Candidate recommendation model construction method and device and terminal equipment

Abstract

The invention discloses a candidate recommendation model construction method and device and terminal equipment, and belongs to the field of artificial intelligence, and the method comprises the steps: calculating the matching confidence coefficient of each resume text according to the resume feature vector of each candidate and the post feature vector of a target post, and taking the resume text with the high matching confidence coefficient as an alternative resume text; performing rule verification on each alternative resume text, and calculating to obtain a rule matching degree; screening out the alternative resume texts of which the matching confidence and the rule matching degree both exceed corresponding thresholds as candidate resume texts; and according to the intention data of each candidate, the historical business transaction data and post feature vector of the target post, and the resume feature vector and rule verification result corresponding to each candidate resume text, training the reinforcement learning model to obtain a candidate recommendation model. By implementing the method and the device, the problems of long time consumption and incomplete talent coverage caused by the fact that the traditional headhunting service depends on manual processing of unstructured texts can be solved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and in particular to a method, apparatus, and terminal device for constructing a candidate recommendation model. Background Technology

[0002] With the increasing demand for mid-to-senior level talent from enterprises, headhunting services have become an important part of the human resources market. Industry data shows that traditional headhunting services typically require screening 50-100 resumes to complete a single mid-to-senior level talent referral, with the entire recruitment cycle lasting 30-60 days. The core reason for this inefficiency is that key data such as resumes and job descriptions are primarily unstructured text. Humans cannot process unstructured text on a large scale, and the efficiency of manually extracting key information from unstructured text is too low. This results in a time-consuming candidate screening process and incomplete talent coverage, failing to meet the needs of enterprises for rapid recruitment of mid-to-senior level talent. Summary of the Invention

[0003] This invention provides a method, apparatus, and terminal device for constructing a candidate recommendation model. The method can solve the problems of long processing time and incomplete talent coverage caused by the reliance on manual processing of unstructured text in traditional headhunting services.

[0004] To address the aforementioned technical problems, one embodiment of the present invention provides a method for constructing a candidate recommendation model, comprising: Obtain historical business transaction data, job feature vectors, and target job rules for the target position, as well as the resume text, resume feature vectors, and intention data of each candidate; Based on the feature vectors of each resume and the feature vector of the job position, the matching confidence of each resume text is calculated, and the resume texts whose matching confidence meets the first screening condition are selected as candidate resume texts; wherein, the first screening condition is: the matching confidence exceeds the preset confidence threshold. For each candidate resume text, rule validation is performed on the candidate resume text according to the target job rules, a rule validation result is generated, and the rule matching degree is calculated based on the rule validation result. Select candidate resume texts whose matching confidence level meets the first screening condition and whose rule matching degree meets the second screening condition; wherein, the second screening condition is: the rule matching degree exceeds the preset matching degree threshold. Based on the intention data, the historical business transaction data, the job feature vector, and the resume feature vector and rule verification results corresponding to each candidate resume text, a preset reinforcement learning model is trained, and the trained reinforcement learning model is used as the candidate recommendation model.

[0005] Further, the step of training a preset reinforcement learning model based on the intention data, the historical business transaction data, the job feature vector, and the resume feature vectors and rule verification results corresponding to each candidate resume text, and using the trained reinforcement learning model as the candidate recommendation model, includes: For each candidate resume text, the job feature vector, the resume feature vector corresponding to the candidate resume text, and the historical recommendation success rate of the job are combined to form a fused feature vector; Based on the fused feature vector, the historical business transaction data, the rule verification results, and the intention data, the training operation is repeatedly performed on the preset reinforcement learning model until the preset number of iterations is reached, and the trained reinforcement learning model is obtained. The trained reinforcement learning model is used as the candidate recommendation model; The training operation includes: Update the real-time iteration count; The fused feature vector is input into the current value network to obtain the state value output by the current value network; wherein, during the first training operation, the current value network is the initial value network after the reinforcement learning model is initialized; For each candidate's resume text, the total reward for the current candidate is calculated based on the historical business transaction data, the rule verification results, and the current candidate's intention data. The fused feature vector is input into the current policy network to obtain the action probability output by the current policy network; wherein, during the first training operation, the current policy network is the initial policy network after the reinforcement learning model is initialized. The current objective function value is calculated based on the action probability output by the policy network during the previous training operation, the action probability output by the current policy network, the total reward, and the state value. Based on the current objective function value, the current policy network is adjusted to obtain the updated policy network; Determine whether the current real-time iteration count has reached the preset iteration count. If so, terminate the training operation, save the updated policy network, and obtain the completed reinforcement learning model; If not, the updated policy network will be used as the policy network for the next training operation.

[0006] Furthermore, the historical business transaction data includes the historical job recommendation success rate and the historical average payment rate for the job; the intention data includes the estimated acceptance rate. The step of calculating the total reward for the current candidate based on the historical transaction data, the rule verification results, and the current candidate's intention data includes: The business reward value is calculated based on the historical recommendation success rate of the position, the historical average payment rate of the position, and the estimated acceptance rate of the current candidate. The rule reward value is calculated based on the rule verification results. The total reward for the current candidate is calculated based on the business reward value and the rule reward value.

[0007] Further, the step of calculating the current objective function value based on the action probability output by the policy network during the previous training operation, the action probability output by the current policy network, the total reward, and the state value includes: The current probability ratio is calculated based on the action probability output by the policy network during the previous training operation and the action probability output by the current policy network. The advantage function value is calculated based on the total reward and the state value. The current objective function value is calculated based on the current probability ratio and the advantage function value.

[0008] Furthermore, the generation process of the job feature vector and the resume feature vector includes: Obtain the job description text for the target position; The job description texts are vectorized using a feature encoding model to generate job feature vectors corresponding to each job description text. The resume texts are vectorized using a feature encoding model to generate resume feature vectors corresponding to each resume text. The construction process of the feature encoding model includes: Obtain a person-job matching annotation dataset; wherein, the person-job matching annotation dataset includes historical resume texts, historical business adaptability annotation weights, and historical cross-industry adaptability annotation weights for several candidates, historical job description texts, historical business requirement ability annotation weights, and historical cross-industry requirement ability annotation weights for several positions, as well as several person-job matching annotation data; The historical job description texts are structured to obtain the historical basic job feature vectors corresponding to each historical job description text. By combining the historical business requirement capability labeling weights and historical cross-industry requirement capability labeling weights of each position, the feature dimensions of each historical basic position feature vector are expanded to generate several historical position feature vectors. The historical resume texts are processed in a structured manner to obtain the historical basic resume feature vectors corresponding to each historical resume text. By combining the historical business adaptability labeling weights and historical cross-industry adaptability labeling weights of each candidate, the feature dimensions of each historical basic resume feature vector are expanded to generate several historical resume feature vectors. Based on the historical job feature vectors of each candidate, the historical resume feature vectors of each job, and the person-job matching annotation data, a preset neural network model is trained, and the trained neural network model is used as the feature encoding model.

[0009] Further, the step of calculating the matching confidence of each resume text based on the feature vector of each resume and the feature vector of the job posting includes: For each resume text, calculate the cosine similarity between the resume feature vector corresponding to the current resume text and the job feature vector; determine the matching confidence of the current resume text based on the cosine similarity.

[0010] Furthermore, each target job rule corresponds to a priority level, and different priorities correspond to different weighting coefficients; The step of performing rule validation on the candidate resume texts according to the target job rules, generating rule validation results, and calculating the rule matching degree based on the rule validation results includes: The matching degree between each candidate resume text and the rules of each target position is identified one by one, and a rule verification result corresponding to each target position rule is generated; wherein, the rule verification result is a binary discriminant value; The rule matching degree of the candidate resume text is calculated based on the weight coefficient of each target position rule and the rule verification result.

[0011] Furthermore, the resume feature vector includes a basic resume feature vector, a resume business feature vector, and a resume implicit feature vector; The candidate recommendation model construction method further includes: Calculate the difference between the rule matching degree and the matching confidence degree; Filter out the rule matching degree and matching confidence that exceed the preset threshold, and use them as the rule matching degree and matching confidence to be optimized; Based on the matching degree of the rule to be optimized and the matching confidence of the rule to be optimized, the corresponding resume feature vector is located and used as the resume feature vector to be optimized. The resume business feature vector and the resume latent feature vector in the resume feature vector to be optimized are adjusted to obtain the adjusted resume feature vector.

[0012] An embodiment of the present invention also provides a candidate recommendation model construction apparatus, comprising: The data acquisition module is used to acquire historical business transaction data, job feature vectors, and target job rules for the target position, as well as the resume text, resume feature vectors, and intention data of each candidate. The first screening module is used to calculate the matching confidence of each resume text based on the feature vector of each resume and the feature vector of the job position, and to select the resume texts whose matching confidence meets the first screening condition as candidate resume texts; wherein, the first screening condition is: the matching confidence exceeds the preset confidence threshold. The matching degree calculation module is used to perform rule verification on each candidate resume text according to the target job rules, generate rule verification results, and calculate the rule matching degree based on the rule verification results. The second filtering module is used to filter out candidate resume texts whose matching confidence meets the first filtering condition and whose rule matching degree meets the second filtering condition, as candidate resume texts; wherein, the second filtering condition is: the rule matching degree exceeds the preset matching degree threshold; The model training module is used to train a preset reinforcement learning model based on the intention data, the historical business transaction data, the job feature vector, the resume feature vector corresponding to each candidate resume text, and the rule verification results, and to use the trained reinforcement learning model as the candidate recommendation model.

[0013] This application also provides a terminal device, including: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the candidate recommendation model construction method as described in the above embodiments of the invention.

[0014] The following benefits can be obtained by implementing the present invention: This invention provides a method, apparatus, and terminal device for constructing a candidate recommendation model. The method first calculates the matching confidence score of each resume text based on the feature vectors of each resume and the feature vector of the job posting, and selects resume texts whose matching confidence scores meet a first screening condition as candidate resume texts; thus, resumes that are initially suitable for the general direction of the job posting are initially selected. Then, for each candidate resume text, rule verification is performed according to the rules of the target job posting, generating rule verification results, and the rule matching degree is calculated based on the rule verification results, thereby verifying whether the resume meets the basic recruitment requirements of the job posting from a hard indicator perspective; furthermore, candidate resume texts whose matching confidence scores meet the first screening condition and whose rule matching degrees meet the second screening condition are selected as candidates. The resume text is used to avoid overlooking potential candidates who are only slightly better than the candidates in terms of soft fit, and to exclude invalid resumes that only meet the rules but have low actual fit. Finally, based on the intention data, the historical business transaction data, the job feature vector, and the resume feature vector and rule verification results corresponding to each candidate resume text, a preset reinforcement learning model is trained. The trained reinforcement learning model is used as a candidate recommendation model. This candidate recommendation model can accurately output a "recommend / not recommend" decision based on the multi-dimensional features of the job and resume, combined with the patterns of historical recommendation data. This allows for the rapid filtering of poorly matched resumes and the identification of mid-to-high-level talents in the recruitment screening process, which improves the accuracy and conversion efficiency of resume recommendation, and reduces the time cost and decision bias of manual screening. Attached Figure Description

[0015] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating a candidate recommendation model construction method provided in a certain embodiment of this application; Figure 2 This is a schematic diagram of the structure of a candidate recommendation model construction device provided in a certain embodiment of this application; Figure 3 This is a schematic diagram of the structure of a terminal device provided in a certain embodiment of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0019] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0020] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0021] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0022] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0023] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0024] See Figure 1 To address the issues of time-consuming processes and incomplete talent coverage resulting from the reliance on manual processing of unstructured text in traditional headhunting services, an embodiment of the present invention provides a method for constructing a candidate recommendation model, comprising: S1. Obtain historical business transaction data, job feature vectors, and target job rules for the target position, as well as the resume text, resume feature vectors, and intention data of each candidate; To illustrate, firstly, it is necessary to obtain the job description text of the target position and the resume text of each candidate; the job description text has been pre-processed by a preset BERT-LLM hybrid model to perform word segmentation, entity recognition, and semantic understanding, and to complete entity standardization; similarly, the resume text has been pre-processed by the same BERT-LLM hybrid model to perform word segmentation, entity recognition, and semantic understanding, and to complete entity standardization. Specifically, the original resume text and original job description text are first formatted and irrelevant characters are removed; second, the BERT-LLM hybrid model is used for word segmentation (e.g., "Generative AI Engineer" is segmented into "Generative AI" and "Engineer") and entity recognition (extracting "Skills: PromptEngineering", "Experience: 3 years", "Salary Expectation: 40K / month", "Job Requirements: Familiar with LLM fine-tuning"); finally, the entities are standardized (e.g., "PYTHON" and "python" are unified into "Python", and "5 years of work experience" is unified into "Experience: 5 years"). It should be noted that the BERT-LLM hybrid model is based on the BERT-base model with fine-tuning. The LLM, which is pre-trained with industry corpus, has the ability to perform deep semantic parsing of unstructured text in the headhunting field. It can accurately identify core entity information such as skills, experience, and salary in resumes and job descriptions. Since the fine-tuning and application methods of such pre-trained language models have been publicly disclosed in existing technologies, they will not be elaborated here.

[0025] To ensure alignment between recommendation strategies and business objectives during subsequent reinforcement learning model training, it is necessary to obtain historical business transaction data for the target position and the intention data of each candidate. Furthermore, it is also necessary to obtain the position feature vector and the resume feature vector for each candidate. The position feature vector is 1088-dimensional (768-dimensional basic position feature vector + 256-dimensional position business feature vector + 64-dimensional position latent feature vector). The 768-dimensional vector is the basic position feature vector extracted from the job description text using a BERT-LLM hybrid model; the 256-dimensional vector is the position business feature vector generated by weighted labeling of business requirements and capabilities; and the 64-dimensional vector is the position latent feature vector generated by weighted labeling of cross-industry requirements and capabilities. The weighted labeling of business requirements and capabilities and the weighted labeling of cross-industry requirements and capabilities were pre-annotated by headhunters based on their industry experience and business data. To better illustrate the 256-dimensional job business feature vector added to the job feature vector, we assume that the business requirement capability labeling weights include quantitative weight items for dimensions such as job matching weight, salary budget sensitivity, and job responsibility priority. Each quantitative weight item is pre-labeled with a specific weight value by headhunters based on their industry recruitment and talent matching experience. For example, the specific weight value for job responsibility is 0.4, the specific weight value for job matching is 0.3, and the specific weight value for salary budget sensitivity is 0.3. Then, the quantitative weight items are concatenated to finally obtain the 256-dimensional job business feature vector in the job feature vector. To better illustrate the 64-dimensional implicit feature vector added to the job feature vector, we assume that the cross-industry required capability labeling weights include quantitative weight items for dimensions such as cross-industry migration acceptance, industry skill interoperability, and corporate culture adaptation preference. Each quantitative weight item is pre-calibrated by headhunters based on industry experience and historical job matching data to obtain a specific weight value. For example, the specific weight value for cross-industry migration acceptance is 0.5, the specific weight value for industry skill interoperability is 0.3, and the specific weight value for corporate culture adaptation preference is 0.2. Then, the quantitative weight items are concatenated to finally obtain the 64-dimensional implicit feature vector of the job feature vector.

[0026] Similarly, each candidate's resume feature vector is 1088-dimensional (768-dimensional basic resume feature vector + 256-dimensional resume business feature vector + 64-dimensional resume latent feature vector). The 768-dimensional vector is the basic resume feature vector extracted from the resume text using the BERT-base model. The 256-dimensional vector is the resume business feature vector generated based on the business adaptability label weights. The 64-dimensional vector is the resume latent feature vector generated based on the cross-industry adaptability label weights. The business adaptability label weights and cross-industry adaptability label weights are obtained in advance through headhunting industry experience and candidate qualification assessment. Similarly, the 256-dimensional resume business feature vector added to each resume feature vector has the same dimension definition and weight labeling rules as the corresponding 256-dimensional job business feature vector in the job feature vector. In addition, the 64-dimensional resume implicit feature vector added to each resume feature vector has the same dimension definition and weight labeling rules as the corresponding 64-dimensional job implicit feature vector in the job feature vector. This ensures that the feature vectors of both the person and the job can be directly used for similarity calculation and matching evaluation within the same feature space.

[0027] In a preferred embodiment, the process of generating the job feature vector and the resume feature vector includes: The job description texts are vectorized using a feature encoding model to generate job feature vectors corresponding to each job description text. The resume texts are vectorized using a feature encoding model to generate resume feature vectors corresponding to each resume text. The construction process of the feature encoding model includes: Obtain a person-job matching annotation dataset; wherein, the person-job matching annotation dataset includes historical resume texts, historical business adaptability annotation weights, and historical cross-industry adaptability annotation weights for several candidates, historical job description texts, historical business requirement ability annotation weights, and historical cross-industry requirement ability annotation weights for several positions, as well as several person-job matching annotation data; The historical job description texts are structured to obtain the historical basic job feature vectors corresponding to each historical job description text. By combining the historical business requirement capability labeling weights and historical cross-industry requirement capability labeling weights of each position, the feature dimensions of each historical basic position feature vector are expanded to generate several historical position feature vectors. The historical resume texts are processed in a structured manner to obtain the historical basic resume feature vectors corresponding to each historical resume text. By combining the historical business adaptability labeling weights and historical cross-industry adaptability labeling weights of each candidate, the feature dimensions of each historical basic resume feature vector are expanded to generate several historical resume feature vectors. Based on the historical job feature vectors of each candidate, the historical resume feature vectors of each job, and the person-job matching annotation data, a preset neural network model is trained, and the trained neural network model is used as the feature encoding model. As an illustration, in the process of generating job feature vectors and resume feature vectors, it is necessary to vectorize each job description text through a feature encoding model to generate a job feature vector corresponding to each job description text. Similarly, it is necessary to vectorize each resume text through a feature encoding model to generate a resume feature vector corresponding to each resume text. The feature encoding model needs to be trained based on job-person matching labeled data to have accurate feature mapping capabilities. Therefore, before training the feature encoding model, it is necessary to construct a number of job-person matching labeled data. Specifically, the training process of the feature encoding model is as follows: First, a person-job matching annotation dataset is obtained. The person-job matching annotation dataset includes historical resume texts, historical business adaptability annotation weights, and historical cross-industry adaptability annotation weights for several candidates; historical job description texts, historical business requirement ability annotation weights, and historical cross-industry requirement ability annotation weights for several positions; and several person-job matching annotation data (labeled "match / not match"). Among them, the person-job matching annotation data are the matching result labels between candidates and positions, including the binary label "match / not match". Subsequently, the BERT-LLM hybrid model is used to perform structured processing on the historical job description texts to obtain the historical basic job feature vectors corresponding to each historical job description text; similarly, the BERT-LLM hybrid model is used to perform structured processing on the historical resume texts to obtain the historical basic resume feature vectors corresponding to each historical resume text. Subsequently, combining the historical business requirement capability labeling weights and historical cross-industry requirement capability labeling weights of each position, the feature dimensions of each historical basic position feature vector are expanded to generate several historical position feature vectors. The specific feature dimension expansion process is the same as the feature dimension expansion process of the 256-dimensional position business feature vector and the 64-dimensional position implicit feature vector added in the above position feature vectors, and will not be repeated here. Similarly, by combining the historical business adaptability labeling weights and historical cross-industry adaptability labeling weights corresponding to each candidate, the feature dimensions of each historical basic resume feature vector are expanded to generate several historical resume feature vectors. The specific feature dimension expansion process is the same as the feature dimension expansion process of the 256-dimensional resume business feature vector and the 64-dimensional resume latent feature vector added to the resume feature vector above, and will not be repeated here. Finally, based on the historical job feature vectors of each candidate, the historical resume feature vectors of each job, and the aforementioned person-job matching annotation data (labeled "match / mismatch"), a pre-defined neural network model is trained, and the trained neural network model is used as the feature encoding model. The following is a brief description of the training process: Using the "match / non-match" binary labels in the job matching annotation data as supervision signals, a loss function based on cross-entropy loss is constructed. The feature vectors of historical resumes and the feature vectors of historical job positions are input into a preset neural network model. The attention weights and fully connected layer parameters of the model are iteratively adjusted through the backpropagation algorithm. When the error between the job matching prediction data (also labeled "match / non-match") output by the model and the corresponding job matching annotation data converges to a preset threshold, the training is terminated, and the trained neural network model is obtained, which serves as the feature encoding model.

[0028] S2. Based on the feature vectors of each resume and the feature vector of the job position, calculate the matching confidence of each resume text, and select the resume texts whose matching confidence meets the first screening condition as candidate resume texts; wherein, the first screening condition is: the matching confidence exceeds the preset confidence threshold. In a preferred embodiment, calculating the matching confidence score of each resume text based on each resume feature vector and the job feature vector includes: For each resume text, calculate the cosine similarity between the resume feature vector corresponding to the current resume text and the job feature vector; determine the matching confidence of the current resume text based on the cosine similarity. Specifically, for each resume text, the cosine similarity between the resume feature vector corresponding to the current resume text and the job feature vector is calculated, and the specific calculation formula is as follows: ; In the formula, Represents the resume feature vector (1088 dimensions); Let Sim represent the feature vector of a job position. For example, if the dot product of the feature vector of a candidate's resume and the feature vector of a job position is 980, the magnitude of the candidate's feature vector is 32, and the magnitude of the feature vector of the job position is 31, then Sim = 980 / (32 × 31) ≈ 0.99.

[0029] Then, latent feature weights are introduced. The higher the weight of the latent feature, the greater its impact on the final confidence level. In this embodiment, The value is set to 0.3; then the latent feature matching degree is calculated. This refers to the cosine similarity between the candidate's resume latent feature vector (64-dimensional) and the job's latent feature vector (64-dimensional) (the principle is the same as the cosine similarity between the resume feature vector and the job feature vector, and will not be repeated here). In the example, =0.82, therefore, the specific formula for calculating the final matching confidence score is as follows: Substituting the data calculated in the above embodiments, the final matching confidence level is obtained: (Keep three decimal places); Then, the resume texts that meet the first screening criteria with matching confidence are used as candidate resume texts. Taking 500 resume texts as an example, after screening, only 85 resume texts that meet the first screening criteria with matching confidence are left. These 85 resume texts are used as candidate resume texts. It should be noted that the preset reliability threshold can be adjusted according to the actual situation. In this embodiment, the preset reliability threshold is 0.6.

[0030] S3. For each candidate resume text, perform rule verification on the candidate resume text according to the target job rules, generate rule verification results, and calculate the rule matching degree based on the rule verification results. This illustrates how building a knowledge graph can provide structured knowledge support for the formulation of job rules. First, it is necessary to acquire the information required for knowledge graph construction, such as industry knowledge (e.g., skill level association graphs), corporate rules (e.g., hard requirements such as salary budgets and educational thresholds), and implicit headhunting experience (e.g., cross-industry talent ability transfer judgment criteria). Specifically, the system automatically crawls industry reports, job requirements from leading companies, and headhunter experience documents for the target position over the past year using a pre-set BERT-CRF model. During the knowledge crawling process, information is extracted using the "Named Entity Recognition + Relation Extraction" algorithm to extract entities (such as "Skills: Python, Position: Product Manager"), relationships (such as "Python → Machine Learning" and "Product Manager → Requires Requirements Analysis Ability"), and rules (such as "Product Manager Experience ≥ 3 Years" and "Salary ≤ 35K / Month"). Then, through a standardized mapping method between entities and relationships, a multi-dimensional knowledge graph is constructed. It should be noted that since the specific entity alignment and relationship reasoning of the knowledge graph are within the scope of existing technologies, they will not be elaborated here.

[0031] As an illustration, after constructing a knowledge graph, rules are extracted and logically transformed from the knowledge graph to generate several job rules related to the requirements of the target job, forming a rule base; Specifically, first, entities related to the target job requirements (such as candidate skills, years of experience, salary expectations, and job requirements) and relationships between entities (such as "candidate skills include Python", "years of experience ≥ 3 years", and "salary expectations ≤ job budget") are filtered from the knowledge graph; the filtered "entity-relationship" pairs are then mapped to the condition and conclusion items of rules: Condition item: Composed of multiple entity relationships, connected by "AND / OR" logic (e.g., candidate skills include Python AND experience ≥ 3 years). Conclusion: The judgment result when the corresponding conditions are met (e.g., meeting the basic requirements for technical positions).

[0032] Then, according to the preset rule syntax format (such as IF {condition} THEN {conclusion}), the condition and conclusion items of the mapped rules are transformed into machine-recognizable logical rules. At the same time, each rule is assigned a rule priority label, and finally a structured job rule library is formed. Each job rule in the job rule library is labeled with a corresponding priority. For example, budget rule priority 1, experience rule priority 2, skill rule priority 3, where budget rule priority 1 is higher than experience rule priority 2, and so on. For example, before transforming the logical rules, the knowledge graph triples are: (Technical position, core skill: Python); (Technical position, experience requirement: ≥3 years); After being converted into logical rules, the logical rules are as follows: If the candidate's skills include Python and they have ≥3 years of experience, then they meet the basic requirements for technical positions.

[0033] In a preferred embodiment, each target job rule corresponds to a priority, and different priorities correspond to different weight coefficients. The step of performing rule validation on the candidate resume texts according to the target job rules, generating rule validation results, and calculating the rule matching degree based on the rule validation results includes: The matching degree between each candidate resume text and the rules of each target position is identified one by one, and a rule verification result corresponding to each target position rule is generated; wherein, the rule verification result is a binary discriminant value; The rule matching degree of the candidate resume text is calculated based on the weight coefficient of each target position rule and the rule verification result. Specifically, the matching degree between the candidate resume texts and the rules of each target position is identified one by one, and rule verification results corresponding to the rules of each target position are generated, wherein the rule verification results are binary discriminant values; Taking the target position as smart hardware product manager as an example, the target position rules include target position rule R1, target position rule R2, and target position rule R3, as follows: Target position rule R1 (budget rule priority 1): Salary expectation ≤ 35K / month; Target position rule R2 (experience rule priority 2): Product manager experience ≥ 3 years; Target position rule R3 (skill rule priority 3): Familiar with prototyping tools such as Axure; Therefore, the matching degree between the candidate resume texts and the rules of each target position is identified one by one, and a rule verification result corresponding to each target position rule is generated. When they match, the rule verification result = 1; when they do not match, the rule verification result = 0. For example, in a candidate's alternative resume text, the salary expectation is 33K / month (match, R1=1), the product manager experience is 2 years (mismatch, R2=0), and the candidate is familiar with Axure (match, R3=1). By combining the weight coefficients of the rules for each target position and the rule verification results, the rule matching degree of the candidate resume text is calculated, where the weight coefficient of the budget rule is... The weighting coefficient of the empirical rule is 0.4. The weighting coefficient for skill rules is 0.35. The value is 0.25. Therefore, the rule matching degree of the candidate resume text is calculated using the following formula. : ; In the formula, Indicates the first The weighting coefficients of the target job rules; Indicates the first The rule validation results for each target position rule; This indicates the total number of rules for the target position; Therefore, the rule matching degree of the candidate resume texts : .

[0034] To illustrate, after calculating the rule matching degree and matching confidence of each resume feature vector, in order to optimize the representation accuracy of the business position feature vector and improve the reliability of the subsequent candidate recommendation model, it is necessary to calculate the difference between the rule matching degree and matching confidence of each resume feature vector. The difference is used to judge the consistency of the matching evaluation results (the smaller the difference, the higher the consistency; the larger the difference, the more significant the deviation). This allows for adjustments to the resume business feature vector and resume latent feature vector in the resume feature vector, reducing invalid recommendations or omissions of high-quality candidates due to feature representation deviations. This makes the matching results output by the subsequently trained candidate recommendation model more closely match the actual recruitment needs of enterprises and the judgment logic of headhunting businesses. In a preferred embodiment, the resume feature vector includes a basic resume feature vector, a resume business feature vector, and a resume implicit feature vector; The candidate recommendation model construction method further includes: Calculate the difference between the rule matching degree and the matching confidence degree; Filter out the rule matching degree and matching confidence that exceed the preset threshold, and use them as the rule matching degree and matching confidence to be optimized; Based on the matching degree of the rule to be optimized and the matching confidence of the rule to be optimized, the corresponding resume feature vector is located and used as the resume feature vector to be optimized. The resume business feature vector and the resume latent feature vector in the resume feature vector to be optimized are adjusted to obtain the adjusted resume feature vector; Specifically, the difference between the rule matching degree and the matching confidence degree is calculated to determine the rule matching degree and the matching confidence degree to be optimized; Specifically, the difference between the matching confidence P and the rule matching degree Q for each match is calculated. Set a preset threshold When the difference Exceed When the condition is met, a difference analysis is triggered. For example, if a candidate has P=0.88, Q=0.55, and Δ=0.33≥0.2, a difference analysis is triggered; if another candidate has P=0.72, Q=0.68, and Δ=0.04<0.2, optimization is not triggered. Specifically, it is necessary to identify the source of the discrepancy. For example, if the discrepancy stems from the low weight of experience-related features in the resume's business feature vector, then the contribution of the experience dimension to the matching confidence can be increased by adjusting the weight of the experience dimension in the resume's business feature vector, thereby narrowing the difference between the matching degree and the rule matching degree. Similarly, if the discrepancy is due to the low weight of soft skills such as cross-industry capabilities in the resume's implicit feature vector, the difference between the rule matching degree and the matching confidence degree can also be reduced by optimizing the weight of the implicit features. For example, if the specific weight value of the "hardware experience" feature in a resume's business feature vector is only 0.1, resulting in a high P, but the target job rule R2 requires "hardware experience" to be "greater than 5 years", resulting in a low Q, the reason for the difference is determined to be that the weight allocation of the "hardware experience" feature in the resume's business feature vector is insufficient. Therefore, by adjusting the specific weight value of the "hardware experience" feature in the resume's business feature vector, the contribution value of the "hardware experience" feature to the matching confidence can be increased.

[0035] S4. Select candidate resume texts whose matching confidence meets the first screening condition and whose rule matching degree meets the second screening condition, and use them as candidate resume texts; wherein, the second screening condition is: the rule matching degree exceeds the preset matching degree threshold. Specifically, candidate resume texts that meet the first screening condition in terms of matching confidence and the second screening condition in terms of rule matching are selected as candidate resume texts. The preset matching threshold can be set according to the actual situation. In this embodiment, the preset matching threshold is set to 0.5. Thus, the selected candidate resume texts ensure that they cover the key hard indicators in the recruitment requirements, avoiding invalid screening due to the lack of hard conditions. The matching confidence that meets the first screening condition (such as a preset confidence threshold ≥ 0.6) verifies the implicit suitability between the resume and the job (such as the fit of career development path, soft skills matching, project experience relevance, etc.), eliminating resumes that only meet the surface rules but have low actual suitability. The final selected candidate resume texts not only ensure the efficiency of recruitment screening but also improve the accuracy of the candidate group, providing a high-quality sample for subsequent screening.

[0036] S5. Based on the intention data, the historical business transaction data, the job feature vector, and the resume feature vector and rule verification results corresponding to each candidate resume text, train the preset reinforcement learning model, and use the trained reinforcement learning model as the candidate recommendation model. Specifically, the state space S of the reinforcement learning model is defined as including a 1088-dimensional resume feature vector, a 1088-dimensional job feature vector, and the historical recommendation success rate of the job (assuming the historical recommendation success rate of the target job is 65%, standardized to 0.65), and the action space A (recommendation = 1, non-recommendation = 0). The value network has the same structure as the policy network, and the output of the value network is the state value. In solving the problem, the PPO algorithm is used to train the policy network in the reinforcement learning model through iterative training, so that after training, the reinforcement learning model can maximize the total reward in action decision-making. The PPO algorithm parameters are specifically set as follows: batch size 32, preset number of iterations 100, clip parameter ε = 0.2, and discount factor γ = 0.95.

[0037] In a preferred embodiment, the step of training a preset reinforcement learning model based on the intention data, the historical business transaction data, the job feature vector, and the resume feature vectors and rule verification results corresponding to each candidate resume text, and using the trained reinforcement learning model as a candidate recommendation model, includes: For each candidate resume text, the job feature vector, the resume feature vector corresponding to the candidate resume text, and the historical recommendation success rate of the job are combined to form a fused feature vector; Based on the fused feature vector, the historical business transaction data, the rule verification results, and the intention data, the training operation is repeatedly performed on the preset reinforcement learning model until the preset number of iterations is reached, and the trained reinforcement learning model is obtained. The trained reinforcement learning model is used as the candidate recommendation model; The training operation includes: Update the real-time iteration count; The fused feature vector is input into the current value network to obtain the state value output by the current value network; wherein, during the first training operation, the current value network is the initial value network after the reinforcement learning model is initialized; For each candidate's resume text, the total reward for the current candidate is calculated based on the historical business transaction data, the rule verification results, and the current candidate's intention data. The fused feature vector is input into the current policy network to obtain the action probability output by the current policy network; wherein, during the first training operation, the current policy network is the initial policy network after the reinforcement learning model is initialized. The current objective function value is calculated based on the action probability output by the policy network during the previous training operation, the action probability output by the current policy network, the total reward, and the state value. Based on the current objective function value, the current policy network is adjusted to obtain the updated policy network; Determine whether the current real-time iteration count has reached the preset iteration count. If so, terminate the training operation, save the updated policy network, and obtain the completed reinforcement learning model; If not, use the updated policy network as the policy network for the next training operation and perform the next training operation. Specifically, firstly, the job feature vector, the resume feature vector corresponding to the candidate resume text, and the historical recommendation success rate of the job are combined to form a fused feature vector, assuming the state space of a certain candidate. The fused feature vectors (including a 1088-dimensional resume feature vector, a 1088-dimensional job feature vector, and a historical job recommendation success rate of 0.65) are input into the current value network to obtain the state value output by the current value network. =0.9; In a preferred embodiment, the historical business transaction data includes the historical job recommendation success rate and the historical average payment rate for the job; the intention data includes the estimated acceptance rate. The step of calculating the total reward for the current candidate based on the historical transaction data, the rule verification results, and the current candidate's intention data includes: The business reward value is calculated based on the historical recommendation success rate of the position, the historical average payment rate of the position, and the estimated acceptance rate of the current candidate. The rule reward value is calculated based on the rule verification results. The total reward for the current candidate is calculated based on the business reward value and the rule reward value. Specifically, the business reward value is calculated based on the historical recommendation success rate of the position, the historical average payment rate of the position, and the estimated acceptance rate of the current candidate. It is assumed that the historical recommendation success rate of the target position is 60% (normalized to 0.6, i.e., ...). =0.6), after initial communication with the headhunter, the estimated acceptance rate of the current candidate is 75% (standardized to 0.75); the historical average return rate for the position is 100,000 yuan, standardized to 1. Substituting these values ​​into the business bonus calculation formula, the business bonus value D1 is calculated: ; Specifically, based on the rule verification results, the rule reward value is calculated. Then, by counting the number of rule verification results that match, and combining this with the total number of rule verification results, the rule satisfaction level is calculated. ; The rule reward value D2 is then calculated: ; In the formula, This represents the first reward weighting coefficient; This indicates the second reward weighting coefficient; the first and second reward weighting coefficients can be set according to the actual situation. Therefore, the total reward D = D1 + D2 = 0.85 + 0.5 = 1.35; The fused feature vector is then input into the current policy network to obtain the action probability output by the current policy network, namely the probability P(A=1) of a candidate being recommended (action A=1) and the probability P(A=0) of not being recommended (action A=0). In a preferred embodiment, calculating the current objective function value based on the action probability output by the policy network during the previous training operation, the action probability output by the current policy network, the total reward, and the state value includes: The current probability ratio is calculated based on the action probability output by the policy network during the previous training operation and the action probability output by the current policy network. The advantage function value is calculated based on the total reward and the state value. The current objective function value is calculated based on the current probability ratio and the advantage function value. Specifically, the current probability ratio is calculated based on the action probabilities output by the policy network during the previous training operation and the action probabilities output by the current policy network. Assuming the policy network output an action probability of 0.6 during the previous training operation, and the current policy network outputs an action probability of 0.66, =1.1; Then, based on the total reward D and the state value The dominant function value was calculated. : ; The current policy network is updated using the PPO algorithm, and the current objective function value is calculated. for: ; Calculate the current objective function value The gradient of the current policy network parameters is calculated, and the network weights and biases of the current policy network are updated along the negative direction of the gradient to obtain the updated policy network. Determine whether the current real-time iteration count has reached the preset iteration count. If yes, terminate the training operation, save the updated policy network, and obtain the trained reinforcement learning model. If no, use the updated policy network as the policy network for the next training operation, and simultaneously update the current value network to a version that matches the policy network parameters, and then proceed with the next training operation. Finally, the trained reinforcement learning model, as a candidate recommendation model, can accurately output "recommend / not recommend" decisions based on the multi-dimensional features of job positions and resumes, combined with the patterns of historical recommendation data. This allows for the rapid filtering of low-fit resumes and the identification of high-potential candidates during the recruitment screening process, improving the accuracy and conversion efficiency of resume recommendations while reducing the time cost and decision bias of manual screening.

[0038] See Figure 2 This is a candidate recommendation model construction device provided in an embodiment of the present invention, comprising: The data acquisition module is used to acquire historical business transaction data, job feature vectors, and target job rules for the target position, as well as the resume text, resume feature vectors, and intention data of each candidate. The first screening module is used to calculate the matching confidence of each resume text based on the feature vector of each resume and the feature vector of the job position, and to select the resume texts whose matching confidence meets the first screening condition as candidate resume texts; wherein, the first screening condition is: the matching confidence exceeds the preset confidence threshold. The matching degree calculation module is used to perform rule verification on each candidate resume text according to the target job rules, generate rule verification results, and calculate the rule matching degree based on the rule verification results. The second filtering module is used to filter out candidate resume texts whose matching confidence meets the first filtering condition and whose rule matching degree meets the second filtering condition, as candidate resume texts; wherein, the second filtering condition is: the rule matching degree exceeds the preset matching degree threshold; The model training module is used to train a preset reinforcement learning model based on the intention data, the historical business transaction data, the job feature vector, the resume feature vector corresponding to each candidate resume text, and the rule verification results, and to use the trained reinforcement learning model as the candidate recommendation model.

[0039] See Figure 3 One embodiment of this application also provides a terminal device, including: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the candidate recommendation model construction method as described above.

[0040] The processor controls the overall operation of the terminal device to complete all or part of the steps of the candidate recommendation model construction method described above. The memory stores various types of data to support the operation of the terminal device; this data may include, for example, instructions for any application or method operating on the terminal device, as well as application-related data. The memory can be implemented using any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0041] In an exemplary embodiment, the terminal device may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the candidate recommendation model construction method as described in any of the foregoing embodiments and achieve the same technical effect as the methods described above.

[0042] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for constructing a candidate recommendation model, characterized in that, include: Obtain historical business transaction data, job feature vectors, and target job rules for the target position, as well as the resume text, resume feature vectors, and intention data of each candidate; Based on the feature vectors of each resume and the feature vector of the job position, the matching confidence of each resume text is calculated, and the resume texts whose matching confidence meets the first screening condition are selected as candidate resume texts; wherein, the first screening condition is: the matching confidence exceeds the preset confidence threshold. For each candidate resume text, rule validation is performed on the candidate resume text according to the target job rules, a rule validation result is generated, and the rule matching degree is calculated based on the rule validation result. Select candidate resume texts whose matching confidence level meets the first screening condition and whose rule matching degree meets the second screening condition; wherein, the second screening condition is: the rule matching degree exceeds the preset matching degree threshold; Based on the intention data, the historical business transaction data, the job feature vector, and the resume feature vector and rule verification results corresponding to each candidate resume text, a preset reinforcement learning model is trained, and the trained reinforcement learning model is used as the candidate recommendation model.

2. The candidate recommendation model construction method as described in claim 1, characterized in that, The step of training a pre-defined reinforcement learning model based on the intention data, the historical business transaction data, the job feature vector, and the resume feature vectors and rule verification results corresponding to each candidate resume text, and using the trained reinforcement learning model as the candidate recommendation model, includes: For each candidate resume text, the job feature vector, the resume feature vector corresponding to the candidate resume text, and the historical recommendation success rate of the job are combined to form a fused feature vector; Based on the fused feature vector, the historical business transaction data, the rule verification results, and the intention data, the training operation is repeatedly performed on the preset reinforcement learning model until the preset number of iterations is reached, and the trained reinforcement learning model is obtained. The trained reinforcement learning model is used as the candidate recommendation model; The training operation includes: Update the real-time iteration count; The fused feature vector is input into the current value network to obtain the state value output by the current value network; wherein, during the first training operation, the current value network is the initial value network after the reinforcement learning model is initialized; For each candidate's resume text, the total reward for the current candidate is calculated based on the historical business transaction data, the rule verification results, and the current candidate's intention data. The fused feature vector is input into the current policy network to obtain the action probability output by the current policy network; wherein, during the first training operation, the current policy network is the initial policy network after the reinforcement learning model is initialized. The current objective function value is calculated based on the action probability output by the policy network during the previous training operation, the action probability output by the current policy network, the total reward, and the state value. Based on the current objective function value, the current policy network is adjusted to obtain the updated policy network; Determine whether the current real-time iteration count has reached the preset iteration count. If so, terminate the training operation, save the updated policy network, and obtain the completed reinforcement learning model; If not, the updated policy network will be used as the policy network for the next training operation.

3. The candidate recommendation model construction method as described in claim 2, characterized in that, The historical business transaction data includes the historical job recommendation success rate and the historical average payment rate for the job; the intention data includes the estimated acceptance rate. The step of calculating the total reward for the current candidate based on the historical transaction data, the rule verification results, and the current candidate's intention data includes: The business reward value is calculated based on the historical recommendation success rate of the position, the historical average payment rate of the position, and the estimated acceptance rate of the current candidate. The rule reward value is calculated based on the rule verification results. The total reward for the current candidate is calculated based on the business reward value and the rule reward value.

4. The candidate recommendation model construction method as described in claim 3, characterized in that, The step of calculating the current objective function value based on the action probability output by the policy network during the previous training operation, the action probability output by the current policy network, the total reward, and the state value includes: The current probability ratio is calculated based on the action probability output by the policy network during the previous training operation and the action probability output by the current policy network. The advantage function value is calculated based on the total reward and the state value. The current objective function value is calculated based on the current probability ratio and the advantage function value.

5. The candidate recommendation model construction method as described in claim 1, characterized in that, The generation process of the job feature vector and the resume feature vector includes: Obtain the job description text for the target position; The job description texts are vectorized using a feature encoding model to generate job feature vectors corresponding to each job description text. The resume texts are vectorized using a feature encoding model to generate resume feature vectors corresponding to each resume text. The construction process of the feature encoding model includes: Obtain a person-job matching annotation dataset; wherein, the person-job matching annotation dataset includes historical resume texts, historical business adaptability annotation weights, and historical cross-industry adaptability annotation weights for several candidates, historical job description texts, historical business requirement ability annotation weights, and historical cross-industry requirement ability annotation weights for several positions, as well as several person-job matching annotation data; The historical job description texts are structured to obtain the historical basic job feature vectors corresponding to each historical job description text. By combining the historical business requirement capability labeling weights and historical cross-industry requirement capability labeling weights of each position, the feature dimensions of each historical basic position feature vector are expanded to generate several historical position feature vectors. The historical resume texts are processed in a structured manner to obtain the historical basic resume feature vectors corresponding to each historical resume text. By combining the historical business adaptability labeling weights and historical cross-industry adaptability labeling weights of each candidate, the feature dimensions of each historical basic resume feature vector are expanded to generate several historical resume feature vectors. Based on the historical job feature vectors of each candidate, the historical resume feature vectors of each job, and the person-job matching annotation data, a preset neural network model is trained, and the trained neural network model is used as the feature encoding model.

6. The candidate recommendation model construction method as described in claim 1, characterized in that, The step of calculating the matching confidence score of each resume text based on the feature vectors of each resume and the feature vector of the job position includes: For each resume text, calculate the cosine similarity between the resume feature vector corresponding to the current resume text and the job feature vector; determine the matching confidence of the current resume text based on the cosine similarity.

7. The candidate recommendation model construction method as described in claim 1, characterized in that, Each target job rule corresponds to a priority, and different priorities correspond to different weight coefficients; The step of performing rule validation on the candidate resume texts according to the target job rules, generating rule validation results, and calculating the rule matching degree based on the rule validation results includes: The matching degree between each candidate resume text and the rules of each target position is identified one by one, and a rule verification result corresponding to each target position rule is generated; wherein, the rule verification result is a binary discriminant value; The rule matching degree of the candidate resume text is calculated based on the weight coefficient of each target position rule and the rule verification result.

8. The candidate recommendation model construction method as described in claim 1, characterized in that, The resume feature vector includes a basic resume feature vector, a resume business feature vector, and a resume implicit feature vector; The candidate recommendation model construction method further includes: Calculate the difference between the rule matching degree and the matching confidence degree; Filter out the rule matching degree and matching confidence that exceed the preset threshold, and use them as the rule matching degree and matching confidence to be optimized; Based on the matching degree of the rule to be optimized and the matching confidence of the rule to be optimized, the corresponding resume feature vector is located and used as the resume feature vector to be optimized. The resume business feature vector and the resume latent feature vector in the resume feature vector to be optimized are adjusted to obtain the adjusted resume feature vector.

9. A candidate recommendation model construction apparatus, characterized in that, include: The data acquisition module is used to acquire historical business transaction data, job feature vectors, and target job rules for the target position, as well as the resume text, resume feature vectors, and intention data of each candidate. The first screening module is used to calculate the matching confidence of each resume text based on the feature vector of each resume and the feature vector of the job position, and to select the resume texts whose matching confidence meets the first screening condition as candidate resume texts; wherein, the first screening condition is: the matching confidence exceeds the preset confidence threshold. The matching degree calculation module is used to perform rule verification on each candidate resume text according to the target job rules, generate rule verification results, and calculate the rule matching degree based on the rule verification results. The second filtering module is used to filter out candidate resume texts whose matching confidence meets the first filtering condition and whose rule matching degree meets the second filtering condition, as candidate resume texts; wherein, the second filtering condition is: the rule matching degree exceeds the preset matching degree threshold; The model training module is used to train a preset reinforcement learning model based on the intention data, the historical business transaction data, the job feature vector, the resume feature vector corresponding to each candidate resume text, and the rule verification results, and use the trained reinforcement learning model as the candidate recommendation model.

10. A terminal device, characterized in that, include: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the candidate recommendation model construction method as described in any one of claims 1-8.

Images