Data management system, search system, data management method, and data management program
By using extraction and organization models in the data management system to extract text-based datasets from image data, the problem of inefficiently retrieving relevant data for material research or development in existing technologies is solved, achieving efficient and accurate data retrieval.
Patent Information
- Application Number
- CN202610190476.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2025-05-14
- Filing Date
- 2026-02-10
- Publication Date
- 2026-11-17
AI Technical Summary
Existing data management systems struggle to efficiently retrieve data related to materials research or development, especially extracting textual information from image data.
By using an extraction model to extract textual datasets from image data, and combining this with a processing model to generate high-precision document data stored in a database, efficient retrieval can be achieved.
It enables efficient retrieval of data related to materials research or development, improving the retrieval accuracy and efficiency of the data management system.
Smart Images

Figure CN123115452A_ABST
Abstract
Description
Technical Field
[0001] One aspect of the present invention relates to a data management system, a retrieval system, a data management method, and a data management program. Background Technology
[0002] A computer system that uses databases to support research or development is known. For example, Patent Document 1 describes a research and development support system comprising: one or more databases storing various data and information related to research and development; and a management system that accesses the one or more databases for information processing. The one or more databases store information related to the researchers' research and development (i.e., research and development related information), information related to the devices used in the research and development (i.e., device related information), and information on the researchers' device usage history (i.e., usage history information). The management system generates and outputs information supporting research and development based on the usage history information, device related information, and research and development related information.
[0003] Patent Document 1: Japanese Patent Application Publication No. 2020-52602 The goal is to develop a mechanism for efficiently retrieving data related to materials research or development. Summary of the Invention
[0004] The data management system according to one aspect of the present invention includes at least one processor. The at least one processor performs the following processing: acquiring image data of a document related to the material; extracting data from the image by means of a language model, i.e., an extraction model, which extracts the descriptions reflected in the image as text data, the processing including the image data and the first dataset for extraction input data, which represents the descriptions in the document extracted from the image data in text form through processing different from the extraction model, and extracting a second dataset from the image data representing the descriptions in the document in text form; and storing the document data based on the second dataset in a database.
[0005] In this approach, while the extraction model extracts a second dataset in text form from image data of documents related to the material, a first dataset in text form extracted from the image data through a different processing method is also provided to the extraction model. The extraction model extracts the second dataset by referencing not only the image data but also the first dataset, thus enabling high-precision extraction of descriptions within documents from the image data. Since the document data based on the high-precision extracted descriptions is stored in a database, data related to the research or development of the material can be retrieved efficiently using this database.
[0006] Invention Effects According to one aspect of the present invention, data related to the research or development of materials can be retrieved efficiently. Attached Figure Description
[0007] Figure 1 This is a diagram representing an example of a retrieval system.
[0008] Figure 2 This is a diagram illustrating an example of the data structure of the materials database and the auxiliary materials database.
[0009] Figure 3 This is a diagram illustrating an example of the data structure of a supplier database.
[0010] Figure 4 This is a diagram illustrating an example of a portal website.
[0011] Figure 5 This is a flowchart representing an example of the actions of the retrieval system.
[0012] Figure 6 This is a diagram illustrating an example of a data management system.
[0013] Figure 7 This is a flowchart illustrating an example of the actions of a data management system.
[0014] Figure 8 This is a diagram representing an example of image data related to a document.
[0015] Figure 9 This is a diagram illustrating an example of extracting input data.
[0016] Figure 10 This is a diagram illustrating an example of organizing input data.
[0017] Figure 11 This is a diagram representing an example of document data. Detailed Implementation
[0018] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the description of the drawings, the same or equivalent elements are labeled with the same symbols and repeated descriptions are omitted.
[0019] [System Overview] The retrieval system involved in this invention is a computer system for retrieving data related to the research or development of materials. The retrieval system accesses a database storing this data and extracts data from the database corresponding to specified search criteria. Users can utilize this retrieval system to obtain information for conducting materials research or development.
[0020] In one example, the search system possesses at least one of the following search functions: material search, supplier search, reagent search, and auxiliary material search. Materials are substances used to manufacture certain products, such as various organic or inorganic compounds. A supplier is an entity that can provide materials to others. Examples of suppliers include material manufacturers and material wholesalers. Reagents are chemical substances used for the detection or quantification of substances through chemical methods, for substance synthesis experiments, or for the determination of physical properties. Auxiliary materials are spare parts or consumables needed for the maintenance, repair, or operation of equipment.
[0021] In one example, the retrieval system displays a portal on the user's terminal that allows access to more than two search functions. The portal includes a user interface for providing users with access to each of the more than two search functions. Therefore, by simply visiting the portal, the user can efficiently obtain the information they need using more than two search functions.
[0022] In one example, the retrieval system, in addition to or replacing the aforementioned retrieval functions, also possesses a batch retrieval function. Batch retrieval refers to the process of searching each database based on common search criteria, that is, the process of simultaneously searching across multiple databases.
[0023] [System Structure] A retrieval system consists of one or more computers. When multiple computers are used, they are connected through communication networks such as the Internet and intranets to construct a logically unified retrieval system.
[0024] A computer constituting a retrieval system typically includes a processor, memory, and a communication interface as hardware. The processor is, for example, a CPU. The memory consists of storage devices such as flash memory and hard disks. The various functional modules of the retrieval system are implemented by the processor executing programs stored in the memory.
[0025] A retrieval program used to enable a computer to function as a retrieval system includes program code for implementing the various functional modules of the retrieval system. This retrieval program can be provided on a tangible recording medium such as a CD-ROM, DVD-ROM, or semiconductor memory, without being temporarily stored. Alternatively, the retrieval program can be provided via a communication network as a data signal superimposed on a carrier wave. The provided retrieval program may be stored, for example, in memory.
[0026] Figure 1This diagram illustrates a retrieval system 1 as an example. In this example, retrieval system 1 includes a server 10 providing a portal website and retrieval functions, and an internal database 20 storing data related to materials technology. The server 10 is connected to the internal database 20 via a communication network. The server 10 can also be connected to an external database 30 via a communication network. The external database 30 is located on a different computer system than retrieval system 1 and stores data related to materials technology. Furthermore, the server 10 is connected to one or more user terminals 40 via a communication network. The communication network associated with retrieval system 1 can be configured to include at least one of the Internet and an intranet.
[0027] Server 10 includes a processor 101 and a memory 102. The memory 102 stores page data of web pages describing the portal website using markup language. The processor 101 functions as a page management unit 11 and a retrieval unit 12. The page management unit 11 is a functional module that reads page data corresponding to a request from user terminal 40 from memory 102 and displays a web page based on that page data on user terminal 40. The retrieval unit 12 is a functional module that searches the internal database 20 or the external database 30 according to the search criteria received from user terminal 40 and sends the search results to user terminal 40. Sending search results to user terminal 40 is an example of processing that provides search results to the user terminal.
[0028] The internal database 20 includes a materials database 21, a supplier database 22, and a secondary materials database 23. The materials database 21 is a device for non-temporarily storing materials-related data. The supplier database 22 is a device for non-temporarily storing supplier data related to suppliers. The secondary materials database 23 is a device for non-temporarily storing secondary materials-related data.
[0029] External database 30 includes reagent database 31. Reagent database 31 is a device for non-temporarily storing reagent-related reagent data.
[0030] User terminal 40 is a computer operated by a user of retrieval system 1. User terminal 40 can be various types of computers such as personal computers, workstations, tablet terminals, smartphones, and wearable terminals.
[0031] In one example, retrieval system 1 provides retrieval functionality to user terminal 40 via a web application. In this case, various request signals sent from user terminal 40 to server 10 can be considered HTTP requests, and various data sent from server 10 to user terminal 40 in response to these request signals can be considered HTTP responses.
[0032] [Data Structures] Figure 2This is a diagram illustrating an example of the data structure of the materials database 21 and the auxiliary materials database 23. In one example, the materials database 21 and the auxiliary materials database 23 each store multiple document files through a directory structure. Figure 2 The directory structure shown includes a root directory 200 and at least one subdirectory 210 located at the next level below the root directory 200. Each subdirectory 210 contains one or more document files 220. Figure 2 The directory structure shown is three-level, but the directory structure can have two or four levels, or even more complex structures such as different hierarchical structures for each subdirectory.
[0033] A document file refers to electronic data containing text of one or more sentences. Document file 220 is generated in a manner that allows for the retrieval of strings within the file. Regarding at least one of the material database 21 and the auxiliary material database 23, at least a portion of the multiple document files 220 can be generated using optical character recognition (OCR). For example, OCR can be performed on each of multiple printed materials such as books, catalogs, and internal documents to generate document file 220 for each printed material. Alternatively, regarding at least one of the material database 21 and the auxiliary material database 23, at least a portion of the multiple document files 220 can be generated using a text creation application. Regarding document files generated using this application, strings within the file can be retrieved from the outset.
[0034] The materials database 21 stores multiple document files 220 related to various materials. Each document file 220 can record one type of material or two or more types of materials. The auxiliary materials database 23 stores multiple document files 220 related to various auxiliary materials. Each document file 220 can record one type of auxiliary material or two or more types of auxiliary materials. Document files 220 recording one or more types of materials and one or more types of auxiliary materials can be stored in the materials database 21 and the auxiliary materials database 23, respectively.
[0035] Figure 3 This is a diagram illustrating an example of the data structure of supplier database 22. As shown in the diagram, supplier database 22 stores supplier data related to multiple suppliers through data tables. A data table is a format in which data is represented by rows and columns. Each row of a data table corresponds to a data record, and each column of a data table corresponds to a data item. Therefore, supplier data is a collection of more than one data record corresponding to more than one supplier.
[0036] exist Figure 3 In the example, each data record in the supplier data includes a supplier ID and supplier attributes as data items. The supplier ID is a unique identifier that identifies each supplier. Supplier attributes are information representing the characteristics or properties of the supplier, and are represented as a collection of one or more data items.
[0037] In one example, supplier attributes include basic information, material function, compound classification, processing method / equipment, and analytical instrumentation. Basic information may include the supplier's name, address, URL, etc. Material function is material information from a functional perspective, represented by one or more data items. Compound classification is compound information from a structural perspective, represented by one or more data items. Processing method / equipment is information related to the method or equipment used to manufacture the material, represented by one or more data items. Analytical instrumentation is information related to the machine used to analyze the material, represented by one or more data items.
[0038] In one example, for each data item belonging to material function, compound classification, processing method / equipment, and analytical instrument, a binary indicator is set to indicate whether a supplier operates the goods or services shown for that data item. For instance, for a given data item, a "1" is set if a supplier operates the goods or services shown for that data item, and a "0" is set if the supplier does not operate the goods or services shown for that data item. Figure 3 In the example, supplier Sa deals in heat-resistant resins and initiators but not pressure-sensitive adhesives, polymers and oligomers but not phenols, SUS reactions but not GL reaction kilns and ultra-high temperature reactions, and AAS and GC / MS but not ICP / OES.
[0039] [Portal Website] Figure 4 This is a diagram illustrating an example of a portal website. In this example, portal website 300 is named "Development Procurement Website". Portal website 300 includes a homepage 310 that is initially displayed. Homepage 310 includes an input field 311 for handling batch searches of multiple databases and links 321-327 to a search page or external website for performing a search on a specific database.
[0040] Links 321 and 322 are both hyperlinks to search pages used for performing material searches on the materials database 21. Link 321 directs the user terminal 40 to search page 331, which performs searches related to the materials catalog on the materials database 21. Search page 331 includes an input field 331a for accepting search criteria related to the materials catalog and a display field 331b for displaying search results related to the materials catalog. Link 322 directs the user terminal 40 to search page 332, which performs searches related to sustainable materials on the materials database 21. Search page 332 includes an input field 332a for accepting search criteria related to sustainable materials and a display field 332b for displaying search results related to sustainable materials.
[0041] Links 323 and 324 are both hyperlinks to external websites used to perform reagent searches on reagent database 31. Link 323 directs user terminal 40 to the homepage 403 of external website A, which is capable of performing searches on reagent database 31. Link 324 directs user terminal 40 to the homepage 404 of external website B, which is capable of performing searches on reagent database 31. Therefore, links 323 and 324 can be considered external links.
[0042] Link 325 is a hyperlink to the search page used to perform supplier searches on supplier database 22. Link 325 directs user terminal 40 to search page 335, which performs supplier searches. Search page 335 includes an input field 335a for accepting supplier-related search criteria and a display field 335b for supplier-related search results.
[0043] Links 326 and 327 are both hyperlinks to search pages used for performing secondary material searches on the secondary material database 23. Link 326 directs the user terminal 40 to search page 336, which performs searches related to the secondary material catalog on the secondary material database 23. Search page 336 includes an input field 336a for accepting search criteria related to the secondary material catalog and a display field 336b for displaying search results related to the secondary material catalog. Link 327 directs the user terminal 40 to search page 337, which performs searches related to sustainable secondary materials on the secondary material database 23. Search page 337 includes an input field 337a for accepting search criteria related to sustainable secondary materials and a display field 337b for displaying search results related to sustainable secondary materials.
[0044] exist Figure 4 In the example, multiple links and search pages are provided for each of the materials database 21, the auxiliary materials database 23, and the reagent database 31, while one link and search page is provided for the supplier database 22. However, the number of links and search pages may be one for at least one of the materials database 21, the auxiliary materials database 23, and the reagent database 31, while multiple links and search pages may be provided for the supplier database 22.
[0045] Figure 4 The input fields for search criteria and the display fields for search results are simply displayed, but the structure of these fields can be implemented through various user interfaces. For example, checkboxes can be used to create input fields corresponding to supplier attributes represented by flag information.
[0046] [System Actions] refer to Figure 5 The operation of the retrieval system 1 will be explained, and the retrieval method involved in this embodiment will also be explained. Figure 5This is a flowchart representing the actions of the retrieval system 1 as a processing flow S1. The server 10 executes processing flow S1 based on each connection to the materials database 21, supplier database 22, auxiliary materials database 23, and reagent database 31. In the following description, this is to provide the user terminal 40 with... Figure 4 The portal website shown is 300 as a prerequisite.
[0047] In step S11, the page management unit 11 displays the homepage 310 of the portal website 300 on the user terminal 40. The user performs an operation on the user terminal 40 to access the homepage 310. In response to this operation, the user terminal 40 sends a request signal to the server 10 to request the homepage 310. The page management unit 11 receives the request signal, reads the page data of the homepage 310 from the memory 102 according to the signal, and sends the page data as a response to the user terminal 40. The user terminal 40 receives and displays the page data.
[0048] In step S12, the retrieval unit 12 receives operation data representing user actions on the homepage 310 from the user terminal 40. The user can request a batch search by entering search criteria in the input field 311. Alternatively, the user can click or tap one of links 321-327 to select the database to use. The user terminal 40 sends a request signal based on the user action to the server 10, and the retrieval unit 12 receives the request signal.
[0049] As shown in step S13, the subsequent processing is determined based on the type of user operation indicated by the request signal. During batch retrieval, the process proceeds to step S14; when a specific database is selected, the process proceeds to step S16.
[0050] (Batch search) In step S14, the retrieval unit 12 performs a batch retrieval. The retrieval unit 12 retrieves data from a set of two or more databases that correspond to the search criteria entered in the input field 311.
[0051] In one example, the retrieval unit 12 accesses the materials database 21, supplier database 22, and auxiliary materials database 23, which belong to the internal database 20, respectively, and retrieves data from these databases that correspond to the search criteria for the batch search. The search criteria can be a set of more than one keyword.
[0052] The retrieval unit 12 performs a full-text search based on search criteria on multiple document files within the material database 21 and extracts material data corresponding to those search criteria. The retrieval unit 12 also performs a full-text search based on search criteria on multiple document files within the auxiliary material database 23 and extracts auxiliary material data corresponding to those search criteria. In one example, the full-text search is the process of extracting all document files describing one or more keywords indicated by the search criteria. Therefore, both the extracted material data and the auxiliary material data can be a collection of one or more document files.
[0053] The retrieval unit 12 performs a record retrieval based on search criteria in the supplier database 22. In one example, the record retrieval is the process of extracting records that have all data items corresponding to one or more keywords shown in the search criteria. Regarding data items with set flag information, the retrieval unit 12 can determine that the data item corresponds to the keyword when the value of the data item corresponding to the keyword is "1", and determine that the data item does not correspond to the keyword when the value is "0".
[0054] In one example, the retrieval unit 12, in addition to the internal database 20 (materials database 21, supplier database 22, and auxiliary materials database 23), also extracts reagent data corresponding to the search criteria for batch retrieval from the reagent database 31, which belongs to the external database 30. For example, the retrieval unit 12 sends search criteria to an external system that manages the reagent database 31 and requests retrieval processing from that external system. The external system responds to the request by searching the reagent database 31 and sends the search results to the server 10. The retrieval unit 12 receives the search results. Alternatively, the retrieval unit 12 can directly access the reagent database 31 to extract reagent data corresponding to the search criteria. The retrieval method for the reagent database 31 depends on the data structure of the reagent data; for example, it can be a full-text search or a record search.
[0055] The search unit 12 can retrieve data corresponding to the search criteria for batch retrieval from one or more of the following databases: the materials database 21, the supplier database 22, the auxiliary materials database 23, and the reagent database 31. Alternatively, the search unit 12 may not retrieve data corresponding to the search criteria from any database.
[0056] In step S15, the search unit 12 sends the search results to the user terminal 40. The search unit 12 can send results from searches of the materials database 21, the supplier database 22, and the auxiliary materials database 23. Alternatively, in addition to these databases, the search unit 12 can also send results from searches of the reagent database 31. The user terminal 40 receives and displays the search results. The portal website 300 can display the search results on the homepage 310 or in another window.
[0057] (Retrieval from a specific database) As shown in step S16, when a specific database is selected, the subsequent processing is determined based on that database. When internal database 20 is selected, the process proceeds to step S17; when external database 30 is selected, the process proceeds to step S20.
[0058] In step S17, the page management unit 11 displays the search page corresponding to the selected database on the user terminal 40. The page management unit 11 reads page data corresponding to the received request signal from the memory 102 and sends the page data to the user terminal 40. When link 321 is activated, the page management unit 11 sends the page data for search page 331. When link 322 is activated, the page management unit 11 sends the page data for search page 332. When link 325 is activated, the page management unit 11 sends the page data for search page 335. When link 326 is activated, the page management unit 11 sends the page data for search page 336. When link 327 is activated, the page management unit 11 sends the page data for search page 337. The user terminal 40 receives and displays the sent page data.
[0059] In step S18, the retrieval unit 12 performs a search on the selected database. The user requests a search by entering search criteria in the input field of the search page displayed on the user terminal 40 in step S17. These search criteria can be a set of one or more keywords. Based on the user's operation, the user terminal 40 sends a request signal representing the search criteria to the server 10. The retrieval unit 12 receives the request signal, accesses the corresponding database, and retrieves data from the selected database that corresponds to the search criteria.
[0060] When a search condition is entered into the input field 331a of the search page 331, the search unit 12 performs a full-text search based on the search condition on multiple document files stored in the material database 21 and belonging to the material catalog. The search unit 12 can extract one or more document files as material data corresponding to the search condition.
[0061] When a search condition is entered into the input field 332a of the search page 332, the search unit 12 performs a full-text search on multiple document files stored in the material database 21 that belong to sustainable materials, based on the search condition. The search unit 12 can extract one or more document files as material data corresponding to the search condition.
[0062] When a search condition is entered into the input field 335a of the search page 335, the search unit 12 performs a record search on the supplier database 22 based on that search condition. The search unit 12 can extract one or more data records as supplier data corresponding to the search condition.
[0063] When a search condition is entered into the input field 336a on the search page 336, the search unit 12 performs a full-text search based on the search condition on multiple document files stored in the secondary material database 23 and belonging to the secondary material catalog. The search unit 12 can extract one or more document files as secondary material data corresponding to the search condition.
[0064] When a search condition is entered into the input field 337a on the search page 337, the search unit 12 performs a full-text search based on the search condition for multiple document files stored in the secondary material database 23 that belong to sustainable secondary materials. The search unit 12 can extract one or more document files as secondary material data corresponding to the search condition.
[0065] In step S19, the retrieval unit 12 sends the retrieval results to the user terminal 40. The user terminal 40 receives the retrieval results and displays them in the display bar of the corresponding retrieval page.
[0066] In step S20, the page management unit 11 directs the user terminal 40 to an external website corresponding to the selected database, i.e., an external website capable of performing reagent searches on the reagent database 31. As a result of this redirection, the user terminal 40 displays the homepage 403 or 404 of the external website. The user performs a reagent search operation on the external website, and the external website displays the search results for the reagent data in response to the operation.
[0067] After the homepage 310 is displayed on the user terminal 40 in step S11, the user can perform various operations on the homepage 310. For example, the user can repeatedly perform batch searches, or display a different search page on the user terminal 40 to search other databases, or distinguish between batch searches and searches of specific databases as needed. Based on such user operations, the search system 1 can repeatedly execute the processing after step S12.
[0068] In one example, within an organization, various printed materials held by relevant personnel are collected, and OCR is performed on each printed material to obtain multiple document files 220. Furthermore, within this organization, a large number of document files 220 can also be created using a text creation application. These document files 220 are then stored in a materials database 21 or a secondary materials database 23. Additionally, within this organization, supplier data is compiled and stored in a supplier database 22. Thus, various information scattered throughout the organization is aggregated into the retrieval system 1. Furthermore, access to a reagent database 31 belonging to an external database 30 is established. Relevant personnel, as users of the retrieval system 1, can access the compiled information from these prepared databases via a portal 300 to obtain the information they need for materials research or development.
[0069] [Data Management for Materials Database] In one example, the data management system involved in this invention manages a materials database. The data management system is a computer system that generates text-based document data from image data of documents related to materials and stores this document data as material data in the materials database. As mentioned above, a document refers to a record containing one or more sentences, such as a book, catalog, internal document, or other printed material. The image data of a document represents the image of the document's description through a set of pixels. The document data represents the document's description in text form. Therefore, the data management system can also be described as a computer system that transforms at least a portion of the image-based document's description into text form.
[0070] To generate document data, the data management system uses an extraction model, which is a language model for extracting descriptions mapped from images as text data. The data management system not only processes the image data of documents related to the material using the extraction model, but also processes a first dataset representing the descriptions within the documents extracted from the image data through a different processing method than the extraction model. The extraction model also uses the first dataset to extract a second dataset representing the descriptions within the documents as text from the image data. The data management system generates document data based on the second dataset and stores this document data in a material database. Because the extraction model references not only the image data but also the first dataset to extract the second dataset, it is able to extract descriptions within documents from image data with high accuracy. Since document data based on the high-precision extracted descriptions is stored in the material database, data related to the research or development of materials can be retrieved efficiently using this material database.
[0071] (System structure) A data management system consists of one or more computers. When multiple computers are used, they are connected through communication networks such as the Internet and intranets to form a logically unified data management system.
[0072] A computer constituting a data management system typically includes a processor, memory, and communication interfaces as hardware components. The processor is, for example, a CPU. Memory consists of storage devices such as flash memory and hard disks. The various functional modules of the data management system are implemented by the processor executing programs stored in memory.
[0073] The data management program used to enable a computer to function as a data management system includes program code for implementing the various functional modules of the data management system. This data management program can be provided on a tangible recording medium such as a CD-ROM, DVD-ROM, or semiconductor memory, without being temporarily stored. Alternatively, the data management program can be provided via a communication network as a data signal superimposed on a carrier wave. The provided data management program may be stored, for example, in memory.
[0074] Figure 6 This is a diagram illustrating a data management system 5 involved in an example. In this example, the data management system 5 generates document data for storage in the material database 21 of the aforementioned retrieval system 1.
[0075] In this example, the data management system 5 includes a server 50 that provides functions for managing document data. The server 50 is connected to the materials database 21 via a communication network. The communication network associated with the data management system 5 can be configured to include at least one of the Internet and an intranet.
[0076] Server 50 includes processor 501. Processor 501 functions as an acquisition unit 51, a first extraction unit 52, a second extraction unit 53, and a generation unit 54. Acquisition unit 51 is a functional module that acquires image data of documents related to materials. First extraction unit 52 is a functional module that extracts a first dataset from the image data. Second extraction unit 53 is a functional module that processes the image data and the first dataset using extraction model 55, and extracts a second dataset from the image data. Generation unit 54 is a functional module that uses processing model 56 to generate document data based on the first and second datasets, and stores the generated document data in material database 21. Processing model 56 is a language model that integrates at least two datasets, each represented in text form, to generate other datasets.
[0077] The language models extracted (model 55) and organized (model 56) are examples of generative AI (Generative Artificial Intelligence). A language model is a technique that processes input data to generate output data representing predictions, responses, etc. This input data is also called cue words. Language models are constructed using neural network models, such as deep learning models. As an example of a language model, a large-scale language model (LLM) can be described, built using large datasets and deep learning, scaling up the three elements of computation, data volume, and the number of model parameters. Another example is a visual language model or a large-scale visual language model that integrates visual data (images) and text data to generate output data. In one example, both extracted model 55 and organized model 56 are achieved by fine-tuning a pre-learned language model. Fine-tuning refers to the technique of retraining the learned model using additional datasets and fine-tuning the parameters of the learned model to obtain the desired output.
[0078] (Storage of document data) refer to Figure 7 The operation of the data management system 5 will be explained, and the data management method involved in this embodiment will also be explained. Figure 7 This is a flowchart representing an example of the actions of the data management system 5 as a processing flow S2.
[0079] In step S21, the acquisition unit 51 acquires image data of documents related to the material. The acquisition unit 51 may read image data from a designated storage device in response to an instruction from a user of the data management system 5, or it may accept image data input by the user. Alternatively, the acquisition unit 51 may receive image data from another computer.
[0080] Figure 8 This is a diagram illustrating an example of image data related to materials. In this example, image data 600 includes descriptions of major categories 601, medium categories 602, and minor categories 603 as classifications related to materials. Furthermore, image data 600 contains descriptions of products 604 belonging to any of these categories. These descriptions can be extracted by the data management system 5.
[0081] Return to Figure 7 In step S22, the first extraction unit 52 performs OCR on the image data to extract the first dataset from the image data. OCR is an example of a process different from the extraction model 55.
[0082] In step S23, the second extraction unit 53 causes the extraction model 55 to process extraction input data including image data and the first dataset, and extracts the second dataset from the image data. The extraction input data also includes instruction information indicating instructions to the extraction model 55. The second extraction unit 53 inputs the extraction input data to the extraction model 55. The extraction model 55 receives the extraction input data, refers to the instructions and the first dataset, and extracts the second dataset from the image data. The extraction model 55 outputs the second dataset, and the second extraction unit 53 acquires the second dataset.
[0083] Figure 9 This is a diagram illustrating an example of input data for extraction. In this example, the input data 700 for extraction includes instruction information 710 for extracting the description of a document from image data. Instruction information 710 includes instructions 711 to 713. Instruction 711 is an instruction for generating a summary of the document's description. Instruction 712 is an instruction for retrieving material-related information from the document's description through semantic retrieval. Instruction 713 is an instruction for extracting material information without extracting prescribed unnecessary information.
[0084] When the extraction model 55 processes the extraction input data 700 in the second extraction unit 53, the extraction model 55, in response to instruction 711, extracts a summary of the document's description represented by the image data from the image data. Furthermore, in response to instruction 712, the extraction model 55 performs semantic retrieval on the document's description and extracts the material information retrieved through this semantic retrieval from the image data. Semantic retrieval refers to a retrieval technique that seeks information highly relevant to the meaning of the retrieval query. By using semantic retrieval, for example, even if there are more than two terms for a specific material due to label differences, these terms can be extracted as information related to that material. Moreover, in response to instruction 713, the extraction model 55 extracts material information from the image data without extracting unnecessary information. Figure 9 In the example, the company's location information was specified as unnecessary information. Unnecessary information can be information unrelated to the material information or information with low relevance to the material information. Other examples of unnecessary information include product delivery or packaging methods, information about information windows, etc. As a result of these processes, extraction model 55 extracts a second dataset from the image data, containing the summary extracted according to instruction 711, the material information extracted according to instruction 712, and the material information extracted according to instruction 713, and outputs this second dataset.
[0085] exist Figure 9In the example, the input data 700 for extraction includes three types of instructions 711 to 713. However, the input data for extraction may include one or two of these three types of instructions, or may not include the remaining types of instructions. Alternatively, the input data for extraction may replace at least some of the three types of instructions or include another type of instruction in addition to at least some of the three types of instructions.
[0086] Return to Figure 7 In step S24, the generation unit 54 causes the sorting model 56 to process the sorting input data containing the first dataset and the second dataset, and generates document data from the first dataset and the second dataset. The sorting input data also includes instruction information indicating instructions to the sorting model 56. The generation unit 54 inputs the sorting input data into the sorting model 56. The sorting model 56 receives the sorting input data, refers to the first dataset and the second dataset, and generates document data from these datasets. The sorting model 56 outputs the document data, and the generation unit 54 acquires the document data.
[0087] Figure 10 This is a diagram illustrating an example of processing input data. In this example, the processing input data 800 includes instruction information 810 for integrating two datasets in text form to generate a new dataset. Instruction information 810 includes instructions 811 for extracting material-related search tags. These search tags are used when retrieving material information. Material-related search tags can be keywords representing the material's name, type, use, manufacturer, model, etc. When the generation unit 54 causes the processing model 56 to process the processing input data 800, the processing model 56, in response to instruction 811, generates document data containing the search tags. The document data contains at least a portion of the description of the document represented by image data.
[0088] like Figure 10 As shown, the input data for organizing can include other types of instructions besides those related to the search tags. Alternatively, the input data for organizing can also include another type of instruction instead of those related to the search tags.
[0089] Figure 11 This is a diagram representing an example of document data. The document data 900 shown in this example is from... Figure 8The data generated from the image data 600 shown. Document data 900 contains descriptions of major categories 901, intermediate categories 902, minor categories 903, and products 904, organized in a hierarchical manner reflecting the positional relationships within the document. The major categories 901, intermediate categories 902, minor categories 903, and products 904 in document data 900 correspond to the major categories 601, intermediate categories 602, minor categories 603, and products 604 in image data 600, respectively. A portion of the descriptions reflected in image data 600 can be extracted from document data 900. Alternatively, document data 900 can contain descriptions generated by rewriting a portion of words or phrases reflected in image data 600 while preserving their meaning.
[0090] Return to Figure 7 In step S25, the generation unit 54 stores the generated document data in the material database 21. For example, the generation unit 54 determines a subdirectory 210 corresponding to the document data and stores the document data as a document file 220 in that subdirectory 210.
[0091] (Retrieving from materials databases) The retrieval system 1 can retrieve document data stored in the material database 21 through the data management system 5 as follows: First, the retrieval unit 12 receives material-related search criteria from the user terminal 40. As described above, these search criteria can be a set of one or more keywords entered into input fields 311, 331a, or 332a. The retrieval unit 12 performs retrieval processing on the material database 21 based on the received search criteria. The retrieval unit 12 can perform retrieval processing based on at least one of vector-based retrieval and keyword retrieval. The retrieval unit 12 sends the results of the retrieval processing (retrieval results) to the user terminal 40. As described above, the user terminal 40 receives and displays the retrieval results.
[0092] [Variation Example] The technology involved in this invention has been described in detail above with reference to various examples. However, this invention is not limited to the above examples. Various modifications can be made to the technology involved in this invention without departing from its spirit.
[0093] The aforementioned retrieval system 1 is constructed as a client-server system, but the retrieval system can also be installed on a standalone computer.
[0094] In the aforementioned data management system 5, the first extraction unit 52 extracts the first dataset from the image data using OCR. However, the first extraction unit 52 may also use methods other than OCR as a different processing method than the extraction model. Alternatively, the data management system may obtain the first dataset from another computer.
[0095] In the aforementioned data management system 5, the generation unit 54 generates document data using the organization model 56. However, the generation unit can also generate document data using methods other than the organization model. For example, the generation unit can generate document data based on the second dataset instead of the first dataset. The generation unit can organize the second dataset into a prescribed format and store the organized second dataset as document data in the database. Alternatively, the generation unit can directly store the second dataset as document data in the database. Thus, the method for generating document data can be various, but in any case, the generation unit will store the document data based on the second dataset in the database.
[0096] The processing order of a method executed by at least one processor is not limited to the example above. For example, some of the steps described above may be omitted, or the steps may be executed in a different order. Furthermore, any two or more of the steps described above may be combined, or some of the steps may be modified or deleted. Alternatively, other steps may be executed in addition to those described above.
[0097] In the comparison of the magnitude of two values in this invention, either "above" or "greater" can be used, or either "below" or "less than" can be used.
[0098] In this invention, the statement "at least one processor executes the first process, executes the second process, ... executes the nth process," or its corresponding statement, represents the concept of a situation where the processor, the executing entity of the n processes from the first process to the nth process, changes midway. That is, this statement represents the concept of two scenarios: one where all n processes are executed by the same processor, and another where the processor changes arbitrarily among the n processes.
[0099] [appendix] As can be understood from the above examples, the present invention includes the following methods.
[0100] (Appendix 1) A data management system having at least one processor, The at least one processor performs the following processing: Acquire image data from documents related to the materials; A language model, or extraction model, is used to extract data from images by reproducing descriptions in the images as text. This model processes the image data and input data for extraction, including the first dataset representing the descriptions within the document extracted from the image data in text form through processing different from the extraction model. It also extracts a second dataset representing the descriptions within the document in text form from the image data. Document data based on the second dataset will be stored in the database.
[0101] (Appendix 2) According to the data management system described in Appendix 1, wherein, The at least one processor performs the following processing: using optical character recognition as a processing different from the extraction model, it extracts the first dataset from the image data.
[0102] (Appendix 3) According to the data management system described in Appendix 1 or 2, wherein, The input data for extraction includes instructions for generating a summary of the description of the document represented by the image data. The at least one processor extracts the second dataset from the image data, which contains a summary of the description of the document.
[0103] (Appendix 4) According to any one of Appendices 1 to 3, in the data management system, wherein, The input data for extraction includes instructions for retrieving material information related to the material from the description of the document represented by the image data through semantic retrieval. The at least one processor extracts the second dataset from the image data, which contains the material information retrieved through the semantic retrieval.
[0104] (Appendix 5) According to any one of Appendices 1 to 4, in the data management system, The input data for extraction includes instructions for extracting material information related to the material without extracting prescribed unnecessary information. The at least one processor extracts the second dataset from the image data, which includes the material information but excludes the specified unnecessary information.
[0105] (Note 6) According to any one of Appendices 1 to 5, in the data management system, The at least one processor performs the following processing: causing the organization model to process the organization input data containing the first dataset and the second dataset to generate the document data, wherein the organization model is a language model that integrates at least two datasets, each represented in text form, to generate other datasets.
[0106] (Note 7) According to the data management system described in Appendix 6, wherein, The input data for processing includes instructions for extracting material-related search tags used when retrieving material information related to the material. The at least one processor generates the document data containing the search tags.
[0107] (Postscript 8) A retrieval system comprising at least one processor, wherein, The at least one processor performs the following processing: Receive search criteria related to the material from the user terminal; The data management system described in any one of Appendices 1 to 7 performs a retrieval process on the database storing the document data, based on the retrieval conditions, using at least one of vector-based retrieval and keyword retrieval; and The results of the retrieval process are sent to the user terminal.
[0108] (Note 9) A data management method, which is executed by a data management system having at least one processor. The data management method includes the following steps: Acquire image data from documents related to the materials; A language model, or extraction model, is used to extract data from images by reproducing descriptions in the images as text. This model processes the image data and input data for extraction, including the first dataset representing the descriptions within the document extracted from the image data in text form through processing different from the extraction model. It also extracts a second dataset representing the descriptions within the document in text form from the image data. Document data based on the second dataset will be stored in the database.
[0109] (Postscript 10) A data management program that causes a computer to perform the following steps: Acquire image data from documents related to the materials; A language model, or extraction model, is used to extract data from images by reproducing descriptions in the images as text. This model processes the image data and input data for extraction, including the first dataset representing the descriptions within the document extracted from the image data in text form through processing different from the extraction model. It also extracts a second dataset representing the descriptions within the document in text form from the image data. Document data based on the second dataset will be stored in the database.
[0110] According to Appendices 1, 9, and 10, when the extraction model extracts a second dataset in text form from image data of documents related to the material, a first dataset in text form extracted from the image data through a different processing method is also provided to the extraction model. The extraction model extracts the second dataset by referencing not only the image data but also the first dataset, thus enabling high-precision extraction of descriptions within documents from the image data. Since the document data based on the high-precision extracted descriptions is stored in a database, data related to the research or development of the material can be retrieved efficiently using this database.
[0111] According to Appendix 2, Dataset 1 was extracted from image data using Optical Character Recognition (OCR). By using OCR, strings within documents can be extracted with high accuracy, which serves as Dataset 1. Using Dataset 1, descriptions within documents can be extracted from image data with even higher accuracy using an extraction model.
[0112] According to Appendix 3, a second dataset containing summaries of document descriptions is extracted from image data. Using this second dataset, document data consisting of the key points of the document can be generated. By using a database storing this document data, data related to the research or development of materials can be retrieved more efficiently. Furthermore, by limiting the size of the document data, storage space within the database used to store the document data can be saved.
[0113] According to Appendix 4, a second dataset containing material information retrieved through semantic retrieval is extracted from image data. By using semantic retrieval, even if there are discrepancies in the descriptions related to the same material, material information can be obtained by absorbing these differences. By using this second dataset containing the material information thus obtained, document data can be generated while preventing or suppressing the loss of information about the material. By using a database storing this document data, data related to the research or development of materials can be retrieved more efficiently.
[0114] According to Appendix 5, a second dataset containing material information but excluding unnecessary information is extracted from image data. Using this second dataset, document data with organized content can be generated. By using a database storing this document data, data related to material research or development can be retrieved more efficiently. Furthermore, by limiting the size of the document data, storage space within the database used to store the document data can be saved.
[0115] According to Appendix 6, document data is generated by integrating the first and second datasets through model organization. This process allows for the generation of document data that complements the content of both datasets, thus enriching the information within the document data. By using a database storing this document data, data related to the research or development of materials can be retrieved more efficiently.
[0116] According to Note 7, by storing document data containing search tags used when retrieving material information in a database, material information can be retrieved accurately or efficiently. Therefore, data related to material research or development can be retrieved more efficiently.
[0117] According to Appendix 8, a retrieval process based on at least one of vector retrieval and keyword retrieval is performed on a database storing document data generated from descriptions extracted with high precision through a data management system. This mechanism enables the acquisition of highly accurate retrieval results; that is, it allows for the efficient retrieval of data related to the research or development of materials.
[0118] Symbol Explanation 1-Retrieval System, 5-Data Management System, 10-Server, 11-Page Management Department, 12-Retrieval Department, 20-Internal Database, 21-Materials Database, 22-Supplier Database, 23-Supply Material Database, 30-External Database, 31-Reagent Database, 40-User Terminal, 50-Server, 51-Acquisition Department, 52-First Extraction Department, 53-Second Extraction Department, 54-Generation Department, 55-Extraction Model, 56-Processing Model, 220-Document Files, 300-Portal Website, 600-Image Data, 700-Input Data for Extraction, 800-Input Data for Processing, 900-Document Data.
Claims
1. A data management system having at least one processor, The at least one processor performs the following processing: Acquire image data from documents related to the materials; A language model, or extraction model, is used to extract data from images by reproducing descriptions in the images as text. This model processes the image data and input data for extraction, including the first dataset representing the descriptions within the document extracted from the image data in text form through processing different from the extraction model. It also extracts a second dataset representing the descriptions within the document in text form from the image data. Document data based on the second dataset will be stored in the database.
2. The data management system according to claim 1, wherein, The at least one processor performs the following processing: using optical character recognition as a processing different from the extraction model, it extracts the first dataset from the image data.
3. The data management system according to claim 1 or 2, wherein, The input data for extraction includes instructions for generating a summary of the description of the document represented by the image data. The at least one processor extracts the second dataset from the image data, which contains a summary of the description of the document.
4. The data management system according to claim 1 or 2, wherein, The input data for extraction includes instructions for retrieving material information related to the material from the description of the document represented by the image data through semantic retrieval. The at least one processor extracts the second dataset from the image data, which contains the material information retrieved through the semantic retrieval.
5. The data management system according to claim 1 or 2, wherein, The input data for extraction includes instructions for extracting material information related to the material without extracting prescribed unnecessary information. The at least one processor extracts the second dataset from the image data, which includes the material information but excludes the specified unnecessary information.
6. The data management system according to claim 1 or 2, wherein, The at least one processor performs the following processing: causing the organization model to process the organization input data containing the first dataset and the second dataset to generate the document data, wherein the organization model is a language model that integrates at least two datasets, each represented in text form, to generate other datasets.
7. The data management system according to claim 6, wherein, The input data for processing includes instructions for extracting material-related search tags used when retrieving material information related to the material. The at least one processor generates the document data containing the search tags.
8. A retrieval system comprising at least one processor, wherein, The at least one processor performs the following processing: Receive search criteria related to the material from the user terminal; The data management system according to claim 1 or 2 performs a retrieval process on the database storing the document data according to the retrieval conditions, based on at least one of vector retrieval and keyword retrieval. and The results of the retrieval process are sent to the user terminal.
9. A data management method, which is executed by a data management system having at least one processor. The data management method includes the following steps: Acquire image data from documents related to the materials; A language model, or extraction model, is used to extract data from images by reproducing descriptions in the images as text. This model processes the image data and input data for extraction, including the first dataset representing the descriptions within the document extracted from the image data in text form through processing different from the extraction model. It also extracts a second dataset representing the descriptions within the document in text form from the image data. Document data based on the second dataset will be stored in the database.
10. A data management program that causes a computer to perform the following steps: Acquire image data from documents related to the materials; A language model, or extraction model, is used to extract data from images by reproducing descriptions in the images as text. This model processes the image data and input data for extraction, including the first dataset representing the descriptions within the document extracted from the image data in text form through processing different from the extraction model. It also extracts a second dataset representing the descriptions within the document in text form from the image data. Document data based on the second dataset will be stored in the database.
Citation Information
Patent Citations
Research and development support system
JP2020052602A
