An artificial intelligence-based human resource management method and system
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIAN FANHUA TECH DEV CO LTD
- Filing Date
- 2025-12-26
- Publication Date
- 2026-06-19
Smart Images
Figure CN122243429A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of human resource management technology, and in particular to a human resource management method and system based on artificial intelligence. Background Technology
[0002] As businesses expand and market competition intensifies, traditional human resource management models present numerous inconveniences. The recruitment process relies on manual resume screening, which is inefficient and susceptible to subjective factors, making it difficult to accurately match job requirements with candidate capabilities. Employee training often adopts a "one-size-fits-all" approach, failing to develop personalized training programs based on employees' existing skill gaps and career development needs, resulting in wasted training resources and poor effectiveness. Performance evaluation standards are vague, and the evaluation process relies on managers' subjective judgment, lacking data support, which can easily lead to employee dissatisfaction. Employee turnover warnings are delayed, making it impossible to identify employee turnover risks in advance and take intervention measures, increasing the company's talent loss costs.
[0003] In existing technologies, human resource management systems introduce simple data statistics functions but do not deeply integrate artificial intelligence technology, thus failing to achieve intelligent data analysis and decision support. Recruitment systems can only screen resumes based on keywords and cannot identify the candidate's implicit abilities and their suitability for the position. Some training systems can only push fixed courses and cannot dynamically adjust training content. The AI models of existing systems mostly use general algorithms, resulting in serious problems of algorithm homogenization. Summary of the Invention
[0004] The purpose of this invention is to propose a human resource management method and system based on artificial intelligence to solve the above-mentioned problems.
[0005] This invention provides an artificial intelligence-based human resource management method, comprising:
[0006] Data Acquisition and Preprocessing: Collect multi-dimensional data on candidates, employees, and positions; clean and standardize the collected structured data; and convert unstructured data into analyzable feature vectors using natural language processing and computer vision technologies.
[0007] Artificial intelligence model construction and training: Construct a multi-scenario AI model with a human resources scenario-based multimodal collaborative Transformer model as the core. The multimodal collaborative Transformer model adopts a three-layer architecture of "scenario-based feature extraction layer + cross-modal collaborative fusion layer + dynamic weight decision layer". Based on the multimodal collaborative Transformer model, a recruitment matching sub-model, a personalized training recommendation sub-model, an intelligent performance evaluation sub-model, and a turnover risk warning sub-model are derived.
[0008] Intelligent management and decision-making output: Input the pre-processed data into the corresponding scenario sub-model, and output recruitment matching results, personalized training programs, performance evaluation reports and turnover risk warning information;
[0009] Model optimization and iteration: Collect user feedback data and actual business data, and use reinforcement learning to optimize and update the parameters of the multimodal collaborative Transformer model and its sub-models.
[0010] In some embodiments, the contextual feature extraction layer of the multimodal collaborative Transformer model includes:
[0011] The contextualized text feature extraction module constructs a dedicated vocabulary containing 5,000+ core human resources terms and a "job-skill-achievement" knowledge graph. It also incorporates a time decay factor into the self-attention mechanism, which assigns a weight of 0.3-0.4 to data from 3 years ago and a weight of 0.7-0.9 to data from the past year.
[0012] The scenario-based visual feature extraction module customizes a human resources micro-expression and behavioral feature library that includes the logic of interview answers and job suitability behaviors, and dynamically adjusts the weight of visual features through industry tags.
[0013] The scenario-based numerical feature extraction module is designed with an intelligent outlier correction unit. It identifies and corrects data anomalies based on job characteristics, and at the same time constructs a "numerical-semantic association mapping" to transform performance scores and attendance data into semantic ability scores.
[0014] In some embodiments, the cross-modal collaborative fusion layer of the multimodal collaborative Transformer model includes:
[0015] The modal semantic alignment unit, based on the contextualized knowledge graph, constructs a multimodal semantic mapping matrix to uniformly map textual, visual, and numerical features to the semantic space of "ability dimension + score".
[0016] A dynamic collaborative attention mechanism dynamically allocates the weights of each modality according to the scenario type. In the recruitment scenario, the weights of the text modality are 0.4-0.6, the visual modality is 0.2-0.4, and the numerical modality is 0.1-0.3. In the resignation warning scenario, the weights of the numerical modality are 0.3-0.5, the text modality is 0.2-0.4, and the visual modality is 0.2-0.4.
[0017] The cross-modal feature enhancement unit enhances the accuracy of visual feature judgment with detailed features of text data and optimizes the semantic mapping results of numerical features with behavioral features of visual data through modal mutual enhancement learning.
[0018] In some embodiments, the dynamic weight decision layer of the multimodal collaborative Transformer model includes:
[0019] The scene label recognition unit automatically identifies the scene type for online campus recruitment and R&D performance evaluation, and calls the pre-configured scene weight template;
[0020] The feedback iteration unit collects decision feedback from HR and hiring departments, and adjusts the weight parameters of each modality feature through reinforcement learning algorithms, with an iteration cycle of 1-3 months.
[0021] In some embodiments, the output of the recruitment matching sub-model includes:
[0022] The matching degree between candidates and positions is scored, with a scoring accuracy of 1%.
[0023] The analysis includes matching strengths and weaknesses, and the analysis includes the identification results of the candidate's implicit skills;
[0024] The candidate ranking list is arranged in descending order of matching degree and supports displaying the core feature comparison of the top 10-20 candidates.
[0025] According to claim 1, the human resource management method based on artificial intelligence is characterized in that the output of the turnover risk early warning sub-model includes:
[0026] The risk level of employee turnover is divided into three levels: low, medium, and high.
[0027] Risk cause analysis should clearly indicate the specific values and percentages of risk characteristics such as attendance abnormalities, decreased communication frequency, and performance decline;
[0028] Intervention measures recommended for high-risk employees include job adjustments, salary negotiations, and psychological counseling.
[0029] This invention provides an artificial intelligence-based human resource management system, comprising:
[0030] The data acquisition module includes a resume import interface, an interview video recording unit, an attendance system integration unit, and a communication record capture unit, which is used to collect multi-dimensional data on candidates, employees, and positions.
[0031] The data preprocessing module includes a structured data cleaning unit, an NLP processing subunit, and a CV processing subunit, which are used to transform the collected data into standardized feature vectors.
[0032] The AI model library module stores multimodal collaborative Transformer models and sub-models for various scenarios, and provides model calling and training interfaces;
[0033] The scenario-based management module includes sub-modules for recruitment management, training management, performance management, and employee relations, which call the corresponding scenario sub-models to implement functions.
[0034] The engineering support module includes a scenario-based model configuration platform, a multimodal data access SDK, and a model performance monitoring panel, which are used for the implementation and maintenance of multimodal collaborative Transformer models.
[0035] This invention addresses the core issues of homogeneity in existing system algorithms and insufficient depth of multimodal fusion through a three-layer customized architecture of a multimodal collaborative Transformer model, effectively improving the model's accuracy in recognizing implicit skills.
[0036] AI models can automatically complete repetitive tasks such as resume screening, performance evaluation, and training recommendations, reducing manual operations and freeing HR personnel from tedious tasks so they can focus on strategic work.
[0037] Based on the deep fusion of multi-source data and scenario-based decision-making capabilities of the multimodal collaborative Transformer model, it avoids interference from subjective human factors and improves the accuracy of recruitment matching, the objectivity of performance evaluation, and the timeliness of resignation warning;
[0038] Based on the characteristics and needs of different candidates and employees, and relying on the semantic alignment and dynamic weighting capabilities of the multimodal collaborative Transformer model, we provide customized recruitment matching, training programs, and intervention measures to improve candidate experience and employee satisfaction.
[0039] Reduce the costs of recruitment mismatch, training resource waste, and talent loss, and improve the return on investment of human resource management;
[0040] The analysis reports generated by the decision support module, combined with multimodal data insights from the multimodal collaborative Transformer model, provide data support for enterprises to formulate talent strategies and help them achieve talent-driven development. Attached Figure Description
[0041] The accompanying drawings illustrate, by way of example and not limitation, the various embodiments discussed herein.
[0042] Figure 1 This is a schematic diagram of the human resource management method based on artificial intelligence according to an embodiment of the present invention. Detailed Implementation
[0043] In order to gain a more detailed understanding of the features and technical content of the embodiments of this application, the implementation of the embodiments of this application will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for reference and illustration only and are not intended to limit the embodiments of this application.
[0044] like Figure 1 As shown, the artificial intelligence-based human resource management system includes:
[0045] The data acquisition module includes a resume import interface, an interview video recording unit, an attendance system integration unit, and a communication record capture unit, which is used to collect multi-dimensional data on candidates, employees, and positions.
[0046] The data preprocessing module includes a structured data cleaning unit, an NLP processing subunit, and a CV processing subunit, which are used to transform the collected data into standardized feature vectors.
[0047] The AI model library module stores multimodal collaborative Transformer models and sub-models for various scenarios, and provides model calling and training interfaces;
[0048] The scenario-based management module includes sub-modules for recruitment management, training management, performance management, and employee relations, which call the corresponding scenario sub-models to implement functions.
[0049] The engineering support module includes a scenario-based model configuration platform, a multimodal data access SDK, and a model performance monitoring panel, which are used for the implementation and maintenance of multimodal collaborative Transformer models.
[0050] The specific implementation steps of the artificial intelligence-based human resource management method of this invention are as follows:
[0051] Step 1: Data collection and preprocessing. Collect multi-dimensional human resources data, including candidate data, employee data, and job data. Clean and standardize the collected structured data, and convert unstructured data into analyzable feature vectors using natural language processing and computer vision technologies.
[0052] Step 2: Artificial intelligence model construction and training. Construct a multi-scenario AI model with the human resources scenario-based multimodal collaborative Transformer model as the core. The multimodal collaborative Transformer model adopts a three-layer architecture of "scenario-based feature extraction layer + cross-modal collaborative fusion layer + dynamic weight decision layer". Based on the multimodal collaborative Transformer model, a recruitment matching sub-model, a personalized training recommendation sub-model, an intelligent performance evaluation sub-model, and a turnover risk warning sub-model are derived.
[0053] Step 3: Intelligent management and decision output. Input the preprocessed data into the corresponding scenario sub-model and output recruitment matching results, personalized training programs, performance evaluation reports and turnover risk warning information.
[0054] Step 4: Model optimization and iteration. Collect user feedback data and actual business data, and use reinforcement learning to optimize and update the parameters of the multimodal collaborative Transformer model and its sub-models.
[0055] In a preferred embodiment of the present invention, the scene-based feature extraction layer of the multimodal collaborative Transformer model in step two includes three types of customized modules:
[0056] The contextualized text feature extraction module constructs a dedicated vocabulary containing 5,000+ core human resources terms and a "job-skill-achievement" knowledge graph. It also incorporates a time decay factor into the self-attention mechanism, which assigns a weight of 0.3-0.4 to data from 3 years ago and a weight of 0.7-0.9 to data from the past year.
[0057] The scenario-based visual feature extraction module customizes a human resources micro-expression and behavioral feature library that includes the logic of interview answers and job suitability behaviors, and dynamically adjusts the weight of visual features through industry tags.
[0058] The scenario-based numerical feature extraction module is designed with an intelligent outlier correction unit. It identifies and corrects data anomalies based on job characteristics, and at the same time constructs a "numerical-semantic association mapping" to transform performance scores and attendance data into semantic ability scores.
[0059] In a preferred embodiment of the present invention, the cross-modal collaborative fusion layer of the multimodal collaborative Transformer model in step two includes:
[0060] The modal semantic alignment unit, based on the contextualized knowledge graph, constructs a multimodal semantic mapping matrix to uniformly map textual, visual, and numerical features to the semantic space of "ability dimension + score".
[0061] The modal semantic alignment unit, based on the contextualized knowledge graph, constructs a multimodal semantic mapping matrix to uniformly map textual, visual, and numerical features to the semantic space of "ability dimension + score".
[0062] A dynamic collaborative attention mechanism dynamically allocates the weights of each modality based on the scenario type. In the recruitment scenario, the text modality has a weight of 0.4-0.6, the visual modality 0.2-0.4, and the numerical modality 0.1-0.3. In the resignation warning scenario, the numerical modality has a weight of 0.3-0.5, the text modality 0.2-0.4, and the visual modality 0.2-0.4.
[0063] The cross-modal feature enhancement unit enhances the accuracy of visual feature judgment with detailed features of text data and optimizes the semantic mapping results of numerical features with behavioral features of visual data through modal mutual enhancement learning.
[0064] In a preferred embodiment of the present invention, the dynamic weight decision layer of the multimodal collaborative Transformer model in step two includes:
[0065] The scene label recognition unit automatically identifies the scene type for online campus recruitment and R&D performance evaluation, and calls the pre-configured scene weight template;
[0066] The feedback iteration unit collects decision feedback from HR and hiring departments, and adjusts the weight parameters of each modality feature through reinforcement learning algorithms, with an iteration cycle of 1-3 months.
[0067] In a preferred embodiment of the present invention, the output of the recruitment matching sub-model in step three includes:
[0068] The matching degree between candidates and positions is scored, with a scoring accuracy of 1%.
[0069] The analysis includes matching strengths and weaknesses, and the analysis includes the identification results of the candidate's implicit skills;
[0070] The candidate ranking list is arranged in descending order of matching degree and supports displaying the core feature comparison of the top 10-20 candidates.
[0071] In a preferred embodiment of the present invention, the output of the resignation risk early warning sub-model in step three includes:
[0072] The risk level of employee turnover is divided into three levels: low, medium, and high.
[0073] Risk cause analysis should clearly indicate the specific values and percentages of risk characteristics such as attendance abnormalities, decreased communication frequency, and performance decline;
[0074] Intervention measures recommended for high-risk employees include job adjustments, salary negotiations, and psychological counseling.
[0075] This invention also relates to an artificial intelligence-based human resource management system, comprising:
[0076] The data acquisition module includes a resume import interface, an interview video recording unit, an attendance system integration unit, and a communication record capture unit, which is used to collect multi-dimensional data on candidates, employees, and positions.
[0077] The data preprocessing module includes a structured data cleaning unit, an NLP processing subunit, and a CV processing subunit, which are used to transform the collected data into standardized feature vectors.
[0078] The AI model library module stores multimodal collaborative Transformer models and sub-models for various scenarios, and provides model calling and training interfaces;
[0079] The scenario-based management module includes sub-modules for recruitment management, training management, performance management, and employee relations, which call the corresponding scenario sub-models to implement functions.
[0080] The engineering support module includes a scenario-based model configuration platform, a multimodal data access SDK, and a model performance monitoring panel, which are used for the implementation and maintenance of multimodal collaborative Transformer models.
[0081] Example 1: Recruitment of Internet Product Managers
[0082] This implementation method is based on a medium-sized Internet company with 500 employees. The company's core business is Internet product development and operation. It has the characteristics of large recruitment demand, rapid iteration of employee skills, performance evaluation that needs to be adapted to agile development model, and high risk of core talent turnover.
[0083] The specific steps are as follows:
[0084] 1. Job posting and data collection:
[0085] Enterprise HR logs into the recruitment management sub-module of the scenario-based management module, enters the job requirements for "Senior Product Manager": the job requirements are "more than 5 years of Internet product experience, have led more than 2 projects with more than 5 million users, and are familiar with user needs analysis", and selects the industry tag "Internet".
[0086] The data collection module batch imports 120 resumes submitted by candidates for this position within 3 days;
[0087] High-definition cameras were deployed in the company's interview room to record 15-minute initial interviews with 30 candidates who passed the initial screening. The videos were automatically uploaded to the MongoDB database.
[0088] 2. Data Preprocessing and Model Analysis
[0089] Clean up candidates' written test scores;
[0090] The resume text is processed to extract keywords such as "5 years of product experience" and "led a project worth tens of millions" and generate text feature vectors.
[0091] Analyze interview videos to extract visual features such as "fluctuations in speaking speed when answering questions about user needs" and "confidence level in micro-expressions when mentioning project details";
[0092] The preprocessed data is input into the AI model library module.
[0093] Multimodal collaborative Transformer model startup:
[0094] Contextualized feature extraction unit: Based on the Internet industry thesaurus, the resume text "tens of millions of yuan projects" is mapped to "business contribution score of 90 points", and the weight of project experience three years ago is adjusted by the time decay factor (weight of the past 2 years of experience is 0.8).
[0095] Cross-modal collaborative fusion unit: Construct a semantic mapping matrix to unify text features (experience matching degree 88 points), visual features (expression ability 85 points), and written test scores (82 points → professional ability 82 points) into the semantic space, and calculate the comprehensive score according to the recruitment scenario weight (text 0.5, visual 0.3, numerical 0.2);
[0096] One candidate was found to have "led projects worth tens of millions of yuan on their resume, but was unable to describe the details during the interview" (modal conflict). The conflict resolution model was invoked, prioritizing visual features (project authenticity score of 60) and reducing the text weight to 0.3.
[0097] Based on the output results, a candidate matching ranking is generated: the top 10 candidates all have a matching score of ≥85, among which candidate A has a matching score of 92.
[0098] 3. Interview Decision-Making and Model Optimization
[0099] HR can view the matching ranking and analysis report, select the top 5 candidates to enter the second round of interviews, and automatically send interview invitations;
[0100] One month later, the recruitment process ended. Candidate A joined the company and passed the probationary period, and HR marked it as a "valid match"; Candidate B failed the probationary period due to "insufficient interview expression skills", and was marked as "visual feature assessment needs to be strengthened".
[0101] The model optimization module automatically feeds the feedback data into the dynamic weight decision unit, and through reinforcement learning, increases the visual modality weight in the recruitment scenario from 0.3 to 0.35, thereby improving the subsequent matching accuracy.
[0102] Example 2: Early Warning of Turnover Risk in R&D Positions
[0103] 1. Real-time data acquisition and preprocessing
[0104] The data acquisition module captures real-time data from an employee in the R&D department.
[0105] Data preprocessing module processes data
[0106] Correct attendance records that were abnormal due to server failure;
[0107] Extract keywords related to job loss risks, such as "project direction" and "career development," from communication records;
[0108] Analyze the employee's team meeting videos and extract the visual feature of "increased frequency of low mood";
[0109] 2. Risk assessment and early warning notification
[0110] Data is input into the turnover risk early warning sub-model of the AI model library module;
[0111] Contextualized feature extraction unit: Transforms attendance anomalies into "stability risk 75 points" and performance decline into "job suitability risk 80 points";
[0112] Cross-modal collaborative fusion unit: The comprehensive risk score calculated based on the weight of the resignation warning scenario is 76 points (high risk);
[0113] Generate an early warning report: Risk level "high", risk reasons "abnormal attendance + declining performance + decreased communication frequency", recommended intervention measures "technical lead communicates career development path + HR assesses the possibility of job adjustment";
[0114] The Employee Relations submodule of the Scenario-based Management module automatically sends alert notifications to HR and the R&D Director.
[0115] The technical solutions described in the embodiments of this application can be combined arbitrarily without conflict.
[0116] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A human resource management method based on artificial intelligence, characterized in that, include: Data Acquisition and Preprocessing: Collect multi-dimensional data on candidates, employees, and positions; clean and standardize the collected structured data; and convert unstructured data into analyzable feature vectors using natural language processing and computer vision technologies. Artificial intelligence model construction and training: Construct a multi-scenario AI model with a human resources scenario-based multimodal collaborative Transformer model as the core. The multimodal collaborative Transformer model adopts a three-layer architecture of "scenario-based feature extraction layer + cross-modal collaborative fusion layer + dynamic weight decision layer". Based on the multimodal collaborative Transformer model, a recruitment matching sub-model, a personalized training recommendation sub-model, an intelligent performance evaluation sub-model, and a turnover risk warning sub-model are derived. Intelligent management and decision-making output: Input the pre-processed data into the corresponding scenario sub-model, and output recruitment matching results, personalized training programs, performance evaluation reports and turnover risk warning information; Model optimization and iteration: Collect user feedback data and actual business data, and use reinforcement learning to optimize and update the parameters of the multimodal collaborative Transformer model and its sub-models.
2. The artificial intelligence-based human resource management method according to claim 1, characterized in that, The contextualized feature extraction layer of the multimodal collaborative Transformer model includes: The contextualized text feature extraction module constructs a dedicated vocabulary containing 5,000+ core human resources terms and a "job-skill-achievement" knowledge graph. It also incorporates a time decay factor into the self-attention mechanism, which assigns a weight of 0.3-0.4 to data from 3 years ago and a weight of 0.7-0.9 to data from the past year. The scenario-based visual feature extraction module customizes a human resources micro-expression and behavioral feature library that includes the logic of interview answers and job suitability behaviors, and dynamically adjusts the weight of visual features through industry tags. The scenario-based numerical feature extraction module is designed with an intelligent outlier correction unit. It identifies and corrects data anomalies based on job characteristics, and at the same time constructs a "numerical-semantic association mapping" to transform performance scores and attendance data into semantic ability scores.
3. The artificial intelligence-based human resource management method according to claim 1, characterized in that, The cross-modal collaborative fusion layer of the multimodal collaborative Transformer model includes: The modal semantic alignment unit, based on the contextualized knowledge graph, constructs a multimodal semantic mapping matrix to uniformly map textual, visual, and numerical features to the semantic space of "ability dimension + score". A dynamic collaborative attention mechanism dynamically allocates the weights of each modality according to the scenario type. In the recruitment scenario, the weights of the text modality are 0.4-0.6, the visual modality is 0.2-0.4, and the numerical modality is 0.1-0.
3. In the resignation warning scenario, the weights of the numerical modality are 0.3-0.5, the text modality is 0.2-0.4, and the visual modality is 0.2-0.
4. The cross-modal feature enhancement unit enhances the accuracy of visual feature judgment with detailed features of text data and optimizes the semantic mapping results of numerical features with behavioral features of visual data through modal mutual enhancement learning.
4. The artificial intelligence-based human resource management method according to claim 1, characterized in that, The dynamic weight decision layer of the multimodal collaborative Transformer model includes: The scene label recognition unit automatically identifies the scene type for online campus recruitment and R&D performance evaluation, and calls the pre-configured scene weight template; The feedback iteration unit collects decision feedback from HR and hiring departments, and adjusts the weight parameters of each modality feature through reinforcement learning algorithms, with an iteration cycle of 1-3 months.
5. The artificial intelligence-based human resource management method according to claim 1, characterized in that, The output of the recruitment matching sub-model includes: The matching degree between candidates and positions is scored, with a scoring accuracy of 1%. The analysis includes matching strengths and weaknesses, and the analysis includes the identification results of the candidate's implicit skills; The candidate ranking list is arranged in descending order of matching degree and supports displaying the core feature comparison of the top 10-20 candidates. According to claim 1, the human resource management method based on artificial intelligence is characterized in that the output of the turnover risk early warning sub-model includes: The risk level of employee turnover is divided into three levels: low, medium, and high. Risk cause analysis should clearly indicate the specific values and percentages of risk characteristics such as attendance abnormalities, decreased communication frequency, and performance decline; Intervention measures recommended for high-risk employees include job adjustments, salary negotiations, and psychological counseling.
6. An artificial intelligence-based human resource management system, characterized in that, include: The data acquisition module includes a resume import interface, an interview video recording unit, an attendance system integration unit, and a communication record capture unit, which is used to collect multi-dimensional data on candidates, employees, and positions. The data preprocessing module includes a structured data cleaning unit, an NLP processing subunit, and a CV processing subunit, which are used to transform the collected data into standardized feature vectors. The AI model library module stores multimodal collaborative Transformer models and sub-models for various scenarios, and provides model calling and training interfaces; The scenario-based management module includes sub-modules for recruitment management, training management, performance management, and employee relations, which call the corresponding scenario sub-models to implement functions. The engineering support module includes a scenario-based model configuration platform, a multimodal data access SDK, and a model performance monitoring panel, which are used for the implementation and maintenance of multimodal collaborative Transformer models.