Information processing system, information processing method, and program

Abstract

Provided is an information processing system for assisting in translation in which consistency with a past document and a translated document thereof is ensured. This information processing system comprises: a search means for searching for a reference document corresponding to a document to be translated; a reference information generation means for generating reference information including the reference document and a translation thereof; and a translation generation means for generating a translation of the document to be translated through machine translation on the basis of instruction information including the reference information. The information processing system also comprises a vectorization means for vectorizing the document to be translated and the reference document. The search means searches for the reference document on the basis of the degree of similarity between the vectorized documents.

Description

Information Processing System, Information Processing Method, and Program

[0007]

[0001] The present invention relates to an information processing system, an information processing method, and a program.

[0002] Conventionally, in machine translation systems, there are technologies that have been widely developed for the purpose of improving the efficiency of document translation work (see, for example, Patent Document 1).

[0003] Japanese Patent Application Laid-Open No. 7-210558

[0004] However, in recent years, there has been a demand to solve problems such as the difficulty of ensuring the consistency of terms and expressions between documents created in the past and their translated documents. However, in the conventional technologies including the technology of Patent Document 1, such demands cannot be sufficiently met. ​​​​​​​​​​​​​This figure shows an overview of the service that can be realized by an information processing system to which a server according to one embodiment of the information processing system of the present invention is applied. This figure shows an example of the configuration of an information processing system to which a server according to one embodiment of the information processing system of the present invention is applied. This is a block diagram showing an example of the hardware configuration of the server in the information processing system of Figure 2. This is a functional block diagram showing an example of the functional configuration of the server in Figure 3 that constitutes the information processing system of Figure 2. This figure shows a translation processing flow according to one embodiment of the present invention. This figure shows a document structuring process according to one embodiment of the present invention. This figure shows a similarity calculation processing flow according to one embodiment of the present invention. This figure shows the main screen of a translation support tool according to one embodiment of the present invention. This figure shows an example of a document to be translated according to one embodiment of the present invention. This figure shows a vectorization result according to one embodiment of the present invention. This figure shows a first example of similarity search results according to one embodiment of the present invention. This figure shows a second example of similarity search results according to one embodiment of the present invention. This figure shows an example of prompt generation according to one embodiment of the present invention. This figure shows a first example of a translation result according to one embodiment of the present invention. This figure shows a second example of a translation result according to one embodiment of the present invention.

[0010] Embodiments of the present invention will be described below with reference to the drawings.

[0011] First, with reference to Figure 1, an overview of the services (hereinafter referred to as "the Services") that can be realized by an information processing system (see Figure 2, described later) to which a server according to one embodiment of the information processing device of the present invention is applied will be described. In this specification, "paragraph number" refers to identification numbers such as

[0012] ,

[0013] , etc., assigned to each paragraph in a patent application document, and "structured database" refers to a database that stores paragraph content and its translation in association with the paragraph number, using the paragraph number as an index.

[0012] Figure 1 shows an overview of the service that can be realized by an information processing system to which a server according to one embodiment of the information processing device of the present invention is applied. The service is a translation system that ensures consistency between past documents and their translated documents. The service utilizes a predetermined artificial intelligence model (large-scale language model, LLM, hereinafter referred to as "external AI3").

[0013] Specifically, as shown in Figure 1, when user A sends a patent application document JD to be translated to server 1, the vectorization unit (corresponding to the vectorization unit 102 in Figure 4, described later) within server 1 vectorizes the document, and the search unit (corresponding to the search unit 103 in Figure 4, described later) extracts similar reference documents RD from the document database. Next, the context generation unit (corresponding to the context generation unit 104 in Figure 4, described later) integrates the extracted reference documents RD and their translations to generate context information, and the translation generation unit (corresponding to the translation generation unit 105 in Figure 4, described later) provides this context information to the external AI 3 to generate the translated document ED.

[0014] More specifically, Server 1 uses a patent-specific embedding model to perform vectorization at the document, paragraph, and sentence levels, and employs a technology-adaptive vectorization function to perform vectorization tailored to each technology field, for example, based on the IPC code (International Patent Classification). In the search process, a multidimensional similarity calculation function calculates and integrates similarity across multiple dimensions, including not only simple cosine similarity, but also the degree of agreement of technical terms, similarity of sentence structure, contextual role, and paragraph length.

[0015] In the context generation process, paragraph-by-paragraph correspondence is performed using parallel text pair extraction technology, and the extracted paragraphs are classified into types such as "claims," ​​"background of the invention," and "examples" using the paragraph type classification function. In addition, specialized terminology correspondence is automatically extracted from similar documents using the parallel text table generation function, and a dictionary for each technical field is constructed. In the translation generation process, a multi-stage translation process (e.g., initial translation, terminology standardization, syntax adjustment, final finishing, etc.) is employed, and the reliability of the output of the external AI3 is evaluated using a reliability scoring function.

[0016] The generated translated document ED is sent to the user terminal 2, where it is reviewed by experts and then added to the database by the learning unit (corresponding to 107 in Figure 4, described later). In other words, the learning unit receives a training file that has been reviewed by experts and has feedback, and saves it to the database. The learning unit then updates the instruction indicators (e.g., prompts described later) that the AI ​​should use to generate translations, and the translation table that the AI ​​should refer to, based on the training file. This update itself can also be performed by the learning unit on an external AI 3. This system enables consistent, high-quality translations that refer to past translation examples.

[0017] In this way, it is possible to create translations that ensure consistency with past documents and their translations.

[0018] Next, with reference to Figure 2, the configuration of an information processing system to which a server according to one embodiment of the information processing device of the present invention is applied will be described.

[0019] The information processing system shown in Figure 2 is configured to include a server 1, a user terminal 2, and an external AI 3. The server 1, user terminal 2, and external AI 3 are interconnected via a network N such as the Internet.

[0020] Server 1 is an information processing device managed by the service provider of this service (Figure 1). Server 1 performs various processes to realize this service while communicating with the user terminal 2 and the external AI 3 as appropriate. The user terminal 2 is an information processing device operated by the user, and consists of a smartphone, tablet, personal computer, etc.

[0021] Figure 3 is a block diagram showing an example of the server hardware configuration in the information processing system shown in Figure 2.

[0022] Server 1 comprises a CPU (Central Processing Unit) 11, a ROM (Read Only Memory) 12, a RAM (Random Access Memory) 13, a bus 14, an input / output interface 15, an input unit 16, an output unit 17, a storage unit 18, a communication unit 19, and a drive 20.

[0023] The CPU 11 executes various processes according to the program recorded in the ROM 12 or the program loaded into the RAM 13 from the storage unit 18. The RAM 13 also stores data and other information necessary for the CPU 11 to execute various processes.

[0024] The CPU 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output interface 15 is also connected to this bus 14. An input / output interface 15 is connected to an input unit 16, an output unit 17, a storage unit 18, a communication unit 19, and a drive 20.

[0025] The input unit 16 is composed of, for example, a keyboard, and takes in various types of information. The output unit 17 is composed of, for example, a display such as an LCD or a speaker, and outputs various types of information as images or sounds. The storage unit 18 is composed of, for example, DRAM (Dynamic Random Access Memory), and stores various types of data. The communication unit 19 communicates with other devices (for example, terminals 2-1, 2-2, and 2-N in Figure 2) via a network N including the Internet.

[0026] A removable media 21, such as a magnetic disk, optical disk, magneto-optical disk, or semiconductor memory, is appropriately mounted on the drive 20. Programs read from the removable media 21 by the drive 20 are installed in the storage unit 18 as needed. The removable media 21 can also store various data stored in the storage unit 18, just like the storage unit 18.

[0027] Although not shown in the diagram, the user terminal 2 and external AI in Figure 2 can have essentially the same hardware configuration as shown in Figure 3. Therefore, a description of the hardware configuration of the user terminal 2 and external AI 3 will be omitted.

[0028] Through the cooperation of various hardware and software components that make up the information processing system in Figure 2, including Server 1 in Figure 3, various processes for providing the service in Figure 1 can be executed.

[0029] Figure 4 is a functional block diagram showing an example of the functional configuration of the server in Figure 3 within the information processing system shown in Figure 2.

[0030] As shown in Figure 4, the CPU 11 of Server 1 functions as follows: Translation Unit 100, Translation Target Reception Unit 101, Vectorization Unit 102, Search Unit 103, Context Generation Unit 104, Translated Text Generation Unit 105, Bilingual Table Generation Unit 106, Learning Unit 107, Feedback Reception Unit 108, Context Correction Unit 109, and Bilingual Table Correction Unit 110. In addition, a document database (Document DB 71) and a translation database (Translation DB 72) are provided in one area of ​​the storage unit 18 of Server 1. Each functional unit from Translation Target Reception Unit 101 to Bilingual Table Generation Unit 106 is integrated as Translation Unit 100. Each functional unit of Feedback Reception Unit 108, Context Correction Unit 109, and Bilingual Table Correction Unit 110 is integrated as Learning Unit 107.

[0031] In the CPU 11 of Server 1, in addition to the above-mentioned functional units, the functional units from the prompt optimization unit 111 to the report generation unit 141 also function. Details of these functional units will be described later.

[0032] The translation target reception unit 101 receives documents to be translated. The received documents to be translated are stored in the document database 71. The translation target reception unit 101 also receives reference documents and their translations. The reference documents and their translations are stored in the translation DB 72. The vectorization unit 102 vectorizes the documents to be translated and the reference documents. By using a patent-specific embedding model and a model that has undergone additional pre-training with a large number of patent documents based on a general-purpose language model, vectorization that is robust to patent-specific phrasing and technical terms is achieved. In addition, a technology field-adaptive vectorization function applies a vectorization model specialized for each technology field, for example, based on IPC codes. The vectorized documents are stored in the translation DB 72. This enables efficient similar document searching.

[0033] The search unit 103 searches for reference documents corresponding to the document to be translated. Using a multidimensional similarity calculation function, it calculates and integrates similarities in multiple dimensions, including not only simple cosine similarity, but also the degree of agreement of technical terms, similarity of sentence structure, and contextual role. Furthermore, a weighted technical term matching function assigns higher weights to technical terms within the text, enabling similarity calculations that prioritize the agreement of technical terms over general expressions. The retrieved reference document information is extracted from the translation DB 72. This ensures consistency with past translation examples. Additionally, past translation examples include not only machine translations but also translations by experienced human translators, reflecting the unique translation styles of patent translation.

[0034] It should be noted that vectorization is not always necessary when searching for similar paragraphs. Normalized text can be compared, and if an exact match is found, a similarity score of 1.0 is immediately assigned. If an exact match is not found, the text can be divided into sentences, and the similarity score for each sentence can be calculated using a text comparison library. In this process, the ratio of sentences should also be considered when calculating the final similarity score.

[0035] The context generation unit 104 generates reference information including a reference document and its translation. Paragraph-by-paragraph correspondence is performed using parallel text pair extraction technology, and the extracted paragraphs are classified into types such as "claims," ​​"background of the invention," and "examples" in the case of a patent application document. The context may be generated so that multiple paragraphs are translated together based on type or reference number. This ensures that the translation form, such as active or passive voice, is consistent among paragraphs that are translated together. Furthermore, the prompt configuration function configures prompts that include not only the sentence to be translated but also the contextual information before and after it. The generated reference information is stored in the translation DB 72. This provides appropriate instruction information to the translation AI.

[0036] The translation generation unit 105 generates a translated text by machine translation based on instruction information including reference information. In one embodiment of the present invention, a database contains a bilingual table for standardizing terminology and examples of expression formats specific to patent documents. A multi-stage translation process is employed in which the user refers to these, specifies the text via prompts, and then has an external AI 3 perform the translation.

[0037] Specifically, the translation is performed in the following four stages: (1) Initial translation stage: A prompt including the document to be translated, similar reference documents, translations of similar reference documents, a bilingual table, and examples of expressions specific to patent documents is sent to an external AI3 to generate the translated text; (2) Terminology standardization stage: Standardization of technical terms based on the bilingual table as post-editing; (3) Syntax adjustment stage: Adjustment to expressions specific to patent documents as post-editing; (4) Final finishing stage: Style standardization and quality check as post-editing.

[0038] In the first stage, a prompt is generated by integrating a translation table extracted from the translation DB72 (e.g., corresponding to specialized terms such as "semiconductor substrate" and "etching") and examples of expression formats specific to patent documents (e.g., standardized expressions such as "characterized by", "consisting of", and "comprising"), and is sent to the external AI3 along with the document to be translated, similar reference documents, and the translations of the similar reference documents. As a result, consistency in specialized terminology and expression formats specific to patent documents are applied from the initial translation stage, and high-quality translations that follow past translation methods are generated.

[0039] Stages 2 through 4 are performed as post-editing processes, with processing time appropriately allocated to each stage in the order of initial translation, terminology standardization, syntax adjustment, and final refinement. Furthermore, a confidence scoring function numerically evaluates the confidence level of the AI-generated output on a scale of 0.0 to 1.0, and automatically re-translates any low-confidence portions below 0.8. The generated translations are stored in the translation DB72. This ensures consistent, high-quality translations.

[0040] In particular, a series of descriptions indicated consecutively with parenthetical numbers such as (1), (2), and (3) will be translated as a single unit. This makes it easier to use consistent English expressions for related technical content, improving the consistency of the translation.

[0041] The translation table generation unit 106 creates a translation table from a reference document and its translation. It has a function to automatically generate a translation table from documents picked up by similarity search, and extracts correspondences of specialized terms such as "semiconductor substrate → semiconductor substrate" and "etching → etching" from translation pairs to build a dictionary for each technical field. As another method for building a dictionary for each technical field, for example, correspondences of semiconductor-related specialized terms are automatically extracted from a group of patents under the IPC classification "H01L" (semiconductor equipment). The created translation table is stored in the translation DB 72. This ensures consistency in specialized terminology.

[0042] The learning unit 107 corrects reference information (here, for example, the way to translate a specific sentence, translation results, etc.) and the translation table (more specifically, for example, a dictionary, typical technical texts and their translations) based on the feedback. With the automatic expansion function of the translation database, new translation cases are automatically added to the translation database, and with the error pattern learning function, error patterns are automatically learned from the parts corrected by humans. The corrected information is stored in the translation DB 72. This realizes continuous improvement of translation quality.

[0043] Furthermore, it is also possible to have the external AI 3 correct the prompt and the translation table (dictionary) when performing translation based on the comments of the translation requester on the delivered translation or the results of the review of the delivered translation by the translation requester. More specifically, a predetermined part of the prompt is managed as a separate file, and by rewriting the content of this file, continuous improvement of translation quality is realized. That is, although the details will be described later, the prompt for giving instructions to the external AI 3 is automatically enhanced through the reflection of feedback on the translation.

[0044] The feedback reception unit 108 receives feedback from the user on the translation result. This feedback includes comments from the translation requester and the results of the review of the delivered translation. The received feedback information is utilized as important input information for improving the translation quality of the entire system.

[0045] The context correction unit 109 corrects the reference information generated by the context generation unit 104 based on the feedback received by the feedback reception unit 108. Specifically, a designated part of the prompt is managed as a separate file, and by rewriting the content of this file, continuous improvement of translation quality is realized. The corrected context information is stored in the translation DB 72 and reflected in future translation processing.

[0046] The translation table correction unit 110 corrects the translation table created by the translation table generation unit 106 based on the feedback received by the feedback reception unit 108. It learns the correspondence of technical terms from the translation parts corrected by humans and constructs a more appropriate translation table. The corrected translation table is stored in the translation DB 72, contributing to the improvement of the translation accuracy of technical terms.

[0047] In the translation DB 72, in addition to the translation results, reference information, and translation table, an expression form example database specific to patent documents is constructed. The translation table stores the correspondence of technical terms classified by technical field and applicant. For example, term correspondences such as "semiconductor substrate → semiconductor substrate", "etching → etching", "sputtering → sputtering" are automatically extracted and accumulated.

[0048] In the expression form example database, the stereotypical expressions specific to patent documents are systematically organized. As specific examples, expression correspondences such as "characterized by / characterized by", "consisting of / consisting of", "comprising / comprising", "wherein / wherein", "characterized in that / characterized in that", "including / including", "preferably / preferably", "specifically / specifically", "for example / for example" are stored.

[0049] These databases are efficiently managed by an index by technical field based on IPC classification (International Patent Classification) and an index by applicant based on the applicant identifier. The optimal translation table and expression form examples are automatically selected according to the technical field of the document to be translated and the applicant information. The update of the database is continuously executed based on the accumulation of new translation cases and manual feedback.

[0050] The document database 71 is a database that stores documents to be translated and reference documents. Documents are stored in a structured format based on the paragraph numbers of the patent documents, and vectorized document information by the vectorization unit 102 is also stored, enabling efficient searching and referencing. The translation database 72 is a database that stores various data related to the translation process, such as translation results, reference information, and translation tables. In particular, in the translation database 72, each paragraph of the translated text is stored linked to the corresponding paragraph of the document to be translated, and is managed in a structured format based on paragraph numbers, similar to the document database 71. These databases work together to support continuous learning and quality improvement.

[0051] The specific vectorization process performed by the vectorization unit 102 will be described in detail. The vectorization process is performed in the following steps. First, as a preprocessing step, formal symbols such as paragraph numbers

[0052] are removed and normalization is performed based on a specialized terminology dictionary. Second, tokenization is performed using a pre-trained language model (e.g., 768 dimensions) to convert each paragraph into a numerical vector. Third, the magnitude of each vector is unified to 1 by L2 normalization.

[0052] As a concrete example, the paragraph "Lithium metal batteries are secondary batteries that use lithium metal as the negative electrode, and can be high-capacity batteries." is transformed into a multidimensional vector. This vectorization maps semantically similar paragraphs to nearby positions, enabling efficient similarity searches.

[0053] Figure 5 is a diagram showing the translation processing flow according to one embodiment of the present invention. This diagram shows the flow of processing performed by the functional block shown in Figure 4. As shown in Figure 5, the translation processing is performed in the following steps. First, in step S101, the translation target reception unit 101 receives input of the document to be translated, and in step S102, the vectorization unit 102 vectorizes the document to be translated in document units, paragraph units, and sentence units. Similarly, in step S103, the reference document stored in the document database 71 is read from the storage unit 18, and in step S104, the vectorization unit 102 vectorizes the reference document in document units, paragraph units, and sentence units.

[0054] Next, in step S105, the search unit 103 identifies similar sections by calculating similarity. This process involves a multi-stage similarity calculation process, as shown in Figure 7, which will be described later. At this stage, the context generation unit 104 extracts reference documents corresponding to the document to be translated and the translated documents of those reference documents from the translation DB 72. In step S106, the parallel translation table generation unit 106 generates a translation table (e.g., a dictionary) based on the reference documents and their translations. Furthermore, in step S107, the context generation unit 104 integrates the document to be translated, the reference documents corresponding to the document to be translated, and the translated documents of those reference documents as translation information to generate context information. Finally, in step S108, the translated text generation unit 105 controls the generation AI of the external AI 3 to generate a translated text based on the context information and the translation table (e.g., a dictionary), and in step S109, the translated document is output.

[0055] Figure 6 shows a document structuring process according to one embodiment of the present invention. This process is performed by the vectorization unit 102. As shown in Figure 6, the input document is vectorized at three structuring levels—document level, paragraph level, and sentence level—through steps S201 to S209. At the document level, the entire document is represented as a single vector; at the paragraph level, each paragraph is vectorized individually; and at the sentence level, each sentence is vectorized individually. This generates a document vector, a group of paragraph vectors, and a group of sentence vectors. This vector information is stored in the translation DB 72 in Figure 4 and used in the subsequent similarity calculation process.

[0056] Figure 7 is a diagram showing the similarity calculation process flow according to one embodiment of the present invention. This process is executed by the search unit 103 and corresponds to step S105 in Figure 5. First, in step S301, a vector comparison of the entire document is performed in document-level similarity calculation, and in step S302, a comparison with a first threshold is performed. If it is equal to or greater than the first threshold, in step S303, paragraph-level similarity is calculated through vector comparison of each paragraph. In step S304, a comparison with a second threshold is performed. In step S304, if it is equal to or greater than the second threshold, in step S305, sentence-level similarity is calculated through vector comparison of each sentence. In step S306, a comparison with a third threshold is performed. Finally, in step S307, weighting calculation is performed by integrating the similarity information. The result of this integration process is used as input for context information generation by the context generation unit 104 (step S107 in Figure 5).

[0057] The specific algorithm for calculating similarity will now be explained. The search unit 103 calculates cosine similarity using the following formula: similarity(A, B) = (A・B) / (||A|| × ||B||) where A is the vector of the paragraph to be translated and B is the vector of the reference paragraph. Regarding threshold settings, the first threshold, second threshold, and third threshold are set to predetermined values ​​based on experimental verification. These thresholds are determined to achieve the optimal balance between translation accuracy and processing speed. Note that these thresholds (predetermined values) are not limited to fixed values. For example, values ​​statistically calculated using a ground truth dataset (ground truth) from past translation cases to maximize the F-value, which is the harmonic mean of the precision and recall of the search results, may be adopted. Alternatively, the user may be able to arbitrarily change the settings according to the technical field and importance of the document to be translated, or they may be variable values ​​that are dynamically adjusted according to the system's resource status and the number of search hits. By defining the logic for determining predetermined values ​​from multiple perspectives in this way, it becomes possible to set optimal thresholds according to the situation.

[0058] Furthermore, in weighted similarity calculation, a weighting coefficient α based on the frequency of occurrence of technical terms is set to an appropriate range and applied, resulting in a more precise similarity evaluation using the formula: weighted_similarity = α × term_similarity + (1-α) × semantic_similarity.

[0059] Figure 8 shows the main screen of a translation support tool according to one embodiment of the present invention. This main screen is the interface provided by the translation target receiving unit 101. On this main screen, 11 processing modules from 1. File registration to 11. Display are selectable, and the user can sequentially execute the necessary processing. This screen corresponds to the functions of the translation target receiving unit 101.

[0060] Figure 9 shows an example of a document to be translated according to one embodiment of the present invention. This is an example of a document received by the translation document receiving unit 101. In this example, paragraphs

[0001] to

[0011] of a patent document are shown, and the technical content relating to lithium metal batteries is described. This document is received in step S101 in Figure 5 and is vectorized by the vectorization unit 102 in Figure 5 in step S102 in Figure 5.

[0061] Figure 10 shows the vectorization result according to one embodiment of the present invention. This is an example of a vector representation generated by the vectorization unit 102. The vectorization unit 102 converts the document content in Figure 9 into a numerical vector, and the original Japanese text is stored in the text field and the numerical array in the vector field. This vectorization result corresponds to the output of the document structuring process shown in Figure 6 and is stored in the translation DB 72 as the processing result of steps S102 and S104 in Figure 5.

[0062] Figures 11 and 12 show first and second examples of similarity search results according to one embodiment of the present invention. These show the results of a similarity search performed by the search unit 103. In Figure 11, reference documents with similarity scores of 0.9939 and 0.6574 are extracted by the search according to the process in Figure 7, and their corresponding English translations are shown. In Figure 12, a reference document with a similarity score of 0.8136 is shown. These processes correspond to the functions of the search unit 103 (step S105 in Figure 5). The extracted similar reference documents and their corresponding translations serve as basic data when the context generation unit 104 generates context information.

[0063] Figure 13 shows an example of prompt generation according to one embodiment of the present invention. This is an example of a prompt generated by the context generation unit 104. The context generation unit 104 integrates information such as the document to be translated, similar reference documents, and a translation table to form a prompt for the generating AI. This prompt includes the instruction, "You are an expert in patent specification translation," and information to be referenced. Here, the information to be referenced consists of information such as the document to be translated, similar reference documents, English translations of similar reference documents (for example, an English translation by an expert, an English translation previously translated and established as a model), and a translation table. This prompt is the result of context generation in step S107 of Figure 5 and is transmitted to the external AI 3 by the translation generation unit 105 in step S108.

[0064] Figures 14 and 15 show first and second examples of translation results according to one embodiment of the present invention. These are examples of translations generated by the translation generation unit 105 using external AI 3. In Figure 14, since the content of the paragraph to be translated and the content of the reference paragraph are the same, the translation result is exactly the same as the English translation of the reference. In Figure 15, an original translation has been made, but the English translation examples of similar reference documents have been followed as much as possible. These processes correspond to the functions of the translation generation unit 105 and step S108 in Figure 5. The generated translation results are sent to the user terminal 2 as output in step S109 of Figure 5.

[0065] This system can employ a parallel processing architecture as a performance optimization function. To efficiently process large-scale patent documents, for example, a 100-page patent specification can be processed in 10 parallel processes, with each processing unit independently performing "vectorization, similarity search, prompt generation, and AI translation." The cache optimization function improves processing efficiency by caching and reusing translation results of frequently occurring patterns and specialized terms. The adaptive resource allocation function dynamically allocates computing resources according to the importance and complexity of each part of the document. These performance optimization functions are designed to improve the overall operational efficiency of the translation unit 100 shown in Figure 4 and are controlled by the CPU 11 of the server 1.

[0066] To improve the processing efficiency of this system, multiple optimization techniques have been implemented. Firstly, a vectorization result caching function reduces the reprocessing time for the same document by more than 90%. Secondly, parallel processing of similarity calculations reduces the processing time for large documents (100 pages or more) by 60% compared to conventional methods. Thirdly, pre-calculation of frequently occurring translation patterns enables real-time translation response. As a concrete performance indicator, efficient processing time has been achieved for large patent documents. Furthermore, it has been confirmed that the system maintains an average response time of less than 3 seconds when operating with approximately 100 simultaneous users, functioning as a practical translation support system.

[0067] Furthermore, as a multimodal extension function, the drawing reference linkage function enables understanding by linking drawing references in the specification with actual drawings, and the mathematical and chemical formula processing function accurately processes mathematical and chemical formulas in the patent document. Specifically, the drawing reference linkage function first extracts drawing reference descriptions such as "shown in Figure 1" and coded terms such as "search unit 103" through text analysis of the specification. In parallel, image recognition technology (e.g., OCR and object detection model) is used to identify the character string and the graphic area (e.g., bounding box) that the character indicates from the corresponding drawing image data. Then, the character extracted from the text is compared with the character extracted from the drawing to generate mapping data that defines the correspondence between terms in the text and specific areas in the drawing. When instructing the external AI3 (particularly a multimodal model capable of processing images and text simultaneously, i.e., artificial intelligence with multimodal processing capabilities capable of processing multiple modalities including drawings and text) to perform translation, in addition to the text data to be translated, related drawing image data is provided as input. At this time, the prompt generated by the context generation unit 104 includes specific instructions based on the mapping data, such as, for example, "The rectangular area indicated by the reference numeral '103' in the image of Figure 1 corresponds to the 'search unit' in the text. Please translate the paragraph based on this visual structural information and correspondence." This enables the external AI 3 to comprehensively understand visual information such as the connection relationships and arrangements between components in the drawing and the textual explanation, and to generate an accurate translation that is consistent with the drawing without misinterpreting what the reference numerals point to. Furthermore, as an intellectual property strategy linkage function, the scope of rights consideration translation function performs translation adjustments to maintain and optimize the scope of patent rights, the prior art differentiation function selects expressions that clearly differentiate from prior art in translations at the examination stage, such as in response to rejection grounds, and the international application optimization function performs translation adjustments that take into account the examination trends of each country when filing PCT international applications or entering national phases. These extended functions extend the functions of the context generation unit 104 and the translation generation unit 105, and provide specialized processing to achieve particularly high-quality translations.

[0068] Furthermore, the learning unit 107 corrects reference information using the context correction unit 109 and updates the translation table using the translation table correction unit 110 based on user feedback received through the feedback reception unit 108. This continuously improves the overall translation quality of the system. The translation results generated through the process shown in Figure 5, and the feedback received therefrom, are stored in the translation DB 72 to improve the accuracy of future translation processing.

[0069] In addition, as a function of the translation generation unit 105, the claim-priority translation learning function prioritizes processing the claims, which are the most important elements in the patent document, and reflects the translation quality of the claims throughout the entire specification.

[0070] Specifically, the claim-priority translation process is performed in the following step-by-step flow. In the first step, when the translation target receiving unit 101 receives a patent application document, it automatically identifies the claim portion using its document analysis function and extracts it sequentially starting from claim 1. In the second step, for each extracted claim, the vectorization unit 102 performs patent claim-specific vectorization, and the search unit 103 searches for similar claims from a claim-specific database.

[0071] In the third stage, the context generation unit 104 generates reference information that emphasizes "clarity of the claim," "appropriateness of the technical scope," and "consistency of dependency relationships" using a prompt template specifically for claims, and integrates this as context information. In the fourth stage, the translation generation unit 105 generates an initial translation by applying claim-specific translation rules (for example, differentiating between "comprising" and "consisting of," and ensuring consistency of technical terminology).

[0072] In the fifth stage, the generated claim translations are presented to the user through the feedback reception unit 108 for revision and approval by experts. In the sixth stage, the translation table generation unit 106 automatically generates a claim-specific translation table from the approved claim translations, and this translation table is referenced with the highest priority in the translation of the entire specification. Through this step-by-step process, the accuracy of the claim translations is carried over to the entire specification, ensuring consistency as a technical document.

[0073] As a variation, it is effective to translate only the claims first and then receive feedback from the user on the translation results. It is effective to save the claims to be translated, the user-modified translations of those claims, and the translation table created based on those translations in the document database 71 and the translation database 72, and then translate the entire specification. This is because the claims contain the core technology of the patented invention and its translation, and that translation can be reflected throughout the entire specification. For example, there is a very high probability that descriptions similar to those in the claims are contained in the specification. Then, using a translation of the claims, for example by an expert, as a reference document, it becomes possible to translate sentences in the specification that are similar to the descriptions in the claims.

[0074] The translation generation unit 105 includes a translation quality optimization engine, which comprehensively evaluates translation quality by combining multiple evaluation criteria such as the BLEU score, TER, and METEOR, which are evaluation indices for machine translation. It also achieves hybrid translation that leverages the characteristics of each engine through a function to cooperate with multiple commercial machine translation engines. Furthermore, it automatically verifies the clarity of the claim description, the consistency of terminology, and the appropriateness of the description format through a quality check function specific to patent documents.

[0075] As specific quality evaluation indicators, the terminology matching rate (degree of consistency of terminology within the same document), syntactic consistency score (degree of conformity to the standard style of patent documents), and translation fluency index (degree of natural English expression) are quantified to calculate an overall quality score. If the overall quality score falls below a set threshold (usually 0.85), a re-translation process is automatically executed. This set threshold is not limited to a fixed value; a value statistically calculated based on the correlation analysis between past quality evaluation data and the amount of manual correction may also be adopted. For example, a score at which the probability of the amount of correction being below a predetermined level is 95% or higher can be set as the threshold. Furthermore, the user may be able to adjust the threshold depending on the purpose of the translation (e.g., for viewing or for filing applications). These functions are applied when generating the translated text in step S108 of Figure 5.

[0076] The system of the present invention is particularly suitable for translating patent application manuscripts that are divided into paragraphs according to their content and further assigned paragraph numbers. Generally, it is desirable that the RAG (Retrievable Augmented Generation) database used to create context in generative AI be structured. Since patent application manuscripts are already structured with paragraphs and paragraph numbers, they are an ideal input format for the system of the present invention. The basic application in the structured original language and the application translated from that basic application can both be linked by paragraph numbers and stored in a database. For example, the translation target reception unit 101 shown in Figure 4 receives the document to be translated and stores it in the document database 71 and the translation DB 72.

[0077] This section explains the specific mechanism for improving translation accuracy using paragraph numbers. By utilizing the unique paragraph numbering structure of patent documents, such as

[0079] ,

[0080] , etc., and performing paragraph-level correspondence, translation accuracy is improved. In this system, structured data is constructed in the document database 71 and translation DB 72 using paragraph numbers as an index. This allows for efficient searching of the corresponding paragraph in a reference document based on content similarity for any paragraph in the document to be translated (e.g.,

[0081] ). Since the position and context of each paragraph within the document are clearly identified by the paragraph number, the correspondence between the searched reference paragraph and its translation can be accurately grasped, and the context generation unit 104 can generate reference information that appropriately combines the three elements of the paragraph to be translated, similar reference paragraphs, and their translated sentences. This structuring enables precise translation control at the paragraph level compared to conventional translation methods that target the entire document.

[0078] Specifically, the basic patent application document is stored in a structured state in the document database 71, and the translated patent application document, with corresponding paragraphs linked, is stored in the translation database 72. This system extracts paragraphs from reference documents that are similar in content to the paragraph to be translated, extracts the translations linked to those paragraphs in the reference documents, and uses the three elements of the content of the paragraph to be translated, the paragraphs in the reference documents, and the translation of the paragraphs in the reference documents as context data to execute a process in which the external AI 3 translates the paragraph to be translated. For example, the vectorization unit 102 shown in Figure 4 vectorizes the document, the search unit 103 searches for similar reference documents, the context generation unit 104 generates context information, and the translation generation unit 105 controls the external AI 3 to generate the translated text.

[0079] Regarding patent application manuscripts and their translations stored as data for RAG, when limited to a specific technical field, practical translation accuracy can be achieved with a dataset of several dozen items, and it is preferable to prepare a dataset of around 100 items if higher accuracy is required. On the other hand, even when a single database integrating all technical fields is constructed (for example, the document database 71 and translation DB 72 in Figure 4), appropriate reference data can be extracted by the similar paragraph detection process performed by the search unit 103, so efficient translation processing is possible even if 1,000 or 10,000 application documents are stored. In particular, by performing vectorization processing in advance using the vectorization unit 102, the processing time of the similarity search process by the search unit 103 can be significantly reduced.

[0080] It is effective to create a document database 71 and a translation database 72 for each applicant and each applicant's technical field, respectively. For example, it is possible to reflect the specific expressions and translation styles of each applicant and technical field. Even for the same technical content, the expression often differs from applicant to applicant, so it is possible to respond to such situations and realize custom-made translations. For example, the translation table generation unit 106 shown in Figure 4 generates translation tables specific to the applicant and technical field, and the learning unit 107 continuously improves them based on feedback from the feedback receiving unit 108.

[0081] As a specific technique for constructing applicant-specific databases, the database structuring function assigns a unique identifier (e.g., applicant code) to each applicant, and based on this identifier, a logical partition dedicated to the applicant is created in the document database 71 and translation DB 72. The technology classification function manages data in a two-dimensional matrix structure using a combination of IPC codes (e.g., H01L, G06F, etc.) and applicant codes.

[0082] The applicant-specific terminology learning function automatically extracts applicant-specific technical terms, expression patterns, and translation styles when new documents are registered, and continuously updates the applicant's dedicated translation table (dictionary data). For example, if applicant A always translates "processing device" as "processing apparatus" and applicant B translates it as "processing device," different translation pairs will be learned in each applicant's database.

[0083] The specific learning algorithm extracts applicant-specific translation patterns using machine learning. For frequently occurring terms, the frequency of occurrence, contextual information, and translation consistency are comprehensively evaluated to automatically construct an applicant-specific translation dictionary. For example, once a predetermined number of training data points have been accumulated, the applicant-specific translation model is activated, and further data accumulation enables highly accurate individualized responses. This individualized response has improved translation consistency within the same applicant from the conventional 70% to over 98%, resulting in a significant improvement in applicant satisfaction.

[0084] The translation consistency management function statistically analyzes translation styles (active / passive voice preference, consistent rules for technical terminology, and stylistic characteristics) from past translations of the same applicant, and automatically applies those styles to new translations. The version control function manages the update history of the applicant database, allowing for tracking of changes in translation quality. These technologies enable a customized translation system tailored to the specific translation requirements of each applicant.

[0085] Furthermore, in the semiconductor field, for example, the definitions of source and drain may be reversed from company to company. In such cases, if a general-purpose machine translation system simply performs translation processing despite the lack of standardized terminology definitions among applicants, it may produce technically incorrect translation results. Such differences in technical field-specific terminology can be flexibly handled by using databases built for each applicant.

[0086] As a specific example, in the translation of "source electrode" and "drain electrode" in the semiconductor field, applicant X may use "source electrode → source electrode" and "drain electrode → drain electrode," while applicant Y may have the opposite correspondence: "source electrode → drain electrode" and "drain electrode → source electrode." This system automatically selects the appropriate translation table based on the applicant identification information, achieving accurate translations in accordance with the applicant's specific technical definitions.

[0087] This ensures consistent translations based on each applicant's past translation examples, guaranteeing uniformity of terminology and consistency of expression within the same applicant. As an effect of improving translation accuracy, it has been confirmed that using an applicant-specific database improves the accuracy of specialized term translation by 35-50% compared to using a general-purpose database, enabling translations without technical inconsistencies. For example, the context correction unit 109 shown in Figure 4 corrects the reference information, and the parallel text table correction unit 110 updates the parallel text table to respond to applicant-specific translation requests.

[0088] (Modified Version) This is a functional block diagram showing an example of the functional configuration of the server in Figure 3 that constitutes the information processing system in Figure 2. Furthermore, the CPU 11 of server 1 further includes a prompt optimization unit 111, a drawing analysis unit 112, a distributed translation management unit 113, a terminology uniformity monitoring unit 114, a translation integration unit 115, a translation appropriateness evaluation unit 116, a warning presentation unit 117, an alternative expression suggestion unit 118, a range identification unit 119, an examination trend analysis unit 120, a claim amendment generation unit 121, a quality evaluation unit 122, The following functions are involved: score calculation unit 123, improvement suggestion unit 124, dictionary update unit 125, history recording unit 126, visualization unit 127, quality control unit 128, multilingual translation unit 129, consistency verification unit 130, interlingual unification unit 131, cost optimization unit 132, encryption unit 133, access control unit 134, confidentiality protection unit 135, verification management unit 136, approval automation unit 137, workflow control unit 138, data collection unit 139, analysis unit 140, and report generation unit 141.

[0089] The prompt optimization unit 111 automatically optimizes instruction prompts to the external AI 3 based on feedback received by the feedback receiving unit 108. The drawing analysis unit 112 analyzes the drawings included in the document to be translated and identifies the correspondence between elements in the drawings and descriptions in the document to be translated. The distributed translation management unit 113 divides the document to be translated into multiple sections and assigns them to multiple translators. The terminology consistency monitoring unit 114 monitors the translation results from multiple translators in real time and detects terminology inconsistencies. The translation integration unit 115 integrates the translation results from multiple translators and generates a single completed translated document.

[0090] The translation appropriateness evaluation unit 116 obtains the translated text generated by the translated text generation unit 105 and evaluates whether it is appropriate from one or more predetermined viewpoints. The warning display unit 117 presents the problems detected by the translation appropriateness evaluation unit 116 to the user terminal 2. The alternative expression suggestion unit 118 proposes alternative expressions corresponding to the detected problems. The scope identification unit 119 extracts the technical matters described in the specification and identifies the scope that can be described as claims. The examination trend analysis unit 120 analyzes past examination cases in the country where the application is to be filed and extracts the examination trend. The claim amendment generation unit 121 generates a proposed amendment to the claim that conforms to the examination trend analyzed by the examination trend analysis unit 120, within the scope that can be described as claims identified by the scope identification unit 119.

[0091] The quality evaluation unit 122 acquires the translated text generated by the translated text generation unit 105 and quantitatively evaluates the quality of the translated text using multiple evaluation indicators. The score calculation unit 123 calculates an overall quality score based on the multiple evaluation indicators. The improvement suggestion unit 124 generates suggestions for improving translation quality based on the overall quality score. The dictionary update unit 125 dynamically updates the specialized terminology dictionary corresponding to the technical field and applicant. The history recording unit 126 records the processing details at each stage of the translation process. The visualization unit 127 visualizes the recorded processing details in chronological order. The quality control unit 128 manages the changes in translation quality in a traceable format.

[0092] The multilingual translation unit 129 performs simultaneous translation from a single source text into multiple languages. The consistency verification unit 130 verifies the consistency of the translated content across multiple languages. The interlingual unification unit 131 ensures terminology consistency across multiple languages. The cost optimization unit 132 optimizes the combination of AI translation and human translation based on the required translation quality level. The encryption unit 133 encrypts the data of the document to be translated. The access control unit 134 controls access to the translation process. The confidentiality protection unit 135 protects the confidentiality of the translation history.

[0093] The Verification Management Unit 136 manages a multi-stage verification process for translation results. The Approval Automation Unit 137 automates the approval flow in the verification process. The Workflow Control Unit 138 controls the workflow for quality assurance. The Data Collection Unit 139 collects and stores translation performance data. The Analysis Unit 140 analyzes the translation performance data to extract quality trends. The Report Generation Unit 141 generates reports based on the quality trends.

[0094] The learning function in this embodiment analyzes user feedback and automatically optimizes instruction prompts and translation tables for the external AI3. This reduces manual adjustment work and enables the system to autonomously learn each applicant's translation style, continuously improving translation quality.

[0095] Specifically, as shown in Figure 4, the learning unit 107 includes a feedback receiving unit 108, a context correction unit 109, and a translation table correction unit 110. In this embodiment, however, the learning unit 107 further includes a prompt optimization unit 111. The prompt optimization unit 111 automatically optimizes the instruction prompts to the external AI 3 based on the feedback received by the feedback receiving unit 108. Specifically, the prompt optimization unit 111 analyzes the parts of the translation result that the user has corrected and uses machine learning to extract what correction patterns occur frequently. For example, if the user has corrected "composing" to "including" in multiple translation results, the prompt optimization unit 111 learns this correction pattern. Then, for the next translation, it automatically adds an instruction to the prompt sent to the external AI 3 stating, "Use 'including' preferentially over 'composing' in certain contexts."

[0096] The optimization process by the prompt optimization unit 111 is performed in the following steps. First, as a feedback data storage process, user modifications are stored in the translation DB 72. Second, as a modification pattern extraction process, frequently occurring modification patterns are statistically extracted from the stored modification data. Third, as a prompt element identification process, the corresponding locations within the prompt are identified. Fourth, as a prompt update process, the prompt descriptions at the identified locations are modified. For prompt management, instead of managing the entire prompt as a single file, each element of the prompt is managed separately as a different file. For example, it is divided into "basic instruction section," "translation table reference instruction section," "expression format instruction section," "technical field specific instruction section," etc. This allows the prompt optimization unit 111 to rewrite only the element files that need modification, achieving efficient prompt management.

[0097] Furthermore, the prompt optimization unit 111 can also request the external AI 3 itself to improve the prompt. Specifically, it sends a meta-prompt to the external AI 3 stating, "Please provide prompt improvement suggestions to improve translation quality based on the following feedback data," and obtains improvement suggestions from the external AI 3. The prompt optimization unit 111 evaluates the validity of the obtained improvement suggestions and reflects them in the actual prompt if deemed appropriate. In addition, the prompt optimization unit 111 optimizes prompts by applicant and by technical field. For example, optimization based on feedback from applicant A is reflected in a prompt element file dedicated to applicant A and is applied only when translating for applicant A. This allows the system to automatically learn each applicant's unique translation style and terminology preferences, enabling the provision of customized translation services. As a measure of the effectiveness of prompt optimization, indicators such as BLEU score, terminology consistency rate, and number of user corrections are automatically recorded before and after optimization. This allows for quantitative evaluation of whether prompt optimization actually contributes to improving translation quality.

[0098] The multimodal translation function in this embodiment analyzes drawings, mathematical formulas, and chemical formulas and utilizes them in translation. By recognizing text and structures in drawings and automatically associating them with descriptions in the specification, accurate translation referencing drawing information is achieved, improving the understanding of the technical content. Specifically, the information processing system in this embodiment has a multimodal translation function. The translation text generation unit 105 includes a drawing analysis unit 112. The drawing analysis unit 112 analyzes the drawings included in the document to be translated and identifies the correspondence between elements in the drawings and descriptions in the document to be translated. The processing by the drawing analysis unit 112 is performed in the following procedure. First, as a drawing acquisition process, drawing files (PDF, JPEG, PNG, etc.) are extracted from the document to be translated. In the case of a patent application document, multiple drawings such as [Figure 1] and [Figure 2] are included, so these are extracted individually. Second, as an OCR process, text (symbols, explanatory text, etc.) written in the drawings is extracted by optical character recognition. For example, it recognizes strings such as "Server 1," "CPU 11," and "Search Unit 103" in the drawing. Thirdly, as a structural recognition process, it recognizes the positional relationships and connection relationships of each element (blocks, arrows, lines, etc.) in the drawing. For example, in a functional block diagram like Figure 4, it recognizes the arrangement of each block and the connection relationships between blocks indicated by arrows. Image recognition technology is used to extract rectangular areas as blocks and arrows as connection relationships.

[0099] Fourth, as an element mapping process, each element in the drawing is mapped to the description in the specification. Specifically, a reference numeral in the drawing (e.g., "search unit 103") is mapped to the description in the specification (e.g., paragraph [00XX], "The search unit 103 searches for a reference document corresponding to the document to be translated"). This mapping is performed based on the matching of reference numerals, the similarity of neighboring text, etc. The mapping relationships identified by the drawing analysis unit 112 are used in the translation process by the translation generation unit 105. Specifically, the drawing information is included in the prompt that the context generation unit 104 sends to the external AI 3. For example, a prompt such as "Please translate the paragraph corresponding to the search unit 103 in Figure 4. In Figure 4, the search unit 103 receives input from the vectorization unit 102 and sends output to the context generation unit 104. Please translate while considering this structure" is generated.

[0100] Furthermore, it is possible to send the drawing itself as image data to the external AI3. If the external AI3 is a multimodal AI (an AI capable of processing both images and text), providing both the drawing image and text as input simultaneously enables more contextually relevant translation. For example, a drawing image can be sent with a prompt such as, "Please translate the following paragraphs while referring to the attached drawing." Similar processing is possible for chemical formulas and mathematical formulas. The drawing analysis unit 112 includes a chemical formula recognition engine and a mathematical formula recognition engine, and recognizes chemical formulas and mathematical formulas in the drawing or text. The recognized chemical formulas and mathematical formulas are converted into structured data (e.g., SMILES format, LaTeX format) and used in the translation process. For example, if the chemical formula "CH3-CH2-OH" is described in the specification, the drawing analysis unit 112 recognizes it and converts it to SMILES format. During translation, the external AI3 is instructed to appropriately translate this chemical formula as the standard English notation "ethanool" or as a structural formula. Multimodal translation provides the following benefits. Firstly, translations that maintain consistency between drawings and text are achieved. Secondly, referring to drawings deepens understanding of the technical content, enabling more accurate translations. Thirdly, the accuracy of translations of chemical formulas and mathematical formulas is improved.

[0101] Another example involves first sending the entire document and drawings to a generating AI, which then identifies the drawing corresponding to each paragraph. The drawing number is then included in the translation context, associated with the paragraph to be translated. This ensures that even if there is no explicit mention of a drawing within a paragraph, the generating AI can reliably compare the content of the paragraph to be translated with the drawing during the translation process.

[0102] As another example, the entire document and drawings are first sent to the generation AI, which identifies the paragraphs corresponding to each drawing and groups these paragraphs together. Then, the generation AI is instructed to translate each group using API functions or similar methods. This ensures that the translation is accurate and consistent with the drawings.

[0103] The collaborative translation function in this embodiment allows multiple translators to work simultaneously, with AI integrating and adjusting each translation. By dividing the document and processing it in parallel, and monitoring terminology consistency in real time, it shortens the translation time for large documents while ensuring consistency.

[0104] Specifically, the information processing system of this embodiment has a real-time collaborative translation function. This function allows multiple translators to work on the same document simultaneously and complete the translation efficiently. Server 1 includes a distributed translation management unit 113, a terminology uniformity monitoring unit 114, and a translation integration unit 115. These are implemented as software modules that function on the CPU 11 of Server 1. The distributed translation management unit 113 divides the document to be translated into multiple sections and assigns them to multiple translators. The unit of division is determined based on the structure of the document, such as paragraph numbers, chapters, and sections. For example, a patent specification consisting of 100 paragraphs is assigned paragraphs

[0001] to

[0033] to translator A, paragraphs

[0034] to

[0066] to translator B, and paragraphs

[0067] to

[0100] to translator C. Each translator is notified of the translation task for their assigned section via the user terminal 2. The translators perform the translation work on their respective user terminals 2. The translation results are sent to server 1 in real time and stored in translation DB 72.

[0105] The terminology consistency monitoring unit 114 monitors the translation results from multiple translators in real time and detects inconsistencies in terminology. Specifically, it automatically detects when different translations are used for the same source term (e.g., "search unit"), such as "search unit" by translator A and "searching unit" by translator B. Detected terminology inconsistencies are immediately notified to all translators involved. The notification screen displays the message, "There is an inconsistency in the translation of 'search unit'. Translator A: 'search unit', Translator B: 'searching unit'. Which will you adopt?" After the translators discuss and decide on a translation, that decision is shared with all translators, and the unified translation is used thereafter. The terminology consistency monitoring unit 114 works in conjunction with the translation table generation unit 106. The decided translation is automatically added to the translation table and stored in the translation database 72. This ensures consistency in terminology throughout the document.

[0106] The translation integration unit 115 integrates the translation results from multiple translators to generate a single completed translated document. During the integration process, stylistic adjustments at section boundaries, verification of paragraph number consistency, and verification of cross-reference accuracy are automatically performed. For example, if translator A's section frequently uses the passive voice and translator B's section frequently uses the active voice, the translation integration unit 115 detects the inconsistency in stylistics and proposes unifying them to one style. This proposal is generated using an external AI 3. Specifically, the translated texts from both sections are sent to the external AI 3 along with the prompt, "Please unify the stylistics of the following two sections," and a unified translation is obtained. The real-time collaborative translation system provides the following benefits: Firstly, the translation time for large documents is significantly reduced. A task that previously took one translator 10 days can be completed in 3-4 days by having three translators process it in parallel. Secondly, translation quality is improved because terminology consistency among multiple translators is automatically ensured. Thirdly, communication between translators becomes more efficient.

[0107] The translation appropriateness evaluation function in this embodiment evaluates the generated translation from multiple perspectives (legal systems, technical fields, examination practices, etc.) and provides warnings and alternative suggestions if problems are found. This enables translation expressions that are appropriate to the legal systems and technical characteristics of each country, thereby reducing the risk of rejection.

[0108] Specifically, the information processing system of this embodiment has a translation appropriateness evaluation function. Server 1 includes a translation appropriateness evaluation unit 116, a warning presentation unit 117, and an alternative expression suggestion unit 118. The translation appropriateness evaluation unit 116 acquires the translated text generated by the translated text generation unit 105 and evaluates whether it is appropriate from one or more predetermined viewpoints. The predetermined viewpoints are stored in advance in the translation DB 72 and are selected according to the characteristics of the translated document to be evaluated (country of application, technical field, applicant, etc.). Examples of predetermined viewpoints include the following: Firstly, conformity to the legal system of the country of application. For example, conformity to Section 112 of the U.S.C. Patent Act (clarity requirement), conformity to Section 84 of the European Patent Convention (EPC) (clarity and brevity), etc. are evaluated. Secondly, conformity to the characteristics of the technical field. Technical fields are classified into computer software, machinery and structure, chemistry and materials, bio and pharmaceuticals, electrical and semiconductor fields, etc.

[0109] In the computer software field, compliance with Section 101 of the U.S. Patent Act (patent eligibility) is particularly important. The translation appropriateness evaluation unit 116 assesses whether there is a risk that the generated translation will be judged as an abstract idea. Specifically, it checks whether specific hardware configurations (processor, memory, etc.) are clearly stated in the translation, and whether the technical effects are described. If these are lacking, it issues a warning. In the mechanical and structural fields, clarity of structural features is important. The translation appropriateness evaluation unit 116 checks whether ambiguous expressions such as "substantially" and "approximately" are used. Since these expressions are likely to be pointed out as unclear in U.S. examination, if detected, it issues a warning and suggests clearer expressions (such as limiting by numerical range). In the chemical and materials field, the method of identifying compounds and materials is important. Simply stating "organic solvent" is insufficient in scope; therefore, it is recommended to specify concrete examples (in Markush format) or limit the scope by chemical structure. The translation appropriateness evaluation unit 116 evaluates whether such limitations are made appropriately.

[0110] Another perspective is suitability to examination practices. The translation appropriateness evaluation unit 116 analyzes past examination data from each country and statistically evaluates whether certain expressions are likely to be rejected. For example, if the “means for…” format (means plus function) is used in the United States, it calculates the rejection rate based on examination data from the past five years and issues a warning if the risk is high. Another perspective is consistency with the specification. In the translation of the claims, it checks whether the terms used are supported in the specification. For example, if “motor” is translated in the claims, but only “stepping motor” is described in the examples in the specification, it warns of the risk of violating the support requirement. A fifth perspective is consistency of translation. It checks whether different translations are used for the same concept within the same document. For example, if the term "acquisition unit" is translated differently in different paragraphs as "acquisition unit," "acquisition unit," and "obtaining unit," the system will detect the inconsistency and issue a warning.

[0111] The warning display unit 117 presents the problems detected by the translation appropriateness evaluation unit 116 to the user terminal 2. The warning screen displays the location of the problem, the type of problem, its importance (high, medium, low), and the specific details of the problem. For example, the message "Symbol (the symbol is ! inside a triangle) Importance: High - Risk of Section 101 of the U.S. Patent Act. The translation of paragraph

[0023] lacks sufficient description of the specific hardware configuration. There is a risk that it will be judged as an abstract idea." is displayed. The alternative expression suggestion unit 118 proposes alternative expressions that address the detected problems. The suggestions are presented as multiple alternatives, and for each alternative, the aspects of improvement, impact on the scope of rights, recommendation level, etc. are also displayed. For example, in response to the Section 101 risk mentioned above, alternatives such as "A system comprising: a memory storage instruction; and a processor configured to execute the instructions to..." (explicitly specifying the hardware configuration) are proposed as alternative 1, and "A non-transitor computer-readable medium storage instruction..." (changing to a medium claim) are proposed as alternative 2. Each alternative is accompanied by evaluation information such as "Section 101 risk: low, scope of rights: somewhat narrow, recommendation level: high". External AI3 is used to generate alternative expressions. The alternative expression suggestion unit 118 sends a prompt containing the problematic section and its content to the external AI3 and obtains multiple alternative expression suggestions. The translation appropriateness evaluation unit 116 re-evaluates the obtained alternatives to confirm whether the problem has been resolved. The translation appropriateness evaluation system significantly improves the quality of the translated work and reduces the risk of rejection during review in each country.

[0112] The country-specific optimized claim generation function in this embodiment automatically generates and translates draft claims optimized for the examination trends of the target country within the scope of the specification. This supports amendment strategies during PCT national phase transitions, improves the rate of patent acquisition in each country, and enables optimal claim expression for the same invention in each country.

[0113] Specifically, the information processing system of this embodiment has a country-specific optimized claim generation function. This function is particularly useful in situations where claims need to be optimized according to the examination trends of each country, such as during the national phase entry of a PCT international application. Server 1 includes a scope identification unit 119, an examination trend analysis unit 120, and a claim amendment generation unit 121. The scope identification unit 119 extracts technical matters described in the specification and identifies the scope that can be described as a claim. Specifically, it analyzes each paragraph of the specification and extracts technical components, method steps, materials, numerical ranges, etc. The extracted technical matters are linked to paragraph numbers and stored in the translation DB 72. For example, if paragraph

[0023] of the specification states, "The search unit 103 searches for reference documents based on the similarity between documents vectorized by the vectorization unit 102," the scope identification unit 119 extracts technical elements such as "search unit," "vectorization," "similarity," and "reference documents." These elements are recognized as the scope that can be described as a claim.

[0114] The Examination Trend Analysis Unit 120 analyzes past examination cases in the country where the patent application is to be filed and extracts examination trends. The examination case data to be analyzed is obtained from public databases of national patent offices, commercial patent databases, etc. The analysis items include the format of the claims, the choice of terms, the number of dependent claims, the rejection rate by technical field, etc. For example, the following is extracted as a US examination trend: In software-related inventions, claims that specify the concrete hardware configuration have a high examination pass rate (approximately 82%). When using functional descriptions, the "configured to" format has a higher examination pass rate than the "means for" format (the former 78%, the latter 45%). If the technical effect is clearly stated in the claim, the examination pass rate improves by 12%. The following is extracted as a European examination trend: In method claims, if the technical problem and technical effect are clearly stated, the examination pass rate is high. Method claims tend to be examined more strictly than apparatus claims (apparatus 75%, method 62%). The following is extracted as a Chinese examination trend: Applications with a large number of dependent claims (10 or more) offer greater room for amendment during the examination process and have a higher final patent acquisition rate. Claim descriptions that are faithful to the embodiments are preferred.

[0115] The claim amendment generation unit 121 generates a proposed claim amendment that conforms to the examination trends analyzed by the examination trend analysis unit 120, within the range that can be described as a claim as identified by the scope identification unit 119. The specific processing procedure is as follows: First, as an analysis of the original claim, the current claim's constituent elements, description format, technical scope, etc., are analyzed. Second, as an examination risk assessment, the risk of the current claim being rejected in the country where the application is to be filed is assessed. Third, as a specification support confirmation, it is confirmed whether the expansion of the claim's scope or changes in expression are supported in the specification. Fourth, as an amendment generation, a proposed claim amendment that conforms to the examination trends and is supported in the specification is generated. For example, if the original claim in Japanese is "an information processing device comprising an acquisition unit for acquiring data and a processing unit for processing the acquired data," the following country-specific optimization is performed.

[0116] The US-optimized version will include specific hardware configurations and technical effects. For example, it will be described as follows: "An information processing apparatus, compiling: a memory storing instructions; and a processor configured to execute the instructions to: acquire data; and process the acquired data to achieve [effects described in specification

[0008] ], while [features described in specification

[0034] ]." In this way, by explicitly stating the technical effects described in paragraph

[0008] of the specification in the claims and adding the characteristic configuration described in paragraph

[0034] , the risk of Section 101 (patent eligibility) is reduced and differentiation from the prior art is strengthened. In the optimized version for Europe, the technical challenges are explicitly stated and conversion to method claims is also considered. For example, it would be written as follows: "A method for processing information, the method comprising: acquiring data; processing the acquired data to solve the technical problem of [the problem described in specification

[0005] ]; while the processing provides the technical effect of [the effect described in specification

[0008] ]." In the optimized version for China, the number of dependent claims is increased to ensure room for correction. In addition to one independent claim, ten or more dependent claims are generated that reflect the features described in each paragraph of the specification. This increases the likelihood of obtaining rights by amending dependent claims to make them independent, even if independent claims are rejected during the examination process. The generated proposed claim amendments are translated by the translation generation unit 105. During translation, the translation system of the basic application (search unit 103, context generation unit 104, etc.) is utilized. In addition, by combining this with the translation appropriateness evaluation function, it is automatically confirmed whether the translated text of the amended claims conforms to the legal system of the country where the application is to be filed. The country-specific optimized claim generation system improves the examination pass rate in each country and enables efficient acquisition of rights.

[0117] The paragraph dependency analysis function in this embodiment analyzes the citation and reference relationships between paragraphs and automatically ensures translation consistency of demonstrative pronouns such as "the above" and "as described above." It also automatically handles changes in paragraph order to guarantee logical consistency in translation.

[0118] Specifically, the information processing system of this embodiment has a function for analyzing inter-paragraph dependencies. The context generation unit 104, which serves as a reference information generation means, includes a dependency analysis unit 222. The dependency analysis unit 222 analyzes the reference relationships between paragraphs in the document to be translated and identifies the content of the preceding paragraph referenced by the demonstrative pronoun. In patent documents, reference expressions such as "the search unit," "the process described above," and "as shown in Figure 3" are frequently used, so accurately understanding these reference relationships greatly affects translation quality. The processing by the dependency analysis unit 222 is performed in the following steps. First, as a demonstrative pronoun detection process, expressions containing demonstrative pronouns such as "the," "as described above," "the said," and "that" are extracted from the document to be translated. Second, as a reference target identification process, the preceding paragraph or preceding text referenced by each demonstrative pronoun is identified. For this identification, the anaphora resolution method in natural language processing technology is used. For example, if paragraph

[0045] contains the statement "The search unit calculates the similarity between vectorized documents," the dependency analysis unit 222 identifies the preceding paragraph that "the search unit" is referring to. If paragraph

[0023] contains the statement "The search unit 103 searches for reference documents corresponding to the document to be translated," and this is the first occurrence of this statement, then it is determined that "the search unit" in paragraph

[0045] is referring to paragraph

[0023] .

[0119] Thirdly, as a reference graph generation process, the inter-paragraph reference relationships of the entire document are represented as a graph structure. A graph is generated in which each paragraph is a node and the reference relationship is an edge. This graph visualizes which paragraphs reference which other paragraphs. Fourthly, as a translation order determination process, the optimal translation order is determined, taking into account the reference relationships. Basically, translation is performed in paragraph number order, but if a referenced paragraph has not yet been translated, the referenced paragraph is translated first. This allows paragraphs containing demonstrative pronouns to be translated while referring to the translation results of the referenced paragraphs, ensuring consistency of terminology. The reference relationships identified by the dependency analysis unit 222 are used in prompt generation by the context generation unit 104. Specifically, the prompt when translating a paragraph containing a demonstrative pronoun includes the original text and the translated text of the referenced paragraph. For example, when translating paragraph

[0045] "The search unit calculates the similarity between vectorized documents," the prompt includes the following information. In paragraph

[0023] , “search unit 103” is translated as “search unit 103”. Please translate “the search unit” in paragraph

[0045] . To maintain consistency, please translate it as “the search unit”. By explicitly including the translated expression in the referenced source in the prompt in this way, the external AI3 can generate a consistent translation. Furthermore, the dependency analysis unit 222 also responds to changes in paragraph order. For example, if it becomes necessary to rearrange the order of paragraphs in the translated document (e.g., changes in structure according to the review standards of each country), it adjusts the references of demonstratives based on the reference relationship graph to ensure that the references are correctly maintained. Specifically, if the paragraph referenced by “the foregoing” moves to the later part of the document, it changes “the foregoing” to “details to be described later.” Paragraph dependency analysis provides the following benefits: Firstly, consistency of terminology in the translation of demonstratives is ensured. Secondly, logical consistency between paragraphs is maintained. Thirdly, even when changes in paragraph order are necessary, translation that maintains reference relationships becomes possible.

[0120] The claim tree function in this embodiment visualizes the hierarchical structure of independent and dependent claims, and displays the technical features added by each dependent claim linked to the corresponding paragraphs in the specification. This helps prevent oversights during translation and supports overall understanding.

[0121] Specifically, the information processing system of this embodiment has a claim tree visualization function. Server 1 includes a claim structure analysis unit 223 and a claim visualization unit 224. The claim structure analysis unit 223 analyzes the hierarchical structure of independent and dependent claims in the claims and extracts the technical features added by each dependent claim. The analysis process is performed in the following steps. First, as a claim classification process, it is determined whether each claim is an independent claim or a dependent claim. In Japanese claims, this is determined by the presence or absence of the phrase "as described in claim ○". In English claims, this is determined by the phrase "The appearance of claim ○", etc. Second, as a dependency relationship identification process, it is determined which claim each dependent claim is dependent on. For example, if claim 2 is described as "the information processing system described in claim 1", it is determined that claim 2 is dependent on claim 1. Furthermore, if claim 5 is described as "an information processing system according to any one of claims 1 to 4," then claim 5 is determined to be a multi-dependent claim that is dependent on all of claims 1, 2, 3, and 4.

[0122] Thirdly, as an additional feature extraction process, the technical features added by each dependent claim are extracted. Novel technical elements not described in the dependent claim are extracted from the body of the dependent claim. For example, if claim 1 is described as "comprising a search means, a reference information generation means, and a translated text generation means," and claim 2 is described as "the search means performs a search based on the similarity between vectorized documents," then the additional features of claim 2 are extracted as "vectorization" and "search based on similarity." Fourthly, as a specification mapping process, the paragraphs of the specification that support the extracted technical features are identified. For example, paragraph

[0031] is identified as the supporting basis for "vectorization," and paragraph

[0032] is identified as the supporting basis for "search based on similarity." The claim visualization unit 224 visualizes and displays the hierarchical structure and technical features analyzed by the claim structure analysis unit 223. The visualization is represented as a tree structure (tree diagram) and displayed on the screen of the user terminal 2. An example of the tree structure display is as follows. An independent claim (e.g., claim 1) is placed at the top level, followed by directly dependent claims (e.g., claims 2, 3, and 4). Further below, secondary dependent claims (e.g., claim 5 is dependent on claim 2) are placed.

[0123] Each claim node displays the following information: claim number, added technical features (in keyword format), supporting specification paragraph number, translation status (untranslated / translated), and translation quality score (evaluation result for draft 4-A). By clicking on any claim node in the tree, the user can view detailed information about that claim, including the original text, translated text, and supporting specification paragraph. The claim tree visualization function has the following effects on the translation process: Firstly, it allows for an overview of the overall claim structure, making it easier to grasp the progress of the translation. Secondly, the additional features of each dependent claim become clearer, reducing the chances of overlooking details during translation. Thirdly, the support relationship with the specification is visualized, allowing for the early detection of risks of violating support requirements. The claim visualization unit 224 is also used to check for consistency in terminology during translation. For example, if "search means" is translated as "search means" in claim 1, and "search unit" is translated as "searching unit" in claim 5 (indirectly dependent on claim 1), the use of different translations for the same concept is visually detected. In this case, a warning is displayed, prompting the translator to unify the terminology. Furthermore, the claim tree is also useful for considering translation strategies. When there are many multi-dependent claims, it provides information to support strategic decisions, such as whether to eliminate multi-dependencies and change them to single dependencies, depending on the examination practices of each country.

[0124] The translation system with competitive analysis functionality in this embodiment adds patent documents and translations of competitors in the same technical field to the reference data. It distinguishes between terms that should be differentiated and industry standard terms that should be standardized, and proposes strategic translation expressions.

[0125] Specifically, the information processing system of this embodiment has a competitor document reference translation function. This function enables more strategic translation by referring not only to the company's own translations but also to translation examples of competitors. The search unit 103 searches for patent documents of competitors and their translations belonging to the same technical field as the document to be translated as reference documents. The patent documents of competitors are obtained from the public patent gazette database. The technical field is determined based on IPC classification, FI terms, keywords, etc. For example, if the IPC classification of the document to be translated is G06F (electrical digital data processing), then patent documents of competitors with the same G06F classification will be searched. Among the searched competitor documents, those with a high degree of similarity to the document to be translated are selected as reference documents. The context generation unit 104 includes a term classification unit 225. The term classification unit 225 distinguishes and extracts industry standard terms and differentiating terms from the translations of competitor patent documents.

[0126] Industry standard terminology refers to common technical terms widely used in the relevant technical field, where translations are consistent among competitors. For example, if "semiconductor substrate" is consistently translated as "semiconductor substrate" in translations by multiple competitors, it is determined to be industry standard terminology. Differentiating terminology refers to unique terms used by each company to express its own unique technologies or products. For example, if one competitor uses "adaptive processing" to describe its unique processing technology, and another competitor uses "intelligent processing" to describe a similar technology, these are determined to be differentiating terms. The term classification unit 225 classifies terms according to the following criteria. The first criterion is frequency of occurrence. If the same translation is used frequently in translations by multiple competitors, it is determined to be industry standard terminology. The second criterion is frequency of use in technical literature. Terms frequently used in IEEE and other technical literature databases such as academic papers are considered industry standard terms. The third criterion is originality. Terms used exclusively by a specific applicant are considered differentiating terms.

[0127] The classified terminology information is used in the translation process. Specifically, the prompts generated by the context generation unit 104 include the following instructions: For industry standard terms, the instruction "Translate 'semiconductor substrate' as the industry standard 'semiconductor substrate'. Competitors A, B, and C also use the same translation" is added. This ensures consistency in terminology within the industry and makes it easier to compare with technical documents and other companies' patents. For differentiating terms, more strategic instructions are given. For example, the instruction "For the term 'adaptive processing', competitor A translates it as 'adaptive processing', and competitor B translates it as 'intelligent processing'. Please propose a translation that takes your company's technical characteristics into consideration and differentiates it from other companies" is added. Based on these instructions, the external AI3 proposes alternative translations that differ from those of competitors, such as "context-aware processing." The proposed translations are evaluated for their consistency with the content of the specification, technical accuracy, and effectiveness in differentiation, and the most suitable translation is selected.

[0128] Furthermore, the competitor document reference function also contributes to reducing the risk of patent infringement. By comparing the claims of competitors with those of your own company and avoiding excessively similar expressions, you can avoid overlapping scope of rights. Specifically, it calculates the similarity between your company's draft claims and those of competitors, and issues a warning if the similarity exceeds a predetermined threshold (e.g., 0.85). The warning screen displays a message such as, "Your claim 1 is highly similar to claim 1 of competitor C's patent no. XXX (similarity: 0.89). Please consider differentiating your expression to avoid overlapping scope of rights." In addition, it is possible to analyze the translation quality of competitors. Translation expressions from patents that have passed examination and been registered by competitors are learned as high-quality examples and reflected in your own translations. Conversely, translation expressions from rejected patents are recorded as expressions to be avoided. The competitor document reference translation function provides the following effects. Firstly, the unification of industry standard terminology promotes understanding within the technical community. Secondly, the strategic use of differentiating terms clarifies the uniqueness of one's own technology. Thirdly, it avoids overlapping scope of rights with competitors, reducing the risk of patent infringement.

[0129] The voice translation support function in this embodiment efficiently improves translations by allowing translators to give correction instructions via voice. Abstract instructions such as "make it more formal" are converted into specific corrections, and hands-free operation is achieved by learning voice patterns. Specifically, the information processing system in this embodiment has a translation instruction function via voice input. Server 1 includes a voice recognition unit 226 and an intent interpretation unit 227. The voice recognition unit 226 recognizes the translator's voice input via the microphone of the user terminal 2 and converts it into text data. Existing voice recognition APIs (e.g., Google Speech-to-Text, Azure Speech Services, etc.) are used for voice recognition. The recognized voice text is sent to the intent interpretation unit 227 as translation correction instructions. Examples of voice correction instructions include "make this paragraph more formal," "change technical terms to more specialized expressions," "change passive voice to active voice," and "split this sentence into two sentences." The intent interpretation unit 227 interprets the speech text recognized by the speech recognition unit 226 and converts it into specific translation correction content. External AI 3 is used for this conversion process.

[0130] The specific processing steps are as follows: First, as part of the instruction target identification process, the scope of the translated text targeted by the correction instruction is identified. The target of the correction is identified from instruction words such as "this paragraph," "this sentence," and "here." If the translator has selected a specific paragraph or sentence on the screen, that selection will be the target of the correction. Second, as part of the correction intent classification process, the intent of the correction instruction is classified. The correction intent is classified into changes in writing style (formalization, casualization, etc.), changes in expression (specialized / easy-to-understand, active / passive voice, etc.), changes in structure (sentence splitting, sentence joining, etc.), and changes in terminology (synonym substitution, etc.). Third, as part of the specific correction content generation process, the specific corrected text is generated using the external AI3. For example, in response to the instruction "Make this paragraph more formal," the following prompt is sent to the external AI3: "Please rewrite the following English text in a more formal style. Avoid colloquial expressions and use an academic / technical style. [Target translated text]" Based on this prompt, the external AI3 generates the corrected translated text. For example, modifications such as changing "get" to "obtain" and "find out" to "determine" may be made.

[0131] Fourth, as a revision confirmation process, the generated revised text is presented to the translator and approval is requested. If the translator responds with voice commands such as "OK" or "Approved," the revision is finalized. If the translator responds with voice commands such as "Redo" or "Another suggestion," an alternative revision is generated. The speech recognition unit 226 has a function to learn the translator's unique voice patterns. It registers revision instruction expressions that the same translator repeatedly uses (e.g., "the usual formal conversion") and makes them available as shortened commands. It also records the history of voice instructions and analyzes frequently occurring revision patterns. For example, if a particular translator frequently makes the correction "change passive voice to active voice," the system adjusts the prompt to prioritize the active voice from the initial translation stage. The following are anticipated use cases for the voice input function: Firstly, to reduce the burden of keyboard input when reviewing a large volume of translations. Secondly, to allow hands-free revision instructions when performing other tasks (such as referring to materials) in parallel with translation. Thirdly, in team work, when multiple translators discuss revisions verbally, the content of those discussions can be directly input as revision instructions. The voice input translation instruction function offers the following benefits: Firstly, it reduces the burden of keyboard input and improves work efficiency. Secondly, it enables more intuitive instructions, reducing stress for translators. Thirdly, voice pattern learning allows the system to be optimized for each individual translator.

[0132] The collaborative support function in this embodiment allows multiple stakeholders to check the translation progress and quality in real time. It provides a collaborative environment where feedback can be given immediately through the comment function, enabling translators, checkers, patent attorneys, and clients to work simultaneously.

[0133] Specifically, the information processing system of this embodiment has a real-time translation progress sharing function. Server 1 includes a progress sharing unit 228 and a comment management unit 229. The progress sharing unit 228 shares the translation progress and quality evaluation in real time with multiple users. Multiple users include all parties involved in the translation project, such as translators, checkers (persons in charge of checking translation quality), patent attorneys, and clients (applicants). Each user can access Server 1 via user terminal 2 and check the translation progress on a dedicated dashboard screen. The following information is displayed on the dashboard. Firstly, the overall progress rate shows what percentage of the entire document to be translated has been completed. The progress rate is displayed using multiple indicators, such as paragraph-based, character-based, and page-based. Secondly, the progress status by paragraph is displayed in a color-coded table format or Gantt chart format, showing the status of each paragraph (not started, in translation, completed, being checked, checked, etc.). Thirdly, as a quality evaluation score, the quality score calculated by the translation appropriateness evaluation function of Proposal 4-A for the completed translation portion is displayed. The score is aggregated at the paragraph level, section level, and document-wide level.

[0134] Fourth, as part of the person in charge information, information such as who is translating each paragraph and who is checking it is displayed. In conjunction with the real-time collaborative translation function of Proposal 3, the status when multiple translators are working in parallel is visualized. Progress information is updated in real time. When a translator completes the translation of a paragraph, the status of that paragraph on the dashboard is immediately changed to "Translation Complete," and all stakeholders can confirm the change. The comment management unit 229 receives comments from multiple users and displays them linked to the translated text. Each user can add comments to any part of the translated text. The method of using the comment function is as follows: First, as a selection of the comment target, the user selects any paragraph, sentence, or word in the translated text with the mouse. Second, as a comment input, the user enters a comment for the selected part into the text box. Examples of comment content include, "Wouldn't 'component' be more appropriate than 'unit' for this translation?" and "Please check the support relationship with specification

[0045] ."

[0135] Thirdly, as a comment notification, when a comment is posted, a notification is sent to all relevant parties (or designated personnel). Notifications are made via system notifications, email notifications, mobile app notifications, etc. Fourthly, as a comment response, the person who received the comment posts a reply comment or makes corrections to resolve the comment. Comments are assigned a status (unresolved, in progress, resolved) and managed accordingly. Comments are displayed as an icon or highlight in the relevant section of the translated text. When a user clicks on the icon, the comment details (poster, posting date and time, content, reply history, etc.) are displayed. In addition, the comment management unit 229 has a comment classification function. Comments are automatically classified according to their content into categories such as "questions about terminology," "quality comments," "requests for confirmation of specification support," and "schedule-related communications." This makes it easier to prioritize and search for comments. Furthermore, the progress sharing unit 228 has a milestone management function. Important milestones in the translation project (e.g., claim translation completion, first draft completion, final delivery, etc.) are set, and the achievement status of each milestone is tracked. When a milestone is reached, an automatic notification is sent to all stakeholders. The real-time translation progress sharing feature provides the following benefits: First, it improves the overall transparency of the project and eliminates information gaps among stakeholders. Second, problems and questions are identified and shared early, enabling quick responses. Third, clients can check the progress in real time, increasing their sense of security and satisfaction.

[0136] The translation system with quality check function in this embodiment automatically detects typographical errors and omissions in the source text and suggests corrections. For source texts with ambiguous meanings, it generates multiple interpretation patterns and corresponding translations, simultaneously supporting source text improvement and translation.

[0137] Specifically, the information processing system of this embodiment has a source text quality check function and a function to handle ambiguous interpretations. Server 1 includes a typographical error detection unit 230, a source text revision proposal generation unit 231, and an ambiguous interpretation unit 232. The typographical error detection unit 230 detects typographical errors in the document to be translated. The detection process is performed in the following steps. First, as a dictionary matching process, each word in the document to be translated is compared with a specialized terminology dictionary and a general dictionary. Words that do not exist in the dictionary are extracted as potentially typographical errors. For example, if "search means" is misspelled as "dangerous search means," the word "dangerous search" does not exist in the dictionary, so it is detected as a candidate for a typographical error. Second, as a contextual consistency check process, the validity of a word is determined from the surrounding context. Using a language model in natural language processing technology, the probability of each word appearing in the sentence is calculated, and words with extremely low probabilities are detected as candidates for typographical errors. Third, as a notation variation detection process, cases where different notations are used for the same concept within the same document are detected. For example, if "database," "data base," and "DB" are all used interchangeably, it will be detected as a variation in spelling.

[0138] The original text amendment generation unit 231 generates amendments for the detected typos and missing characters. The following methods are used to generate the amendments. The first method is candidate generation based on the edit distance. For the word detected as a typo candidate, words in the dictionary with a small edit distance (Levenshtein distance) are extracted as correction candidates. For example, for "險索", "検索" with an edit distance of 1 is presented as a correction candidate. The second method is candidate selection based on context. When there are multiple correction candidates, the most appropriate candidate is selected from the surrounding context. For example, for the mis-conversion of "こうさく", candidates such as "工作", "耕作", "交錯", etc. exist, but from the context of "検索部は情報を~する", it is determined as "検索" (originally, "抽出", etc. would be appropriate, but "工作" is inappropriate). The third method is amendment generation using the external AI 3. The entire sentence containing the typo is sent to the external AI 3, and amendments are obtained with a prompt of "There may be a typo in this sentence. Please present the correct sentence." The generated amendments are presented to the user in the following format. "Symbol (the symbol is! inside △) Possibility of typo: '險索手段' in paragraph

[0023] is considered a miswriting of '検索手段'. Original text after correction: '検索手段は、翻訳対象文書に対応する参照文書を検索する。' Translated text after correction: 'The search means searches for reference documents corresponding to the translation target document.'" The user can choose to approve, reject, or enter their own corrections for the presented amendments. The approved corrections are reflected in the original text, and translation is executed based on the original text after the correction.

[0139] The ambiguous interpretation unit 232 detects expressions with ambiguous meanings in the document to be translated and generates multiple interpretation patterns. Ambiguity in meaning mainly arises from ambiguity in dependency relationships. For example, the expression "a woman holding a bouquet of red roses" has the following two interpretations: Interpretation A: "A woman holding a red (bouquet of roses)" (the entire bouquet is red). Interpretation B: "A woman with a (bouquet of red roses)" (the roses are red). The ambiguous interpretation unit 232 generates multiple interpretation patterns in the following steps: First, as an ambiguity detection process, it detects places where dependency relationships cannot be uniquely determined by syntactic analysis. Second, as an interpretation pattern generation process, it enumerates all possible dependency relationship patterns. Third, as a validity evaluation process, it evaluates the validity of each interpretation pattern based on the context and other descriptions in the specification. The generated multiple interpretation patterns are presented to the user. "Symbol (the symbol is an exclamation mark inside a triangle) Ambiguity of interpretation: The “vectorized similarity between documents” in paragraph

[0045] may be interpreted in the following ways: Interpretation A: “vectorized (similarity between documents)” (the similarity is vectorized). Corresponding translation A: “the vectorized similarity between documents” Interpretation B: “(vectorized documents) similarity” (the documents are vectorized). Corresponding translation B: “the similarity between vectorized documents” Recommendation: Interpretation B (consistent with the description in specification

[0031] )" The translation generation unit 105 generates translations corresponding to each of the revised proposals generated by the source text revision proposal generation unit 231 and the multiple interpretation patterns generated by the ambiguous interpretation unit 232. This ensures that an appropriate translation can be obtained even if there are problems with the source text. The source text quality check and ambiguity handling functions provide the following benefits: Firstly, typographical errors and grammatical errors in the source text are corrected before translation, preventing mistranslations. Secondly, for source texts with ambiguous meanings, multiple interpretations and translations are presented, allowing translators to make the appropriate choice. Thirdly, improved source text quality leads to improved translation quality.

[0140] The verification function in this embodiment automatically verifies consistency with the drawings after translation is complete. It checks whether the number of elements in the drawings matches the singular / plural forms in the translated text, and verifies symbol correspondence, terminology consistency, numerical consistency, etc., to guarantee final quality.

[0141] Specifically, the information processing system of this embodiment has a post-translation quality verification function. Server 1 includes a drawing consistency verification unit 233, a symbol correspondence verification unit 234, a terminology unification verification unit 235, and a numerical consistency verification unit 236. These verification processes are executed after all translations are completed by the translated text generation unit 105. The drawing consistency verification unit 233 verifies the consistency of the translated text after translation completion with the drawings included in the document to be translated. In particular, it checks whether the number of elements in the drawing matches the singular / plural form of the translated text. The verification process is executed in the following steps. First, as a drawing element extraction process, the elements drawn in the drawing and their number are recognized. The functions of the drawing analysis unit 112 of Proposal 2 are utilized to extract blocks, arrows, symbols, etc. in the drawing. For example, if multiple user terminals are drawn as "user terminal 2-1, 2-2, 2-N" in Figure 2, it is recognized that there are multiple user terminals. Second, as a translated text singular / plural form extraction process, it is determined whether each noun in the translated text is singular or plural. In English, you need to determine whether it's "user terminal" (singular) or "user terminals" (plural), or whether it's "a user terminal" (singular) or "user terminals" (plural), etc.

[0142] Thirdly, as a consistency verification process, the number of elements in the drawing is checked to see if the singular and plural forms of the translated text match. For example, if multiple user terminals are depicted in the drawing, but the translated text is "The system comprises a user terminal" (singular), this is detected as a mismatch. Fourthly, as a warning generation process, if a mismatch is detected, a warning message is generated and presented to the user. An example of a warning message is as follows: "Symbol (symbol is ! inside a triangle) Mismatch with drawing: Figure 2 depicts multiple user terminals (2-1, 2-2, 2-N), but the translated text in paragraph

[0015] is "a user terminal" (singular). Please consider correcting it to "user terminals" (plural)." The symbol correspondence verification unit 234 verifies the correspondence between the symbols in the specification and claims and the symbols in the drawing. Firstly, all symbols in the specification are extracted. For example, symbols such as "Server 1" and "Search Unit 103" are extracted. Secondly, all symbols in the drawing are extracted. Thirdly, it is checked whether symbols in the specification exist in the drawing, or conversely, whether symbols in the drawing are explained in the specification. If a mismatch is detected, the following warnings are issued: "Symbol (symbol is ! inside a triangle) Symbol mismatch: "Search Unit 103" described in paragraph

[0045] does not exist in Figure 4." "Symbol (symbol is ! inside a triangle) Unexplained symbol: "Database 80" in Figure 4 is not explained in the specification."

[0143] The terminology consistency verification unit 235 verifies whether different translations are used for the same concept throughout the entire translated text. First, as a source term extraction process, it extracts technical terms from the document to be translated and identifies groups of terms that represent the same concept. For example, "search unit" and "search means" represent the same concept. Second, as a translation matching process, it checks whether the translations for the same concept are consistent. For example, if "search unit" is translated as "search unit" and "search means" is translated as "searching means," it is detected as inconsistency. Third, as a section-by-section verification process, it also checks for terminology consistency in each part of the claim, specification, and abstract. In particular, it focuses on checking whether the translations used in the claim are consistently used in the specification. The numerical consistency verification unit 236 verifies whether the numerical values ​​match between the source text and the translated text. First, as a source numerical extraction process, it extracts all numerical values ​​(integers, decimals, fractions, percentages, etc.) from the source text. Secondly, as part of the translation numerical extraction process, all numerical values ​​are extracted from the translated text. Thirdly, as part of the numerical matching process, it is checked whether the numerical values ​​in the corresponding source text and the translated text match. If a mismatch is detected, a warning like the following is issued: "Symbol (symbol is ! inside a triangle) Numerical mismatch: The source text of paragraph

[0056] is "0.85", but the translated text is "0.58".

[0144] Furthermore, if unit conversion is required (e.g., mm → inch), the accuracy of the conversion is also checked. For example, if the original text is "10 mm" and the translated text is "0.39 inches," it is correct (10 mm = 0.3937 inches, mathematical symbol (symbol meaning almost equal to) 0.39 inches), but if it is "3.9 inches," it will be detected as an error. In addition, as part of the paragraph number correspondence verification process, the correspondence of cross-references ("see paragraph

[0023] ," etc.) is also verified. It is checked whether the part in the original text that says "see paragraph

[0023] " is correctly translated as "see paragraph

[0023] ," and whether the paragraph numbers match. The results of all verification processes are output as a verification report. The verification report lists all detected warnings, the severity of the warnings (high, medium, low), and recommended corrections. The translator or checker will perform final checks and revisions based on this report. The post-translation quality verification function provides the following benefits: First, it prevents inconsistencies between drawings and text. Second, it detects missing or incorrect symbols. Third, it automatically identifies inconsistencies in terminology, improving quality. Fourth, it prevents simple errors such as transcription mistakes. Fifth, it strengthens quality assurance before final delivery.

[0145] Although one embodiment of the present invention has been described above, the present invention is not limited to the embodiments described above, and any modifications, improvements, etc. that can achieve the objectives of the present invention are considered to be included in the present invention.

[0146] Furthermore, the system configuration shown in Figure 2 and the hardware configuration of Server 1 shown in Figure 3 are merely illustrative examples for achieving the objectives of the present invention and are not particularly limited.

[0147] Furthermore, the functional block diagram shown in Figure 4 is merely illustrative and not particularly limiting. In other words, it is sufficient that the information processing system in Figure 2 has the functionality to execute the various processes described above as a whole, and the functional blocks and databases used to realize this functionality are not particularly limited to the example in Figure 4.

[0148] Furthermore, the location of the functional blocks and database is not limited to Figure 4 and can be arbitrary. For example, at least a portion of the functional blocks and database located on the server 1 side may be provided on the terminal 2-1 side, terminal 2-2 side, terminal 2-N side, or other information processing device not shown.

[0149] Furthermore, the series of processes described above can be executed by hardware or by software. Also, a single functional block may consist of hardware alone, software alone, or a combination of both.

[0150] When a series of processes are executed by software, the programs that make up that software are installed on a computer or other device from a network or storage medium. The computer may be a computer built into dedicated hardware. Alternatively, the computer may be a computer capable of performing various functions by installing various programs, such as a server, a general-purpose smartphone, or a personal computer.

[0151] Such recording media containing programs may consist not only of removable media (not shown) distributed separately from the main unit of the device to provide the program to the user, but also of recording media provided to the user in a state where they are pre-installed in the main unit of the device.

[0152] In this specification, the step of describing a program to be recorded on a recording medium includes not only processes that are performed chronologically in that order, but also processes that are not necessarily performed chronologically, but are executed in parallel or individually.

[0153] To summarize, an information processing device to which the present invention applies only needs to have the following configuration, and can take various forms. That is, an information processing device to which the present invention applies (for example, Server 1 in Figures 2 to 4) is an information processing system that supports the translation of a document to be translated, and only needs to include: a search means for searching for reference documents corresponding to the document to be translated (for example, the search unit 103 in Figure 4, S104 in Figure 5, S301 to S307 in Figure 7, and the similarity search results in Figures 11 and 12); a reference information generation means for generating reference information including the reference document and its translated text (for example, the context generation unit 104 in Figure 4, S105 in Figure 5, the document structuring processing in Figures S201 to S209, and the prompt generation example in Figure 13); and a translation text generation means for generating a translated text of the document to be translated by machine translation based on instruction information including the reference information (for example, the translation text generation unit 105 in Figure 4, S107 in Figure 5, and the translation results in Figures 14 and 15).

[0154] In this way, translations that ensure consistency with past documents and their translations become possible. Conventional machine translation systems process the documents to be translated independently, making it difficult to ensure consistency in terminology and expression with documents from the same applicant or documents in similar technical fields that have been translated in the past. In contrast, the present invention achieves consistency with past translations by combining structured data using paragraph numbers and RAG (Retrievable Augmented Generation) technology.

[0155] Specifically, the system uses patent-specific embedded representations by the vectorization unit 102 and multidimensional similarity calculations by the search unit 103 to identify the most suitable reference paragraph for the paragraph to be translated. By utilizing the translation examples, parallel tables, and examples of expression formats specific to patent documents, the system simultaneously achieves consistency in technical terminology and stylistic coherence.

[0156] As a result of utilizing the bilingual table and example expressions, the accuracy of specialized terminology translation has improved by 40-60% compared to conventional methods, and the application rate of expression formats specific to patent documents has increased from 85% to over 98%. Experimental results show that translations using this system have a 15-30% improvement in BLUE scores compared to conventional general-purpose machine translation, and a particularly significant improvement of 25-45% in the accuracy of specialized terminology translation has been confirmed. In addition, the rate of terminology consistency within the same applicant has improved from 60% to over 95%, and it has become possible to reduce the amount of manual correction work required after translation by 70-80% compared to conventional methods. In terms of processing time, an overall reduction of 75-85% has been achieved compared to conventional manual translation.

[0157] Furthermore, the system includes vectorization means for vectorizing the document to be translated and the reference document (for example, the vectorization unit 102 in Figure 4, S102 and S103 in Figure 5, the document structuring process in Figure 6, and the vectorization examples in Figures 9 and 10), and the search means can search for the reference document based on the similarity between the vectorized documents.

[0158] This enables efficient similar document searching and faster computation processing.

[0159] Furthermore, the search means can search for the reference documents by comparing normalized texts.

[0160] This makes it possible to provide a flexible search method that does not depend on vectorization.

[0161] Furthermore, the search can be performed at least one level, such as a document, a paragraph, or a sentence.

[0162] This makes it possible to perform multifaceted similarity evaluations based on the hierarchical structure of documents.

[0163] Furthermore, the search means can calculate similarity for each document, paragraph, and sentence unit of the document to be translated, extract paragraphs from documents in which the similarity of the document unit is equal to or greater than a first predetermined value, and extract sentences from the sentences contained in the extracted paragraphs in which the similarity of the sentence unit is equal to or greater than a third predetermined value.

[0164] This enables highly accurate identification of similar areas through gradual refinement and improves processing efficiency.

[0165] Furthermore, the reference information generation means calculates an overall similarity score weighted based on the similarity scores of each document unit, paragraph unit, and sentence unit, and can include the portion where the overall similarity score is equal to or greater than a predetermined value in the reference information along with the corresponding translated text.

[0166] This will enable more precise similarity assessment and improved translation quality.

[0167] Furthermore, when processing on a paragraph-by-paragraph basis, the reference information generation means can generate reference information that includes three elements: the original text of the paragraph to be translated, the original text of similar paragraphs extracted by the search, and the translated text corresponding to those similar paragraphs. Here, the translated text may be a translation that was translated as a model translation, or a translation that was specially translated for a particular user in the past.

[0168] This allows the generating AI to translate while referring to the correspondence between the original text and the translated text of similar paragraphs, thereby improving translation accuracy.

[0169] Furthermore, the system includes a translation table generation means (for example, the translation table generation unit 106 in Figure 4, S106 and S108 in Figure 5, and the translation table in the prompt in Figure 13) that creates a translation table, such as a dictionary, from a reference document and its translation, and the reference information generation means can include the translation table in the reference information.

[0170] This will enable consistency in technical terminology and improve the accuracy of translations of specialized terms.

[0171] Furthermore, the system may further include a feedback receiving means for receiving feedback on documents translated by the system, and a learning means (for example, the learning unit 107, feedback receiving unit 108, context correction unit 109, and translation table correction unit 110 in Figure 4) for correcting at least one of the reference information and the translation table based on the feedback.

[0172] This enables continuous improvement of translation quality and self-evolving of the system.

[0173] Furthermore, the system may further include multimodal processing means for generating translations that take into account the drawings included in the document to be translated.

[0174] This will enable more accurate translations utilizing drawing information and improve the accuracy of patent document translations.

[0175] Furthermore, the translation generation means can generate the translation using artificial intelligence technology.

[0176] This enables the learning of translation patterns through the use of learning capabilities, simultaneous consideration of reference document information (which was previously difficult) through the processing of large amounts of data, ensuring consistency of specialized terminology through contextual understanding capabilities, and handling document-specific expressions through flexible responses.

[0177] Furthermore, the translation generation means can perform translation processing on multiple entries indicated consecutively by parenthetical numbers as a single unit.

[0178] This makes it easier to use consistent translation expressions for related technical content, improving translation consistency.

[0179] An information processing method according to one aspect of the present invention is an information processing method performed by an information processing system that translates a patent application document to which paragraph numbers have been assigned, and includes the steps of generating reference information which includes: the content corresponding to a predetermined paragraph number of the patent application document to be translated, the content of a paragraph in a reference patent application document having similar content to the aforementioned content, and a translated text of the content of the paragraph in the reference patent application document, and integrating the reference information as context into a translation instruction prompt for a machine translation system (for example, external AI 3 in Figure 1) or an external AI (artificial intelligence) (for example, context generation unit 104 in Figure 4, prompt generation example in Figure 13); and causing the external AI 3 to generate a translated text of the content corresponding to the predetermined paragraph number of the patent application document to be translated based on the context and the translation instruction prompt (for example, translation text generation unit 105 in Figure 4, translation results in Figures 14 and 15).

[0180] This enables precise translation by utilizing the paragraph numbering structure unique to patent application documents, and by referring to similar past translation examples, consistent and high-quality translations can be achieved.

[0181] Furthermore, the information processing method may also include a step of searching a predetermined storage unit (for example, the translation DB 72 in Figure 4) for the content of a paragraph in a reference patent application document having similar content to the content described above, and a translation of the content of the paragraph in the reference patent application document (for example, the search unit 103 in Figure 4).

[0182] This allows for the efficient extraction of similar paragraphs from a vast patent document database, and improves the consistency and accuracy of translations through structured search processing based on paragraph numbers.

[0183] The information processing system is an information processing system that supports the translation of a patent application document to which paragraph numbers are assigned, and comprises a database structured based on the paragraph numbers of the patent application document (for example, the document database 71 and translation DB 72 in Figure 4), a search means for searching the database for reference paragraphs similar to the paragraph to be translated (for example, the search unit 103 in Figure 4), a reference information generation means for generating reference information including the reference paragraph and its translated text (for example, the context generation unit 104 in Figure 4), and a translated text generation means for generating a translated text of the paragraph to be translated by machine translation based on instruction information including the reference information (for example, the translated text generation unit 105 in Figure 4).

[0184] This enables precise translation control at the paragraph level by utilizing the paragraph numbering structure unique to patent documents, resulting in a significant improvement in translation accuracy compared to conventional translation methods that target the entire document.

[0185] Furthermore, in an information processing system, the translation generation means can generate a translated text through a multi-stage process consisting of an initial translation stage using prompts that include a bilingual table and examples of expression formats specific to patent documents, and post-editing stages such as a terminology unification stage, a syntax adjustment stage, and a final finishing stage. At each stage, an automatic quality check based on quality evaluation indicators is performed, and if the result falls below a set threshold, a re-translation process is carried out (for example, the multi-stage translation process in the translation generation unit 105 in Figure 4).

[0186] This process ensures consistent translation quality, with uniformity of terminology and application of patent document-specific expression formats in the initial translation stage, and quality checks at each stage guarantee consistent translation quality.

[0187] Furthermore, the information processing system is equipped with a learning means (for example, the learning unit 107 in Figure 4) that constructs an applicant-specific database based on applicant identification information and learns the applicant's unique translation style and correspondence with specialized terminology. The learning means can also extract translation patterns from past translation examples of the same applicant and automatically generate a translation table specific to the applicant (for example, the translation table generation unit 106 in Figure 4 generates an applicant-specific translation table).

[0188] This improves translation consistency within the same applicant from the previous 70% to over 98%, enabling customized translations that correspond to the applicant's unique technical definitions and style of expression.

[0189] Furthermore, in an information processing system, the search means can search for similar reference paragraphs using weighted cosine similarity, which applies weighting coefficients based on the frequency of occurrence of technical terms, and the weighting coefficients can be set to values ​​optimized for each technical field (for example, the weighted similarity calculation function in the search unit 103 in Figure 4).

[0190] This enables similarity calculations that prioritize the matching of technical terms over general expressions, allowing for highly accurate searches of reference paragraphs that appropriately reflect technical terminology specific to the field.

[0191] Furthermore, the information processing system includes a database (for example, the database of example expressions in the translation DB 72 in Figure 4) that stores translation tables classified by technical field and applicant, and examples of expression formats specific to patent documents. The reference information generation means can also extract appropriate translation tables and examples of expression formats from the database based on the technical field and applicant information of the paragraph to be translated, and integrate them into a prompt (for example, the prompt integration process by the context generation unit 104 in Figure 4).

[0192] This enables efficient management using technology-specific indexes and applicant identifiers based on IPC classifications, and allows for the automatic selection of optimal translation support information according to the characteristics of the document to be translated.

[0193] Furthermore, the learning means may include a prompt optimization means (for example, the prompt optimization unit 111 in Figure 4) that automatically optimizes instruction prompts to the external AI based on the feedback received by the feedback receiving means.

[0194] This reduces manual adjustment work and enables the system to autonomously learn each applicant's translation style, continuously improving translation quality.

[0195] Furthermore, the translated text generation means includes a drawing analysis unit (for example, the drawing analysis unit 112 in Figure 4) that analyzes the drawings included in the document to be translated and identifies the correspondence between the elements in the drawings and the descriptions in the document to be translated, and can generate the translated text based on the correspondence.

[0196] This ensures that the consistency between drawings and text is maintained in the translation, leading to a deeper understanding of the technical content and enabling more accurate translations.

[0197] Furthermore, the information processing system may include a distributed translation management unit (for example, the distributed translation management unit 113 in Figure 4) that divides the document to be translated into multiple sections and assigns them to multiple translators, and a terminology consistency monitoring unit (for example, the terminology consistency monitoring unit 114 in Figure 4) that monitors the translation results by multiple translators in real time and detects inconsistencies in terminology.

[0198] This significantly reduces the translation time for large documents and improves translation quality by automatically ensuring consistency in terminology among multiple translators.

[0199] Furthermore, the system may include: acquisition means for acquiring a translated text generated by a translated text generation means; translation appropriateness evaluation means (for example, translation appropriateness evaluation unit 116 in Figure 4) for evaluating whether the translated text is appropriate from one or more predetermined viewpoints; warning presentation means (for example, warning presentation unit 117 in Figure 4) for presenting the problem if the translated text is evaluated as inappropriate from the predetermined viewpoints; and alternative expression suggestion means (for example, alternative expression suggestion unit 118 in Figure 4) for suggesting alternative expressions corresponding to the problem.

[0200] This makes it possible to create translations that are appropriate for the legal systems and technological characteristics of each country, thereby reducing the risk of rejection.

[0201] Furthermore, the system may further include: acquisition means for acquiring information on the claims, specifications, and the country where the application is to be filed (for example, the translation target receiving unit 101 in Figure 4); scope identification means for identifying the scope that can be described as a claim from the technical matters described in the specification (for example, the scope identification unit 119 in Figure 4); examination trend analysis means for analyzing the examination trends of the country where the application is to be filed (for example, the examination trend analysis unit 120 in Figure 4); claim amendment generation means for generating proposed amendments to the claims that are in line with the examination trends of the country where the application is to be filed within the scope that can be described as a claim (for example, the claim amendment generation unit 121 in Figure 4); and translation generation means for generating translated text of the proposed amendments to the claims (for example, the translation generation unit 105 in Figure 4).

[0202] This will improve the approval rate in each country and enable more efficient acquisition of rights.

[0203] Furthermore, the system may include: acquisition means for acquiring the translated text generated by the translated text generation means; quality evaluation means for quantitatively evaluating the quality of the translated text using a plurality of evaluation indicators (for example, the quality evaluation unit 122 in Figure 4); score calculation means for calculating an overall quality score based on the plurality of evaluation indicators (for example, the score calculation unit 123 in Figure 4); and improvement suggestion means for generating suggestions for improving translation quality based on the overall quality score (for example, the improvement suggestion unit 124 in Figure 4).

[0204] This allows for a multifaceted evaluation of translation quality and enables improvements based on objective data.

[0205] Furthermore, the system may include: identification means for identifying the technical field and the applicant (for example, the functions of the translation target reception unit 101 and the context generation unit 104 in Figure 4); dictionary update means for dynamically updating a dictionary of specialized terminology corresponding to the technical field and the applicant (for example, the dictionary update unit 125 in Figure 4); and translation generation means for generating a translated text using the updated dictionary of specialized terminology (for example, the translation generation unit 105 in Figure 4).

[0206] This enables highly accurate translations that always reflect the latest technical terminology and applicant-specific terminology.

[0207] Furthermore, the system may include a history recording means (for example, a history recording unit 126 in Figure 4) for recording the processing details of each of the multiple processing stages in the translation process, a visualization means (for example, a visualization unit 127 in Figure 4) for visualizing the recorded processing details in chronological order, and a quality control means (for example, a quality control unit 128 in Figure 4) for managing the changes in translation quality in a traceable format.

[0208] This will increase transparency in the translation process and enable more efficient quality control.

[0209] Furthermore, the system may include a multilingual translation means (for example, the multilingual translation unit 129 in Figure 4) that performs simultaneous translation from one source text into multiple languages, a consistency verification means (for example, the consistency verification unit 130 in Figure 4) that confirms the consistency of the translated content among the multiple languages, and an interlingual unification means (for example, the interlingual unification unit 131 in Figure 4) that ensures the standardization of terminology among the multiple languages.

[0210] This makes it possible to ensure consistency in translation content and uniformity of terminology even when expanding to multiple languages.

[0211] Furthermore, the system may include a request acquisition means for acquiring a translation quality requirement level (for example, input from the main screen in Figure 8), a cost optimization means for optimizing the combination of AI translation and human translation based on the translation quality requirement level (for example, the cost optimization unit 132 in Figure 4), and a translation execution means for executing translation processing according to the optimized combination (for example, the translation unit 100 in Figure 4).

[0212] This makes it possible to provide translation services that optimize the balance between required quality and cost.

[0213] Furthermore, the system may include encryption means for encrypting the data of the document to be translated (for example, the encryption unit 133 in Figure 4), access control means for controlling access to the translation process (for example, the access control unit 134 in Figure 4), and confidentiality means for protecting the confidentiality of the translation history (for example, the confidentiality unit 135 in Figure 4).

[0214] This makes it possible to safely translate even highly confidential patent documents.

[0215] Furthermore, the system may include verification management means (for example, verification management unit 136 in Figure 4) for managing a multi-stage verification process for translation results, approval automation means (for example, approval automation unit 137 in Figure 4) for automating the approval flow in the verification process, and workflow control means (for example, workflow control unit 138 in Figure 4) for controlling the workflow for quality assurance.

[0216] This will streamline the translation quality assurance process and enable reliable quality control.

[0217] Furthermore, the system may include data collection means for collecting and accumulating translation performance data (for example, the data collection unit 139 in Figure 4), analysis means for analyzing the translation performance data and extracting quality trends (for example, the analysis unit 140 in Figure 4), and report generation means for generating a report based on the quality trends (for example, the report generation unit 141 in Figure 4).

[0218] This allows us to understand the overall trends in translation work and use that information to continuously improve quality and streamline operations.

[0219] 1...Server, 2-1...Terminal, 2-2...Terminal, 2-N...Terminal, 3...Terminal, N...Network, 11...CPU, 12...ROM, 13...RAM, 14...Bus, 15...Input / Output Interface, 16...Input Unit, 17...Output Unit, 18...Storage Unit, 19...Communication Unit, 20...Drive, 21...Removable Media, 100...Translation Unit, 101...Translation Target Reception Unit, 102...Vectorization Unit, 103...Search Unit, 104...Context Generation Unit, 105...Translated Text Generation Unit, 106...Bilingual Table Generation Unit, 107...Learning Unit, 108...Feedback Reception Unit, 109...Context Correction Unit, 110...Bilingual Table Correction Unit, 111...Prompt Optimization Unit, 112...Drawing Analysis Unit, 113...Distributed Translation Management Unit, 114...Terminal Unification Monitoring Unit, 115...Translation Integration Unit, 116... Translation Appropriateness Evaluation Unit, 117...Warning Display Unit, 118...Alternative Expression Proposal Unit, 119...Scope Identification Unit, 120...Examination Trend Analysis Unit, 121...Claim Amendment Proposal Generation Unit, 122...Quality Evaluation Unit, 123...Score Calculation Unit, 124...Improvement Proposal Unit, 125...Dictionary Update Unit, 126...History Recording Unit, 127...Visualization Unit, 128...Quality Control Unit, 129...Multilingual Translation Unit, 130...Consistency Confirmation Unit, 131...Interlingual Unification Unit, 132...Cost Optimization Unit, 133...Encryption Unit, 134...Access Control Unit, 135...Confidentiality Protection Unit, 136...Verification Management Unit, 137...Approval Automation Unit, 138...Workflow Control Unit, 139...Data Collection Unit, 140...Analysis Unit, 141...Report Generation Unit, 222...Dependency Analysis Unit, 223...Claim Structure Analysis Unit, 224...Claim Visualization Unit, 225...Term Classification Unit,226...Speech Recognition Unit, 227...Intent Interpretation Unit, 228...Progress Sharing Unit, 229...Comment Management Unit, 230...Typo Detection Unit, 231...Original Text Revision Proposal Generation Unit, 232...Multi-meaning Interpretation Unit, 233...Drawing Consistency Verification Unit, 234...Code Correspondence Verification Unit, 235...Terminology Unification Verification Unit, 236...Numerical Consistency Verification Unit, 71...Document Database, 72...Translation Database, JD...Patent Application Document to be Translated, RD...Reference Document, ED...Translated Document, A...User, S...Information Processing System

Claims

1. An information processing system that supports the translation of a document to be translated, comprising: search means for searching for reference documents corresponding to the document to be translated; reference information generation means for generating reference information including the reference documents and their translated texts; and translation text generation means for generating a translated text of the document to be translated by machine translation based on instruction information including the reference information.

2. The information processing system according to claim 1, comprising vectorization means for vectorizing the document to be translated and the reference document, wherein the search means searches for the reference document based on the similarity between the vectorized documents.

3. The information processing system according to claim 1, wherein the search means searches for the reference document by comparing normalized texts.

4. The information processing system according to claim 1, wherein the search is performed at least one unit of document units, paragraph units, and sentence units.

5. The information processing system according to claim 4, wherein the search means calculates similarity for each document, paragraph, and sentence unit of the document to be translated, extracts paragraphs from among the paragraphs in documents in which the similarity of the document unit is equal to or equal to a first predetermined value, the paragraphs in which the similarity of the paragraph unit is equal to or equal to a second predetermined value, and extracts sentences from among the sentences in the extracted paragraphs, the sentences in which the similarity of the sentence unit is equal to or equal to a third predetermined value.

6. The information processing system according to claim 5, wherein the reference information generation means calculates an overall similarity weighted based on the similarity of each document unit, paragraph unit, and sentence unit, and includes in the reference information, together with the corresponding translated text, the portion where the overall similarity is equal to or greater than a predetermined value.

7. The information processing system according to claim 1, wherein the reference information generation means generates reference information that includes three elements when processing on a paragraph basis: (1) the original text of the paragraph to be translated, (2) the original text of a similar paragraph extracted by search, and (3) the translated text corresponding to the similar paragraph.

8. The information processing system according to claim 1, further comprising a translation table generation means for creating a translation table from a reference document and its translation, wherein the reference information generation means includes the translation table in the reference information.

9. The information processing system according to claim 8, further comprising a feedback receiving means for receiving feedback on a document translated by the information processing system, and a learning means for modifying at least one of the reference information and the translation table based on the feedback.

10. The information processing system according to claim 1, further comprising multimodal processing means for generating a translation that takes into account the drawings contained in the document to be translated.

11. The information processing system according to claim 1, wherein the translated text generation means generates the translated text using artificial intelligence technology.

12. The information processing system according to claim 1, wherein the translated text generation means performs translation processing on a plurality of descriptions indicated consecutively by parenthetical numbers as a single unit.

13. An information processing method to be performed by an information processing system that assists in the translation of a document to be translated, comprising: a search step of searching for a reference document corresponding to the document to be translated; a reference information generation step of generating reference information including the reference document and its translated text; and a translation text generation step of generating a translated text of the document to be translated by machine translation based on instruction information including the reference information.

14. A program that causes a computer to execute control processing including: a search step of searching for reference documents corresponding to a document to be translated; a reference information generation step of generating reference information including the reference documents and their translated texts; and a translation text generation step of generating a translated text of the document to be translated by machine translation based on instruction information including the reference information.

15. An information processing method to be performed by an information processing system that translates a patent application document to which paragraph numbers have been assigned, comprising: a step of integrating the content corresponding to a predetermined paragraph number of the patent application document to be translated, the content of a paragraph in a reference patent application document having similar content to the content thereto, and a translated text of the content of the paragraph in the reference patent application document as context to a translation instruction prompt for a machine translation system or an external AI (artificial intelligence); and a step of causing the generating AI system to generate a translated text of the content corresponding to the predetermined paragraph number of the patent application document to be translated based on the context and the translation instruction prompt.

16. The information processing method according to claim 15, further comprising the step of retrieving from a predetermined storage unit the content of a paragraph in a reference patent application document having similar content to the content described above, and a translation of the content of the paragraph in the reference patent application document.

17. An information processing system for supporting the translation of a patent application document to which paragraph numbers are assigned, comprising: a database structured based on the paragraph numbers of the patent application document; a search means for searching the database for reference paragraphs similar to the paragraph to be translated; a reference information generation means for generating reference information including the reference paragraph and its translated text; and a translation text generation means for generating a translated text of the paragraph to be translated by machine translation based on instruction information including the reference information.

18. The information processing system according to claim 17, wherein the translated text generation means generates a translated text in a multi-stage process consisting of an initial translation stage using a translation table and prompts including examples of expression formats specific to patent documents, a terminology unification stage, a syntax adjustment stage and a final finishing stage performed as post-editing, and performs an automatic quality check based on quality evaluation indicators at each stage, and performs a re-translation process if it falls below a set threshold.

19. The information processing system according to claim 17, further comprising: an information processing system that constructs an applicant-specific database based on applicant identification information; a learning means for learning the applicant's unique translation style and specialized terminology correspondence; the learning means for extracting translation patterns from past translation examples of the same applicant and automatically generating a translation table specific to the applicant.

20. The information processing system according to claim 17, wherein the search means searches for similar reference paragraphs using weighted cosine similarity, which applies weight coefficients based on the frequency of occurrence of technical terms, and the weight coefficients are set to values ​​optimized for each technical field.

21. The information processing system according to claim 1, comprising a database storing translation tables classified by technical field and applicant, and examples of expression formats specific to patent documents, wherein the reference information generation means extracts appropriate translation tables and examples of expression formats from the database based on the technical field and applicant information of the paragraph to be translated, and integrates them into prompts.

22. The information processing system according to claim 9, wherein the learning means includes a prompt optimization means that automatically optimizes instruction prompts to the external AI based on feedback received by the feedback receiving means.

23. The information processing system according to claim 1, wherein the translated text generation means includes a drawing analysis unit that analyzes a drawing included in the document to be translated and identifies a correspondence between elements in the drawing and descriptions in the document to be translated, and generates the translated text based on the correspondence.

24. The information processing system according to claim 1, further comprising: a distributed translation management unit that divides the document to be translated into multiple sections and assigns them to multiple translators; and a terminology consistency monitoring unit that monitors the translation results by the multiple translators in real time and detects inconsistencies in terminology.

25. An information processing system for supporting the translation of a document to be translated, comprising: acquisition means for acquiring a translated text generated by a translated text generation means; translation appropriateness evaluation means for evaluating whether the translated text is appropriate from one or more predetermined viewpoints; warning presentation means for presenting the problems if the translated text is evaluated as inappropriate from the predetermined viewpoints; and alternative expression suggestion means for proposing alternative expressions corresponding to the problems, as described in claim 1.

26. An information processing system for supporting the translation of a patent application document, further comprising: acquisition means for acquiring information on claims, specifications, and countries where the application is to be filed; scope identification means for identifying the scope that can be described as a claim from the technical matters described in the specification; examination trend analysis means for analyzing the examination trends of the countries where the application is to be filed; claim amendment generation means for generating proposed amendments to the claims that are in line with the examination trends of the countries where the application is to be filed, within the scope that can be described as a claim; and translation generation means for generating a translated text of the proposed amendments to the claims.

27. An information processing system for supporting the translation of a document to be translated, comprising: an acquisition means for acquiring a translated text generated by a translated text generation means; a quality evaluation means for quantitatively evaluating the quality of the translated text using a plurality of evaluation indicators; a score calculation means for calculating an overall quality score based on the plurality of evaluation indicators; and an improvement suggestion means for generating suggestions for improving the translation quality based on the overall quality score, as described in claim 1.

28. An information processing system for supporting the translation of a document to be translated, comprising: identification means for identifying a technical field and an applicant; dictionary update means for dynamically updating a dictionary of specialized terminology corresponding to the technical field and the applicant; and translation generation means for generating a translated text using the updated dictionary of specialized terminology.

29. An information processing system for supporting the translation of a document to be translated, comprising: a history recording means for recording the processing content of each of the multiple processing stages in the translation process; a visualization means for visualizing the recorded processing content in chronological order; and a quality control means for managing the changes in translation quality in a format that allows tracking.

30. An information processing system for supporting the translation of a document to be translated, comprising: a multilingual translation means for performing simultaneous translation from one source text into multiple languages; a consistency verification means for confirming the consistency of the translated content between the multiple languages; and a language unification means for ensuring terminology uniformity between the multiple languages, as described in claim 1.

31. An information processing system for supporting the translation of a document to be translated, comprising: a request acquisition means for acquiring a translation quality requirement level; a cost optimization means for optimizing a combination of AI translation and human translation based on the translation quality requirement level; and a translation execution means for executing a translation process according to the optimized combination, as described in claim 1.

32. An information processing system for supporting the translation of confidential documents, comprising: encryption means for encrypting data of documents to be translated; access control means for controlling access to the translation process; and confidentiality protection means for protecting the confidentiality of the translation history, as described in claim 1.

33. An information processing system for supporting the translation of a document to be translated, comprising: verification management means for managing a multi-stage verification process for translation results; approval automation means for automating the approval flow in the verification process; and workflow control means for controlling a workflow for quality assurance, as described in claim 1.

34. An information processing system for supporting the translation of documents to be translated, comprising: data collection means for collecting and accumulating translation performance data; analysis means for analyzing the translation performance data and extracting quality trends; and report generation means for generating a report based on the quality trends, as described in claim 1.

Images