Knowledge management system, document classification apparatus, work restructuring support system, expert system generation apparatus, and document retrieval system
The integration of a large-scale language model in a knowledge management system facilitates internal technical information searches and business reform by aggregating and visualizing data, addressing the lack of effective utilization in existing technologies.
Patent Information
- Application Number
- JP2023071727
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-04-25
- Publication Date
- 2024-11-07
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Large-scale language models have not been effectively utilized for internal technical information searches within companies.
A knowledge management system utilizing a large-scale language model to store and search project attributes, employee data, and customer information, with features like project search, employee search, and document classification, supported by a business reform system that aggregates and visualizes information from existing systems.
Enables efficient search and classification of technical information within a company, promoting knowledge sharing and business improvement by leveraging existing systems and enhancing data utilization.
Smart Images

Figure 2024157390000001_ABST
Abstract
Description
[Technical field]
[0001] The present invention relates to a knowledge management system, a document classification device, a business process improvement support system, an expert system generation device, and a document search system. [Background technology]
[0002] Large-scale language models (LLMs) such as ChatGPT (registered trademark) are known. Large-scale language models can realize various natural language processing tasks such as text classification, sentiment analysis, information extraction, text summarization, text generation, and question answering. [Prior art documents] [Patent documents]
[0003] [Patent Document 1] JP 2023-033202 A Summary of the Invention [Problem to be solved by the invention]
[0004] However, in conventional techniques, large-scale language models have not been utilized for technical information searches within companies.
[0005] The present disclosure provides a technique for utilizing large-scale language models for internal technical information search within a company. [Means for solving the problem]
[0006] One aspect of the present disclosure is a knowledge management system comprising: a first storage means for storing project attributes including attributes related to the project's technical field and attributes related to the project's customer industry, and project data including identification information of employees who participated in the project; a second storage means for storing employee data including identification information of employees; a reception means for receiving input of search conditions for project data; a project search means for searching the first storage means using a large-scale language model based on the search conditions received by the reception means; a first output means for outputting the project data searched by the project search means; an employee search means for searching the second storage means using a large-scale language model based on the identification information of employees included in the project data searched by the project search means; and a second output means for outputting the employee data searched by the employee search means.
[0007] One aspect of the present disclosure includes an input / output device having a display for inputting and outputting data, a design information database for saving and reading documents, a keyword extraction means for acquiring keywords from documents displayed or executed on the display by using a large-scale language model, a task list associating tasks with related words, a classification means for classifying documents stored in the design information database based on the tasks listed in the task list, a classification result database for saving and calling up the classification results by the classification means, a business process template database for saving and calling up business process template data in which each business procedure, its deliverables, and reference materials are associated with each other and their storage addresses, and a processing unit for reading the business process template data stored in the business process template database. a task support means for supporting the user in classifying stored documents into predetermined categories and presenting them, and a task support means for supporting the user in classifying stored documents into predetermined categories and presenting them, the task support means comprising: a task list table for listing keywords extracted by the keyword extraction means; a task list table for listing tasks that match; a task list table for listing tasks that match; a task list table for listing tasks that match; a task list table for listing tasks that match; a task list table for listing tasks that match; a task list table for listing tasks that match; a task list table for listing tasks that match; a task list table for listing tasks that match; a task list table for listing tasks that match; a task list table for listing tasks that match; a task list table for listing tasks that match; a task list table for listing tasks that match; a task list table for listing tasks that match; a task list table for listing tasks that match; a task list table for listing tasks that match;
[0008] One aspect of the present disclosure is a business reform support system that utilizes existing systems within a company, which is characterized by data integration with host computers, including personal computers, and a data warehouse / integrated database (DB) currently in operation within the company, and data integration with existing systems, channels, and external information containing various types of data as input information sources, in which a large-scale language model centrally aggregates the various pieces of information scattered across the existing systems, channels, external information, host computer, and data warehouse / integrated DB based on extraction conditions to extract insights, and by visualizing the extracted insights, improves the company's business operations.
[0009] One aspect of the present disclosure is an expert system generation device for building an expert system using a large-scale language model, comprising: a case acquisition means for accumulating situations and associated results observed from an external system for which the expert system is to be built; a rule generation means for automatically learning the accumulated situations and results and deriving, as knowledge, a statistical tendency for the result to arise from the situation; and a consulting means for automatically determining a result for an observed situation from the derived statistical tendency, characterized in that the case acquisition means, the rule generation means, and the consulting means are operated independently or in conjunction with each other to update the situation and the result, automatically updating them to more appropriate knowledge.
[0010] One aspect of the present disclosure is a document retrieval system including a document retrieval device that searches for documents using keywords, a translation service device, and a document database that stores a plurality of documents to be searched, the document retrieval device including a keyword receiving means that receives one or more keywords as input keywords, a keyword translation means that acquires translated keywords for a plurality of other languages in which each of the input keywords is translated into other languages using a large-scale language model, a keyword score determination means that determines a keyword score for each of the input keywords and the translated keywords, a document retrieval means that searches for documents based on the input keywords and the translated keywords, and acquires a plurality of search result documents, and a document score calculation means that calculates a document score for each of the search result documents based on the keyword score. and a search result output means for outputting each of the search result documents in association with the corresponding document score, wherein the translation service device generates a plurality of translation keywords having a ranking corresponding to each of the input keywords, the translation service device represents the ranking of the translation keywords based on the order in which the translation keywords are listed, the keyword score determination means determines a translation score for all combinations of each of the input keywords and each of the translation keywords based on the ranking, the keyword score determination means determines the keyword score for each of the translation keywords based on all of the associated translation scores, and the keyword score of the input keyword is higher than any of the keyword scores of the translation keywords corresponding to the input keyword. Effect of the Invention
[0011] The present disclosure enables large-scale language models to be utilized for internal technical information search within a company. [Brief description of the drawings]
[0012] [Figure 1] 1 is a diagram showing an overall configuration of a knowledge management system according to an embodiment of the present invention. [Diagram 2] 2 shows an example of data stored in a database server. [Diagram 3] An example of the project registration / editing screen is shown below. [Figure 4] 1 illustrates an example of a data structure of a project table and a project attribute table. [Diagram 5] An example of the contact registration / editing screen is shown. [Figure 6] 3 shows an example of the data structure of a contact table and a contact attribute table. [Figure 7] 13 shows an example of a course list screen. [Figure 8] 1 shows an example of the data structure of an employee table, a course table, and a course application table. [Figure 9] An example of the customer registration / editing screen is shown below. [Figure 10] 2 illustrates an example of the data structure of a customer table and a customer attribute table. [Figure 11] An example of the new agent registration / editing screen is shown. [Figure 12] 1 shows a data structure of an agent table. [Figure 13] 13 shows an example of a data structure of a forum management table. [Figure 14] 13 illustrates an example of a data structure of a community management table. [Figure 15] 13 shows an example of a project search screen. [Figure 16] 13 shows an example of a contact report search screen. [Figure 17] 1 is a block diagram showing a system configuration of the present embodiment. [Figure 18] FIG. 2 is an explanatory diagram illustrating a business process template configuration. [Figure 19] FIG. 13 is an explanatory diagram illustrating a display of a predecessor-successor relationship of a business process. [Figure 20] FIG. 11 is an explanatory diagram illustrating an example of a business process template execution. [Figure 21] FIG. 11 is a flow diagram showing a classification process of design information. [Figure 22] FIG. 11 is a flow diagram showing a classification process of design information. [Figure 23] FIG. 11 is a flow diagram showing a classification process of design information. [Figure 24] FIG. 13 is an explanatory diagram showing an example of an array in which the title and destination address of each document are stored. [Diagram 25] FIG. 13 is an explanatory diagram showing a task list. [Figure 26] FIG. 11 is an explanatory diagram showing an example of a reference material. [Figure 27] FIG. 11 is an explanatory diagram showing a classification result. [Figure 28] FIG. 1 is an explanatory diagram illustrating a conventional search system. [Figure 29] FIG. 11 is an explanatory diagram showing an example of product registration. [Diagram 30] FIG. 13 is an explanatory diagram showing an example of progress information display. [Diagram 31] FIG. 11 is an explanatory diagram showing an example of progress information registration. [Diagram 32] FIG. 11 is an explanatory diagram showing an example of progress information registration. [Diagram 33] FIG. 13 is an explanatory diagram illustrating an editing menu of a business process template. [Diagram 34] FIG. 13 is an explanatory diagram illustrating a new addition screen for a business process name. [Diagram 35] FIG. 13 is an explanatory diagram illustrating an editing screen for a business process name. [Diagram 36] FIG. 4 is a diagram illustrating an example of a filter table used in the present embodiment. [Figure 37] FIG. 13 is a diagram showing a method for acquiring data from Excel (all cells method). [Figure 38] FIG. 13 is a diagram showing a method for acquiring data from Excel (designated cell method). [Figure 39] FIG. 13 is a diagram showing a method for acquiring data from Word. [Diagram 40] 1 is a block diagram showing an example of a basic overall configuration of an embodiment of the present invention; [Diagram 41] 1 is a block diagram showing an example of the configuration of a business process reform support system using an existing system within a company according to an embodiment of the present invention. [Diagram 42] FIG. 1 is a conceptual diagram of an awareness information construction system. [Diagram 43] 1 is a flow block diagram showing an example of the configuration of an awareness information construction system and the flow of information. [Diagram 44] FIG. 1 is a block diagram showing an example of the configuration of an awareness information construction system. [Diagram 45] 13 is a flowchart illustrating an example of the operation of an input linkage interface. [Figure 46] 13 is a flowchart showing an example of the operation of a notice information extraction unit. [Figure 47] 13 is a flowchart showing an operation example of an output linkage interface. [Figure 48] 13 is a screen diagram showing a specific example of an observation information extraction unit. FIG. [Figure 49] 11 is a data flow diagram showing an example of extraction of notice information. [Figure 50] FIG. 4 is a workflow diagram showing a part of an operation example of the present embodiment. [Figure 51] FIG. 4 is a workflow diagram showing a part of an operation example of the present embodiment. [Figure 52] FIG. 4 is a workflow diagram showing a part of an operation example of the present embodiment. [Diagram 53] FIG. 4 is a workflow diagram showing a part of an operation example of the present embodiment. [Figure 54] 11A to 11C are part of screen diagrams showing screen changes in each scene in an operation example of this embodiment. [Figure 55] 11A to 11C are part of screen diagrams showing screen changes in each scene in an operation example of this embodiment. [Figure 56] 11A to 11C are part of screen diagrams showing screen changes in each scene in an operation example of this embodiment. [Figure 57] 11A to 11C are part of screen diagrams showing screen changes in each scene in an operation example of this embodiment. [Figure 58] 11A to 11C are part of screen diagrams showing screen changes in each scene in an operation example of this embodiment. [Figure 59] 11A to 11C are part of screen diagrams showing screen changes in each scene in an operation example of this embodiment. [Figure 60]FIG. 1 is a schematic diagram showing a future outlook for the present embodiment. [Figure 61] FIG. 2 is a block diagram of an expert system generating device according to an embodiment. [Figure 62] FIG. 2 is an explanatory diagram of the function of a consultation tree generation block. [Figure 63] FIG. 2 is an explanatory diagram of the function of a tree check block. [Figure 64] FIG. 2 is a block diagram of a consulting means. [Figure 65] 1 is a flowchart for generating a decision tree. [Figure 66] 13 is a detailed flowchart of step S300. [Figure 67] 13 is a detailed flowchart of step S305. [Figure 68] 13 is a detailed flowchart of step S3051. [Figure 69] 13 is a detailed flowchart of step S30511. [Figure 70] 13 is a detailed flowchart of step S3055. [Figure 71] 13 is a detailed flowchart of step S3056. [Figure 72] 13 is a detailed flowchart of step S30561. [Figure 73] 13 is a detailed flowchart of step S400. [Figure 74] 13 is a detailed flowchart of step S402. [Figure 75] 13 is a detailed flowchart of step S403. [Figure 76] 1 is a flowchart regarding consulting. [Figure 77] 13 is a detailed flowchart of step S700. [Figure 78] 13 is a detailed flowchart of step S701. [Figure 79] 13 is a detailed flowchart of step S702. [Figure 80] FIG. 13 illustrates a case editing panel. [Figure 81] 1 is a display screen showing a generated decision tree. [Figure 82] 13 is a display screen showing the decision tree after pruning is completed. [Figure 83] This is a display screen showing a decision tree obtained for the task of calcining test pieces of incineration ash from sludge treatment. [Figure 84] 11A and 11B show a screen display before and after exception handling of an element. [Figure 85] 1 is a portion of a display showing a decision tree prior to exception of a date element. [Figure 86] 13 is a portion of a display showing a decision tree after the date element has been excluded. [Figure 87] FIG. 1 is a schematic diagram of a recovered solvent refining plant. [Figure 88] FIG. 1 is a diagram showing a configuration of a document search system. [Figure 89] 89 is a flowchart illustrating the operation of the document retrieval device in the document retrieval system of FIG. 88. [Figure 90] FIG. 13 is a diagram showing an example of a correspondence relationship between an input keyword and a translation keyword. [Figure 91] 13 is a diagram showing an example of a correspondence relationship between the ranking of translation keywords and the translation score based on the ranking. FIG. [Figure 92] FIG. 13 is a diagram showing an example of the correspondence between the translation score based on the ranking for each keyword and the keyword score finally assigned to each keyword. [Figure 93] 11 is a diagram showing an example of information indicating the number of times each keyword appears in the text data of a search result document. FIG. [Figure 94] FIG. 13 is a diagram showing an example of a calculation result of a document score for a search result document. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0013] In the following, five embodiments will be described in which large-scale language models such as ChatGPT are used in technical information searches within a company. The five embodiments are use as a knowledge management system, automatic classification of technical information, automation of duties and tasks, an expert system (a system that incorporates the knowledge and experience of experts into software to automatically make judgments and diagnoses), and support for multiple languages. However, examples of use of technical information searches within a company are not limited to these. In addition to ChatGPT, there are various other large-scale language models, such as Dolly & Dolly 2.0, BLOOM, PaLM, and OPT (all of which are registered trademarks).
[0014] [First embodiment] In this embodiment, an embodiment using a large-scale language model as a search engine in a knowledge management system will be described. As an example of a knowledge management system, the description will be given with reference to Japanese Patent Application Laid-Open No. 2013-109516.
[0015] Hereinafter, a knowledge management system according to an embodiment of the present invention will be described with reference to the drawings.
[0016] 1 is a diagram showing the overall configuration of a knowledge management system according to this embodiment. In this system, a database server 1001, an authentication server 1003, one or more user terminal devices 1005, and an administrator terminal device 1007 are connected to each other via a network 1009.
[0017] The database server 1001, the authentication server 1003, one or more user terminal devices 1005, and the administrator terminal device 1007 are all configured, for example, by a general-purpose computer system, and the individual components or functions of the database server 1001, the authentication server 1003, the user terminal device 1005, and the administrator terminal device 1007 described below are realized, for example, by executing a computer program. The user terminal device 1005 and the administrator terminal device 1007 may be realized on different computers or on the same computer.
[0018] The authentication server 1003 authenticates users and administrators who access this system. For example, the authentication server 1003 receives an authentication request from a user terminal device 1005 and authenticates the user who uses the user terminal device 1005. The authentication server 1003 also receives an authentication request from an administrator terminal device 1007 and authenticates the administrator who uses the administrator terminal device 1007. The users and administrators authenticated by the authentication server 1003 can access the database server 1001 within the scope of their respective given authorities.
[0019] The user terminal device 1005 has a predetermined input device and output device. The user terminal device 1005 provides various interfaces for the user to access data in the database server 1001. For example, each registration / editing screen described below is displayed on each user terminal device 1005, and the user can input to the screen. Also, each user terminal device 1005 displays search results in response to a search request from each search screen described below.
[0020] The administrator terminal device 1007 has a predetermined input device and output device, and provides the system administrator with various interfaces for system management.
[0021] The database server 1001 stores various data. This system mainly has data on projects ordered by customers, data on contacts with customers (potential customers) in sales activities aimed at receiving orders for the projects, data on employees, data on customers, and other data. The project-related data, contact data, employee data, and customer data are all associated with one another.
[0022] 2 shows an example of data stored in the database server 1001. The database server 1001 has a project table 1010 and a project attribute table 1011 as data related to projects, a contact table 1014 and a contact attribute table 1015 as data related to contacts, an employee table 1018, a course table 1019, and a course application table 1020 as data related to employees, and a customer table 1024 and a customer attribute table 1025 as data related to customers. The database server 1001 further has an agent table 1028 related to agents, a forum management table 1030 and a forum data storage unit 1031 related to forums, and a community management table 1034 and community data 1035 related to communities as other data. The database server 1001 further has a search engine 1040 (an application example of a large-scale language model) for searching each of the above data.
[0023] <Project related data> Fig. 3 shows an example of the project registration / editing screen 2000. Fig. 4 shows an example of the data structure of the project table 1010 and the project attribute table 1011 (the project table 1010 and the project attribute table 1011 are examples of a first storage means).
[0024] When an order for a project is received from a client, project-related data is registered on the project registration / edit screen 2000 before or after the start of the project. The data entered here is saved in the project table 1010 and the project attribute table 1011. The project-related data includes project attributes, including attributes related to the technical field of the project and attributes related to the industry of the client of the project, and identification information of employees who participated in the project. The data saved in the project table 1010 and the project attribute table 1011 can be updated from the project registration / edit screen 2000 at any time thereafter. Note that the input items and data items shown here are merely examples, and other input items and data items may be included, and some items may be omitted.
[0025] As shown in FIG. 3, the project registration / editing screen 2000 has input areas for each of the following items: project ID 2001, customer name 2002, contract ID 2003, start date 2004, end date 2005, other party person in charge 2006, potential theme 2007, industry 2008, member 2009, forum name 2010, amount 2011, and contact ID 2012.
[0026] The start date 2004 and the end date 2005 are the start date and the end date of the project, respectively. At the time of registration, the planned start date and the planned end date are registered, and may be corrected to the actual start and end dates afterwards. In the other party person in charge 2006, data related to the person in charge on the customer side is set. In the theme candidate 2007, a theme code indicating the technical field of the project is set. In the industry 2008, an industry code indicating the industry of the customer is set. In the member 2009, identification information (employee ID, name, etc.) of employees participating in the project is input. In the forum name 2010, a forum name set for each project is set. In the amount 2011, the contract amount of the project. In the contact ID 2012, an ID associated with contact-related data described later.
[0027] Fig. 4A shows an example of the data structure of the project table 1010, and Fig. 4B shows an example of the data structure of the project attribute table 1011. The project table 1010 and the project attribute table 1011 store data related to projects ordered by customers.
[0028] The project table 1010 has, as data items, a project ID 1100, a contract ID 1102, a start date 1103, an end date 1104, an amount 1105, and a contact ID 1106. In each data item of the project table 1010, data inputted in the project ID 2001, the customer name 2002, the contract ID 2003, the start date 2004, the end date 2005, the amount 2011, and the contact ID 2012 on the project registration / editing screen 2000 is saved.
[0029] The project attribute table 1011 has the following data items: project ID 1110, attribute ID 1111, attribute value ID 1112, and attribute value 1113. The attribute ID 1111 is an ID that identifies the type of each attribute. The attribute value ID 1112 is an ID that is assigned to each attribute value when a predetermined value is set as the attribute value. The attribute value 1113 stores an attribute value freely input by the user.
[0030] The project attribute table 1011 stores attribute data related to various projects associated with each other by a project ID 1110. For example, an attribute ID is assigned to each of the other party person in charge 2006, topic candidate 2007, industry 2008, member 2009, and forum name 2010 on the project registration / editing screen 2000, and values (data) entered in each input item are stored in the attribute value ID 1112 or attribute value 1113.
[0031] The manager can refer to the project-related data from the manager terminal device 1007 and modify it as necessary. For example, the manager can understand the contents of the project by hearing from the members or by referring to the proposal, and modify the data set in the theme candidate 2007 and industry 2008 if they are not appropriate. This eliminates data variations due to differences in the person who inputs the project-related data, making it possible to make the data highly useful.
[0032] <Contact-related data> 5 shows an example of the contact registration / edit screen 2400. FIG.
[0033] When contact is made with a customer (prospect) in sales activities, data related to the contact is registered from the contact registration / edit screen 2400. Data entered on this screen is saved in the contact table 1014 and the contact attribute table 1015. Contact-related data includes data related to contacts made in sales activities with prospective customers. The data saved in the contact table 1014 and the contact attribute table 1015 can be updated from the contact registration / edit screen 2400 at any time thereafter. Note that the input items and data items shown here are merely examples, and other input items and data items may be included, and some items may be omitted.
[0034] As shown in FIG. 5 , the contact registration / editing screen 2400 has input areas for each of the following items: contact ID 2401, customer name 2402, case name 2403, contact date 2404, contact method 2405, other party's person in charge 2406, reporter 2407, accompanying person 2408, inquiry status 2409, competition information 2410, business content 2411, desired personnel for member recruitment 2412, and amount 2413.
[0035] Contact ID 2401 is identification information given to a series of predetermined sales activity units such as a product or service. For example, when the same customer is contacted multiple times regarding a certain product or service, those activities are registered as data having a common contact ID. Contact method 2405 sets the method of making contact, such as an individual visit, meeting at a seminar or study group, introduction by an acquaintance, or telephone sales. Reporter 2407 and accompanying person 2408 set identification information (name, employee ID, etc.) of the employee who made the contact. Inquiry status 2409 sets the reason for making the contact and the reaction, etc. For example, it is registered that the other party is proactive, that it is a push to seminar attendees who have not responded, or that it is an introduction from an existing customer. Competition information 2410 is information about whether each contact was made in a competition. It indicates whether it is a case or not. In the business content 2411, specific business content such as proposals made through contact and contents of discussions is set. In the desired personnel 2412, conditions of personnel desired as members in case of receiving an order for a project from this contact are set. In the amount 2413, an expected contract amount is set.
[0036] FIG. 6A shows an example of the data structure of the contact table 1014, and FIG. 6B shows an example of the data structure of the contact attribute table 1015.
[0037] The contact table 1014 has, as data items, a contact ID 1140, a contact date 1141, an employee ID 1142, a customer ID 1143, an amount 1144, and a reporter flag 1145.
[0038] The contact ID 1140 and contact date 1141 store the data set in the contact ID 2401 and contact date 2404 on the contact registration / edit screen 2400. The employee ID 1142 stores the employee ID of the employee set in the reporter 2407 or accompanying person 2408 on the contact registration / edit screen 2400. The customer ID 1143 stores the customer ID of the customer set in the customer name 2402. The amount 1144 stores the amount set in the amount 2413. The reporter flag 1145 is set in the case of an employee who is the reporter 2407.
[0039] The contact attribute table 1015 has a contact ID 1150, an attribute ID 1151, an attribute value ID 1152, and an attribute value 1153. The contact attribute table 1015 stores attribute data related to various contacts associated with the contact ID 1150. As in the case of the project attribute table 1011, an attribute item on the contact registration / editing screen 2400 that is not stored in the contact table 1014 is set as an attribute ID 1151, and the input value (data) is stored as an attribute value ID 1152 or attribute value 1153.
[0040] The administrator can refer to the above contact-related data from the administrator terminal device 1007 and modify it as necessary.
[0041] <Employee-related data> Fig. 7 shows an example of a course list screen 2900. Fig. 8 shows an example of the data structure of the employee table 1018, the course table 1019, and the course application table 1020.
[0042] Many courses (training programs) are available for employees within a company. With this system, it is possible to manage course enrollment applications and attendance status for each employee. A course list screen 2900 in FIG. 7 is a screen for each employee to apply for courses that they will take and check their attendance status. In addition, this system may have a screen that allows managers to apply for courses for their subordinates and check their attendance status.
[0043] As shown in FIG. 7, the course list screen 2900 displays a course ID 2901, a course name 2902, a start date 2903, a target student 2904, a status 2905, and a course attendance date 2906. The course list screen 2900 further includes an application area 1907 for accepting application input for each course. The start date 2903 indicates the scheduled date for each course. If multiple courses are scheduled, all the dates for which the course is to be held are displayed in the start date 2903. The status 2905 displays the attendance status of each course for each employee. The attendance date 2906 is set to the date when the application was confirmed or the date when the course was attended. When the application button displayed in the application area 1907 is pressed, the application for that course is accepted.
[0044] FIG. 8A shows an example of the data structure of an employee table 1018, FIG. 8B shows an example of the data structure of a course table 1019, and FIG. 8C shows an example of the data structure of a course application table 1020 (an example of a second storage means).
[0045] The employee table 1018 is a table that stores employee attributes, and has the following data items: employee ID 1180, department 1181, job title 1182, year of joining the company 1183, and email address 1184. The data in the employee table 1018 is preset. For example, data may be acquired from another system within the company and set in the employee table 1018.
[0046] The course table 1019 is a table that stores course attributes, and has the following data items: course ID 1190, course name 1191, date 1192, and target audience 1193. Data for each item is registered in advance for each course in the course table 1019. The course ID 2901, course name 2902, date 2903, and target audience 2904 on the course list screen 2900 are displayed based on the data stored in the course table 1019.
[0047] The course application table 1020 has the following data items: course ID 1200, employee ID 1201, status 1202, and attendance date 1203. The course application table 1020 stores data entered on the course list screen 2900. That is, the course application table 1020 stores the status 1202 and attendance date 1203 data for each employee. Here, the status 1202 is set to, for example, attended, application confirmed, application in progress, not attended, absent, etc. The attendance date 1203 is the attendance date when the application is confirmed. The confirmation of course attendees and management of attendance status may be performed by another system (not shown), or may be performed manually by an employee. The latest attendance status is set in the course application table 1020.
[0048] <Customer related data> Fig. 9 shows an example of a customer registration / editing screen 3400. Fig. 10 shows an example of the data structure of the customer table 1024 and the customer attribute table 1025.
[0049] 9, the customer registration / editing screen 3400 has input fields for customer name 3401, reading 3402, customer ID 3403, organization type 3404, business code 3405, and head office location 3406. For example, the company type (joint stock company, limited liability company) is set in the organization type 3404. A code indicating the business type of the customer is set in the business code 3405.
[0050] FIG. 10A shows an example of the data structure of the customer table 1024, and FIG. 10B shows an example of the data structure of the customer attribute table 1025.
[0051] The customer table 1024 has a customer ID 1240 as a data item.
[0052] The customer attribute table 1025 has the following data items: customer ID 1250, attribute ID 1251, attribute value ID 1252, and attribute value 1253. The customer attribute table 1025 stores attribute data related to various customers associated with each other by the customer ID 1250. As in the case of the project attribute table 1011 and the contact attribute table 1015, an attribute item on the customer registration / editing screen 3400 that is not stored in the customer table 1024 is set as attribute ID 1251, and the input value (data) is stored as attribute value ID 1252 or attribute value 1253.
[0053] <Agent> 11 shows an example of an agent new registration / edit screen 2800. FIG.
[0054] The Agent acts on behalf of the User and operates the System in accordance with the conditions specified by the User. The agent is generated based on the conditions specified by the user on the new agent registration / edit screen 2800.
[0055] As shown in FIG. 11, the new agent registration screen 3800 has input areas for each of the following items: registrant 3801, matter name 3802, search target 3803, keyword 3804, reporter or member 3805, customer 3806, and destination 3807. In registrant 3801, identification information (name, employee ID, etc.) of the employee who registers the agent is set. In matter name 3802, the name of this agent is set. In search target 3803, the target from which the agent will collect data is set. In keyword 3804, a keyword used when performing a search. In reporter or member 3805, the employee who will send the data collected by this agent. In customer 3806, identification information (name, customer ID, etc.) of the customer to be searched is set. In destination 3807, the email address of reporter or member 3805 is set.
[0056] 12 shows an example of the data structure of agent table 1028. Agent table 1028 has, as data items, agent ID 1280, name 1281, search target 1282, keyword 1283, employee ID 1284, search target customer ID 1285, and destination email address 1286. Data entered on new agent registration screen 3800 is stored in the corresponding items of agent table 1028.
[0057] Database server 1001 generates an agent by referring to agent table 1028. The generated agent searches the data in database server 1001, for example, once a day, to collect data matching keywords 1283 or search target customer ID 1285. The agent then distributes the collected data to destination email address 1286. This allows a user who has registered an agent to obtain data in the field in which he or she is interested.
[0058] <Forum> 13 shows an example of the data structure of the forum management table 1030. The forum management table 1030 has the following data items: project ID 1300, forum ID 1301, and forum name 1302. A forum is set for each project, and is used as a place for information exchange between members participating in each project. Therefore, the project to which each forum is associated is specified by the project ID 1300. When the forum name 2010 is set on the project registration / editing screen 2000 in FIG. 3, the data is saved in the forum management table 1030.
[0059] The forum data storage unit 1031 stores various data within a forum for each forum ID. For example, the employee IDs of employees participating in the forum, data posted in the forum, etc. are stored for each forum ID.
[0060] <Community> 14 shows an example of the data structure of the community management table 1034. The community management table 1034 has, as data items, a community ID 1340 and a name 1341. A community is a place where employees within the system can discuss various topics and exchange information.
[0061] Community data 1035 stores various data within a community for each community ID. For example, employee IDs of employees participating in the community, data posted in the community, etc. are stored for each community ID.
[0062] <Data Search> In this embodiment, the user can search for data registered using each of the above-mentioned registration / editing screens using various search screens from their own user terminal device 1005. For example, project-related data, contact reports, customer data, employee data, etc. can be searched. Furthermore, based on the search results obtained by performing a first search on any of the data, a second search on other data can be performed. First, the search for project-related data and contact reports will be described.
[0063] FIG. 15 shows an example of a project search screen 4000. The project search screen 4000 (an example of a receiving means) has an area for receiving input of search conditions for project data. In the example of the figure, the search conditions can be set as customer names, words in the data, potential themes, industries, start dates, end dates, amounts, etc. When a search request is received on the project search screen 4000, the search engine 1040 (here, an example of the project search means 1040a) searches the project table 1010 and the project attribute table 1011 according to the conditions specified here, and extracts projects that match the conditions. The first output means 1041 outputs data related to the extracted projects as search results. The project-related data output here includes members who participated in the project. The system can further output personal information of specific members who participated in the project. For example, the search engine 1040 (here, an example of the employee search means 1040b) searches the employee table 1018, the course table 1019, and the course application table 1020 using the employee ID of the specific member as a key to acquire employee attribute data. The second output means 1042 outputs the acquired employee data as a search result.
[0064] Alternatively, the search engine 1040 (here, an example of the project search means 1040a) may further search the project table 1010 and the project attribute table 1011, or the contact table 1014 and the contact attribute table 1015, using the employee ID of the specific member as a key, and extract data related to projects and contacts in which the specific member currently or previously participated. As a result, the first output means 1041 further outputs data related to projects and contacts in which the specific member currently or previously participated.
[0065] In this system, when performing a KnowWho search, first, the above-mentioned project search is used to search for projects related to the technical field (potential topic) or industry of interest. It is then assumed that among the members involved in the projects extracted by this search, there are people with extensive knowledge of the technical field or industry of interest. At this time, the system can also search for additional information, such as know-how possessed by specific employees. In other words, the system performs a KnowWho search via a project search.
[0066] The background to this KnowWho search is the high accuracy of project-related data. In other words, in the case of a normal knowledge management system, employees are required to register their own experiences, skills, etc., but in reality, it is difficult to ensure that this is done thoroughly. In contrast, in the case of this system, project-related data is important data that is also used in systems other than this system. Therefore, reporters are forced to input data related to the project whether they like it or not, and the accuracy of the data is improved by having the administrator review the input data. Therefore, in this system, the input of project-related data is guaranteed, and the input data has high reliability. This allows for an effective KnowWho search to be performed by searching for projects that are closely related to the desired information and identifying employees involved in the project.
[0067] FIG. 16 shows an example of a contact report search screen 5000. As shown in the figure, the contact report search screen 5000 has an area for accepting input of search conditions such as customer name, words in the data, inquiry category, industry, contact date, and amount. When a search request is accepted on the contact report search screen 5000, the search engine 1040 searches the contact table 1014 and the contact attribute table 1015 according to the conditions specified here, extracts contact data that matches the conditions, and outputs contact data related to the contact. Again, the contact data output includes the members who participated in the contact. The system can further search the employee attributes of a specific member who participated in the contact by the search engine 1040. For example, the search engine 1040 searches the employee table 1018, the course table 1019, and the course application table 1020 using the employee ID of the specific member as a key to acquire the employee attributes. Alternatively, the search engine 1040 may further search the project table 1010 and the project attribute table 1011, or the contact table 1014 and the contact attribute table 1015, using the employee ID of the specific member as a key, to extract data related to projects and contacts in which the specific member currently or previously participates. This further outputs data related to projects and contacts in which the specific member currently or previously participates. Therefore, the above-mentioned KnowWho search is also possible by searching the contact report, promoting knowledge sharing within the company.
[0068] Also, instead of or in addition to the contact data in this embodiment, data related to research and development activities may be held in a similar data structure. This research and development data can be registered and edited in the same way as contact data, and can also be used as a search target. In other words, by searching the research and development data, it is possible to search for employees who have a lot of knowledge and experience about a specific technology or theme in research and development, and it is possible to promote knowledge sharing within the company in the research and development field. Furthermore, by associating each research and development theme with a project, it is possible to easily understand which research and development theme is specifically used in which project.
[0069] The above-described embodiments of the present invention are merely illustrative examples of the present invention, and are not intended to limit the scope of the present invention to these embodiments. Those skilled in the art can implement the present invention in various other forms without departing from the gist of the present invention.
[0070] [Second embodiment] In this embodiment, an embodiment using a large-scale language model as a keyword extraction means in an automatic classification technique for technical information will be described. As an example of an automatic classification technique for technical information, the description will be made with reference to Japanese Patent Publication No. 2011-108067 (Patent No. 5297351).
[0071] [System Configuration] 17 is a block diagram showing the configuration of an information retrieval system according to an embodiment of the present invention. The information retrieval system according to this embodiment (an example of a document classification device) can be constructed on a program running on a computer, and has an input / output device 151 for inputting and outputting data, which is composed of a keyboard, mouse, display, etc., where a user inputs data and displays the processing results, a business process template database 152 for storing business process template data in which each business procedure, its deliverables, and reference materials are described in association with the titles and storage addresses of the storage destinations, and a design information database 153 for storing design information such as reference materials, study documents, and accident information.
[0072] The system also includes a business support processing device 170 that includes business support means 155 for reading business process template data stored in the business process template database 152 from the design information database 153 and providing business support such as guiding the user through the business, presenting reference materials, and managing deliverables, and business process template editing means 156 for editing the business process template data.
[0073] The system also includes keyword extraction means 157 (an application example of a large-scale language model) which acquires keywords from a document which is called from the design information database 153 and is being displayed or executed on the display, classification category determination means 160 which determines the classification categories of design information to be displayed to a user and their display order based on the keywords extracted by the keyword extraction means 157, classification category priority determination means 164, and classification result display means 161 which displays the design information of the classification categories determined by the classification category determination means 160 in a manner that the classification categories with higher priority are displayed at the top based on the order determined by the classification category priority determination means 164.
[0074] The information retrieval system further has a document creation source business process discrimination means 158 that discriminates the original business procedure from which a document displayed on the display or being executed was created as a deliverable, based on design information such as reference materials, review documents, and accident information in the design information database 153 that is read into the business support means 155. A classification category priority determination means 164 determines the priority of classification categories to be displayed to the user based on the number of times that issues and related words listed in the issue list 162 appear in the document displayed on the display or being executed, and the distance between the business procedure acquired by the document creation source business process discrimination means 158 and the business procedure under consideration.
[0075] The information search system also has a task list 162 listing tasks and their related words, a task list editing means 163 for editing the task list 162, a classification means 159 for classifying design information stored in the design information database 153 according to the task categories listed in the task list 162, and a classification result database 154 for storing the classification results by the classification means 159.
[0076] [Business process template] The structure of a business process template is shown in Figure 18. A business process template represents a series of business information as a collection of ordered business processes, and each business process can be further divided into more detailed business processes. In other words, a business process template represents business information as a collection of hierarchically structured business processes, and each business process is made up of individual business steps. In Figure 18, the business process template is made up of the following three types of data:
[0077] The first is common data 201, which is composed of data of a template ID, a template name, and a project name. The template ID is an ID that is assigned by the system when a user creates a template. This template ID must not overlap with other template IDs.
[0078] The second is unit business process data 202, which is composed of data on business process ID, business process name, comment, person in charge, deadline, registered work time, start date, end date, achievement rate, actual work time, actual start date, actual end date, preceding process, title of deliverable and storage address in the design information database, title of reference material and storage address in the design information database.
[0079] The business process ID is an ID that is assigned by the system when a business process is defined, and does not overlap with other business process IDs. The business process name is a name that is assigned to a business process by the user when the business process is defined. The registered work hours, start date, and end date are the work hours, start date, and end date that were planned before the project started, and the work hours, start date, and end date when the work was actually performed are the actual work hours, actual start date, and actual end date. A preceding process is a business process that must be completed before the business process is executed, and the relevant business process is considered using the deliverables of the preceding process. The business process ID of the preceding process is entered in the business process template.
[0080] The third is process order / hierarchy data 203, which defines the order in which business processes are executed and the parent-child relationships between business processes for the above-mentioned unit business process data 202. In the example of Fig. 18, business processes with business process IDs 1 and 5 are defined as the top-level business processes in the process order / hierarchy data 203. Furthermore, business processes with business process IDs 2 and 3 are defined in a lower hierarchy than the business process with business process ID 1.
[0081] Next, when this business process template is loaded into the business support means 155, the business processes are displayed hierarchically on the display of the input / output device 151 as shown in Fig. 19. Fig. 19 shows an example of executing the business process template for the "A model development design" project. The top layer displays "A model development design," and the layers below that display the business processes of "product planning," "concept design," "detailed design," "prototype and testing," and "drawing creation." Furthermore, the layers below that display even more detailed business processes.
[0082] In addition, when the user presses a process display button (not shown) to switch to the arrow display mode and selects any business process (displayed with a circle), the predecessor-successor relationship of the business process is displayed with an arrow. In this embodiment, the business procedure refers to the order of business related by the predecessor-successor relationship of this business process.
[0083] The left-facing arrow indicates that the business process "Create report on heat study results" 302 is the preceding process of the user-selected business process "Create report on strength study results" 303. That is, in "Create report on strength study results", a shape that satisfies the strength design conditions is considered within the range that satisfies the thermal design conditions in the upstream process "Create report on heat study results" 302, so "Create report on heat study results" is displayed as the preceding process.
[0084] Also, the business process with the right arrow, "Mechanism Study" 304, represents the subsequent process of the selected business process, "Strength Study Result Report Creation" 303. In other words, a subsequent process means a business process that must be started after the selected business process is executed. These preceding and subsequent processes are displayed using the information of the preceding process in the unit business process data 202 of the business process template data.
[0085] FIG. 20 shows the entire screen displayed by the processing of the business support means 155. The left frame contains the business process display area 401 shown in FIG. 19. When the "Output" tab 402 in the right frame is selected, a list of deliverables registered for each business process is displayed. In the example of FIG. 20, "Thermal Study Result Report" is displayed on the display for the business process of "Creating Thermal Study Result Report". Furthermore, when the "Reference" tab 403 is selected, a list of reference materials associated with each business process is displayed. By selecting the reference materials list, it is possible to refer to the deliverables and reference materials stored in the design information database. 404 is a process display button.
[0086] [Automatic Classification of Design Information] This system periodically identifies issues that the user is considering from design information such as a document currently being displayed or executed on the display, and automatically presents the results of classifying the design information stored in the design information database 153 based on the identified issues. The process will be explained below with reference to the flow charts in Figures 21, 22, and 23.
[0087] In the system of this embodiment, the keyword extraction means 157 in Fig. 17 periodically obtains the number N of active documents and the title and destination address of each document using a function for obtaining information on active documents, such as the getObject function provided in VBA (Visual Basic (registered trademark) for Application), as shown in Fig. 21, and creates an N-row, 2-column array as shown in Fig. 24 in the memory of the keyword extraction means 157, and stores the title 801 and destination address 802 of each document (S501). Next, a text output program is executed using the destination addresses stored in the array in the memory as input, and text information is obtained from the active documents and output as a text file (S502).
[0088] The keyword extraction means 157 executes a morphological analysis program using this text file as input, and breaks down the text information into words (S503). Of the broken down words, only nouns are extracted and stored in memory as keywords (S504). For example, the text information "consider the strength of parts" is broken down into "parts," "of," "strength," "for," and "consider." When noun words are extracted from this, "parts" and "strength" are obtained as keywords.
[0089] Next, the classification category determination means 160 matches the acquired keywords with related words of each assignment listed in the assignment list 162, and counts the number of occurrences Xn of the related words in the document for each assignment (S505). In the assignment list 162, related words 903 are associated with each assignment 902 as shown in Fig. 25. 901 is the assignment number.
[0090] An issue with Xn equal to or greater than 1 means an issue in the business process currently being executed. For example, when design information such as that shown in Fig. 26 is displayed or executed on a display, keywords such as "strength," "stress," and "proof strength" are extracted. Comparing these keywords with the issue list in Fig. 25, the related words for "strength" 907 match, and "strength" becomes an issue in the business process currently being executed.
[0091] Next, the document creation source business process determination means 158 acquires the business process template data loaded in the business support means 155 (S601), as shown in the flow diagram of Fig. 22, and then acquires the business process selected by the business support means 155 (S602). Next, it compares whether the title of the deliverable and reference materials of the unit business process data of the acquired business process template data match the title of the active document stored in the array in memory (S603). If they match, it acquires the business process name whose title of the deliverable and reference materials of the unit business process data of the business process template data match the title of the acquired active document (S604).
[0092] The business process ID of the preceding process of the unit business process data of the business process template data is traced from the business process selected by the business support means, and the number of steps Yn up to the acquired business process is acquired (S605). If the title of the deliverable of the unit business process data of the business process template data does not match the title of the active document, Yn = 0 is set. Xn and Yn are calculated only for the number of documents displayed, N.
[0093] [Relevance to the issue P] Next, the classification category priority order determination means 164 calculates the relevance P between each task and the currently executing business process (S701) as shown in the flow diagram of Fig. 23. The relevance P is calculated using formula (1). This formula means that the more related words a task contains and the related words are contained in the deliverable of a business process that is closer to the currently executing business process, the higher the relevance P of the task to the currently executing business process. Degree of association P = ΣXn × Zn (1) Xn: The number of occurrences of words related to the task in the document displayed on the screen Zn: The relevance of the document displayed on the screen to the currently running business process Yn: The business process in which the displayed document is registered, Relevance of currently running business processes M: Total number of steps defined in the template Zn=M / (Yn+1): Title of the document in progress and the deliverable of the business process or If the reference titles match Zn=1: Title of the document being run and the deliverables of the business process or If the reference titles do not match For example, if the "Thermal Study Result Report" that describes the results of the study on the temperature distribution inside the product is being displayed or executed on the display, it can be seen from the business process template data that the business process to which this design information is registered is "Creating a report on the thermal study results," as in the example of Figure 20. If the business process currently being studied is "Creating a report on the strength study results," as can be seen from Figure 19, "Creating a report on the thermal study results" is the business process one step upstream of "Creating a report on the strength study results," so Yn is 1 and Zn is M / 2.
[0094] Furthermore, if "Strength Study Reference Material" as shown in Figure 26 is being displayed or executed at the same time, "Strength" will be extracted as an issue. If this "Strength Study Reference Material" is registered as reference material for the currently executing Business Process "Strength Study Result Report Creation", the number of steps Yn between the Business Process in which the displayed document is registered and the currently executing Business Process will be 0, and Zn = M. Therefore, when Xn is the same, "Strength" will have a higher degree of association than "Heat".
[0095] Next, the design information stored in the design information database 153 is classified in advance by the tasks listed in the task list 162, and the results classified by the tasks are presented in order of relevance (S702), as shown in Fig. 27. Fig. 27 is an example of the display of results classified by "strength" 1102 and "heat" 1106. The relevance calculated by formula (1) is displayed to the right of each classification, and "strength" 1102, which is a task with a high relevance of 67, is displayed at the top.
[0096] Next, at a lower level of classification, a title 1107 of the design information classified into that issue is displayed, and by selecting this title on the screen, design information 1108 is displayed.
[0097] [Relevance of design information Q] Next, for each piece of design information classified by task, the relevance Q between the design information and the currently executing business process selected by the business support means 155 is calculated, and the relevance values are displayed in descending order. Furthermore, the relevance Q is displayed for each piece of design information. To find the relevance Q, first, a document vector is generated for the business process name and the document, with the frequency of occurrence of the words contained therein as components. If the document vector generated from the business process name is vector A and the document vector generated from the document is vector B, the relevance Q for the design information is found using formula (2). The closer the value found using formula (2) is to 100, the more relevant the information is. When generating the document vector for the business process name, comment information from the unit business process data 202 of the business process template data may be added. Furthermore, each component of the document vector may be weighted according to the importance of the words.
[0098]
number
[0099] This document classification method is just one example, and other methods may be used instead of the above. Also, as shown in Figure 27, the classification results by issue may be further classified by product name or part name, and the product name or part name may be obtained from the keywords contained in the design information displayed or executed on the screen, and the display may be limited to issues related to that product or part.
[0100] [Classification of trade-offs] In this system, in addition to presenting the results of classifying design information stored in the design information database based on the issues the user is considering, the system also classifies and presents issues that are trade-offs (opposite relationships) with issues. Issues that are trade-offs with a certain issue are listed with the number assigned to each issue in the trade-off relationship column 904 in the issue list in Fig. 25. For example, since the issue number that is a trade-off with strength 907 is 5 (906), the issue to be considered that is a trade-off is weight reduction 905. The classification results for the issues that are trade-offs obtained in this way are displayed in a lower hierarchical level than the strength classification results, as shown in 1105 in Fig. 27.
[0101] Conventionally, as shown in FIG. 28, a user would input a search keyword into a search system (1201), perform a search, and look for required design information from search results (1202) ordered based only on the relevance to the search keyword. In contrast, in the system of this embodiment, the classification categories for the issue the user is considering can be displayed together with the relevance to the business process being executed, as shown in FIG. 27, without the user having to input business process information. Furthermore, the classified design information can be displayed in descending order of relevance to the business process being executed, and the relevance can be displayed for each piece of design information. This allows the user to efficiently search for information related to the business process being executed without hassle.
[0102] Furthermore, whereas in the past, users were unable to obtain relevant information unless they consciously specified search keywords, with this system it is possible to obtain classification results for design information not only for issues that the user is considering, but also for issues that involve trade-offs that the user is not aware of.
[0103] [Business support method] When business process template data is loaded in the business support means 155 shown in Fig. 17, the user can perform the business without any omissions by performing the business according to the business process displayed in the business process display area 401 in Fig. 20. Furthermore, when the "Reference" tab 403 is selected, a list of reference materials associated with each business process is displayed. By selecting a reference material from the list, the reference material stored in the design information database 153 can be referenced. This allows the user to efficiently proceed with the business while referring to the reference materials associated with each business process.
[0104] Furthermore, in the business support means 155, deliverables such as design review reports and progress information can be registered and managed for each business process. The registered deliverables are displayed in a list when the "Output" tab 402 in the right frame of FIG.
[0105] Fig. 29 shows an example of registering a deliverable. When registering a deliverable, select a business process and press the register button 1321. Then, on the deliverable registration screen 1322 that is displayed, specify the file that will be the deliverable stored on the PC, enter a title, and press the register button 1303 to register the deliverable. The registered deliverable is stored in the design information database 153, and the entered title of the deliverable and the storage address in the design information database of the deliverable are written to the business process template.
[0106] 30, when the mouse cursor 1402 is placed on a specific business process, the progress information is displayed on a progress information display screen 1403. The information displayed includes the person in charge, working hours, start date, end date, achievement rate, actual working hours, actual start date, actual end date, etc. When registering progress information, a progress information registration screen 1503 is displayed for each business process.
[0107] 31 and 32 show examples of progress information registration. On progress information registration screen 1503, the screen can be switched between a "General" tab 1501 and a "Details" tab 1502. On the screen displayed by selecting "General" tab 1501, information such as the achievement rate, actual work hours, actual start date, and actual end date can be registered. Also, on the screen displayed by selecting "Details" tab 1502, information such as the deadline, standard work hours, standard start date, and standard end date can be registered, as shown in FIG. 32.
[0108] [Editing Business Process Templates] The business process template editing means 156 allows editing of the business process name and the hierarchical structure of the business process.
[0109] 33 shows the edit menu for a business process template. When a business process displayed in a business process display area 1701 is selected, an edit menu including create new business process 1702, edit business process name 1703, copy 1704, cut 1705, paste 1706, and delete 1707 is displayed.
[0110] For the create new business process 1702, there are submenus for sibling process 1708 and child process 1709. Similarly, for the paste 1706, there are further submenus for sibling process 1710 and child process 1711.
[0111] When sibling process 1708 is selected, business process input screen 1801 in FIG. 34 is displayed. By entering a business process name in business process input field 1802 and pressing register button 1803, a business process with the entered business process name can be added to the same hierarchical level of the selected business process. When child process 1709 is selected, business process input screen 1801 is similarly displayed. By entering a business process name in business process input field 1802 and pressing register button 1803, a business process with the entered business process name can be added to a lower hierarchical level of the selected business process.
[0112] When edit business process name 1703 is selected, edit business process name screen 1901 in Fig. 35 is displayed, and the existing business process name is displayed in business process input field 1902. The business process name can be edited by changing it and pressing register button 1903.
[0113] The selected business process can be copied by selecting Copy 1704. The copied business process can be pasted as a child process of the selected business process or as a process at the same hierarchical level by selecting Paste 1706. Whether to paste as a process at the same hierarchical level or as a child process is determined by selecting Sibling Process 1710 or Child Process 1711 from the submenu.
[0114] Selecting Cut 1705 cuts the business process, and selecting Paste 1706 allows it to be pasted in the same manner as Copy 1704. Selecting Delete 1707 allows the business process to be deleted from the template.
[0115] [Third embodiment] In this embodiment, an embodiment using a large-scale language model as a means for extracting information in automating a job or task will be described. As an example of a job or task automation technology, the embodiment will be described with reference to Japanese Patent Application Laid-Open No. 2021-86403.
[0116] This embodiment was born from the improvement of problems such as cost reduction including efficiency through business improvement, and provides a system that uses related systems and data depending on the business, aims to solve problems by reducing system-related costs within a company and improving work efficiency, and supports business reform in the company. In order to reduce system-related costs, it is desirable to continue using existing systems that are familiar to users and are "stable, secure, and safe" for as long as possible, so the point of this embodiment is to use the existing system as it is.
[0117] The business reform support system proposed in this embodiment builds an information network that can consolidate and utilize information assets, including various data (Excel, Word, PowerPoint, etc.) created by individuals and buried in their personal computers, ledgers stored on paper, and case-related materials, and as a core platform for the company's business systems, consolidates and manages data groups from in-house systems (Excel, Word, PowerPoint, etc.) and utilizes them as an intranet cloud. In today's society where information is overflowing, valuable information collected within companies is stored in databases (DB), and most people are unable to find it. This embodiment has been made in light of this situation. The fusion of systems through the collection and management of data files (Excel, Word, etc.) from various systems within the company leads to the creation of new business systems. For example, just as human induced pluripotent stem (iPS) cells are "regenerative medicine," the business reform support system, which is an integration pluripotent stem (IPS) system, is a business reform trunk in systems, and is a groundbreaking system that opens up business reform by utilizing existing systems and obtaining new applications. It realizes business reform by utilizing data and information from systems that employees have used in their work so far, and utilizing past analysis and future predictions as new applications. The groundbreaking integration pluripotent stem (IPS) system, which obtains new applications, can also be said to be a "trunk system" that realizes system regeneration.
[0118] The business process reform support system of this embodiment, which utilizes existing systems, utilizes information and data from existing systems within the company and is operated under the management of a system administrator. Specifically, the system aims to support business process reform by integrating, analyzing, and updating systems scattered throughout the company for the purposes of "increasing sales," "increasing gross profit," and "reducing expenses."
[0119] The systems used within a company include existing data that is simply stored without being used, various data buried in personal computers (including computers) created by individuals, ledgers and case-related materials stored on paper, etc. This embodiment extracts and consolidates these information assets as awareness information, thereby unifying and sharing information, and positioning the information network not just as a communication tool but as the core of corporate strategy. This embodiment is a so-called cockpit system, the leading platform of the future, and a pioneer of intranet clouds that manage the consolidation and management of in-house data groups stored in various systems.
[0120] Recently, companies have come to encourage businesses to simply increase the number of visits made in a limited time, and instead encourage businesses to increase the time spent at important customers. In this way, the extraction of "awareness information" in this embodiment is used to derive information about customers that salespeople should visit. For example, in the past, by linking with a map system or an existing system, the number of visits made by salespeople to customers was increased and the quality of visits was improved by visualizing information on a map. According to this embodiment, for example, in visualizing information on a map, further "awareness information" can be provided, so the quality of visits can be further improved. Therefore, it is important to have a mechanism for extracting and aggregating "awareness information" that leads to business reform in companies.
[0121] Regarding the business of a company, for example, to explain the changes in the environment in financial institutions, financial institutions used to only need to collect a lot of deposits. In other words, if they collected deposits, they had a mountain of borrowers, and if they deposited the surplus funds in Shinkin Central Banks or the Bank of Japan and managed them, they would generate profits, but in recent years, this business model has become unsustainable. Therefore, there is a demand to move away from only low-profit sales activities. In order to increase profitability, it is necessary to sell products such as loans, investment trusts, and insurance, so the behavior of salespeople needs to change, and there is a demand to visualize how much profit activities are bringing in and place more emphasis on profit-making activities. In addition, companies need to change their old evaluation methods. Until now, evaluation methods have been centered on deposits (sales), but it is necessary to change to evaluation methods such as how many corporations have increased their profits by more than 1 million yen per year, or how many individuals have increased their profits by more than 100,000 yen per year.
[0122] The role (function) of the awareness information management tool of the business reform support system required in this embodiment is a mechanism to know revenue, not the ranking of transaction amount for each customer. For example, it reflects annual interest income (gross profit), commission income, and miscellaneous income. The objective is to grasp the profitability per customer (client), and prepare a data table of the number of cases handled, transaction volume, revenue, and ranking, like the personal medical record of a sales representative, and a list of activity levels for each item, graphing each value, automatically registering regular collection destinations, and registering destinations with needs, etc. individually. In addition, the actual values are converted into a skill graph with average values to visualize people's strengths and weaknesses.
[0123] Furthermore, in this embodiment, the contents to be used for the assessment are configured. Daily report input information is input by each sales (public relations) representative, and the revenue performance and activity list are compiled in the chart, the file is updated in real time, and the information is disclosed to the superior, so that the progress of the case can be understood and the sales process can be evaluated and guided. As an awareness information management tool of this business reform support system, it is linked to accounting systems and other necessary existing systems to obtain data more reliably, prevent erroneous input, and speed up compilation.
[0124] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.
[0125] Examples of keywords for various customer-related information data scattered within the existing system include the following: "number of loans," "number of visits," "number of interviews," "number of prospects," "number of successful deals," "sales target," and "sales performance." From the various customer-related information and data scattered throughout the system, insightful information related to the above keywords is extracted and stored (linked) in a filter. In order to share the information obtained from the keyword list information items with the evaluator for each purpose, it is classified into "shared filters (leading to information and proposals)," "shared filters (leading to increased revenue)," "shared filters (leading to gaining trust and preventing mistakes)," and so on, as shown in Figure 36. In addition to the above, shared filters can be created for each purpose and keywords can be set.
[0126] The keyword list is a group of keywords for extracting information necessary for awareness from various information scattered within existing systems such as Excel and Word. Information extracted from Excel, Word, etc. using the keyword list becomes "various data 519" (see Fig. 40). "Various data 519" becomes one piece of information from other existing systems 500, and furthermore, "existing systems 500", "channels 600", and "external information 700" become information sources.
[0127] FIG. 37 shows a method for acquiring data (FIG. 37(B)) from all cell information (FIG. 37(A)) in Excel (hereinafter referred to as the "all cell method"). This all cell method can acquire all character information, but does not know what information the characters are. In the example of FIG. 37, for example, it is known that information such as "April 4, 2019" exists, but it is not known that the "report date" is April 4, 2019. In addition, with the all cell method, there are no delimiters for all information just by reading it, so it is just a simple string of characters. To solve this, it is necessary to use a known morphological analysis or full-text search engine to make the information meaningful before using it. "Morphological analysis" is the process of dividing a sentence written in a natural language into morphemes, which are the smallest units of language, and determining the parts of speech and changes of each morpheme. In the IT field, it is used as one of the natural language processing methods by computer, such as kana-kanji conversion, full-text search, and machine translation. In addition, "full-text search engines" are sold as middleware products that enable high-speed searches of databases, etc., and by extracting index information from large volumes of data, creating search clusters, and searching data via the created clusters, searches can be performed much more efficiently and quickly than direct searches of the data source. As an example, the text can be broken down into nouns and verbs using "grain and particle" to improve accuracy. For example, if "I visited XX company on March 24th," a morphological analysis is performed to obtain "I visited XX company on July 24th." (broken down by " / "), and by obtaining "XX company" and "visit" as nouns, it is possible to determine whether the file in question matches the filter.
[0128] In the all-cell method, a shared folder is prepared on a server in the network, and processing is carried out on that server. When each person creates a document on their own PC, they save the file in the shared folder, rather than saving the document on their own PC. There are well-known methods for saving files on the server without the PC user being aware of it, such as using a thin client or setting the OS so that documents can only be saved in the shared folder.
[0129] FIG. 38 shows a method of acquiring data (FIG. 38(B)) from designated cell information (FIG. 38(A)) in Excel (hereinafter referred to as the "designated cell method"). In this designated cell method, the cell number and the information written in the cell are registered in advance in the master, and when reading the file, the cell numbers registered in the master are sequentially specified and the contents written in the cells are sequentially read. In the example of FIG. 38, the contents of the report date are written in cell N4, so the contents of cell N4 "April 4, 2019" are read. Similarly, the contents of the department are written in cell B5, so the contents of cell B5 "Sales Headquarters Sales Division 2" are read, and the contents of the name are written in cell B6, so the contents of cell B6 "Saito △o" are read. Compared to the all-cell method described in FIG. 37, the information to be read is known in advance, so the information is meaningful. The designated cell method cannot be used unless it is determined what information exists in which cell, so it is optimal for fixed formats. Excel / Word can also be used to visualize the awareness information required in this embodiment. By writing the desired insight information into Excel / Word using the method described below in the reverse order of reading, the user can open Excel / Word and obtain the insight information in the form of text, tables, or graphs.
[0130] In the designated cell method, a shared folder is prepared on a server in the network and processing is performed on that server. A method of storing files on the server without the user of the computer being aware of it can be performed using a known OS.
[0131] FIG. 39 shows a case where data is obtained from Word. As with the all-cell method, all character information can be obtained from Word, but it is not clear what information the characters are. Also, as shown in FIG. 39(A), simply reading the information does not separate all the information, so it is just a simple string of characters. To solve this, as with the all-cell method in Excel, it is necessary to use a known morphological analysis or full-text search engine to make the information meaningful before using it. Since Word does not have cells like Excel, it is not possible to obtain specific information. In the example of FIG. 39(B), for example, it is clear that there is information such as "January 1, 2017", but it is not clear that this is "January 1, 2017" as the "contract date".
[0132] 40 shows the linkage relationship of the business reform support system 100 according to this embodiment of a financial institution as an example of a company, in which an awareness information construction system 200 is integrated and included in the business reform support system 100, and a data warehouse (DWH) / integrated database (DB) 400 links data, while an existing system 500, a channel 600, and external information 700 link data with the business reform support system 100 via an interface (I / F). If the DWH / integrated DB 400 has already been built within the company, or is built in conjunction with the introduction of the business reform support system 100, data linkage between this embodiment and the DWH / integrated DB 400 makes it possible to use data in real time between the existing system using the DWH / integrated DB 400 and the business reform support system 100 according to this embodiment.
[0133] The existing systems 500 currently in operation in a company include an international system 501, an exchange control system 502, a fund and securities system 503, a management control system 504, an office centralized system, a CRM and SFA 506, a loan approval system 507, a financial analysis system 508, a collateral evaluation system 510, a bond management system 511, an examination system 512, a credit rating and self-assessment system 513, a risk control system 514, an investment trust and insurance counter sales system 515, a marketing system 516, a groupware 517, a contract system 518, and various data (Excel, Word, PowerPoint, etc.) 519. These existing systems 500 and the business reform support system 100 are linked by a known method. In this embodiment, Word, Excel, etc. are involved in this linkage. Word and Excel data are treated as one of the existing systems (various data 519) and are used as a source of information for awareness information and as an output destination. The channels 600 include a branch office counter 601, an ATM (automated cash withdrawal and deposit machine) 602, a call center 603, the Internet 604, direct mail 605, email 606, and a social networking service (SNS) 607. These channels 600 and the business reform support system 100 share data using a known method. Furthermore, the external information 700 includes company information 701, external rating information 702, personal credit information 703, registration information 704, statistical information 705, real estate information 706, market information 707, and compliance information 708. These external information 700 and the business reform support system 100 share data using a known method.
[0134] 41, which includes a CPU (including an MPU, an MCU, etc.) 101 that performs overall control and arithmetic processing, to which a ROM 102 that stores programs, etc., a RAM 103 that stores data and parameters for arithmetic processing, and a database (DB) 104 that stores various data and information are connected. The business process restructuring support system 100 includes a data input unit 109 for inputting data and instructions from a numeric keypad, a mouse, a scanner, etc., an input unit 112 for inputting data and information from the outside via an interface 111, a visualization unit 113 for visualizing data and information on a screen (display) or a printer, etc., and an output unit 114 for outputting data and information to the outside via an interface 115, and is connected to the Internet 604 via a router 116.
[0135] In addition, the business reform support system 100 is provided with an awareness information construction unit 210 including an input linkage interface (I / F) 220 and an output linkage interface (I / F) 260 in cooperation with the CPU 101, etc., and the input linkage interface 220 imports information in data-linkage with the DWH / integrated DB 400, existing system 500, channel 600 and external information 700, and the output linkage interface 260 outputs information in data-linkage with the DWH / integrated DB 400 and existing system 500.
[0136] FIG. 42 shows a conceptual configuration of the awareness information construction system 200. The awareness information construction system 200 collects a wide range of information using the existing system 500, the channel 600, and the external information 700 as input information sources, and extracts awareness information based on extraction conditions such as the wealthy class, negotiation history, and registration information. The extraction conditions are registered in the awareness information extraction master in the DB 104. Types of awareness information include shared information that leads to proposals, information that leads to increased revenue, information on gaining trust, and information on preventing mistakes. Shared information that leads to proposals includes visit plans, negotiation history, the wealthy class, visit frequency, deadline management, registration information, information on ideas, business matching, relations, and visualization on a map. Information that leads to increased revenue includes negotiation history, the wealthy class, visit frequency, registration information, information on ideas, business matching, relations, and the like. Information on gaining trust and preventing mistakes includes visit plans, deadline management, and visualization on a map. The notice information extracted by the notice information construction system 200 is notified to a person in charge, etc., or is visualized on a map, CRM, Excel / Word, etc.
[0137] The schematic configuration and information flow of the awareness information construction system 200 are as shown in Fig. 43. The awareness information construction system 200 collects information from a plurality of existing systems, such as a CRM 506, a marketing system 516, and a groupware 517, scattered as an existing system 500, via an input linkage interface 220. The input linkage interface 220 is composed of a system connection interface 221 that takes in information, and a unified aggregated storage device interface 222 that sends information to a unified aggregated storage device 230. The input linkage interface 220 analyzes or simplifies data or information of different formats, the system connection interface 221 can increase the number of variations depending on the connected system, and the unified aggregated storage device interface 222 has the function of storing simplified information in the unified aggregated storage device 230, changing the stored contents, and deleting the stored contents.
[0138] The information sent from the centralized storage device interface 222 of the input linkage interface 220 is input to and stored in the centralized storage device 230. The centralized storage device 230 includes a customer section 231 that stores information such as a visit plan and a negotiation history, a date and time section 232 that stores date and time data such as deadline management and campaigns, and an address section 233 that stores information such as real estate and land prices. Key information is identified from each piece of information and data, and the identified key information and information subject to the extraction conditions are stored in the centralized storage device 230. Examples of this key information are customers (account numbers), dates and times, addresses, etc., and these key information are determined by the settings of the centralized storage device interface 222 and the awareness information extraction section 240 (an example of application of a large-scale language model). The decomposed information is stored separately in the centralized storage device 230, such as the customer section 231, the date and time section 232, and the address section 233, but the key information is used to link the separate information extracted by the extraction conditions to make it meaningful awareness information.
[0139] The data and information stored in the centralized aggregation storage device 230 are read into the awareness information extraction unit 240 described below, and the awareness information extraction unit 240 equipped with a large-scale language model extracts awareness information from the input information source based on extraction conditions such as the wealthy class, negotiation history, registration information, etc. This makes it possible to centrally aggregate various information scattered across the DWH / integrated DB 400, existing system 500, channel 600, and external information 700 and extract awareness information.
[0140] The extraction of information that can be noticed includes visit plans (selection of customers who should be visited as a priority), negotiation history (discovery of proposal ideas and latent needs), wealthy customers (destinations for financial products such as assets held in custody), due date management (avoidance of missed follow-ups), registration information (substitutes for mortgages and apartment loans), business matching (business relationships and connections with stakeholders), relationships (preventing the outflow of deposits outside the business area due to inheritance), promotional product information (whether or not there are customers for each product such as personal loans and investment trusts), investment trusts and insurance products that are easy to acquire through non-face-to-face channels or at the counter, destinations for large transfer deposits and withdrawals and specific ATMs. This is based on the user (household), frequency of visit (revealing long-unvisited customers), asset management clients in the past month, clients requiring caution (clients managed for mid- to long-term delinquency, clients managed for complaints, etc.), past visit history of executives and branch managers, clients scheduled for real estate collateral inspection (reevaluation), new real estate properties, subdivision development information, clients visited by sales representatives from other banks, clients of other banks with outflows of funds, main clients of other banks, internet and call center transaction users (high frequency clients), multiple owners of real estate properties, owners of real estate targeted for productive green space, companies with good performance, companies with improved COSMOS (registered trademark: company profile database) ratings, companies with changed representatives, etc. The notice information extraction unit 240 judges whether or not the extraction conditions for notice information extraction are met, and extracts those that meet the extraction conditions, and the extraction conditions can be set in the notice information extraction master in DB104 for each notice.
[0141] The notice information extraction unit 240 prepares logic for combining multiple items according to the purpose of extraction, and extracts, for example, customers who are wealthy, have not traded any products in the promotional product information, and have not visited for more than one month in the visit frequency category, as "notice information."
[0142] The notice information extracted by the notice information extraction unit 240 is sent to the notice aggregation storage device 250 and stored therein. As with the centralized aggregation storage device 230, the memory includes a customer section 251 that stores information such as a visit plan and a negotiation history, a date and time section 252 that stores date and time data such as deadline management and campaigns, and an address section 253 that stores information such as real estate and road prices. Key information is identified from each piece of information that matches the extraction conditions, and each piece of information that matches the extraction conditions based on the key information is stored in the notice aggregation storage device 250. Examples of this key information are the customer (account number), date and time, and address, and these key information are determined by the settings of the notice information extraction unit 240 and the notice aggregation storage device interface 261 in the output linkage interface 260. These key information are used to link information separately stored in the customer section 251, date and time section 252, address section 253, etc., to make it into meaningful notice information.
[0143] The notice information extraction unit 240 extracts information such as "above a certain level of wealth, with recent updates to registration information and no negotiation history for more than one month," and this extracted wealthy class information, registration information, and negotiation history information is copied from the centralized centralized storage device 230 to the notice centralized storage device 250. When copying, the notice information extraction master identifier is also stored, making it possible to identify what notice the information stored in the notice centralized storage device 250 is.
[0144] The awareness information stored in the awareness aggregation storage device 250 is input to the output linkage interface 260 and visualized on a map, CRM, etc. It is indicated by an icon on a map, or linked to a list of potential visiting destinations in CRM. The output linkage interface 260 is composed of an awareness aggregation storage device interface 261 and a system connection interface 262, and processes the wealthy information, registration information, and negotiation history information stored in the awareness aggregation storage device 250 into information linked with "customer," and then processes it into data for a map system with addresses as keys. The awareness aggregation storage device interface 261 extracts the awareness information according to the awareness visualization type, and the system connection interface 262 outputs the simplified awareness information in accordance with the visualization unit.
[0145] The awareness information constructing unit 210 is made up of an input link interface 220 , a unified aggregated storage device 230 , an awareness aggregated storage device 250 , and an output link interface 260 .
[0146] 44 is a block diagram showing the interrelationships of the awareness information construction system 200, in which the DWH / integrated DB 400, the existing system 500, the channel 600, and the external information 700 are connected to the system connection interface 221 in the input linkage interface 220, and information etc. taken in from the system connection interface 221 is stored in the unified aggregated storage device 230 via the unified aggregated storage device interface 222. The awareness information extraction unit 240 reads out the information stored in the unified aggregated storage device 230, extracts the awareness information, and stores the extracted awareness information in the awareness aggregated storage device 250. The awareness information stored in the awareness aggregated storage device 250 is read out by the awareness aggregated storage device interface 261, and transmitted to the DWH / integrated DB 400 and the existing system 500 via the system connection interface 262.
[0147] 45 shows an example of the operation of the input linkage interface 220. First, existing system information to be linked is acquired (step S210), and from the acquired existing system information, data acquired from the existing system 500 connected via the system connection interface 221 is decomposed into decryption units (step S211). Then, one piece of decomposed data is acquired (step S212), one decomposition condition is acquired from the existing system information (step S213), and it is determined whether the decomposed data matches the decomposition condition (step S214). If the decomposed data matches the decomposition condition, it is determined whether there are other decomposition conditions (step S215).
[0148] If it is determined in step S214 that the decomposition data does not match the decomposition conditions, the process returns to step S212. If it is determined in step S215 that there are other decomposition conditions, the process returns to step S213.
[0149] If it is determined in step S215 above that there are no other decomposition conditions, the acquired decomposition data is recorded in the centralized storage device 230 (step S216), and it is determined whether there is next decomposition data (step S217), and if there is no next decomposition data, an extraction request is made to the notice information extraction unit 240 (step S218). If it is determined in step S217 that there is next decomposition data, the process returns to step S212 above.
[0150] FIG. 46 shows an example of the operation of the notice information extraction unit 240. First, user information is acquired (step S220), extraction conditions are acquired from the notice information extraction master in the DB 104 (step S221), and a data list is acquired from the centralized storage device 230 (step S222). Then, one data is acquired from the data list (step S223), and one extraction condition is acquired from the extraction condition list (step S224). Then, it is determined whether the data matches the extraction condition (step S225), and if the data matches the extraction condition, it is determined whether there are other extraction conditions (step S226). If it is determined in the above step S225 that there is no match, it returns to the above step S223, and if it is determined in the above step S226 that there are other extraction conditions, it returns to the above step S224.
[0151] In the above step S226, if it is determined that there are no other extraction conditions, the data is recorded in the awareness aggregation storage device 250 (step S230), and it is determined whether there is next data in the data list (step S231), and if there is no next data, an output request is made to the output linkage interface 260 (step S232). In step S231, if it is determined that there is next data, the process returns to the above step S223.
[0152] Next, an example of the operation of the output link interface 260 will be described with reference to the flowchart of FIG.
[0153] First, the visualization system information to be linked is acquired (step S240), and the awareness information data is acquired from the awareness aggregation storage device 250 (step S241). Then, the awareness information data is converted from the visualization system information (step S242), the converted awareness information data is input to the system connection interface 262 (step S243), and the awareness information data is output to the visualization unit (step S244). In this example, the awareness information visualization unit visualizes the information on a map, CRM, and Excel / Word.
[0154] Fig. 48 shows an example of a screen in the notice information extraction unit 240, taking as an example the management task of documents entrusted to the customer (Fig. 48(A)). In this case, the processing date of the case is stored in the date and time unit 232 of the centralized integrated storage device 230, and the notice information extraction unit 240 extracts the notice information shown in Fig. 48(C) by comparing this date and time information with the notification (deadline) master set in advance in DB 104 as shown in Fig. 48(B). Since the master registration can be set for each notice, it is possible to extract, for example, items that are more than 30 days old or items that have been processed for one week.
[0155] Documents, etc. (seal certificate, mortgage certificate, etc.) required for the sale and purchase of automobiles, houses, etc., and loan applications, etc. are temporarily kept by customers and stored, but the necessary processing must be performed on the day of the case processing of the sale and application, and measures such as periodic inventory are also required. If this is managed by a paper ledger, the burden of work will be large due to the entry at the time of deposit, daily confirmation, inventory, etc. When the awareness information construction system 200 of this embodiment is introduced, as shown in FIG. 48 (A), the necessary information is automatically stored in the centralized storage device 230 by simply registering the documents, etc. at the time of deposit on the screen (existing system cooperation unit 140), and the awareness information extraction unit 240 extracts and stores it in the awareness aggregation storage device 250. Then, when the case processing day of the sale and purchase or application arrives, a notification is made via the output cooperation interface 260, and if there are documents that have been kept for a long time, the person in charge is notified via the output cooperation interface 260 on a screen as shown in FIG. 48 (C). This significantly reduces the burden and effort placed on staff in charge of checking ledgers and conducting inventory.
[0156] If information about documents that have already been processed remains in the notification, it becomes difficult to know which documents have been processed or not. In the case of conventional paper ledger management, a check is made in a "processed" column to distinguish whether or not a document has been processed, but in the case of this embodiment, by deleting processed documents from the notification list, it becomes possible to distinguish the notification list as unprocessed documents.
[0157] Figure 49 shows the data flow illustrating an example of the extraction of notice information by the notice information extraction unit 240. The extraction conditions for this example are as described above for "customers who are in the wealthy bracket, have not traded any products in the promoted product information, and have not visited for more than a month in terms of visit frequency."
[0158] 49, the contents of customer database 301 of accounting system host 300 serving as a host computer (including a PC) are stored in customer database 230-1 in centralized aggregated storage device 230 via input linkage interface 220. Also, product list 506-1, product purchase history 506-2, and visit history 506-3 in CRM 506 are stored in product list 230-2, product purchase history 230-3, and visit history 230-4 in centralized aggregated storage device 230 via existing system linkage unit 140 (system connection IF 141, centralized aggregated storage device IF 142). For the product list 230-2, product purchase history 230-3, and visit history 230-4 in the centralized centralized storage device 230, the extraction conditions for the notice information are set as "wealthy class (savings balance of 10 million yen or more)" as customer information, "no purchase history of product name "xxxxx"" as recommended product, and "visit frequency of not visiting for more than one month" as visit history, so data matching the extraction conditions is extracted. That is, relevant customers matching the extraction condition "wealthy class (savings balance of 10 million yen or more)" are extracted from the customer database 230-1 in the centralized centralized storage device 230, relevant products and purchase histories matching the extraction condition "no purchase history of product name "xxxxx"" are extracted from the product list 230-2 and product purchase history 230-3 in the centralized centralized storage device 230 to create a relevant product list and a list of no purchase history, and relevant visit histories matching the extraction condition "visit frequency of not visiting for more than one month" are extracted from the visit history 230-4 in the centralized centralized storage device 230. From the data extracted in this way, data that matches all the conditions is extracted and generated as noticed information. In this example, the noticed information is "customers who are in the wealthy bracket, have not traded any products in the promoted product information, and have not visited for more than one month in the visit frequency category," and the data are "customer number," "customer number and name," "deposit balance," "product number," "product name," "purchase history," and "last visit date."
[0159] Next, an example of the overall operation of this embodiment will be described with reference to the operation flow diagrams of FIGS. 50 to 53 and the screen diagrams of FIGS.
[0160] In the business process restructuring support system 100, the person in question is identified based on the contents of the hearing sheet (step SB1), the consultation is accepted (step SB6), the application form is filled out, explanatory documents are submitted (step SB10), the application details are entered, the explanatory documents are registered (step SB11), and the application is approved (step SB13). In addition, in the business process restructuring support system 100, identity identification (step SB2) is also performed for a mortgage pre-screening application (step SC1) submitted by mail or in person. In the case of an application submitted by mail or in person, the application form is filled out and identity verification documents are copied (step SC2), the application details are entered, the explanatory documents are registered (step SB12), and the application is approved (step SB13). In the case of an online loan application (step SC3), the application details are entered and identity verification documents are attached (step SC4), and the application is approved (step SB13).
[0161] The business reform support system 100 then sends a review request to the review system (step SB20), the review system makes a review decision and sends the decision result (step SC10), and the business reform support system 100 receives the review result (step SB21). If the bank's own review is negative, the business reform support system 100 requests a review from a guarantee company, which is an external review agency (step SB30), the guarantee company makes a review decision and sends the decision result (step SC11), and the business reform support system 100 receives the review result (step SB35).
[0162] Thereafter, the business restructuring support system 100 approves the loan (step SB40) and sends an inquiry for a compliance check to the compliance check system (step SB41). The compliance check system checks whether the received inquiry content has any compliance problems and sends the inquiry result to the business restructuring support system 100 (step SC33).
[0163] The business restructuring support system 100 receives the inquiry results (step SB42), and a bank employee checks the results, contacts the customer by phone or e-mail, and registers the contact details. The business restructuring support system 100 uses the accounting system host linkage function to send linkage data to the accounting system host (a host computer including a PC), and the accounting system host 300 receives the linkage data (step SC40). At timing T1, the customer's employment confirmation and linkage to the accounting system host 300 are completed, so the employment confirmation results are registered (step SB50). After that, the approval conditions are lifted (step SB51), and the lifting of the approval conditions is confirmed (step SB52).
[0164] Then, the principal's consent is obtained, and the BUSINESS REFORM SUPPORT SYSTEM 100 performs data linking and obtains the principal's consent via email, the Web, etc. (STEP SC50). The BUSINESS REFORM SUPPORT SYSTEM 100 receives the consent result (STEP SB53), and the bank employee executes the loan using the execution terminal (STEP SC60). At this timing T3, the execution conditions have been met.
[0165] Thereafter, the post-conditions removal work is performed in cooperation with the collateral evaluation system 510 (step SB60), and the collateral evaluation system 510 performs an evaluation of the mortgage collateral property and returns the evaluation result to the business reform support system 100 (step SC70). Thereafter, the removal of the post-conditions is confirmed and the process is completed.
[0166] 54 to 59 are screen examples of FIG. 50 to FIG. 53, respectively, where screen DP1 corresponds to process SB1, screen DP2 corresponds to process SB6, screen DP3 corresponds to process SB11, and screen DP4 corresponds to process SB13. Screen MD1 is an noticed information display screen of this system used when opening screen DP4. Screen DP6 corresponds to process SB40, and screen MD3 corresponds to an noticed information display screen when opening screen DP6. Screen DP7 corresponds to process SB43, screen MD4 corresponds to an noticed information display screen when opening screen DP7, screen DP9 corresponds to process SB51, and screen MD6 corresponds to an noticed information display screen when opening screen DP9. Screen PL5 indicates process SC50, screen DP11 corresponds to process SB60, screen MD8 corresponds to an noticed information display screen when opening screen DP11, screen DP12 corresponds to process SB62, and screen MD9 corresponds to an noticed information display screen when opening screen DP12.
[0167] Figure 60 is a future outlook for the business process reform support system, which will serve as an entry point for AI as a result of the fusion. In other words, it will be possible to automate the extraction of insight information using AI, analyze the success patterns of top performers and reflect this in the extraction of insight information, acquire potential customers, and predict new markets and products. When the business process reform support system is applied to a personal loan business process reform support system, it will be possible to extract insight information from negotiation history, visit frequency, story information, and business matching, making it possible to propose additional products (increased revenue) and change visiting plans (gain trust), which will result in an increase in the number of loans and the optimization of improved loan quality.
[0168] Companies have common challenges such as reducing system-related costs, improving work efficiency, and resolving problems, and the applications of business reform support systems that integrate existing systems are endless. They can also be applied to linking with other industries, map data, and digital libraries, and to building systems that use 2D and 3D CAD data.
[0169] Furthermore, while the above example is described taking a financial institution as an example of a company, the invention can be similarly applied to a variety of other companies (such as automobile manufacturers, beverage manufacturers, and daily necessities retailers).
[0170] [Fourth embodiment] In this embodiment, an embodiment using a large-scale language model as an expert system will be described. As an example of the expert system, the description will be given with reference to Japanese Patent Application Laid-Open No. 2005-339529.
[0171] <Summary> The expert system generation device of the present invention is an expert system generation device for constructing an expert system, and comprises: case acquisition means for accumulating situations and associated results observed from an external system for which the expert system is to be constructed; rule generation means for automatically learning the accumulated situations and results and deriving, as knowledge, a statistical tendency for the results to arise from the situations; and consulting means for automatically determining the results for an observed situation from the derived statistical tendency, and is characterized in that the situation and the results are updated and automatically updated to more appropriate knowledge by operating the case acquisition means, the rule generation means, and the consulting means independently or in conjunction with each other.
[0172] In the expert system generating device of the present invention, the case acquisition means first accumulates the situations and the associated results that can be observed from the external system for which the expert system is to be constructed. Then, the rule generation means automatically learns the accumulated situations and results, and derives the statistical tendency of the results from the situations as knowledge. Here, the external system refers to all systems to which the expert system is applied, such as natural changes and human actions, as well as artificially manufactured devices such as manufacturing devices and processing devices. The knowledge as the statistical tendency thus derived is not immutable, and the situation and results are updated by operating the case acquisition means, the rule generation means, and the consulting means independently or in cooperation with each other, and the knowledge is automatically updated to a more appropriate knowledge. Therefore, even if the knowledge automatically determined by the consulting means at first lacks validity, the situation and results are updated thereafter, and the knowledge is automatically updated, so that the knowledge can change to one with high validity. Therefore, the burden of employing a skilled knowledge engineer to obtain highly valid knowledge from the beginning, accurately extracting specialized knowledge from experts, organizing the specialized knowledge, and selecting and discarding it is reduced. It also reduces the burden of interviewing experts and organizing the problems from the records, and accurately confirming facts and rules, questions and answers, causal links, premises and conclusions, probability functions, etc. Therefore, it becomes possible to generate an expert system even if the knowledge engineer is not well trained and skilled. Also, whereas the burden on knowledge engineers in the past was continuous, this expert system generation device significantly reduces this continuous burden. Furthermore, the effort, time, and cost required to build an expert system are reduced.
[0173] Therefore, the expert system generating device of the present invention places less of a burden on knowledge engineers, makes it possible to generate appropriate expert systems without much effort or time, and is highly versatile.
[0174] In the expert system generation device of the present invention, the case acquisition means acquires and accumulates, as data, situations and results relating to an external system for which an expert system is to be constructed, the rule generation means automatically extracts statistical trends in the form of rules from the data acquired and accumulated by the case acquisition means, and the consulting means refers to the rules generated by the rule generation means and automatically determines the results for a situation from the current data of the external system, the case acquisition means, the rule generation means and the consulting means can be operated independently, and the rule generation means can cooperate with the case acquisition means and the consulting means to automatically update the rules as the data accumulates, thereby updating them to rules that are more suited to the situation.
[0175] In this case, data on the external system for which the expert system is to be constructed is first acquired by the case acquisition means, and the rule generation means automatically extracts the causal relationships between the data in the form of rules. Here, the data on the external system means numerical values, symbolic values, etc., that express the state of the external system. The rules thus generated are not immutable, but are updated automatically as subsequently acquired data accumulates, so that they become more suitable for the situation. Therefore, even if the rules initially generated lack validity, the rules can gradually become more valid as subsequent data is acquired. This reduces the burden on knowledge engineers, makes it possible to create appropriate expert systems without much effort or time, and is highly versatile.
[0176] In such an expert system generating device, data acquired by the case acquisition means is treated as a case consisting of attributes and attribute values, and a conclusion arrived at by the consulting means can be obtained as a conclusion attribute value.
[0177] The rules generated by the rule generating means in the expert system generating device of the present invention can be generated in the form of a decision tree, a Prolog expression, a logical expression, a function in a statistical analysis method, or the like.
[0178] When the rule generating means generates rules in the form of a decision tree, it is preferable that the decision tree can be automatically generated regardless of whether there are too many or too few cases. When trying to actually obtain cases from an industrial site, some of the cases may be unknown. In addition, it may not be possible to accurately determine in advance what kind of cases are necessary and what kind of cases are unnecessary. For this reason, if it is possible to automatically generate a decision tree regardless of whether there are too many or too few cases, such as cases with unknown condition attribute values or cases with unknown conclusion attribute values, it is possible to generate a decision tree for the time being and later update it to a more valid decision tree by the automatic decision tree updating means. For this reason, the burden on knowledge engineers can be further reduced and the range of applicability can be further expanded. As a specific method for automatically generating a decision tree regardless of whether there are too many or too few cases, there is a method of skipping unknown values so as not to affect the probability when calculating the amount of information.
[0179] It is also preferable that the rule generating means can set the conclusion attribute value to a fixed value for a given case. In this case, when the attribute value of the conclusion attribute is obviously wrong, it is possible to correct it to an appropriate value. Also, when the conclusion is uniquely determined under any condition, it is possible to set the attribute value of the conclusion attribute to an appropriate value. Therefore, the inference rules obtained by the expert system generating device of the present invention can be made more tailor-made and more suited to the current situation.
[0180] Furthermore, when prior knowledge is available regarding the distribution of attribute values of cases to be acquired, it is preferable that the rule generating means adds cases including an unknown conclusion attribute value derived as a result of tracing a decision tree once generated, In this case, by generating missing cases based on condition attributes that appear on the decision tree until the unknown conclusion attribute value is reached, it becomes possible to generate an appropriate decision tree even when a small number of conclusion attribute values are unknown.
[0181] It is also preferable that the consulting means automatically acquires the condition attribute values from outside and derives the conclusion attribute values from the already acquired decision tree. In this case, consulting can be automated. An example of such consulting means is a means for automatically starting an external system in the form of an executable program and converting the output value of the external system into the condition attribute value and acquiring it. In this case, for example, if a program for accessing the outside via a network exists on the computer on which the expert system generating device is running, it is possible to control the external program of another device connected via the network via that system and acquire cases.
[0182] Furthermore, it is preferable that the consulting means applies the conclusion attribute value to the external system. In this case, it is possible to automatically control the external system based on the result of the consulting. As an example of such a consulting means, the consulting means may automatically start the external system in an executable program format, convert the conclusion attribute value obtained by the consulting means into an input value for the external system, and deliver it to the external system.
[0183] In addition, it is preferable that the rule generating means evaluates the validity of the generated decision tree and updates it to a more valid decision tree based on the result. Examples of such rule generating means include those that allow exceptions to cases, exceptions to condition attribute items, and pruning of branching branches. If exceptions to cases and exceptions to condition attribute items are possible, a more realistic and valid decision tree can be generated. If branching branches are possible, the decision tree is simplified and analysis becomes easier when the case becomes unnecessarily complicated and difficult to analyze. Here, it is preferable that branching branches can be pruned both between any condition attribute items and between the condition attribute and the conclusion attribute. In this case, it is possible to change, simplify, or delete only a part of the decision tree, and it is possible to edit the decision tree to be more realistic. In addition, it is also possible that, when multiple conclusion attribute values are derived for one condition case as a result of tracing the decision tree, the decision tree can be regenerated with the corresponding case as an exception. In this way, it is possible to generate a decision tree even when the conclusion attribute values in the current case are inconsistent and conflict with each other.
[0184] It is preferable that the case acquisition means is capable of automatically acquiring cases. In this case, the decision tree can be automatically updated and the labor required by the operator to acquire cases can be significantly reduced. An example of the case acquisition means is a means for automatically starting an external system in the form of an executable program and converting the output value of the external system into a condition attribute value to acquire the case. In this case, for example, if a program for accessing the outside via a network exists on the computer on which the expert system generation device is running, the external program of another device connected via the network can be controlled via the system to acquire the case.
[0185] When the rule generation means generates rules in the form of a decision tree, it is also preferable to accumulate information that experts have gained through experience as cases and extract, as a decision tree, knowledge having a multiple knowledge structure including the same knowledge elements, a hierarchical knowledge structure in which knowledge elements constituting one piece of knowledge are a higher-level concept or a lower-level concept of another piece of knowledge, or a multi-viewpoint knowledge structure which is a special knowledge structure of the multiple knowledge structure and is a collection of knowledge from many different perspectives.
[0186] In this way, it is possible to extract knowledge having a multiple knowledge structure, a hierarchical knowledge structure, or a multi-perspective knowledge structure as a decision tree, making it easier to grasp knowledge elements from various aspects and in a multi-layered manner, and enabling the generation of more accurate decision trees.
[0187] In this case, the rule generating means is preferably configured as follows to generate a decision tree according to the purpose: (1) to be able to arbitrarily select either a condition attribute or a conclusion attribute for each attribute item of a case; and (2) to generate only decision trees corresponding to a specific conclusion attribute.
[0188] The rule generating means can have a plurality of conclusion attribute items for a case base in which cases are accumulated, and can be provided with a user interface for comparing decision trees corresponding to each conclusion attribute. In this way, it becomes easy to determine which of the plurality of conclusion attribute items is an important conclusion attribute.
[0189] Furthermore, the rule generating means may be capable of expanding and concatenating another decision tree to the conclusion attribute value of the generated decision tree, which enables further expansion of causal relationships and allows for the development of an analysis of causal relationships.
[0190] The rule generating means may also be capable of detailed concatenation, which connects a condition attribute item of a generated decision tree with another decision tree having all attribute values of the condition attribute item as conclusion attribute values. In this case, a more detailed decision tree can be generated for any condition attribute value, making it possible to obtain more detailed causal relationships.
[0191] Furthermore, when the conclusion attribute value is a numerical value, the rule generating means can approximate the conclusion attribute value with a regression line and obtain the approximation residual as the range of the conclusion attribute. In this way, the numerical value of the conclusion attribute can be statistically divided into a numerical range, and a rule can be constructed as the conclusion of the decision tree, so that it becomes possible to estimate the range of the conclusion value that has not been observed as a case.
[0192] The expert system generator of the present invention can be used to implement the method of implementing the expert system generator of the present invention.
[0193] Furthermore, in the method of implementing the expert system generation device, the actions of experts can be automated using the rule-based knowledge of the experts. This makes it possible to absorb and utilize the knowledge of experts in all industries, including manufacturing, mining, agriculture, forestry, fisheries, transportation, medical care, finance, and commerce. For example, the rule-based knowledge of experts can be used to support human work and educate people.
[0194] The expert can be an expert on industrial processes. In this case, the knowledge of the expert on industrial processes can be used as an asset to automatically operate the industrial process, to support the operator of the industrial process, to educate the operator of the industrial process, and the like. Here, the industrial process includes petrochemical plants such as refining plants using a rectification tower, robots, steel production sites, ceramic product production sites, power plants, medical sites, and other equipment and machines in various industrial sites. Refining plants using a rectification tower have traditionally been operated by skilled experts based on their many years of experience and intuition, based on the temperature inside the rectification tower, the sound generated during boiling, and the state of bubbles observed through an observation window in the rectification tower. If the implementation method of the expert system generation device of the present invention is used for such refining plants, it is possible to control them in the same way as a skilled expert, which is preferable.
[0195] In the case where the industrial process is a refining plant using a rectification tower, the acquisition of cases from the refining plant can be performed via at least one of a temperature sensor, a pressure sensor, a flow rate sensor, a sound sensor, and an image sensor installed in the rectification tower. In this case, it becomes possible to automatically capture as cases the temperature, pressure, flow rate, sound generated during boiling, and bubble conditions generated during boiling at various positions in the rectification tower, which are important attributes and attribute values for controlling the refining plant using the rectification tower. In addition, as a method of acquiring cases in the case acquisition means, a device for data acquisition that has already been instrumented can be directly connected to the case acquisition means. In this case, past data recorded in a device for data acquisition installed in the industrial process can be effectively used.
[0196] <Description of the embodiment> Hereinafter, an embodiment embodying an expert system generating device and an implementation method thereof will be described with reference to the drawings.
[0197] As shown in FIG. 61, the expert system generating device of the embodiment includes a case acquisition means 10, a rule generation means 20, and a consulting means 30. The case acquisition means 10 includes a case acquisition interface 10a for acquiring cases observed in an external system that is the target of case acquisition. The external system is the target for building an expert system. Data on various cases sent from various sensors and the like provided in the external system is automatically stored in a data file as cases consisting of attributes and attribute values of each data via the case acquisition interface 10a, forming a case base. The case acquisition means 10 is also capable of directly inputting data from a keyboard, and further, it is also possible to directly access the storage unit of the control device via the case acquisition interface 10a, and automatically acquire various recorded operating data as cases (situations and associated results) by controlling the sequencer of the external system, or read cases from a text file in a specified format.
[0198] The rule generation means 20 is composed of a generation preparation block 21, a tree generation block 22, and a tree check block 23. The rule generation means 20 automatically learns accumulated situations and results, and derives statistical tendencies that lead to the results from the situations as knowledge. The generation preparation block 21 checks whether a decision tree can be generated from a case (for example, when the case does not have a conclusion attribute (hereinafter referred to as "target element"), or when there is no case), and determines the boundary value of the numeric conclusion attribute value (hereinafter referred to as "target element value") by using a regression line.
[0199] As shown in Fig. 62, the tree generation block 22 is made up of a part related to a block function associated with ID3 and a part related to a block function associated with branch generation of a decision tree, and performs calculations for generating a decision tree based on the data created by the generation preparation block 21. The block functions associated with ID3 are made up of a numeric type binarization block function and an information amount calculation block function. The numeric type binarization block function binarizes two numeric type or date type element values and divides the binarized element values into two clusters. The information amount calculation block function performs calculations related to information amount, such as the information amount and information entropy of each element.
[0200] Furthermore, the tree check block 23 checks the decision tree generated by the tree generation block 22, such as whether or not there are duplicated target element values, as shown in FIG.
[0201] As shown in Fig. 64, the consulting means 30 is composed of a consulting preparation block 31, a consulting block 32, and an output block 33. The consulting means 30 automatically judges the results for a certain observed situation from the derived statistical tendency.
[0202] The consulting preparation block 31 is a block for acquiring information necessary for consulting (data to be traced to a decision tree, information on detailed connections and expanded connections, information on fix target cases described below, etc.). The consulting block 32 is a block for tracing consulting data to a decision tree and performing consulting. The output block 33 is a block for outputting the results of consulting performed by the consulting block 32, and includes the output of target elements and their values obtained in consulting, element information on conditional branches after a tracing failure, and the output of tracing results. Use force.
[0203] The expert system generation device updates the situation and the results by operating the case acquisition means, the rule generation means, and the consulting means independently or by linking them together, and automatically updates them to more appropriate knowledge.
[0204] The case acquisition means 10, the rule generation means 20 and the consulting means 30 are configured using a large-scale language model. Also, the expert system generated by the expert system generation device based on the cases can be realized using a large-scale language model.
[0205] <Flowchart> The expert system generating device configured as above generates a decision tree in the following procedure, and further performs consulting based on the decision tree. (Generating decision trees) The generation of a decision tree is performed according to the flowchart shown in FIG. 65. That is, in step S000, the exception flag is checked based on the cases acquired from the external system by the case acquisition means 10, and the cases other than the exception case are registered in the memory. Next, in step S100, it is checked whether or not there are cases, and whether or not all the cases are exceptional cases (here, the exceptional cases refer to cases in which the target element value determined exceptionally is adopted regardless of the value of the target element value based on the generated rule), and it is determined whether or not the generation of a decision tree is possible. As a result, if it is determined that the generation of a decision tree is impossible, the process is terminated. On the other hand, if it is determined that the generation of a decision tree is possible, in step S200, the range of the numerical target is obtained when the target element is a numerical target. That is, a simple regression analysis is performed on the condition attribute value (hereinafter referred to as "element value") and the target element value in each case, and a regression line is obtained. Then, the obtained regression line is used to obtain the ratio of the difference between each target element value and the regression value, and a boundary value is obtained from the ratio.
[0206] Then, in step S300, a decision tree is generated. Step S300 is shown in more detail in Fig. 66. That is, first, in step S301, it is checked whether a decision tree already exists, and if it does exist, all nodes in the decision tree are deleted in step S302, and if it does not exist, tree information such as the theme name and target of the target element is registered in the tree database. Then, in step S304, the top node is registered in the tree database, and then in step S305, calculation is performed using ID3.
[0207] FIG. 67 is a flowchart showing the calculation by ID3 in step S305 in more detail. First, in step S3051, preprocessing is performed to calculate the amount of information. Specifically, as shown in FIG. 68, it is determined whether the case data is of logical type (i.e., either numeric type or date type) or not, and if it is not of logical type, binarization is performed in step S30511. FIG. 69 is a flowchart showing step S30511 in more detail. That is, first, in step S305111, each element value is counted, and further, in step S305112, an expected value is calculated. Then, if the expected value is equal to or less than a predetermined value, a Fisher's test is performed (step S305113), and if the expected value exceeds the predetermined value, a chi-square test is performed (step S305114). As a result, if binarization is possible, the boundary value is obtained (step S305115), and the above process is repeated from step S30511 for another element value. Furthermore, if binarization is not possible, the above process is repeated from step S30511 for another element value without obtaining the boundary value. After repeating the process (number of cases - 1) times in this manner, a cluster is created from the obtained boundary value in S305116. Then, as shown in FIG. 68, in step S30512, the number of element values and the number of target element values are counted for each element value in the case. In this case, if an unknown value exists in the element value, the unknown value is not counted. If the value of the element value is a numeric type or a date type, counting is performed for each value.
[0208] Then, as shown in FIG. 67, in step S3052, an information element is obtained from each element, and this is looped for the number of elements to obtain the overall set entropy and information amount. Furthermore, in step S3053, the acquisition of the element with the largest amount of information is repeated for the number of elements. Then, in step S3054, the element and the element value are registered. Furthermore, in step S3055, if the flag for pruning of the branch is on and the number of nodes from the top matches the specified number of pruning nodes, the target element and the target element value are registered. That is, in step S3055, as shown in FIG. 70, first, in step S30551, it is determined whether the value is the same as the value of a target element (hereinafter referred to as a "fix target element") whose value is set in advance, and if it is not the same, it is determined in step S30552 whether a value exists. Then, if a value exists, in step S30553, the operation of acquiring the target value of the matching case is repeated for the number of cases. Also, in step S30551, if the value is the same as the value of the fix target element, it is determined whether or not the value exists. If the value exists, in step S30553, the operation of obtaining the target value of the matching case is repeated the number of times corresponding to the number of cases. In this manner, the fix target value is obtained and step S3055 is terminated. On the other hand, if the value does not exist in step S30552, step S3055 is terminated without performing steps S30553 and S30554.
[0209] On the other hand, if the flag for pruning of branches in Fig. 67 is on and at least one of the conditions that the number of nodes from the top matches the specified number of pruned nodes is not met, a target reaching determination is made in step S3056 as shown in Fig. 67. That is, in this step S3056, as shown in Fig. 71, if the entropy of the element value is zero and the amount of information of all elements is zero or the number of element values is 1 or less, the target is reached on the previous branch. Also, if the entropy of the element value is zero and the amount of information of all elements is other than zero or the number of element values is 1 or less, the target is reached.
[0210] On the other hand, if the entropy of the element value is not zero, in step S30561, a determination is made as to whether the target has been reached based on the amount of information. That is, in step S30561, as shown in Fig. 72, if the element has the most amount of information and the amount of information of the element is zero or less, the target is reached at the previous branch. Also, for elements other than the element with the most amount of information, it is determined whether the amount of information of the element is greater than zero, and if the amount of information of the element is greater than zero, the next condition element is searched for, and if not, the target is reached.
[0211] Then, if the target has not been reached as a result of determining in step S3056 shown in FIG. 67, then in step S3057, examples for the remaining elements are reconstructed for each element value of the previous element, and further in step S3058, calculations are performed based on ID3.
[0212] After the decision tree is thus created in step S300 shown in Fig. 65, the target values are checked in step S400. That is, as shown in Fig. 73, first, the number of target values of the node is obtained in step S401. Then, if the number of target values is two or more, processing of multiple target values is performed in step S402. A flowchart showing step S402 in more detail is shown in Fig. 74. That is, the following options are available: (1) display multiple target values as they are, (2) invalidate the values and delete the target value node, (3) stop generating the decision tree, or (4) forcefully determine the target values.
[0213] Also, if the number of target values is 0 (i.e., the target value is unknown), processing without a target value is performed in step S403. That is, as shown in Fig. 75, a selection is made as to whether (1) the target value is to be ignored as invalid, (2) generation of a decision tree is to be stopped, (3) a missing case is to be generated based on elements that appear on the tree until the unknown target value is reached, or (4) the target value is to be forcibly assigned by a value selection requester.
[0214] Finally, in step S500 shown in FIG. 65, a bit is set only for those that are to be generated for the exception processing, and the generation of the decision tree is completed.
[0215] (consulting) Based on the decision tree generated as described above, consulting is performed according to the flowchart shown in Fig. 76. That is, first, in step S600, the example file to be consulted, the consulting environment setting file (a file in which selection information on whether to automatically obtain element values from outside, obtain them manually, output remaining elements that have not been traced, or generate a style again is registered), and the history file (a file that records elements that have already been traced in the process of tracing the tree) are read. Next, in step S700, the decision tree is traced. In step S700, as shown in Fig. 77, if the node of the tree is an element, the element value is traced in step S701, and if the node of the tree is a target element, the target element value is traced in step S702.
[0216] The tracing of element values in step S701 is performed according to the flowchart shown in FIG. 78. That is, first, in step S7011, it is determined whether values can be obtained from the history file called in in step S600, and if values can be obtained, a log is taken in step S7012 and the tree is traced. On the other hand, if values cannot be obtained, it is determined whether a detailed connection flag is present, and if the detailed connection flag is present, a tree related to the detailed connection is traced in step S7013. On the other hand, if the detailed connection flag is not present, it is determined whether or not to obtain the values using an external program, and if so, the values are obtained from the external program, and a log is taken in step S7012 and the tree is traced. Also, if values are not obtained from an external program, one of the following is selected and performed: (1) a standard deviation is obtained to obtain values, and then a target element value is traced, (2) values are manually input and the tree is traced, (3) the remaining elements are obtained and output to a file, or (4) the tree is reconstructed again.
[0217] As shown in Fig. 79, the tracing of the target element value in S702 involves comparing the fix case with the consulting process (the condition attribute passed through while tracing the tree) in S7021, and determining whether there is an expansion connection, and if there is no expansion connection, starting an external program based on the obtained target element value in step S7022. If there is an expansion connection, tracing the tree based on it in step S7023. Thus, the consulting ends.
[0218] The following will provide a more detailed explanation of the case where a decision tree is generated using this expert system generation device and consulting is carried out.
[0219] <Case acquisition process S_1> First, in the case acquisition step S_1, a theme to be performed by the expert system is determined, and elements (condition attributes) and element values (condition attribute values) that can be used for the theme are registered. Furthermore, a target element (conclusion attribute) and a target element value (conclusion attribute value) are registered. Registration can be performed automatically from various sensors provided in an external system via a case acquisition interface 10a shown in FIG. 61. Data can also be input directly from a keyboard, and if the case base is in a text file format, it can be loaded. Furthermore, it is also possible to directly access the storage unit of the control device via the case acquisition interface 10a and automatically acquire various recorded operation data as cases. The cases registered in this way are displayed on the computer screen as a case editing panel as shown in FIG. 80, and are also filed in the computer. Furthermore, the registered elements and target elements can be freely added and deleted. Furthermore, "cases" consisting of combinations of elements, element values, target elements, and target element values can be freely added and deleted. Furthermore, elements and target elements can be freely interchanged. Furthermore, if the conditions indicate that the target element value is clearly incorrect, or if the target element value is determined under certain conditions, a fixed value can be registered as a fixed target element value. (You can also select a case you want to make an exception to in the case editing panel, and click "Make exception" in the edit menu to select any case as an exception.
[0220] <Decision tree generation process S_2> Next, by selecting "Generate decision tree" from the function menu on the case editing panel displayed on the computer screen, the user can select whether to generate decision trees corresponding to all target elements or to select only the currently selected decision tree. By selecting "Display decision tree" from the function menu, the decision tree display panel shown in Figure 81 is displayed. Furthermore, by selecting any node in the decision tree display panel with the mouse and clicking "Prune" from the function menu, the decision tree is pruned (Figure 82). Here, pruning a decision tree is the same operation as making an exception of the corresponding case. After that, by moving it to the case editing panel, cases that will become unnecessary when the decision tree is reconstructed are automatically processed as exceptions. Pruning can be performed not only between elements, but also between elements and target elements.
[0221] In addition, when analyzing the generated decision tree, it is possible to generate a decision tree based on attributes other than the top node by making the top node an exception in the style editing panel. For example, Figure 83 shows the top node of a decision tree that results from the determination of the melting grade from the gloss and degree of melting of a test piece of incineration ash from sludge treatment, and the collection of various data on sludge treatment. Here, by making the top node element "incineration plant" an exception, a decision tree that is not dependent on the incineration plant can be derived. As a result, as shown in Figure 84, the top node has the element "MgO", and it can be seen that the content of magnesium oxide is an important factor in addition to the incineration plant element.
[0222] In addition, when generating a decision tree, if the majority of the decision tree is occupied by the same element, the element is likely to be noise that specializes the rule. In such a case, by excluding the same element as an exception, it becomes easier to find important elements that were overlooked. For example, by making the "date" element an exception in the decision tree shown in Figure 85, the "SiO2" element appears as shown in Figure 86. Since silicon oxide has the function of increasing the melting of materials, it is presumed that it has a large effect on the melting grade value. This inference can be easily judged by observing the decision tree when the "SiO2" element is made an exception. In this way, by repeating inference and rule induction, the analysis of the decision tree can be carried out efficiently.
[0223] Also, another decision tree is connected to the target element value of the generated decision tree. When performing an expand connection, a target element value selection requester is displayed using the edit menu, and a target element value to be expanded is selected from a list of expandable target element values. When the expand connection is completed in this way, an expand connection mark is displayed in front of the target element value name in the style edit panel (not shown). In this way, a decision tree can be further expanded and generated for any target element value, making it possible to further expand causal relationships and develop the analysis of causal relationships. Furthermore, when performing detailed concatenation by connecting the element values of the generated decision tree to another decision tree having all element values as target element values, an attribute item selection requester is displayed using the edit menu, and the element to be detailed is selected by clicking on it from among the displayed elements. When detailed concatenation is completed in this way, a detailed concatenation mark is displayed in front of the connected element name in the style editing panel (not shown). In this way, a more detailed decision tree can be generated for any element value, making it possible to obtain more detailed causal relationships. <Consulting process S_3> After the decision tree is generated in the decision tree generation step S_2, select "Style-Consulting" from the menu at the top of the screen in the case editing panel (not shown). This displays a target element selection dialog, allowing the selection of the target element. When the target element is selected, consulting according to the decision tree for that target element is started. The consulting means 30 is capable of automatically starting an external system in an executable program format, converting the output value of the external system into an element value, and acquiring it. The consulting means 30 is also capable of automatically starting an external system in an executable program format, converting the target element value acquired by the consulting means 30 into an input value for the external system, and handing it over. Then, answers according to the element type, such as logical, numeric, or date type, are selected or input, and the value of the target element is derived as a conclusion when all answers have been input. Here, when the element is logical, input is merely a selection. Even for a numeric element such as age, if the age of a person is unknown, it can be treated in the same way as a logical element, with "age unknown". In this way, even when there is a lack of data, it is possible to generate a decision tree and provide consulting.
[0224] (Operating a solvent recovery plant using a distillation column) The expert system generating device according to the embodiment was used to operate a solvent recovery purification plant using a rectification column.
[0225] This solvent recovery plant is for separating and recovering tetrahydrofuran (THF) and methyl ethyl ketone (MEK) from a recovered solvent, and is equipped with a storage tank 11 for storing a recovered solvent in which THF and MEK are mixed, and two rectification towers 12 and 13, as shown in FIG. 34. Pipes 11a and 11b are connected to the storage tank 11, and the other ends of the pipes 11a and 11b are connected to the ninth stages of the rectification towers 12 and 13. Heat exchangers 12a and 13a are provided at the lower ends of the rectification towers 12 and 13, and the lower ends of the rectification towers 12 and 13 can be heated by supplying steam from the outside. In addition, bottom pipes 12b and 13b are connected to the lower ends of the rectification towers 12 and 13, and the bottoms accumulated at the lower ends of the rectification towers 12 and 13 can be taken out. Furthermore, distillation pipes 12c, 13c are connected to the upper ends of the distillation columns 12, 13, and the distillation pipes 12c, 13c are connected to one ends of the condensers 14, 15. The other ends of the condensers 14, 15 are connected to catcher tanks 16b, 17b via recovery pipes 16, 17. Furthermore, reflux pipes 16a, 17a branch off from the middle of the recovery pipes 16, 17 and are connected to bottom pipes 12b, 13b, so that a portion of the bottoms can be refluxed to the distillation columns 12, 13.
[0226] As information transmission means for grasping the operating status of this solvent recovery refining plant and obtaining information related to cases, thermometers TR1 to 11, flow meters V1 to 8, and pressure gauges P1 and P2 (not shown) are installed at the positions shown in Table 1.
[0227] [Table 1] The solvent recovery plant was operated by three experienced operators.
[0228] (Acquisition of cases) The hourly values of the thermometers TR1-11, the flowmeters V1-8, and the pressure gauges P1 and P2, as well as the temperature and date and time, were automatically imported into the case acquisition means 10 via the automatic acquisition interface 1a from the recording section of a control device (not shown). In addition, the weather, the operator, whether or not an instrument check was performed, etc. were directly entered into the case acquisition means 10 from the keyboard. In this way, the case acquisition means 10 created a case base divided and shown in Tables 2-4.
[0229] [Table 2]
[0230] [Table 3]
[0231] [Table 4] (Decision tree generation) Furthermore, a decision tree was generated by the rule generating means 20 based on the case base.
[0232] (consulting) Furthermore, consulting was provided based on the decision tree thus generated.
[0233] By applying this expert system generating device to a solvent recovery refining plant, a case base is automatically formed by the case acquisition means 10 using information from the thermometers TR1-11, the flowmeters V1-8, and the pressure gauges P1 and P2, and a decision tree can be automatically generated from this case base by the rule generation means 20. The consulting means 30 can automatically obtain the target element value of the target element by referring to this decision tree. Therefore, the decision tree initially generated by the rule generation means 20 is automatically updated continuously as the case base is accumulated, and is updated to a more suitable decision tree in terms of case learning. Even if the causal relationship changes over time due to some cause, an appropriate decision tree is generated following the change. Therefore, the target element value obtained by the consulting means 30 can also be a more suitable value.
[0234] In the above embodiment, information from the thermometers TR1-11, flow meters V1-8 and pressure gauges P1 and P2 is imported from the memory unit of the control device, but this information can also be transmitted in real time to the case acquisition means 10 via the automatic acquisition interface 1a.
[0235] This embodiment can be used in any industry by utilizing the knowledge of experts in manufacturing, mining, agriculture, forestry, fisheries, transportation, medical care, finance, commerce, and any other industry.
[0236] [Fifth embodiment] In this embodiment, an embodiment will be described in which a large-scale language model is used as a keyword translation means for translating keywords into multiple languages in a document search system. An example of the document search system will be described with reference to Japanese Patent Publication No. 2012-074087 (Patent No. 4945015).
[0237] In this embodiment, when a keyword is input in one language for searching a document database containing documents written in various languages such as Japanese, English, French, and Chinese, the input keyword is converted into another language by a translation engine, and a search is performed using both the input keyword and the keyword converted into the other language. Keywords are prioritized by assigning a score to each keyword, and this priority is reflected in the priority of the document search results that are output. In this way, a document search method that supports multiple languages is realized.
[0238] This embodiment will now be described with reference to the accompanying drawings. <Embodiment 1> 88 shows the configuration of a document search system 150 according to this embodiment. The document search system 150 is a system used to search for documents using keywords.
[0239] The document retrieval system 150 includes a document retrieval device 9 that searches for documents using keywords.
[0240] The document search device 9 is an information processing device, and has a configuration as a well-known computer. The document search device 9 includes an input device 29 that is used by a user to input keywords. This input device 29 is, for example, a mouse or a keyboard. The document search device 9 also includes a display device 40 that displays the results of the search process to the user. The display device 40 is, for example, a display or a printer. The document search device 9 also includes a calculation device 50 that performs calculations. The calculation device 50 is, for example, a CPU (Central Processing Unit).
[0241] Although not shown, the document search device 9 includes a memory and a HDD (hard disk drive) as storage means for storing information. The document search device 9 also includes a network interface for transmitting and receiving information to and from other information processing devices.
[0242] The storage means of the document search device 9 stores a document search program that prescribes the operations of the document search device 9 and the arithmetic device 50. By executing this document search program, the arithmetic device 50 functions as keyword receiving means 51, keyword translation means 52 (an application example of a large-scale language model), keyword score determination means 53, document search means 54, document score calculation means 55, and search result output means 56 shown in Fig. 88, and each function will be described in detail later.
[0243] Furthermore, the document search device 9, which is a computer, realizes other functions described in this specification by the calculation device 50 executing the document search program or another program.
[0244] The document search system 150 includes a translation service device 110 communicably connected to the document search device 9. The translation service device 110 translates keywords. The translation service device 110 receives a phrase expressed in one language, translates it into another language, and outputs it. In other words, the translation service device 110 has a function of generating a keyword (translated keyword) in which an input keyword (input keyword) is translated into another language based on the input keyword. Note that "translation" here can also be considered as conversion from a keyword in one language to a keyword in another language.
[0245] The translation service device 110 translates into a plurality of languages. For example, for an input keyword in Japanese, it generates and outputs a translated keyword in English and a translated keyword in French.
[0246] Furthermore, the translation service device 110 generates a plurality of translation keywords with rankings for one input keyword. That is, for a certain word, the translation keywords are ranked in order from the most frequently used translation depending on the frequency with which the corresponding translations are used, and a list of translation keywords is generated. This list represents the ranking of each translation keyword by, for example, arranging the translation keywords according to their ranking, but the ranking of each translation keyword may also be represented by associating the translation keyword with a numerical value or the like representing the ranking. A known configuration can be used for the translation service device 110. For example, the translation service device 110 stores a dictionary file that associates one or more translations with each of a large number of words and phrases, and performs translation by referring to this dictionary file.
[0247] The document search system 150 includes a document database 120 communicably connected to the document search device 9. The document database 120 stores a plurality of documents to be subjected to search processing by the document search device 9.
[0248] Document database 120 receives an input of one or more keywords, extracts all stored documents that contain any of the keywords, and outputs the extracted documents or a list thereof.
[0249] The operation of the document retrieval system 150 configured as above will be described with reference to the flow chart in FIG. 89 and data examples in FIGS. 90 to 94.
[0250] 89 is a flowchart for explaining the operation of the document search device 9 in the document search system 150. First, the keyword receiving means 51 receives one or more input keywords to be used in search from the user via the input device 29 (step S1, keyword receiving step). In this example, it is assumed that two input keywords in Japanese, "sensei" (teacher) and "kyoushi" (teacher), have been received.
[0251] Next, the keyword translation means 52 uses the translation service device 110 to translate the input keywords into translated keywords (step S2, keyword translation step). In this step S2, the keyword translation means 52 passes the input keywords to the translation service device 110, and the translation service device 110 generates translated keywords for each of the received input keywords and returns them to the keyword translation means 52. In this way, the keyword translation means 52 obtains the translated keywords.
[0252] FIG. 90 shows an example of the correspondence between input keywords and translated keywords. In this example, there are two types of translated keywords: those in English as shown in FIG. 90(A) and those in French as shown in FIG. 90(B). In the table of FIG. 90(A), the input keyword "teacher" is associated with three translated keywords in English: "teacher" ranked 1, "instructor" ranked 2, and "master" ranked 3. In this way, the translation service device 110 stores a plurality of ranked translated keywords in association with each input keyword.
[0253] 90(B), the input keyword "teacher" is also associated with two French translation keywords, "professeur" ranked 1 and "instructeur" ranked 2. In this way, the keyword translation means 52 acquires translation keywords for a plurality of languages other than the language of the input keyword.
[0254] The document searching device 9 may store the input keywords, the obtained translated keywords, and the correspondence relationships shown in FIG. 90 in a format such as a table in the storage means.
[0255] Next, the keyword score determination means 53 determines a keyword score for each of the input keyword and the translated keyword (step S3, keyword score determination step). Here, the keyword score determination means 53 determines the keyword score based on the correspondence relationships shown in FIGS. 91 and 92.
[0256] Figure 91 shows an example of the correspondence between the ranking of translated keywords and the translation score based on the ranking. The keyword score determination means 53 determines the keyword score of each translated keyword based on this translation score. The document retrieval device 9 stores the correspondence shown in Figure 91 in advance in the form of a table or the like in its storage means, and this correspondence can be changed as appropriate by a user or administrator of the document retrieval device 9. The input keyword is always given a fixed score, for example 100 (note that this score is treated slightly differently from the translation score, as will be described later, and is therefore shown in parentheses in FIG. 91). Also, a different translation score is given to the translation keyword depending on its ranking. The score given decreases by a specified value, for example 10, for each rank decrease, so that 90 is given to ranking 1, 80 to ranking 2, 70 to ranking 3, and so on.
[0257] The larger the score, the more importance is attached to the document containing the keyword in the search results (i.e., the higher the ranking of the document in the search results). In this way, the keyword score determination means 53 determines the keyword score of the translation keyword based on the ranking of the translation keyword.
[0258] The relationship between the rank and the translation score is not limited to that shown in Figure 91. The translation score for rank 1 may be any value lower than the keyword score for the input keyword. The translation scores for ranks 2 and below may be any value determined by a function that monotonically decreases as the rank decreases (i.e., in this example, as the number representing the rank increases).
[0259] The translation service device 110 generally ranks translation keywords according to the frequency with which they are used as translations. When information such as syntax and context is not taken into consideration, a translation that is actually used more frequently as a translation among a plurality of translations listed in a dictionary or the like for a certain word can be said to be a more appropriate translation. A document that includes a more appropriate translation is more likely to be what the user desires than a document that includes only unsuitable translations. In other words, a translation keyword with a higher ranking can be said to be a more likely keyword. The keyword score determination means 53 determines the translation score based on the ranking of each translation keyword, so that a translation keyword that is more likely can have a higher translation score, and more accurate search results can be obtained.
[0260] It is not necessary for the translation service device 110 to rank each keyword according to its frequency of use through strict statistical analysis. Since general dictionaries and the like usually determine the order of appearance by taking into account the frequency of use of translated terms to some extent, the use of a well-known general dictionary can have a certain degree of effect in improving the accuracy of search results.
[0261] FIG. 92 shows an example of the correspondence between the translation score based on the ranking for each keyword and the keyword score finally given to each keyword.
[0262] The keyword score determination means 53 always assigns a keyword score of 100 to the input keywords as described above. For the translation keywords, the translation score is first determined based on the ranking for all combinations of each input keyword and each translation keyword. In Fig. 92, a translation score based on the ranking is assigned to all combinations of two input keywords and five translation keywords (a total of 10 combinations).
[0263] For example, as shown in FIG. 90(A), the translated keyword "master" has a rank of 3 for the input keyword "teacher", so in FIG. 91 it is given a translation score of 70, which corresponds to the rank of 3. Also, this translated keyword "master" has a rank of 2 for the input keyword "teacher", so in FIG. 91 it is given a translation score of 80, which corresponds to the rank of 2. Note that if a translation keyword does not have a rank for any of the input keywords, that is, if the translation keyword is not a translation of the input keyword, the translation score for that combination is 0. However, the translation score in this case does not have to be 0, and it may be any value that is smaller than the translation scores of any of the other translation keywords that correspond to the input keyword.
[0264] Based on the translation scores thus determined, the keyword score determination means 53 further determines a final keyword score for each translated keyword. In the example of Fig. 92, the average of the translation scores given to the translated keyword is set as the keyword score for that translated keyword.
[0265] In this manner, the keyword score determination means 53 determines a keyword score for each translated keyword based on all of the associated translation scores.
[0266] The document retrieval device 9 may store the correspondence shown in FIG. 92 in the form of a table or the like in its storage means.
[0267] Here, the keyword score given to the input keyword is always 100, as described above. Also, since all translation scores are 90 or less (i.e., less than or equal to the translation score for rank 1), the average keyword score (keyword score of the translated keyword) is always less than 90. Therefore, the keyword score given to the input keyword in the native language is higher than any of the keyword scores given to the translated keywords in other languages.
[0268] Since there is no risk of mistranslation or inappropriate translation of input keywords in one's native language, documents that contain the input keywords are more likely to be what the user wants than documents that contain only translated keywords. In other words, the input keywords can be said to be more likely keywords. In this way, by giving more weight to the more likely input keywords by setting a higher score and relatively lowering the score of the translated keywords, more accurate search results can be obtained.
[0269] In addition, when a translation keyword corresponds to multiple input keywords, such as the translation keyword "master" in this example, a document that contains that translation keyword is more likely to be what the user wants than a document that contains only other translation keywords. In other words, such a translation keyword can be said to be a more likely keyword.
[0270] Here, the keyword score determination means 53 determines a keyword score based on all the translation scores related to a certain translation keyword, thereby increasing the keyword score of a translation keyword that simultaneously corresponds to a plurality of input keywords. For example, the translation keyword "master" in FIG. 92 corresponds to both the input keywords "teacher" and "teacher", and has a translation score that is not 0 for each of them. However, the translation keyword "instructor" corresponds to the input keyword "teacher" but not to "teacher", and the translation score for "teacher" is 0. As a result, the keyword score of the translation keyword "master" becomes higher. In this way, by giving weight to the score of a more likely translation keyword by setting it high and setting the score of a translation keyword relatively low, more accurate search results can be obtained.
[0271] Next, the document search means 54 uses the document search system 150 to search for documents based on the input keyword and the translated keyword, and obtains multiple search result documents (step S4, document search step). In this step S4, the document search means 54 passes the input keyword and the translated keyword to the document database 120, and the document database 120 extracts all documents that contain either the input keyword or the translated keyword from among the documents stored therein, and returns the extracted documents to the document search means 54 as search result documents.
[0272] Here, the document search means 54 performs a search using input keywords in the native language and translated keywords in other languages, so that results can be obtained in a single search even when searching a document database 120 that contains documents in multiple languages.
[0273] The search result document obtained in step S4 includes information that identifies the body data of the document (title, date and time, author, etc.), but does not necessarily have to include the body data. If the search result document does not include body data, the body data itself may be output from document database 120 in response to a separate request from the user.
[0274] Additionally, each of the search result documents is associated with information indicating the number of times each keyword appears in the text data.
[0275] FIG. 93 shows an example of such information. In this example, documents A to J are extracted as search result documents. For example, in document A, the translated keyword "teacher" appears 12 times, the translated keyword "instructor" appears 10 times, and the translated keyword "master" appears 6 times, which indicates that the total number of times all keywords appear in document A is 28 times. In this way, document database 120 measures the number of times each input keyword and each translated keyword appears for each search result document, associates this with each search result document, and returns it to document search means 54. Note that in FIG. 93, the search result documents are sorted in order of the number of times each keyword appears.
[0276] The document retrieval device 9 may store the correspondence shown in FIG. 93 in the form of a table or the like in its storage means.
[0277] In the example of FIG. 93, the number of times the keyword appears is used, but instead, the number of times the keyword appears may be calculated taking into account the recognition rate of character recognition.
[0278] In documents in which character strings are represented by character codes (text data, data for word processing programs, etc.), the number of times a keyword appears can be counted accurately using a character code collation process. In contrast, in documents in which character strings are represented by image data, character recognition processing must be performed to convert the image into character codes, but the accuracy of this character recognition processing is not always high. In this regard, during character recognition processing, the degree to which character recognition was performed for the document according to a predetermined standard may be evaluated as a recognition rate, and this recognition rate may be taken into consideration. For example, the value representing the number of times a keyword appears may be reduced according to the recognition rate. Specifically, for documents with a recognition rate of 100%, the number of times a keyword appears can be used as is, and for documents with a recognition rate of 50%, the number of times a keyword appears can be halved and used.
[0279] Here, the method for calculating the recognition rate may be any method known in the art for character recognition processing.
[0280] Next, the document score calculation means 55 calculates a document score for each of the search result documents based on the keyword score determined by the keyword score determination means 53 (see Figure 92) and the number of times each of the input keywords and translated keywords appears (see Figure 93) (step S5, document score calculation step).
[0281] In step S5, the document score is calculated by multiplying the keyword score of each keyword by the number of times the keyword appears in the search result document, and then summing the results for all keywords. This document score can be said to represent the possibility (likelihood) that the search result document is what the user wants.
[0282] Figure 94 shows an example of the calculation results using this calculation method. In document A, the translated keyword "teacher", which has a keyword score of 90, appears 12 times, so the multiplication result is 90 x 12 = 1080. Similarly, the multiplication result for the translated keyword "instructor" is 400, and the multiplication result for the translated keyword "master" is 450. Furthermore, no other input keywords or translated keywords appear in document A, so the multiplication result is 0. The document score for document A is the sum of all these values, or 1930.
[0283] The document retrieval device 9 may store the correspondence shown in FIG. 94 in the form of a table or the like in its storage means.
[0284] For documents in which character strings are represented by image data, the document score calculation means 55 may calculate the document score based on the recognition rate in character recognition processing of the search result document, in addition to the keyword score and the number of occurrences.
[0285] Here, since the keyword score is a different value for each keyword, the document score of a document in which a keyword appears many times does not necessarily have a high document score. For example, among the search result documents, the document in which the keyword appears the most (28 times, see Figure 93) is document A, but the document with the highest document score is document C (2500, see Figure 94), and the rankings are reversed. The reason for this is that all the keywords that appear in document C are input keywords, so the keyword score for each keyword is relatively high, whereas all the keywords that appear in document A are translation keywords, so the keyword score for each keyword is relatively low. Also, since the keyword scores differ between each translation keyword, weight is given to the more likely translation keyword.
[0286] In this way, the document score calculation means 55 takes into account the quality of each keyword when calculating the document score of each search result document, thereby enabling a more accurate evaluation compared to methods that simply calculate the document score based on the number of times a keyword appears.
[0287] Next, the search result output means 56 outputs the search result documents (i.e. documents A to J) in association with the document scores calculated for each of them by the document score calculation means 55 (step S6, search result output step). This output is provided to the user via the display device 40, allowing the user to know the search results. At this time, the search result output means 56 ranks the search result documents in descending order of document score, and outputs according to this ranking.
[0288] As described above, the document search device 9 according to embodiment 1, the document search method executed by the document search device 9, and the document search system 150 determine keyword scores for each of the input keywords and translated keywords, and calculate document scores based on these keyword scores, so that it is possible to appropriately determine the priority order of documents output as search results.
[0289] In the above-mentioned first embodiment, the language expressing the input keywords is Japanese, and the languages expressing the translation keywords are English and French, but these may be other languages, including Chinese, for example. The language expressing the input keywords may be set according to the language used by the user, and the other language expressing the translation keywords may be set according to the language of the documents included in the document database 120.
[0290] The language expressing the translation keyword may be a single language (for example, only English). Furthermore, the translation service device 110 may output a single translation keyword for the input keyword, or may output multiple translation keywords without ranking them. Even with this configuration, as long as the keyword scores differ between the input keyword and the translation keywords, it is possible to obtain more accurate results than with conventional searches.
[0291] In the example of the first embodiment, an OR search (logical sum search) is performed, and all documents in which any one of the multiple input keywords and multiple translation keywords appears are obtained as search result documents. Alternatively, an AND search (logical product search) can also be performed.
[0292] In this case, in step S4 in Fig. 89, the document search means 54 passes the input keyword and the translated keyword to the document database 120 and instructs it to perform an AND search. The document database 120 extracts all documents that satisfy the following conditions i and ii from among the documents stored therein, and returns the extracted documents to the document search means 54 as search result documents.
[0293] - Condition i: For the input keyword "teacher", at least one of the input keyword itself and its corresponding translation keywords "teacher", "instructor", "master", "professeur", and "instructeur" appears. - Condition ii: For the input keyword "teacher", at least one of the input keyword itself and its corresponding translated keywords "teacher", "master", and "professeur" appears.
[0294] In other words, the document search means 54 and document database 120 create a keyword group for each input keyword by linking the input keyword and its corresponding translated keyword with an OR condition, and then link all of these keyword groups with an AND condition to create the final search conditions.
[0295] As a result of a search based on these conditions, for example, among the documents shown in Fig. 93 as search result documents in the first embodiment, document H does not satisfy condition ii and is not extracted because it does not contain the input keyword "teacher" or the corresponding translated keywords "teacher," "master," or "professeur." Similarly, document J does not satisfy condition ii and is not extracted.
[0296] In this example, the translated keywords "teacher," "master," and "professeur" correspond to both of the two input keywords "sensei" and "kyoushi," so documents in which any of these translated keywords appear will be extracted. For example, document E contains the translated keyword "teacher," and this translated keyword satisfies both conditions i and ii, so document E is extracted.
[0297] Even in the case of such an AND search, the processes from step S5 onwards can be performed in the same way as in the OR search. That is, the document scores are calculated and the search results are output in the same way as in embodiment 1. However, in this example, since documents H and J are not extracted in step S4, the processes from step S5 onwards are not executed for documents H and J.
[0298] In addition, in the first embodiment, a search using translated keywords is always performed by the document search means 54, but this may be switchable, and for example, the user may be able to specify that a search be performed using only the input keywords without using translated keywords. In this way, it is possible to perform processing equivalent to conventional document searches using only the input keywords, as necessary.
[0299] The document database 120 may store each document to be searched in association with language information indicating the language in which the document is written, and the translation service device 110 may store each translation keyword in association with language information indicating the language in which the translation keyword is written. In this case, the input keyword can always be treated as being in a specific language that corresponds to the mother tongue.
[0300] For example, a certain keyword in Japanese may be written the same way (a character string represented by the same character code) when translated into Chinese. For such keywords, the keyword score of the input keyword can be applied to documents in Japanese, and the keyword score of the translated keyword can be applied to documents in Chinese. In other words, for input and translated keywords that are written the same way in different languages, the keyword score of the keyword whose language information matches that of the search result document may be used when calculating the document score of the search result document.
[0301] In this way, even if keywords that are written the same in multiple languages are included, the likelihood of each keyword can be appropriately evaluated.
[0302] In the first embodiment, the number of times a keyword appears in a search result document is measured by the document database 120, but this may be measured by another component. For example, the text data of the search result document may be passed from the document database 120 to the document searching device 9, and the document searching means 54 or the document score calculation means 55 of the document searching device 9 may measure the number of times a keyword appears in a search result document.
[0303] The translation service device 110 and the document database 120 may be any device capable of transmitting and receiving appropriate information with the document search device 9 regarding keyword translation and document search, and may be configured, for example, as a computer, and may realize the functions of the translation service device 110 and the document database 120 by executing a program stored in each storage means. In this case, the program of the document search device 9, the program of the translation service device 110, and the program of the document database 120 function as document search programs, causing these computers to function as the document search system 150.
[0304] In the hardware configuration of the first embodiment, the document searching device 9, which is a single computer, includes keyword receiving means 51, keyword translation means 52, keyword score determination means 53, document searching means 54, document score calculation means 55, and search result output means 56, and in addition to this, a translation service device 110 and a document database 120 are provided, each as a single computer. However, the hardware configuration may be different from this. For example, the computer constituting the document searching device 9 may have the function of the translation service device 110 and the function of the document database 120. [Explanation of symbols]
[0305] 9 Document Search Device 10 Means of acquiring cases 20 Rule Generation Method 30 Consulting methods 50 Arithmetic unit 100 Business Reform Support System 110 Translation service equipment 120 Document Database 151 I / O Devices 152 Business Process Template Database 153 Design Information Database 154 Classification Results Database 155 Business support means 156 Business Process Template Editing Method 157 Keyword Extraction Method 158 Document creation source business process identification means 159 Classification means 160 Classification Category Determination Method 161 Classification result display means 162 Task List 163 Task List Editing Method 164 Classification Category Priority Determination Method 170 Business support processing device 400 Data Warehouse (DWH) / Integrated Database (DB) 500 Existing System 600 Channels 700 External Information 1001 Database Server 1003 Authentication Server 1005 User terminal device 1007 Administrator terminal device
Claims
1. 1. A knowledge management system comprising: A first storage means for storing project attributes including attributes related to the technical field of the project and attributes related to the industry of the project's client, and project data including identification information of employees who participated in the project; a second storage means for storing employee data including identification information of the employee; A receiving means for receiving input of search conditions for project data; a project search means for searching the first storage means using a large-scale language model based on the search conditions received by the reception means; a first output means for outputting the project data searched by the project search means; an employee search means for searching the second storage means using a large-scale language model based on employee identification information included in the project data searched by the project search means; a second output means for outputting the employee data searched by the employee search means; A knowledge management system that includes:
2. an input / output device having a display for inputting and outputting data; a design information database for storing and retrieving documents; A keyword extraction means for acquiring keywords from a document being displayed or executed on a display by using a large-scale language model; a problem list in which problems are associated with related words, and a classification means for classifying documents stored in the design information database based on the problems listed in the problem list; a classification result database for storing and retrieving the classification results by the classification means; A business process template database that stores and retrieves business process template data in which each business procedure, its deliverables, and reference materials are associated with the titles and storage addresses of the associated procedures; A document classification device for an information retrieval system having a task support means for reading task process template data stored in the task process template database and providing task support including task guidance, presentation of reference materials, and task product management, which classifies stored documents into predetermined categories and presents them, a classification category determination means for comparing the keywords extracted by the keyword extraction means with the tasks and related words in the task list, and determining a matching task as a classification category to be displayed to a user; a document creation source business process determination means for acquiring a title of a document being displayed or executed on the display, comparing it with a title of a deliverable of the business process template data, and searching for an original business procedure from which the deliverable was created; A classification category priority determination means for determining the degree of relevance of the problem to the business procedure under consideration and determining the priority of the classification category; A document classification device for an information retrieval system comprising:
3. A business reform support system that utilizes existing systems within a company, characterized in that it connects data with host computers, including personal computers, and a data warehouse / integrated database (DB) currently in operation within the company, and also connects data with existing systems, channels, and external information containing various types of data as input information sources, and a large-scale language model centrally consolidates the various information scattered across the existing systems, channels, external information, host computers, and data warehouse / integrated DB based on extraction conditions to extract insight information, and improves the company's business by visualizing the extracted insight information.
4. An expert system generation device for constructing an expert system using a large-scale language model, comprising: A case acquisition means for accumulating situations and associated results observed from an external system for which the expert system is to be constructed; A rule generating means for automatically learning the accumulated situations and results and deriving statistical tendencies for causing the results from the situations as knowledge; and a consulting means for automatically determining a result for an observed situation from the derived statistical tendency; An expert system generating device characterized in that the situation and the results are updated by operating the case acquisition means, the rule generation means, and the consulting means independently or in cooperation with each other, thereby automatically updating them to more appropriate knowledge.
5. a document search device for searching documents using keywords; A translation service device; a document database that stores a plurality of documents to be searched, The document search device includes: A keyword receiving means for receiving one or more keywords as input keywords; a keyword translation unit that acquires translated keywords for a plurality of other languages, each of the input keywords being translated into other languages by using a large-scale language model; a keyword score determination means for determining a keyword score for each of the input keywords and the translated keywords; a document retrieval means for retrieving documents based on the input keyword and the translation keyword, and acquiring a plurality of search result documents; a document score calculation means for calculating a document score for each of the search result documents based on the keyword scores; a search result output means for outputting each of the search result documents in association with the corresponding document score, the translation service device generates a plurality of translation keywords having a ranking corresponding to each of the input keywords; the translation service device indicates a ranking of the translation keywords based on an order in which the translation keywords are listed; the keyword score determination means determines a translation score for each combination of the input keywords and each translation keyword based on the ranking; The keyword score determination means determines, for each of the translation keywords, the keyword score based on all of the associated translation scores; the keyword score of the input keyword is higher than any of the keyword scores of the translation keywords corresponding to the input keyword; Document retrieval system.
Citation Information
Patent Citations
Dialogue model training method and device therefor
JP2023033202A
Cited By
User question response method and system, computer equipment and readable storage medium
CN119202142A
Sales management system and sales management program
JP7685177B1
