Enterprise human resource intelligent scheduling and dynamic configuration optimization system and method

Abstract

The invention discloses an enterprise human resource intelligent scheduling and dynamic configuration optimization system and method, and relates to the technical field of talent resource information processing, and the scheme comprises a data processing module which is used for carrying out the analysis of posts issued by all enterprises and resumes uploaded by all job seekers, obtaining corresponding post portrait vectors and talent portrait vectors; the talent-post matching module is used for mapping the post portrait vector and the talent portrait vector to the same semantic vector space by utilizing a pre-training language model, and acquiring a candidate resume set of a to-be-recruited post based on the similarity between the post portrait vector and the talent portrait vector; and the multi-task recommendation resource scheduling module is used for executing an intelligent resource scheduling mechanism based on the reinforcement learning model so as to perform optimal resource allocation on the candidate resume set of the to-be-recruitment post, and updating pre-training language model parameters based on a result output by the reinforcement learning model. According to the technology provided by the invention, the problem of unreasonable enterprise human resource scheduling is solved.

Description

Technical Field

[0001] This invention relates to the field of human resource information processing technology, and in particular to an intelligent scheduling and dynamic configuration optimization system and method for enterprise human resources. Background Technology

[0002] As businesses expand and the digitalization of the human resources market accelerates, recruitment faces the dual challenges of massive job demand and massive resume supply. To improve the efficiency of matching people to jobs, existing recruitment management systems typically integrate Natural Language Processing (NLP), Optical Character Recognition (OCR), and recommendation algorithms, aiming to assist in talent screening through automation.

[0003] In existing technical solutions, the conventional implementation process mainly includes the following steps: First, at the data parsing level, OCR technology is used to extract text from resume documents, and keyword matching algorithms based on rules or statistical machine learning are used to compare the skills, education, and other entities in the resume with the job description; Second, at the person-job matching level, the system calculates the similarity score between candidates and jobs, generates a recommendation list, and pushes it to recruiters; Third, at the resource scheduling level, when facing multiple concurrent job postings, the system usually uses static rules for resource allocation, such as according to the time order of job postings or by manually setting a fixed number of recommendations; Finally, at the model iteration level, the system collects historical click, communication, and hiring data from recruiters and trains the model offline to optimize the accuracy of recommendations.

[0004] For example, Chinese invention patent CN116433201B discloses an intelligent matching system and method for talent resource information based on big data, which includes: Step 1: predicting the recruitment requirements of recruiting companies at various delay points; Step 2: filtering talent information in the talent pool based on the obtained relevant professional names and recruitment professional names; Step 3: performing secondary filtering of talent information; Step 4: intelligently matching talents according to the constructed talent map and sending the matched talent information to the company's recruitment end.

[0005] For example, Chinese invention patent application CN120494775A discloses an intelligent identification method based on a talent digital intelligence brain system, which includes: first, constructing a talent database based on the talent digital intelligence brain system; then collecting enterprise recruitment demand texts, historical recruitment data, and industry trend information, and extracting core indicators such as key skills, experience requirements, and cultural fit; combining enterprise historical data and industry dynamics, and identifying potential needs not explicitly stated by enterprises through association rule mining algorithms; integrating explicit and potential needs to form a structured enterprise recruitment demand set; intelligently screening candidates based on the talent digital intelligence brain system and the enterprise recruitment demand set; and generating a personalized recommendation report after screening.

[0006] The above-mentioned technology has at least the following technical problems: Existing technologies typically employ static allocation strategies such as first-come, first-served or manually preset allocation. Due to a lack of dynamic awareness and global optimization capabilities regarding job deadlines and talent scarcity, the system cannot adaptively adjust resource allocation. This makes it difficult for urgent positions to obtain sufficient recommended resources, resulting in an unreasonable allocation of recommended resources.

[0007] Furthermore, existing matching models suffer from significant data bias during training. Popular job postings or top-ranked resumes generate a large number of clicks due to their high exposure, causing the model to overfit these popular samples during training. Conversely, high-quality long-tail resumes that go unviewed due to poor display positions lack effective feedback samples. This data-level bias directly impacts the matching model, making it difficult to learn and identify potential talent. Moreover, as the model iterates, this preference for popular samples is continuously reinforced, ultimately causing the system to lose its ability to identify long-tail and specialized talent, revealing problems with the company's HR allocation. Summary of the Invention

[0008] This invention provides an intelligent scheduling and dynamic configuration optimization system and method for enterprise human resources, which can improve the accuracy of talent-job matching, thereby improving the accuracy of enterprise human resource scheduling. The technical solution provided by this application is as follows: According to the first aspect of this application, an intelligent scheduling and dynamic configuration optimization system for enterprise human resources is provided. This system includes: a data processing module for monitoring job posting requests from various enterprises and resume upload requests from job seekers through a preset interface, parsing the job postings from each enterprise and the resumes uploaded by each job seeker to obtain corresponding job profile vectors and talent profile vectors; a talent-job matching module for mapping the job profile vectors and talent profile vectors to the same semantic vector space using a pre-trained language model, and obtaining a set of candidate resumes for the job to be recruited based on the similarity between the job profile vectors and talent profile vectors; and a multi-task recommendation resource scheduling module for triggering an intelligent resource scheduling mechanism based on a reinforcement learning model if there are multiple concurrent tasks and recommendation resource constraints, to optimally allocate resources to the set of candidate resumes for the job to be recruited, and updating the parameters of the pre-trained language model based on the output of the reinforcement learning model; otherwise, outputting talent-job recommendation results based on the set of candidate resumes for the job to be recruited.

[0009] According to another aspect of this application, a method for intelligent scheduling and dynamic configuration optimization of enterprise human resources is provided. This method is applied to an enterprise human resources intelligent scheduling and dynamic configuration optimization system. The method includes: S1, monitoring job posting requests from various enterprises and resume upload requests from job seekers through a preset interface, parsing the job postings from various enterprises and the resumes uploaded by job seekers respectively to obtain corresponding job profile vectors and talent profile vectors; S2, using a pre-trained language model to map the job profile vectors and talent profile vectors to the same semantic vector space, and obtaining a set of candidate resumes for the job to be recruited based on the similarity between the job profile vectors and talent profile vectors; S3, if there is multi-task concurrency and recommendation resource constraints, triggering an intelligent resource scheduling mechanism based on a reinforcement learning model to optimally allocate resources to the set of candidate resumes for the job to be recruited, and updating the parameters of the pre-trained language model based on the output of the reinforcement learning model; otherwise, outputting talent-job recommendation results based on the set of candidate resumes for the job to be recruited.

[0010] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. By analyzing job postings from various companies and resumes uploaded by job seekers, unstructured text information is transformed into structured key features. These key features are then quantified into vector forms, enabling mathematical calculation and comparison of job and talent information. This provides a feasible data foundation for talent-job matching. Existing technologies primarily rely on keyword matching or simple rule matching, while this invention maps job profile vectors and talent profile vectors to the same semantic vector space. This allows for a more accurate understanding of the semantic meaning of jobs and talents, avoiding matching errors caused by different keywords with the same meaning, and improving the accuracy of talent-job matching. The similarity between vectors is calculated, and a candidate resume set is selected based on the similarity. This set includes job seekers with a high degree of matching with the job openings, narrowing the recommendation scope and improving recommendation efficiency. Finally, by judging the concurrency of multiple tasks and resource constraints, different resource scheduling methods can be flexibly selected according to the actual situation. The intelligent resource scheduling mechanism based on the reinforcement learning model can dynamically allocate resources according to the real-time task status and resource conditions, achieving optimal resource utilization and improving the overall performance of human resource scheduling. The reinforcement learning model parameters are updated according to the actual task execution, enabling the model to continuously learn and optimize, thereby improving the accuracy and efficiency of subsequent resource scheduling.

[0011] 2. By using reinforcement learning models to predict recommendation success rates and expected conversion cycles, we can better understand the recruitment difficulty and time required for each position in advance, providing a scientific basis for subsequent resource allocation and thus improving recruitment efficiency and quality. Then, by calculating the comprehensive recommendation score and obtaining the urgency level, we can more comprehensively evaluate the recruitment value of each position, highlighting its time urgency. Next, by distributing positions according to their urgency level to different priority scheduling pools and sorting them within the pools based on the comprehensive recommendation score, we achieve classified management and orderly processing of positions, preventing important and urgent positions from being delayed. Finally, while existing technologies use fixed resource allocation methods, such as allocation according to position order or random allocation, this invention traverses the scheduling queue according to priority order and dynamically allocates resources based on the remaining allocable quota and the minimum number of resumes allocated daily. This ensures that each position receives reasonable resource support while avoiding problems of excessive resource concentration or uneven distribution.

[0012] 3. By judging the historical view rate, if the historical view rate does not meet the standard, sample weights are introduced to weight the loss function before training the pre-trained language model. This allows the pre-trained language model to focus more on learning from specific samples during training, optimize model parameters, and improve the accuracy of the pre-trained language model in predicting the matching degree between job and talent. Especially for resumes that have not been viewed, if the historical view rate meets the standard, the ranking decay function is triggered for jobs in an inefficient processing state, thereby reducing the comprehensive recommendation score of jobs in an inefficient processing state, so that jobs with high processing efficiency can be processed first. Finally, within the preset ranking recovery period, the comprehensive recommendation score of the corresponding job is gradually increased until the initial comprehensive recommendation score is restored. This gives jobs an opportunity to improve and ensures the dynamic balance of the scheduling queue, avoiding a job being in a disadvantageous position for a long time due to a single judgment.

[0013] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0014] The accompanying drawings are provided for a better understanding of this solution and do not constitute a limitation of this application. Wherein: Figure 1 This is an interface diagram of the recruitment mini-program provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an intelligent scheduling and dynamic configuration optimization system for enterprise human resources provided in an embodiment of the present invention; Figure 3 This is an architecture diagram of the pre-trained language model provided in an embodiment of the present invention; Figure 4 This is a flowchart of the architecture of the reinforcement learning model provided in the embodiments of the present invention; Figure 5 This is a flowchart of resource scheduling provided in an embodiment of the present invention; Figure 6 This is a flowchart for determining the resume history view rate provided in an embodiment of the present invention. Detailed Implementation

[0015] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of this application, including various details to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0016] like Figure 1The interface diagram of the recruitment mini-program shown in this embodiment of the invention provides an intelligent scheduling and dynamic configuration optimization system for enterprise human resources, applied to a recruitment mini-program. Figure 2 The diagram shows the structure of an intelligent scheduling and dynamic configuration optimization system for enterprise human resources. The processing flow of this system may include the following steps: The data processing module monitors job posting requests from companies and resume upload requests from job seekers through preset interfaces. This ensures the system can obtain the latest job and resume information in a timely manner, guaranteeing the real-time nature and freshness of the data. This provides a foundation for subsequent accurate matching and recommendations. The module parses the job postings from companies and the resumes uploaded by job seekers to obtain corresponding job profile vectors and talent profile vectors. It transforms complex unstructured text information into structured key features, improving the processability and analyzability of the information. The vector representation allows job and talent information to be calculated and compared at the mathematical level, providing a feasible data foundation for talent-job matching.

[0017] Specifically, the process involves analyzing the job postings from various companies and the resumes uploaded by job seekers. The specific steps are as follows: This study utilizes a pre-trained language model to perform named entity recognition and dependency parsing on published job description texts to extract job profile entities. Word embedding technology is then used to map these extracted entities into structured job profile vectors, which mathematically quantify the job's required features. Named entity recognition identifies key entities such as job title, required skills, and educational requirements from the job description text. Dependency parsing analyzes the dependency relationships between words in a sentence, revealing its grammatical structure. Job profile entities represent key entities extracted from the job description text that represent the job's required features, such as job title, required professional skills, work experience requirements, educational requirements, and work location. This approach accurately and comprehensively extracts key information from job description texts, avoiding the subjectivity and omissions inherent in manual extraction and improving the accuracy and efficiency of information extraction. Furthermore, dependency parsing allows for a better understanding of the logical relationships between various pieces of information in the job description, providing a more accurate foundation for subsequent job profile vector construction.

[0018] When analyzing and optimizing resumes, each identified text block in the image has corresponding coordinates, usually including horizontal and vertical coordinates, to determine its specific location. Obtaining the coordinate information of the text blocks helps to locate and analyze the text in the resume, providing an important basis for restoring the logical reading order and accurately extracting the talent profile entity.

[0019] The LayoutLM layout analysis model is introduced to process the layout of resume images and identify the types of text regions in the resume, such as education background and work experience, which helps to accurately extract relevant information. The layout analysis model can automatically identify different text region types in the resume, improving the accuracy and efficiency of resume parsing. By accurately identifying each region, talent profile entities can be extracted from the corresponding regions in a more targeted manner, avoiding information confusion and incorrect extraction.

[0020] Coordinate calculations are performed based on the coordinate information of each text block to reconstruct the logical reading order of the resume. Specifically: if the average difference in the horizontal coordinates of the center points of all text blocks is less than a preset limit, all text blocks are determined to belong to the same column; otherwise, they are determined to belong to different columns, and the column order is defined according to the horizontal coordinates. Within the same column, the vertical arrangement order of each text block is defined according to the vertical coordinates. The average difference in the horizontal coordinates of the center points of all text blocks represents the result obtained by calculating the distance between the horizontal coordinate of each text block and the average horizontal coordinate, and then taking the average value. The limit for the horizontal coordinate difference is set by the preset personnel based on the layout characteristics of the resume (such as column width). The logical reading order refers to the natural order followed when reading a document, usually from top to bottom and from left to right. Reconstructing the logical reading order of the resume allows the computer to process resume information according to people's normal reading habits, improving the accuracy and rationality of information extraction, and providing a more accurate foundation for subsequent talent profile entity extraction and vector construction.

[0021] Talent profile entities are extracted from the text corresponding to the logical reading order after the coordinate information of each text block is restored. Talent profile entities include educational background entities, work experience entities, and project experience. Similar to the mapping method of job profile vectors, each extracted talent profile entity is mapped into a structured talent profile vector. The talent profile vector is used to quantitatively represent the job seeker's ability characteristics at the mathematical level. Extracting talent profile entities according to the logical reading order can more accurately obtain information such as the job seeker's educational background, work experience, and project experience, avoiding extraction errors caused by disordered information order.

[0022] like Figure 3 The diagram shows the architecture of a pre-trained language model. The input layer consists of [CLS], python, 3 years of experience, and [SEP]. [CLS] is the classification label, a special token of the BERT model used to represent the aggregate representation of the entire sequence. python represents the text content in the job description, 3 years of experience represents the text content in the job description, and [SEP] is the delimiter used to separate different text segments. The embedding layer includes E... [CLS] , E1 to E n E [SEP]This represents the embedding vector of the input token, converting text into a numerical vector; the BERT core layer includes Transformer and BERT. Transformer is the basic building block of the BERT model, containing multiple layers of self-attention mechanisms, while BERT is the overall model name, representing a bidirectional encoder based on Transformer; the output layer includes C, T1 to T... n T [SEP] Job entity classification header and Mask LM, C, T1 to T n and T [SEP] This represents the token after BERT processing. The job entity classification header is the classifier that outputs the entity type. Mask LM is a masked language model, one of the pre-training tasks of BERT.

[0023] The talent-job matching module utilizes pre-trained language models such as BERT / ERNIE, trained on general text data (e.g., the entire Wikipedia, web pages, books, etc.), to map job profile vectors and talent profile vectors to the same semantic vector space. Specifically, the job profile vector or talent profile vector is input into the pre-trained language model, which converts it into characters or words. The encoder of the pre-trained language model outputs the semantic representation vector corresponding to the job profile vector or talent profile vector, eliminating the semantic differences between job and talent information. This makes similarity calculation more accurate and reasonable, improving the precision of talent-job matching. Based on the cosine similarity of the job profile vector and talent profile vector, and using approximate nearest neighbor search technology for filtering, a preset number of candidate resumes with the highest cosine similarity (set by a preset number of experienced personnel, e.g., 10) are selected as the candidate resume set for the job openings. This narrows the recommendation scope, focusing attention on job seekers with high job matching, improving recommendation efficiency and targeting.

[0024] The multi-task recommendation resource scheduling module is used to trigger an intelligent resource scheduling mechanism based on a reinforcement learning model when there are multiple concurrent tasks and recommendation resource constraints. Specifically, if the number of pending recommendation tasks exceeds a pre-set threshold for concurrent tasks (pre-defined by experienced personnel), and the system's resource utilization (e.g., CPU utilization, memory utilization) exceeds a pre-set threshold for resource utilization (pre-defined by experienced personnel), then the module optimizes resource allocation across the set of candidate resumes for the job openings. It also updates the parameters of the pre-trained language model based on the reinforcement learning model's output, enabling the model to continuously learn and optimize, making more reasonable resource allocation decisions in subsequent tasks. Otherwise, it outputs talent-job recommendation results based on the set of candidate resumes for the job openings, ensuring the efficiency and real-time performance of intelligent human resource scheduling. The module flexibly selects different processing methods according to the system's actual situation, improving the adaptability and flexibility of intelligent human resource scheduling and enabling it to cope with different business scenarios and resource conditions. Under conditions of multi-task concurrency and resource constraints, it can dynamically adjust resource allocation strategies to achieve optimal resource utilization, improving the overall performance and recommendation effectiveness of intelligent human resource scheduling.

[0025] like Figure 4 The flowchart of the reinforcement learning model shown illustrates the following: The state feature input layer includes job urgency, talent scarcity, and resume processing efficiency. These features are input into the intermediate reinforcement learning model training module, which employs a neural network structure. The V-branch (state value branch) and A-branch (action advantage branch) of Dueling DQN predict the job recommendation success rate and conversion cycle, respectively, forming an action space. The action output and priority processing module outputs a comprehensive recommendation score based on the predicted job recommendation success rate and conversion cycle, classifies urgency levels, and generates a scheduling queue based on the comprehensive recommendation score and urgency level. The resource scheduling module scans the queue, calculates the remaining allocable quota, and allocates resumes according to the acquired quota for this round. The environmental feedback and parameter update module obtains the state value data of the next state, evaluates the performance assessment value, calculates the temporal difference error, and updates the network parameters through backpropagation, thus forming a complete reinforcement learning loop.

[0026] like Figure 5 The flowchart shown illustrates the resource scheduling process. Specifically, the intelligent resource scheduling mechanism based on a reinforcement learning model includes: This involves establishing a connection between a reinforcement learning model (agent) and an enterprise human resource intelligent scheduling and dynamic configuration optimization system (environment). Initially, the neural network parameters within the model are randomly initialized. The reinforcement learning model and the environment then interact cyclically. The reinforcement learning model perceives the current environmental state, such as an urgent and scarce position. Based on the current strategy (network parameters), it evaluates the value of each possible action under this environmental state, outputs a predicted Q-value, executes the action with the highest predicted Q-value, and feeds the result back to the reinforcement learning model. Finally, the reinforcement learning model outputs a target Q-value. The predicted Q-value and the target Q-value are compared, and the difference between them is calculated. Based on this difference, optimization algorithms such as gradient descent are used to adjust the weights of the model's neural network, making the predicted Q-value closer to the target Q-value. This process utilizes a trained reinforcement learning model, such as a Deep Q-Network (DQN) or Dueling... DQN outputs predicted values ​​based on current state characteristics. These predicted values ​​include the expected recommendation success rate, the expected conversion period, and state value data. The recommendation success rate refers to the probability of successfully recommending suitable candidates to the positions available and ultimately securing employment. The expected conversion period refers to the estimated time required from recommending a candidate to the position until that candidate is successfully converted into an employed employee. The reinforcement learning model can automatically learn patterns and rules from historical data and make dynamic predictions based on current state characteristics, gaining an understanding in advance of the recruitment difficulty and time required for each position, thus providing a scientific basis for subsequent resource allocation.

[0027] Status characteristics include job urgency, talent scarcity, and resume processing efficiency. Job urgency is the reciprocal of the time difference between the job application deadline and the current date (if the denominator is 0, the job urgency is set to the maximum value of job urgency in the historical time period). Talent scarcity is the reciprocal of the historical application rate (if the denominator is 0, the talent scarcity is set to the maximum value of talent scarcity in the historical time period). The historical application rate represents the ratio of the number of resumes received to the number of job postings for a specific position in the past recruitment process, used to measure the level of attention that talents pay to the position.

[0028] The comprehensive recommendation score for a job posting is defined as the product of the job's recommendation success rate and its importance level weight. Simultaneously, a corresponding urgency level is obtained, quantifying the urgency of the job's recruitment needs. The job importance level weight represents a weight assigned by pre-defined personnel based on factors such as the job's importance to the company, its strategic value, and its business impact. This weight is used to reflect the varying degrees of importance of different jobs when calculating the comprehensive recommendation score. The comprehensive recommendation score considers both the job's recommendation success rate and importance, providing a more comprehensive assessment of the job's recruitment value. The urgency level classification highlights the job's time-sensitive nature, enabling the company's intelligent human resource scheduling system to prioritize important and urgent positions, ensuring that the recruitment needs of key positions are met in a timely manner.

[0029] All job openings are distributed to corresponding priority scheduling pools based on their urgency level. Within each priority scheduling pool, positions are sorted in descending order based on their comprehensive recommendation score, generating a scheduling queue. Each priority scheduling pool represents a different set categorized by job urgency level, storing job openings with the same urgency level for priority processing during resource allocation. The scheduling queue represents a sequence of job openings within each priority scheduling pool, arranged from highest to lowest comprehensive recommendation score, guiding the order of resource allocation. By distributing job openings to different priority scheduling pools based on urgency level and sorting them within each pool according to comprehensive recommendation score, categorized management and orderly processing of job openings are achieved.

[0030] The scheduling queues are traversed sequentially from highest to lowest priority. Starting with the highest priority scheduling pool, each priority scheduling pool is traversed in turn to dynamically allocate resources to the positions to be recruited. For the currently traversed position, the remaining allocable quota for that position in the current recruitment cycle is calculated, and the corresponding resumes are allocated to that position according to the quota allocated in this round. By traversing the scheduling queues in priority order and dynamically allocating resumes to each position according to the quota allocated in this round, it is ensured that each position receives reasonable resource support, thereby flexibly adjusting according to the actual situation and avoiding the problems of excessive concentration or uneven distribution of resources.

[0031] The remaining allocable quota is the difference between the maximum number of resumes that can be received for this position and the current number of resumes that have been allocated for this position. The allocation quota for this round is the minimum of the minimum number of resumes that can be allocated for this position per day and the remaining allocable quota. The maximum number of resumes that can be received represents the upper limit of the number of resumes that the position can receive during the recruitment period, which is determined by the preset personnel based on factors such as the position requirements and the company's processing capacity. The current number of resumes that have been allocated represents the number of resumes that have been allocated to this position as of the current moment. The minimum number of resumes that can be allocated per day represents the minimum number of resumes that need to be allocated for each position to ensure the recruitment progress.

[0032] The specific methods for obtaining the emergency level are as follows: If the urgency level of a position to be recruited falls within the first urgency range, or the talent scarcity level falls within the first talent scarcity range, then the urgency level is recorded as the first urgency level. The numerical range of position urgency is divided into different intervals, each corresponding to a different urgency level, to quickly determine the urgency range of the position. Similarly, the numerical range of talent scarcity is divided into different intervals, each corresponding to a different talent scarcity level, to facilitate determining the talent scarcity situation. Positions with extremely high urgency or extremely high talent scarcity are marked as the first urgency level, highlighting the urgency and importance of these positions. During the allocation of human resources related to talent recruitment, these positions can receive priority attention and processing, ensuring that key positions can be filled with talent in a timely manner.

[0033] If a job's urgency falls within the second urgency range, or its talent scarcity falls within the second talent scarcity range, then the urgency level is recorded as the second urgency level. For jobs with moderate urgency and talent scarcity, they are also marked as the second urgency level. While prioritizing the handling of jobs with the first urgency level, resources can be rationally allocated for recruiting these jobs, ensuring an orderly recruitment process and avoiding excessive concentration or dispersion of resources. In other words, as long as a job falls within the first urgency range or the first talent scarcity range, it is classified as the first urgency level; if either of these ranges is in the second urgency range or the second talent scarcity range, it is classified as the second urgency level.

[0034] If neither of the above two conditions is met, the urgency level will be recorded as Level 3. Positions with relatively low urgency and talent scarcity will also be classified as Level 3. These positions can be allocated recruitment resources relatively later. While ensuring the recruitment needs of the first two levels of positions are met, recruitment work can be gradually advanced based on the actual situation to improve the efficiency of recruitment resource utilization.

[0035] The lower limit of the emergency zone for the first position is higher than the lower limit of the emergency zone for the second position. If the urgency of a position is greater than the urgency limit for the first position, then the urgency of the position belongs to the emergency zone for the first position. If the urgency of a position is not greater than the urgency limit for the first position, but is greater than the urgency limit for the second position, then the urgency of the position belongs to the emergency zone for the second position. If the urgency of a position is not greater than the urgency limit for the second position, then the urgency of the position belongs to the emergency zone for the third position. The urgency limits for the first and second positions are set by the pre-set personnel based on the actual application situation.

[0036] The lower limit of the first talent scarcity interval is higher than the lower limit of the second talent scarcity interval. If the talent scarcity is greater than the first talent scarcity limit, then the talent scarcity belongs to the first talent scarcity interval. If the talent scarcity is not greater than the first talent scarcity limit but is greater than the second talent scarcity limit, then the talent scarcity belongs to the second talent scarcity interval. If the talent scarcity is not greater than the second talent scarcity limit, then the job urgency belongs to the third talent scarcity interval. The first and second talent scarcity limits are set by the preset personnel based on the actual application situation.

[0037] The level of urgency decreases progressively from the first to the third emergency level.

[0038] By dividing positions into first, second, and third urgency zones and talent scarcity zones, the urgency level of positions can be determined more accurately, providing a more detailed basis for the allocation of recruitment resources and further improving the efficiency of human resource utilization and the rationality of the recruitment process.

[0039] Specifically, the parameters of the pre-trained language model are updated based on the output of the reinforcement learning model. The specific process is as follows: The reinforcement learning model outputs predicted values ​​based on the current state features. The predicted values ​​include the expected recommendation success rate, the expected conversion cycle, and state value data. In reinforcement learning, the state value represents the expected cumulative reward that the agent can obtain by following the current policy in a certain state.

[0040] By collecting relevant records, we can identify the number of people successfully hired during the actual recruitment process, the number of resumes recommended to positions, and the time it takes for each resume to go from recommendation to hiring. We can also statistically analyze the actual number of hires, the current number of resumes assigned, and the time from resume recommendation to successful hiring to evaluate recruitment effectiveness and the performance of reinforcement learning models.

[0041] The ratio of actual hires to the number of resumes currently assigned is recorded as the conversion rate metric, measuring the efficiency of converting recommended resumes into successful hires. The reciprocal of the time from resume recommendation to successful hire is then mapped to the range of 0-1 using a Sigmoid function to obtain the recruitment efficiency metric; the shorter the time from resume recommendation to successful hire, the higher the recruitment efficiency. The conversion rate metric and the recruitment efficiency metric are weighted and summed to obtain a performance evaluation value, which integrates the conversion effect and efficiency of recruitment, and is used to comprehensively evaluate the overall effectiveness of the recruitment process. The weights of the conversion rate metric and the recruitment efficiency metric are set by preset personnel based on the actual application scenario.

[0042] The process involves acquiring the state value data for the next state, and then calculating the temporal difference error based on the performance evaluation value, the current state value data, and the next state value data. This temporal difference error is used as a gradient signal to update the reinforcement learning model's parameters. The temporal difference error represents an error metric used in reinforcement learning to update state value estimates, measuring the difference between the current state value estimate and the prediction based on the next state value estimate and immediate rewards. By minimizing the temporal difference error, the model can progressively improve its state value estimation, thereby optimizing the strategy. Through continuous parameter updates, the reinforcement learning model can gradually learn better recruitment strategies. As a gradient signal, the temporal difference error guides the model to adjust its parameters in a direction that maximizes long-term cumulative rewards, thereby improving the expected recommendation success rate, shortening the expected conversion cycle, and ultimately enhancing the efficiency of intelligent human resource scheduling for enterprises.

[0043] like Figure 6 The flowchart shown for determining resume history view rate should be supplemented with the following: Based on the output of the reinforcement learning model, the parameters of the pre-trained language model are updated, followed by: For resumes that have not been viewed within a preset time period, the historical view rate of the resume is obtained based on the statistical data of the positions applied for. The historical view rate represents the ratio of the number of times the resume was viewed to the number of times the resume was applied for within the preset time period, providing an important basis for subsequent resume processing.

[0044] If the historical view rate is not less than the lower limit of the view rate, the sample weight of the corresponding resume is set to the reciprocal of the historical view rate, and the resume processing efficiency of the monitored job openings is judged. The lower limit of the view rate represents a pre-set threshold for judging the degree of attention a resume receives. The lower the historical view rate of a resume, the greater its sample weight, and the more attention it receives during the training process. Judging the resume processing efficiency helps to identify problems in the recruitment process in a timely manner, such as slow resume screening, so that corresponding measures can be taken to optimize it.

[0045] If the historical view rate is less than the lower limit, the sample weight of the corresponding resume should be set to the maximum weight to make the pre-trained language model pay more attention to this type of resume. The maximum weight can be set to the maximum value of the sample weights within the historical time period. If the historical view rate is extremely low, taking the reciprocal of the historical view rate will cause the model to fail to converge. Setting the sample weight of the corresponding resume to the maximum weight can control the weights of all samples within a reasonable numerical range. In machine learning, sample weight is used to measure the importance of each sample in the training process. The larger the weight, the greater the impact of the sample on the model training. By setting sample weights, the importance of unviewed resumes can be highlighted, making the pre-trained language model pay more attention to these samples during training and improving the pre-trained language model's ability to process this type of resume.

[0046] By introducing sample weights, the loss function in the pre-trained language model, which maps job profile vectors and talent profile vectors to the same semantic vector space, is weighted. This weighted loss function is then used to train the pre-trained language model, increasing the learning weight of unviewed samples. In machine learning, the loss function measures the difference between the model's predictions and the actual results. Minimizing the loss function optimizes the parameters of the pre-trained language model, improving its performance. The weighted loss function guides the model to focus more on unviewed resumes, optimizing the mapping relationship between job profile vectors and talent profile vectors, and improving the model's accuracy and efficiency in resume matching and job recommendation.

[0047] Specifically, the efficiency of resume processing for monitored job openings is assessed, and the specific process is as follows: Resume processing efficiency includes average resume processing time and resume backlog rate. The resume backlog rate is the ratio of the current number of resumes to be processed to the preset daily processing capacity limit. Average resume processing time refers to the average time spent processing all resumes received for the job posting (such as initial screening and evaluation), and is used to measure the speed of resume processing. By clearly defining the indicators of resume processing efficiency, including average resume processing time and resume backlog rate, we can comprehensively and objectively measure the efficiency of resume processing for the job, and provide an accurate basis for subsequent judgment on whether the job is in an inefficient processing state.

[0048] If the average resume processing time exceeds the limit, or the resume backlog rate exceeds the limit, the position is determined to be in an inefficient processing state, triggering a ranking decay function to lower its ranking priority in the scheduling queue. If neither of these conditions is met, the position is determined not to be in an inefficient processing state and will not be processed. The ranking decay function is used to adjust the ranking priority of positions in the scheduling queue. Determining a position to be in an inefficient processing state can promptly identify inefficient links in the recruitment process, preventing slow processing of individual positions from affecting the overall recruitment progress. Triggering the ranking decay function to lower the ranking priority of positions allows the enterprise's intelligent human resource scheduling system to prioritize higher-efficiency positions, rationally allocate resources, and improve overall recruitment efficiency.

[0049] Specifically, the function that triggers the sorting decay is as follows: The product of the overall recommendation score of the position in the scheduling queue and the attenuation factor preset by the preset personnel based on the actual application scenario is used as the updated overall recommendation score of the position. The ranking priority of the position in the scheduling queue is then re-determined based on the updated overall recommendation score.

[0050] Within the preset sorting recovery period, if the average processing time for resumes corresponding to positions in an inefficient processing state does not exceed the resume average processing time limit, and the resume backlog rate does not exceed the resume backlog rate limit, then the overall recommendation score for that position will be gradually increased until it is restored to the original overall recommendation score of that position in the scheduling queue, thus gradually restoring its sorting priority to its original level. Otherwise, the overall recommendation score for that position will not be increased. Gradually restoring the overall recommendation score of positions within the preset sorting recovery period allows positions sufficient time to improve. This ensures that the enterprise's intelligent human resource scheduling system incentivizes efficiency improvements while avoiding a situation where a single misjudgment leads to a prolonged period of unreasonable position sorting priorities.

[0051] Specifically, based on the set of candidate resumes for the positions to be filled, the talent-job recommendation results are output. The specific process is as follows: The matching score between each candidate resume and the job opening is obtained. The matching score reflects the similarity between the talent profile vector of the candidate resume and the job profile vector of the job opening. It is calculated using cosine similarity. By calculating cosine similarity, the matching degree between talent and job is transformed into a specific numerical value, making the matching results measurable and comparable. Talent profile vector and job profile vector contain information from multiple dimensions. Cosine similarity calculation can comprehensively consider the relationship between these dimensions and comprehensively evaluate the matching degree between candidates and jobs, rather than relying on a single factor.

[0052] Based on the matching score, all candidate resumes in the candidate resume set are sorted in descending order; resuming candidate resumes with higher matching scores at the top can quickly locate the candidate resumes that best match the job and improve screening efficiency.

[0053] Based on the maximum number of resumes that can be accepted for a job opening, the system selects a target number of candidate resumes from the sorted pool as the final recommendation output. The target number cannot exceed the maximum number of resumes that can be accepted. If the number of candidate resumes is not less than the maximum number of resumes that can be accepted, then the target number equals the maximum number of resumes that can be accepted; if the number of candidate resumes is less than the maximum number of resumes that can be accepted, then the target number equals the number of candidate resumes. Selecting recommendations based on the maximum number of resumes that can be accepted for a job opening ensures that the number of recommended candidates matches the actual needs of the position, avoiding recommending too many or too few resumes and improving the efficiency of recruitment resource utilization. Only selecting the top-ranked target number of candidate resumes for recommendation reduces the number of resumes that recruiters need to process, achieving dynamic allocation of the company's human resources.

[0054] This invention provides a method for intelligent scheduling and dynamic configuration optimization of enterprise human resources, including: S1 monitors job posting requests from various companies and resume upload requests from job seekers through a preset interface. It parses the job postings from each company and the resumes uploaded by each job seeker to obtain corresponding job profile vectors and talent profile vectors. By monitoring requests in real time through the preset interface, it can obtain the latest information from companies and job seekers in a timely manner, ensuring the timeliness of the data. It parses job postings and resumes and converts them into vectors, transforming unstructured text information into structured numerical information, which facilitates subsequent computer processing and analysis, improving the efficiency and accuracy of information processing.

[0055] S2 utilizes a pre-trained language model to map job profile vectors and talent profile vectors to the same semantic vector space. Based on the similarity between the job profile vectors and talent profile vectors, it obtains a set of candidate resumes for the job openings. By using a pre-trained language model to map vectors to the same semantic vector space, it can accurately capture the semantic relationship between jobs and talents, overcoming the limitations of traditional keyword-based matching methods and improving the accuracy and comprehensiveness of matching. Obtaining a set of candidate resumes based on similarity can quickly filter out the job seekers who best match the job, reducing the amount of data to be processed subsequently and improving the efficiency of intelligent scheduling and dynamic configuration optimization of enterprise human resources.

[0056] S3. If there are multiple concurrent tasks and resource constraints, an intelligent resource scheduling mechanism based on a reinforcement learning model is triggered to optimally allocate resources to the set of candidate resumes for the job openings. This mechanism can dynamically allocate resources according to the actual situation, achieve optimal utilization of human resources, avoid waste and unreasonable allocation of resources, and update the parameters of the pre-trained language model based on the output of the reinforcement learning model. This improves the adaptive capability and long-term performance of the enterprise's intelligent human resource scheduling and dynamic configuration optimization. Otherwise, the talent-job recommendation result is output based on the set of candidate resumes for the job openings.

[0057] Example 2: If there is an abnormal time, Example 1 cannot respond to the emergency in a timely manner. Therefore, Example 2 needs to be used for resource scheduling based on Example 1.

[0058] The intelligent resource scheduling mechanism based on reinforcement learning models also includes: The genetic algorithm generates a globally optimal scheduling scheme based on a pre-defined fitness function. This scheme includes talent-job recommendation results and priority ranking. In this application, the pre-defined fitness function evaluates the merits of each scheduling scheme, generating a globally optimal scheme through iterative evolution. The fitness function considers multiple factors, such as talent-job matching, recruitment progress, and resource utilization, to assess the quality of the scheduling scheme; a higher function value indicates a better scheme. The talent-job recommendation results represent recommending suitable talent for each job based on the talent's abilities, experience, and skills, as well as the job's needs, responsibilities, and requirements, or recommending a matching job for each talent, thus establishing a talent-job correspondence. Priority ranking involves sorting the matching pairs in the talent-job recommendation results in descending order of matching degree to determine the processing order. By simulating biological evolution, the genetic algorithm can search for near-globally optimal scheduling schemes in a complex solution space. By comprehensively considering multiple key factors in the recruitment process, the generated scheduling scheme achieves a balance in talent-job matching, resource utilization efficiency, and recruitment progress, improving the quality and efficiency of enterprise recruitment and human resource scheduling.

[0059] Real-time monitoring of status characteristics enables timely detection of abnormal changes in the system, ensuring the scheduling system's rapid response to emergencies. If an abnormal event is detected, the affected subset of tasks is identified for precise problem localization, avoiding large-scale adjustments to the entire scheduling scheme. An insertional heuristic search is performed on the task subset to find the nearest available time window in the original scheduling queue, outputting the globally optimal scheduling scheme after reallocation to adapt to changes brought about by the abnormal event. An abnormal event refers to an event caused by unexpected state changes during the execution of the globally optimal scheduling scheme, leading to its failure. This includes, but is not limited to, the insertion of new urgent positions and the suspension of recruitment for existing positions. For example, the insertion of a new urgent position means that suitable candidates need to be found quickly, and the original globally optimal scheduling scheme cannot meet this requirement. This urgent need; Insertion-based heuristic search refers to searching the solution space using heuristic rules (rules based on experience or specific goals) to find solutions that meet the new conditions by inserting new operations or elements into the original problem. In this application, when an abnormal event occurs, the algorithm is used on the affected subset of tasks to find a suitable adjustment scheme in the original scheduling queue, i.e., the insertion position is calculated by prioritizing the earliest available time window strategy; the available time window represents a time period in the original scheduling queue that meets specific conditions (such as not conflicting with other tasks, resources being available, etc.) and is used to reschedule tasks affected by the abnormal event; by finding the nearest available time window, the impact on other tasks is minimized, ensuring the stability and flexibility of the enterprise human resource intelligent scheduling system in abnormal situations.

[0060] If no abnormal events are detected, the talent-job recommendation results are output based on the set of candidate resumes for the positions to be recruited, providing recruiters with timely and accurate talent recommendations and improving the quality of talent recommendations.

[0061] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims can be performed in a different order than that shown in the embodiments and still achieve the desired results. Additionally, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are also possible or may be advantageous.

[0062] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the method, device, and storage medium embodiments are basically similar to the system embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the system embodiments.

[0063] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0064] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A smart scheduling and dynamic allocation optimization system for enterprise human resources, characterized in that, The system includes: The data processing module is used to monitor job posting requests from various companies and resume upload requests from job seekers through preset interfaces. It parses the job postings from various companies and the resumes uploaded by job seekers to obtain the corresponding job profile vectors and talent profile vectors. The talent-job matching module uses a pre-trained language model to map job profile vectors and talent profile vectors to the same semantic vector space, and obtains a set of candidate resumes for the job to be recruited based on the similarity between the job profile vectors and talent profile vectors. The multi-task recommendation resource scheduling module is used to trigger an intelligent resource scheduling mechanism based on a reinforcement learning model if there are multiple concurrent tasks and recommendation resource constraints. This mechanism optimizes the allocation of resources to the set of candidate resumes for the job openings and updates the parameters of the pre-trained language model based on the output of the reinforcement learning model. Otherwise, it outputs talent-job recommendation results based on the set of candidate resumes for the job openings.

2. The enterprise human resource intelligent scheduling and dynamic configuration optimization system as described in claim 1, characterized in that, The intelligent resource scheduling mechanism based on the reinforcement learning model specifically includes: Using a reinforcement learning model, the recommendation success rate and expected conversion cycle of job openings are predicted based on current state features; The status characteristics include job urgency, talent scarcity, and resume processing efficiency. The product of the recommendation success rate of the job opening and the weight of the job importance level is defined as the comprehensive recommendation score of the job opening. At the same time, the corresponding urgency level is obtained. The urgency level is used to quantify the urgency of the job recruitment needs. All open positions are distributed to the corresponding priority scheduling pools according to their urgency level. Within each priority scheduling pool, positions are sorted in descending order based on the comprehensive recommendation score to generate a scheduling queue. The scheduling queue is traversed sequentially in descending order of priority. Resources are dynamically allocated to the positions to be recruited. For the position currently being traversed, the remaining allocable quota for that position in the current recruitment cycle is calculated and dynamically allocated according to the allocation quota for this round.

3. The enterprise human resource intelligent scheduling and dynamic allocation optimization system as described in claim 2, characterized in that, The specific method for obtaining the emergency level is as follows: If the urgency of the position to be recruited is in the first urgency range, or the talent scarcity is in the first talent scarcity range, then the urgency level will be recorded as the first urgency level. In addition, if the urgency of a position falls within the second urgency range, or the talent scarcity falls within the second talent scarcity range, then the urgency level will be recorded as the second urgency level. Otherwise, it will be recorded as the third emergency level; The lower limit of the emergency interval for the first position is higher than the lower limit of the emergency interval for the second position. The lower limit of the first talent scarcity interval is higher than the lower limit of the second talent scarcity interval; The urgency level decreases progressively from the first emergency level to the second emergency level and the third emergency level.

4. The enterprise human resource intelligent scheduling and dynamic configuration optimization system as described in claim 2, characterized in that, The intelligent resource scheduling mechanism based on the reinforcement learning model also includes: A globally optimal scheduling scheme is generated based on a preset fitness function in the genetic algorithm. The globally optimal scheduling scheme includes talent-job recommendation results and priority ranking. Real-time monitoring of status characteristics; if an abnormal event is detected, the subset of tasks affected by the abnormal event is locked, an insertional heuristic search is performed on the subset of tasks, the nearest available time window is found in the original scheduling queue, and the globally optimal scheduling scheme after reallocation is output. If no abnormal events are detected, output talent-job recommendation results based on the set of candidate resumes for the positions to be recruited; The aforementioned abnormal event refers to an event that causes the global optimal scheduling scheme to fail due to an unexpected state change during the execution of the global optimal scheduling scheme.

5. The enterprise human resource intelligent scheduling and dynamic configuration optimization system as described in claim 1, characterized in that, The specific process for updating the pre-trained language model parameters based on the output of the reinforcement learning model is as follows: The reinforcement learning model outputs a predicted value based on the current state features. The predicted value includes the expected recommendation success rate, the expected conversion cycle, and state value data. Statistics include the actual number of hires, the number of resumes currently assigned, and the time from resume recommendation to successful hiring. The ratio of the actual number of hires to the number of resumes currently assigned is recorded as the conversion rate indicator. Recruitment efficiency metrics are calculated based on the time from resume submission to successful hiring. The effectiveness evaluation value is obtained by weighting and coupling the conversion rate indicator and the recruitment efficiency indicator; Obtain the state value data of the next state. Based on the performance evaluation value, the state value data of the current state, and the state value data of the next state, obtain the temporal difference error. Then, use the temporal difference error as a gradient signal to input into the reinforcement learning model to update the model parameters.

6. The enterprise human resource intelligent scheduling and dynamic configuration optimization system as described in claim 5, characterized in that, The process of updating the pre-trained language model parameters based on the output of the reinforcement learning model also includes: For resumes that have not been viewed within a preset time period, the historical view rate of the resume is obtained based on the statistical data of the positions applied for. If the historical view rate is not less than the lower limit of the view rate, the sample weight of the corresponding resume will be set to the reciprocal of the historical view rate, and the resume processing efficiency of the monitored job openings will be judged. If the historical view rate is less than the lower limit of the view rate, the sample weight of the corresponding resume will be set to the maximum weight. The sample weights are introduced to weight the loss function in the pre-trained language model that maps job profile vectors and talent profile vectors to the same semantic vector space. The pre-trained language model is then trained using the weighted loss function.

7. The enterprise human resource intelligent scheduling and dynamic configuration optimization system as described in claim 6, characterized in that, The process for determining the efficiency of resume processing for monitored job openings is as follows: The resume processing efficiency includes the average resume processing time and the resume backlog rate; If the average resume processing time exceeds the resume average processing time limit, or the resume backlog rate exceeds the resume backlog rate limit, the position is determined to be in an inefficient processing state, triggering the sort decay function. If the above conditions are not met, the position is deemed not to be in an inefficient processing state and no processing is performed. The trigger sorting decay function is specifically as follows: The product of the overall recommendation score of the position in the scheduling queue and the preset decay factor is used as the updated overall recommendation score of the position. The ranking priority of the position in the scheduling queue is re-determined based on the updated overall recommendation value score. Within the preset sorting recovery period, if the average processing time of resumes for positions in an inefficient processing state is not greater than the average resume processing time limit and the resume backlog rate is not greater than the resume backlog rate limit, then the overall recommendation score for that position will be gradually increased until it is restored to the original overall recommendation score of that position in the scheduling queue; otherwise, the overall recommendation score for that position will not be increased.

8. The enterprise human resource intelligent scheduling and dynamic configuration optimization system as described in claim 1, characterized in that, The process of analyzing the job postings from various companies and the resumes uploaded by job seekers is as follows: A pre-trained language model is used to perform named entity recognition and dependency parsing on the published job description text to extract job profile entities, and the extracted job profile entities are mapped into structured job profile vectors. The resume is parsed and optimized, and the OCR is used to recognize the resume image corresponding to the resume and obtain the coordinate information of each text block in the resume image. A layout analysis model is introduced to process the layout of resume images and identify the types of text regions in the resume; Talent profile entities are extracted from the text corresponding to the logical reading order restored based on the coordinate information of each text block, and each extracted talent profile entity is mapped into a structured talent profile vector.

9. The enterprise human resource intelligent scheduling and dynamic configuration optimization system as described in claim 1, characterized in that, The process of outputting talent-job recommendation results based on the set of candidate resumes for the positions to be filled is as follows: Obtain the matching score between each candidate resume and the job opening, whereby the matching score reflects the similarity between the talent profile vector of the candidate resume and the job profile vector of the job opening. Sort all candidate resumes in the candidate resume set in descending order according to the matching score; Based on the maximum number of resumes that can be received for the job opening, a target number of candidate resumes are selected from the sorted candidate resume set as the final recommendation result output, wherein the target number does not exceed the maximum number of resumes that can be received.

10. A method for intelligent scheduling and dynamic allocation optimization of enterprise human resources, characterized in that, The method includes: S1 monitors job posting requests from various companies and resume upload requests from job seekers through a preset interface, and parses the job postings from various companies and the resumes uploaded by job seekers to obtain the corresponding job profile vector and talent profile vector. S2 utilizes a pre-trained language model to map job profile vectors and talent profile vectors to the same semantic vector space, and obtains a set of candidate resumes for the job openings based on the similarity between the job profile vectors and talent profile vectors. S3. If there are multiple concurrent tasks and resource constraints for recommendations, an intelligent resource scheduling mechanism based on a reinforcement learning model is triggered to optimally allocate resources to the set of candidate resumes for the job openings. The parameters of the pre-trained language model are updated based on the output of the reinforcement learning model. Otherwise, talent-job recommendation results are output based on the set of candidate resumes for the job openings.

Images