Processing unit

Abstract

To provide a processing device that can improve the quality of translations. [Solution] The processing device comprises an input means for inputting the original text and its translation, an acquisition means for checking the translation against the original text using a large-scale language model and obtaining the corrected translation if corrections are needed, and an output means for outputting the corrected translation. The output means outputs the difference between the input translation and the corrected translation, either character-level or word-level.

Description

Technical Field

[0001] This invention relates to a processing device.

Background Art

[0002] Conventionally, there has been a translation system that automatically translates text in a first language into text in a second language.

[0003] For example, Patent Document 1 below discloses a translation system, a translation device, and a program that perform translation considering the characteristics of the text to be translated.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] An object of this invention is to provide a processing device capable of improving the quality of translation.

Means for Solving the Problems

[0006] The present invention has been made to solve the above problems, and provides a processing device described in the claims attached hereto.

Modes for Carrying Out the Invention

[0007] Examples of the present invention will be described below. Needless to say, this example is merely an example and can be changed into various forms without departing from the gist of the present invention. [Example] 1. Outline of System Configuration The translation checking system according to this embodiment (hereinafter referred to as "this system") is a processing device that checks the source text and its translation input by the user using a Large Language Model (hereinafter referred to as "LLM"), obtains the corrected translation as necessary, and presents the difference to the user. This system consists of multiple modules (means) as shown below. Each module will be described below. 1. Input method - The system provides an interface for users to upload or manually input the source text to be translated and the corresponding translated text. Specifically, this includes forms that operate on a web browser and text input fields in desktop applications. - The source text and translated text can be entered via text files, Word files, or a dedicated input screen. Users can input the source text and its translation by associating each part (such as a segment) of the translated text, or they can input the entire text and its translation together. 2. Acquisition method - This system works in conjunction with LLM and has a function to determine whether "translation checking" and "correction" are necessary based on the content of the input source text and the translated text. - Inquiries to the LLM are made with the original text and the translated text together. The LLM assesses whether the translated text is faithful to the meaning of the original text in terms of grammar, terminology, nuance, and style. If necessary, it generates a more natural and accurate translation and returns it as the "revised translation." Specifically, a request such as "Please compare the following source text with the translated text and make corrections if there are any mistranslations or stylistic issues in the translated text" is sent as a prompt to the LLM. The system receives the correction results returned by the LLM and retrieves the corrected translated text. If no corrections are needed, a response of "No corrections" is received, and the translated text is used as is. 3. Output means - This system outputs the differences between the input translation and the revised translation, character by character or word by word. The difference output includes visually clear highlighting so that users can see at a glance where and what kind of changes have been made. - Furthermore, this embodiment also includes a function to output the differences in the modified parts as an HTML document. In the HTML-formatted difference report, the text before and after the modification is displayed using color coding such as red or green, and it is also possible to implement a feature that allows users to refer to translation intent and additional explanations by hovering the mouse over the text. - In addition, detailed results of the LLM check (for example, comments such as "there is a tense error" or "some nuances are missing") can sometimes be obtained at the same time, and this system provides these check results to the user so that they can better understand the correction process and improvements.

[0008] This configuration offers users the advantage of being able to perform simple and quick translation checks during the translation process, revise the translated text as needed, and easily compare and verify the results. 2. Operation Details 1. Input the original text and the translated text. - The user opens the system's input screen and enters the original text and its translation. For example, the following UI is envisioned. - Original text input field (multi-line text field) on the left side - Translation input field (multi-line text field) on the right side - The entered text is sent to the system's server and temporarily stored in a database or memory.

[0009] Alternatively, data (e.g., JSON data) containing the original text divided into predetermined units and their corresponding translations may be input. Batch processing may also be performed by sequentially inputting the data containing the original text divided into predetermined units and their corresponding translations.

[0010] Alternatively, you can input the entire source text and the entire translated text, automatically divide them into predetermined units, and perform batch processing on each unit. 2. Translation Check and Acquisition of Correction Contents - The system sends a pair of the original text and the translated text (which may be the whole pair or a partial pair for the whole text) to the LLM and instructs it to comprehensively check accuracy, naturalness, style, etc. - The LLM compares with the context, meaning, and style of the original text and presents appropriate amendments if there are translation errors in the translated text. Additionally, it may generate "comments (check contents)" as needed. - The system analyzes the "corrected translation text" received from the LLM and extracts the difference from the input translated text. This difference extraction is performed at the character level or word level. General text difference algorithms (e.g., diff algorithms based on Longest Common Subsequence (LCS)) can be used for difference extraction. - If no correction is made at all, obtain the "no correction required" status from the LLM output and handle it as a state with no difference. 3. Display of Differences (Including HTML Format) - Differences are generated as an HTML document as in the following example. ```html Original text: This is a sample sentence. Original translation: This is a sample sentence. Revised translation: This is a sample sentence. Difference: this is This is Sample The text sentence It is. ``` - As described above, strikethroughs and color-coding can be used to explicitly show the changed points. This allows the user to compare the text before and after correction and visually grasp what replacements, additions, or deletions were made. Note that as long as the differences are understandable, it is not limited to HTML documents, and text documents or Markdown documents may also be used. - On the user interface of the system, since HTML documents can be directly displayed in a web browser, the user can view the correction contents without using additional software. 4. Display of Check Results (Optional) - In this system, it is also possible to obtain the comments and evaluation points generated by the LLM as "check results" and present them to the user. For example, the following information can be displayed. - Comparison with translation examples in similar past contexts - Basis for the section deemed to be a mistranslation - Evaluation regarding stylistic consistency - Possible omissions or missing parts in the translation - By presenting these check results to the user through an output method, the user can understand not only the corrections made but also the reasons for those corrections, thereby obtaining information that contributes to improving translation quality. When checking each part of a document, it is desirable to send information about previously processed parts along with the part being checked, along with prompts, in order to check consistency with preceding parts. Sending information about all previously processed parts would increase the amount of data too much, so it may be acceptable to send information about the past few processed parts (1 to n times, an arbitrary number depending on the system), along with prompts, to the LLM. [Specific example] - "Input means for inputting the original text and its translation" refers to the user's input screen or input form in this embodiment. Uploading a file may also be used as an input means. Input may be done in parts. - The "acquisition means that uses a large-scale language model to check the translated text against the source text and obtain the corrected translated text if corrections are needed" corresponds to a processing module that sends the source text and translated text to the LLM and receives the proposed corrections. The text may be sent in parts. Alternatively, processing may be performed internally by the LLM instead of sending the text. - The "output means for outputting the corrected translated text" corresponds to the translation result output function, including the difference display function of this embodiment. Furthermore, as stated, "the output means outputs the character-level or word-level difference between the input translated text and the modified translated text," this embodiment displays the character-level or word-level difference using a difference extraction algorithm. The difference display is preferably at the character-level or word-level, but it may also be at the line-level. - The statement says, "The output means outputs the modified parts between the input translated text and the modified translated text as an HTML document." However, in this embodiment, an HTML document showing the differences is generated, providing a mechanism that allows the user to easily view it on a web browser. Specifically, this refers to a function that uses HTML tags to highlight areas that have been modified, such as by using color coding or strikethrough. - The documentation states that "the acquisition means further acquires the results of checking the translated text, and the output means further outputs the results of checking," but in this embodiment, comments and evaluation information (justifications and advice for corrections) provided by the LLM are acquired as "check results" and made available for display to the user. Therefore, users can understand not only the changes made, but also the reasons behind them, providing information that contributes to improving translation quality. 4. Effects According to this embodiment, the following effects can be obtained. 1. Improving translation work efficiency Traditionally, checking and correcting translated texts was often done manually by experts, which was time-consuming. In this embodiment, correction suggestions can be automatically generated using LLM, significantly improving the efficiency of the translation process. 2. Clarification of the revisions - The difference output function allows users to easily see which words and characters were changed before and after the revision. The HTML format difference report, in particular, is highly visual and makes it easy for users to perform additional checks and fine-tuning. 3. Improving translation quality - By providing the check results simultaneously, users can understand why the corrected translation is more appropriate. This can also lead to a learning effect, allowing users to prevent similar errors in the future. 4. Flexible expandability - This system utilizes LLM as an external service, allowing for easy LLM version upgrades and switching to new language models, thus enabling flexible adaptation to advancements in the translation field. Of course, processing can also be done using the internal LLM instead of an external one. - Furthermore, the output method can be extended to display differences and generate reports in formats other than HTML (such as PDF, JSON, and Word documents). The differences could also be output as a change history in a word processing program such as Word. 5. Summary This embodiment relates to a processing device that checks the source text and translated text entered by the user using LLM, generates a revised translated text as necessary, and clearly presents the differences in a format such as HTML. In addition to satisfying the constituent requirements shown in the claims, it has the excellent effect of contributing to the efficiency and quality improvement of translation work.

[0011] The present invention is not limited to these embodiments, and those skilled in the art can make various changes and modifications within the scope of the technical concept disclosed herein. [others] Translation checking can be done sentence by sentence, or in larger units such as paragraphs, chapters (or even entire documents). This allows for flexible operation depending on the user's work style and the length and format of the translated text. [Other examples] The following describes other embodiments using machine translation (including translation by neural networks and translation by LLM). These can be applied by further incorporating them into the embodiments described above, and various modifications can be made without departing from the spirit of the present invention. 1. System Configuration Overview In this example, the user only inputs the source text (text to be translated), the system automatically performs machine translation, and then uses a large-scale language model (LLM) to check the machine translation result, and retrieves and presents the corrected translation as needed.

[0012] Specifically, the process proceeds as follows: 1. Input of the original text - The user inputs the source text to be translated into this system (for example, the source text when translating from English to Japanese, Japanese to English, etc. The supported languages ​​are optional). - The input method is the same as in previous examples, such as uploading a text file, directly entering text into a text field in a browser, or copying and pasting from another application. As with the embodiments described above, you can enter the entire text or enter it in parts. 2. Perform machine translation - Upon receiving user input, the system invokes its internally built-in machine translation engine (or a machine translation service accessed via an external API) to automatically generate a translated text from the source text. As with the embodiments described above, the entire text may be processed at once, or it may be processed in batches, part by part. - While machine translation engines often use, for example, neural machine translation (NMT) based engines, this invention does not limit the type of engine. - While the machine translation results obtained at this stage generally possess a certain level of quality, final adjustments by human experts or LLMs may be desirable in terms of contextual understanding and expressiveness. 3. Translation check by LLM - This system sends the correspondence between the machine-translated text and the original text to the LLM for review and necessary corrections (as described in the embodiment above). - LLM analyzes the context and meaning of the original text, reads the machine-translated text, detects mistranslations, stylistic inconsistencies, and awkward expressions, and generates suggested revisions as needed. - At this stage, the LLM will improve the translation from multiple perspectives, including paraphrasing, adding nuances, and standardizing domain-specific terminology. In some cases, they may determine that there are "no particular issues (no corrections needed)" and respond accordingly. 4. Obtaining the revised translation and displaying the differences. - The system retrieves the "corrected translation" returned from LLM and extracts the difference between it and the translation generated by the machine translation engine. - Differences are extracted character by character or word by word, using common difference comparison algorithms (such as those based on Longest Common Subsequence (LCS)). - The extracted differences are presented to the user in an easy-to-understand format, such as HTML or text. This allows them to visually understand what has been corrected or improved from the machine translation and how. The reasons why a format other than HTML is acceptable are as described above. 5. Displaying check results (optional) - Furthermore, if the LLM provides feedback regarding translation deficiencies or reasons for improvement, the results of that check can also be presented to the user. - For example, by displaying comments such as "There is a mix of polite and informal language" or "The idioms in the original text have been translated literally, but there are more natural expressions," users can not only receive the corrected results but also gain a deeper understanding of the background and rationale behind them. 2. Advantages of this embodiment 1. Efficient translation workflow - In this embodiment, machine translation and LLM checking are performed consistently within the system, significantly reducing the amount of work required. 2. Improving translation quality - While machine translation engines are well-suited for fast and large-volume translations, they can have limitations in capturing subtle nuances and stylistic adjustments. In contrast, LLM (Language Literacy) can perform additional checks and corrections on machine translation results to produce more natural and accurate translations. 3. Transparency through differential display - By visualizing the differences between the machine translation result and the corrected translation from LLM, users can easily track "what parts were changed and to what extent." - This high level of transparency allows users to review the revisions proposed by the LLM and make final decisions regarding the translation. 4. Learning effect of check results - Because users can refer to comments and ratings from LLM (Learning Leadership Module), it is easier to identify common mistranslation patterns and stylistic errors in machine translation. This facilitates learning of points to be careful about and areas for improvement in future translation work. 5. Scalability and Versatility - This implementation can be handled using essentially the same processing flow regardless of the machine translation engine or service used. The LLM checking portion is also independent of any specific vendor or model, and translation quality can be continuously improved by switching to the latest high-performance LLM. 3. Combination with other embodiments - Check at the segment level As previously mentioned, translation checks can be performed at various levels of granularity, such as sentence by sentence, paragraph by paragraph, or chapter by chapter (or even document by document). The same applies to this additional embodiment, allowing for flexible configuration to suit the user's operational style, such as checking machine translation output sentence by sentence or correcting paragraphs as a whole. Machine translation can also be performed at various levels of granularity, such as sentence by sentence, paragraph by paragraph, or chapter by chapter (or even document by document). - Displaying differences using HTML documents As mentioned earlier, differences can be visualized using HTML color coding, etc. This works similarly with machine translation, offering the advantage of intuitively understanding the differences between the translation generated by the machine translation engine and the translation corrected by LLM. - Additional output of check results Machine translation is particularly prone to issues with handling technical terms and losing context. This system provides an interface that allows users to easily review comments from the LLM (Language Licensing Manager) highlighting such problems. This enables users to ultimately consider more appropriate terminology and consistent writing style. 4. Summary In this embodiment, the result of machine translation is used as the initial translation, which is then re-checked by the LLM to present the final revised translation. - User input of original text → Batch translation by machine translation engine → Check and correction by LLM → Display of differences and evaluation comments This process can be automated, and users can review and evaluate the changes, resulting in an efficient and high-quality translation workflow.

[0013] Note that translation checks and revisions may be performed more than once. Furthermore, these checks may be performed using different LLMs (Language-Language Models) or the same LLM.

[0014] When translation is performed by an LLM, the LLM who performs the translation and the LLM who checks the translation may be the same person or different people. Using different LLMs allows for the absorption of the individual characteristics of each LLM, making it possible to provide a more average translation.

[0015] For translation, any method may be used, including neural network translation, machine translation, or LLM-based translation.

[0016] As described above, this system combines the characteristics of machine translation and LLM to enable the translation of large amounts of text in a short time while improving the naturalness and accuracy of the final expression. This embodiment is extremely useful because users can simultaneously enjoy the advantages of machine translation (speed and large-capacity processing) and the advantages of LLM proofreading (contextual understanding and correction of subtle nuances).

[0017] Additionally, the output may include the original text, the translated text, the revised translated text, the differences, and check comments (the processing results for the entire original text may be output at once, or the processing results for parts of the text may be displayed repeatedly to show the overall results).

[0018] Alternatively, you may choose to output only the corrected translation (you may output only the corrected translation as a single file). If you have performed translation checks on a sentence-by-sentence, paragraph-by-paragraph, or chapter-by-chapter basis, you may choose to output only the corrected translations together (you may output them as a single file). [Regarding the usage of LLM] The LLM (and / or machine translation engine) used in this system may utilize an external service (LLM on the cloud) via an API, or it may utilize a local LLM installed in an on-premises environment. The advantages and effects of each configuration are illustrated below. 1. When using an external LLM API - Access to the latest models: External services often constantly update their models, making it easy to use the latest high-performance LLMs. - Reduced maintenance burden: Since the service provider handles model tuning and version control, users can significantly reduce their infrastructure management burden. - Scalability: It is easy to scale out on the cloud, making it easy to create an environment that can handle a large volume of requests. 2. When using LLM installed locally - Security & Privacy: There is no need to send highly confidential documents or personal information externally, reducing the risk of data breaches and information control issues. - Customizability: By having our own models trained on specific domains and company-specific terminology, we can perform more optimized translation checks. - Offline use: It can be operated even in environments with limited internet access, and can be used stably during outages or while on extended business trips. [Various variations of difference display] This system can utilize a variety of methods to help users intuitively understand the differences (exclusions) between the input translation and the corrected translation. 1. Text color coding - Red and green highlighting: Deleted parts are displayed in red, and added parts in green. - Background color change: A light red background will be applied to the deleted portion, and a light green background to the added portion. 2. Adding markup (symbols) - Strikethrough and underline: Deleted parts are indicated with strikethrough (~~text~~), added parts with underline ( text This is shown by ). - Inserting special characters: For example, insert arrow symbols such as "←" or "→" before and after the deleted part, or enclose it like "[-Deleted part-]" or "{+Added part+}". 3. Side-by-side (comparison) display - This format displays the original text on the left and the revised text on the right, side by side. It highlights the changes while making it easy to get an overview of the entire document. 4. Difference table with row numbers - Similar to comparing program source code, it employs a table-style UI with line numbers assigned to each line, indicating changes for each line. Changes are indicated by icons or colored backgrounds. 5. Balloons and pop-ups - When you hover your cursor over a section with a difference (mouse over it), a pop-up window or balloon display will show "Before Correction → After Correction." This allows you to check the differences without interrupting the flow of the text. 6. Change tracking style (word processor style) - Similar to the "Track Changes" feature in Word, deletes will be displayed with a strikethrough, while added sections will be inserted in a different color. A "List of Changes" will be displayed at the top of the page or in the side panel for easy review. 7. Collapse display - To make documents with many revisions easier to read, a collapsible interface will be provided. Normally, the full text will be displayed, and the "Diff Details" can be expanded or collapsed with a click or tap. 8. Emphasis through animation - Implement an animation that highlights the changes in order, allowing users to view the modifications sequentially without manual intervention. 9. Accessibility features such as audio and Braille - It is also possible to combine accessibility-conscious methods, such as using speech synthesis to read out the differences, or displaying the text before and after the correction on a Braille display.

[0019] As described above, differences can be represented not only through simple color coding but also in various other ways, making it possible to build an optimal interface tailored to the user's needs and environment. [Example Hardware Configuration] To realize this system, the following hardware configuration can be exemplified. Note that this is merely an example, and in implementing the present invention, it is possible to add or omit other components, or to integrate each component. 1. Server (Main Processing Unit) -CPU(Central Processing Unit) - Equipped with a multi-core CPU capable of high-speed computation, it is responsible for processing data related to translation checking and coordinating with other modules. - GPU(Graphics Processing Unit) / TPU(Tensor Processing Unit) - Secure the computational resources required when running large-scale language models locally, thereby accelerating deep learning and inference processes. - Main memory (RAM) - Ensure sufficient workspace is available throughout the entire system. Install ample capacity to accommodate situations where large amounts of data need to be temporarily stored when using LLM. - Storage (HDD / SSD) - Records user input text, translation results, difference information, check results, etc. Processing efficiency improves by enabling high-speed reading and writing using large-capacity SSDs, etc. - Network Interface - When using external LLM APIs or machine translation services, an internet connection is required. Even when linking multiple devices on an internal network, a gigabit-class or higher communication bandwidth is ensured. 2. User terminal - A device (such as a PC, tablet, or smartphone) used by users to operate this system. - Users input source and translated text, view and edit results, and check differences through a browser or dedicated application. - Since the majority of processing is assumed to be done on the server side, a user terminal with standard performance is sufficient. However, a terminal with a large screen and high-speed network connection may be desirable in some cases to improve efficiency for large amounts of text input and verification. 3. Accessories and peripherals (optional) - Display devices / projectors - Used when multiple people need to share and review translation results and differences in an office or meeting room. - Input devices such as keyboards, mice, and touch panels - Used by users to correct and edit translated text. On touchscreen devices, pen input is also possible, allowing users to add handwritten text and annotations. - Audio input / output device - Used when performing operations via text-to-speech or voice input, aiming to improve accessibility. 4. Operation form Any of the following environments is acceptable. - On-premises environment - To prioritize security, all servers are located in the company's internal data center or on a closed network. - Cloud environment - This system is built on virtual servers (virtual machines) or containers, making it easy to scale out and add resources flexibly. - Hybrid configuration - It's also possible to use a hybrid approach, such as keeping sensitive data and local LLM inference environments on-premises while using external LLM APIs and machine translation services for supplementary purposes.

[0020] By combining the hardware configurations described above, this system can ensure high processing performance and flexible expandability while enabling operation tailored to the user's usage patterns.

[0021] The processing in the above-described embodiment may be performed by software or by hardware circuitry. Furthermore, a program for executing the processing in the above-described embodiment can be provided, or the program may be recorded on a recording medium such as a CD-ROM, flexible disk, hard disk, ROM, RAM, or memory card and provided to the user. The program is executed by a computer such as a CPU. Alternatively, the program may be downloaded to the device via a communication line such as the Internet.

[0022] The embodiments described above should be considered in all respects to be illustrative and not restrictive. The scope of the present invention is indicated by the claims rather than by the foregoing description, and all modifications within the meaning and scope equivalent to the claims are intended to be included.

Claims

1. An input method for entering the original text and its translation, An acquisition means that uses a large-scale language model to check the translated text against the source text and, if corrections are needed, obtains the corrected translated text. The system includes an output means for outputting the modified translation, The output means is a processing device that outputs the difference between the input translated text and the modified translated text, on a character-by-character or word-by-word basis.

2. The processing apparatus according to claim 1, wherein the output means outputs the modified parts of the input translated text and the modified translated text as an HTML document.

3. The processing apparatus according to claim 1 or 2, wherein the acquisition means further acquires the result of checking the translated text, and the output means further outputs the result of checking the text.