and its translation are registered in a first database (FIG. 10).
[0390] After the process in step S109, in step S110, a process is performed to create a second database (FIG. 11) from the contents of the first database and the sentence to be translated. In step S111, a process is performed to encrypt the word or the word + reference code. In this process, the second database is used.
[0391] In step S131 of FIG. 13, a process of converting a code into a word (translated word) or a code into a word (translated word)+reference code is performed using the second database (first decryption).
[0392] The first and second databases may be operated by different types of programs. For example, the first database may be a CSV file or an Excel file, and the second database may use software such as SQLite or MySQL.
[0393] In this embodiment, a word + reference code is registered in the database as one unit together with a code, but a preposition + word, or a preposition + word + reference code may be registered in the database as one unit together with a code. For example, if "Si" in "Si substrate 300" is a preposition and there is a preposition before the word "substrate", "Si substrate" is registered as one unit together with a code, or "Si substrate 300" is registered as one unit together with a code.
[0394] Also, after performing full-text encryption processing of the translation source text using the database of word and code combinations (numbers 0 to 3504) in Fig. 11, a database of word + reference code and code combinations (numbers 4012 to 4016) may be created by searching for code + reference code portions in the document. After that, encryption is performed again using the database of word + reference code and code combinations.
[0395] It is desirable to perform processing to prevent duplicate registration of words and their translations in the database, such as processing to prevent the same pair of words and their translations from being registered, or processing to prevent the same word or translation from being registered in either the database.
[0396] [others]
[0397] The flowchart in this embodiment may be executed by multiple software programs. For example, a part may be executed by a word processing software (macro, etc.), a part may be executed by a script language or a compiled executable file, or a part may be executed by a human being.
[0398] In addition, processing such as converting full-width characters to half-width characters, deleting spaces, and reducing the number of spaces may be performed as appropriate (in any of the processes).
[0399] In the above embodiment, the words in the first language are replaced with words in the second language before encryption (steps S105, S107, etc.), but instead, the words in the first language may be directly encrypted using a database and then sent to an external computer resource.
[0400] You can replace all the words (and / or words + references) in a document with a cipher, or just some of them.
[0401] [Sixth embodiment]
[0402] The hardware configuration of the translation system in the sixth embodiment of the present invention is the same as that in the first to fifth embodiments, and therefore the description will not be repeated here. In the sixth embodiment, the translation system compares an original text and a translation of a sentence translated by the method according to the above-mentioned embodiment (or by another method), and evaluates the result of the translation.
[0403] A translation is evaluated based on whether information present in the original text is retained in the translation and / or whether information not present in the original text is present in the translation. The former judges whether there is any loss of information in the original text, while the latter judges whether unnecessary information has been added to the translation (so-called "information spillover").
[0404] For example, the text after the first encryption illustrated in the fifth embodiment is as follows:
[0405] "
[0023]
[0406] "HSD" has "HSE" and "HEF", and "HSE" is connected to "HEG" and "HEF". When the input of "HEG" is high, "HEF" displays a warning."
[0407] and when this sentence is translated by an external computer resource, the result is
[0408] "
[0023] .
[0409] "HSD" includes "HSE" and "HEF". "HSE" is connected to "HEG" and "HEF". When an input of "HEG" is high, "HEF" displays warning."
[0410] Here, the text after the first encryption contains one "HSD", two "HSE", three "HEF", and two "HEG".
[0411] The translated text also contains one "HSD", two "HSE", three "HEF", and two "HEG".
[0412] This indicates that all pre-translation word information remains in the translated text without being lost (the word loss rate is determined to be 0%). It also indicates that the translated text does not contain words that were not in the pre-translation text (the word emergence rate is determined to be 0%).
[0413] When there is a word missing, for example, a message such as "One "HSE" is missing," "Both "HSE"s are missing," "One of the two "HSE"s in the original text is missing," "An "HSD" in the original text is missing," or "The word missing rate in this text is XX%" may be presented to the user (displayed on the display, announced by voice, etc.). Also, when there is no missing word, a message such as "No missing words" or "0% missing rate" may be presented to the user.
[0414] The warning message is displayed only if there is a defect, and if there is no defect, no message may be displayed.
[0415] When there is word overflow (unnecessary information added to the translation), messages such as "One more 'HSE' has been added," "Two more 'HSE' have been added," "An 'HSD' has been added that was not in the original," "There is one 'HSE' in the original, but two in the translation," and "The word overflow rate for this sentence is XX%" can be displayed to the user. Also, when there is no loss, messages such as "There is no overflow" and "Overflow rate is 0%" can be displayed to the user.
[0416] The warning message is displayed only when there is a leak, and if there is no leak, no message needs to be displayed.
[0417] The warning message may be inserted into the translation (for example, before the beginning or after the end of the sentence) so that the user can see which sentence the warning is directed to, or the sentence and the content of the warning may be made clear to the user by underlining, shading, adding a comment to the sentence, etc.
[0418] Also, rather than presenting a code such as "HSE," the corresponding words "electric vehicle 100" (or "electric vehicle" without the reference number) may be read from the database and presented, or the words "electric car 100" (or "electric car" without the reference number) may be read from the database and presented.
[0419] A message including both the original and translated words, such as "The translation does not contain the word 'electric vehicle 100', which is the translation for the original word 'denki judo 100', may be displayed to the user as a warning.
[0420] In addition, when translating an unencrypted sentence such as "Electric vehicle 100 includes engine 101 and display unit 102. engine 101 is connected to signal input unit 103 and display unit 102. When an input of signal input unit 103 is high, display 102 displays warning," the translation may be evaluated by determining whether there are the same number of translated words ("electric vehicle 100" or "electric vehicle") corresponding to the pre-translation words ("electric vehicle 100" or "electric vehicle") or whether they are different.
[0421] To make such a judgment, it is necessary to determine a translated word from a pre-translation word (or conversely, it is necessary to determine a pre-translation word from a translated word), which can be done by determining corresponding words using databases such as those shown in Figures 3 and 10. Alternatively, a translation may be determined from a word by using machine translation or other dictionaries.
[0422] It is possible to check whether a code word such as "HSE" is included in the original and translated texts without using a database. In other words, it is possible to check the number of occurrences of each word (each character string) by finding three capital letters of the alphabet using regular expressions and counting the number of occurrences for each code word.
[0423] In this way, the result of translating a sentence written in a first language into a sentence written in a second language can be evaluated. That is, the sentence written in the first language contains a plurality of character strings. It is determined whether a character string corresponding to a character string contained in the sentence written in the first language is contained in the sentence written in the second language, and the determination result is displayed.
[0424] For example, in Japanese, the distinction between singular and plural forms of nouns is not clear, whereas in English, the distinction between singular and plural forms of nouns is clear. For this reason, even when a correct translation is performed, the singular and plural forms of nouns may differ between the original translation and the translation. Due to such differences in singular and plural forms, when a word (e.g., "apparatus") corresponding to a word (e.g., "apparatus") contained in a sentence written in a first language is correctly contained in a sentence written in a second language, it is undesirable that "apparatus" and "apparatuses" are determined to not match (a word is determined to be missing) when the word corresponding to "apparatus" is registered in the database.
[0425] Therefore, for example, plural words may be converted to singular (or vice versa), or singular and plural words may be treated as the same, thereby matching words contained in the pre-translation sentence with words contained in the translated sentence (determining whether they exist in both).
[0426] Similarly, when dealing with languages in which verbs change depending on the person of the subject, etc., when determining whether a verb word contained in a sentence written in a first language is contained in a sentence written in a second language, it is possible to match the words contained in the sentence before translation with the words contained in the translated sentence by, for example, converting the verb to its original form, or treating the verb as the same even if the word form changes.
[0427] As mentioned above, with current machine translation, some information in the original text may be lost (missing) after translation, or information not in the original text may appear in the translation (information leaks). Checking such information leaks or leaks by human eyes requires a great deal of effort (it is necessary to check both the original text and the machine translation).
[0428] When evaluating machine-translated text, the above-mentioned judgment is made for each unit that makes up the document (such as a sentence, section, phrase, paragraph, line, page, multiple sentences, multiple sections, multiple phrases, multiple paragraphs, multiple lines, multiple pages, or the entire document) and the judgment results are presented for each unit that makes up the document, allowing the user to see whether there are any missing words in the translation (loss of information) or unnecessary information has been added for that unit, and if so, which words. This reduces the effort required for users to check translations.
[0429] In other words, missing or added words can be shown to the user in association with the translation result of a single unit, so the user can easily determine whether or not to check that unit in detail. The user's attention is drawn only to unit parts that have missing translations (loss of information) or unnecessary added information (the user can check those unit parts in detail and revise the translation). For units for which no warning was issued, the user's effort in checking and revising is reduced.
[0430] Furthermore, when evaluating a translated text, a judgment may be made for the entire document, and the judgment result for the entire document may be presented. This makes it possible to judge whether there are any missing translations (loss of information) or unnecessary information has been added (evaluation of the translation) for the entire document. This judgment result can be used to evaluate the quality of the translation. For example, it becomes possible to evaluate machine translations, machine translation engines, and human translations.
[0431] For example, the process of evaluating each sentence (the unit of a sentence from the beginning to the end separated by a period (or a full stop, etc.)) can be performed as follows.
[0432] (1) The sentence sent in step S125 of FIG. 5 is compared and evaluated with the translation result of the sentence received in step S127. If there are any missing parts in the translation (loss of information) or any unnecessary information has been added, information about the missing character strings (codes, words, etc.) or added character strings (codes, words, etc.) (character strings and their number) is presented to the user.
[0433] (2) Prepare an original document with all translations completed and a document with its translation, read each document one sentence at a time (the unit of a sentence from the beginning of a sentence to the end separated by a full stop (or period, etc.)), compare and evaluate the two, and if there are any missing translations (loss of information) or unnecessary information added, present the user with information about those strings (strings and their number).
[0434] [Seventh embodiment]
[0435] The hardware configuration of the translation system in the seventh embodiment of the present invention is the same as that in the first to sixth embodiments, and therefore the description will not be repeated here. In the seventh embodiment, the translation system compares an original text and a translation of a sentence translated by the method according to the above-mentioned embodiment (or by another method), evaluates the translation result, and automatically corrects the translation.
[0436] A translation is evaluated based on whether information present in the original text is retained in the translation and / or whether information not present in the original text is present in the translation. The former judges whether there is any loss of information in the original text, while the latter judges whether unnecessary information has been added to the translation (so-called "information spillover").
[0437] For example, the text after the first encryption illustrated in the fifth embodiment is as follows:
[0438] "
[0023]
[0439] "HSD" has "HSE" and "HEF", and "HSE" is connected to "HEG" and "HEF". When the input of "HEG" is high, "HEF" displays a warning."
[0440] and when this sentence is translated by an external computer resource, the result is
[0441] "
[0023] .
[0442] "HSX" includes "HSE" and "HEF". "HSE" is connected to "HEG" and "HEF". When an input of "HEG" is high, "HEF" displays warning."
[0443] Here, the text after the first encryption contains one "HSD", two "HSE", three "HEF", and two "HEG".
[0444] In the translated text, one "HSD" is missing, and one "HSX" that was not in the original text has appeared (emerged). This is most likely due to the machine translation, which replaced the word that should have been processed with another word. There is also a one-to-one relationship between the number of missing words and the number of words that appeared (and also between the type of missing word and the type of word that appeared). This makes it possible to match the two. In this case, it is assumed that "HSD" in the document before translation has been replaced with "HSX" by machine translation, and a process is performed to replace "HSX" in the translated text with "HSD".
[0445] Similarly, for example, if two "HSD"s are missing during the translation process and two "HSX"s that were not in the original sentence appear, the number of missing words and the number of words that appear are one-to-one, and the types of missing words and the types of words that appear are also one-to-one, so the two can be matched up. Therefore, a process is performed to replace both "HSX"s in the translated sentence with "HSD".
[0446] In this way, when there is one type of missing word and the number of generated words is equal to the number of missing words, the generated words are replaced with the missing words, making it possible to provide an accurate translation (automatic correction of mistranslations in the translation).
[0447] For example, when there are two or more types of missing words, it is difficult to determine which missing word should replace the generated word, regardless of the number of generated words. In this case, automatic replacement may not be performed, and a warning may be output as in the fifth embodiment. Based on information about words around the generated word, it may be automatically determined which missing word should replace the generated word, and the replacement may be performed.
[0448] Also, when there is one type of missing word and the number of generated words is different from the number of missing words, a process of replacing the generated words with the missing words may be performed, or a warning may be issued. A warning (attention call) may be issued after the replacement process is performed.
[0449] In the above embodiment, an example was shown in which a generated word (code) for a code is replaced with a missing word (code). In a translation device that corrects a result of translating a text written in a first language into a text written in a second language, the text written in the first language may contain a plurality of character strings in the first language, and it may be determined whether a character string in the second language corresponding to a character string contained in the text written in the first language is contained in the text written in the second language, and the text written in the second language may be corrected (the generated word may be replaced with a missing word) based on the determination result.
[0450] The present invention is not limited to evaluation of the results of machine translation and correction of translated text, but can also be applied to evaluation of the results of translation performed by a human being and correction of translated text.
[0451] In the above embodiment, the translation may be performed using an external server, or may be performed within a stand-alone PC or within a PC in a LAN.
[0452] The above-described multiple embodiments and the elements contained therein (part of the configurations, part of the processes) can be combined or replaced to create new embodiments.
[0453] The processes in the above-described embodiments may be performed by software or by using a hardware circuit. Also, a program for executing the processes in the above-described embodiments may be provided, or the program may be recorded on a recording medium such as a CD-ROM, a flexible disk, a hard disk, a ROM, a RAM, or a memory card and provided to a user. The program is executed by a computer such as a CPU. Also, the program may be downloaded to the device via a communication line such as the Internet.
[0454] The above-described embodiments should be considered as illustrative and not restrictive in all respects. The scope of the present invention is defined by the claims, not the above description, and is intended to include all modifications within the meaning and scope of the claims. [Explanation of symbols]
[0455] 100 Computers 101 CPU 103 Communications Department 109 RAM 111 Storage device 200 Other company's computer resources 203 Search and Replace 205 Document Editing and Word Registration Section 207 Database 209 Memory 211 Display / Output section 213 Input section 400 Internet 500 in-house computer resources
Claims
1. A translation device that corrects a result of translating a text written in a first language into a text written in a second language, comprising: the text written in the first language includes a plurality of character strings; a determining means for determining whether unnecessary information has been added to the translation by determining whether the sentence written in the second language contains a character string corresponding to a character string contained in the sentence written in the first language, and whether the sentence written in the second language contains a character string that does not correspond to a character string contained in the sentence written in the first language; and a correction means for correcting the sentence written in the second language by replacing the generated word with a character string corresponding to the missing word based on the determination result of the determination means.
2. 2. The translation device according to claim 1, wherein the text written in the second language is obtained by machine translating the text written in the first language.
3. 3. The translation device according to claim 1, wherein the character string is either a character string indicating a word, a character string indicating a word and a reference sign immediately following it, or a cryptographic character string, or a character string including either a character string indicating a word, a character string indicating a word and a reference sign immediately following it, or a cryptographic character string.
4. acquiring a sentence written in the second language, which is a machine translation result of a sentence in which a character string indicating a certain word or a character string including a character string indicating a certain word in the sentence written in the first language is replaced with another character string corresponding to the character string; 4. The translation device according to claim 1, wherein the determining means determines whether the replaced character string is included in a sentence written in the second language.
5. The determination means performs a determination for each unit constituting the document, 5. The translation apparatus according to claim 1, wherein said correction means corrects each of the units constituting the document.
6. The determination means performs a determination on the entire document, 6. The translation apparatus according to claim 1, wherein said correction means corrects an entire document.
7. 7. The translation device according to claim 1, wherein the determination result includes at least one of a character string not included in the sentence written in the second language and a number of character strings not included in the sentence written in the second language.
8. 8. The translation device according to claim 1, wherein the determining means determines whether a character string corresponding to a character string included in the sentence written in the second language is included in the sentence written in the first language.
9. A control program for a translation device for correcting a result of translating a text written in a first language into a text written in a second language, comprising: The translation device includes a computer. the text written in the first language includes a plurality of character strings; The control program for the translation device includes: a determining step of determining whether unnecessary information has been added to the translation by determining whether the sentence written in the second language contains a character string corresponding to a character string contained in the sentence written in the first language, and whether the sentence written in the second language contains a character string that does not correspond to a character string contained in the sentence written in the first language; A control program for a translation device that causes a computer to execute a correction step of correcting the sentence written in the second language by replacing the generated word with a character string corresponding to the missing word based on the judgment result of the judgment step.
10. A translation evaluation method using a translation device that corrects a result of translating a text written in a first language into a text written in a second language, comprising: The translation device includes a computer. the text written in the first language includes a plurality of character strings; The translation evaluation method includes: a determining step of determining whether unnecessary information has been added to the translation by determining whether the sentence written in the second language contains a character string corresponding to a character string contained in the sentence written in the first language, and whether the sentence written in the second language contains a character string that does not correspond to a character string contained in the sentence written in the first language; and a correction step of correcting the sentence written in the second language by replacing the generated word with a character string corresponding to the missing word based on the judgment result of the judgment step.
Citation Information
Patent Citations
Translation aid method and system
JP2005071017A
Translation request method, translation request terminal and program
JP2006244029A
Machine translation device, and program
JP2007094626A
Translation device, translation model learning device, method, and program
JP2015121992A
Translation method, translation apparatus, and translation system
JP2018120584A
