AI chat robot control system and method for human resources

Abstract

The invention discloses an AI chat robot control system and method for human resources, and the method comprises the following steps: S1, obtaining and preprocessing recruitment data, and forming a recruitment data set; s2, performing semantic analysis by using a hierarchical Transform model, and generating position-job seeker feature representation; s3, searching and optimizing the Transform model through a neural architecture to obtain an optimized model; s4, calculating position matching scores based on the optimization model, and generating a candidate sorting list in combination with the knowledge graph; s5, calling the MX chat robot to perform intelligent interview on the candidates, dynamically adjusting the questioning difficulty, and obtaining interview data; s6, fusing the job seeker interview score and the position matching score to generate a final candidate recommendation list; and S7, generating a recruitment decision support report based on the recommendation list. According to the invention, the intelligent recruitment method of hierarchical Transform, neural architecture search optimization and MX chat robot cooperative work is adopted, and accurate optimization of position matching and automatic decision support of job seeker interview evaluation are realized.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an AI chat robot control system and method for human resources. Background Art

[0002] In the current field of intelligent recruitment technology, job matching and job applicant screening are one of the most critical links in the recruitment process. Traditional recruitment systems usually rely on job recommendation methods based on keyword matching or simple rule engines. These methods usually match the keywords in the job applicant's resume with the requirements in the job description to generate a recommendation list. However, this method has major limitations, mainly reflected in insufficient semantic understanding ability, low job matching accuracy, and incomplete job applicant portraits. In addition, the traditional recruitment process relies on manual screening and interview decisions, resulting in low recruitment efficiency and difficulty in quickly screening out the best candidates for the job in a large-scale talent pool.

[0003] In terms of job matching, existing technologies mainly rely on job applicant screening methods based on rules or shallow machine learning models. These methods usually cannot fully understand the deep semantic relationship between job descriptions and job applicant resumes, and it is difficult to explore the potential match between job requirements and job applicant capabilities. For example, a job applicant may have skills that meet the job requirements, but because specific keywords are not used in the resume, traditional matching algorithms may mistakenly lower their match scores. In addition, many job matching systems ignore the importance weight of the skills required for the position, resulting in recommended job applicants that do not meet the recruitment requirements, affecting recruitment accuracy and job search success rate.

[0004] In recent years, deep learning-based recruitment recommendation systems have been gradually applied to corporate recruitment, using neural network models to vectorize and model job positions and job seeker features to improve matching accuracy. However, existing deep learning recruitment systems still have deficiencies in model architecture optimization. For example, most systems use deep neural networks with fixed structures and cannot adaptively adjust the model structure according to the characteristics of recruitment data, resulting in limited model performance. In addition, the hyperparameter tuning of existing systems mostly relies on manual experiments or grid search methods, lacks automated optimization mechanisms, and is difficult to maintain efficiency and adaptability in a constantly changing recruitment environment.

[0005] In the aspect of intelligent interviews, existing recruitment systems mainly rely on traditional manual interviews or rule-based automated question-and-answer systems. These methods cannot dynamically adjust the interview process and are also difficult to optimize interview questions based on the real-time performance of job seekers. Current intelligent interview systems usually adopt a preset question set and lack the ability to dynamically adjust interview questions based on job seekers' resumes and real-time feedback, resulting in a rigid interview process and making it difficult to effectively evaluate the comprehensive abilities of job seekers. In addition, traditional interview scoring methods often rely solely on the subjective judgment of interviewers and lack an objective multi-modal scoring mechanism, leading to lower stability and fairness of interview results.

[0006] In the aspect of the integration of job matching and interview scoring, existing recruitment systems usually calculate job matching scores and interview scores separately, lacking a unified scoring framework and making it difficult to reasonably integrate multi-dimensional information. Current systems mainly conduct preliminary screening based on job matching scores, while interview scoring serves as a subsequent evaluation link, resulting in a lack of end-to-end optimization in the entire recruitment process and affecting the accuracy of the final recruitment decision. At the same time, most existing scoring integration methods adopt simple weighted averaging or linear combination, failing to fully consider the weight differences in job requirements and the impact of job seekers' historical interview performances on scoring.

[0007] In the aspect of recruitment decision support, existing recruitment systems usually only provide a list of job recommendations and lack systematic decision support tools. For example, most systems only recommend candidates based on job matching scores, ignoring key factors such as job seekers' interview performances, career development potential, and prediction of recruitment success rates. In addition, the utilization degree of recruitment data by existing systems is relatively low, and they fail to conduct in-depth analysis and prediction based on historical recruitment data, resulting in recruitment decisions lacking data support and HR still having to rely on experience for the final screening, reducing the scientific and automated levels of recruitment.

[0008] Therefore, how to provide an AI chatbot control system and method for human resources is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0009] An object of the present invention is to propose an AI chatbot control system and method for human resources. Through hierarchical semantic analysis, neural architecture optimization, and intelligent interview interaction methods, the present invention realizes the precise optimization of job matching and automated support for job seeker evaluation. By adopting the Transformer model combined with neural architecture search technology, it deeply extracts features from job descriptions and job seekers' resumes and enhances the accuracy of job matching based on a knowledge graph. At the same time, combined with the MX chatbot, it realizes dynamic intelligent interviews, automatically adjusts interview questions, and real-time analyzes the text, voice, and video performances of job seekers to provide comprehensive interview scores. This method has the advantages of high matching degree, intelligent interaction, optimized recommendation, and automatic decision-making, providing an innovative solution for intelligent recruitment.

[0010] An AI chatbot control method for human resources according to an embodiment of the present invention includes the following steps:

[0011] S1. Obtain and preprocess recruitment data to obtain a recruitment data set;

[0012] S2. Use a hierarchical Transformer model to perform hierarchical semantic analysis on the recruitment data set to form job-seeker feature representations;

[0013] S3. Optimize the network structure of the hierarchical Transformer model through neural architecture search to obtain an optimized Transformer model;

[0014] S4. Based on the optimized Transformer model, perform job matching on the job-seeker feature representations, calculate the similarity between the job feature vector and the job-seeker feature vector, and enhance the matching result by combining a knowledge graph, and generate a candidate ranking list according to the matching score;

[0015] S5. Invoke the MX chatbot to conduct intelligent interviews on the job-seekers in the candidate ranking list, generate personalized interview questions by combining the optimized Transformer model, dynamically adjust the questioning difficulty according to the job-seekers' answers, and obtain job-seeker interview data;

[0016] S6. Based on the job-seeker interview data, calculate the comprehensive interview score of the job-seeker, and combine the interview score with the job matching score in the candidate ranking list to generate a final candidate recommendation list;

[0017] S7. Generate a recruitment decision support report based on the final candidate recommendation list.

[0018] Optionally, the S3 specifically includes:

[0019] S31. Define the search space of neural architecture search, set the search range of the structural parameters of the hierarchical Transformer model, and set the search parameter set;

[0020] S32. Construct a search objective function for the Transformer model to optimize the ability of job-seeker portrait analysis, and define the objective optimization function:

[0021]

[0022] Among them, Θ represents the search parameter set, F(Θ) represents the target optimization function, N represents the number of job samples evaluated by the Transformer model, i represents the i-th job sample, Acc(·) represents the job matching accuracy rate, Eff(·) represents the computing efficiency, Comp(·) represents the model complexity, and ω1, ω2, ω3 are adjustment weight coefficients;

[0023] S33. Based on the neural architecture search strategy, sample the parameter combinations in the search space and adopt the policy gradient optimization method:

[0024]

[0025] Among them, Θ * represents the optimal Transformer model parameters obtained by NAS search, η is the learning rate, T is the number of search rounds, Θ t represents the parameter, λ is the search step adjustment factor, and t represents the t-th round;

[0026] S34. Based on the optimized Transformer model, perform feature encoding on the job-seeker feature representation and calculate the job matching score through the multi-layer attention of the Transformer:

[0027]

[0028] Among them, S(P,C) represents the matching score between job P and job seeker C, L represents the number of layers of the Transformer model, l represents the l-th layer, H represents the number of attention heads of the multi-head self-attention mechanism, h represents the h-th head, is the query matrix of the l-th layer and h-th head, is the key matrix of the l-th layer and h-th head, is the value matrix of the l-th layer and h-th head, d is the dimension of the hidden layer, ω is the dynamic adjustment coefficient, M represents the total number of candidate job seekers, m represents the m-th job seeker, γ represents the temperature scaling parameter, represents the average matching score, S n represents the matching scores of all job seekers, W m represents the global weight of the m-th job seeker, and ∈ represents the numerical stability factor;

[0029] S35. Based on the trained and optimized Transformer model, calculate the final job matching score using the job feature vector and the job seeker feature vector, and define the final matching scoring function:

[0030]

[0031] Among them, S *(P, C) is the finally calculated job matching score, where P is the job, C is the job seeker, α is the matching score adjustment coefficient, v is the feature vector, β is the regularization parameter, τ is the normalization coefficient, and H is the number of multi-head attention heads.

[0032] Optionally, S4 specifically includes:

[0033] S41. Calculate the job feature vector and the job seeker feature vector based on the optimized Transformer model, and construct a job matching score matrix;

[0034] S42. Optimize the job matching score matrix based on multi-scale attention and combined with the global attention enhancement factor;

[0035] S43. Calculate the job matching similarity, and adjust the matching score using the Euclidean distance and information entropy:

[0036]

[0037] Among them, S * (P, C) is the finally calculated job matching score, where P is the job, C is the job seeker, α is the matching score adjustment coefficient, v is the feature vector, β is the regularization parameter, τ is the normalization coefficient, H is the number of multi-head attention heads, and M' P,C is the optimized job matching score matrix, M represents the total number of candidate job seekers, m represents the m-th job seeker, and p m Probability distribution of job seekers in different category information;

[0038] S44. Calculate the candidate ranking based on the final job matching score, and form a candidate ranking list:

[0039]

[0040] Among them, C * is the finally recommended candidate ranking list, is to sort the job seekers with the highest matching score for operation selection;

[0041] S45. Optimize the job matching in combination with the knowledge graph, and calculate the optimized candidate ranking list based on the Graph Convolutional Network.

[0042] Optionally, S5 specifically includes:

[0043] S51. Call the MX chatbot for an intelligent interview based on the candidate ranking list, and construct an interview interaction information matrix;

[0044] S52. The feature extraction layer based on the Transformer model calculates the text response feature vector of the job seeker and generates the text response score in combination with the attention mechanism;

[0045] S53. Calculate the speech performance score of the job seeker based on speech emotion analysis and optimize it in combination with spectral features:

[0046]

[0047] Among them, S V (C) is the speech performance score of the job seeker, C is the job seeker, T is the number of interview questions, j is the index of the interview question, MFCC(·) is the Mel Frequency Cepstral Coefficient, I is the interview interaction information matrix, M is the number of candidate job seekers, n is the index of all job seekers, γ is the temperature scaling factor, and A is the speech emotion intensity of the job seeker when answering the question;

[0048] S54. Calculate the video performance score of the job seeker based on facial expression analysis and optimize it in combination with temporal information:

[0049]

[0050] Among them, S F (C) is the video performance score of the job seeker, F(t) is the facial emotion intensity of the job seeker at time step t, and W m is the weighting coefficient of the job seeker in facial emotion analysis;

[0051] S55. Integrate the text, speech, and video scores to calculate the final interview score of the job seeker and optimize the score in combination with the job seeker's historical performance:

[0052]

[0053] Among them, is the final interview data matrix of all job seekers, and S T (C) is the text response score of the job seeker, α1, α2, α3, α4 are weight factors, K is the number of the job seeker's historical interview data, and H k is the score of the job seeker's historical interview k.

[0054] Optionally, the specific content of S6 includes:

[0055] S61. Calculate the comprehensive score of text, speech, and video based on the job seeker's interview data and construct an interview score matrix;

[0056] S62. Calculate the weighted integration of the job matching score and the interview score and introduce a dynamic weight adjustment mechanism;

[0057] S63. Calculate the optimal job - job seeker matching assignment based on the optimal transport theory:

[0058]

[0059] Among them, B * is the optimal job - job seeker assignment matrix, B is the job - job seeker assignment matrix, is the sorting of job - job seeker pairs with the lowest matching score for operation selection, P is the job, C is the job seeker, π ij is the assignment weight between the job and the job seeker, v is the feature vector, λ is the weight factor, γ is the temperature scaling factor, K is the number of the job seeker's historical interview data, H k is the score of the job seeker's historical interview k;

[0060] S64. Calculate the final candidate recommendation score based on multi - objective optimization and generate a candidate recommendation list:

[0061]

[0062] Among them, S R (C) is the final candidate recommendation score, S C (P,C) is the comprehensive score of the job seeker and the job, is the sorting of job seekers with the highest recommendation score for operation selection, ρ1, ρ2, ρ3 are the weight coefficients, τ is the attenuation factor;

[0063] S65. Optimize the final recommendation ranking in combination with the knowledge graph and generate a final candidate recommendation list.

[0064] Optionally, the S7 specifically includes:

[0065] S71. Based on the final candidate recommendation list, construct a recruitment decision - making support matrix and calculate the candidate decision feature vector;

[0066] S72. Based on the recruitment decision - making support matrix, use the hierarchical analysis method to calculate the candidate comprehensive evaluation score and obtain the final comprehensive evaluation score of the job seeker;

[0067] S73. Calculate the prediction of the recruitment success rate, optimize the recruitment decision - making support, and use the Bayesian inference method to optimize the analysis of the advantages and disadvantages of candidates to obtain the evaluation of the job seeker's career development potential;

[0068] S75. Combine the recruitment decision - making support matrix, the final comprehensive evaluation score of the job seeker, the predicted probability of being hired, and the evaluation of career development potential to generate a recruitment decision report and provide optimization suggestions.

[0069] The AI chatbot control system for human resources according to the embodiment of the present invention includes the following modules:

[0070] A data collection module, which is used to collect job descriptions, job seeker resumes, recruitment history data, and interview feedback data, construct a job-skill-experience relationship network, and generate a recruitment dataset;

[0071] A data preprocessing module, which is used to perform text cleaning, standardization, and data denoising on the recruitment dataset, construct structured representations of jobs and job seekers by combining multi-level feature extraction methods, and generate job-seeker feature representations using a feature mapping method based on hierarchical Transformer;

[0072] A hierarchical Transformer analysis module, which is used to perform semantic analysis at the industry level, job level, skill level, and experience level on the job-seeker feature representations, extract multi-dimensional features, and calculate job matching scores using a multi-head self-attention mechanism;

[0073] A neural architecture search optimization module, which is used to optimize the network structure of the hierarchical Transformer model using neural architecture search, adjust hyperparameters, and select the optimal model architecture;

[0074] A job matching calculation module, which is used to calculate the matching scores between jobs and job seekers based on the optimized Transformer model, calculate job matching similarities by combining the job-skill-experience relationship graph, and generate a sorted candidate list through an optimization algorithm;

[0075] The MX chatbot interview module is used to call the MX chatbot to conduct intelligent interviews on job seekers in the sorted candidate list, dynamically adjust questions, construct an interview interaction information matrix, and record the text answers, voice expressions, and video behavior data of job seekers in real time;

[0076] An interview scoring calculation module, which is used to calculate the final interview score based on the text answer score, voice performance score, and video performance score of job seekers, and construct a job seeker interview data matrix by combining historical interview data;

[0077] A job seeker recommendation optimization module, which is used to fuse the job matching score and the interview score, calculate the comprehensive candidate score, and optimize the job-seeker matching based on the optimal transport theory to adjust the recommended sorted list;

[0078] A knowledge graph optimization module, which is used to optimize the candidate recommendation list by combining the job-skill-experience relationship graph and adjust the candidate ranking through a graph neural network;

[0079] A recruitment decision support module, which is used to generate a recruitment decision report based on the final matching score, interview score of job seekers, and historical recruitment data, and provide candidate strength and weakness analysis, career development prediction, and recruitment success rate prediction;

[0080] A storage and feedback module, which is used to store job-seeker matching data, interview data, recruitment decision data, and job-seeker feedback data, support the continuous optimization of the model, and improve the automation and intelligence level of the recruitment process through a feedback loop.

[0081] The beneficial effects of the present invention are as follows:

[0082] First of all, through hierarchical Transformer for multi-level semantic analysis of job descriptions and job-seeker resumes, and combining neural architecture search to optimize the model structure, the present invention realizes the precise optimization of job matching. Compared with the traditional recruitment system based on keyword matching, the present invention can deeply explore the semantic relationship between the job and the job-seeker, and use the knowledge graph to enhance the relevance of job-seeker features, making job recommendations more accurate and intelligent, and effectively solving the problems of low matching accuracy and insufficient job-seeker portrait description in the traditional system. In addition, the introduction of neural architecture search enables the Transformer model to dynamically adjust the structure according to the characteristics of recruitment data, improving the adaptability of the model in different recruitment scenarios and avoiding the problem of insufficient generalization ability caused by the traditional deep learning method relying on a fixed architecture.

[0083] Secondly, the present invention conducts intelligent interviews through the MX chatbot, combining the text, voice, and video features of the job-seeker to achieve multi-modal evaluation of the job-seeker. The MX chatbot can not only dynamically adjust interview questions based on the optimized HT model, but also adaptively adjust the interview difficulty according to the real-time performance of the job-seeker, improving the pertinence and intelligence level of the interview interaction. Different from the existing automatic interview system based on a fixed question set, the intelligent interview of the present invention can dynamically adjust the questioning strategy according to the performance of the job-seeker, so as to more comprehensively evaluate the comprehensive ability of the job-seeker, avoiding the limitations of the traditional system in job-seeker evaluation. At the same time, combined with multi-modal feature analysis, the system can calculate the final interview score of the job-seeker, improving the objectivity and consistency of the interview results and solving the problem that manual interviews are easily affected by subjective factors.

[0084] In addition, the present invention adopts a multi-objective optimization strategy to fuse the job matching score and the interview score, and optimizes the job-seeker matching assignment based on the optimal transport theory, so that the finally recommended candidates not only meet the job requirements but also achieve the best matching degree in terms of interview performance. Compared with the traditional recruitment system that only relies on the job matching score to screen candidates, the present invention can comprehensively consider the interview performance and historical recruitment data of the job-seeker, construct a more reasonable candidate recommendation list, and improve the accuracy and matching quality of recruitment. At the same time, the system optimizes the final recommendation ranking through the knowledge graph, so that the recommended candidates not only match the current job but also have good career development potential, helping enterprises optimize talent reserves and improve the recruitment success rate.

[0085] In terms of recruitment decision - making support, the present invention constructs a decision - making support matrix based on recruitment data, combines job requirements, job - seeker portraits, and recruitment success rate prediction to generate an intelligent recruitment decision - making report, which helps HR quickly screen highly - matched candidates and optimize the recruitment process. Compared with traditional recruitment systems that only provide a list of candidate recommendations, the present invention can provide more comprehensive data support for HR, including the job - matching degree of job - seekers, interview performance, historical job - seeking experiences, and predicted recruitment success rate, etc., enhancing the scientificity and accuracy of recruitment decisions. By optimizing the matching strategy through the optimal transport theory, the system can quickly adjust the matching results in the case of changing recruitment requirements, improving the flexibility and adaptability of enterprise recruitment, and overcoming the problem of the lack of optimization mechanism in the existing technology in a dynamic recruitment environment.

[0086] Finally, the present invention combines the historical interview performance of job - seekers, uses multi - objective optimization methods and knowledge graphs to optimize the candidate recommendation ranking, and realizes an efficient and accurate intelligent recruitment process. Compared with the manual screening and rule - matching methods of traditional recruitment systems, the intelligent recruitment method of the present invention can automatically complete job matching, job - seeker evaluation, interview interaction, and recruitment decision - making support, significantly improving recruitment efficiency, reducing the workload of HR, and enhancing the accuracy and fairness of recruitment decisions. Through continuous adaptive optimization, the system can operate efficiently under changing recruitment requirements and job - seeker characteristics, providing a more intelligent talent screening and recruitment solution for enterprises. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used in conjunction with the embodiments of the present invention to explain the present invention, but do not constitute a limitation to the present invention. In the drawings:

[0088] Figure 1 is a flowchart of the AI chat - robot control method for human resources proposed by the present invention;

[0089] Figure 2 is a schematic diagram of the AI chat - robot control system for human resources proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0090] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic way, so they only show the components related to the present invention.

[0091] Reference Figure 1 , the AI chat - robot control method for human resources includes the following steps:

[0092] S1. Obtain and pre - process recruitment data to obtain a recruitment data set;

[0093] S2. Perform hierarchical semantic analysis on the recruitment dataset using a hierarchical Transformer model to form job - candidate feature representations;

[0094] S3. Optimize the network structure of the hierarchical Transformer model through neural architecture search to obtain an optimized Transformer model;

[0095] S4. Based on the optimized Transformer model, perform job matching on the job - candidate feature representations, calculate the similarity between the job feature vector and the candidate feature vector, and enhance the matching result by combining with a knowledge graph. Generate a candidate ranking list according to the matching scores;

[0096] S5. Invoke the MX chatbot to conduct intelligent interviews with the candidates in the candidate ranking list, generate personalized interview questions by combining with the optimized Transformer model, and dynamically adjust the question - asking difficulty according to the candidates' answers to obtain candidate interview data;

[0097] S6. Based on the candidate interview data, calculate the comprehensive interview score of the candidates, and combine the interview score with the job matching score in the candidate ranking list to generate a final candidate recommendation list;

[0098] S7. Generate a recruitment decision - making support report based on the final candidate recommendation list.

[0099] The AI chatbot control method for human resources provided by the present invention can improve the intelligent level of the recruitment process and the accuracy of job matching. By using a hierarchical Transformer model for semantic analysis and combining neural architecture search to optimize the model structure, job matching is made more accurate, avoiding the limitations of traditional keyword - based matching. At the same time, using a knowledge graph to enhance job - candidate matching ensures that the recommendation results are more in line with recruitment requirements. The MX chatbot of the present invention can dynamically generate personalized interview questions and adjust the question - asking difficulty according to the real - time performance of the candidates, making the interview more flexible and accurate. Combining text, voice, and video analysis, the system calculates the comprehensive interview score and integrates the job matching score to achieve accurate candidate recommendation. In addition, the system generates a recruitment decision - making support report to provide intelligent analysis for HR and improve the scientific nature of recruitment decisions. Overall, the present invention can optimize job matching, improve interview efficiency, shorten the recruitment cycle, and enhance the recruitment success rate, providing an efficient and intelligent talent recruitment solution for enterprises.

[0100] In this embodiment, the specific content of S3 includes:

[0101] S31. Define the search space of neural architecture search, set the search range of the structural parameters of the hierarchical Transformer model, and set the search parameter set;

[0102] S32. Construct a Transformer model search objective function to optimize the ability of job seeker portrait analysis, and define the objective optimization function:

[0103]

[0104] Among them, Θ represents the set of search parameters, F(Θ) represents the objective optimization function, N represents the number of job samples evaluated by the Transformer model, i represents the i-th job sample, Acc(·) represents the job matching accuracy, Eff(·) represents the computing efficiency, Comp(·) represents the model complexity, and ω1, ω2, ω3 are adjustment weight coefficients;

[0105] S33. Based on the neural architecture search strategy, sample the parameter combinations in the search space and adopt the policy gradient optimization method:

[0106]

[0107] Among them, Θ * represents the optimal Transformer model parameters obtained by NAS search, η is the learning rate, T is the number of search rounds, Θ t represents the parameter, λ is the search step adjustment factor, and t represents the t-th round;

[0108] S34. Based on the optimized Transformer model, perform feature encoding on the job-seeker feature representation and calculate the job matching score through the multi-layer attention of the Transformer:

[0109]

[0110] Among them, S(P,C) represents the matching score between job P and job seeker C, L represents the number of layers of the Transformer model, l represents the l-th layer, H represents the number of attention heads of the multi-head self-attention mechanism, h represents the h-th head, is the query matrix of the l-th layer and h-th head, is the key matrix of the l-th layer and h-th head, is the value matrix of the l-th layer and h-th head, d is the dimension of the hidden layer, ω is the dynamic adjustment coefficient, M represents the total number of candidate job seekers, m represents the m-th job seeker, γ represents the temperature scaling parameter, represents the average matching score, S n represents the matching scores of all job seekers, W m represents the global weight of the m-th job seeker, and ∈ represents the numerical stability factor;

[0111] S35. Based on the optimized Transformer model through training, calculate the final job matching score using the job feature vector and the job seeker feature vector, and define the final matching scoring function:

[0112]

[0113] Among them, S * (P, C) is the finally calculated job matching score, P is the job, C is the job seeker, α is the matching score adjustment coefficient, v is the feature vector, β is the regularization parameter, τ is the normalization coefficient, and H is the number of multi-head attention heads.

[0114] In the present invention, the hierarchical Transformer model is optimized through neural architecture search to improve the accuracy and computational efficiency of job matching. By setting the search space and optimizing the search strategy, the automatic adjustment of model parameters is realized to optimize the job seeker portrait analysis ability. Based on the policy gradient optimization method, the optimal model parameters are efficiently searched to ensure the stability and reliability of job matching. In the process of job matching calculation, a multi-layer self-attention mechanism is adopted and combined with global attention weight adjustment to make the job and job seeker feature representations more accurate. The matching score is adjusted by temperature scaling to ensure computational stability, and a numerical stability factor is introduced to reduce computational errors. In the calculation of the final job matching score, the distance metric between the job and job seeker feature vectors is combined to make the matching result more reasonable and ensure the scientific nature of the recommended candidate ranking.

[0115] In this embodiment, the S4 specifically includes:

[0116] S41. Calculate the job feature vector and the job seeker feature vector based on the optimized Transformer model, and construct a job matching score matrix;

[0117] S42. Optimize the job matching score matrix based on multi-scale attention and in combination with the global attention enhancement factor;

[0118] S43. Calculate the job matching similarity, and adjust the matching score using the Euclidean distance and information entropy:

[0119]

[0120] Among them, S * (P, C) is the finally calculated job matching score, P is the job, C is the job seeker, α is the matching score adjustment coefficient, v is the feature vector, β is the regularization parameter, τ is the normalization coefficient, H is the number of multi-head attention heads, and M' P,C is the optimized job matching score matrix, M represents the total number of candidate job seekers, m represents the m-th job seeker, and p m is the probability distribution of the job seeker in different category information;

[0121] S44. Calculate the candidate ranking based on the final job matching score and form a candidate ranking list:

[0122]

[0123] where C * is the candidate ranking list of the finally recommended candidates, is the ranking of the job seeker with the highest operation selection matching score;

[0124] S45. Optimize job matching in combination with the knowledge graph and calculate the optimized candidate ranking list based on the Graph Convolutional Network.

[0125] In the present invention, the feature vectors of the job and the job seeker are calculated through the optimized Transformer model, a job matching score matrix is constructed, and the matching accuracy is improved by combining the multi-scale attention and the global attention enhancement factor. The Euclidean distance and information entropy are used to adjust the matching score to ensure that the matching calculation not only considers the feature similarity but also optimizes the information distribution weight, making the score more reasonable. In the candidate ranking stage, the present invention generates a candidate ranking list based on the final job matching score to ensure that the recommended job seekers have the highest matching degree. Further, the matching relationship is optimized in combination with the knowledge graph, and the graph convolutional network is used to optimize the ranking, making the job-job seeker matching more accurate. Through the multi-layer optimization strategy, the job matching score is made more stable, avoiding the limitations of single-feature matching.

[0126] In this embodiment, the S5 specifically includes:

[0127] S51. Call the MX chatbot for an intelligent interview based on the candidate ranking list and construct an interview interaction information matrix;

[0128] S52. Calculate the text answer feature vector of the job seeker based on the feature extraction layer of the Transformer model and generate a text answer score in combination with the attention mechanism;

[0129] S53. Calculate the voice performance score of the job seeker based on voice emotion analysis and optimize it in combination with spectral features:

[0130]

[0131] where S V (C) is the voice performance score of the job seeker, C is the job seeker, T is the number of interview questions, j is the index of the interview question, MFCC(·) is the Mel Frequency Cepstral Coefficient, I is the interview interaction information matrix, M is the number of candidate job seekers, n is the index of all job seekers, γ is the temperature scaling factor, and A is the voice emotion intensity of the job seeker when answering the question;

[0132] S54. Calculate the video performance score of the job applicant based on facial expression analysis and optimize it by combining temporal information:

[0133]

[0134] Among them, S F (C) is the video performance score of the job applicant, F(t) is the facial emotion intensity of the job applicant at time step t, and W m is the weighting coefficient of the job applicant in facial emotion analysis;

[0135] S55. Calculate the final interview score of the job applicant by integrating text, voice, and video scores, and optimize the score by combining the historical performance of the job applicant:

[0136]

[0137] Among them, is the final interview data matrix of all job applicants, and S T (C) is the text answer score of the job applicant, α1, α2, α3, α4 are weight factors, K is the number of historical interview data of the job applicant, and H k is the score of the kth historical interview of the job applicant.

[0138] Through the MX chatbot intelligent interview, the present invention combines text, voice, and video multimodal analysis to achieve a comprehensive evaluation of job applicants. The Transformer model is used to extract text answer features, and the attention mechanism is combined to generate text answer scores, improving the ability to understand interview questions. The voice performance score of the job applicant is calculated through Mel Frequency Cepstral Coefficients and voice emotion analysis, and optimized by combining spectral features to make voice evaluation more accurate. The video performance score uses facial expression analysis and optimizes the detection of the emotional fluctuations of the job applicant by combining temporal information, improving the stability of interview performance evaluation. The final score integrates text, voice, and video data, and is optimized by combining the historical interview performance of the job applicant to ensure that the scoring system takes into account both current capabilities and historical trends. Integrating multimodal information to optimize interview scores makes interview recommendation results more scientific and improves the reliability of recruitment decisions.

[0139] In this embodiment, the S6 specifically includes:

[0140] S61. Calculate the comprehensive scores of text, voice, and video based on the interview data of the job applicant and construct an interview score matrix;

[0141] S62. Calculate the weighted integration of the job matching score and the interview score, and introduce a dynamic weight adjustment mechanism;

[0142] S63. Calculate the optimal job-applicant matching assignment based on the optimal transport theory:

[0143]

[0144] Among them, B * is the optimal position-job seeker allocation matrix, B is the position-job seeker allocation matrix, is the ranking of the position-job seeker with the lowest matching score for operation selection, P is the position, C is the job seeker, and π ij is the allocation weight between the position and the job seeker, v is the eigenvector, λ is the weight factor, γ is the temperature scaling factor, K is the number of historical interview data of the job seeker, and H k is the score of the k-th historical interview of the job seeker;

[0145] S64. Calculate the final candidate recommendation score based on multi-objective optimization and generate a candidate recommendation list:

[0146]

[0147] Among them, S R (C) is the final candidate recommendation score, and S C (P, C) is the comprehensive score of the job seeker and the position, is the ranking of the job seeker with the highest recommendation score for operation selection, ρ1, ρ2, and ρ3 are weight coefficients, and τ is the attenuation factor;

[0148] S65. Optimize the final recommendation ranking in combination with the knowledge graph and generate a final candidate recommendation list.

[0149] In the present invention, by fusing the comprehensive scores of the text, voice, and video of the job seeker, constructing an interview scoring matrix, and performing weighted fusion in combination with the position matching score, the dynamic optimization of the position-job seeker matching is realized. The optimal transport theory is used to calculate the optimal position-job seeker allocation matrix to make the matching allocation more accurate, and the final candidate recommendation score is calculated based on the multi-objective optimization strategy to ensure the best recommendation result. Through the dynamic weight adjustment mechanism, the weights of the position matching and the interview score are adaptively adjusted, so that the system can be optimized according to different recruitment scenarios, improving the flexibility and adaptability of the recommendation. In addition, the final candidate ranking is optimized in combination with the knowledge graph to ensure that the recommendation result better meets the enterprise recruitment needs.

[0150] In this embodiment, the S7 specifically includes:

[0151] S71. Based on the final candidate recommendation list, construct a recruitment decision support matrix and calculate the candidate decision eigenvector;

[0152] S72. Based on the recruitment decision support matrix, use the hierarchical analysis method to calculate the comprehensive evaluation score of the candidate and obtain the final comprehensive evaluation score of the job seeker;

[0153] S73. Calculate the recruitment success rate prediction, optimize the recruitment decision support, and use the Bayesian inference method to optimize the analysis of the advantages and disadvantages of candidates, so as to obtain the evaluation of the career development potential of job seekers;

[0154] S75. Generate a recruitment decision report by combining the recruitment decision support matrix, the final comprehensive evaluation score of job seekers, the predicted probability of being hired, and the evaluation of career development potential, and provide optimization suggestions.

[0155] Through constructing a recruitment decision support matrix, calculating the candidate decision feature vector, and using the hierarchical analysis method to calculate the final comprehensive evaluation score, the present invention ensures the comprehensiveness and scientificity of candidate evaluation. Based on the recruitment success rate prediction, combined with the Bayesian inference method to optimize the analysis of the advantages and disadvantages of candidates, accurately evaluate the career development potential of job seekers, and improve the rationality of recruitment decisions. In addition, the present invention integrates the comprehensive evaluation score of job seekers, the predicted probability of being hired, and the career development potential to generate a recruitment decision report and provide optimization suggestions, enabling HR to make more accurate recruitment decisions based on data-driven.

[0156] Reference Figure 2 , an AI chatbot control system for human resources, including the following modules:

[0157] A data collection module, used to collect job descriptions, job seeker resumes, recruitment historical data, and interview feedback data, construct a job-skill-experience relationship network, and generate a recruitment dataset;

[0158] A data preprocessing module, used to perform text cleaning, standardization, and data denoising on the recruitment dataset, construct a structured representation of the job and the job seeker in combination with a multi-level feature extraction method, and generate a job-job seeker feature representation using a feature mapping method based on hierarchical Transformer;

[0159] A hierarchical Transformer analysis module, used to perform semantic analysis at the industry level, job level, skill level, and experience level on the job-job seeker feature representation, extract multi-dimensional features, and calculate the job matching score using a multi-head self-attention mechanism;

[0160] A neural architecture search optimization module, used to optimize the network structure of the hierarchical Transformer model using neural architecture search, adjust hyperparameters, and select the optimal model architecture;

[0161] A job matching calculation module, used to calculate the matching score between the job and the job seeker based on the optimized Transformer model, calculate the job matching similarity in combination with the job-skill-experience relationship graph, and generate a sorted candidate list through an optimization algorithm;

[0162] The MX Chatbot Interview Module is used to call the MX Chatbot to conduct an intelligent interview with job seekers in the candidate sorted list, dynamically adjust questions, construct an interview interaction information matrix, and record the text answers, voice expressions, and video behavior data of job seekers in real time;

[0163] The Interview Scoring Calculation Module is used to calculate the final interview score based on the text answer score, voice performance score, and video performance score of job seekers, and construct a job seeker interview data matrix in combination with historical interview data;

[0164] The Job Seeker Recommendation Optimization Module is used to fuse the job matching score and the interview score, calculate the comprehensive candidate score, and optimize the job-seeker matching based on the optimal transport theory to adjust the recommended sorted list;

[0165] The Knowledge Graph Optimization Module is used to optimize the candidate recommendation list in combination with the job-skill-experience relationship graph, and adjust the candidate ranking through a graph neural network;

[0166] The Recruitment Decision Support Module is used to generate a recruitment decision report based on the final matching score, interview score of job seekers, and historical recruitment data, and provide analysis of the advantages and disadvantages of candidates, career development prediction, and recruitment success rate prediction;

[0167] The Storage and Feedback Module is used to store job-seeker matching data, interview data, recruitment decision data, and job seeker feedback data, support the continuous optimization of the model, and improve the automation and intelligence level of the recruitment process through a feedback loop.

[0168] The human resource AI chatbot control system provided by the present invention combines hierarchical Transformers, neural architecture search, and the MX Chatbot to achieve the intelligence, automation, and precise optimization of the entire recruitment process. Through deep learning and intelligent optimization methods, the intelligence and automation of the recruitment process are realized, and the accuracy of job matching and recruitment efficiency are improved. The system uses hierarchical Transformers to perform in-depth semantic analysis on job and job seeker characteristics, and combines neural architecture search to optimize the model structure, making job matching more accurate and avoiding the inefficient screening problems in traditional recruitment methods. Based on multi-modal data analysis, the system not only considers the resume information of job seekers, but also integrates the text, voice, and video performances in the intelligent interview to construct a comprehensive job seeker evaluation system, improving the scientificity of talent screening. The intelligent interview mechanism of the system can dynamically adjust interview questions, and analyze the language, expression, and voice emotion characteristics of job seekers in combination with the attention mechanism to ensure the stability and accuracy of the evaluation results. In addition, the system predicts the recruitment success rate based on historical recruitment data, and combines career development potential analysis to generate a recruitment decision report to assist HR in making more scientific recruitment decisions and optimizing the enterprise's talent acquisition process.

[0169] Example 1:

[0170] To verify the feasibility of the present invention in implementation, the present invention is applied to the talent recruitment system of a large Internet enterprise, which needs to recruit a large number of talents for technical, product, operation and other positions every year. Due to the large demand for talents and long recruitment cycle of this enterprise, HR needs to process a huge amount of resumes and interview arrangements, and at the same time, it is necessary to accurately match suitable candidates to avoid recruitment mistakes caused by inefficient screening and human subjective errors. The traditional recruitment system mainly relies on keyword matching for job recommendations, and the selected candidates often deviate greatly from the job requirements. At the same time, the interview process depends on manual operation, resulting in low efficiency. The recruitment decision highly relies on the experience of HR and lacks objective data support. In addition, during the recruitment process, it is difficult for HR to effectively measure the long-term development potential of job seekers, resulting in poor matching after some positions are hired, and even a high turnover rate. Therefore, this enterprise urgently needs an efficient and intelligent recruitment method to improve the accuracy of job matching, optimize the interview process, and enhance the scientific nature of recruitment decisions.

[0171] In this scenario, the present invention is applied to the enterprise's recruitment system, covering key links such as job matching, intelligent interviews, optimization of candidate recommendations, and recruitment decision support. First, the system constructs a recruitment dataset from the enterprise's existing recruitment data, job descriptions, and job seeker resumes, conducts in-depth semantic analysis of job and job seeker information based on the hierarchical Transformer (HT) model, generates job-seeker feature representations, and enhances the relationship between job and job seeker features by combining a knowledge graph. Subsequently, the HT model is optimized using neural architecture search (NAS) to improve the accuracy of job matching, and the matching score is calculated based on the similarity between the job feature vector and the job seeker feature vector to form a candidate ranking list. Next, the system calls the MX chatbot to conduct intelligent interviews with the top 100 candidates in the ranking, adopts a dynamic question generation mechanism, adjusts the question difficulty according to the job seeker's answer, and records the text, voice, and video performance data of the job seeker in real time. Finally, the system integrates the job matching score and the interview score, uses a multi-objective optimization method to calculate the final candidate recommendation list, and optimizes the recommendation ranking based on the prediction of the recruitment success rate to generate a recruitment decision support report to help HR make more accurate recruitment decisions.

[0172] The system has demonstrated remarkable effects in the actual recruitment process. Taking the recruitment data in the first quarter of 2024 as an example, during the recruitment of 1,500 positions, the system processed a total of 350,000 resumes and screened out 150,000 candidates who met the basic requirements. In the job matching session, based on the optimized model of HT + NAS of the present invention, the accuracy rate of job matching has increased by 24.7%. Compared with the traditional keyword matching method, the matching degree between candidates and positions has increased by 18.3% on average. In the intelligent interview session, the system arranged a total of 40,000 automatic interviews, and the interview efficiency has increased by 64.5%. The average interview time for a single candidate has been shortened to 12 minutes, while the traditional manual interview takes an average of 25 minutes. In the candidate recommendation session, the matching degree between the candidates finally confirmed by the HR for employment and the top 20 candidates recommended by the system has reached 92.5%, which is a significant increase compared to the matching degree (about 75%) of the previous HR manual screening and recommendation. In addition, in terms of the recruitment cycle, it originally took an average of 21 days from resume screening to the final interview. After using the present invention, the average recruitment cycle has been shortened to 12 days, and the overall recruitment efficiency has increased by 42.8%. In the follow-up data one year after employment, the average passing rate of the probation period for employees recruited based on the present invention has increased by 16.2%, and the turnover rate within half a year of employment has decreased by 9.7%, indicating that the intelligent recruitment method of the present invention not only improves the accuracy of talent matching but also enhances the long-term adaptability between employees and positions.

[0173] To more intuitively demonstrate the implementation effects of the present invention, the following are the key recruitment data statistics of a large Internet enterprise before and after applying the present invention in the first quarter of 2024:

[0174] Table 1 Comparison Table of Recruitment Effects before and after the Application of the Intelligent Recruitment System

[0175]

[0176]

[0177] The experimental results show that, with the support of the intelligent recruitment system of the present invention, the accuracy rate of job matching, the efficiency of interview screening, and the accuracy of the final recruitment decision have all been significantly improved. During the recruitment process for multiple positions such as technology, product, and operation, the job matching algorithm optimized by the present invention based on the hierarchical Transformer (HT) and neural architecture search (NAS) has increased the matching degree between job seekers and positions by 18.3%, and the job matching accuracy rate has been increased to 90.5%. In the interview stage, through the automated process of the MX chatbot intelligent interview, the interview efficiency has been increased by 64.5%, the screening speed of candidates has been greatly accelerated, and the interview process has become more fluent. In addition, in the final candidate recommendation link, the matching degree between the candidates confirmed by the HR for employment and the candidates recommended by the system reaches 92.5%, which is 23% higher than the traditional method, fully demonstrating the advantages of the present invention in terms of recruitment accuracy.

[0178] For example, in the large-scale recruitment of technology positions in an Internet company, the traditional recruitment system requires HR to spend a lot of time screening resumes, while the intelligent recruitment system of the present invention conducts in-depth semantic analysis of job and job seeker information through the HT model, significantly improving the automated screening efficiency of job matching. In the intelligent interview session for candidates, the MX chatbot of the present invention can dynamically adjust the question content according to the real-time answers of job seekers to ensure the pertinence of interview questions, enabling a more comprehensive assessment of the abilities of job seekers. In the recruitment of operation positions, the system combines multi-modal data such as voice, text, and video for job seeker performance analysis, making the interview scores more objective, avoiding the subjective errors of traditional manual interviews, and improving the scientific nature of candidate recommendation.

[0179] Through comparative experiments, it is verified that the intelligent recruitment system provided by the present invention has shown significant advantages in job matching, interview evaluation, and recruitment decision optimization. Its job matching model optimized by HT+NAS effectively improves the accuracy of job recommendation. The MX chatbot realizes the automation and intelligence of the interview process, improving the recruitment screening efficiency. The knowledge graph optimization and multi-objective optimization strategy ensure the scientific nature of the final candidate recommendation. The experimental results prove that the present invention can effectively solve the problems of insufficient job matching accuracy, low interview screening efficiency, and recruitment decision relying on experience in traditional recruitment systems, providing an enterprise with a more efficient and intelligent recruitment solution.

[0180] As described above, it is only the preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, should be covered by the protection scope of the present invention.

Claims

1. AI chatbot control method for human resources, characterized in that, It includes the following steps: S1. Obtain and preprocess recruitment data to obtain a recruitment dataset; S2. Use a hierarchical Transformer model to perform hierarchical semantic analysis on the recruitment dataset to form job-seeker feature representations; S3. Optimize the network structure of the hierarchical Transformer model through neural architecture search to obtain an optimized Transformer model; S4. Based on the optimized Transformer model, perform job matching on the job-seeker feature representations, calculate the similarity between the job feature vector and the job-seeker feature vector, and enhance the matching result by combining a knowledge graph. Generate a sorted candidate list according to the matching scores; S5. Call the MX chatbot to conduct intelligent interviews with the job-seekers in the sorted candidate list, generate personalized interview questions by combining the optimized Transformer model, and dynamically adjust the questioning difficulty according to the job-seekers' answers to obtain job-seeker interview data; S6. Based on the job-seeker interview data, calculate the comprehensive interview scores of the job-seekers, and combine the interview scores with the job matching scores in the sorted candidate list to generate a final candidate recommendation list; S7. Generate a recruitment decision support report based on the final candidate recommendation list.

2. The AI chatbot control method for human resources according to claim 1, wherein, The specific content of S3 includes: S31. Define the search space of neural architecture search, set the search range of the structural parameters of the hierarchical Transformer model, and set the search parameter set; S32. Construct a search objective function for the Transformer model to optimize the ability of job-seeker portrait analysis, and define the objective optimization function: Among them, Θ represents the search parameter set, F(Θ) represents the objective optimization function, N represents the number of job samples evaluated by the Transformer model, i represents the i-th job sample, Acc(·) represents the job matching accuracy rate, Eff(·) represents the computing efficiency, Comp(·) represents the model complexity, and ω1, ω2, ω3 are adjustment weight coefficients; S33. Based on the neural architecture search strategy, sample the parameter combinations in the search space, and adopt the policy gradient optimization method; where Θ * represents the optimal Transformer model parameters obtained by NAS, η is the learning rate, T is the number of search rounds, and Θ t represents the parameters, λ is the search step adjustment factor, and t represents the t-th round; S34. Based on the optimized Transformer model, perform feature encoding on the job-seeker feature representations, and calculate the job matching scores through the multi-layer attention of the Transformer; Among them, S(P, C) represents the matching score between position P and job seeker C, L represents the number of layers of the Transformer model, l represents the l-th layer, H represents the number of attention heads of the multi-head self-attention mechanism, h represents the h-th head, is the query matrix of the h-th head in the l-th layer, is the key matrix of the h-th head in the l-th layer, is the value matrix of the h-th head in the l-th layer, d is the dimension of the hidden layer, ω is the dynamic adjustment coefficient, M represents the total number of candidate job seekers, m represents the m-th job seeker, γ is the temperature scaling parameter, represents the average matching score, S n represents the matching scores of all job seekers, W m represents the global weight of the m-th job seeker, ∈ represents the numerical stability factor; S35. Based on the trained and optimized Transformer model, calculate the final job matching score using the job feature vector and the job-seeker feature vector, and define the final matching scoring function; Among them, S * (P, C) is the finally calculated job matching score, where P is the job, C is the job seeker, α is the matching score adjustment coefficient, v is the feature vector, β is the regularization parameter, τ is the normalization coefficient, and H is the number of multi-head attention heads.

3. The AI chatbot control method for human resources according to claim 1, characterized in that The specific content of S4 includes: S41. Calculate the job feature vector and the job-seeker feature vector based on the optimized Transformer model, and construct a job matching scoring matrix; S42. Based on multi-scale attention, combine the global attention enhancement factor to optimize the job matching scoring matrix; S43. Calculate the job matching similarity, and adjust the matching scores using the Euclidean distance and information entropy; Among them, S * (P, C) is the finally calculated job matching score. P is the job, C is the job seeker, α is the matching score adjustment coefficient, v is the feature vector, β is the regularization parameter, τ is the normalization coefficient, H is the number of multi-head attention heads, M' P,C is the optimized job matching score matrix, M represents the total number of candidate job seekers, m represents the m-th job seeker, p m The probability distribution of job seekers in different categories of information; S44. Calculate the candidate rankings based on the final job matching scores and form a sorted candidate list: Among them, C * is the sorted list of the finally recommended candidates, is the sorting of the job seekers with the highest matching score for operation selection; S45. Optimize job matching by integrating a knowledge graph and calculate an optimized candidate ranking list based on the Graph Convolutional Network.

4. The AI chatbot control method for human resources according to claim 1, wherein, The specific steps of S5 are as follows: S51. Invoke the MX chatbot for intelligent interviews based on the candidate ranking list and construct an interview interaction information matrix. S52. Calculate the text response feature vector of the job seeker based on the feature extraction layer of the Transformer model and generate a text response score by combining the attention mechanism. S53. Calculate the voice performance score of the job seeker based on voice emotion analysis and optimize it by combining spectral features. Among them, S V (C) is the speech performance score of the job seeker, C is the job seeker, T is the number of interview questions, j is the index of the interview question, MFCC(·) is the Mel Frequency Cepstral Coefficient, I is the interview interaction information matrix, M is the number of candidate job seekers, n is the index of all job seekers, γ is the temperature scaling factor, and A is the speech emotion intensity of the job seeker when answering questions; S54. Calculate the video performance score of the job seeker based on facial expression analysis and optimize it by combining temporal information. Among them, S F (C) is the video performance score of the job seeker, F(t) is the facial emotion intensity of the job seeker at time step t, and W m is the weighting coefficient of the job seeker in facial emotion analysis; S55. Integrate the text, voice, and video scores to calculate the final interview score of the job seeker and optimize the score by combining the job seeker's historical performance. Among them, is the final interview data matrix for all job seekers, S T (C) is the score for the text answers of the job seeker, α1, α2, α3, α4 are weight factors, K is the number of historical interview data of the job seeker, H k is the score for the k-th historical interview of the job seeker.

5. The AI chatbot control method for human resources according to claim 1, wherein, The specific steps of S6 are as follows: S61. Calculate the comprehensive scores of text, voice, and video based on the job seeker's interview data and construct an interview score matrix. S62. Calculate the weighted fusion of the job matching score and the interview score and introduce a dynamic weight adjustment mechanism. S63. Calculate the optimal job-seeker matching assignment based on the optimal transport theory. Among them, B * is the optimal position-job seeker assignment matrix, B is the position-job seeker assignment matrix, is the ranking of the position-job seeker with the lowest matching score for operation selection, P is the position, C is the job seeker, and π ij is the assignment weight between the position and the job seeker, v is the eigenvector, λ is the weight factor, γ is the temperature scaling factor, K is the number of historical interview data of the job seeker, and H k is the score of the historical interview k of the job seeker; S64. Calculate the final candidate recommendation score based on multi-objective optimization and generate a candidate recommendation list. Among them, S R (C) is the recommended score of the final candidate, S C (P, C) is the comprehensive score of the job seeker and the position, is the ranking of the job seekers with the highest recommended score for operation selection, ρ1, ρ2, ρ3 are weight coefficients, and τ is the attenuation factor; S65. Optimize the final recommendation ranking by integrating the knowledge graph and generate the final candidate recommendation list.

6. The AI chatbot control method for human resources according to claim 1, wherein The specific steps of S7 are as follows: S71. Based on the final candidate recommendation list, construct a recruitment decision support matrix and calculate the candidate decision feature vector. S72. Based on the recruitment decision support matrix, use the hierarchical analysis method to calculate the comprehensive evaluation score of the candidate and obtain the final comprehensive evaluation score of the job seeker. S73. Calculate the prediction of the recruitment success rate, optimize the recruitment decision support, and use the Bayesian inference method to optimize the analysis of the candidate's strengths and weaknesses to obtain the evaluation of the job seeker's career development potential. S75. Generate a recruitment decision report by combining the recruitment decision support matrix, the final comprehensive evaluation score of the job seeker, the predicted probability of being hired, and the career development potential evaluation, and provide optimization suggestions.

7. An AI chatbot control system for human resources, which executes the AI chatbot control method for human resources according to any one of claims 1 to 6, characterized in that, It includes the following modules: Data collection module, which is used to collect job descriptions, job seeker resumes, recruitment historical data, and interview feedback data, construct a job-skill-experience relationship network, and generate a recruitment dataset. Data preprocessing module, which is used to perform text cleaning, standardization, and data denoising on the recruitment dataset, construct a structured representation of jobs and job seekers by combining multi-level feature extraction methods, and generate job-seeker feature representations using a feature mapping method based on hierarchical Transformer. Hierarchical Transformer analysis module, which is used to perform semantic analysis at the industry level, job level, skill level, and experience level on the job-seeker feature representations, extract multi-dimensional features, and calculate the job matching score using the multi-head self-attention mechanism. Neural architecture search optimization module, which is used to optimize the network structure of the hierarchical Transformer model using neural architecture search, adjust hyperparameters, and select the optimal model architecture. A job matching calculation module, which is used to calculate the matching score between a job and a job seeker based on the optimized Transformer model, calculate the job matching similarity in combination with the job-skill-experience relationship graph, and generate a sorted list of candidates through an optimization algorithm; The MX chatbot interview module, which is used to call the MX chatbot to conduct intelligent interviews on the job seekers in the sorted list of candidates, dynamically adjust questions, construct an interview interaction information matrix, and record the text answers, voice expressions and video behavior data of the job seekers in real time; An interview scoring calculation module, which is used to calculate the final interview score based on the text answer score, voice performance score and video performance score of the job seeker, and construct a job seeker interview data matrix in combination with historical interview data; A job seeker recommendation optimization module, which is used to integrate the job matching score and the interview score, calculate the comprehensive score of the candidate, and optimize the job-job seeker matching based on the optimal transport theory to adjust the recommended sorted list; A knowledge graph optimization module, which is used to optimize the candidate recommendation list in combination with the job-skill-experience relationship graph, and adjust the candidate ranking through a graph neural network; A recruitment decision support module, which is used to generate a recruitment decision report based on the final matching score, interview score of the job seeker and historical recruitment data, and provide candidate strength and weakness analysis, career development prediction and recruitment success rate prediction; A storage and feedback module, which is used to store job-job seeker matching data, interview data, recruitment decision data and job seeker feedback data, support the continuous optimization of the model, and improve the automation and intelligence level of the recruitment process through a feedback loop.

Images