JP7743098B2 - Recruitment page generator - Google Patents

Description

The present invention relates to a recruitment page generation device.

There are workplaces where working patterns are complex and the turnover of personnel is relatively high, such as in nursing and caregiving. In workplaces where the turnover of personnel is relatively high, recruitment is frequent, and the work involved in recruitment can become a burden on the workplace. Therefore, there is a need for means to support recruitment.

(Prior Art)
As an example of prior art related to supporting recruitment, Patent Document 1 discloses an employment support device that calculates the similarity between behavioral characteristic information (first behavioral characteristic information) generated by inputting aggregate information on feature amounts of email sending and receiving histories of employees of a company into a predetermined mathematical model, and behavioral characteristic information (second behavioral characteristic information) generated by inputting aggregate information on feature amounts of email sending and receiving histories of a job seeker into the above-mentioned predetermined mathematical model. The technology in Patent Document 1 allows job seekers to obtain information on suitable companies without any conscious effort, and companies can collect information on job seekers that is useful for making hiring or rejection decisions.

JP 2010-231685 A

In workplaces with complex working patterns, job posting conditions (job requirements) can vary widely, including employment type, working hours, and required skills. Creating a job posting page that reflects these diverse job requirements requires a high level of skill in creating the job posting page. In addition to the diverse job posting conditions, when labor turnover is relatively high and job postings are frequent, the high level of skill required can place a huge burden on advertising personnel, who have few replacement personnel. The technology in Patent Document 1 is only able to provide convenience to job seekers and companies through similarity calculations, and there is room for further improvement in terms of reducing the burden on advertising personnel.

The present invention was made in light of these circumstances, and its purpose is to reduce the burden on recruiters involved in creating recruitment pages by, for example, providing them with model recruitment pages.

After extensive research into solving the above-mentioned problems, the inventors discovered that the above-mentioned objectives could be achieved by generating job postings using a large-scale language model that had undergone machine learning using proofread job postings, etc. As a result, the inventors were able to complete the present invention. Specifically, the present invention provides the following:

The present invention provides a recruitment page generation device comprising: a conditions receiving unit that receives job requirements; a manuscript generation unit that provides input including the job requirements to a large-scale language model and causes the large-scale language model to generate a first job requirement that reflects the job requirements; a proofreading receiving unit that receives a proofread version of the first job requirement that has been proofread; and a machine learning unit that uses training data that includes the job requirements and the first job requirement as explanatory variables and the proofread version as a target variable to train the large-scale language model to generate a job requirement that reflects the job requirements, taking the job requirements as input.

If there were a device that could automatically generate recruitment pages, such a device could reduce the burden on recruiters by allowing them to use the generated recruitment pages as is or as a model for recruiters. One possible procedure for automatically generating recruitment pages is to generate a recruitment page by combining standard phrases corresponding to each individual condition included in the recruitment requirements. However, when creating a recruitment page that reflects a wide variety of recruitment requirements, a huge amount of standard phrases would need to be prepared in advance, which could actually increase the burden on recruiters.

A large-scale language model is a language model that has undergone pre-training using large amounts of text, including recruitment pages. Through this pre-training, the large-scale language model has the emergent ability to generate a recruitment page that reflects the job requirements when they are input. However, if an attempt is made to generate a recruitment page that reflects a variety of job requirements using only the above-mentioned emergent ability, there may be deficiencies in any of the various requirements, such as the match between the job requirements and the recruitment page, or the ability to appeal to job seekers, and the resulting recruitment page may not be exemplary.

The present invention receives a proofread version of the first job posting generated by the present invention in a proofread version receiving unit, and performs machine learning using training data that includes the job requirements and the first job posting as explanatory variables and the received proofread version as a target variable, thereby teaching the large-scale language model the capabilities that are insufficient with the emergent capabilities described above. As a result, the present invention can have the large-scale language model that has undergone machine learning generate a model recruitment page and provide it to recruiters. Therefore, the present invention can reduce the burden on recruiters involved in generating recruitment pages.

The present invention can reduce the burden on recruiters involved in creating recruitment pages by, for example, providing them with model recruitment pages.

FIG. 1 is a block diagram showing the hardware and software configurations of a system S according to this embodiment. FIG. 2 is an example of the job posting database 122 . FIG. 3 is a main flowchart showing an example of a preferred flow of the generation process executed by the generation device 1 of this embodiment. FIG. 4 is a continuation of FIG.

The following describes in detail an example embodiment of the present invention, with reference to the drawings.

<System S>
1 is a block diagram showing the hardware and software configurations of a system S according to this embodiment. An example of a preferred embodiment of the hardware and software configurations of the system S according to this embodiment will be described below with reference to FIG.

The system S is composed of a recruitment page generation device 1 and a terminal T configured to be able to communicate with the generation device 1 via a network N.

[Generation device 1]
The generating device 1 includes a control unit 11, a storage unit 12, and a communication unit 13. There are no particular limitations on the type of the generating device 1. Examples of the type include a server, a cloud server, and the like.

The generation device 1 is configured to perform a process of causing a large-scale language model that has been pre-trained using a large amount of text to perform machine learning, and a process of causing the large-scale language model that has undergone machine learning to generate a recruitment page (job posting). These processes are performed, for example, via an API related to the use of the large-scale language model. The following is a more detailed example of the configuration of the generation device 1.

[Control unit 11]
The control unit 11 includes a CPU (Central Processing Unit), a RAM (Random Access Memory), a ROM (Read Only Memory), and the like.

The control unit 11 cooperates with the memory unit 12 and/or communication unit 13 as necessary. The control unit 11 then implements the software components of the program of this embodiment executed by the generation device 1, such as the condition receiving unit 111, draft receiving unit 112, manuscript generation unit 113, proofread receiving unit 114, evaluation receiving unit 115, and machine learning unit 116. The functions provided by each of the software components of the program of this embodiment will be explained in the description of the preferred flow of the generation process described below.

[Storage unit 12]
The memory unit 12 is a device in which data and/or files are stored, and includes a storage unit that stores data non-temporarily using a hard disk, a semiconductor memory, a recording medium, a memory card, or the like.

The storage unit 12 may have a mechanism that enables connection to storage devices or storage systems such as NAS (Network Attached Storage), SAN (Storage Area Network), cloud storage, file servers, and/or distributed file systems via the network N.

The memory unit 12 stores programs executed by the microcomputer, a large-scale language model 121, a job posting database 122, etc.

(Large-scale language model 121)
The large-scale language model 121 (LLM) in this embodiment is a type of probabilistic model used in natural language processing, and is a model for probabilistically predicting how likely a given word or sentence is to occur in natural language. The generation device 1 provides a sentence of generation data (described later) as an input to the large-scale language model 121, predicts a sentence of an adoption page that is likely to occur following the sentence of the generation data, and generates the predicted adoption page sentence.

The large-scale language model 121 may be stored on an external cloud server or the like and executed via an API (Application Programming Interface) related to the use of the large-scale language model 121, or may be stored in the memory unit 12.

By implementing the large-scale language model 121 via an API, the generation device 1 has a simpler and more cost-effective configuration than a configuration in which the large-scale language model 121 is stored in the memory unit 12.

By storing the large-scale language model 121 in the storage unit 12, the generation device 1 can execute processing related to the large-scale language model 121 without having to communicate point-by-point with an external device such as a cloud server. This allows the generation device 1 to reduce processing delays and information security risks associated with communication. The following explanation focuses on the large-scale language model 121 being stored in the storage unit 12, but those skilled in the art will understand that a similar implementation can be implemented in which the large-scale language model 121 is executed via an API on an external cloud server or the like.

The large-scale language model 121 is trained on a large amount of text, at least 500 GB (gigabytes) or more. Examples of such models include ChatGPT, GPT-3.5, and GPT-4 from OpenAI (registered trademark), and LLaMa from Meta AI (registered trademark).

As the amount of training data increases, the performance of a language model increases in accordance with the broken neural scaling law (BNSL), and it is thought that the model acquires emergent capabilities when transitioning between segments in BNSL. Therefore, the large-scale language model 121 of this embodiment acquires the emergent capabilities, acquired according to the amount of text used in pre-training, to generate job postings that reflect the job posting conditions when they are input.

When counting tokens, which are semantic blocks of text (words, etc.), the large-scale language model 121 preferably has an upper limit of 30,000 or more contexts used to generate text. This allows the large-scale language model 121 to generate job postings that take into account, as context, job requirements written using a large number of tokens to include numerous conditions. This also allows the large-scale language model 121 to perform machine learning that takes into account, as context, draft text written using a large number of tokens.

The large-scale language model 121 is preferably trained on text containing 10 trillion or more tokens. This is believed to give the large-scale language model 121 the emergent ability to generate job postings that reflect nuances appropriate to the job type and other aspects of the workplace where the job is being posted.

The large-scale language model 121 is preferably a model with a large number of parameters, at least 10 billion or more. This allows the large-scale language model 121 to perform processing that fully reflects the emergent capabilities obtained from large amounts of text.

(Recruitment script database 122)
The job posting database 122 stores information that associates data used to generate job postings (generation data) with the generated job postings. This associated information includes, for example, information associating job posting requirements with a first job posting generated based on the job posting requirements, and information associating a draft of a job posting with a second job posting generated based on the draft. Job postings such as the first job posting and the second job posting are preferably stored in association with proofread versions of the job postings and improvement information such as evaluations of the job postings. This allows the generation device 1 to perform machine learning based on the improvement information.

The job posting database 122 is configured to extract job postings that meet specified conditions. This allows the generation device 1 to extract job postings that meet specified conditions, such as job posting conditions.

The job postings stored in the job posting database 122 are not particularly limited. These job postings include various types of advertisements, exemplified by, for example, text suitable for transmission via email, short message, or posting to social media, modified text (e.g., HTML (HyperText Markup Language)) that allows for greater expressiveness than text when transmitted via email, a combination of text and images (e.g., PDF (Portable Document Format), web pages, etc.), audio suitable for transmission via voice calls or voice messages, image suitable for image advertisements, and expressive video.

The job postings stored in the job posting database 122 preferably contain text. Including text in the job postings makes it easier to extract text based on the conditions of the text format generated by the large-scale language model 121. If the job posting contains multimedia information such as images, audio, or video, it is preferable that the job posting be associated with text indicating the content of the multimedia information generated by various recognition processes for the multimedia information, in order to make it easier to extract text based on the conditions of the text format.

Figure 2 is an example of a job posting database 122. In the example shown in Figure 2, the job posting identified by ID "A0001" reads, "We are looking for a registered nurse (new graduates and experienced candidates welcome). Educational background and experience are not required, and the position is 64 years old or younger. Childcare leave, family care leave, and nursing care leave are available, and work-life balance is emphasized. Working hours are 9:00-18:00 (60-minute break), and no overtime. Salary is △△ yen to △△ yen. Bonuses and allowances are generous. Skills can also be improved through extensive training. Why not put your skills to good use on our team? We look forward to your application." is used as data to generate job posting conditions: "1. Qualifications, etc. (1) Educational background and experience are not required (registered nurse required) (2) Age 64 or younger *Retirement age is 65 2. Working conditions (1) Childcare leave, family care leave, and nursing care leave are available (2) Working hours are 9:00-18:00 (60-minute break), and no overtime" (3) In response to "Salary: △△ million yen - △△ million yen," the revised draft as improvement information is linked to the following: "[Registered nurse wanted (day shift)] We are looking for a registered nurse. No educational background or experience required, eligible applicants are 64 years of age or younger. We have a track record of taking parental leave, caregiving leave, and nursing care leave, so you can balance work with family and childcare. Working hours (day shift only): 9:00-18:00 (60-minute break), no overtime. Salary: △△ million yen - △△ million yen. Generous bonuses and allowances. Why not show us your skills on our team? We look forward to receiving your application."

Also, in the example shown in Figure 2, the job posting identified by ID "A0002" reads: "[New nurse wanted] No educational background or experience required, must be able to commute to △△ city △△ town, three-shift system. Day shift 8:45-17:00 (60-minute break), late night shift 16:30-24:00 (45-minute break), late night shift 24:00-32:45 (90-minute break). Must be able to work at least five days a week, employment period 3-6 months or 6 months or more. 8-10 late night and night shifts per month. Salary △△ yen - △△ yen. Comprehensive benefits. Why not grow in a rewarding workplace? We look forward to your application." corresponds to the proposed text used as data to be generated, "[Nurses wanted] Educational background and experience are not required, and nurses who can commute to △△ town in △△ city. There are three shifts, with day shifts from 8:45 to 17:00 (60-minute break), late-night shifts from 16:30 to 24:00 (45-minute break), and late-night shifts from 24:00 to 22:45 (90-minute break). Nurses who can work at least five days a week for either 3-6 months or more than 6 months. There are 8-10 late-night and late-night shifts per month. Monthly salary is △△ yen to △△ yen.", and is associated with the assessment used as improvement information, "Since there are few young people in △△ city and it is important for veterans to return to work, it would be better to write in a way that does not limit the application to new nurses, even if educational background and experience are not required."

This allows the generation device 1 to store in the job posting database 122 the job posting script generated by the large-scale language model 121 based on input including generation data indicating a variety of job posting conditions, and provide it to the terminal T, etc. This also allows the generation device 1 to train the large-scale language model 121 to generate job postings based on the job postings and improvement information stored in the job posting database 122 through machine learning.

[Communication unit 13]
The communication unit 13 is not particularly limited as long as it connects the generation device 1 to the network N and enables communication with the terminal T, etc. Examples of the communication unit 13 include a wireless device compatible with a mobile phone network, a device connectable to a wireless LAN, and a network card compatible with the Ethernet standard.

[Network N]
The type of the network N is not particularly limited as long as it allows communication between the generating device 1 and the terminal T, etc. The type of the network N is, for example, the Internet, a mobile phone network, a wireless LAN, etc.

[Terminal T]
Terminal T is not particularly limited. Terminal T may be any of various terminals used by people involved in job postings, such as those who create job postings, such as recruiters, and those who proofread job postings. Terminal T is configured to execute processes such as sending generation data to generation device 1, receiving and displaying the generated job posting from generation device 1, and sending improvement information to generation device 1. In the process of sending improvement information to generation device 1, it is preferable to send information identifying the job posting corresponding to the improvement information (identification information) together with the improvement information.

[Main flowchart of generation process]
Fig. 3 is a main flowchart showing an example of a preferred flow of the generation process executed by the generation device 1 of this embodiment. Fig. 4 is a diagram continuing from Fig. 3. The following is an example of a preferred flow of the generation process executed by the generation device 1 of this embodiment using Figs. 3 and 4.

The generation process includes a series of steps (steps S1 to S2) related to generating a job posting based on the received job requirements. This allows the generation device 1 to automatically generate a job posting for a recruitment page that reflects a variety of job requirements.

[Step S1: Determine whether job conditions have been received]
The control unit 11 executes the condition receiving unit 111 in cooperation with the storage unit 12 and the communication unit 13. The control unit 11 then executes a process to determine whether the job offer conditions have been received from the terminal T or the like by the condition receiving unit 111 (step S1, condition receiving step). If it is determined that the job offer conditions have been received, the control unit 11 stores the received job offer conditions in the storage unit 12 and proceeds to step S2. If it is not determined that the job offer conditions have been received, the control unit 11 proceeds to step S3.

(Recruitment conditions)
The recruitment conditions received in the conditions receiving step are not particularly limited as long as they can be provided as input text to the large-scale language model 121. The recruitment conditions include a wide variety of conditions, exemplified by, for example, job content, work location, whether or not there is a transfer and details thereof, working hours (including non-standard desired hours such as desired hours that vary depending on the day of the week or desired hours that vary only on specific days), working days (days of the week, etc.), contract period, employment type (contract employee, full-time employee, etc.), salary and salary increase system, whether or not there is social insurance and details thereof, whether or not there is a childcare facility and details thereof, whether or not there is childcare support and details thereof, whether or not there is retraining support and details thereof, whether or not there is support for obtaining qualifications and details thereof, compliance efforts (measures against industrial accidents, measures against sexual harassment, measures against power harassment, health management support, etc.), workplace atmosphere, supervisor's hobbies, workplace customs, etc.

The job requirements received in the requirements receiving step are preferably presented in a hierarchy based on semantic relationships, with the various requirements described above divided into categories. For example, the hierarchy of job requirements may be presented in a hierarchy that shows multiple major categories, intermediate categories below each of the major categories, and, as necessary, small categories below each of the intermediate categories.

The following is an explanation of the above-mentioned hierarchy using Figure 2. The job posting identified by ID "A0001" in the example shown in Figure 2 and the corresponding job requirements are an example of hierarchical job requirements divided into categories. In these job requirements, "1. Qualifications, etc." and "2. Working Conditions" are listed as major items corresponding to the highest level. Under the major item "1. Qualifications, etc.", the sub-items listed are (1) conditions related to educational background, etc., and (2) conditions related to age, etc., and under the major item "2. Working Conditions," the sub-items listed are (1) conditions related to leave, etc., (2) conditions related to working hours, and (3) conditions related to salary. Based on this example, those skilled in the art can easily conceive of various types of hierarchy, such as a hierarchy that further includes minor items based on the above example, or a hierarchy that further includes detailed items, etc., under the minor items.

By hierarchizing the job requirements, in the first draft generation step described below, the large-scale language model 121 can appropriately grasp the job requirements based on the job requirements whose semantic relationships have been organized through hierarchization, and generate a first job draft that appropriately reflects the job requirements thus grasped.

[Step S2: Generate a job posting that reflects the received job requirements]
The control unit 11 cooperates with the storage unit 12 and the communication unit 13 to execute the manuscript generation unit 113. Then, the control unit 11 executes a process in which the manuscript generation unit 113 generates a job posting (first job posting) that reflects the job posting conditions received in step S1 in the large-scale language model 121 (step S2, first posting generation step). The control unit 11 then proceeds to step S5.

Through the above-mentioned pre-training of the large-scale language model 121, the large-scale language model 121 acquires the emergent ability to understand job requirements. Then, by using the large-scale language model 121 that has undergone pre-training and acquired the emergent ability, the generation device 1 generates a job posting that reflects the job requirements, based on the job requirements. In this way, the generation device 1 can reduce the workload of recruiters.

[Step S3: Determine whether a draft has been received]
The control unit 11 cooperates with the storage unit 12 and the communication unit 13 to execute the draft receiving unit 112. The control unit 11 then executes a process to determine whether a draft of a job posting has been received from the terminal T or the like using the draft receiving unit 112 (step S3, draft receiving step). If it is determined that the draft has been received, the control unit 11 stores the received draft in the storage unit 12 and proceeds to step S4. If it is not determined that the draft has been received, the control unit 11 proceeds to step S5.

(Draft text of recruitment advertisement)
The draft text of the job posting received in the draft text receiving step is not particularly limited as long as it can be provided as input text to the large-scale language model 121. The draft text includes draft texts of job postings used in a wide variety of recruitment pages, exemplified by the recruitment page reflecting the above-mentioned recruitment conditions.

[Step S4: Generate a job posting that reflects the job requirements indicated in the draft]
The control unit 11 executes the manuscript generation unit 113 in cooperation with the storage unit 12 and the communication unit 13. Then, the control unit 11 executes a process in which the manuscript generation unit 113 generates, in the large-scale language model 121, a job posting (second job posting) that reflects the job requirements indicated in the draft text, based on the draft text received in step S3 (step S4, second manuscript generation step). The control unit 11 then proceeds to step S5.

Through the above-mentioned pre-training in the large-scale language model 121, the large-scale language model 121 acquires the emergent ability to understand the job requirements set forth in the draft text. Then, by using the large-scale language model 121 that has undergone this pre-training and acquired this emergent ability, the generation device 1 understands the job requirements set forth in the draft text based on the draft text, and generates a job posting that reflects the job requirements set forth in the draft text. In this way, the generation device 1 can reduce the workload of recruiters.

[Step S5: Submit job posting]
The control unit 11, in cooperation with the storage unit 12 and the communication unit 13, executes a process of providing the job posting created in step S2 or step S4 to the terminal T or the like (step S5, job posting providing step). The control unit 11 then proceeds to step S6.

The destination to which the job posting is provided in the job posting providing step is not particularly limited. The destination to which the job posting is provided is, for example, terminal T used by the recruiter who will create the recruitment page. In this way, generation device 1 provides the recruiter with a job posting that can serve as a reference for creating the recruitment page or a job posting that can serve as a model for creating a recruitment page, thereby reducing the workload of the recruiter. Therefore, generation device 1 can reduce the burden on the recruiter related to generating the recruitment page by, for example, providing the recruiter with a model recruitment page.

The generation process includes a series of processes (steps S6 to S8) related to machine learning based on the received proofread manuscript. This allows the generation device 1 to learn based on the content and intent of the proofread manuscript, and generate a job posting that is more appropriate for the job requirements.

[Step S6: Determine whether a proof has been received]
The control unit 11 operates the proofreading receiving unit 114 in cooperation with the storage unit 12 and the communication unit 13. The control unit 11 then executes a process to determine whether a proofreading, which is a job posting that has been proofread, has been received from the terminal T or the like (step S6, proofreading receiving step). If it is determined that the proofreading has been received, the control unit 11 associates the received proofreading with the job posting corresponding to the proofreading and stores it in the job posting database 122, and proceeds to step S7. If it is not determined that the proofreading has been received, the control unit 11 proceeds to step S9. Note that the "job posting" in the proofreading receiving step is not particularly limited as long as it is a job posting generated by the generation device 1, and may be, for example, a first job posting, a second job posting, etc.

The means for identifying the job posting corresponding to the proofread draft in the proofread draft receiving step is not particularly limited. Such means may be, for example, a procedure for identifying the job posting using information related to identifying the job posting (identification information) sent along with the proofread draft (identification information usage procedure), or a procedure for identifying the job posting using the history of job postings sent to terminal T that sent the proofread draft (context usage procedure). When the means for identifying the job posting includes the identification information usage procedure, generation device 1 can reliably identify the job posting. When the means for identifying the job posting includes the context usage procedure, generation device 1 can identify the job posting without requiring recruiters, etc., to expend the effort of adding identification information.

[Step S7: Obtaining job requirements]
The control unit 11, in cooperation with the storage unit 12, executes a process of acquiring the job requirements corresponding to the proofread manuscript received in step S6 from the job posting database 122 (step S7, job requirements acquisition step). The control unit 11 then proceeds to step S8.

[Step S8: Machine learning to generate job postings]
The control unit 11 executes the machine learning unit 116 in cooperation with the storage unit 12 and the communication unit 13. The control unit 11 then executes a process in which the machine learning unit 116 uses training data that includes the job requirements acquired in step S7 and the job posting identified in step S6 as explanatory variables and the proofread version received in step S6 as a target variable to have the large-scale language model 121 perform machine learning to generate a job posting that reflects the job requirements as input (step S8, first machine learning step). The control unit 11 then proceeds to step S9.

By automatically generating recruitment pages, the generation device 1 can reduce the burden on recruiters by allowing them to use the generated recruitment pages as is or to use the generated recruitment pages as a model. However, large-scale language model 121 is a language model that has undergone pre-training using large amounts of text, including recruitment pages. However, if the pre-training is insufficient, there is a concern that the generated recruitment pages will be insufficient in terms of various requirements, such as the match between the recruitment requirements and the job posting, or the ability to appeal to job seekers, and will not serve as a model.

In this embodiment, the generation device 1 receives a proofread draft of the job posting generated by the generation device 1 in the proofread draft receiving unit 114, and performs machine learning using training data that includes the job requirements and the job posting as explanatory variables and the received proofread draft as a target variable, thereby allowing the large-scale language model 121 to learn the capabilities that are insufficient with just the above-mentioned emergent capabilities. This allows the generation device 1 to perform machine learning using the proofread draft as a model. The generation device 1 then allows the large-scale language model 121, on which the machine learning has been performed, to generate a model recruitment page and provide it to the recruiter. Therefore, the generation device 1 can reduce the burden on the recruiter involved in generating a recruitment page.

The generation process includes a series of processes (steps S9 to S11) related to machine learning based on the received evaluations. This allows the generation device 1 to learn based on the content and intent of the evaluations, and generate job postings that are more appropriate for the job requirements.

[Step S9: Determine whether evaluation has been received]
The control unit 11 operates the evaluation receiving unit 115 in cooperation with the storage unit 12 and the communication unit 13. The control unit 11 then executes a process to determine whether an evaluation on the first job posting has been received from the terminal T or the like using the evaluation receiving unit 115 (step S9, evaluation receiving step). If it is determined that an evaluation has been received, the control unit 11 associates the received evaluation with the first job posting corresponding to the evaluation and stores it in the job posting database 122, and proceeds to step S10. If it is not determined that an evaluation has been received, the control unit 11 returns to step S1 and repeats steps S1 to S11.

The evaluations include, for example, evaluations showing the effect of the recruitment page based on the generated job posting for various job requirements categorized by industry, age, work location, etc. (number of applications for the job related to the recruitment page, number of replies to the recruitment page on social media, number of likes for the recruitment page on social media, etc.), as well as comments added by recruiters to the generated job posting.

By including the effects of the recruitment page based on the generated job posting in the evaluation, the generation device 1 can perform machine learning based on the reactions to the actual recruitment page, i.e., machine learning that reflects the actual situation in the recruitment field, in the second machine learning step described below. This allows the generation device 1 to improve the generation of job postings so as to elicit more favorable reactions.

By including the recruiter's comments attached to the generated job posting in the evaluation, the generation device 1 can perform machine learning that reflects the recruiter's specialized knowledge in the second machine learning step described below. This allows the generation device 1 to improve the generation of job postings so that they evoke a more desirable response from recruiters.

[Step S10: Obtain job posting]
The control unit 11, in cooperation with the storage unit 12, executes a process of retrieving the job posting and job requirements corresponding to the evaluation received in step S9 from the job posting database 122 (step S10, job posting retrieval step). The control unit 11 then proceeds to step S11.

[Step S11: Machine learning to generate job postings]
The control unit 11 executes the machine learning unit 116 in cooperation with the storage unit 12 and the communication unit 13. Then, the control unit 11 executes a process in which the machine learning unit 116 uses the training data including the evaluations received in step S9 and the job posting and job requirements acquired in step S10 as explanatory variables to cause the large-scale language model 121 to perform machine learning to generate a job posting that reflects the job requirements as input (step S11, second machine learning step). The control unit 11 returns the process to step S1 and repeats steps S1 to S11.

As mentioned above, if the large-scale language model 121 does not have enough prior training, there is a concern that it may lack any of the various requirements, such as the match between the job requirements and the job posting, or the ability to appeal to job seekers, and may generate a job posting that is not suitable as a model.

In this embodiment, the generation device 1 receives evaluations of the job posting generated by the generation device 1 in the evaluation receiving unit 115, and performs machine learning using training data including the evaluations, job requirements, and job postings as explanatory variables, thereby allowing the large-scale language model 121 to learn the capabilities that are insufficient with the emergent capabilities described above. Because the target of machine learning is the large-scale language model 121 that has acquired the emergent capability to generate job postings on its own, the generation device 1 can perform machine learning using training data that does not include a target variable. This allows the generation device 1 to revise the job posting generation policy based on the evaluations. The generation device 1 can then cause the large-scale language model 121, which has undergone machine learning, to generate a model recruitment page and provide it to the recruiter. Therefore, the generation device 1 can reduce the burden on the recruiter involved in generating recruitment pages.

<Usage example>
The following is an example of how the generation device 1 of this embodiment is used.

[Generating job postings]
The generation device 1 receives generation data such as job requirements or draft text from a recruiter or the like via a terminal T or the like. The generation device 1 generates a job posting from the generation data using a large-scale language model 121. The generation device 1 provides the generated job posting to the terminal T or the like.

[Creating improvement information]
The recruiter or the like checks the job posting generated and provided by generation device 1. Then, the recruiter or the like creates improvement information (proofreads, evaluations, etc.) to improve the job posting, as necessary.

[If the improvement information is a proofread draft]
The recruiter or the like proofreads the job posting and creates a proofread draft, which is the proofread job posting. The recruiter or the like then sends the proofread draft to the generation device 1 via a terminal T or the like. The generation device 1 receives the proofread draft and performs machine learning based on the proofread draft.

The following is an explanation of the creation of a proofread version as improvement information and machine learning based on the proofread version, using the job posting identified by ID "A0001" (hereinafter also referred to as "job posting No. 1") in the example job posting database 122 shown in Figure 2 as an example.

The first job posting reads, "We are looking for a registered nurse (new graduates and experienced candidates welcome). Educational background and experience are not important, and the position is open to those under 64 years of age. We have a track record of providing parental leave, caregiving leave, and nursing care leave, and we value work-life balance. Working hours are 9:00-18:00 (60-minute break), and there is no overtime. Salary is △△ yen - △△ yen. Generous bonuses and allowances. You can also improve your skills through extensive training. Why not put your skills to good use on our team? We look forward to receiving your application."

After reviewing Job Posting No. 1, the recruiter felt that it should be emphasized that the position was for a "registered nurse working day shifts," and that rather than emphasizing it by enclosing "[New graduates and experienced candidates welcome]" in square brackets, it should be changed to "[Registered nurse wanted (day shift)]." Furthermore, given the local circumstances, where there are many nurses in the area who have taken time off work to raise children, the recruiter felt that rather than simply stating "emphasis on work-life balance," it should be more specific, stating that "it is possible to balance work with family and childcare." Furthermore, the recruiter felt that the skills development training offered was not particularly comprehensive compared to other workplaces, and felt that the phrase "it is possible to develop your skills through comprehensive training" should be deleted.

The recruiter then proofread Job Posting Script No. 1 based on these ideas, and created the following proofread version of Job Posting Script No. 1 as improvement information: "[Registered nurse wanted (day shift)] We are looking for a registered nurse. Educational background and experience are not required, and applicants must be under 64 years old. There is a track record of taking parental leave, caregiving leave, and nursing care leave, and work can be balanced with family and childcare. Working hours (day shift only): 9:00-18:00 (60-minute break), no overtime. Salary: △△ yen - △△ yen. Generous bonuses and allowances. Why not put your skills to good use on our team? We look forward to receiving your application.", and sent this to Generation Device 1.

The generation device 1 received this proofread and performed machine learning from this proofread and the job postings and job requirements stored in the job posting database 122.

Specifically, the generation device 1 has machine-learned that when the job requirements include "1. Qualifications, etc. (1) Educational background and experience not required (registered nurse required)" and "(2) Working hours 9:00-18:00 (60-minute break), no overtime," it is better to use "Registered nurse wanted (day shift)" as the heading to be emphasized in brackets rather than "New graduates and experienced candidates welcome." Furthermore, the generation device 1 has machine-learned that when the job requirements include "2. Working conditions (1) History of taking childcare leave, family care leave, and nursing care leave," it is better to specifically state "Able to balance work-life and family life" rather than simply stating "Priority on work-life balance." Furthermore, the generation device 1 has machine-learned that when the job requirements do not include information suggesting training, it is better not to write "Skill development is also possible through comprehensive training." This enables the generation device 1 to generate more appropriate and exemplary job postings.

In this way, because the generation device 1 of this embodiment performs machine learning based on proofread drafts, recruiters can improve the generation of job postings in the generation device 1 through intuitive operations, such as creating proofread drafts of job postings and sending them to the generation device 1, just as in the above explanation using the No. 1 job posting as an example. Therefore, the generation device 1 of this embodiment uses machine learning to generate job postings through improvement information (proofreads) provided through intuitive and familiar operations, and can provide recruiters with model recruitment pages, thereby reducing the burden on recruiters involved in generating recruitment pages.

[If the improvement information is evaluation]
The recruiter or the like creates an evaluation of the job posting. Then, the recruiter or the like transmits the evaluation to the generation device 1 via a terminal T or the like. The generation device 1 receives the evaluation and performs machine learning based on the evaluation.

The following is an explanation of the creation of a proofread version as improvement information and machine learning based on the proofread version, using the job posting identified by ID "A0002" (hereinafter also referred to as "No. 2 job posting") in the example job posting database 122 shown in Figure 2 as an example.

Job posting No. 2 reads, "[New nurses wanted] No educational background or experience required, must be able to commute to △△ City/△△ Town, three-shift system. Day shift 8:45-17:00 (60-minute break), semi-night shift 16:30-24:00 (45-minute break), late night shift 24:00-22:45 (90-minute break). Must be able to work at least five days a week, employment period 3-6 months or 6 months or more. 8-10 semi-night and late night shifts per month. Salary △△ yen - △△ yen. Comprehensive benefits. Why not grow in a rewarding workplace? We look forward to your application."

After reviewing job posting No. 2, the recruiter decided that since there are few young people in △△ City and therefore few newly graduated nurses, and since it is more important for veteran nurses to return to work, it would be better to write it so that it does not limit the job to new nurses, even if educational background and experience are not required. Based on this idea, the recruiter created an evaluation of job posting No. 2, saying, "Since there are few young people in △△ City and it is more important for veteran nurses to return to work, it would be better to write it so that it does not limit the job to new nurses, even if educational background and experience are not required." as improvement information, and sent it to generation device 1.

The generation device 1 received this evaluation and performed machine learning based on this evaluation and the job postings and job requirements stored in the job posting database 122.

Specifically, generation device 1 has learned through machine learning that when a draft sentence used as data to be generated includes "Those who can commute to XX City..." in addition to "No educational background or experience required," it is better to use a headline that is not limited to new nurses, such as "Nurses Wanted" or "New graduates and experienced candidates welcome," instead of using the typical headline "(New nurses wanted)" when no educational background or experience is required.

In this way, because the generation device 1 of this embodiment performs machine learning based on evaluations, recruiters can improve the generation of job postings in the generation device 1 through intuitive operations, such as creating an evaluation of the job posting and sending it to the generation device 1, as in the above explanation using the Type 2 job posting as an example. Therefore, the generation device 1 of this embodiment uses machine learning to generate job postings through improvement information (evaluations) provided through intuitive and familiar operations, and can provide recruiters with model recruitment pages, thereby reducing the burden on recruiters involved in generating recruitment pages.

It should be noted that within the scope of the concept of the present invention, those skilled in the art will be able to come up with various modifications and alterations. Therefore, it is understood that these modifications and alterations fall within the scope of the present invention. For example, even if a person skilled in the art appropriately adds or deletes components or modifies the design of the above-described embodiment, or adds or omits processes or modifies conditions, these modifications are also included within the scope of the present invention as long as they maintain the essence of the present invention.

S System 1 Generation device 11 Control unit 111 Condition receiving unit 112 Draft text receiving unit 113 Manuscript generation unit 114 Proofread manuscript receiving unit 115 Evaluation receiving unit 116 Machine learning unit 12 Storage unit 121 Large-scale language model 122 Job posting manuscript database 13 Communication unit N Network T Terminal

Claims (2)

a condition receiving unit for receiving job conditions;
a script generation unit that provides an input including the job requirements to a large-scale language model and causes the large-scale language model to generate a first job requirements script that reflects the job requirements;
a proofreading receiving unit that receives the first job posting proofread by a recruiter;
a rating receiving unit that receives a rating for the first job posting;
a machine learning unit that uses training data including the evaluation, the job requirements, and the first job posting as explanatory variables and the proofread version corresponding to the first job posting as a target variable to train a large-scale language model to generate a job posting that reflects the job requirements as input;
Equipped with
The evaluation receiving unit receives an evaluation including the number of replies on SNS to a recruitment page based on the first job posting and/or the number of likes on the recruitment page on SNS.
Recruitment page generator. A draft receiving unit for receiving a draft of the job posting is further provided,
the manuscript generation unit provides an input including the draft text to a large-scale language model, and causes the large-scale language model to generate a second job posting that reflects the job requirements included in the draft text;
The generating device of claim 1 .