Human Capital Management Scoring System and Program

Abstract

We provide a human capital management scoring system and program that can recommend effective measures to users who want to improve their evaluation of human capital management. [Solution] In the health management scoring system 10, which is a human capital management scoring system, the perspective evaluation request means 33 obtains perspective evaluation result data from a large language model (LLM) that indicates whether or not each of multiple perspectives is met with a binary value. Furthermore, the score calculation means 34 calculates the health management navigation score, which is a human capital management navigation score, using the perspective evaluation result data with TF-IDF, and the output means 45 displays information related to human capital management (health management) on the screen using the human capital management navigation score (health management navigation score).

Description

Technical Field

[0001] The present invention relates to a human capital management scoring system and program configured by a computer that presents information related to health management or other human capital management. For example, for the purpose of enabling a business owner company of a health insurance association to obtain a certification as an excellent company in health management or a health management brand, it can be used when analyzing data describing the efforts related to the health management of a company and presenting measures considered effective for the health management issues faced by each company.

Background Art

[0002] The health management survey form published by the Ministry of Economy, Trade and Industry not only quantifies the evaluation of the health management status of each company, but also describes the content of issues, the results of measure implementation, the results of effect verification, etc. in text (see FIG. 2 described later). In addition, the comprehensive evaluation of the health management level is calculated by weighting four items, namely "Aspect 1 Management Philosophy and Policy", "Aspect 2 Organizational System", "Aspect 3 System and Measure Implementation", and "Aspect 4 Evaluation and Improvement", at a ratio of 3:2:2:3. Since it is considered difficult to significantly change the management philosophy and the like, it is necessary to improve the evaluation value of "Aspect 4 Evaluation and Improvement" in order to increase the comprehensive evaluation. In addition, "Aspect 4 Evaluation and Improvement" is mainly evaluated from the text information on the content of issues, the results of measure implementation, and the results of effect verification of each company. Therefore, it is important for each company to implement health management measures that lead to an increase in the comprehensive evaluation and "Aspect 4 Evaluation and Improvement" in order to obtain a certification as an excellent company in health management or a health management brand.

[0003] Furthermore, a dashboard system developed by the applicant of this application is known as a system that analyzes information related to health management and presents the results to users (see Patent Document 1). In this dashboard system, topic estimation processing is performed for each theme using multiple response data obtained from questionnaire data such as health management questionnaire data, the topic distribution of each response data is determined, a topic vector is obtained using the response vector of the response data with the highest topic value, the similarity between the corporate document vector of multiple corporate document data created from corporate document data such as integrated report data and the topic vector of each sub-theme is calculated as a correlation score, and the corporate document data is displayed on the screen using the correlation score.

[0004] While the dashboard system described in Patent Document 1 performs topic estimation processing, the present invention differs in that, as will be described in detail later, it does not perform topic estimation but instead utilizes large language models (LLMs) such as Chat Generative Pre-trained Transformer (ChatGPT) to obtain binary answers indicating whether or not each of the set multiple tasks is satisfied, and further uses these binary answers to calculate Term Frequency-Inverse Document Frequency (TF-IDF) to obtain a health management navigation score or other human capital management navigation score (the human capital management navigation score is a higher-level concept term than the health management navigation score).

[0005] Furthermore, an information processing device is known that enables the visualization of various performance indicators of personnel (see Patent Document 2). This information processing device includes a first acquisition unit that acquires information indicating the user's motivation from the user's responses to a survey input by a terminal device, a second acquisition unit that acquires information indicating the user's health condition acquired by the terminal device in the survey, a calculation unit that calculates an indicator associated with the organization based on the information indicating the user's motivation and the information indicating the user's health condition, and an output unit that outputs the indicator.

[0006] The information processing device described in Patent Document 2 differs from the present invention in that it does not perform processing using a large-scale language model (LLM) or processing to calculate TF-IDF. Furthermore, this information processing device does not perform clustering processing. [Prior art documents] [Patent Documents]

[0007] [Patent Document 1] Japanese Patent Publication No. 7534511 (Abstract) [Patent Document 2] Japanese Patent Publication No. 2024-135377 (Abstract) [Overview of the project] [Problems that the invention aims to solve]

[0008] In order to obtain certification as a Health Management Excellent Corporation or to acquire the Health Management Brand, it is necessary to provide recommendations that meet the following three conditions to users (companies or their personnel who wish to improve the evaluation of their health management initiatives).

[0009] The first condition is to recommend effective measures and their results for addressing the health management challenges faced by the company.

[0010] The second condition is to recommend the results of various implemented measures. This is because users are more likely to choose and implement measures that are better suited to their company's situation if they can select one from a variety of recommended options.

[0011] The third condition is to recommend the results of policy implementations that receive high overall evaluations. This is because, in order to receive "Certification as a Health Management Excellent Corporation," one must rank highly in the overall evaluation.

[0012] One possible approach in this case is to simply recommend the results of measures implemented by other companies that address challenges similar to those of one's own company. However, this approach fails to satisfy the second and third conditions of the three conditions mentioned above.

[0013] Furthermore, the overall evaluation score for health management is calculated based on the content of the issues, the results of the implemented measures, and the results of the effectiveness evaluation that companies have answered in the health management survey. Therefore, in order to raise the overall evaluation score for health management, it is important to receive high evaluations for the content of the issues, the results of the implemented measures, and the results of the effectiveness evaluation. However, since these are written in text, quantitative evaluation is extremely difficult. In other words, the overall evaluation is calculated using the criteria adopted by the surveying body (in the case of the health management survey, the Ministry of Economy, Trade and Industry), but it is not always clear what kind of content will result in what kind of evaluation, or in other words, what kind of content will receive a high evaluation. Therefore, if it is possible to quantitatively grasp the evaluation of each company's initiatives as written in text, and clarify the relationship between these quantified values ​​and the overall evaluation score, it will be possible to recommend effective measures.

[0014] The above points also apply to surveys on human capital management in a broader sense, not limited to health management, where questionnaires answered by each company are collected, and evaluations are conducted by the surveying body based on the collected questionnaires. In this application, since human capital management includes the non-financial situation of companies, health management is a concept included in human capital management.

[0015] The objective of this invention is to provide a human capital management scoring system and program that can recommend measures that are deemed effective for users who wish to improve their evaluation of human capital management. [Means for solving the problem]

[0016] The present invention relates to a human capital management scoring system comprising a computer that presents information related to health management or other human capital management, A survey data storage means stores, in association with company response identification information for identifying a company, or, when multiple responses from a single company are permitted, company and response data, the following: issue content data consisting of text indicating the content of issues, measure implementation result data consisting of text indicating the results of the measures implemented, and effect verification result data consisting of text indicating the results of the effect verification, obtained as responses from each company to a survey on health management or other human capital management; issue content data consisting of text indicating the content of issues, measure implementation result data consisting of text indicating the results of the effect verification; and effect verification result data consisting of text indicating the results of the effect verification. A combining means that combines the issue content data, policy implementation result data, and effectiveness verification result data stored in this survey data storage means, or combines the issue content data and policy implementation result data to create combined text data. In order to determine whether the content expressed in the combined text data of each company created by this combining means satisfies each of several perspectives for evaluating each company's efforts in human capital management, a perspective evaluation request data is created for each of the combined text data of each company, and a request sentence asking for a binary answer indicating whether the content expressed in the combined text data satisfies each of the several perspectives is created and input into a chat-generative pre-trained transformer (ChatGPT) or other large-scale language model. The perspective evaluation request means stores the perspective evaluation result data, which indicates whether each of the several perspectives is satisfied in one of two ways, output from this large-scale language model, in association with company response identification information and in a perspective evaluation result storage means. A score calculation means that uses binary values ​​constituting the perspective evaluation result data for each company's combined text data stored in the perspective evaluation result storage means to treat each of the multiple perspectives as a word and each company's combined text data as a document to calculate the term frequency-inverse document frequency (TF-IDF), and stores the sum or average of the TF-IDF values ​​for each of the multiple perspectives obtained for each of the combined text data, or the value obtained by normalizing or other processing these sums or averages, as a health management navigation score or other human capital management navigation score, in association with company response identification information and stored in the score storage means. An output means that performs at least one of the following processes as a process for displaying information related to human capital management on the screen: a process for displaying the human capital management navigation score stored in the score storage means along with at least one of the following on the screen: the company name corresponding to the company response identification information, the issue content data, the measure implementation result data, the effectiveness verification result data, and the combined text data; a process for displaying at least one of the following on the screen according to the results of selecting, extracting, sorting, or grouping using the human capital management navigation score: the company name corresponding to the company response identification information, the issue content data, the measure implementation result data, the effectiveness verification result data, and the combined text data; and a process for displaying the distribution of human capital management navigation scores for each company or the change in human capital management navigation scores for each company over time. It is characterized by having the following features.

[0017] Here, "health management or other human capital management" indicates that "health management" is a term encompassing the concept of "human capital management." Similarly, "health management level survey data or other survey data related to human capital management" indicates that "health management level survey data" is a term encompassing the concept of "human capital management survey data." Furthermore, "health management Navi score or other human capital management Navi score" indicates that "health management Navi score" is a term encompassing the concept of "human capital management Navi score." Therefore, the "survey data" handled by the human capital management scoring system of the present invention includes cases where only data related to health management is used, where data related to health management and data related to financial status are mixed, or where only data related to financial status is used.

[0018] Furthermore, "health management survey data or other human capital management survey data" refers to survey data in which responses to predetermined themes are organized. The survey is not limited to national or local government agencies such as the Ministry of Economy, Trade and Industry, but can also be conducted by private research organizations. In short, it is sufficient if the data includes responses from each company in accordance with predetermined themes. Therefore, even if the name of the survey currently given by the Ministry of Economy, Trade and Industry as the "health management survey" is changed later, the survey data after the name change is acceptable as long as it includes responses to similar themes. Survey data that conforms to the guidelines of ISO 30414 (Guidelines for Disclosure of Information on Human Capital) is also acceptable.

[0019] Furthermore, the phrase "corporate response identification information for identifying a company, or for identifying a company and response data when multiple responses from a single company are permitted" indicates that, in cases where a single company can only record (input) one response data (one issue content data, one corresponding measure implementation result data, and one effectiveness verification result data) in the questionnaire data, the "corporate response identification information" in this application is information for identifying a company. However, since there is a one-to-one correspondence between companies and response data, the "corporate response identification information" in this case is also information for identifying the response data. On the other hand, if the survey data allows for multiple responses from a single company (for example, in the health management survey conducted by the Ministry of Economy, Trade and Industry, one company can provide (input) two responses), the "company response identification information" of this application is information that can identify a company and also identify the response data. In this case, the "company response identification information" may be identification information obtained by combining company identification information and response identification information managed separately, or it may be an integrated form of such identification information. Specifically, this is described in detail in <Company Response Identification Information: Figure 3> in [Modes for Carrying Out the Invention] described later.

[0020] In addition, the "normalization process or other processing" in the "score calculation means" means that the normalization process is included in the concept of processing, and the "other processing" is, for example, a standardization process, a deviation value calculation process, etc. Specifically, it is detailed in (Calculation of the value of the health management navigation score) in <Configuration of Processing Means 30 / Score Calculation Means 34: FIG. 6> in [Mode for Carrying Out the Invention] described later.

[0021] Furthermore, as the "processing of displaying information related to human capital management on the screen" by the "output means", a screen display process using the human capital management navigation score is listed, and it is described that at least one of these processes is to be performed. However, the meaning is that as long as the function of performing a screen display process using the human capital management navigation score is included in the output means, and it is not intended to exclude the function of performing a screen display process that does not use the human capital management navigation score (for example, a screen display process using the value of the comprehensive evaluation or the comprehensive deviation value). [[ID=)5]]

[0022] In such a human capital management scoring system of the present invention, the viewpoint evaluation request means obtains viewpoint evaluation result data indicating whether each of a plurality of viewpoints is satisfied or not in either of two values from a large language model (LLM). Furthermore, the score calculation means calculates the human capital management navigation score by TF-IDF using the viewpoint evaluation result data, and the output means can perform the process of displaying information related to human capital management on the screen using the human capital management navigation score.

[0023] Therefore, a Human Capital Management Navi Score that quantifies the evaluation of the problem content data, policy implementation result data, and effectiveness verification result data, which are free texts described freely by each company, is obtained, and information regarding human capital management can be displayed on a screen using this Human Capital Management Navi Score. At this time, it is confirmed that there is a correlation between the obtained Human Capital Management Navi Score and the value of the comprehensive evaluation calculated by the evaluation subject of the questionnaire data, and the usefulness of the Human Capital Management Navi Score is confirmed. In the case of the embodiment described later (when nine viewpoints are set for the Health Management Degree Questionnaire by the Ministry of Economy, Trade and Industry, and normalization processing by dividing by the maximum value is performed), the correlation coefficient is about 0.39, and the confirmation of the correlation can also be visually confirmed using the graph in FIG. 21 described later. However, the graph itself shown in FIG. 21 of the present application described later is an image shown for the purpose of functional explanation and is not drawn by strictly tracing the graph output by the actual machine using actual accurate numerical values.

[0024] Also, even if a value (a value indicating how many viewpoints are satisfied among a plurality of viewpoints, or a value proportional thereto), which is the total value or average value for each combined text data calculated using the viewpoint evaluation result data (binary values indicating whether or not each of a plurality of viewpoints is satisfied), is used as the evaluation result score, a correlation with the value of the comprehensive evaluation is recognized, but the evaluation result score is a numerical value that does not consider the difficulty level of achieving each of the plurality of viewpoints. In contrast, in the present invention, since the Human Capital Management Navi Score is calculated using the TF-IDF value, it is a numerical value that takes the difficulty level into consideration, and the usefulness of the evaluation result score is further enhanced.

[0025] And by using the Human Capital Management Navi Score obtained in this way, it becomes possible to recommend measures that are considered effective to users who want to improve the evaluation regarding human capital management, and the above object is achieved thereby.

[0026] <Configuration for extracting a plurality of viewpoints for evaluation using a large language model (LLM)>

[0027] Furthermore, in the aforementioned human capital management scoring system, The system includes a perspective extraction means that prepares numerous combinations of two combined text data selected from the combined text data of each company created by the combining means, and for each of these numerous combinations, creates perspective extraction request data that includes two combined text data and a request for a response asking which of the two combined text data expresses the best initiative and why that conclusion was reached, and inputs this into a large-scale language model, and stores the text data showing why that conclusion was reached, output from the large-scale language model, in a perspective extraction result storage means as perspective data that shows from what perspective the comparison was made and from what perspective the superiority or inferiority was determined. The method for requesting evaluation from a specific perspective is: It is desirable that the system be configured to use multiple perspectives set using perspective data stored in a perspective extraction result storage means, as multiple perspectives for evaluating each company's efforts regarding human capital management.

[0028] In this configuration, where multiple evaluation perspectives are extracted using a large-scale language model (LLM), the perspective extraction means performs so-called pairwise evaluation to extract perspectives from the LLM. Subsequently, the perspective evaluation request means uses the extracted perspectives to have the LLM evaluate the content of the activities expressed in the combined text data using the LLM's set perspectives. This makes it possible to evaluate the combined text data in a manner consistent with the LLM's thinking characteristics (decision processing algorithm).

[0029] <This configuration generates a set of recommendation candidates from the top k clusters of policy implementation results, and selects the policy implementation result data with the highest Human Capital Management Navi Score within each policy implementation result cluster as the policy to be recommended.>

[0030] Furthermore, in the aforementioned human capital management scoring system, A company response vectorization means that vectorizes the issue content data and the implementation result data of each company stored in the survey data storage means, and stores the obtained issue content vectors and implementation result vectors in a company response vector storage means in association with company response identification information. A policy implementation result clustering means that performs clustering using the policy implementation result vectors of each company stored in the company response vector storage means, assigns a policy implementation result cluster number to each company's policy implementation result vector to identify the classified policy implementation result cluster, and stores the assigned policy implementation result cluster number in the company response vector storage means in association with company response identification information. A recommendation candidate set generation means calculates the similarity between the issue content vector of a company specified by a user receiving a policy recommendation service and the issue content vectors of several other companies, sorts the company response identification information for several other companies in descending order of the calculated issue similarity, and selects company response identification information in order from the one with the highest issue similarity until the number of policy implementation result clusters for the policy implementation result cluster numbers corresponding to the sorted company response identification information is equal to the number of required policies specified by the user or a predetermined number of required policies, thereby generating a recommendation candidate set consisting of policy implementation result data belonging to the top k policy implementation result clusters. This system includes a recommendation target measure selection means that selects the measure with the highest human capital management navigation score among the top k measure implementation result clusters selected by this recommendation candidate set generation means as the measure to be recommended. The output means is, It is desirable that the system be configured to display the data showing the results of the measures selected by the recommended measures selection method as the recommended measures on the screen.

[0031] In this configuration, a set of recommendation candidates is generated from the top k clusters of policy implementation results, and the policy implementation result data with the highest human capital management navigation score within each policy implementation result cluster is selected as the policy to be recommended. This makes it possible to recommend a diverse range of policy implementation result data.

[0032] In other words, the policy implementation result clustering means creates policy implementation result clusters by performing clustering using policy implementation result vectors. Then, the recommendation candidate set generation means first sorts the company response identification information by the similarity of issues using issue content vectors (therefore, the policy implementation result data, policy implementation result vectors, and policy implementation result cluster numbers corresponding to the company response identification information are also sorted). Next, the top k policy implementation result cluster numbers (therefore, the top k policy implementation result clusters) are selected to generate a recommendation candidate set. Furthermore, the recommendation target policy selection means selects the policy implementation result data with the highest human capital management navigation score among each policy implementation result cluster as the policy to be recommended, and displays it on the screen using the output means, making it possible to recommend a variety of policy implementation result data.

[0033] Therefore, simply sorting company response identification information by the similarity of the issues and outputting the policy implementation result data corresponding to the top-ranked company response identification information may result in a large amount of similar policy implementation result data being output. However, by selecting the top k policy implementation result clusters and selecting and outputting the policy implementation result data with the highest human capital management navigation score within each policy implementation result cluster, it is possible to prevent a large amount of similar policy implementation result data (i.e., policy implementation result data belonging to the same policy implementation result cluster) from being output, and to output diverse policy implementation result data.

[0034] <Configuration for displaying press releases on the screen>

[0035] Furthermore, in a configuration where a set of recommendation candidates is generated from the top k clusters of policy implementation results mentioned above, and the policy implementation result data with the highest human capital management navigation score within each policy implementation result cluster is selected as the policy to be recommended, A press release acquisition means that uses keywords stored in a keyword storage means to acquire data from multiple press releases from a press release distribution system and stores it in a press release storage means, This press release acquisition means vectorizes the title data of each press release acquired by this press release acquisition means, and stores the resulting title vectors in a press release storage means in association with the acquired press releases. The system includes a press release selection means that calculates the similarity between each title vector stored in the press release storage means and the policy implementation result vector stored in the company response vector storage means for the policy implementation result data selected by the recommendation target policy selection means, and selects a predetermined number of press releases or a number specified by the user that have a high similarity between the calculated title and policy. The output means is, It is desirable that the system be configured to display data on the implementation results of the measures that were selected as recommended measures, and to also display data on a predetermined number of press releases selected by the press release selection means, or a number of press releases specified by the user.

[0036] By displaying press releases on the screen in this manner, it becomes possible to present users with more information about human capital management.

[0037] <Configuration equipped with means for extracting policy names and means for creating policy tags>

[0038] Furthermore, in a configuration where a set of recommendation candidates is generated from the top k clusters of policy implementation results mentioned above, and the policy implementation result data with the highest human capital management navigation score within each policy implementation result cluster is selected as the policy to be recommended, A policy name extraction means extracts policies related to health management or other human capital management from the content expressed in the policy implementation result data of each company stored in the survey data storage means, by creating one policy extraction request data for each company's policy implementation result data, which includes the policy implementation result data and a request to extract policies related to health management or other human capital management from the text described in the policy implementation result data, and inputting this into a large language model, and storing the text data indicating one or more extracted policies output from the large language model as policy name data in the policy extraction result storage means in association with company response identification information, This system includes a policy name extraction means that vectorizes each of the policy name data extracted by this policy name extraction means, performs clustering using the obtained policy name vectors, and for each of the multiple policy name clusters that have been classified, creates policy tag naming request data that includes a set of policy name data belonging to the same policy name cluster and a request to name this set, and inputs this into a large-scale language model, and a policy tag creation means that assigns as a policy tag the text data indicating the name of the same policy name cluster output from the large-scale language model to the policy name data belonging to the same policy name cluster, and stores the assigned policy tags in a policy extraction result storage means in association with company response identification information. The output means is, It is desirable that the system be configured to display the implementation result data of the measures that were selected as recommended measures on the screen, and to execute a process that displays the measure tags stored in the measure extraction result storage means, which are associated with the company response identification information corresponding to the measure implementation result data displayed on the screen.

[0039] In this configuration, which includes means for extracting policy names and means for creating policy tags, the number of policies (policy name data) extracted from policy implementation result data by the Large-Scale Language Model (LLM) is too large. Therefore, clustering is performed to create policy name clusters, and then the Large-Scale Language Model (LLM) assigns names to each policy name cluster. These names are then used as policy tags and can be displayed on the screen along with the policy implementation result data of the recommended policies. As a result, it is possible to understand the content of the policy implementation result data of the recommended policies by looking at the policy tags, which are fewer in total number than the policy name data, that is, by looking at the policy tags, which are more cohesive units than the policy name data. Furthermore, by displaying the policy tags on the screen, it becomes possible to perform various processes such as searching using those policy tags.

[0040] <Configuration equipped with a means for searching related policies by tagging them>

[0041] Furthermore, in the case of a configuration that includes the above-mentioned means for extracting policy names and means for creating policy tags, The system includes a policy tag-related policy search means that, using a policy tag specified by the user, extracts company response identification information stored in the policy extraction result storage means associated with the specified policy tag and / or policy implementation result data stored in the survey data storage means associated with the said company response identification information, and from the extracted company response identification information and / or policy implementation result data, selects a predetermined number or a number specified by the user, in descending order of the human capital management navigation score stored in the score storage means associated with the extracted company response identification information. The output means is, It is desirable that the system be configured to display the company name and / or policy implementation result data corresponding to the company response identification information selected by the policy tag-related policy search means.

[0042] In this configuration, which includes a means for searching related policies using policy tags, it is possible to search for and display related policy implementation result data using policy tags.

[0043] <Invention of a program>

[0044] Furthermore, the program of the present invention is intended to enable a computer to function as the human capital management scoring system described above.

[0045] Furthermore, the above program or any part thereof can be recorded and stored or distributed on recording media such as magneto-optical disks (MO), compact discs (CD), digital versatile disks (DVD), flexible disks (FD), magnetic tape, read-only memory (ROM), electrically erasable and rewritable read-only memory (EEPROM), flash memory, random access memory (RAM), hard disk drives (HDD), solid state drives (SSD), and flash disks. It can also be transmitted using transmission media such as wired networks like local area networks (LAN), metropolitan area networks (MAN), wide area networks (WAN), the Internet, intranets, extranets, wireless communication networks, or combinations thereof, and can also be transmitted on carrier waves. Moreover, the above program may be part of another program, or may be recorded on a recording media together with a separate program. [Effects of the Invention]

[0046] As described above, according to the present invention, a perspective evaluation result data is obtained from a large-scale language model (LLM) that indicates whether or not each of multiple perspectives is satisfied with a binary value. Furthermore, a human capital management navigation score is calculated using TF-IDF with this perspective evaluation result data, and information related to human capital management can be displayed on the screen using this human capital management navigation score. This has the effect of recommending effective measures to users who want to improve their evaluation of human capital management. [Brief explanation of the drawing]

[0047] [Figure 1] An overall configuration diagram of a health management scoring system, one embodiment of the human capital management scoring system of the present invention. [Figure 2] An example diagram of the data from the health management assessment questionnaire of the above embodiment. [Figure 3] A diagram illustrating the relationships between various data derived from the response data of each company recorded in the health management survey form of the above embodiment. [Figure 4] An explanatory diagram illustrating the process for extracting perspectives for evaluating each company's efforts regarding health management in the above embodiment. [Figure 5] A diagram illustrating the evaluation process for each company's efforts regarding health management from multiple (nine) perspectives set out in the above embodiment. [Figure 6] A diagram illustrating the calculation process of the Health Management Navi Score (Human Capital Management Navi Score) using TF-IDF in the above embodiment. [Figure 7] A diagram illustrating the vectorization process and the process for creating the policy implementation result cluster in the above embodiment. [Figure 8] A diagram illustrating the process of extracting policy names and creating policy tags in the above embodiment. [Figure 9] A diagram illustrating the process for generating the recommendation candidate set according to the above embodiment. [Figure 10] This diagram illustrates the process of selecting a recommended measure from a set of recommendation candidates and selecting a press release to be displayed on the screen according to the above embodiment. [Figure 11]A flowchart illustrating the flow of preparation processes before service provision using the health management scoring system of the above embodiment. [Figure 12] A flowchart illustrating the processing flow when providing services to a user using the health management scoring system of the above embodiment. [Figure 13] A diagram illustrating the overall screen display processing by the output means of the above embodiment. [Figure 14] An illustrative diagram of the upper part of the corporate profile screen of the above embodiment. [Figure 15] An illustrative diagram of the lower part of the corporate profile screen in the above embodiment. [Figure 16] An illustrative diagram of the condition setting section and the upper part of the "Search by Company Name" tab on the policy recommendation screen of the above embodiment. [Figure 17] An illustrative diagram of the lower part of the "Search by Company Name" tab on the policy recommendation screen of the above embodiment. [Figure 18] An example diagram of the "Search from entered issues" tab on the policy recommendation screen of the above embodiment. [Figure 19] An example diagram of the "Policy Tag List" tab on the policy recommendation screen of the above embodiment. [Figure 20] An illustrative diagram of the similar company search screen of the above embodiment. [Figure 21] An illustrative diagram of the Health Management Navi Score screen of the above embodiment. [Figure 22] An illustrative diagram of the display screen for the policy tag-related policy search results of the above embodiment. [Figure 23] An illustrative diagram of the detailed analysis result display screen of the above embodiment. [Modes for carrying out the invention]

[0048] One embodiment of the present invention will be described below with reference to the drawings. Figure 1 shows the overall configuration of a health management scoring system 10, which is one embodiment of a human capital management scoring system. Figure 2 shows an example of data entry in a health management questionnaire, and Figure 3 shows the relationships between various data derived from the response data of each company entered in the health management questionnaire. Figure 4 is an explanatory diagram of the process of extracting perspectives for evaluating each company's efforts regarding health management, Figure 5 is an explanatory diagram of the process of evaluating each company's efforts regarding health management based on multiple (nine) set perspectives, Figure 6 is an explanatory diagram of the process of calculating the health management navigation score (human capital management navigation score) using TF-IDF, Figure 7 is an explanatory diagram of the vectorization process and the process of creating a cluster of policy implementation results, Figure 8 is an explanatory diagram of the process of extracting policy names and creating policy tags, Figure 9 is an explanatory diagram of the process of generating a set of recommendation candidates, and Figure 10 is an explanatory diagram of the process of selecting recommended policies from the set of recommendation candidates and selecting press releases to be displayed on the screen. Furthermore, Figure 11 shows a flowchart illustrating the preparation process before service provision by the health management scoring system, and Figure 12 shows a flowchart illustrating the processing flow when providing services to users by the health management scoring system. Also, Figure 13 is an overall explanatory diagram of the screen display processing by the output means 45, Figure 14 shows an example of the upper part of the company profile screen 200, Figure 15 shows an example of the lower part of the company profile screen 200, Figure 16 shows an example of the condition setting section 310 and the upper part of the "Search by company name" tab 301 of the policy recommendation screen 300, Figure 17 shows an example of the lower part of the "Search by company name" tab 301 of the policy recommendation screen 300, and Figure 18 shows Figure 19 shows an example of the "Search from entered issues" tab 302 on the policy recommendation screen 300, Figure 20 shows an example of the "Policy Tag List" tab 303 on the policy recommendation screen 300, Figure 21 shows an example of the Health Management Navi Score screen 500, Figure 22 shows an example of the policy tag-based related policy search results display screen 600, and Figure 23 shows an example of the detailed analysis results display screen 700.

[0049] <Overall structure of the Health Management Scoring System 10: Figure 1>

[0050] In Figure 1, the health management scoring system 10 includes a scoring server 20 that performs various processes to present health management information to users and stores various data necessary for performing these processes, and a service provision system 70 using a large-scale language model (LLM), a vectorization processing system 71, and a press release provision system 80 connected to the scoring server 20 via network 1. In addition, an administrator terminal 90 operated by the system administrator and a user terminal 91 operated by a person in charge at a user company are connected to the scoring server 20 via network 1.

[0051] Here, Network 1 is an external network primarily composed of the Internet, but it may also be a combination of the Internet and internal networks such as LANs or intranets, and it does not matter whether it is wired, wireless, or a hybrid of wired and wireless; in short, it is sufficient if it can transmit information at a reasonable speed between multiple locations (regardless of distance).

[0052] The scoring server 20 is composed of one or more computers and includes processing means 30 that perform various processes to present users with various information related to health management, and storage means 50 that store various data necessary for performing the various processes.

[0053] The processing means 30 includes a combining means 31, a perspective extraction means 32, a perspective evaluation request means 33, a score calculation means 34, a company response vectorization means 35, a policy implementation result clustering means 36, a recommendation candidate set generation means 37, a recommendation target policy selection means 38, a press release acquisition means 39, a title vectorization means 40, a press release selection means 41, a policy name extraction means 42, a policy tag creation means 43, a policy tag related policy search means 44, and an output means 45.

[0054] Here, each of the means 31 to 45 constituting the processing means 30 is realized by a central processing unit (CPU) located inside the scoring server 20, one or more programs that define the operating procedure of this CPU, and working memory such as main memory and cache memory. Details of each of these means 31 to 45 will be described later.

[0055] The memory means 50 includes a survey sheet data storage means 51, a company information storage means 52, a perspective extraction result storage means 53, a set perspective storage means 54, a perspective evaluation result storage means 55, a score storage means 56, a company response vector storage means 57, a keyword storage means 58, a press release storage means 59, a policy extraction result storage means 60, and a policy name / policy tag correspondence relationship storage means 61.

[0056] Here, the hardware of each storage means 51 to 61 constituting the storage means 50 can be, for example, non-volatile memory such as a hard disk drive (HDD) or solid-state drive (SSD). Details of the data storage methods of each of these storage means 51 to 61 (such as the database table structure or the data recording method in a file) will be described later.

[0057] The service provision system 70 using a Large-Scale Language Model (LLM) is a computer system that provides cloud API (Application Programming Interface) services. In this embodiment, the Chat Generative Pre-trained Transformer (ChatGPT, Azure OpenAI ChatCompletion API GPT-4) is used, but it is not limited to this. For example, OpenAI's GPT-3.5, Google's Palm and Palm2, Amazon Web Services (AWS)'s Titan, Meta Platforms' Llama, etc. may also be used.

[0058] The vectorization processing system 71 is an external service provision system via a cloud API connected to network 1, and is a computer system that performs vectorization processing of text data. In this embodiment, as an example, the Azure OpenAI Embeddings API service provision system (Azure OpenAI Embeddings API text-embedding-ada-002 version2) can be used, but is not limited to this, and the embedding method may be Doc2Vec, BERT, Transformer, etc. The vector has a fixed length, but the number of dimensions is arbitrary, and in this embodiment, for example, it is 1536 dimensions.

[0059] The press release distribution system 80 is an external service provision system via a cloud API connected through network 1. It is a computer system that, when a keyword is entered, outputs press release data (including title data) related to that keyword.

[0060] The administrator terminal 90 and user terminal 91 are computer-based and include display means such as an LCD display and input means such as a mouse and keyboard. These administrator terminal 90 and user terminal 91 may also be portable devices such as laptop computers, tablets, and smartphones.

[0061] <Structure of data from the health management survey: Figure 2>

[0062] In Figure 2, the health management survey data published by the surveying body (Ministry of Economy, Trade and Industry) contains the response data of each company to the health management survey. This response data includes the issue theme number (in the example in Figure 2, issue theme number = 4 is selected) indicating the issue theme selected by each company from among 10 issue themes prepared by the surveying body, issue content data in free text format entered by each company, data on the results of the measures implemented, and data on the results of the effectiveness verification. This response data from each company is stored in the survey data storage means 51 (see Figure 3) along with the overall evaluation value calculated by the surveying body for that response data, and data showing evaluation results such as various standard scores.

[0063] <Company response identification information: Figure 3>

[0064] As shown in the survey form data storage means 51 in Figure 3, the response data of each company and the data showing the evaluation results of that response data are stored in association with company response identification information. In this application, "company response identification information" is defined to include cases in the survey form data where a single company can only enter (input) one response data, as already described in [Means for Solving the Problem]. However, in the health management survey form data shown in Figure 2 of this embodiment, a single company can enter (input) multiple (two) response data. Therefore, the "company response identification information" in this embodiment is information that can identify a company and also identify the response data.

[0065] In this embodiment, as shown in Figure 3, for the sake of explanation, the "company response identification information" is treated as an integrated entity of company identification information and response identification information. That is, the first part of the company response identification information functions as company identification information, and the second part functions as identification information for response data within that company. For example, in "00052300" in Figure 3, the first part "000523" indicates the company, and the second part (last two digits) "00" indicates the first response data within the company "000523". The second response data within the company "000523" would be, for example, "00052301" by adding "01" to the end of "000523". Alternatively, it may be written as a sub-number, such as "000523-00" or "000523-01". Furthermore, the company name may be used as company identification information, for example, the company response identification information for the first and second response data of ABC Corporation may be "ABC Corporation (1)", "ABC Corporation (2)", etc.

[0066] Furthermore, a combination of company identification information and separately managed response identification information may be used as company response identification information. For example, company identification information such as "000523" and internal response data identification information such as "00" or "01" may be stored in different columns of the same record, and their combination may function as company response identification information. Alternatively, the response data identification information may not be internal response data identification information, but rather identification information that manages all companies' response data with sequential numbers. For example, company identification information such as "000523" or "ABC Corporation" may be combined with sequential response identification information for all companies such as "002341" or "002342".

[0067] <Configuration of processing means 30 / connecting means 31: Figure 3>

[0068] The combining means 31 performs the process of creating combined text data by combining the issue content data, policy implementation result data, and effect verification result data stored in the survey sheet data storage means 51 (see Figure 3). Although the issue content data and policy implementation result data may be combined while excluding the effect verification result data, it is preferable to combine the effect verification result data as well, as this may have an effect such as reducing the number of evaluation criteria (9 in this embodiment) described later.

[0069] In this process, the merging means 31 removes newline characters and then performs the merging operation. The resulting merged text data may be stored in main memory, but it may also be stored in non-volatile memory (merged text data storage means, not shown).

[0070] <Configuration of processing means 30 / viewpoint extraction means 32: Figure 4>

[0071] The perspective extraction means 32 prepares numerous combinations of two combined text data selected from the combined text data of each company created by the combining means 31, and for each of these numerous combinations, it creates perspective extraction request data (including two combined text data, and a request for a response asking which of the two combined text data expresses the best initiative and why that conclusion was reached), inputs the created perspective extraction request data into the large-scale language model (LLM), and stores the text data output from the large-scale language model (LLM) that shows why that conclusion was reached as perspective data that shows from what perspective the comparison was made and how superiority or inferiority was determined, in the perspective extraction result storage means 53.

[0072] Specifically, as shown in Figure 4, the perspective extraction means 32 selects two combined text data and creates perspective extraction request data. In the example in Figure 4, the state in which perspective extraction request data has been created using the combined text data created by the combining means 31 for company response identification information = 00026400 and the combined text data created by the combining means 31 for company response identification information = 00073400 is shown. The perspective extraction means 32 then transmits the created perspective extraction request data to the service provision system 70 using a large-scale language model (LLM) via the network 1.

[0073] Furthermore, when the perspective extraction means 32 selects two combined text data to be combined, it may select from a large number of combined text data already created by the combining means 31. However, it may also be done in a different order: the perspective extraction means 32 selects two company response identification information to be combined, and the combining means 31 creates two combined text data corresponding to those two company response identification information. In this order, it is not necessary to store a large number of combined text data, and it is sufficient to store only the pairs of company response identification information used for the combination.

[0074] Next, the perspective extraction means 32 receives LLM response data transmitted from the service provision system 70 using a large-scale language model (LLM) via the network 1, and stores the text data containing the reasoning behind that conclusion as perspective data in the perspective extraction result storage means 53, associating it with the pair of company response identification information used in the combination. In the example in Figure 4, the pair of company response identification information, 00026400 and 00073400, and the perspective data, "Because quantitative evaluation is being conducted," are shown to be associated and stored in the perspective extraction result storage means 53.

[0075] The perspective extraction means 32 then repeats the above process for a large number of pairs of company response identification information (combinations of two combined text data). In this embodiment, there are health management survey forms for more than 2000 companies, and each company can enter two response data. Therefore, there are approximately 4200 response data (company response identification information) in total, and there are 10 different theme topics. For each theme, 40 pairs were randomly selected and the perspective extraction process was performed. Thus, the total number of pairs used was 40 pairs × 10 themes, totaling 400 pairs. In other words, a perspective extraction request data was created for each of the 400 pairs, and one response was obtained from the Large-Scale Language Model (LLM) for each of the 400 pairs.

[0076] Furthermore, the viewpoint extraction means 32 uses a large number of viewpoint data (400 in this embodiment) stored in the viewpoint extraction result storage means 53 to set up a plurality of viewpoints (9 in this embodiment) to be used in the processing of the viewpoint evaluation request means 33, which will be described later. It then performs a process to store the set plurality of viewpoint data (9 in this embodiment) in the set viewpoint storage means 54 in association with viewpoint numbers (1 to 9 in this embodiment), or a process to support this setting work.

[0077] In this embodiment, there are 400 extracted (collected) viewpoint data (viewpoint data stored in the viewpoint extraction result storage means 53), but naturally there are duplicates, so a setting operation is performed to eliminate duplicates, resulting in 9 set viewpoint data. This number of 400 is a number that a person can concentrate on and review, in other words, the total number of pairs was set to 400 so that it would be such a number. Therefore, in this embodiment, the viewpoint extraction means 32 outputs (displays on the screen or prints) all (400) viewpoint data stored in the viewpoint extraction result storage means 53 to the administrator terminal 90. Then, the system administrator (system developer) visually checks the output results and determines that there are multiple (9) viewpoint data, and that even if the total number of pairs is increased, no further (10th or beyond) viewpoint data will appear. The system administrator then performs the input work to set multiple (9) viewpoint data, and the viewpoint extraction means 32 accepts the setting input and stores the set multiple (9) viewpoint data in the setting viewpoint storage means 54, associating them with viewpoint numbers (1 to 9 in this embodiment). At this time, if the system administrator (developer) determines that there are approximately multiple (9) viewpoint data among the large number (400) of viewpoint data extracted, the system administrator (developer) performs the work of combining viewpoint data with similar expressions into one expression before inputting the setting into the setting viewpoint storage means 54.

[0078] Furthermore, if the number of viewpoint data extracted and stored in the viewpoint extraction result storage means 53 is very large, the viewpoint extraction means 32 may vectorize each of those viewpoint data, perform clustering using the obtained viewpoint vectors to create multiple (for example, nine) viewpoint clusters, create one viewpoint expression aggregation request data (including multiple viewpoint data belonging to the same viewpoint cluster and a request to combine the expressions of these multiple viewpoint data into a single expression) for each of the created multiple (nine) viewpoint clusters, input it into the large-scale language model (LLM), and store the LLM response data output from the large-scale language model (LLM) in the set viewpoint storage means 54 as set viewpoint data (expression aggregated viewpoint data), associated with the viewpoint number (1 to 9 in this embodiment). This process follows a similar flow to the process using the policy tag creation means 43 described later (see Figure 8), which involves vectorizing policy name data to create policy name vectors, creating multiple policy name clusters by performing clustering using the policy name vectors, requesting a large-scale language model (LLM) to name the policy name clusters using the set of policy name data belonging to the policy name clusters, and using the LLM response data as policy tags.

[0079] Alternatively, the viewpoint extraction process by the viewpoint extraction means 32 may be omitted, and viewpoint data created by the system administrator (developer) may be stored in the configured viewpoint storage means 54 as configured viewpoint data, associated with viewpoint numbers (1 to 9 in this embodiment). In other words, if multiple viewpoint data are stored in the configured viewpoint storage means 54 before the processing of the viewpoint evaluation request means 33 described later, the processing of the viewpoint evaluation request means 33 described later can be performed, so the viewpoint data may be viewpoint data created by human thought. However, it is preferable to store viewpoint data set using viewpoint data obtained by the viewpoint extraction process by the viewpoint extraction means 32 (representation-aggregated viewpoint data) in the configured viewpoint storage means 54, as this allows for the setting of viewpoint data in line with the thinking algorithm of a large-scale language model (LLM), and improves the reliability of the processing of the viewpoint evaluation request means 33 described later.

[0080] <Configuration of processing means 30 / perspective evaluation request means 33: Figure 5>

[0081] The perspective evaluation request means 33 creates one perspective evaluation request data for each of the combined text data of each company created by the combining means 31, in order to determine whether the content expressed in the combined text data of each company satisfies each of the multiple perspectives for evaluating each company's efforts in human capital management. This data includes the combined text data and a request for a binary response indicating whether the efforts expressed in the combined text data satisfy each of the multiple perspectives. The created perspective evaluation request data is input into a large-scale language model (LLM), and the perspective evaluation result data output from the large-scale language model (LLM), which indicates whether each of the multiple perspectives is satisfied in either binary form, is stored in the perspective evaluation result storage means 55 in association with company response identification information.

[0082] Specifically, in this embodiment, as shown in Figure 5, the perspective evaluation request means 33 creates perspective evaluation request data (including combined text data and a request statement asking for a binary response indicating whether the initiatives expressed in the combined text data satisfy any one of the multiple (nine) perspectives) for each of the combined text data of each company created by the combining means 31, and for each of the multiple (nine) perspectives. The created perspective evaluation request data is then transmitted via the network 1 to the service provision system 70 using a large-scale language model (LLM). In the example in Figure 5, the state in which perspective evaluation request data has been created is shown using the combined text data for company response identification information = 00115200 (Company C) = "There are many employees on leave due to mental illness. Mental health training was conducted. The number of people suffering from depression has decreased." and the perspective data with perspective number = 2 = "Is a quantitative evaluation being conducted?".

[0083] Next, the perspective evaluation request means 33 receives LLM response data (perspective evaluation result data indicating whether or not the perspective is met with a binary value of 1 (met) or 0 (not met)) transmitted from the service provision system 70 using a large-scale language model (LLM) via the network 1, and processes the perspective evaluation result data to be stored in the perspective evaluation result storage means 55 in association with the company response identification information and the perspective number. In the example in Figure 5, the perspective evaluation result data of "1", indicating that the perspective of perspective number 2 is met, is shown to be stored in the perspective evaluation result storage means 55 in association with the company response identification information = 00115200 (Company C) and perspective number = 2.

[0084] The perspective evaluation request means 33 then repeats this process for all combined text data (all company response identification information) and all perspective numbers. Therefore, if there are, for example, 4200 combined text data and 9 perspectives (perspective numbers = 1 to 9), 4200 data × 9 perspectives = 37800 perspective evaluation request data will be created, and 37800 perspective evaluation result data (either 1 or 0) will be obtained.

[0085] In this case, as shown in Figure 5, the viewpoint evaluation request means 33 creates viewpoint evaluation request data using multiple (9) viewpoint data stored in the set viewpoint storage means 54. Therefore, if the set viewpoint storage means 54 stores multiple (9) viewpoint data (expression-aggregated viewpoint data) set using viewpoint data extracted by the viewpoint extraction means 32 described above, the viewpoint evaluation request data will be created using viewpoint data that reflects the processing by the viewpoint extraction means 32. Note that in the automatic creation process of viewpoint evaluation request data, the viewpoint evaluation request data may be created by changing the ending of the viewpoint data (for example, deleting "ka.", "ka ni ka.", "ka?", etc.), or the viewpoint data may be used as is to create the viewpoint evaluation request data.

[0086] Furthermore, in this embodiment, the perspective evaluation request means 33 creates one perspective evaluation request data for each of the combined text data of each company, and one for each of the multiple (9) perspectives. However, instead of creating perspective evaluation request data for each perspective individually, it is also possible to create perspective evaluation request data for several perspectives together. For example, it is possible to ask whether or not each of the multiple (9) perspectives is satisfied, or to divide the multiple (9) perspectives into several groups (for example, into three groups of perspective numbers 1-3, 4-6, and 7-9) and ask whether or not each of these perspectives is satisfied. However, it is preferable to create perspective evaluation request data for each perspective individually in order to prevent the perspective evaluation request data from becoming lengthy or to prevent misunderstandings of the request text by the Large-Scale Language Model (LLM). Therefore, although the claims of this application include the phrase "a request to answer with a binary value whether or not each of the multiple viewpoints is satisfied," this phrase does not only refer to cases where all of the multiple (nine) viewpoints are answered at once, but also includes cases where the multiple (nine) viewpoints are answered one by one, as in this embodiment, or cases where the multiple (nine) viewpoints are divided into several parts and answered separately.

[0087] Furthermore, in this embodiment, the "binary value" is defined as 1 (satisfying) or 0 (not satisfying), but it is not limited to this. For example, it could be 10 (satisfying) or 0 (not satisfying), or 100 (satisfying) or 0 (not satisfying), and since the value is adjusted (normalized, etc.) by the score calculation means 34 described later, it is equivalent.

[0088] Furthermore, when creating perspective evaluation request data using the perspective evaluation request means 33, if the combined text data used in the processing by the perspective extraction means 32 described above remains in main memory or is stored in non-volatile memory, the perspective evaluation request data may be created using that combined text data. Alternatively, each time perspective evaluation request data is created using the perspective evaluation request means 33, the combined text data may be created using the response data (issue content data, policy implementation result data, and effect verification result data) stored in the questionnaire data storage means 51 by the combining means 31.

[0089] <Configuration of processing means 30 / score calculation means 34: Figure 6>

[0090] The score calculation means 34 uses the binary values ​​(1 or 0 in this embodiment) that constitute the perspective evaluation result data for each company's combined text data stored in the perspective evaluation result storage means 55 to consider each of the multiple (9 in this embodiment) perspectives as a word and to consider each company's combined text data as a document to calculate TF-IDF (Term Frequency-Inverse Document Frequency). The obtained TF-IDF values ​​for each of the multiple (9) perspectives are summed (or averaged) for each of the combined text data to obtain a total value (or average). Furthermore, after performing a Yeo-Johnson transformation on these total values ​​(or average), these values ​​are normalized using the Yeo-Johnson transformed values ​​of the total values ​​(or averages) for all the combined text data. The normalized value is then stored in the score storage means 56 in association with the company response identification information as the Health Management Navi Score (Human Capital Management Navi Score).

[0091] TF-IDF is generally calculated by multiplying TF (Term Frequency), which indicates the frequency of word occurrences within a document, by IDF (Inverse Document Frequency), which is a value indicating the rarity of a word. Therefore, it represents a value indicating the importance of a word. As described above, since each of the multiple (nine in this embodiment) perspectives is considered a word, and the combined text data of each company is considered a document when calculating TF-IDF, the TF-IDF value in this application is a numerical value that takes into account the difficulty of achieving each of the multiple perspectives.

[0092] Specifically, the score calculation means 34 performs the TF-IDF calculation process as shown in Figure 6. However, the viewpoint evaluation result storage means 55 shown at the top of Figure 6 is the same as the viewpoint evaluation result storage means 55 shown at the bottom of Figure 5. In Figure 6, however, in order to explain the content of the TF-IDF calculation process by performing specific calculations using concrete numerical examples, the number of viewpoints used in the calculation and the number of combined text data (number of company response identification information) have been reduced. That is, in Figure 6, although there are actually 9 viewpoints in this embodiment, the number has been reduced to 3 viewpoints, and although there are actually about 4200 combined text data, the calculation is explained as if there are only 3 data.

[0093] (Calculation of TF value) In this embodiment, for TF, the value (a value of 1 or 0) of the viewpoint evaluation result data shown in the matrix stored in the viewpoint evaluation result storage means 55 is used as is. Therefore, since the value of TF is the value of the matrix stored in the viewpoint evaluation result storage means 55, in the case of this embodiment, the value of TF-IDF is obtained by multiplying the value (a value of 1 or 0) of the viewpoint evaluation result data shown in the matrix by the value of IDF, which will be described later.

[0094] Furthermore, the TF used when calculating TF-IDF (TF-IDF as described in the claims) in the present invention is not limited to the case where the value of the viewpoint evaluation result data (a value of 1 or 0) is used directly as the value of TF, as in this embodiment.

[0095] For example, a standard TF is the value obtained by dividing the number of occurrences of word t in document d by the total number of words in text d (the total number of words when the same word appears multiple times but is counted separately). In this case, the denominator is, in other words, the sum of the occurrences of each word in text d, and is usually expressed using Σ. If this standard TF is adopted, the value obtained is the number of occurrences of viewpoint t in the combined text data d (which is naturally 1 or 0) divided by the sum of the occurrences of each viewpoint in the combined text data d (which is naturally 1 or 0) for all viewpoints. In this case, the denominator is ultimately the number of viewpoints that the combined text data d satisfies. The calculation for the example in Figure 6 is as follows.

[0096] For the combined text data of company response identification information = 00127100 (Company H), the standard TF value is [0 / 1,1 / 1,0 / 1]=[0,1,0]. For the combined text data of company response identification information = 00127200 (Company J), the standard TF value is [1 / 3,1 / 3,1 / 3]=[0.33,0.33,0.33]. For the combined text data of company response identification information = 00127300 (Company K), the standard TF value is [1 / 2,1 / 2,0 / 2]=[0.5,0.5,0].

[0097] Furthermore, when calculating the standard TF value as described above, if the denominator becomes zero (i.e., if there is combined text data that does not satisfy all aspects), you can adjust it by adding "+1" to the denominator, for example.

[0098] Furthermore, it can also be used as a frequency factor (TF) on a logarithmic scale, and the TF is calculated using the following formula. TF = log(1 + number of occurrences of word t in document d)

[0099] Therefore, if we substitute, it becomes as follows: TF = log{1 + number of occurrences of perspective t in the combined text data d (1 or 0)}

[0100] Furthermore, the TF may be calculated using other methods, for example, the TF calculated using the following formula.

[0101] TF = 0.5 + 0.5 × (Number of occurrences of word t in document d / Number of occurrences of the most frequently occurring word in document d)

[0102] Therefore, the substitution is as follows. In this case as well, if the denominator may become zero (if there is combined text data that does not satisfy all aspects), you can adjust it by adding "+1" to the denominator, for example.

[0103] TF = 0.5 + 0.5 × {Number of occurrences of viewpoint t in the combined text data d (1 or 0) / Number of occurrences of the viewpoint with the highest frequency in the combined text data d (1 or 0)}

[0104] (Calculation of IDF value) In this embodiment, the IDF is calculated using the scikit-learn library, so "+1" is added outside the log term in the IDF calculation formula, and "+1" is also added to the numerator and denominator inside the log term, resulting in the following formula.

[0105] The IDF of word t = log{(1 + total number of documents) / (1 + number of documents containing word t)} + 1

[0106] Therefore, if we substitute, it becomes as follows: IDF of viewpoint t = log{(1 + total number of combined text data) / (1 + number of combined text data that satisfy viewpoint t)} + 1

[0107] In the example in Figure 6, the IDF values ​​for each viewpoint (viewpoint numbers = 1, 2, 3) are calculated as follows, as shown near the center of Figure 6.

[0108] IDF(viewpoint number=1)=log{(1+3) / (1+2)}+1=1.29 IDF(viewpoint number=2)=log{(1+3) / (1+3)}+1=1.00 IDF(viewpoint number=3)=log{(1+3) / (1+1)}+1=1.69

[0109] Regarding the IDF, in this embodiment, the IDF value was calculated using the default calculation formula of the scikit-learn library described above. However, the calculation formula is not limited to the one described above, and for example, the adjustment of the "+1" in the numerator or denominator inside the log term, or the adjustment of the "+1" outside the log term, may be omitted.

[0110] (Calculation of TF-IDF value) In this embodiment, the TF-IDF value is calculated by multiplying the value of each element in the matrix stored in the viewpoint evaluation result storage means 55 shown at the top of Figure 6 (which is also the value of TF) by the IDF value mentioned above. Specifically, by multiplying the value of each element in the column for viewpoint number 1 (the first column) stored in the viewpoint evaluation result storage means 55 by 1.29, multiplying the value of each element in the column for viewpoint number 2 (the second column) by 1.00, and multiplying the value of each element in the column for viewpoint number 3 (the third column) by 1.69, the TF-IDF value as shown in the matrix near the center of Figure 6 is obtained.

[0111] (Calculation of Health Management Navi score) Next, the score calculation means 34 obtains the TF-IDF value, and then sums (or averages) the TF-IDF values ​​for each of the multiple (nine) perspectives obtained for each combined text data to obtain a total value (or average). This total value (or average) becomes the Health Management Navi score. However, this is the value before processing such as normalization.

[0112] In the example shown in Figure 6, for the combined text data of company response identification information = 00127100 (Company H), the sum of the TF-IDF values ​​for each perspective number = 1, 2, and 3 is 0.00 + 1.00 + 0.00 = 1.00. This sum of 1.00 is then stored in the score storage means 56 as the Health Management Navi score (however, this is the value before processing such as normalization), associated with the company response identification information = 00127100.

[0113] Furthermore, for the combined text data of company response identification information = 00127200 (Company J), the sum of the TF-IDF values ​​for each perspective number = 1, 2, and 3 is 1.29 + 1.00 + 1.69 = 3.98. This sum of 3.98 is then stored in the score storage means 56 in association with company response identification information = 00127200 as the Health Management Navi score (however, this is the state before processing such as normalization).

[0114] Furthermore, for the combined text data of company response identification information = 00127300 (Company K), the sum of the TF-IDF values ​​for each perspective number = 1, 2, and 3 is 1.29 + 1.00 + 0.00 = 2.29. This sum of 2.29 is then stored in the score storage means 56 in association with company response identification information = 00127300 as the Health Management Navi score (however, this is the state before processing such as normalization).

[0115] Subsequently, the score calculation means 34 performs a Yeo-Johnson transformation (converting to a Gaussian distribution to make it easier to handle as a machine learning feature) on the sum (or mean) of the TF-IDF values ​​for each of the obtained combined text data. Furthermore, it performs a process to normalize the Yeo-Johnson transformed value of the sum (or mean) (hereinafter referred to as the pre-normalized health management navigation score) so that it falls within the range of 0 to 100. This normalization process involves extracting the maximum value from the pre-normalized health management navigation scores, dividing the pre-normalized health management navigation score for each combined text data by that maximum value, and multiplying the result by 100 to obtain the final normalized health management navigation score. The score calculation means 34 then stores the obtained final health management navigation score in the score storage means 56 in association with the company response identification information.

[0116] In this embodiment, a normalization process is performed by dividing by the maximum value (a normalization process in which the value before processing and the value after processing are proportional). However, this is not the only method. For example, the maximum and minimum values ​​may be extracted from the Health Management Navi score before normalization, and a normalization process may be performed in which the maximum value becomes 100 and the minimum value becomes 0 (a normalization process in which the value before processing and the value after processing are not proportional). In this case, the final normalized Health Management Navi score is obtained by subtracting the minimum value from the Health Management Navi score before normalization, dividing the result by the difference between the maximum value and the minimum value, and multiplying the resulting value by 100. Furthermore, the maximum value used for normalization may be the maximum value of the base year rather than the maximum value of each year, and a normalization process may be performed in which the maximum value of the base year is set to 100. In addition, outliers may be excluded before performing these normalization processes.

[0117] Alternatively, instead of normalization, a standardization process may be performed to set the mean of each year to 0 and the variance of each year to 1, or the standard score for each year may be calculated. Such processing of the Health Management Navi Score (the Health Management Navi Score representing the sum or average value of the TF-IDF values) such as normalization, standardization, and calculation of standard scores may be performed using any method that makes it easier for the user to understand the magnitude and meaning of the values. For example, processing to set the median of each year to 50, or processing to set the median of the base year to 50, may also be performed.

[0118] <Configuration of processing means 30 / corporate response vectorization means 35: Figure 7>

[0119] As shown in Figure 7, the corporate response vectorization means 35 vectorizes the issue content data, policy implementation result data, and effectiveness verification result data of each company stored in the questionnaire data storage means 51, and performs the process of storing the obtained issue content vector, policy implementation result vector, and effectiveness verification result vector in the corporate response vector storage means 57 in association with corporate response identification information.

[0120] Although not shown in Figure 7, the corporate response vectorization means 35 also vectorizes the combined text data created by the combining means 31, and stores the resulting combined text vector in the corporate response vector storage means 57 in association with the corporate response identification information.

[0121] Specifically, the corporate response vectorization means 35 transmits the issue content data, the policy implementation result data, the effectiveness verification result data, and the combined text data to the vectorization processing system 71 via the network 1. The vectorization processing system 71 then receives the issue content vector, the policy implementation result vector, the effectiveness verification result vector, and the combined text vector transmitted via the network 1, associates them with corporate response identification information, and stores them in the corporate response vector storage means 57. In this embodiment, all vectors are 1536-dimensional vectors.

[0122] <Configuration of processing means 30 / policy implementation result clustering means 36: Figure 7>

[0123] The policy implementation result clustering means 36 performs clustering using the policy implementation result vectors of each company stored in the company response vector storage means 57 (see Figure 7), assigns a policy implementation result cluster number to each company's policy implementation result vector to identify the classified policy implementation result cluster, and stores the assigned policy implementation result cluster number in the company response vector storage means 57 (see Figure 7) in association with company response identification information.

[0124] Specifically, in this embodiment, the policy implementation result clustering means 36 performs clustering using BERTopic as an example. BERTopic is a type of text clustering algorithm that converts text data into a high-dimensional vector, performs dimensionality reduction (dimensional compression), and then clusters the resulting low-dimensional vector using an algorithm such as Hdbscan. The process of converting text data into a high-dimensional vector (1536 dimensions in this embodiment) is performed by the aforementioned company response vectorization means 35, and therefore the vectorization processing system 71 (in this embodiment, Azure OpenAI Embeddings API text-embedding-ada-002 version 2 is used) is utilized. In this embodiment, the low-dimensional vector created by BERTopic is, as an example, a 5-dimensional vector.

[0125] In order to create clusters of policy implementation results, clustering is performed using policy implementation result vectors. However, classifying policy implementation result vectors also classifies the policy implementation result data corresponding to the policy implementation result vectors, and further classifies the company response identification information corresponding to the policy implementation result data. Therefore, in this description, these are used as equivalent expressions.

[0126] This clustering is performed across the 10 themes of the challenge. The number of clusters representing the results of the implemented measures is determined by the system administrator (developer), but in this embodiment, it is approximately 300, and therefore, the number of cluster numbers representing the results of the implemented measures is also approximately 300. Thus, if there are approximately 4200 vectors representing the results of the implemented measures and corresponding data on the results of the implemented measures and company response identification information, these approximately 4200 items will be classified into approximately 300 groups.

[0127] Policy implementation result vectors and their corresponding policy implementation result data and company response identification information belonging to the same policy implementation result cluster are assigned the same policy implementation result cluster number. In this description, as shown in Figure 7, clustering is performed using policy implementation result vectors, so it is first assumed that the policy implementation result vectors are the direct targets of classification, and the policy implementation result cluster number is stored in the company response vector storage means 57 in association with the company response identification information. However, since the purpose of clustering is to cluster policies, it is also possible to consider that the policy implementation result data is being classified, and store the policy implementation result cluster number together with the policy implementation result data in the survey form data storage means 51 (see Figure 3) in association with the company response identification information, and the storage format is equivalent. Furthermore, a policy implementation result cluster attribution information storage means (not shown) that stores the correspondence between company response identification information and policy implementation result cluster number (i.e., attribution information indicating which policy implementation result cluster the company response identification information belongs to) may be provided, and this is also equivalent. Therefore, the present invention (claim description) states that "the assigned policy implementation result cluster number is stored in the company response vector storage means in association with company response identification information," and this description includes equivalent storage forms as described above.

[0128] <Configuration of processing means 30 / recommendation candidate set generation means 37: Figure 9>

[0129] The recommendation candidate set generation means 37 processes differently depending on whether the user receiving the policy recommendation service inputs a company (usually, the user specifies their own company, as this is done by a user who wants to improve their company's health management initiatives, but any company can be specified) or a problem (usually, the user specifies a problem their own company faces, but any problem can be specified) as a query on the screen displayed by the output means 45.

[0130] (Processing when the received query concerns a company specified by the user) The recommendation candidate set generation means 37 receives, as a query, the company (company identification information) related to the designation of a user who will receive the policy recommendation service (see company name input unit 321 in Figure 16). Furthermore, if the company designated by the user that has been received has entered multiple (two) response data (including issue content data) into the health management survey form, and the output means 45 receives information related to the user's issue selection (company response identification information corresponding to the selected issue content data) (see issue selection unit 322 in Figure 16), it receives the company response identification information as a query from the output means 45.

[0131] The recommendation candidate set generation means 37 then calculates the similarity (issue similarity) between the issue content vector of the company specified by the user received from the output means 45 (issue content vector corresponding to the received company response identification information) and each of the issue content vectors of multiple other companies (issue content vectors corresponding to company response identification information other than the received company response identification information). The company response identification information for the multiple other companies is then sorted in descending order of the calculated issue similarity. The number of policy implementation result clusters for the policy implementation result cluster numbers corresponding to the sorted company response identification information is equal to the required number of policies k specified by the user (see the policy number input unit 311 in Figure 16) or a predetermined required number of policies k. The company response identification information is then selected in order from the one with the highest issue similarity, thereby generating a recommendation candidate set consisting of policy implementation result data belonging to the top k policy implementation result clusters.

[0132] Specifically, as shown in Figure 9, if the query received from the output means 45 is a recommendation for measures to address the issue content data of a company specified by the user (usually the user's own company, which we'll call Company X) = "There is a high rate of employee turnover due to depression" (company response identification information i = 00003300), the recommendation candidate set generation means 37 performs the following processing:

[0133] In Figure 9, the recommendation candidate set generation means 37 first calculates the similarity (issue similarity) between the issue content vector corresponding to the company response identification information i=00003300 of the company (Company X) specified by the user and stored in the company response vector storage means 57 (see Figure 7), and the issue content vectors corresponding to the company response identification information i=00115200 (Company C), 00115300 (Company D), 00115400 (Company E), 00115500 (Company F), 00115600 (Company G), ... of all other companies (Company C, Company D, Company E, Company F, Company G, ...). In this embodiment, cosine similarity is calculated as an example. The calculated issue similarity is stored in the main memory in association with the company response identification information.

[0134] Next, the recommendation candidate set generation means 37 sorts the company response identification information for multiple other companies (Company C, Company D, Company E, Company F, Company G, ...) in descending order of similarity to the calculated issue. This sorting result is shown near the center of Figure 9. After sorting, the data itself indicating the similarity values ​​of the issues (0.94, 0.90, 0.70, 0.60, ...) are arranged from highest to lowest value, and the corresponding company response identification information is also stored in main memory in that order (Company C, Company G, Company D, Company E, ...).

[0135] Furthermore, near the center of Figure 9, the issue content data and policy implementation result data stored in the survey data storage means 51 (see Figure 3), associated with the sorted company response identification information, are also shown. However, the issue content data and policy implementation result data do not necessarily need to be stored in the main memory at this stage; they may be retrieved from the survey data storage means 51 using the company response identification information in subsequent processing. In addition, the policy implementation result cluster numbers obtained from the company response vector storage means 57 (see Figure 7) using the sorted company response identification information are also arranged in the order of the sorted company response identification information and stored in the main memory.

[0136] Furthermore, since there is a one-to-one correspondence between the issue content vector used to calculate the similarity of issues, the issue content data, the policy implementation result data, and the corresponding company response identification information, sorting the company response identification information in descending order of issue similarity is equivalent to sorting the issue content vector, issue content data, and policy implementation result data in that order. Also, there is not a one-to-one correspondence between the company response identification information and the policy implementation result cluster number. For example, there is a relationship between approximately 4200 company response identification information entries and 300 policy implementation result cluster numbers. However, since defining a company response identification information entry determines one policy implementation result cluster number, sorting the company response identification information will also sort the policy implementation result cluster numbers accordingly.

[0137] Therefore, the purpose of sorting during the preparation of policy recommendations by the recommendation candidate set generation means 37 is to rearrange the policies (policy implementation result data) in order of similarity of issues (issue content vectors, i.e., issue content data). As described above, this rearrangement is equivalent to the rearrangement of company response identification information. For this reason, the present invention (claim description) states "rearrange the company response identification information for multiple other companies in order of the degree of similarity of the calculated issues," but this description includes equivalent sorting.

[0138] Then, the recommendation candidate set generation means 37 uses the number of required measures k (see the number of measures input unit 311 in Figure 16) specified by the user and received from the output means 45, and the sorted measure implementation result cluster numbers (measure implementation result cluster numbers sorted in order of the similarity of the issues to the company response identification information) to perform the process of generating a set of recommendation candidates. In this embodiment, the number of required measures k is specified by the user, but it may also be a fixed number of required measures k predetermined by the system.

[0139] In other words, when the company response identification information sorted in descending order of similarity to the issues is arranged in main memory, the corresponding policy implementation result cluster numbers are also arranged in the same order in main memory. Therefore, the recommendation candidate set generation means 37 selects k policy implementation result cluster numbers from the top (in descending order of similarity to the issues) and selects the company response identification information corresponding to them. That is, it selects the company response identification information corresponding to each number that appears until the k-th policy implementation result cluster number appears. In other words, it selects the company response identification information in descending order of similarity to the issues until the number of policy implementation result clusters is equal to the required number of policies k. In this case, for the k-th policy implementation result cluster number (the last number), all of the corresponding company response identification information is selected as long as it has appeared consecutively since its first appearance.

[0140] Then, the set of policy implementation result data (or various data including policy implementation result data) corresponding to the selected company response identification information is defined as the recommendation candidate set. If the corresponding policy implementation result data is not stored in main memory, the set of policy implementation result data obtained from the survey form data storage means 21 using the selected company response identification information is defined as the recommendation candidate set. This makes it possible to generate a recommendation candidate set consisting of policy implementation result data belonging to the top k policy implementation result clusters.

[0141] In this process, if the same cluster number appears multiple times for the first (k-1) cluster numbers before the k-th cluster number appears, all corresponding company response identification information is selected. Therefore, if the required number of measures is k=5, even if the same cluster number appears two or more times for the first to fourth cluster numbers before the fifth cluster number appears, all corresponding company response identification information is selected to generate a set of recommendation candidates.

[0142] Therefore, the first number to appear (the cluster number of the implementation result for the case with the highest similarity of issues) is naturally the first implementation result cluster number. The next number to appear may be the same as the first implementation result cluster number or the second implementation result cluster number, but even if it is the same as the first implementation result cluster number, the corresponding company response identification information is selected to generate a set of recommendation candidates. For example, if the first number appeared 4 times, the second number 3 times, the third number 2 times, and the fourth number 5 times before the fifth implementation result cluster number (the last number) first appeared, then the company response identification information corresponding to those 4+3+2+5=14 appearances is selected. Furthermore, if the fifth number (the last number) appeared 3 times consecutively, including when it first appeared, then the company response identification information corresponding to these 3 appearances is selected, resulting in a total of 14+3=17 company response identification information being selected to generate a set of recommendation candidates.

[0143] In the example in Figure 9, [1] when the number of required measures k=1, only the measure implementation result cluster number=3 (the first number) is selected, and the various data (issue content data, measure implementation result data, etc.) associated with the company response identification information i=00115200,00115600 (Company C, Company G) corresponding to measure implementation result cluster number=3 become the set of recommendation candidates. Note that when k=1, the first measure implementation result cluster number is also the last number, so as long as that last number=3 appears consecutively (in the example in Figure 9, it appears twice consecutively), all the corresponding company response identification information is selected, and the various data (issue content data, measure implementation result data, etc.) associated with them become the set of recommendation candidates.

[0144] [2] If the number of required measures k=2, the measure implementation result cluster numbers = 3 (first number) and 26 (second number) are selected, and the various data (issue content data, measure implementation result data, etc.) associated with the company response identification information i=00115200, 001156000, 0115300 (Company C, Company G, Company D) corresponding to measure implementation result cluster numbers = 3 and 26 become the set of recommendation candidates.

[0145] [3] If the number of required measures k=3, the measure implementation result cluster numbers = 3 (1st number), 26 (2nd number), and 188 (3rd number) are selected, and the various data (issue content data, measure implementation result data, etc.) associated with the company response identification information i=00115200, 001156000, 0115300, 00115400 (Company C, Company G, Company D, Company E) corresponding to measure implementation result cluster numbers = 3, 26, and 188 become the set of recommendation candidates.

[0146] Furthermore, although the processing is performed using the policy implementation result cluster number as described above, that is, processing is performed in units of policy implementation result clusters, the reason why the required number of policies k is used instead of the required number of clusters is that the recommended policy selection means 38, described later, selects only one policy implementation result data belonging to each of the top k selected policy implementation result clusters as the recommended policy.

[0147] (Handling of cases where the received query is a task entered by the user) When the output means 45 receives a task entered by the user as a query (see the health management task input unit 351 in Figure 18), the recommendation candidate set generation means 37 receives the health management task input data (text data) as a query from the output means 45.

[0148] The recommendation candidate set generation means 37 then vectorizes the health management issue input data received from the output means 45 and calculates the similarity (issue similarity) between the obtained health management issue input vector and the issue content vectors of all companies stored in the company response vector storage means 57 (see Figure 7). In this embodiment, cosine similarity is calculated as an example. The calculated issue similarity is stored in the main memory in association with company response identification information.

[0149] Here, the process of vectorizing the health management issue input data is the same as the process of creating the issue content vector using the company response vectorization means 35 described above. This is because, since cosine similarity is calculated, the health management issue input vector to be created must have the same number of dimensions and vectorization algorithm as the issue content vector. Therefore, if Azure OpenAI Embeddings API text-embedding-ada-002 version 2 is used as the vectorization processing system 71 for creating the issue content vector, it should also be used for creating the health management issue input vector.

[0150] Furthermore, the processing after calculating the similarity of the issues is the same as described above (processing when the received query is related to a company specified by the user). However, in the case described above (processing when the received query is related to a company specified by the user), the company specified by the user is entered, so the company response identification information for all other companies is sorted in descending order of issue similarity. In contrast, in this case (processing when the received query is an issue entered by the user), the issue itself is entered, not the company specified by the user, so the company response identification information for all companies is sorted in descending order of issue similarity.

[0151] In addition, the policy extraction result storage means 60 (see Figure 8) stores policy name data and policy tags created from policy implementation result data, associated with company response identification information. However, using the same method as for creating these policy name data and policy tags, issue name data and issue tags may be created in advance from issue content data and stored in the issue extraction result storage means (not shown) associated with company response identification information. In other words, the issue name data, issue tags, and issue extraction result storage means (not shown) correspond to the policy name data, policy tags, and policy extraction result storage means 60 (see Figure 8), respectively. Then, the recommendation candidate set generation means 37 determines whether there is a match between the health management issue input data entered by the user and the issue name data or issue tag stored in the issue extraction result storage means (not shown). If a match is found, it extracts the company response identification information corresponding to the matched issue name data or issue tag, and obtains the policy implementation result data and various data (including overall evaluation or other evaluation values) corresponding to the extracted company response identification information from the survey sheet data storage means 51 (see Figure 3), as well as the health management navigation score stored in the score storage means 56 (see Figure 6), and further, the policy implementation result class stored in the company response vector storage means 57 (see Figure 7). The process involves obtaining a data number, sorting the extracted company response identification information in descending order of Health Management Navi score, overall evaluation, or other evaluation values, selecting the company response identification information when the first policy implementation result cluster number appears, selecting the company response identification information when the second policy implementation result cluster number appears, and so on, until the kth policy implementation result cluster number appears, and continuing this process until the k obtained company response identification information is used to generate a set of recommendation candidates based on the policy implementation result data and various other data corresponding to those company response identification information.

[0152] <Configuration of processing means 30 / recommendation target measure selection means 38: Figure 10>

[0153] The recommendation target measure selection means 38 selects the measure implementation result data (which can be considered as various data including measure implementation result data) with the highest Health Management Navi score, overall evaluation, or other evaluation value (see sort criterion selection unit 312 in Figure 16) from among the top k measure implementation result clusters selected by the recommendation candidate set generation means 37, as the measure to be recommended, and then executes the process of passing the selected measure implementation result data to the output means 45.

[0154] Here, the Health Management Navi score, overall evaluation, or other evaluation values ​​are the criteria selected by the user in the sorting criteria selection section 312 of Figure 16. The sorting criteria in Figure 16 are the criteria for determining recommended measures within the same measure implementation result cluster after sorting by the similarity of issues. Within the same measure implementation result cluster, it is an option to determine what type of evaluation value (health management navigation score, overall evaluation, etc.) to use and select the highest value.

[0155] Specifically, the recommendation target measure selection means 38 receives from the recommendation candidate set generation means 37 multiple company response identification information corresponding to multiple measure implementation result data constituting the recommendation candidate set, and the top k measure implementation result cluster numbers corresponding to these. From among the company response identification information assigned the same measure implementation result cluster number (company response identification information belonging to the same measure implementation result cluster), it selects the one with the highest Health Management Navi score, overall evaluation, or other evaluation value. At this time, the Health Management Navi score is obtained from the score storage means 56 (see Figure 6) using the company response identification information, and the overall evaluation or other evaluation value is obtained from the survey sheet data storage means 51 using the company response identification information. This selection process within the same cluster is then performed for all top k measure implementation result clusters (measure implementation result cluster numbers), and the obtained measure implementation result data and various data corresponding to the k company response identification information are selected as the recommended measures and passed to the output means 45.

[0156] The sorting result shown in Figure 10 is the same as the sorting result shown in Figure 9 mentioned above, but the Health Management Navi score and overall evaluation columns are added. In the example in Figure 10, the policy implementation result cluster with policy implementation result cluster number = 3 (the first number) contains two company response identification information. If the Health Management Navi score is selected as the criterion for deciding on recommended policies, then the company with the highest Health Management Navi score in this policy implementation result cluster is company response identification information i = 00115200 (Company C). Therefore, the policy implementation result data and various data corresponding to 00115200 (Company C) are selected as the policy to be recommended and passed to the output means 45.

[0157] Furthermore, the policy implementation result cluster with policy implementation result cluster number = 26 (the second number) contains only one company response identification information i = 00115300 (Company D). This is the company response identification information with the highest Health Management Navi score, and the policy implementation result data and various data corresponding to this 00115300 (Company D) are selected as the policy to be recommended and passed to the output means 45.

[0158] Similarly, the policy implementation result cluster with policy implementation result cluster number = 188 (the third number) contains only one company response identification information i = 00115400 (Company E). Therefore, this is the company response identification information with the highest Health Management Navi score, and the policy implementation result data and various data corresponding to this 00115400 (Company E) are selected as the policy to be recommended and passed to the output means 45.

[0159] <Configuration of processing means 30 / press release acquisition means 39::Figure 10>

[0160] As shown in Figure 10, the press release acquisition means 39 sequentially transmits multiple keywords (a number of keywords related to health management) stored in the keyword storage means 58 to the press release provision system 80 via the network 1, receives the press release data, as well as its title data and date data, transmitted from the press release provision system 80 via the network 1, and performs the process of storing the received press release data, as well as its title data and date data, in the press release storage means 59. The registration of keywords in the keyword storage means 58 is performed in advance by the system administrator (developer).

[0161] <Configuration of processing means 30 / title vectorization means 40: Figure 10>

[0162] As shown in Figure 10, the title vectorization means 40 vectorizes each title data of the press release acquired by the press release acquisition means 39, and stores the resulting title vector in the press release storage means 59 in association with the acquired press release data.

[0163] Here, the process of vectorizing the title data is the same as the process of creating the policy implementation result vector using the company response vectorization means 35 described above. This is because the cosine similarity with the policy implementation result vector is calculated by the press release selection means 41 described later, so the title vector to be created must have the same number of dimensions and vectorization algorithm as the policy implementation result vector. Therefore, if Azure OpenAI Embeddings API text-embedding-ada-002 version2 is used as the vectorization processing system 71 for creating the policy implementation result vector, it is also used for creating the title vector.

[0164] <Configuration of processing means 30 / press release selection means 41: Figure 10>

[0165] As shown in Figure 10, the press release selection means 41 calculates the similarity (cosine similarity in this embodiment) between each of the title vectors stored in the press release storage means 59 and the policy implementation result data (and corresponding company response identification information) selected by the recommended policy selection means 38, and the policy implementation result vector stored in the company response vector storage means 57 (see Figure 7). The press releases selected by the press release storage means 59 have a high similarity between the title and the policy, and a predetermined number (p items) or a number specified by the user (p items) of press releases. The press release selection means 41 retrieves the data, title data, and date data of the selected press releases from the press release storage means 59 and passes them to the output means 45 for screen display (see press release display unit 331C in Figure 17).

[0166] <Configuration of processing means 30 / policy name extraction means 42: Figure 8>

[0167] The policy name extraction means 42 uses the policy implementation result data of each company stored in the survey data storage means 51 (see Figure 3) to extract policies related to health management from the content expressed in the policy implementation result data. To do this, it creates one policy extraction request data for each company's policy implementation result data (including the policy implementation result data and a request to extract policies related to health management from the text described in this policy implementation result data), inputs the created policy extraction request data into a large-scale language model (LLM), and stores the LLM response data (text data indicating one or more extracted policies) output from the large-scale language model (LLM) as policy name data in the policy extraction result storage means 60 (see Figure 8) in association with company response identification information.

[0168] Specifically, the policy name extraction means 42 uses any one policy implementation result data obtained from the survey data storage means 51 (see Figure 3) to create one policy extraction request data as shown in Figure 8, transmits the created policy extraction request data to the service provision system 70 using a large-scale language model (LLM) via the network 1, receives LLM response data transmitted from the service provision system 70 via the network 1, and stores one or more text data indicating policies included in the received LLM response data as policy name data, associated with the company response identification information, in the policy extraction result storage means 60 (see Figure 8). In the example in Figure 8, using the policy implementation result data = "Aiming to promote exercise habits and communication among employees,..." from company response identification information = 00132200 (Company L), two policy name data, "Walking competition, distribution of pedometers" (pedometer is a registered trademark), are created (extracted). This process is then repeated for all policy implementation result data (all company response identification information).

[0169] <Configuration of processing means 30 / policy tag creation means 43: Figure 8>

[0170] The policy tag creation means 43 vectorizes each of the policy name data extracted by the policy name extraction means 42, performs clustering using the obtained policy name vectors, and then creates policy tag naming request data (including a set of policy name data belonging to the same policy name cluster and a request to name this set) for each of the multiple policy name clusters that have been classified. The created policy tag naming request data is input into the large-scale language model (LLM), the LLM response data output from the large-scale language model (LLM) (text data indicating the name of the policy name cluster) is attached as a policy tag to each of the policy name data belonging to the policy name cluster, and the attached policy tags are stored in the policy extraction result storage means 60 (see Figure 8) in association with the company response identification information.

[0171] Specifically, first, the policy tag creation means 43 vectorizes each of the policy name data extracted by the policy name extraction means 42 and stored in the policy extraction result storage means 60 (see Figure 8) to create a policy name vector. This vectorization process is the same as the process for creating policy implementation result vectors by the company response vectorization means 35 described above. Therefore, the policy tag creation means 43 transmits each policy name data one by one to the vectorization processing system 71 via the network 1, receives the policy name vectors transmitted from the vectorization processing system 71 via the network 1, and stores the received policy name vectors together with the corresponding policy name data in the policy name / policy tag correspondence relationship storage means 61 (see Figure 8). This process is performed for all policy name data. Note that this process is performed for all policy name data, including cases where multiple policy name data correspond to one company response identification information (such as 00132200 (Company L) and 00132600 (Company Q) in the example in Figure 8). In this embodiment, the Azure OpenAI Embeddings API text-embedding-ada-002 version 2 is used as the vectorization processing system 71.

[0172] Next, the policy tag creation means 43 performs clustering using all policy name vectors created from all policy name data. In this embodiment, this clustering process is performed in the same way as the policy implementation result clustering means 36 described above, that is, clustering is performed using BERTopic. Therefore, BERTopic is used to remove dimensions (compress the dimensions) from the high-dimensional policy name vectors (1536 dimensions in this embodiment) stored in the policy name / policy tag correspondence relationship storage means 61 (see Figure 8) to create low-dimensional policy name vectors (for example, 5-dimensional vectors), and clustering is performed using the obtained low-dimensional policy name vectors. Then, policy name cluster numbers for identifying the multiple policy name clusters that have been classified are stored in the policy name / policy tag correspondence relationship storage means 61 (see Figure 8) in association with the policy name data and policy name vectors belonging to each policy name cluster.

[0173] In this process, the number of policy name clusters formed through clustering is determined by the system administrator (developer), and is, for example, around 300. The total number of policy name data is, for example, around 10,000. From these approximately 10,000 policy name data, approximately 300 policy name clusters and policy name cluster numbers are created, and as described later, the same number of approximately 300 policy tags are created.

[0174] Next, the policy tag creation means 43 creates policy tag naming request data (including a set of policy name data belonging to the same policy name cluster and a request to name this set) for each of the multiple policy name clusters that have been classified and formed, as shown in Figure 8. Since policy name data belonging to the same policy name cluster are assigned the same policy name cluster number, the policy tag naming request data is created using a set of policy name data with the same policy name cluster number stored in the policy name / policy tag correspondence relationship storage means 61 (see Figure 8).

[0175] In the example in Figure 8, the three policy name data entries, "Walking Competition," "Walking Rally," and "Walking Event," are all assigned the same policy name cluster number = 41. Therefore, the set of these three policy name data entries constitutes a single policy name cluster.

[0176] Then, the policy tag creation means 43 transmits the created policy tag naming request data to the service provision system 70 using a large-scale language model (LLM) via the network 1, receives LLM response data (text data indicating the name of the policy name cluster) transmitted from the service provision system 70 via the network 1, assigns the received text data as policy tags to each policy name data belonging to the policy name cluster, and stores the assigned policy tags in the policy extraction result storage means 60 (see Figure 8) in association with the company response identification information.

[0177] In the example in Figure 8, a set of three policy name data, "Walking Competition," "Walking Rally," and "Walking Event," yielded the LLM response data (policy tag) "Walking Event." This "Walking Event" is then associated with "Walking Competition," "Walking Rally," and "Walking Event" and stored in the policy name / policy tag correspondence relationship storage means 61 (see Figure 8).

[0178] Furthermore, the correspondence between company response identification information and policy name data is not a one-to-one relationship, but can be a one-to-many relationship. The policy extraction result storage means 60 (see Figure 8) stores one or more policy name data associated with one company response identification information. Therefore, the policy extraction result storage means 60 also stores one or more policy tags attached to one or more policy name data in a state where they are associated with that policy name data. Accordingly, for example, in the policy extraction result storage means 60 shown in Figure 8, with company response identification information = 00132200 (Company L), the policy tag = "Walking Event" is stored in a state where it corresponds to the policy name data = "Walking Competition," and the policy tag = "Step Count Management" is stored in a state where it corresponds to the policy name data = "Distribution of Pedometers" (Pedometer is a registered trademark).

[0179] <Configuration of processing means 30 / policy tag related policy search means 44>

[0180] The policy tag lookup related policy search means 44 receives the policy tag specified by the user from the output means 45 (see evaluation result display unit 331B in Figure 17 and policy tag list display unit 380 in Figure 19), and using the received policy tag, extracts the company response identification information stored in the policy extraction result storage means 60 (see Figure 8) associated with the specified policy tag and / or the policy implementation result data stored in the survey data storage means 51 (see Figure 3) associated with the said company response identification information, and from the extracted company response identification information and / or policy implementation result data, extracts the company response The system selects a predetermined number (u items) of company response identification information and / or policy implementation result data, either in descending order of the Health Management Navi score stored in the score storage means 56 (see Figure 6) associated with the identification information, or in descending order of the overall evaluation or other evaluation values ​​stored in the questionnaire data storage means 51 (see Figure 3), either in descending order of the overall evaluation or other evaluation values, and then passes the selected u items of company response identification information and / or policy implementation result data to the output means 45 for screen display (see Figure 22) as the search result for related policies using policy tags.

[0181] <Configuration of processing means 30 / output means 45>

[0182] The output means 45 performs the process of displaying information related to health management on the screen of the user terminal 91 using the health management navigation score, overall evaluation, or other evaluation values ​​(such as various standard scores). Here, the health management navigation score is stored in the score storage means 56 (see Figure 6), and the overall evaluation or other evaluation values ​​(such as various standard scores) are stored in the questionnaire data storage means 51 (see Figure 3).

[0183] More specifically, the output means 45 performs the following processes: (1) displaying on the screen at least one of the following on the screen along with the health management navigation score or overall evaluation or other evaluation values ​​(various standard scores, etc.): the company name corresponding to the company response identification information (stored in the company information storage means 52), the issue content data, the policy implementation result data, the effect verification result data, and the combined text data; (2) displaying on the screen at least one of the following on the screen on the screen on the company name corresponding to the company response identification information, the issue content data, the policy implementation result data, the effect verification result data, and the combined text data, according to the results of selection, extraction, sorting, or grouping using the health management navigation score or overall evaluation or other evaluation values ​​(various standard scores, etc.); and (3) displaying on the screen the distribution of the health management navigation score or overall evaluation or other evaluation values ​​(various standard scores, etc.) for each company, or the year-on-year changes in the health management navigation score or overall evaluation or other evaluation values ​​(various standard scores, etc.) for each company.

[0184] (Explanation of Figure 13) Figure 13 shows the various screens displayed by the output means 45 and the transitions between these screens. First, the output means 45 displays the top screen 100, and from this top screen 100, it is possible to transition to the company profile screen 200 (see Figures 14 and 15), the policy recommendation screen 300 (see Figures 16 to 19), the similar company search screen 400 (see Figure 20), and the health management navigation score screen 500 (see Figure 21). Furthermore, from the policy recommendation screen 300, it is possible to transition to the policy tag related policy search result display screen 600 (see Figure 22), the analysis result details display screen 700 (see Figure 23), and the press release display screen 800.

[0185] In Figure 13, the company profile screen 200 (see Figures 14 and 15) is equipped with a company information display unit 210, a White 500 display unit 220, a health management brand display unit 230, an annual trend display unit 240, a benchmark comparison display unit 250, a health management excellent corporation / certification standard achievement status display unit 260, and a White 500 acquisition conditions display unit 270.

[0186] The policy recommendation screen 300 (see Figures 16 to 19) includes display sections for when the "Search by Company Name" tab 301, the "Search by Entered Issue" tab 302, and the "Policy Tag List" tab 303 are selected, as well as a condition setting section 310.

[0187] When the "Search by Company Name" tab 301 is selected, the display section (see Figures 16 and 17) includes a company issue extraction display section 320, a policy recommendation result display section 330, and a proposed policy summary display section 340. The policy recommendation result display section 330 includes a policy 1 display section 331 (including a related press release display section), a policy 2 display section 332 (including a related press release display section), ..., and a policy k display section (including a related press release display section). k is the required number of policies specified by the user, and the proposed policy summary display section 340 is a summary of k recommended policies.

[0188] When the "Search from entered issues" tab 302 is selected, the display section (see Figure 18) includes a company issue input section 350, a policy recommendation result display section 360, and a proposed policy summary display section 370. The policy recommendation result display section 360 includes a policy 1 display section 361 (including a related press release display section), ..., and a policy k display section (including a related press release display section). k is the required number of policies specified by the user, and the proposed policy summary display section 370 is a summary of k recommended policies.

[0189] When the "Policy Tag List" tab 303 is selected, the display section (see Figure 19) is provided with a policy tag list display section 380.

[0190] The similar company search screen 400 (see Figure 20) is equipped with a latest year data display section 410 and a year-on-year data display section 420.

[0191] (Explanation of Figure 14) In the company profile screen 200 shown in Figure 14, a company name input section 201 is provided, where the user enters the name of the company whose information they wish to view.

[0192] The company information display unit 210 displays the corporate name (company name), industry name, whether it is listed on the stock exchange, and its overall ranking for each year for the company specified by the user in the company name input unit 201. The corporate name (company name), industry name, and whether it is listed on the stock exchange are stored in the company information storage means 52, and the overall ranking is stored in the survey sheet data storage means 51 (see Figure 3).

[0193] The White 500 display unit 220 and the Health Management Brand display unit 230 show whether or not the company specified by the user has been certified for each fiscal year.

[0194] The annual trend display unit 240 is provided with a vertical axis selection unit 241 for selecting the content to be plotted (the content of the vertical axis of the graph) and a graph display unit 242. The graph display unit 242 shows the annual trend of the content selected in the vertical axis selection unit 241 (in the example in Figure 14, the Health Management Navi score). The annual trends of the company, industry average, and overall average related to the user specified in the company name input unit 201 are shown. The Health Management Navi score for the company is stored in the score storage means 56 (see Figure 6), and the industry average and overall average may be calculated and displayed by the output means 45 each time they are displayed, or pre-calculated values ​​may be stored in the score storage means 56 (for example, virtual subject identification information such as industry average identification information and overall average identification information corresponding to company response identification information may be provided, and the average value of the Health Management Navi score may be stored in the same form as the individual company value in association with this virtual subject identification information), or they may be stored in a statistical information storage means (not shown). If there are multiple (two) response data for a company specified by the user in the company name input section 201, the Health Management Navi score with the highest value for that company may be plotted, or the average Health Management Navi score within that company may be plotted.

[0195] (Explanation of Figure 15) In the company profile screen 200 of Figure 15, the benchmark comparison display unit 250 includes an item selection unit 251 for selecting items to plot, a graph display unit 252 for the items selected by the item selection unit 251, a difference display unit 253 for displaying the difference in standard deviation between the company specified by the user and the industry average for that item, and an advice display unit 254 for the company specified by the user. Various standard deviations are stored in the questionnaire data storage means 51 (see Figure 3), and the industry average may be calculated and displayed by the output means 45 each time it is displayed, or a pre-calculated value may be stored in the questionnaire data storage means 51 (for example, virtual entity identification information such as industry average identification information and overall average identification information corresponding to company response identification information may be provided, and the average value of various standard deviations may be stored in the same form as the values ​​for individual companies in association with this virtual entity identification information), or it may be stored in statistical information storage means (not shown). In the example shown in Figure 15, the difference display unit 253 displays the overall standard score. However, if processing is also performed to calculate a standard score for the Health Management Navi score, the Health Management Navi score (standard score) may be displayed instead of the overall standard score, or together with the overall standard score. The advice display unit 254 may display content automatically generated by a large-scale language model (LLM) using the display content of the difference display unit 253 and the graph display unit 252, or it may display content selected from a set of pre-created fixed sentences based on the content of the difference display unit 253 and the graph display unit 252.

[0196] The Health Management Excellent Corporation Certification Status Display Unit 260 includes a display unit 261 indicating whether or not each certification standard has been achieved, and an advice display unit 262 for the company regarding the user's designation of the achievement of each certification standard. The advice display unit 254 may display content automatically generated by a large-scale language model (LLM) using the content displayed in the display unit 261, or it may display content selected from a set of pre-created fixed sentences based on the content of the display unit 261.

[0197] The acquisition condition display unit 270 of the White 500 displays pre-created and prepared fixed text. This text may be written within the program, or it may be stored externally in memory.

[0198] (Explanation of Figure 16) In the policy recommendation screen 300 of Figure 16, the condition setting unit 310 for extracting policies to recommend includes a policy number input unit 311 that accepts input of the required number of policies k related to the specification of the user (a company or its person in charge that receives the policy recommendation service), a sorting criterion selection unit 312 that selects the criteria for determining recommended policies within the same policy implementation result cluster after sorting by the similarity of issues, a change rate filtering selection unit 313 that selects whether to enable filtering by the rate of change of the overall standard score for the recommended policies (a process that narrows down to policies of companies whose rate of change of the overall standard score is equal to or greater than a set value, i.e., a process that narrows down to policies of companies whose overall standard score has increased compared to the previous year), a listing classification selection unit 314 that selects whether to include listed companies and unlisted companies in the recommended policies, an industry-specific limitation selection unit 315 that selects whether to limit the recommended policies to the same industry, and an issue theme selection unit 316 that individually selects whether to include each of the 10 types of issue themes in the recommended policies. The assignment theme selection section 316 is equipped with "Check all checkboxes" and "Decheck all checkboxes" buttons, making it easy to select the required assignment theme by clicking these buttons.

[0199] In this embodiment, when the rate of change filtering selection unit 313 is selected, the filtering process uses a fixed value predetermined by the system for the rate of change setting (threshold). However, the user may be allowed to arbitrarily set the rate of change setting (threshold).

[0200] By selecting from the change rate filtering selection unit 313, the listing category selection unit 314, the same industry-only selection unit 315, and the issue theme selection unit 316, the user can narrow down the candidates for recommended measures. This makes it possible to recommend measures from companies in similar situations to one's own, or from companies that have significantly increased their evaluation of their measures.

[0201] In the policy recommendation screen 300 of Figure 16, when the "Search by Company Name" tab 301 is selected, the company issue extraction display unit 320 includes a company name input unit 321 that accepts input of a company (usually the company itself) specified by the user, an issue selection unit 322 that selects which issue to recommend policies for when the company entered in the company name input unit 321 has entered multiple (two) response data (including issue content data) in the health management survey form, and a company information display unit 323 that displays information about the company specified by the user (corporate name, industry name, whether it is listed or not, issue theme, issue content data, and policy implementation result data). The corporate name, industry name, and whether it is listed or not are stored in the company information storage means 52, and the issue theme (issue theme number), issue content data, and policy implementation result data are stored in the survey form data storage means 51 (see Figure 3).

[0202] (Explanation of Figure 17) When the "Search by Company Name" tab 301 is selected on the policy recommendation screen 300 in Figure 17, the policy recommendation result display unit 330 displays information on k policies (where k is the required number of policies specified by the user) selected by the recommended policy selection means 38 (including the company's response data related to the policy and various evaluation information and other related information derived from that response data). In this case, the information on the k policies is displayed separately for each policy in the policy 1 display unit 331, policy 1 display unit 332, ..., policy k display unit, and in addition, information integrating the k policies is displayed in the proposed policy summary display unit 340.

[0203] More specifically, the Policy 1 display unit 331, which displays the content of the recommended Policy 1, includes a company response display unit 331A, an evaluation result display unit 331B, and a press release display unit 331C.

[0204] The company response display unit 331A displays the issue theme, issue content data, measure implementation result data, and effectiveness verification result data for the recommended measure 1. This information is stored in the survey form data storage means 51 (see Figure 3).

[0205] The evaluation result display unit 331B displays the overall evaluation, health management navigation score, issue similarity, extracted measures (measure name data), and measure tags. The overall evaluation is stored in the survey form data storage means 51 (see Figure 3), the health management navigation score is stored in the score storage means 56 (see Figure 6), the issue similarity is the similarity of the issue content vectors calculated by the recommendation candidate set generation means 37 (the issue similarity shown in Figure 9), and the extracted measures (measure name data) and measure tags are stored in the measure extraction result storage means 60 (see Figure 8).

[0206] The row for the Health Management Navi Score displays a "Detailed Analysis Results" selection section. When the user clicks here, the detailed analysis results screen 700 (see Figure 23) is displayed.

[0207] Policy tags can be selected by the user with a click. When a policy tag is clicked, processing is performed by the policy tag-related policy search means 44, and the policy tag-related policy search results display screen 600 (see Figure 22) is displayed.

[0208] The press release display unit 331C displays the title data and date data for each press release selected by the press release selection means 41. The title data can be selected by the user by clicking, and when clicked, the press release display screen 800 is displayed, which shows the data content of the press release (including title, date, and body text).

[0209] Furthermore, the Policy 2 display unit 332, which displays the content of the recommended Policy 2, is equipped with a company response display unit, an evaluation result display unit, and a press release display unit, similar to the Policy 1 display unit 331 described above. Moreover, this display is the same up to the Policy k display unit.

[0210] The proposed policy summary display unit 340 displays the content of the LLM response data (summary data) obtained as output from the Large-Scale Language Model (LLM), which is a request to create a summary of the contents of policies 1 to k using the content displayed in the Policy 1 display unit 331, Policy 2 display unit 332, ..., Policy k display unit. This summary is displayed when the "Generate Summary" button is clicked.

[0211] (Explanation of Figure 18) In the policy recommendation screen 300 shown in Figure 18, when the "Search from entered issues" tab 302 is selected, the company issue input section 350 is equipped with a health management issue input section 351 where the user enters issues related to health management (usually issues that the company faces or issues similar to or related to them). Therefore, in this "Search from entered issues" tab 302, instead of the user entering a company, the user enters an issue, and the system recommends policies from each company for issues that are the same as or similar to that issue.

[0212] In the policy recommendation screen 300 shown in Figure 18, when the "Search from entered issues" tab 302 is selected, the policy recommendation result display unit 360 is provided with a policy 1 display unit 361, a policy 2 display unit 362, ..., and a policy k display unit. These are exactly the same as the display format of the policy recommendation result display unit 330 when the "Search by company name" tab 301 is selected in the policy recommendation screen 300 shown in Figure 17. Therefore, for example, the policy 1 display unit 361 is provided with a company response display unit 361A, an evaluation result display unit, and a press release display unit.

[0213] In addition, the display of information on k measures (where k is the required number of measures specified by the user) selected by the recommended measures selection means 38 (including the companies' response data related to those measures and various evaluation information and other related information derived from that response data) is the same as when the "Search by Company Name" tab 301 is selected as described above. However, the content of the selection process for recommended measures by the recommended measures selection means 38 is different. This point has already been explained in detail in the description of the recommended measures selection means 38, so a detailed explanation will be omitted here.

[0214] Furthermore, the summary display section 370 of the proposed measures is the same as when the "Search by Company Name" tab 301 is selected, and is displayed after the same processing as the summary display section 340 of the proposed measures.

[0215] (Explanation of Figure 19) When the "Policy Tag List" tab 303 is selected on the policy recommendation screen 300 in Figure 19, the policy tag list display unit 380 displays all policy tags and the number of companies implementing policies to which those policy tags are attached. These policy tags can be selected by the user with a click. When clicked, processing is performed by the policy tag-related policy search means 44, and the policy tag-related policy search results display screen 600 (see Figure 22) is displayed. The policy tags are stored in the policy extraction result storage means 60 (see Figure 8).

[0216] (Explanation of Figure 20) The similar company search screen 400 in Figure 20 is provided with a company name input section 401 for entering the company specified by the user (usually the user's own company).

[0217] The latest year data display unit 410 displays the results of comparing the company specified by the user entered in the company name input unit 401 (Company Y in the example in Figure 20) with other companies (companies other than Company Y) that have a high similarity, using the scores for the latest year (two types of scores from among the following: lateral deviation score, detailed item deviation score, issue theme deviation score, overall deviation score, health management navigation score, etc.).

[0218] Accordingly, the latest year data display unit 410 is equipped with a vertical axis selection unit 411 for selecting the type of score to be used on the vertical axis, a horizontal axis selection unit 412 for selecting the type of score to be used on the horizontal axis, and a distribution display unit 413 that shows the distribution of two types of scores for a company specified by the user (Company Y) and similar companies (companies other than Company Y) as visual distances. In addition, three axes may be selected and a three-dimensional display may be used so that three types of scores can be compared, and in this case, the three-dimensional display may be made freely rotatable on the screen.

[0219] In this embodiment, the output means 45 calculates the similarity between the two scores selected by the vertical axis selection unit 411 and the horizontal axis selection unit 412, and as an example, uses the root mean square (RMS). That is, the RMS value is calculated by adding the squared difference between the two companies for the vertical axis score (the difference between Company Y and companies other than Company Y) and the squared difference between the two companies for the horizontal axis score (the difference between Company Y and companies other than Company Y), dividing the result by 2, and taking the square root of the result. Furthermore, when using three axes for a three-dimensional display, the RMS value is calculated by adding the squared difference between the two companies for the first axis score (the difference between Company Y and companies other than Company Y), the squared difference between the two companies for the second axis score (the difference between Company Y and companies other than Company Y), and the squared difference between the two companies for the third axis score (the difference between Company Y and companies other than Company Y), dividing the result by 3, and taking the square root of the result. A smaller distance indicated by this RMS value indicates a higher degree of similarity between the two companies. Therefore, companies with small RMS values ​​(a predetermined number of companies selected in ascending order) are designated as similar companies to Company Y and plotted on the distribution display unit 413. Alternatively, similar companies to be plotted may be selected based on distances other than RMS.

[0220] Furthermore, the types of scores used to determine whether or not companies are similar may differ from the types of scores used for the vertical and horizontal axes when plotting on the distribution display unit 413. For example, there may be five types of scores used to determine whether or not companies are similar (i.e., the types of scores used for RMS calculation), and two of these five types of scores may be used for the vertical and horizontal axes when plotting on the distribution display unit 413. Moreover, one or both of the types of scores used for the vertical and horizontal axes when plotting on the distribution display unit 413 do not necessarily have to be included in the types of scores used for RMS calculation. For example, RMS may be calculated using three fixed types of scores to determine similar companies, and plots may be displayed on the distribution display unit 413 using any two types of scores (which do not have to be included in the three fixed types). Furthermore, in order to enable flexible display of these data, in addition to the vertical axis selection unit 411 and horizontal axis selection unit 412 (or, in the case of a 3D display, three axis selection units for selecting the first to third axes) for plotting the distribution display unit 413, a score selection unit for determining similar companies may also be provided for selecting any number of score types for determining similar companies.

[0221] The time-series data display unit 420 includes an item selection unit 421 for selecting the content to be plotted (the type of score to display the time-series change) for the company specified by the user (Company Y in the example of Figure 20) entered in the company name input unit 401, and a time-series change graph display unit 422 for displaying time-series change graphs for the company specified by the user (Company Y) and similar companies (companies other than Company Y) for the item (type of score) selected in the item selection unit 421.

[0222] (Explanation of Figure 21) The Health Management Navi Score screen 500 in Figure 21 includes a correlation display target score selection unit 501 for selecting a target score (vertical axis score) to display the correlation with the Health Management Navi Score (horizontal axis score), a selection unit 502 for selecting and inputting a problem theme to filter by a problem theme, a selection unit 503 for selecting and inputting a policy tag to filter by a policy tag, and a correlation display unit 510 for displaying the correlation between the Health Management Navi Score and the target score selected in the correlation display target score selection unit 501.

[0223] The correlation display unit 510 divides the Health Management Navi score, which takes values ​​from 0 to 100, into 11 ranges, and displays the median (second quartile, 50th percentile) 511, the first quartile (25th percentile) 512, the third quartile (75th percentile) 513, the minimum value 514, and the maximum value 515 of the target score (overall standard score in the example in Figure 21) within each range.

[0224] (Explanation of Figure 22) The policy tag-based related policy search results display screen 600 in Figure 22 includes a display unit 601 for policy tags related to the user specified in the search (in the example in Figure 22, "Recommend Medical Examination" and "Improve Medical Examination Rate"), a display unit 602 for the current sort criteria for this page (in the example in Figure 22, "Health Management Navi Score"), a sort criterion selection unit 603 for selecting (changing) the sort criteria for this page only, and a related policy display unit 610 for displaying policies (company names and response data of companies related to those policies) extracted by the policy tag-based related policy search means 44 using the policy tags related to the user specified in the display unit 601.

[0225] The related policy display unit 610 displays the company name, issue theme, issue content data, policy implementation result data, effectiveness verification result data, and the sorting criteria displayed in the display unit 602, or the sorting criteria selected (changed) thereafter in the sorting criteria selection unit 603, corresponding to the company response identification information obtained by searching using policy tags. Therefore, in the example in Figure 22, the policies extracted by the policy tag related policy search means 44 (company names and response data of companies related to those policies) are displayed in descending order of the Health Management Navi score, which is the current sorting criterion for this page. Furthermore, if a different sorting criterion (for example, overall evaluation) is selected in the sorting criteria selection unit 603, sorting will be performed using the changed sorting criterion (for example, overall evaluation). Note that changing the sorting criterion on this policy tag related policy search result display screen 600 is limited to this page only; to change the overall sorting criterion, use the sorting criterion selection unit 312 in the condition setting unit 310 shown in Figure 16.

[0226] (Explanation of Figure 23) The detailed analysis results screen 700 in Figure 23 displays the perspective evaluation result data for each perspective (9 perspectives in this embodiment) for the company specified by the user (usually the company itself) and the company for the recommended measures 1 to k (where k is the number of measures required by the user). (However, binary data of 1 or 0 are converted to ○ and ×.) The number of perspectives that were met (how many of the 9 perspectives were met), the health management navigation score, and the overall evaluation are also displayed. The perspective evaluation result data is stored in the perspective evaluation result storage means 55 (see Figure 5). The number of perspectives that were met is displayed by the output means 45 by summing the perspective evaluation result data.

[0227] <Configuration of memory device 50 / survey sheet data memory device 51: Figure 3>

[0228] As shown in Figure 3, the survey form data storage means 51 stores the response data of each company, including the issue theme number, issue content data, policy implementation result data, and effect verification result data, as well as data indicating the evaluation results for that response data (overall evaluation value calculated by the evaluating body, various standard scores, for example, overall standard score, aspect 1 management philosophy / policy standard score, aspect 2 organizational structure standard score, aspect 3 system / policy implementation standard score, aspect 4 evaluation / improvement standard score, etc.), in association with company response identification information.

[0229] In addition, the standard scores stored in the survey form data storage means 51 include, for example, detailed item 1_1 standard score for formalization and internal dissemination, detailed item 1_2 standard score for information disclosure and dissemination to other companies, detailed item 2_1 standard score for management involvement, detailed item 2_2 standard score for implementation system, detailed item 2_3 standard score for dissemination to employees, detailed item 3_1 standard score for goal setting and utilization of health checkups and screenings, detailed item 3_2 standard score for building a foundation for practicing health management, detailed item 3_3 standard score for health guidance, detailed item 3_4 standard score for improving lifestyle habits, detailed item 3_5 standard score for other measures, detailed item 4_1 standard score for health checkups and stress checks, detailed item 4_2 standard score for working hours and leave of absence, and detailed item 4_3 standard score for verifying the effectiveness of individual issues and overall measures and improving them.

[0230] Other challenges include: Challenge 1: Disease prevention score for all employees regardless of health status; Challenge 2: Score for preventing the worsening of diseases in high-risk individuals such as those with lifestyle-related illnesses; Challenge 3: Score for preventing, early detection, and response to stress-related illnesses such as mental health problems; Challenge 4: Score for preventing decreased employee productivity and accident prevention; Challenge 5: Score for addressing health-related issues specific to women and maintaining and improving women's health; Challenge 6: Score for returning to work after leave and balancing work and treatment; Challenge 7: Score for optimizing working hours and ensuring work-life balance and personal time; Challenge 8: Score for promoting communication among employees; Challenge 9: Score for preventing infectious diseases (such as influenza) among employees; and Challenge 10: Score for reducing the smoking rate among employees.

[0231] <Configuration of memory means 50 / corporate information memory means 52>

[0232] The corporate information storage means 52 stores corporate information such as the name of the company (in this application, including notations on drawings, it is referred to as corporate name, corporate name, company name, etc.), industry name (industry identification information), whether or not it is listed on the stock exchange, whether or not it is certified as a White 500 company, whether or not it is certified as a Health Management Brand, and whether or not it is certified as a Health Management Excellent Corporation, in association with corporate identification information (in this embodiment, it matches the first part of the corporate response identification information).

[0233] <Configuration of memory means 50 / perspective extraction result memory means 53: Figure 4>

[0234] As shown in Figure 4, the perspective extraction result storage means 53 stores the extracted perspective data in association with the pairs of company response identification information used in the extraction combination.

[0235] <Configuration of memory means 50 / setting perspective memory means 54: Figures 4 and 5>

[0236] As shown in Figures 4 and 5, the configured viewpoint storage means 54 stores multiple (nine in this embodiment) configured viewpoint data (representation-aggregated viewpoint data) in association with viewpoint numbers (1 to 9 in this embodiment).

[0237] <Configuration of memory means 50 / perspective evaluation result memory means 55: Figures 5 and 6>

[0238] The perspective evaluation result storage means 55 stores perspective evaluation result data, which is indicated by either a binary value of 1 (satisfied) or 0 (not satisfied), in association with company response identification information and perspective number.

[0239] <Configuration of memory means 50 / score memory means 56: Figure 6>

[0240] As shown in Figure 6, the score storage means 56 stores the health management navigation score (human capital management navigation score) in association with the company response identification information.

[0241] <Configuration of memory means 50 / corporate response vector memory means 57: Figure 7>

[0242] As shown in Figure 7, the company response vector storage means 57 stores the issue content vector, the policy implementation result vector, the effect verification result vector, as well as the combined text vector (not shown) and the policy implementation result cluster number, in association with the company response identification information.

[0243] <Configuration of memory device 50 / keyword memory device 58: Figure 10>

[0244] The keyword memory device 58 is used to memorize multiple keywords (a number of keywords related to health management). Keywords to be memorized include, for example, health management, collaborative health, presenteeism, absenteeism, human resources, health insurance, data health, healthcare, health tech, 36 agreement, EAP, action plan, incentives, walking events, exercise, hygiene committee, engagement, online seminars, smoking cessation, no overtime days, medical consultation recommendations, stress checks, self-care, and flexible work arrangements.

[0245] <Configuration of memory device 50 / press release memory device 59: Figure 10>

[0246] The press release storage means 59 stores the press release data obtained from the press release provision system 80, the title data, the date data, and the title vector created by the title vectorization means 40 in association with each other.

[0247] <Configuration of memory means 50 / policy extraction result memory means 60: Figure 8>

[0248] As shown in Figure 8, the policy extraction result storage means 60 stores company response identification information, one or more policy name data, and one or more policy tags in association with each other.

[0249] <Configuration of memory means 50 / policy name and policy tag correspondence memory means 61: Figure 8>

[0250] As shown in Figure 8, the policy name / policy tag correspondence storage means 61 stores policy name data, policy name vectors, policy name cluster numbers, and policy tags in association. Therefore, it stores information on which each policy name data belongs to a policy name cluster.

[0251] <Flowchart of preparation processes before service provision using the Health Management Scoring System 10: Figure 11>

[0252] In Figure 11, the health management scoring system 10 performs the following preparatory processes before providing services to users. These preparatory processes are initiated and proceeded upon operation of the administrator terminal 90 by the system administrator (developer).

[0253] First, the combining means 31 combines the issue content data, the policy implementation result data, and the effect verification result data contained in the health management survey data (see Figure 2) stored in the survey data storage means 51 (see Figure 3) to create combined text data (step S1). The details of this process have already been described in the explanation of the combining means 31, so a detailed explanation is omitted here.

[0254] Next, the viewpoint extraction means 32 uses the combined text data to perform pairwise evaluation using a large-scale language model (LLM) to extract viewpoints for evaluation, and stores the extraction results in the viewpoint extraction result storage means 53 (see Figure 4). Furthermore, using the extracted viewpoint data, multiple (9) viewpoint data to be used for evaluation are set and stored in the set viewpoint storage means 54 (see Figure 4) (step S2). The details of this process have already been described in detail in the description of the viewpoint extraction means 32, so a detailed explanation is omitted here.

[0255] Next, the viewpoint evaluation request means 33 uses the combined text data and the multiple (nine) viewpoint data stored in the set viewpoint storage means 54 (see Figures 4 and 5) to evaluate whether the content of the combined text data satisfies each viewpoint using a large-scale language model (LLM), and stores the obtained viewpoint evaluation result data (a binary value of 1 or 0) in the viewpoint evaluation result storage means 55 (see Figure 5) (step S3). The details of this process have already been described in the explanation of the viewpoint evaluation request means 33, so a detailed explanation is omitted here.

[0256] Then, the score calculation means 34 uses the perspective evaluation result data (binary values ​​of 1 or 0) stored in the perspective evaluation result storage means 55 (see Figures 5 and 6) to calculate TF-IDF, and further performs processes such as calculating the total value and normalization to calculate the health management navigation score, which is then stored in the score storage means 56 (see Figure 6) (step S4). The details of this process have already been described in detail in the explanation of the score calculation means 34, so a detailed explanation is omitted here.

[0257] Subsequently, the corporate response vectorization means 35 vectorizes the issue content data, policy implementation result data, and effectiveness verification result data stored in the questionnaire data storage means 51 (see Figure 3) to create issue content vectors, policy implementation result vectors, and effectiveness verification result vectors, and stores them in the corporate response vector storage means 57 (see Figure 7) (step S5). The details of this process have already been described in detail in the explanation of the corporate response vectorization means 35, so a detailed explanation is omitted here.

[0258] Next, the policy implementation result clustering means 36 performs clustering using the policy implementation result vectors of each company stored in the company response vector storage means 57 (see Figure 7), assigns a policy implementation result cluster number to each policy implementation result vector according to the clustering result, and stores it in the company response vector storage means 57 (step S6). The details of this process have already been described in detail in the explanation of the policy implementation result clustering means 36, so a detailed explanation is omitted here.

[0259] Also, the policy name extraction means 42 extracts policies using a large language model (LLM) with respect to the policy implementation result data stored in the questionnaire data storage means 51 (see FIG. 3), and stores the obtained policy name data in the policy extraction result storage means 60 (see FIG. 8) (step S7). Since the details of this process have already been described in detail in the description of the policy name extraction means 42, a detailed description is omitted here.

[0260] Subsequently, the policy tag creation means 43 vectorizes the policy name data stored in the policy extraction result storage means 60 (see FIG. 8) to create a policy name vector, and stores the created policy name vector together with the corresponding policy name data in the policy name - policy tag correspondence storage means 61 (see FIG. 8). Further, clustering is performed using all the obtained policy name vectors, and a policy tag is created using a large language model (LLM) with respect to the set of policy name data belonging to the classified policy name cluster, and is stored in the policy name - policy tag correspondence storage means 61 (see FIG. 8) and the policy extraction result storage means 60 (see FIG. 8) (step S8). Since the details of this process have already been described in detail in the description of the policy tag creation means 43, a detailed description is omitted here.

[0261] Furthermore, the press release acquisition means 39 acquires press release data, its title data, and date data from the press release providing system 8 using the keywords stored in the keyword storage means 58, and stores them in the press release storage means 59 (step S9). Since the details of this process have already been described in detail in the description of the press release acquisition means 39, a detailed description is omitted here.

[0262] Subsequently, the title vectorization means 40 vectorizes the title data stored in the press release storage means 59 to create a title vector, and stores the created title vector in the press release storage means 59 (step S10). Details of this process have already been described in detail in the description of the title vectorization means 40, so a detailed explanation will be omitted here.

[0263] <Flow of processing when providing services to users by the health management scoring system 10: Figure 12>

[0264] In FIG. 12, the health management scoring system 10 displays information related to health management on the screen of the user terminal 91 by the output means 45. Since the transition of the screen displayed by the output means 45 is diverse as shown in FIG. 13, the processing for the main screen display will be described here. [[ID=DNA10]]

[0265] First, the output means 45 accepts the input of the required number of measures k related to the designation of the user (a company or its person in charge receiving the service of measure recommendation) at the measure number input section 311 of the condition setting section 310 of the measure recommendation screen 300 in FIG. 16, and accepts the input of the company (usually the company itself) related to the designation of the user at the company name input section 321 of the company's issue extraction display section 320 of the measure recommendation screen 300 in FIG. 16 (step S21).

[0266] Next, the recommendation candidate set generation means 37 calculates the similarity (similarity of issues, in this embodiment cosine similarity) between the issue content vector of the company specified by the user, stored in the company response vector storage means 57 (see Figure 7), and each of the issue content vectors of multiple other companies. The company response identification information for the multiple other companies is then sorted in descending order of the calculated issue similarity. The company response identification information is then selected in order from the most similar until the number of policy implementation result clusters corresponding to the sorted company response identification information is equal to the number of required policies k specified by the user. This process generates a recommendation candidate set consisting of policy implementation result data (stored in the survey sheet data storage means 51 (see Figure 3)) belonging to the top k policy implementation result clusters (step S22). Details of this process have already been described in the explanation of the recommendation candidate set generation means 37, so a detailed explanation is omitted here.

[0267] Next, the recommendation target measure selection means 38 selects the measure implementation result data with the highest Health Management Navi score, overall evaluation, or other evaluation value from among the top k measure implementation result clusters that constitute the recommendation candidate set generated by the recommendation candidate set generation means 37 as the measure to be recommended (step S23). The details of this process have already been described in detail in the description of the recommendation target measure selection means 38, so a detailed explanation is omitted here.

[0268] Furthermore, the press release selection means 41 calculates the similarity (title-to-measure similarity, in this embodiment cosine similarity) between each of the title vectors of the multiple press releases stored in the press release storage means 59 and the measure implementation result vector (stored in the company response vector storage means 57) for the measure implementation result data selected by the recommended measure selection means 38. Based on the calculated title-to-measure similarity, a predetermined number (p items) or a number specified by the user (p items) of press releases is selected in descending order (step S24). Details of this process have already been described in the description of the press release selection means 41, so a detailed explanation is omitted here.

[0269] Then, the output means 45 displays the top k policy implementation result data selected as recommended policies by the recommended policy selection means 38 on the policy recommendation result display unit 330 of the policy recommendation screen 300 in Figure 17, and also displays the policy tags (stored in the policy extraction result storage means 60) corresponding to the policy implementation result data. Furthermore, it displays the press release information (title data, date data, and body data displayed after clicking the title) selected by the press release selection means 41 on the screen (step S25).

[0270] Next, the policy tags displayed on the policy recommendation result display unit 330 of the policy recommendation screen 300 in Figure 17, and the policy tags displayed on the policy tag list display unit 380 of the policy recommendation screen 300 in Figure 19, can be selected by the user by clicking, and the output means 45 accepts this selection operation. Then, the policy tag lookup related policy search means 44 receives the policy tag specified by the user from the output means 45, extracts the company response identification information stored in the policy extraction result storage means 60 (see Figure 8) associated with the received policy tag, and / or the policy implementation result data stored in the survey data storage means 51 (see Figure 3) associated with this company response identification information, and selects a predetermined number (u items) or a number (u items) specified by the user to be company response identification information and / or policy implementation result data in descending order of the health management navigation score or overall evaluation or other evaluation value (step S26).

[0271] Then, the output means 45 displays on the policy tag related policy search result display screen 600 the company name (stored in the company information storage means 52) and various data such as policy implementation result data that correspond to the company response identification information extracted by the policy tag related policy search means 44 (step S27).

[0272] <Effects of this embodiment>

[0273] This embodiment provides the following advantages. Specifically, the health management scoring system 10 obtains perspective evaluation result data from a large-scale language model (LLM) using a perspective evaluation request means 33, which indicates whether each of several perspectives (nine in this embodiment) is satisfied or not using a binary value (1 or 0 in this embodiment). Furthermore, the score calculation means 34 calculates a health management navigation score using TF-IDF with the perspective evaluation result data, and the output means 45 can display information related to health management on the screen using the health management navigation score.

[0274] Therefore, a Health Management Navi Score is obtained by quantifying the evaluation of the issue content data, policy implementation result data, and effect verification result data, which are free text data freely entered by each company. This Health Management Navi Score can then be used to display information related to health management on the screen. In this case, a correlation has been confirmed between the obtained Health Management Navi Score and the overall evaluation value calculated by the evaluation body of the survey data (in this embodiment, the Ministry of Economy, Trade and Industry), thus confirming the usefulness of the Health Management Navi Score. Furthermore, when nine perspectives are set for the health management survey questionnaire by the Ministry of Economy, Trade and Industry, and normalization processing is performed by dividing by the maximum value, the correlation coefficient is approximately 0.39. The correlation has also been confirmed visually using the correlation display unit 510 in Figure 21.

[0275] Furthermore, even if the evaluation result score is calculated by simply using the sum or average value for each combined text data using the perspective evaluation result data (a binary value indicating whether or not each of multiple perspectives is met) (a value indicating how many of the multiple perspectives were met, or a value proportional thereto), a correlation can be observed with the overall evaluation value. However, this evaluation result score is a numerical value that does not take into account the difficulty of achieving each of the multiple perspectives. In contrast, in this embodiment, the Health Management Navi score is calculated using the TF-IDF value, so it is a numerical value that takes difficulty into account, and the usefulness of the evaluation result score is further increased.

[0276] Furthermore, by using the Health Management Navi score obtained in this way, it is possible to recommend effective measures to users who wish to improve their evaluation of health management.

[0277] Furthermore, in the health management scoring system 10, as shown in Figure 4, the perspective extraction means 32 performs so-called pairwise evaluation to extract perspectives from the large-scale language model (LLM). Then, as shown in Figure 5, the perspective evaluation request means 33 uses the extracted perspectives to have the large-scale language model (LLM) evaluate the content of the initiatives expressed in the combined text data using multiple (9) perspectives. This allows the combined text data to be evaluated in a manner consistent with the thinking characteristics (decision processing algorithm) of the large-scale language model (LLM). Therefore, the reliability of the LLM response data can be improved.

[0278] Furthermore, the health management scoring system 10 includes a means for vectorizing company responses 35, a means for clustering policy implementation results 36, a means for generating a set of recommendation candidates 37, and a means for selecting recommended policy targets 38. Therefore, it can generate a set of recommendation candidates from the top k policy implementation result clusters and select the policy implementation result data with the highest health management navigation score within each policy implementation result cluster as the policy target for recommendation. As a result, it can recommend (suggest) a variety of policy implementation result data.

[0279] Specifically, as shown in Figure 7, the policy implementation result clustering means 36 performs clustering using policy implementation result vectors to create policy implementation result clusters. Then, as shown in Figure 9, the recommendation candidate set generation means 37 first sorts the company response identification information by the similarity of issues using the issue content vector (therefore, the policy implementation result data, policy implementation result vectors, and policy implementation result cluster numbers corresponding to the company response identification information are also sorted). Next, the top k policy implementation result cluster numbers (therefore, the top k policy implementation result clusters) are selected to generate a recommendation candidate set. Furthermore, as shown in Figure 10, the recommendation target policy selection means 38 selects the policy implementation result data with the highest Health Management Navi score among each policy implementation result cluster as the policy to be recommended, and displays it on the screen using the output means 45. Thus, a variety of policy implementation result data can be recommended.

[0280] Therefore, simply sorting company response identification information by the similarity of the issues and outputting the policy implementation result data corresponding to the top-ranked company response identification information may result in a large amount of similar policy implementation result data being output. However, by selecting the top k policy implementation result clusters and selecting and outputting the policy implementation result data with the highest Health Management Navi score within each policy implementation result cluster, it is possible to prevent a large amount of similar policy implementation result data (i.e., policy implementation result data belonging to the same policy implementation result cluster) from being output, and to output diverse policy implementation result data.

[0281] Furthermore, the health management scoring system 10 includes a press release acquisition means 39, a title vectorization means 40, a press release selection means 41, and a press release storage means 59. As shown in Figure 17, the output means 45 can display data on the implementation results of measures that have been designated as recommended measures, and the press release selection means 41 can display data on a predetermined number (p items) or a number (p items) specified by the user for the press releases. This allows the system to present users with more information related to health management.

[0282] Furthermore, since the health management scoring system 10 is equipped with a policy name extraction means 42 and a policy tag creation means 43, as shown in Figure 8, the number of policies (policy name data) extracted from the policy implementation result data by the large-scale language model (LLM) is too large. Therefore, clustering is performed to create policy name clusters, and then the large-scale language model (LLM) is used to name each policy name cluster. These names are then used as policy tags and can be displayed on the screen along with the policy implementation result data of the policies that were selected as recommended policies. As a result, it is possible to understand the content of the policy implementation result data of the policies that were selected as recommended policies by looking at the policy tags, which are a more cohesive unit than the policy name data (for example, about 10,000), which is smaller in total number, rather than the policy name data.

[0283] In addition, since the health management scoring system 10 includes the policy tag-related policy search means 44, as shown in FIG. 22, it is possible to search for and display on the screen the related policy implementation result data using the policy tags.

[0284] <Form of Variation>

[0285] Note that the present invention is not limited to the above-described embodiment, and modifications and the like within the scope capable of achieving the object of the present invention are included in the present invention.

[0286] [[ID=I4]]For example, in the above-described embodiment, the health management scoring system 10 is described as a server-client type system in which the scoring server 20, the administrator terminal 90, and the user terminal 91 are connected via the network 1, but it is not limited thereto, and all or part of it may be a stand-alone type system.

[0287] Specifically, a functional part related to the preparation process before providing services to users (see FIG. 11) among the various functions of the scoring server 20 is provided in the administrator terminal 90, and a functional part related to the process at the time of providing services to users (see FIG. 12) among the various functions of the scoring server 20 is provided in the user terminal 91. The data (such as health management navigation data) prepared in advance by the administrator terminal 90 is transmitted to the user terminal 91 via the network 1 or transferred to the user terminal 91 using a recording medium such as a DVD or a USB memory. Thereafter, the user terminal 91 alone (however, communication with an external system such as the service providing system 70 using a large language model (LLM) is performed) may perform the screen display process of the information related to health management.

[0288] Alternatively, the scoring server 20 may retain only the functional parts related to pre-service preparation (see Figure 11) from among its various functions, while providing the functional parts related to service provision to users (see Figure 12) from among its various functions to the user terminal 91. The server and the administrator terminal 90 may then perform the pre-service preparation, and the data obtained from this preparation (such as health management navigation data) may be transmitted to the user terminal 91 via the network 1, or transferred to the user terminal 91 using a recording medium such as a DVD or USB memory. After that, the user terminal 91 alone may perform the screen display processing of health management-related information (however, communication with external systems such as the service provision system 70 using a large-scale language model (LLM) may be performed). [Industrial applicability]

[0289] As described above, the human capital management scoring system and program of the present invention are suitable for use, for example, in analyzing data describing a company's health management initiatives and suggesting measures that are deemed effective for the health management challenges faced by each company, with the aim of enabling employer companies of health insurance associations to obtain certification as excellent health management companies or health management brands. [Explanation of Symbols]

[0290] 10. Health Management Scoring System 31 Coupling means 32 Perspective Extraction Means 33. Method for requesting evaluation from different perspectives 34. Score Calculation Method 35. Corporate response vectorization method 36. Clustering method for the results of policy implementation 37. Means for generating a set of recommendation candidates 38. Methods for selecting recommended measures 39. Methods for obtaining press releases 40. Title Vectorization Method 41. Press Release Selection Methods 42 Measure name extraction means 43. Methods for creating policy tags 44. Search methods for related policies using policy tags 45 Output means 51. Survey form data storage means 53. Storage means for the extraction of viewpoint results 54 Setting viewpoint storage means 55. Storage means for evaluation results from a specific viewpoint. 56. Score storage means 57 Corporate Response Vector Storage Method 58 Keyword Memory Methods 59 Press release storage means 60 Measure extraction result storage means 61. Measure Name / Measure Tag Correspondence Storage Method

Claims

1. A human capital management scoring system comprising a computer that presents information related to health management or other human capital management, A survey data storage means that stores, in association with company response identification information for identifying a company, or, when multiple responses from a single company are permitted, company and response data, the following: issue content data consisting of text indicating the content of issues, measure implementation result data consisting of text indicating the results of the measures implemented, and effect verification result data consisting of text indicating the results of the effect verification, obtained as responses from each company to a survey on health management or other human capital management; A combining means for combining the issue content data, the policy implementation result data, and the effect verification result data stored in this survey data storage means, or for combining the issue content data and the policy implementation result data to create combined text data, In order to determine whether the content expressed in the combined text data of each company created by this combining means satisfies each of several perspectives for evaluating each company's efforts regarding human capital management, a perspective evaluation request data is created for each of the combined text data of each company, including the combined text data and a request for a binary answer indicating whether the efforts expressed in the combined text data satisfy each of the several perspectives. This data is then input into a chat-generative pre-trade Transformer (ChatGPT) or other large-scale language model. The perspective evaluation request means stores the perspective evaluation result data output from this large-scale language model, which indicates whether each of the several perspectives is satisfied in one of the binary answers, in association with company response identification information and in a perspective evaluation result storage means. A score calculation means that uses the binary values ​​constituting the perspective evaluation result data for the combined text data of each company stored in the perspective evaluation result storage means to treat each of the multiple perspectives as a word and the combined text data of each company as a document to calculate term frequency-inverse document frequency (TF-IDF), and stores the sum or average of the obtained TF-IDF values ​​for each of the multiple perspectives for each of the combined text data, or the value obtained by normalizing or other processing these sum or average values, in association with company response identification information in the score storage means as a health management navigation score or other human capital management navigation score, An output means that performs at least one of the following processes as a process for displaying information related to human capital management on a screen: a process for displaying on a screen the human capital management navigation score stored in the score storage means, along with the company name corresponding to the company response identification information, the issue content data, the policy implementation result data, the effect verification result data, and the combined text data; a process for displaying on a screen the company name corresponding to the company response identification information, the issue content data, the policy implementation result data, the effect verification result data, and the combined text data according to the results of selecting, extracting, sorting, or grouping using the human capital management navigation score; and a process for displaying on a screen the distribution of the human capital management navigation score for each company or the change in the human capital management navigation score for each company over time. A human capital management scoring system characterized by having the following features.

2. The perspective extraction means prepares numerous combinations of two combined text data selected from the combined text data of each company created by the combining means, and for each of these numerous combinations, creates perspective extraction request data including two combined text data and a request for a response asking which of the two combined text data expresses the best initiative and why that conclusion was reached, and inputs this into the large-scale language model, and stores the text data output from the large-scale language model that shows why that conclusion was reached as perspective data that shows from what perspective the comparison was made and superiority or inferiority was determined, in the perspective extraction result storage means. The aforementioned means for requesting evaluation from a particular viewpoint is: The configuration uses multiple perspectives set using the perspective data stored in the perspective extraction result storage means as the multiple perspectives for evaluating each company's efforts regarding human capital management. The human capital management scoring system according to feature 1.

3. A company response vectorization means that vectorizes the issue content data and the policy implementation result data of each company stored in the survey data storage means, and stores the obtained issue content vector and policy implementation result vector in a company response vector storage means in association with company response identification information. A policy implementation result clustering means that performs clustering using the policy implementation result vectors of each company stored in the company response vector storage means, assigns a policy implementation result cluster number to each company's policy implementation result vector to identify the classified policy implementation result cluster, and stores the assigned policy implementation result cluster number in the company response vector storage means in association with company response identification information. A recommendation candidate set generation means generates a set of recommendation candidates consisting of the policy implementation result data belonging to the top k policy implementation result clusters, by calculating the similarity between the issue content vector of a company designated by a user receiving the policy recommendation service and each of the issue content vectors of multiple other companies, sorting the company response identification information for multiple other companies in descending order of the calculated similarity of issues, and selecting company response identification information in order from the one with the highest similarity of issues until the number of policy implementation result cluster numbers corresponding to the sorted company response identification information is equal to the required number of policies k designated by the user or a predetermined required number of policies k, and thereby generating a set of recommendation candidates consisting of the policy implementation result data belonging to the top k policy implementation result clusters. The system comprises a recommendation target measure selection means that selects the measure implementation result data with the highest Human Capital Management Navi Score from among the top k measure implementation result clusters selected by the recommendation candidate set generation means as the measure to be recommended. The output means is The system is configured to execute a process that displays the data showing the implementation results of the measures selected by the recommended measures selection means as the recommended measures on the screen. The human capital management scoring system according to feature 1.

4. A press release acquisition means that uses keywords stored in a keyword storage means to acquire data from multiple press releases from a press release distribution system and stores it in a press release storage means, A title vectorization means that vectorizes the title data of each of the press releases acquired by this press release acquisition means, and stores the resulting title vectors in the press release storage means in association with the acquired press releases, The system comprises a press release selection means that calculates the similarity between each of the title vectors stored in the press release storage means and the policy implementation result vectors stored in the company response vector storage means for the policy implementation result data selected by the recommended policy selection means, and selects a predetermined number of press releases or a number specified by the user that have a high similarity between the calculated titles and policies. The output means is The system is configured to display the data showing the results of the measures that were selected as recommended measures, and to display the data of a predetermined number of press releases selected by the press release selection means or a number of press releases specified by the user. The human capital management scoring system according to feature 3.

5. A policy name extraction means that, in order to extract policies related to health management or other human capital management from the content expressed in the policy implementation result data of each company stored in the survey data storage means, creates one policy extraction request data for each company's policy implementation result data, which includes the policy implementation result data and a request to extract policies related to health management or other human capital management from the text described in the policy implementation result data, and inputs it into the large-scale language model, and stores the text data indicating one or more extracted policies output from the large-scale language model as policy name data, associated with company response identification information, in the policy extraction result storage means. The system includes a policy name extraction means that vectorizes each of the policy name data extracted by this policy name extraction means, performs clustering using the obtained policy name vectors, and for each of the multiple policy name clusters that have been classified, creates policy tag naming request data that includes a set of policy name data belonging to the same policy name cluster and a request to name this set, and inputs this into the large-scale language model, and a policy tag creation means that assigns the text data indicating the name of the same policy name cluster output from the large-scale language model as a policy tag to the policy name data belonging to the same policy name cluster, and stores the assigned policy tag in the policy extraction result storage means in association with company response identification information. The output means is The system is configured to display the data showing the results of the measures that were selected as recommended measures on the screen, and to display the measure tags stored in the measure extraction result storage means, which are associated with the company response identification information corresponding to the displayed measure implementation result data, on the screen. The human capital management scoring system according to feature 3.

6. The system includes a policy tag-related policy search means that, using the policy tag specified by the user, extracts company response identification information stored in the policy extraction result storage means associated with the specified policy tag and / or policy implementation result data stored in the survey data storage means associated with the company response identification information, and from the extracted company response identification information and / or policy implementation result data, selects a predetermined number or a number specified by the user, in descending order of the human capital management navigation score stored in the score storage means associated with the extracted company response identification information. The output means is The system is configured to display on the screen the company name and / or the policy implementation result data corresponding to the company response identification information selected by the policy tag-related policy search means. The human capital management scoring system according to feature 5.

7. A program for causing a computer to function as a human capital management scoring system according to any one of claims 1 to 6.

Images