Information processing device and program for displaying recruitment appeal information.

Abstract

This invention relates to an information processing technology that optimizes the presentation of appealing information in recruitment activities based on candidate attributes and behavioral history. Conventionally, appeals have relied on the experience and intuition of interviewers, making it difficult to flexibly optimize them to reflect responses during the conversation and long-term recruitment results. In this invention, appeal actions, including the content and order of presentation, are determined based on the candidate's context and presented on a terminal, and the learning model is updated using the decision-making information obtained as a result of the execution as a reward. Furthermore, real-time updates are performed based on response estimation from the conversation log, and feedback is provided on result information including retention indicators after joining the company and reasons for declining the offer. As a result, it is possible to continuously generate appeals optimized for each candidate, improving the accuracy of recruitment activities and long-term results.

Description

Technical Field

[0001] The present invention relates to an apparatus, a method, and a program therefor for optimizing the presentation of appeal information in recruitment activities.

Background Art

[0002] In recent years, due to the decrease in the working population, the difficulty of talent recruitment has increased, and suppressing withdrawal from an offer and improving the recruitment fulfillment rate have become important issues. In conventional recruitment activities, appeals to candidates are often made based on the experience and intuition of interviewers, and there is a problem that the optimal information presentation according to individual aspirations is personalized. There is also a method of learning the correlation between candidate profile information and recruitment success or failure and generating a scouting document as in the technology described in Patent Document 1, but a mechanism for capturing the candidate's reaction obtained from the dialogue log during the interview in real time and updating the appeal items, or reflecting the period until settling in the company after joining as a reward in learning has not been sufficiently established. For this reason, there is a need for an information processing technology that can automatically determine an optimal appeal action according to the context based on the attributes and action history of candidates through learning using result information.

Prior Art Documents

Patent Documents

[0005] Furthermore, conventional learning models tend to reward short-term results such as whether or not a candidate is offered a position, and the cycle of using post-employment retention, evaluation, and specific reasons for declining an offer to improve future recruitment efforts is insufficient. Therefore, there is a need for technology that continuously learns the relationship between a candidate's context, recruitment actions, and decision-making outcomes, and automatically updates the model. In order to maximize long-term recruitment results while incorporating considerations for interviewers' strengths and weaknesses and sensitive attributes, it is a crucial challenge to automatically derive the optimal recruitment method for each individual candidate. [Means for solving the problem]

[0006] The information processing device of the present invention acquires candidate context, including candidate attributes and behavioral history, and determines and presents appeal actions, including the content and order of presentation of recommended appeal information, based on that context. Furthermore, it updates the model by learning the relationship between candidate context, appeal actions, and result information obtained as a result of the appeal, using the result information regarding the candidate's decision as a reward. This enables personalized and optimal appeals to individual candidates through real-time updates of appeal items based on dialogue logs, reflection of retention indicators and reasons for declining after joining the company in rewards, text generation by generative AI, and topic presentation according to the interviewer's tendencies, thereby maximizing the results of recruitment activities. [Effects of the Invention]

[0007] According to this invention, it is possible to optimize appeal actions based on candidate attributes and responses during conversations, thereby dramatically increasing the acceptance rate of job offers compared to conventional, person-dependent methods that relied on the interviewer's experience and intuition. In particular, by presenting personalized appeal information using generated AI and providing real-time support that complements the strengths and weaknesses of each interviewer, it is possible to standardize the quality of recruitment operations while minimizing wasted activities, which offers economic advantages. Furthermore, by feeding back retention indicators after joining the company and reasons for declining offers as rewards for learning, it has the technological effect of contributing to long-term improvement in recruitment accuracy and organizational growth, not just short-term job offer acquisition. [Brief explanation of the drawing]

[0008] [Figure 1] This is a block diagram showing the overall configuration of an information processing system according to one embodiment of the present invention. [Figure 2] This figure shows an example of a data structure for candidate context, appeal actions, and outcome information. [Figure 3] This flowchart shows the steps involved in determining recommended appeal actions based on the candidate's context. [Figure 4] This is a conceptual diagram of the process of updating the learning model using result information as a reward. [Figure 5] This is a flowchart illustrating the process of updating and presenting appeal points in real time based on the analysis of the dialogue log. [Figure 6] This diagram shows a configuration that generates persuasive text by combining factual data held in a knowledge base with generative AI. [Figure 7] This diagram illustrates the process of prioritizing recommended topics based on a model of each interviewer's appeal tendencies. [Figure 8] This flowchart shows the steps for setting learning constraints to exclude specific attributes and outputting audit logs. [Figure 9] This figure shows an example of how recommended appeal items and dialogue support information are displayed on the device screen. [Figure 10]It is a diagram showing a configuration that feeds back the settlement rate and evaluation indicators after joining the company as result information for learning.

Embodiment for Carrying Out the Invention

[0009] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[0010] This embodiment relates to an information processing apparatus and a program for optimizing the presentation of solicitation information in the recruitment activity (see FIGS. 1 to 10).

[0011] More specifically, this embodiment relates to a system for improving a method of conveying the attractiveness of a company to candidates in a recruitment process such as an interview.

[0012] In the recruitment process, it is required to present information suitable for candidates' aspirations and attributes.

[0013] By presenting appropriate information, it becomes possible to enhance the motivation of candidates.

[0014] The system of this embodiment supports the presentation of this information based on data.

[0015] FIG. 1 is a block diagram showing the overall configuration of an information processing system 100 according to an embodiment of the present invention.

[0016] As shown in FIG. 1, the information processing system 100 is configured to include a plurality of devices and provides an overall basis for realizing the optimization of solicitation information in the recruitment activity.

[0017] The information processing system 100 includes an appeal optimization server 110 as a main information processing device.

[0018] The information processing system 100 includes a database for storing various types of data.

[0019] The information processing system 100 includes an interviewer terminal 130 for use by the interviewer.

[0020] Each of these devices is connected to the others via a communication network, enabling them to communicate data with one another.

[0021] Communication networks are built using the internet, dedicated lines, virtual private networks, and other similar technologies.

[0022] Alternatively, the communication network may include the company's internal network.

[0023] The appeal optimization server 110 is a server computer that is responsible for the central computing processing of the system, and may be implemented as a virtual server in a cloud computing environment or as a physical server device in an on-premises environment.

[0024] The appeal optimization server 110 consists of a computer system equipped with a processor and memory.

[0025] The processor of the appeal optimization server 110 performs various calculations and control processes.

[0026] The memory of the appeal optimization server 110 temporarily or permanently stores the programs and data necessary for processing.

[0027] The appeal optimization server 110 has the function of selecting recommended appeal information from a list of candidates that should be presented to the candidate.

[0028] This selection will be made according to objective criteria based on data.

[0029] The appeal optimization server 110 transmits the determined appeal action 210 to the interviewer terminal 130.

[0030] As shown in Figure 1, DB is a general term for the various databases used in the information processing system 100.

[0031] The database may be a storage device located inside the appeal optimization server 110.

[0032] Alternatively, the database may be an external database server that is physically independent of the appeal optimization server 110.

[0033] As shown in Figure 2, the database stores a variety of data related to recruitment activities in a structured manner.

[0034] The database contains a table that holds information about candidate context 200.

[0035] The database contains tables that hold information about the interviewers.

[0036] The database includes a table that serves as a knowledge base 500, which holds data such as the persuasive text 530.

[0037] The database contains tables that hold data related to past interview history and results information.

[0038] The interviewer terminal 130 is an information and communication terminal operated by the interviewer in charge of the interview.

[0039] The interviewer terminal 130 is comprised of, for example, a personal computer.

[0040] The interviewer terminal 130 may consist of a mobile device such as a tablet or a smartphone.

[0041] The interviewer terminal 130 is equipped with a display screen.

[0042] The interviewer's terminal 130 displays information received from the appeal optimization server 110.

[0043] The interviewer terminal 130 may be equipped with a microphone for voice input.

[0044] The interviewer terminal 130 may be equipped with a camera for video input.

[0045] The interviewer terminal 130 has the function of acquiring a dialogue log 400, which includes audio and video of the conversation during the interview, and sending it to the appeal optimization server 110.

[0046] The interviewer terminal 130 may be equipped with an application for running a web-based interview system.

[0047] The interviewer terminal 130 may be configured to communicate with the appeal optimization server 110 via a web browser.

[0048] The overall configuration of the information processing system 100 (see Figure 1) seamlessly connects the interview process with the data analysis infrastructure.

[0049] This allows interviewers to receive support in real time (see Figure 5) or in advance (see Figure 9).

[0050] The connection between the appeal optimization server 110 and the interviewer terminal 130 is protected by a secure communication protocol.

[0051] There are multiple interviewer terminals 130, and it is common for multiple interviewers to use individual terminals.

[0052] Data transmitted from multiple interviewer terminals 130 is centrally aggregated by the appeal optimization server 110.

[0053] The aggregated data is stored in the database.

[0054] In this way, the appeal optimization server 110, the interviewer terminal 130, and the database work closely together.

[0055] This collaboration creates a data-driven, continuous improvement loop in recruitment activities (see Figures 4 and 10).

[0056] Figure 2 is a diagram showing an example of the definition of the data structure that forms the basis of learning in this embodiment, and it shows the basic data structure handled in the system.

[0057] As a prerequisite for the operation of the information processing system 100, the input data, output data, and result data are clearly defined (see Figure 3).

[0058] The data structure shown in Figure 2 includes candidate context 200, appeal action 210, and result information 220.

[0059] The combination of these three data sets constitutes the training data for optimization.

[0060] The candidate context 200, appeal action 210, and result information 220 are associated with each other and stored in the database.

[0061] This association is performed using, for example, a unique candidate ID or interview session ID as a key.

[0062] Candidate context 200 is a collection of information about a candidate's attributes and behavioral history.

[0063] Candidate Context 200 functions as an input feature to represent the candidate's current situation and preferences.

[0064] The candidate context 200 includes explicit attribute information.

[0065] Explicit attribute information may include information about the candidate's age.

[0066] Explicit attribute information may include information about the candidate's educational background.

[0067] Explicit attribute information may include information about the candidate's field of study.

[0068] Explicit attribute information may include information about the candidate's place of residence and preferred work location.

[0069] Additionally, the candidate context 200 includes information about the candidate's behavioral history.

[0070] Information regarding behavioral history includes the results of aptitude tests.

[0071] Information regarding the history of actions includes the contents of the application form.

[0072] Information regarding behavioral history includes evaluation history from past interviews.

[0073] Information regarding behavioral history may include the results of an analysis of self-introduction videos.

[0074] Furthermore, the candidate context 200 may include estimated data regarding the candidate's values ​​and motivations.

[0075] This information provides clues to determine what kind of appeal is most likely to resonate with the candidate.

[0076] The candidate context 200 can be dynamically updated as the recruitment process progresses.

[0077] For example, the candidate context 200 at the initial selection stage may differ from the candidate context 200 at the final interview stage.

[0078] As the interview progresses, evaluation comments from the interviewer and other information are added to the candidate context 200.

[0079] The data items for Candidate Context 200 are stored in the Candidate Information table in DB120.

[0080] The data structure of candidate context 200 may be vectorized to make it easier to handle as input for machine learning.

[0081] During the vectorization process, the content of the application form, written in natural language, is converted into numerical data.

[0082] Appeal Action 210 is a data structure that defines the content of the actions that companies take to engage with candidates.

[0083] Appeal Action 210 is data related to appeal information that has been presented or is scheduled to be presented (see Figure 2).

[0084] Appeal Action 210 is an output from the system and also functions as a learning log.

[0085] Appeal Action 210 includes information regarding the content of the appeal information to be presented.

[0086] The content presented consists of topics and subjects designed to specifically convey the company's appeal.

[0087] The categories of topics to be presented include education and training systems, employee benefits and leave policies, job appeal and autonomy, organizational climate and culture, and compensation and evaluation systems.

[0088] The appeal action 210 includes identifiers and tags that represent these topics.

[0089] Furthermore, appeal action 210 includes information regarding the order in which the appeal information is presented.

[0090] The presentation order refers to information indicating which topics were presented first and which were presented later within the limited time available for an interview.

[0091] For example, the order in which the organizational culture is discussed first, followed by the job description, is recorded.

[0092] The data presented in a specific order is stored in the data structure as a time-series list or array.

[0093] Furthermore, appeal action 210 may include information on specific speaking techniques and scripts (see Figure 6).

[0094] Discourse refers to information that indicates how a topic is discussed, including the perspective and expression used.

[0095] For example, different speaking styles, such as calmly conveying facts or passionately sharing employee experiences, can be tagged.

[0096] The data for appeal action 210 also includes information about the channels and media through which the appeal was presented.

[0097] Channel information refers to distinctions such as whether the interview was conducted in person, online, or via email.

[0098] The appeal action 210 is customized and generated for each candidate, and its generation history is stored in the appeal optimization server 110.

[0099] This saved history of appeal actions 210 serves as evidence to verify what appeals were made later.

[0100] The result information 220 is a data structure related to the candidate's decision-making obtained as a result of executing the appeal action 210 (see Figure 2).

[0101] The result information 220 represents the variables and rewards 310 that serve as learning objectives in the information processing system 100 shown in Figure 1.

[0102] Result information 220 includes the facts of the candidates' final decision.

[0103] Typical examples of decision-making events include accepting or declining a job offer.

[0104] Furthermore, withdrawals during the selection process are also included as facts in the results information 220.

[0105] Result information 220 includes not only whether or not a fact exists, but also information about the reasons behind it.

[0106] If a candidate withdraws, the category of the reason for their withdrawal is recorded in the results information 220.

[0107] Reasons for declining an offer include finding employment at another company, mismatch with desired job type, dissatisfaction with work location, or dissatisfaction with salary and other benefits.

[0108] These categories of reasons for refusal are incorporated into the information processing system 100 as teacher information.

[0109] The results information 220 may include data on whether candidates passed or failed each selection stage, or subjective evaluation data such as questionnaire responses.

[0110] Furthermore, as shown in Figure 10, retention rates (910) as post-employment indicators (900) may be fed back into learning as result information (220).

[0111] The results information 220 is registered with the appeal optimization server 110 when the recruitment activities are completed or when each phase is finished.

[0112] These three elements—candidate context 200, appeal action 210, and result information 220—are linked together.

[0113] The collection of linked data records becomes a dataset for measuring the effectiveness of the promotional information.

[0114] For example, the results of performing an appeal action 210 with a specific presentation order on a person with a candidate context 200 of certain attributes are recorded and used to update the learning model 300, as shown in Figure 4.

[0115] If the result information 220, which is acceptance of a job offer, is obtained, that combination is stored as a success case and used to update the learning model 300 (see Figure 4).

[0116] If information 220 is obtained indicating that a job offer was declined, that combination will be stored as a failure example.

[0117] By adopting such a data structure, the information processing system 100 can comprehensively learn from past cases (see Figure 3).

[0118] Detailed definitions of candidate contexts allow for the representation of candidates' diverse personalities as data.

[0119] The detailed definition of the 210 appeal actions allows for accurate recording of variations in the company's outreach efforts.

[0120] The detailed definition of result information 220 allows for quantitative evaluation of performance, including post-hire indicators 900 (see Figure 10).

[0121] These data structures are implemented on the database as either a relational model or a document model.

[0122] The candidate context 200 table is structured with candidate ID as the primary key.

[0123] The table for appeal action 210 is structured with the action ID as the primary key and the candidate ID as the foreign key.

[0124] The table for result information 220 is structured with the result ID as the primary key and the candidate ID as the foreign key.

[0125] This data structure makes it easy to combine and retrieve a series of historical records related to a specific candidate.

[0126] The appeal optimization server 110 reads and writes data to and from the database according to this data structure (see Figure 1).

[0127] The information entered through the interviewer terminal 130 is converted into an appropriate data structure by the appeal optimization server 110.

[0128] For example, when an interviewer selects a topic to promote on the presentation screen 800, the data is parsed as data for a promotional action 210 by referring to the knowledge base 500 (see Figure 6) (see Figure 9).

[0129] The evaluations entered by the interviewer and the reasons for declining the job offer are recorded as data in the results information 220, along with the response estimation results based on the analysis of the dialogue log 400 (see Figure 5).

[0130] Some of the information regarding Candidate Context 200 may be obtained from an external recruiting management system via an API.

[0131] Data obtained from external systems is normalized into the format of candidate context 200 shown in Figure 2.

[0132] This normalized basic data structure makes it easier to exchange data between different systems.

[0133] Furthermore, standardizing the data structure makes it possible to stably supply high-quality training data with less noise.

[0134] The candidate context 200 can include accumulated time-series features and is used for excluding sensitive attributes 700 (see Figure 8) and applying the interviewer model 600 (see Figure 7).

[0135] Appeal Action 210 is also recorded as a chronological action log in accordance with the progress of the interview.

[0136] This allows us to capture dynamic changes across the entire process, rather than just a single static data point.

[0137] The data structure shown in Figure 2 is the core information representation framework of the information processing system 100.

[0138] The attribute fields in the candidate context 200 can be added or removed according to the company's hiring policy, and constraints can be applied to exclude certain sensitive attributes 700, as well as the output of audit logs 710 (see Figure 8).

[0139] The categories of the 210 appeal actions can also be flexibly customized according to changes in a company's business operations and systems.

[0140] The 220 items of results information can also be expanded to match the KPIs that companies prioritize, and these can be used as feedback indicators such as the 900 post-hire indicators and the 910 retention rate (see Figure 10).

[0141] As a result, the information processing system 100 has the versatility to be applied to the recruitment processes of various companies.

[0142] The appeal optimization server 110 periodically scans this structured data stored in the database.

[0143] The scanned data is loaded into memory space for analysis.

[0144] The interviewer terminal 130 exchanges information via the user interface without being aware of the details of these data structures.

[0145] A portion of the candidate context 200 is clearly displayed on the presentation screen 800 of the interviewer's terminal 130 (see Figure 9).

[0146] On the display screen 800, dialogue support information 810 is presented based on the interviewer model 600's tendencies, including suggested appeal actions 210 to be taken next and recommended topics 610, which are recommended based on the candidate context 200 (see Figure 3).

[0147] Then, via the interviewer's terminal 130, a portion of the result information 220 is input after the interview is completed, and this is used as a reward 310 to update the learning model 300 (see Figure 4).

[0148] The overall configuration of the information processing system 100 (see Figure 1) seamlessly integrates the on-site business flow and data collection process.

[0149] Data collection is carried out naturally as an extension of the interview process, and the system is designed not to increase the burden on interviewers.

[0150] The transmission of data from the interviewer terminal 130 to the appeal optimization server 110 may be performed asynchronously.

[0151] Furthermore, the system may also include a function to automatically extract the history of appeal actions 210 using a response estimation unit 410 that processes audio data during the interview in the background and analyzes the dialogue log 400 (see Figure 5).

[0152] In this case, speech recognition technology is used to convert the spoken content into text.

[0153] Keywords corresponding to specific appeal categories are extracted from the transcribed speech content.

[0154] Based on the extracted keywords, data on the content and presentation order of the appeal text 530 and other elements in the appeal action 210 are automatically generated by referring to the factual data 510 in the knowledge base 500 (see Figure 6).

[0155] This eliminates the need for interviewers to manually record the appeal actions 210 (see Figure 7).

[0156] The automatically generated data for appeal action 210 is stored in the database (see Figure 6).

[0157] Regarding result information 220, it is also possible to configure it to be automatically acquired in conjunction with the electronic signature system for the acceptance letter of employment (see Figure 10).

[0158] When a candidate accepts the signature using the electronic signature system, that information is reflected in the database as result information 220.

[0159] Alternatively, a status change in the recruitment management system can trigger the generation of result information 220 (see Figure 3).

[0160] The information processing system 100 ensures data integrity by combining manual input and automatic acquisition.

[0161] These data stored in the database are protected by appropriate access controls.

[0162] Candidate context 200, which includes personal information, is encrypted and stored in the database.

[0163] The appeal optimization server 110 may also be equipped with a function for anonymizing data (see Figure 8).

[0164] After anonymization is performed to make it impossible to identify the candidates, the learning process using the learning model 300 is carried out (see Figure 4).

[0165] The configuration and data structure shown in Figures 1 and 2 form the basis for the operation of the system in this embodiment.

[0166] The close coordination between each component and the data structure enables the optimization of objective appeals that do not rely on empirical rules.

[0167] The processing power of the appeal optimization server 110 and the storage capacity of the database can be scaled up according to the scale of adoption.

[0168] The presentation screen 800 on the interviewer's terminal 130 can be updated to support new methods of presenting appeal actions 210 (see Figures 5 and 9).

[0169] The information processing system 100 supports the improvement of the quality of recruitment activities through this series of mechanisms.

[0170] Next, we will further explain examples of detailed data fields for Candidate Context 200.

[0171] The candidate context 200 data fields include skill set information such as programming language experience and language proficiency.

[0172] Furthermore, the candidate context 200 data fields include information about the candidate's reasons for changing jobs and their motivations for applying, which are stored as a set of keywords extracted from the free-response section of the application form.

[0173] The candidate context 200 data fields contain information about the candidate's career plan.

[0174] Flag information, such as whether the user is management-oriented or specialist-oriented, is retained.

[0175] These detailed fields provide an even higher level of clarity regarding the candidates.

[0176] Next, we will further explain examples of detailed data fields for appeal action 210.

[0177] The data field for appeal action 210 includes information about the time taken for the appeal.

[0178] Time data is stored to show how many minutes a particular topic was discussed.

[0179] The data field for appeal action 210 contains the ID of the material used for the appeal.

[0180] Identification information for slides and videos presented via screen sharing during the interview is retained.

[0181] These detailed fields allow for an analysis of which presentation methods were most effective.

[0182] Next, we will further explain examples of detailed data fields in result information 220.

[0183] The data fields in the results information 220 include information about the time period leading up to the decision.

[0184] The number of days between the offer of employment and the acceptance or rejection of the offer is recorded.

[0185] The data fields in result information 220 include information about the selection status of competitors.

[0186] Information is retained about what other companies the candidate applied to and which companies they chose.

[0187] These detailed fields allow for a deeper understanding of the true reasons for declining and the deciding factors for accepting.

[0188] The overall structure and data structure definition described above serve as a prerequisite for the advanced information processing described later.

[0189] The appeal optimization server 110 uses this data to perform the next step of processing.

[0190] The interviewer terminal 130 continues to function as the frontline interface for collecting this data.

[0191] DB120 continues to store information while maintaining data consistency and availability.

[0192] This completes an information processing system aimed at optimizing persuasive information used in recruitment activities.

[0193] Referring to Figure 3, the decision-making process for the appeal action 210 based on the learning model 300 will be explained.

[0194] Figure 3 is a flowchart showing the steps involved in determining the recommended appeal action 210 based on the candidate context 200.

[0195] The information processing system 100 (see Figure 1) acquires candidate context 200 before the start of the interview or during the interview.

[0196] Candidate Context 200 is a dataset containing diverse information about the candidates (see Figure 2).

[0197] Candidate Context 200 contains information about the candidate's attributes.

[0198] Information regarding attributes could include, for example, the candidate's age.

[0199] Additionally, the candidate's gender may be included as part of their attribute information.

[0200] Furthermore, the candidate's educational background and work experience may be included as attribute information.

[0201] Candidate Context 200 also includes information about the candidate's behavioral history.

[0202] The user's activity history includes the content of job postings they have previously viewed.

[0203] The candidate's behavioral history may include a record of company information sessions they have previously attended.

[0204] Furthermore, the results of aptitude tests and responses to pre-selection questionnaires may be included in the candidate context 200.

[0205] The information processing system 100 preprocesses the acquired candidate context 200.

[0206] As part of the preprocessing, the text data is converted to numerical values ​​and normalized.

[0207] If missing values ​​exist, imputation is performed during preprocessing.

[0208] The pre-processed candidate context 200 is input into the learning model 300.

[0209] Learning model 300 is a mathematical model that has been pre-constructed using machine learning algorithms.

[0210] The learning model 300 is constructed using, for example, a neural network.

[0211] Alternatively, the learning model 300 may be an ensemble learning model using decision trees.

[0212] The learning model 300 evaluates potential appeals to present to candidates based on the input candidate context 200.

[0213] Multiple potential appeals are pre-registered in the information processing system.

[0214] Potential appeal information includes details about a company's employee benefits.

[0215] It also includes information about companies' education and training programs.

[0216] Furthermore, it includes information about the company's organizational culture and career paths.

[0217] The learning model 300 calculates the probability that a candidate will respond positively to each candidate appeal.

[0218] The probability is calculated using data from similar candidates in the past.

[0219] Appeal information whose calculated probability exceeds a predetermined threshold is selected as recommended appeal information.

[0220] Alternatively, the top multiple appeals with the highest calculated probabilities may be selected.

[0221] The information processing system determines the content to be presented regarding the selected appeal information.

[0222] The presentation should include specific messages that should be conveyed as persuasive information.

[0223] Furthermore, if multiple pieces of persuasive information are selected, the information processing system determines the order in which they are presented.

[0224] The presentation order is rearranged to best attract the candidate's attention.

[0225] In this way, the information processing system generates an appeal action 210, which includes the content and order of presentation.

[0226] The determined appeal action 210 is output from the learning model 300.

[0227] The output appeal action 210 is temporarily stored in the memory of the information processing system.

[0228] Subsequently, the information processing system sends the appeal action 210 to the interviewer's terminal 130.

[0229] This procedure makes it possible to automatically determine the appropriate, personalized appeal action for each candidate.

[0230] This has the effect of extracting objective, data-driven appeals without relying on the interviewer's experience or intuition.

[0231] Next, referring to Figure 5, we will explain the real-time processing using the dialogue log 400 during the interview.

[0232] Figure 5 is a flowchart illustrating the process of updating and presenting appeal items in real time based on the analysis of 400 dialogue logs.

[0233] The interview between the interviewer and the candidate is conducted via the information processing system 100 or using the interviewer's terminal 130 (see Figure 1).

[0234] Spoken words during the interview are picked up via a microphone.

[0235] The collected audio data is sequentially converted into text by the information processing system 100.

[0236] This text conversion process generates 400 interview dialogue logs.

[0237] The generated dialogue log 400 is stored in memory in chronological order.

[0238] The information processing system 100 reads the accumulated dialogue log 400 at regular intervals.

[0239] The retrieved dialogue log 400 is input to the response estimation unit 410.

[0240] The response estimation unit 410 is a module that analyzes the dialogue log 400 using natural language processing technology.

[0241] The response estimation unit 410 estimates the candidate's response to the appeal information presented by the interviewer (see Figure 5).

[0242] The frequency of positive and negative words in the candidate's utterances is used to estimate the response.

[0243] In addition, nonverbal elements such as the frequency of verbal affirmations and the tone of speech may also be taken into consideration.

[0244] The response estimation unit 410 classifies the candidates' responses into categories such as high interest, low interest, or concern.

[0245] If the level of interest is estimated to be high, the response estimation unit 410 makes a determination to recommend further exploration of that topic.

[0246] Conversely, if the level of interest is estimated to be low, the response estimation unit 410 determines that the topic should be changed to a different one.

[0247] If it is presumed that a candidate has expressed concerns about a particular system, it will be determined that supplementary information should be provided to address those concerns.

[0248] The information processing system 100 receives the judgment result from the reaction estimation unit 410.

[0249] The information processing system 100 then determines the next appeal item to be presented.

[0250] This decision-making process is carried out in real time as the interview progresses, following the procedure shown in Figure 5.

[0251] For example, a new appeal item is calculated within a few seconds of the previous utterance.

[0252] The updated appeal items are defined as new appeal actions 210 (see Figure 2).

[0253] The information processing system 100 immediately transmits the updated appeal action 210 to the interviewer terminal 130.

[0254] The interviewer's terminal 130 displays the received appeal action 210 on the display screen 800 (see Figure 9), etc.

[0255] The interviewer can dynamically adjust the direction of the conversation while referring to the updated appeal actions 210.

[0256] As shown in Figure 5, the system estimates the candidate's response from the dialogue log 400 and updates the appeal items to be presented next in real time.

[0257] This has the effect of enabling immediate adaptation to changes in the flow of the interview and the candidate's responses.

[0258] By appropriately changing the topic when the candidate's interest wanes, you can maintain a high level of interview quality.

[0259] Next, with reference to Figure 6, we will explain the generation logic of the appeal text 530.

[0260] Figure 6 shows a configuration that generates persuasive text 530 by combining factual data 510 stored in the knowledge base 500 with a generating AI.

[0261] The information processing system 100 is equipped with a knowledge base 500 accumulated within the company.

[0262] Knowledge base 500 is a database in which information that can be appealing to candidates is stored in a structured manner, and it holds a large amount of factual data 510.

[0263] Fact Data 510 consists of factual information such as the official policies of a company and actual working conditions.

[0264] For example, factual data 510 stores accurate figures for paid leave utilization rates and average overtime hours.

[0265] Furthermore, the eligibility requirements for various allowances and the specific contents of the training curriculum are included as factual data 510.

[0266] The information processing system 100 retrieves the relevant factual data 510 from the knowledge base 500 based on the determined appeal action 210.

[0267] The search extracts 510 factual data points related to the target appeal topic.

[0268] The extracted factual data 510 is input into the generating AI.

[0269] Generative AI is a module that generates text using natural language, and is constructed using large-scale language models and other methods.

[0270] The information processing system 100 combines the extracted factual data 510 and the candidate context 200 to construct an input prompt for the generating AI.

[0271] The prompts include instructions to generate text in a tone and style that matches the candidate's preferences and attributes.

[0272] AI520 automatically generates persuasive text 530 based on factual data 510, following the prompt's instructions.

[0273] The generated persuasive text 530 is composed of conversational language that can be read aloud by the interviewer.

[0274] Alternatively, it may be generated as bulleted text to be displayed to candidates on the screen.

[0275] Because the text is generated based on clear evidence in the form of factual data 510, the risk of AI 520 creating content that is contrary to the facts is reduced.

[0276] This hybrid generation logic results in an appeal text 530 that combines accuracy and flexibility.

[0277] The generated appeal text 530 is provided to the interviewer terminal 130 as part of the appeal action 210.

[0278] <​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​

[0288] In the process of determining the appeal action 210, the information processing system refers to the interviewer model 600 of the responsible interviewer.

[0289] Then, the candidates for a plurality of appeal topics to be presented to the candidate are compared with the interviewer's area of expertise.

[0290] The information processing system preferentially selects the topics that the interviewer is good at as the recommended topics 610.

[0291] Regarding the topics that the interviewer is not good at, adjustments are made to replace them with other topics that the interviewer is good at as much as possible (see Fig. 7).

[0292] If it is necessary for the interviewer to present a topic that the interviewer is not good at due to the candidate's orientation, the information processing system 100 performs special control.

[0293] Specifically, control is performed to present a detailed script that the interviewer can just read out as it is.

[0294] The determined recommended topics 610 are transmitted to the interviewer terminal 130.

[0295] The interviewer terminal 130 preferentially displays the received recommended topics 610 at the top of the screen.

[0296] By the control using such an interviewer model 600, interview support optimized for the skill set of each interviewer is realized.

[0297] Since topics that the interviewer can talk about with confidence are selected, there is an effect of increasing the persuasiveness to the candidate.

[0298] It prevents insufficient explanation due to topics that the interviewer is not good at and contributes to improving the overall recruitment ability of the company.

[0299] Finally, referring to FIG. 9, the screen configuration of the interviewer terminal 130 will be described.

[0300] FIG. 9 is a diagram showing an example of the UI configuration of the dialogue support information 810 and the recommended appeal items presented on the screen of the interviewer terminal 130.

[0301] The interviewer terminal 130 receives various types of information transmitted from the information processing system 100 and displays it on the presentation screen 800.

[0302] The presentation screen 800 has a layout organized so that it is easy for the interviewer to view during the interview.

[0303] In the central part of the presentation screen 800, an area indicating the progress of the current dialogue is arranged.

[0304] In this progress area, the history of the topics that have been talked about so far is displayed in the form of icons and short texts.

[0305] The dialogue support information 810 is displayed at a position that is easy to enter the field of view, such as on the right side of the presentation screen 800.

[0306] The dialogue support information 810 includes hints for the actions that the interviewer should take next.

[0307] For example, when the interviewer's speaking time is too long, an alert prompting the candidate to answer a question is displayed as the dialogue support information 810.

[0308] Also, at the bottom of the dialogue support information 810, the recommended appeal items determined by the information processing system 100 are displayed in a list format.

[0309] At the top of the list, the most recommended appeal item, which is the interviewer's area of expertise and is presumed to have a good response from the candidate, is arranged.

[0310] Next to each recommended appeal item, an execution button for selecting that item is provided.

[0311] When the interviewer taps or clicks the execute button, a detailed persuasive text 530 related to that item will be displayed.

[0312] The 530 persuasive statements presented are printed in a large font size, designed to be easily readable at a glance.

[0313] Furthermore, the presentation screen 800 may display an indicator that visually shows the candidate's response.

[0314] The indicator lights up green if the candidate is showing a positive response, and changes to yellow or red if they are showing concern.

[0315] This allows the interviewer to intuitively grasp whether the current conversation is going well.

[0316] When the information processing system estimates a new response from the dialogue log 400, the content displayed on the presentation screen 800 is updated asynchronously in real time without a screen transition.

[0317] Therefore, the interviewer does not need to manually refresh the screen.

[0318] The latest dialogue support information (810) and recommended appeal items are continuously provided without disrupting the interview process.

[0319] This UI configuration allows interviewers to easily utilize the system's support information, even during the cognitively demanding task of conducting an interview.

[0320] Next, with reference to Figure 4, the feedback learning process based on decision-making results in the information processing system according to this embodiment will be described in detail.

[0321] Figure 4 is a conceptual diagram of the process of updating the learning model using result information as a reward.

[0322] In this system, after the entire recruitment process is completed, result information 220 regarding the candidate's final decision is obtained.

[0323] The acquired result information 220 serves as important feedback data indicating the extent to which the persuasive actions 210 presented through the interviewer's terminal 130 were effective.

[0324] The outcome of the appeal actions 210, determined based on the candidate context 200, is reflected in the final success or failure of the hiring process.

[0325] Therefore, a mechanism for feeding this result information 220 back into the learning model 300 is essential for optimizing the content and presentation order of the appeal action 210.

[0326] In this embodiment, result information 220 from interviews and recruitment activities is incorporated as compensation 310.

[0327] This learning using reward 310 is performed based on a so-called reinforcement learning approach.

[0328] In the reinforcement learning approach, the candidate context 200 is observed as a state.

[0329] Then, the interviewer selects an appeal action 210 to be presented on the interviewer's terminal 130 as their action.

[0330] For each selected appeal action 210, result information 220, which represents the candidate's decision-making, is observed.

[0331] As shown in Figure 4, this observed result information 220 is defined as the reward 310 and used to update the learning model 300.

[0332] The learning model 300 is updated to select appeal actions 210 that will yield a higher reward 310 for similar candidate contexts in the future 200.

[0333] In this way, we continuously improve our strategies using triplicate data consisting of candidate context 200, appeal action 210, and reward 310.

[0334] The information processing system 100 continuously stores newly acquired result information 220 in its database.

[0335] The accumulated log data is used as input for the retraining process of the learning model 300.

[0336] The impact of appeal action 210 on candidates is captured as a clear signal in the form of acceptance or rejection of the job offer.

[0337] If the result information 220 indicates acceptance of the job offer, the information processing system 100 awards a positive reward 310.

[0338] By providing a positive reward 310, the learning model 300 updates its evaluation of the appeal action 210 in relation to the candidate context 200 to a higher level.

[0339] This means that in subsequent interviews, if a candidate with similar attributes or preferences appears, the same or similar appeal action 210 will be recommended preferentially.

[0340] On the other hand, if the result information 220 indicates that the job offer was declined, the information processing system 100 needs to relatively lower the evaluation of the appeal action 210.

[0341] However, simply giving a negative reward 310 makes it difficult to identify which part of the appeal action 210 was inappropriate.

[0342] Therefore, in this embodiment, in addition to the binary result of accepting or declining the job offer, detailed information regarding the background of the refusal is incorporated as training data.

[0343] Specifically, as shown in Figure 2, the results information 220 includes categories of reasons for a candidate's refusal of a job offer.

[0344] The categories of reasons for declining an offer are obtained by interviewers or HR personnel entering them into the information processing system 100 via the interviewer terminal 130 or the like.

[0345] Alternatively, the categories of reasons for declining the position could be automatically extracted from the responses of a questionnaire administered to the candidates.

[0346] Reasons for declining an offer may include, for example, dissatisfaction with salary and benefits.

[0347] Additionally, discrepancies in terms of work location or working conditions can sometimes be cited as a reason for declining an offer.

[0348] Furthermore, a discrepancy between one's own career plan and growth opportunities and the environment offered by the company can also be a reason for declining an offer.

[0349] A mismatch with the company's organizational culture is also a significant reason for declining a job offer.

[0350] The information processing system 100 analyzes the acquired categories of reasons for refusal as training information.

[0351] For example, if a candidate declines an offer due to salary concerns, and the salary-related appeal action 210 was not adequately presented during the interview, the system adjusts the learning model 300.

[0352] In subsequent attempts, the system will learn to present attractive compensation and benefits to candidates with similar candidate contexts (200) at an earlier stage.

[0353] Furthermore, in the case of candidates who declined due to a mismatch with the company culture, it is possible that concerns about the culture were not detected in the candidate context 200, such as pre-employment aptitude tests.

[0354] In this case, the system optimizes the model to incorporate specific rhetorical devices and factual data as appeal actions 210 to alleviate cultural anxieties.

[0355] By using the categories of reasons for refusal as learning teacher information, it becomes possible to build a sophisticated appeal strategy that goes beyond simply improving the acceptance rate.

[0356] Learning Model 300 understands what concerns candidates and what factors led them to withdraw their application.

[0357] This leads to a presentation order in which appeal actions 210 are presented that proactively addresses potential concerns of candidates.

[0358] For example, let's say learning model 300 discovers that candidates with specific professional skills tend to decline offers due to a lack of growth opportunities.

[0359] Subsequently, when selecting new candidates with similar professional skills, the learning model 300 prioritizes presenting appealing topics related to growth opportunities and training programs to the interviewer terminal 130.

[0360] This type of optimization helps prevent missing out on top candidates and improves the hiring success rate.

[0361] Furthermore, the categories of reasons for declining an offer also serve as data to visualize the recruitment challenges faced by companies.

[0362] If there is a concentration of refusals in a particular appeal category, the system will prompt you to increase the variations of the appeal actions 210 related to that category.

[0363] Based on trends in reasons for refusal, the appeal optimization server 110 can generate new appeal items or review existing factual data.

[0364] The reason for withdrawal category, as result information 220, contributes not only to personalization for individual candidates but also to improving the overall appeal template.

[0365] Feedback learning using result information 220, which includes reasons for withdrawal, may be carried out sequentially using online learning methods.

[0366] Alternatively, the learning model 300 could be updated in a batch process once a certain period of recruitment activity data has been accumulated.

[0367] When using online learning, a failure in one interview will immediately influence the decision on persuasive action 210 for the next interview.

[0368] This makes it possible to quickly adapt to changes in the recruitment market and shifts in the trend of candidates declining offers due to the actions of competitors.

[0369] On the other hand, when using batch learning, the learning model 300 can be updated based on more statistically stable trends.

[0370] System administrators can combine and apply these two update methods as needed.

[0371] Furthermore, control may be implemented within the feedback loop to select exploratory appeal actions 210 at a certain rate.

[0372] By taking exploratory actions, we can measure the effectiveness of new combinations of appeal items and candidate contexts that have not been tested before.

[0373] This prevents the learning model 300 from getting stuck in a local optimal solution.

[0374] In order to discover unknown success patterns, the system may intentionally present different appeal sequences to the interviewer's terminal 130.

[0375] The proportion of exploration can be dynamically adjusted according to the importance and timing of the recruitment activity.

[0376] For example, in the initial stages of recruitment, a high proportion of exploration is set to collect diverse result information.

[0377] As the recruitment process nears its end, the focus shifts to practical application, and 210 persuasive actions are selected to ensure acceptance.

[0378] In this way, by combining the design of the reward 310 based on the result information 220 with the search control, the optimization of the appeal can be effectively carried out.

[0379] Next, with reference to Figure 10, we will explain the extension of long-term rewards in the learning model 300 of this system.

[0380] Figure 10 shows a diagram illustrating a system in which retention rates and evaluation indicators after joining the company are fed back into learning as result information.

[0381] The outcome information 220 regarding acceptance or rejection of job offers, as explained above, corresponds to short-term recruitment results.

[0382] By using short-term hiring results as a reward (310), the system learns persuasive actions (210) that reliably increase the desire to join the company.

[0383] However, the true purpose of recruitment is not simply to get candidates to join the company, but to ensure that candidates settle in and thrive within the organization after joining.

[0384] Making overly attractive appeals or appeals that are detached from reality may increase the short-term acceptance rate of job offers.

[0385] On the other hand, such appeals may cause a reality shock after joining the company and could lead to early resignation.

[0386] Therefore, in this embodiment, the post-employment indicator 900 is used not only as short-term recruitment results, but also as long-term result information 220.

[0387] The Post-Hiring Indicators 900 include various data that show the behavior and status of candidates after they actually join the company.

[0388] A typical example of the post-hire indicator 900 is the retention rate 910.

[0389] The retention rate of 910 is an indicator that shows whether or not a candidate is still employed after a certain period of time has elapsed since joining the company.

[0390] For example, the retention rate of an employee after six months or one year of employment is recorded as 910.

[0391] In addition to the retention rate (910), the results of post-hire performance evaluations and personnel assessments may also be included in the post-hire indicators (900).

[0392] Furthermore, engagement scores and employee satisfaction obtained from internal surveys, etc., may also be used as evaluation indicators.

[0393] The information processing system 100 (see Figure 1) retrieves these post-employment indicators 900 and links them to the ID of the target candidate.

[0394] The acquired post-hire metrics (900) are associated with the appeal actions (210) previously performed on the candidate, in accordance with the data structure shown in Figure 2.

[0395] Then, a new compensation of 310 is calculated based on the post-employment indicator of 900.

[0396] For example, if a candidate remains with the company for a long period of time and receives high evaluations, the information processing system 100 provides the learning model 300 with an additional positive reward 310.

[0397] Conversely, if a candidate leaves the position early or receives a significantly low evaluation, the information processing system 100 will impose a negative reward of 310.

[0398] As a result, the learning model 300 learns persuasive actions 210 that not only get candidates to accept job offers but also increase the retention rate 910 after they join the company.

[0399] By incorporating long-term rewards (310), the value of persuasive actions (210) that involve deliberately conveying harsh realities and challenges during interviews will be properly evaluated.

[0400] A sincere appeal that reflects the reality may slightly decrease the short-term acceptance rate, but it will significantly improve the long-term retention rate.

[0401] Based on the concept of reward 310 shown in Figure 4, the learning model 300 automatically optimizes trade-offs to maximize such long-term outcomes.

[0402] As shown in Figure 10, the appeal optimization server 110 periodically receives post-hire indicators 900 from the human resources information system, etc.

[0403] The received post-hire performance indicators 900 are stored in the database as part of the results information 220.

[0404] The information processing system 100 includes a function that calculates an integrated reward by combining a short-term reward 310 and a long-term reward 310.

[0405] Integrated compensation is defined, for example, as a weighted sum of immediate compensation upon acceptance of a job offer and delayed compensation received six months after joining the company.

[0406] The weighting of each reward of 310 is determined according to the company's hiring policy and business plan.

[0407] If securing a sufficient number of people quickly is an urgent matter, the weight of the short-term reward of 310 will be increased.

[0408] On the other hand, if improving employee retention is a management challenge, the weight of compensation (310) based on long-term retention rate (910) should be increased.

[0409] By flexibly modifying the design of reward 310 in this way, the behavior of the learning model 300 can be adapted to the company's strategy.

[0410] Updating the learning model 300 using long-term results information 220 has the advantage of consistently handling data from recruitment to post-employment.

[0411] Candidate Context 200 includes not only pre-employment attribute information but also evaluation data from the selection process.

[0412] By linking the history of appeal actions 210 with post-hire indicators 900, it becomes clear what kind of appeals to which candidates with what attributes will increase the retention rate 910.

[0413] For example, it has been learned that emphasizing the degree of autonomy given to candidates who aspire to work autonomously within an organization leads to higher engagement after joining the company.

[0414] These learning results will directly influence the priority of recommended topics presented on the interviewer's terminal 130 in subsequent interviews.

[0415] The interviewer's terminal 130 will now prioritize displaying appealing topics that have a high probability of retention after joining the company, as well as a high probability of short-term acceptance.

[0416] In learning using long-term rewards, it is necessary to address the time delay until reward 310 is received.

[0417] It takes several months to several years from the end of the recruitment process to achieve a post-hire score of 900.

[0418] To address this delayed reward problem, the system may employ a method that interpolates the reward 310 using an intermediate metric.

[0419] For example, the response rate to questionnaires during the pre-employment period or performance during onboarding training immediately after joining the company could be used as interim compensation.

[0420] By utilizing intermediate rewards, it becomes possible to partially update the learning model 300 even before the final retention rate of 910 is determined.

[0421] Alternatively, a proxy model could be constructed that predicts the 900 post-employment indicators in advance using a vast amount of accumulated historical data.

[0422] By using the retention probability predicted by the surrogate model as a pseudo-reward 310, the learning model 300 can be updated at a near real-time speed.

[0423] In this way, the mechanisms shown in Figures 4 and 10 have the effect of improving the quality of appeals from both short-term and long-term perspectives.

[0424] Feedback learning based on result information 220 transforms the persuasion process, which previously relied on the interviewer's subjectivity and experience, into a data-driven one.

[0425] The information displayed on the interviewer's terminal 130 is always based on the latest recruitment results and data on employee performance after joining the company.

[0426] This allows companies to conduct truly valuable recruitment activities, regardless of the interviewer's skills.

[0427] The update process for the learning model 300 is automated by a scheduler that runs periodically.

[0428] When the server detects newly available result information 220, it starts a retraining pipeline in the background.

[0429] The pipeline involves feature extraction, reward calculation (310), model parameter updates, and evaluation of the updated model.

[0430] Deployment to the production environment will only occur if it is confirmed that the updated training model 300 outperforms the previous model.

[0431] This automated cycle allows the system to improve the accuracy of its appeals the longer it continues to operate.

[0432] Learning based on the categories of reasons for declining job offers is particularly effective for positions with high recruitment difficulty.

[0433] Candidates with rare skills often receive job offers from multiple companies, resulting in a variety of complex reasons for declining offers.

[0434] The system processes these complex categories of reasons for declining as multidimensional training data, deriving appeal actions 210 that pinpoint and address each candidate's concerns.

[0435] Furthermore, optimization using 900 post-hire indicators, including retention rate 910, functions as a mechanism to structurally prevent recruitment mismatches.

[0436] By using post-hire performance evaluations of candidates as training data, we can curb excessive appeals to candidates who are attractive during interviews but actually don't fit the company culture.

[0437] Conversely, for candidates who are modest in their interview presentations but are highly valued after joining the company, personalized appeals are strengthened to ensure a more certain acceptance.

[0438] Optimization from a long-term perspective yields concrete results such as reduced recruitment costs and improved organizational productivity.

[0439] The configurations in Figures 4 and 10 connect the entire process, from the start of recruitment activities to success after joining the company, using information processing.

[0440] This connection integrates previously fragmented recruitment and HR data, creating a comprehensive data foundation for learning.

[0441] The correspondence between appeal action 210 and result information 220 can also be visualized on the dashboard as anonymized statistical information.

[0442] However, the core value of this system lies not merely in visualization, but in directly feeding that knowledge back into the learning model 300 and automatically applying it to subsequent decision-making.

[0443] Without having to perform complex data analysis, interviewers can execute data-driven, optimal appeals simply by following the instructions on the interviewer terminal 130.

[0444] Optimization techniques such as reinforcement learning and bandit algorithms provide the mathematical basis for enabling this automated application.

[0445] By setting the objective function to maximize the reward of 310, the system autonomously continues to search for the optimal policy.

[0446] Categorizing reasons for refusal and introducing the retention rate 910 significantly extends this objective function to better reflect the realities of the recruitment domain.

[0447] This framework allows for the simultaneous pursuit of both the short-term goal of accepting a job offer and the long-term goal of succeeding after joining the company.

[0448] As described above, the system according to this embodiment includes a continuous feedback loop based on the result information 220.

[0449] From short-term reasons for declining a position to long-term post-hire indicators (900), a diverse range of outcome information (220) is reflected in the learning model (300) as rewards (310).

[0450] This makes it possible to present 210 appeal actions that address the true needs of candidates while securing and retaining the best talent for the company.

[0451] As the operational period lengthens, the learning model 300 absorbs more result information 220, continuously improving the accuracy of its recommendations (see Figure 4).

[0452] Ultimately, the information provided via the interviewer terminal 130 becomes a crucial support tool that fundamentally enhances a company's recruitment capabilities (see Figure 9).

[0453] The 210 personalized appeal actions tailored to each candidate effectively guide their decision-making.

[0454] As a result, early departures due to mismatches after joining the company decrease, and the retention rate improves.

[0455] The tripartite learning loop of candidate context 200, appeal actions 210, and result information 220 elevates the entire recruitment process into a scientific and efficient one (see Figure 2).

[0456] Updating the learning model 300 to incorporate the 900 post-hire indicators opens up new possibilities for information processing systems in the human resources field (see Figure 10).

[0457] System administrators can flexibly respond to changes in the market environment by regularly reviewing the weighting of the result information 220.

[0458] For example, during periods of increased labor market fluidity, it may be possible to temporarily increase the weight of the compensation (310) related to retention rate (910).

[0459] This ensures that the learning model 300 remains optimized for the latest recruitment challenges that companies face.

[0460] The above is a detailed explanation of feedback learning based on decision-making outcomes and the extension of long-term rewards.

[0461] The types and number of result information 220 obtained, as well as the specific algorithm of the learning model 300, can be appropriately modified depending on the embodiment.

[0462] Further subdivision of rejection reason categories and the addition of 900 new post-hire indicators can be easily handled by modifying the system design.

[0463] The continuous learning and improvement cycle functions as a crucial core feature of this system.

[0464] Technical means are provided to achieve data-driven optimization in the communication domain of interviews, which tends to be highly subjective (see Figures 5, 6, and 7).

[0465] Figure 8 is a flowchart showing the learning constraints for excluding specific attributes and the steps for outputting the audit log 710.

[0466] Ensuring fair recommendations in the information processing system 100 is extremely important (see Figure 1).

[0467] The appeal optimization server 110 determines the appeal action 210 based on the candidate context 200 (see Figure 3).

[0468] In this decision-making process, it is necessary to prevent the candidates' sensitive attributes (700) from having an undue influence.

[0469] Sensitive Attribute 700 refers to personal attributes unrelated to a candidate's abilities or suitability.

[0470] Specifically, gender and age are included in the 700 sensitive attributes.

[0471] Furthermore, nationality and place of origin may also be defined as sensitive attributes (700).

[0472] Information such as religion and beliefs may also fall under the category of sensitive attribute 700.

[0473] When these 700 sensitive attributes are input into the learning model 300, the recommended results may become biased.

[0474] This is because unconscious biases present in past hiring data may be reproduced by the learning model 300.

[0475] Therefore, the information processing system 100 imposes a constraint that excludes sensitive attributes 700 from the learning input.

[0476] The appeal optimization server 110 performs attribute determination processing when it obtains the candidate context 200.

[0477] Various information elements included in the candidate context 200 are checked to determine whether or not they fall under the sensitive attribute 700.

[0478] This inspection is performed using predefined rule-based filtering.

[0479] Alternatively, it is possible to extract sensitive attributes 700 from text data using natural language processing techniques.

[0480] The 700 extracted sensitive attributes will be processed to disable them.

[0481] The invalidation process replaces the value of sensitive attribute 700 with an unidentifiable dummy string.

[0482] Alternatively, the Sensitive Attribute 700 field itself may be completely removed from the data structure.

[0483] This physically removes the sensitive attribute 700 from the input features used by the learning model 300 when it performs inference.

[0484] The algorithm guarantees that sensitive attributes 700 will not influence the decision on the appeal action 210.

[0485] This constraint is consistently applied during both model training and inference.

[0486] During training, sensitive attributes 700 are removed from the candidate context 200 used as training data.

[0487] Similarly, sensitive attributes 700 are excluded from the candidate context 200, which is entered in real time during inference.

[0488] Simply excluding sensitive attribute 700 from the input may not be sufficient in all cases.

[0489] This is because other non-sensitive attributes may function as substitute variables for sensitive attribute 700.

[0490] For example, this applies to cases where a particular educational background or residential area has a strong correlation with a specific gender or age.

[0491] Algorithmic mechanisms are introduced to mitigate such indirect biases through substitute variables.

[0492] The appeal optimization server 110 performs debiasing in the feature representation space of the learning model 300.

[0493] Specifically, the learning process uses an adversarial learning framework to train the model so that it cannot reconstruct information about the 700 sensitive attributes from the features.

[0494] As a result, the internal representation acquired by the learning model 300 becomes independent of the sensitive attribute 700.

[0495] As a result, the bias in the recommended outcome of the 210 appeal actions determined is effectively suppressed.

[0496] Furthermore, the fairness indicators for the output results of the learning model 300 will be continuously evaluated.

[0497] Statistical fairness indicators such as demographic parity and equivalent odds are calculated during the reasoning process.

[0498] The distribution of recommended appeal actions 210 across specific attribute groups is monitored for any significant differences.

[0499] If the difference exceeds a predetermined threshold, retraining of the learning model 300 is triggered.

[0500] These algorithmic mechanisms enable the information processing system 100 to maintain a high level of overall fairness (see Figure 8).

[0501] Audit log 710 is generated to retrospectively prove that impartial recommendations are ensured.

[0502] The primary purpose of Audit Log 710 is to enable post-event verification in system operations.

[0503] The appeal optimization server 110 records a detailed operation history as an audit log 710 at each step of information processing.

[0504] The date and time the candidate context 200 was acquired, along with its data content, is recorded in audit log 710.

[0505] The fact that sensitive attribute 700 was identified and exclusion processing was applied is also clearly stated in audit log 710.

[0506] The algorithmic basis for determining which items were classified as sensitive attribute 700 is also saved.

[0507] The audit log 710 allows us to track that the clean input data after the exclusion process has been passed to the learning model 300.

[0508] The candidate appeal actions 210 output by the learning model 300, along with their confidence levels, will also be recorded.

[0509] The content of the appeal action 210 presented to the interviewer's terminal 130 is ultimately saved in the audit log 710.

[0510] Furthermore, the result information 220 obtained as a result of executing the appeal action 210 is also linked to the audit log 710.

[0511] This makes the entire process, from input to output and results, fully visible and allows for post-mortem verification.

[0512] The generated audit logs 710 are sequentially saved to a secure storage area that is difficult to tamper with.

[0513] Each record in audit log 710 is assigned a hash value based on cryptographic technology.

[0514] By chaining the hash values ​​of preceding and succeeding records, the continuity and completeness of the entire data set are strongly guaranteed.

[0515] During post-event verification, a process is executed in which audit log 710 is output from the appeal optimization server 110.

[0516] This output process can only be executed by specific system administrators or auditors who have been granted the necessary permissions.

[0517] The output audit log 710 is converted into a human-readable format or a format that can be read by a dedicated analysis tool.

[0518] The auditors analyze the generated audit log 710 to check if the system contains any unexpected biases.

[0519] It is also possible to investigate individual cases to determine whether an unfair appeal action 210 has been presented to a specific candidate.

[0520] The operation history on the interviewer's terminal 130 is also managed in synchronization with the audit log 710.

[0521] The interviewer's decision not to adopt any of the 210 recommended persuasive actions will also be recorded in the log.

[0522] It is possible to analyze separately the bias in appeals based on the interviewer's own judgment and the bias in appeals based on system recommendations.

[0523] This makes it easier to determine whether the problem is due to the algorithm or to human operation.

[0524] The process of generating and outputting these audit logs 710 strongly supports corporate compliance.

[0525] As shown in Figure 8, the process of excluding specific attributes is carried out through several further steps.

[0526] First, the appeal optimization server 110 receives candidate context 200 from an external database.

[0527] The received candidate context 200 is expanded into a secure temporary memory area.

[0528] Next, the attribute extraction module begins analyzing the candidate context 200.

[0529] The attribute extraction module compares the data with a pre-configured dictionary of 700 sensitive attributes.

[0530] The judgment dictionary includes a comprehensive list of characteristic keywords and related data item names that indicate gender and age.

[0531] If a matching item is found as a result of the matching process, its location information is temporarily recorded.

[0532] Next, the invalidation processing module is activated and replaces the value of the identified sensitive attribute 700.

[0533] The original data, prior to its invalidation, is retained as a temporary log for auditing purposes and is strictly isolated from subsequent processing data.

[0534] At this stage, the audit log 710 generated stores the original value and the fact that it was invalidated, all encrypted.

[0535] Encryption often employs public-key cryptography, which involves strict management of the decryption key.

[0536] This makes it possible to access the original sensitive attribute 700 information only in the event of a legal audit.

[0537] This protected information is never referenced during normal operation or during the inference process of the learning model 300.

[0538] Only candidate contexts 200 that are in a safe state after the invalidation process has been completed are sent to the next inference step.

[0539] The learning model 300 receives this safe candidate context 200 and begins the process of calculating the appeal action 210.

[0540] The generation of audit logs 710 continues uninterrupted even during the inference process within the learning model 300.

[0541] The version information of the 300 learning models used for inference and the identifiers of the parameter sets used are recorded.

[0542] The internal score value of each appeal action 210 calculated as a result of the inference is written to the audit log 710.

[0543] Once the final appeal action 210 is determined, the system generates metadata explaining the reasoning behind that decision.

[0544] This metadata includes which non-sensitive features strongly contributed to the decision.

[0545] This information regarding contributions is also an important component of the audit log 710, facilitating post-mortem verification.

[0546] Multiple complementary approaches are applied in combination to the algorithmic mechanisms that mitigate bias in recommendation results.

[0547] One complementary approach is dynamic resampling of the training data.

[0548] Before training the learning model 300, statistically adjust for imbalances in the number of data points between specific attribute groups.

[0549] Techniques are employed to either oversample data from minority attribute groups or undersample data from the majority.

[0550] This prevents the learning model 300 from overfitting to the majority trend and making biased judgments.

[0551] Another approach involves introducing a regularization term into the loss function of the learning model 300.

[0552] A term is added to the loss function that penalizes inference results that compromise fairness.

[0553] For example, the system is designed so that losses increase if there is a large difference in the prediction error of the result information 220 between different attribute groups.

[0554] Furthermore, an approach involving group-specific adjustment of score thresholds through post-processing is also employed.

[0555] The threshold for determining the appeal action 210 is optimized based on the score output by the learning model 300.

[0556] This ensures that the distribution of recommended results ultimately presented to the interviewer's terminal 130 meets fair standards.

[0557] These algorithmic mechanisms can be flexibly enabled or disabled depending on the system settings.

[0558] The audit log 710 also records the settings of which approach was applied and how the inference was performed (see Figure 8).

[0559] The appeal optimization server 110 periodically performs self-diagnostic tests to verify that these algorithms are functioning correctly.

[0560] In the self-assessment test, dummy candidate contexts 200, which include known biases, are automatically entered into the system.

[0561] The system automatically evaluates whether it can correctly output unbiased appeal actions 210 even with this dummy data.

[0562] The execution history of this self-assessment and the calculated evaluation score are also securely stored as part of the audit log 710.

[0563] If the test results do not meet the specified criteria, an alert will be immediately sent to the system administrator.

[0564] The history of alert occurrences and the administrator's response to them are also recorded as evidence of the system's health.

[0565] The generated audit log 710 will be used for multifaceted post-event verification within the company.

[0566] The HR department regularly compiles the audit log 710 and objectively analyzes the overall trends in the recommendation results.

[0567] A mechanism is provided that automatically generates monthly or quarterly governance reports from audit log 710.

[0568] The report visualizes and includes the number of times each of the 210 appeal actions was presented and the estimated attribute distribution of candidates.

[0569] If it is discovered that a particular appeal action 210 is unintentionally concentrated on a specific candidate group, corrective action will be taken.

[0570] The system administrator will drill down into the relevant section of audit log 710 to investigate the cause in detail.

[0571] The specific features of the 300 learning models that caused the bias will be identified.

[0572] If necessary, the retraining parameters of the learning model 300 are adjusted, and additional sensitive attributes 700 to be excluded are specified (see Figure 8).

[0573] Furthermore, audit log 710 serves as extremely important verification material during system audits conducted by external third-party organizations.

[0574] It is possible to demonstrate, based on objective data, that the system's operation complies with relevant laws and guidelines.

[0575] The auditing body will verify in advance, using cryptographic proof, that the audit log 710 has not been tampered with.

[0576] Then, the inference process from input to output is randomly sampled and examined in detail.

[0577] The information processing system 100 (see Figure 1) provides a level of processing transparency that can withstand such rigorous external audits.

[0578] Furthermore, it becomes possible to provide logical explanations based on audit log 710 in response to inquiries and requests for information disclosure from the candidates themselves.

[0579] We can clearly present the algorithmic rationale for why a specific appeal action 210 was presented.

[0580] However, when disclosing the system to candidates, parts related to the system's confidential algorithms will be appropriately masked.

[0581] The process of generating this disclosure data itself will also be recorded in audit log 710 as it will be subject to a later audit.

[0582] Strict governance is also applied when updating the learning model 300 using the result information 220 as the reward 310 (see Figure 4).

[0583] The control program works to ensure that the sensitive attribute 700 does not indirectly affect the calculation process of reward 310 (see Figure 10).

[0584] A clear approval process exists before deploying the updated, new learning model 300 to the production environment.

[0585] Shadow testing of the new learning model 300 is automatically performed using the past audit logs 710.

[0586] The appeal actions 210 (see Figures 2 and 3) that the new learning model 300 outputs for past candidate contexts 200 are evaluated.

[0587] Deployment will not be permitted unless data confirms that the fairness metrics have not deteriorated compared to the old model.

[0588] The interviewer's terminal 130 always receives only recommendations from the latest learning model 300 that has gone through this approval process.

[0589] The definition of sensitive attributes 700 may differ depending on the laws and regulations of the region to which the information processing system 100 applies.

[0590] Therefore, the appeal optimization server 110 has the function to dynamically switch constraint rules according to the applicable jurisdiction.

[0591] When settings are changed from the administration screen, the list of 700 sensitive attributes to be excluded is automatically updated.

[0592] The change history of this rule, along with metadata indicating who made what changes and when, is also recorded in audit log 710.

[0593] The strictest exclusion restrictions apply when data is moved across jurisdictions.

[0594] Items that may be considered Sensitive Attribute 700 in all relevant jurisdictions will be uniformly excluded from the data.

[0595] The processing steps related to this cross-border data transfer are also tracked in detail by audit log 710.

[0596] Audit Log 710 will be operated under appropriate lifecycle management.

[0597] Audit logs 710, once their retention period has expired, are automatically anonymized or deleted in a secure manner.

[0598] The fact that deletion or anonymization was performed is also recorded as a simplified system log.

[0599] Measures have also been taken to ensure fairness in the display control of the interviewer's terminal 130.

[0600] The display data transmitted from the appeal optimization server 110 to the interviewer terminal 130 does not contain any sensitive attribute 700.

[0601] Suppose the interviewer enters a phrase corresponding to sensitive attribute 700 into the memo field of the interviewer's terminal 130 during their conversation with the candidate.

[0602] The text analysis function of the interviewer's terminal 130 immediately detects the relevant text.

[0603] A warning alert is displayed on the interviewer's terminal 130 screen, urging them to refrain from making judgments based on sensitive attributes 700.

[0604] The display history of this warning alert is also sent to and recorded in audit log 710 as part of the interviewer's activity log.

[0605] This strengthens governance not only over the system's automatically generated recommendations but also over the interviewer's subjective evaluation process.

[0606] The system helps ensure that a consistent sense of fairness is ingrained in interviewers throughout the entire recruitment process.

[0607] The various audit logs (710) are ultimately centrally aggregated as foundational data for corporate compliance management.

[0608] Furthermore, please note that unless the word "only" is used, such as in "based solely on candidate context 200," it is assumed that additional information may also be considered in this specification.

[0609] Also, please note that the statement "Exclude if sensitive attribute 700 is included" does not necessarily mean "Always exclude if included" unless explicitly stated otherwise.

[0610] Furthermore, the statement "output audit log 710" does not limit the output to only one specific data format.

[0611] Furthermore, for the sake of clarity, even if there are aspects of operation that differ from the operation described herein in some way, program, terminal, device, server, or system (hereinafter referred to as "method, etc."), each aspect of the present invention is intended to cover the same operation as any of the operations described herein.

[0612] It should be added that the existence of operations different from those described herein does not mean that such methods, etc., fall outside the scope of each aspect of the present invention. [Explanation of Symbols]

[0613] 100 Information Processing Systems 110 Appeal Optimization Server 130 Interviewer terminal 200 Candidate Context 210 Appeal Actions 220 Results information 300 Learning Models 310 Rewards 400 Dialogue Logs 410 Reaction estimation unit 500 Knowledge Bases 510 Fact Data 530 Appeal text 600 Interviewer Models 610 Recommended Topics 700 Sensitive Attributes 710 Audit Log 800 presentation screen 810 Dialogue Support Information 900 Post-employment performance indicators 910 Retention rate AI520 element AI520 DB120 Element DB120

Claims

1. An information processing device for optimizing the presentation of appealing information in recruitment activities, Means for obtaining candidate context, which includes information about the candidate's attributes or behavioral history, A preprocessing means for excluding or masking predetermined sensitive attributes, including the candidate's gender or age, included in the candidate context, from the training input for the learning model and the inference input for determining appeal actions, A means for selecting recommended appeal information from a list of appeal information candidates to be presented to the candidate based on the candidate context after preprocessing by the preprocessing means, and determining an appeal action including the content and presentation order of the appeal information, A means for presenting the determined appeal action to the terminal, A means for acquiring result information regarding the candidate's decision-making as a result of executing the appeal action, calculating a reward value based on at least one of the information regarding acceptance or rejection of the job offer and indicators regarding retention or evaluation after joining the company included in the result information, acquiring the category of reason for rejection included in the result information as auxiliary teacher information separate from the reward value if the candidate rejects the offer, learning the relationship between the pre-processed candidate context, the appeal action, and the result information using the reward value and the auxiliary teacher information, and updating the learning model used to determine the selection or presentation order of the appeal information in subsequent appeal actions, A means for outputting an audit log that associates the content of exclusion or masking by the preprocessing means, the appeal action presented to the terminal, and the result information, Information processing device including

2. The means for determining the appeal action estimates the candidate's response to the presented appeal information based on a dialogue log transcribed from the audio of the conversation with the candidate, updates the next appeal item to be presented in real time, and presents it to the terminal. The information processing apparatus according to claim 1.

3. The means for updating the learning model performs learning using a weighted sum of an immediate reward for accepting the job offer and a delayed reward based on an indicator for retention or evaluation after joining the company as the reward value. The information processing apparatus according to claim 1 or 2.

4. The means for updating the learning model updates the weight of appeal items or presentation order corresponding to the reason for refusal based on the category of the reason for refusal obtained as the auxiliary teacher information, and optimizes the appeal action for similar candidates in subsequent attempts. The information processing apparatus according to claim 1.

5. The means for determining the appeal action includes maintaining factual data that can be appealed to the candidate as a knowledge base, selecting appeal items according to the candidate context after preprocessing using a statistical model, and generating text based on the factual data corresponding to the selected appeal items using a generative AI. The information processing apparatus according to claim 4.

6. The system models and maintains the strengths and weaknesses of each interviewer in persuasive communication, and based on these strengths, prioritizes presenting persuasive topics or communication styles that the interviewer is capable of performing to the terminal. The information processing apparatus according to claim 4 or 5.

7. The audit log includes items excluded or masked by the preprocessing means, the version of the learning model, the score of the candidate appeal action, the appeal action presented to the terminal, and the result information corresponding to the appeal action, The information processing apparatus according to claim 1.

8. A program that causes a computer to perform processing to optimize the presentation of appealing information in recruitment activities, wherein the processing is: A step of obtaining candidate context, which includes information about the candidate's attributes or behavioral history, A preprocessing step of excluding or masking predetermined sensitive attributes, including the candidate's gender or age, included in the candidate context, from the training input for the learning model and the inference input for determining appeal actions, Based on the candidate context after preprocessing by the aforementioned preprocessing step, the step of selecting recommended appeal information from the candidates of appeal information to be presented to the candidate, and determining an appeal action including the content and order of presentation of the appeal information, The steps include presenting the determined appeal action to the terminal, The steps include: obtaining result information regarding the candidate's decision-making as a result of executing the appeal action; calculating a reward value based on at least one of the information regarding acceptance or rejection of the job offer and indicators regarding retention or evaluation after joining the company included in the result information; obtaining the category of reason for rejection included in the result information as auxiliary teacher information separate from the reward value if the candidate rejects the offer; learning the relationship between the pre-processed candidate context, the appeal action, and the result information using the reward value and the auxiliary teacher information; and updating the learning model used to determine the selection or presentation order of the appeal information in subsequent appeal actions. A step of outputting an audit log that associates the content of exclusion or masking performed in the preprocessing step, the appeal action presented to the terminal, and the result information, A program that includes this.

9. In the step of determining the appeal action, the system estimates the candidate's response to the presented appeal information based on the dialogue log with the candidate, updates the next appeal item to be presented in real time, and presents it to the terminal. The program according to claim 8.

10. In the step of updating the learning model, the learning is performed using a weighted sum of immediate rewards for accepting the job offer and delayed rewards based on indicators related to retention or evaluation after joining the company as the reward value. The program according to claim 8 or 9.

Images