Human resource vertical large model training method based on hr-rag knowledge enhancement

Abstract

The application discloses a human resource vertical category model training method based on HR-RAG knowledge enhancement, belongs to the technical field of human resource vertical category model training, and comprises the following steps: step 1, knowledge structuring and building, step 2, designing a retrieval enhancement strategy, step 3, correlation coding and fusion, step 4, task alignment and incremental training, and step 5, confrontation verification and optimization. Through the HR vertical knowledge graph and the weighted retrieval strategy, the accurate adaptation of knowledge and business scenarios is realized, the general technical knowledge matching misplacement problem is solved, and the pertinence of model generated content is improved. Meanwhile, relying on the knowledge updating and incremental training mechanism, new rules and system iteration content can be quickly integrated, the knowledge lag problem is solved, and the output content can meet the latest requirements.

Description

Technical Field

[0001] This invention relates to the field of human resource vertical model training technology, and more specifically, to a method for training large human resource vertical models based on HR-RAG knowledge enhancement. Background Technology

[0002] Human resources is the core functional system of "people" in organizational management, covering the entire process of recruitment, training, performance, compensation, employee relations and organizational development. However, general large models lack a deep understanding of structured knowledge and business scenarios in the HR field, which can easily lead to knowledge mismatch, content lag and task deviation, resulting in insufficient credibility and usability in actual business. Therefore, it is necessary to carry out training of large models for human resources verticals.

[0003] The prior art patent document with authorization announcement number CN119273313B discloses a "digitalized human resource management statistical method and system", which receives screening requests from the evaluation end, obtains the application templates of each candidate end and inputs them into the screening training model to screen questions, obtains a customized list of questions corresponding to each candidate end, and generates automatic question data corresponding to each candidate end based on the customized list of questions.

[0004] The patent document with authorization announcement number CN111784308A discloses "A method for training a human resource cost assessment and pricing model based on a work platform", which includes the following steps: the task acquisition server extracts and calculates keywords from task information and transmits the calculation results to the human resource assessment server; the human resource assessment server establishes a human resource cost assessment system and sets multi-level evaluation grades; the human resource cost assessment system generates work task level information and matches personnel information that completes the task based on the task information; and the task level and personnel information are sent to the pricing network model.

[0005] While existing technologies can establish cost assessment systems through keyword extraction, match task personnel information, and train pricing network models to achieve accurate task cost estimation and profit maximization, and improve job matching accuracy and interview efficiency through customized question lists, virtual digital human automatic interviews, and interactive backtracking chains, they lack specific representation and retrieval strategies for structured, semi-structured, and unstructured knowledge in the HR field. This results in knowledge matching mismatches, a lack of dynamic knowledge increment integration mechanisms leading to knowledge lag, and a disconnect between the retrieval and generation stages. Furthermore, the absence of a training system adapted to core HR tasks makes it difficult to meet the precise needs of the entire human resources business process. Summary of the Invention

[0006] This invention mainly provides a training method for large-scale human resource vertical models based on HR-RAG knowledge enhancement, which can solve the problems mentioned in the background technology.

[0007] To achieve the above objectives, this invention provides the following technical solution: a training method for a large-scale human resources vertical model based on HR-RAG knowledge enhancement, comprising: S1. Collect relevant knowledge in the field of human resources, define the core entities and relationships in the knowledge, and form a structured knowledge set; S2. Based on the basic similarity calculation method, multi-dimensional human resources scenario-related weights are introduced to construct a multi-dimensional retrieval strategy for searching and filtering the knowledge set; S3. The retrieved and filtered knowledge is associated with the model input requirements, and a hierarchical coding method is used to achieve deep integration between the two; S4. Design specific training objectives for human resources-related tasks, adopt incremental training, and combine multiple rounds of iterative optimization of model parameters; S5. Optimize the knowledge set and retrieval strategy through adversarial sample verification and knowledge consistency verification.

[0008] Furthermore, in S1, the core entities include positions, employees, skills, performance indicators, compensation items, relevant laws and regulations, and corporate systems; the relationships include positions-skill requirements, employees-skill mastery, performance indicators-compensation linkage, and relevant laws and regulations-system constraints; the knowledge acquisition sources include structured data from the human resources information system, semi-structured data from employee handbooks and related documents, and unstructured data from interview records and exit interviews.

[0009] Furthermore, in S2, the weight dimensions of the multi-dimensional retrieval strategy include job level weight, knowledge timeliness weight, and business relevance weight; the retrieval method adopts a fusion sorting method of semantic retrieval, structured retrieval, and timeliness retrieval.

[0010] Furthermore, in S3, the hierarchical encoding adopts a dual-path encoding structure, including a conventional semantic encoding path for questions and a structured encoding path for retrieved knowledge entities and relationships.

[0011] Furthermore, in S4, human resources-related tasks include resume screening and evaluation, interview question generation, performance report writing, salary calculation, and employee turnover risk prediction; incremental training adopts efficient parameter fine-tuning technology, freezing the underlying parameters of the model and only adjusting the parameters of the upper adaptation layer.

[0012] Furthermore, in S5, adversarial samples are generated for high-risk HR scenarios; knowledge consistency verification compares the model-generated content with authoritative knowledge in the knowledge set; and retrieval strategy optimization adjusts weight values ​​based on retrieval mismatch cases.

[0013] Furthermore, the structured knowledge set is stored using a dynamic knowledge graph for HR vertical categories.

[0014] The beneficial effects of the HR-RAG knowledge-enhanced large-scale human resources model training method of this invention are as follows: Through HR vertical knowledge graphs and weighted retrieval strategies, it can achieve precise adaptation of knowledge to HR business scenarios, covering not only core entities and relationships throughout the entire process but also incorporating multi-dimensional scenario weights and optimizing the fusion and ranking logic, making the model-generated content more targeted and providing precise knowledge support for human resources business. Furthermore, through knowledge update and incremental training mechanisms, it enables the rapid integration of new regulations and iterative system content, breaking down knowledge lag barriers, ensuring the model always meets the latest requirements, improving output compliance, and enhancing practical value. Simultaneously, through semantic association encoding and adversarial verification closed loops, it provides a deep fusion path of retrieval and generation, effectively suppressing unfounded inferences. Based on authoritative knowledge verification and high-risk scenario verification, it strengthens the authority and robustness of generated content, greatly improving the model's decision-making reliability in the entire HR business process. Attached Figure Description

[0015] The present invention will now be described in further detail with reference to the accompanying drawings and specific implementation methods.

[0016] Figure 1 This is a schematic diagram of the method flow for training a large-scale human resources vertical model based on HR-RAG knowledge enhancement, as described in this invention. Detailed Implementation

[0017] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] Example 1 like Figure 1 As shown, a technical solution is provided: a training method for a large-scale human resources vertical model based on HR-RAG knowledge enhancement, including: Step 1: Knowledge Structure Construction Collect relevant knowledge in the field of human resources, define the core entities and relationships within the knowledge, and form a structured knowledge set; Specifically, the core entities include positions, employees, skills, performance indicators, compensation items, relevant laws and regulations, and corporate policies; the relationships include position-skill requirements, employee-skill mastery, performance indicators-compensation linkage, and relevant laws and regulations-system constraints; the knowledge collection sources include structured data from the human resources information system, semi-structured data from employee handbooks and related documents, and unstructured data from interview records and exit interviews.

[0019] For structured data in human resources information systems, the system directly extracts standardized data such as job positions, basic employee information, payroll details, and performance ratings through its built-in data export interface. It simultaneously verifies data format consistency and removes duplicate, missing, or invalid data. For semi-structured data such as employee handbooks, salary systems, and attendance regulations, it uses text segmentation and keyword extraction technology to break down the document's hierarchical structure and extract structured entries such as "system name - applicable objects - implementation standards - responsible department." For unstructured data such as interview evaluation records, exit interview minutes, and employee feedback emails, it uses natural language understanding technology for semantic analysis to extract key information such as employee skill descriptions, job suitability evaluations, core reasons for leaving, and improvement suggestions, transforming them into associative structured data units. Simultaneously, it comprehensively covers the entire process of human resources recruitment, training, performance, compensation, and employee relations, clearly defining the specific attributes and boundaries of each entity. For example, the job entity includes core attributes such as job title, department, qualifications, responsibilities, and reporting relationships; the employee entity covers key information such as start date, educational background, skills certificates, work experience, and performance records; and the relevant legal entity clearly defines the name of the legal regulation, its effective date, and its scope of application, ensuring that the definition of each core entity is accurate and unambiguous, providing a clear foundation for building relationships. Furthermore, based on HR business logic and actual scenario rules, a clear mapping of relationships between core entities is established (e.g., matching the job requirements in the job description with the needs of the skill entities), clarifying the core skills and proficiency required for each position, determining the linkage between performance indicator entities and compensation item entities according to the performance management system, setting the corresponding salary adjustment ratios for different performance levels, establishing a "constraint" relationship between relevant legal entities and corporate system entities, marking the legal basis for the system clauses, and constructing a "mastery" relationship between employee entities and skill entities through employee skill files, recording the time of skill acquisition and proficiency level of employees; Finally, the sorted core entities, attributes, and relationships are imported into the HR vertical dynamic knowledge graph for storage. The graph database uses a node-edge structure, with each core entity as an independent node and the relationships between entities as edges. A knowledge update trigger mechanism is also set up so that when new regulations are introduced, corporate policies are revised, job positions are adjusted, or employee information changes, the graph update process is triggered, and the corresponding nodes and edges are added, modified, or deleted synchronously to ensure that the structured knowledge set can reflect the latest dynamics of HR business in real time.

[0020] Step 2: Design search enhancement strategies Based on the basic similarity calculation method, multi-dimensional human resources scenario-related weights are introduced to construct a multi-dimensional retrieval strategy for searching and filtering the knowledge set; Specifically, the weighting dimensions of the multi-dimensional retrieval strategy include job level weight, knowledge timeliness weight, and business relevance weight; the retrieval method adopts a combined sorting approach of semantic retrieval, structured retrieval, and timeliness retrieval.

[0021] First, the cosine similarity formula is used as the basic similarity calculation method. The input requirements of the model (such as resume screening requirements, salary calculation consultation, etc.) and the knowledge units in the structured knowledge set (such as job skill requirements, salary system clauses, etc.) are transformed into standardized vector forms respectively. The semantic matching degree between the two is accurately calculated by the following formula, as shown below: In the formula, For input requirements, This serves as a knowledge unit, providing an objective basis for judgment in subsequent retrieval and filtering. Then, considering the characteristics of human resources business scenarios, specific application rules for the three categories of weights—job level, knowledge timeliness, and business relevance—are clarified. For example, the weight of job level is set according to the differences in job management level; for instance, the weight coefficient of knowledge related to senior management positions is higher than that of entry-level positions, ensuring that the needs of core positions are matched first. The weight of knowledge timeliness is set according to the knowledge update time (for instance, the weight coefficient of regulations and corporate systems revised within the past year is higher than that of knowledge that has not been updated for more than three years, ensuring that the search results are in line with the latest requirements). The weight of business relevance is set according to the closeness of the connection between knowledge and the current task (for instance, in payroll calculation tasks, the weight coefficient of payroll items and performance-linked rules is higher than that of training-related knowledge). At the same time, a dynamic weight calibration mechanism is established to fine-tune the coefficient values ​​based on actual business feedback. Furthermore, in achieving the synergistic integration of semantic retrieval, structured retrieval, and time-sensitive retrieval, semantic retrieval, based on the aforementioned cosine similarity calculation results, filters out knowledge with a high degree of semantic fit with the input requirements. Structured retrieval, based on the precise matching logic of core entities and relationships, extracts the associated knowledge (e.g., job skill requirements, salary calculation standards) corresponding to entities in the input requirements (e.g., specific positions, salary items). Time-sensitive retrieval is sorted according to the knowledge update timestamp, prioritizing the locking of recently updated dynamic knowledge such as laws and regulations. Subsequently, the three types of retrieval results are assigned corresponding weight values, and a comprehensive ranking score is obtained through weighted summation, achieving the orderly integration of multi-dimensional retrieval results. Finally, a threshold standard for retrieval and filtering was set to remove low-matching knowledge below the threshold in the comprehensive ranking, retaining only highly relevant and effective knowledge units. At the same time, the matching details of each retrieval were recorded, including the similarity score of each knowledge unit, the weight distribution, and the contribution ratio of the retrieval method, forming a retrieval log archive. This provides complete data support for subsequent optimization of retrieval strategies based on mismatch cases, ensuring that the filtered knowledge can accurately meet the model input requirements and laying a solid quality foundation for subsequent deep integration.

[0022] Step 3: Association Coding and Fusion The retrieved and filtered knowledge is correlated with the model input requirements, and a hierarchical coding method is used to achieve deep integration between the two. Specifically, the hierarchical coding adopts a dual-path coding structure, including a conventional semantic coding path for questions and a structured coding path for retrieved knowledge entities and relationships.

[0023] The conventional semantic encoding path for questions involves first performing HR business scenario semantic preprocessing on the natural language requirements input to the model (e.g., "generating interview questions for technical positions" or "calculating quarterly performance-based compensation for employees"). This process identifies the task type, core demands, and key constraints within the requirements. Then, the preprocessed natural language requirements are transformed into standardized semantic vectors. During this transformation, the task intent (e.g., generating interview questions, calculating compensation), core entity references (e.g., technical positions, quarterly performance), and business scenario characteristics (e.g., interview assessment, compensation settlement cycle) are carefully preserved to ensure the encoding results accurately reflect the core semantics of the input requirements, providing a clear intent guide for integration with retrieval knowledge. The structured encoding path for knowledge entities and relationships involves structuring and parsing the effective knowledge units after retrieval and filtering. It extracts core entities (such as job titles, skills, performance indicators, and salary items) and related relationships (such as job-skill requirements and performance indicator-salary linkage). Following the entity-relationship logic of HR vertical knowledge graphs, the knowledge units are transformed into structured vectors containing entity identifiers, relationship types, and attribute information. For example, for the knowledge unit "Technical positions require Python programming skills," the encoding must clearly define the entities "Technical Positions" and "Python Programming Skills," the relationship type "Skill Requirements," and the attribute "Required Skills," ensuring that the encoding results fully preserve the structured features and logical connections of the knowledge. Simultaneously, bidirectional association recognition is performed. By comparing the core entities and task types in the conventional semantic encoding of the question with the entities and relation types in the structured encoding of the retrieval knowledge, a precise mapping relationship between the two is established. For example, if the core entity extracted from the question encoding is "sales position" and the task type is "resume screening," then the structured encoding results containing the entity "sales position" and the relational relationships "skill requirements" and "job qualifications" in the retrieval knowledge encoding are matched. At the same time, the core association points with the highest matching degree (such as "customer negotiation skills" and "sales performance requirements") are marked to ensure that the results of the two encoding paths are highly consistent in terms of business logic and entity association. Finally, a semantic enhancement fusion mechanism is adopted to achieve deep integration. The semantic encoding result of the question is taken as the core requirement, and the structured encoding result of the retrieved knowledge is taken as the knowledge support basis. Through semantic association weight allocation, the structured encoding information with high relevance to the core requirement of the question is integrated to form a unified fusion encoding output. For example, after the question encoding of "generating technical job interview questions" is fused with the structured encoding of "technical job requires Python skills and project experience", the output includes the task intent of "interview question generation", the entity of "technical job", and "Python skills + project experience". This not only retains the task orientation of the input requirement, but also deeply embeds the structured logic of the retrieved knowledge, providing highly adapted encoding input for subsequent model training and task execution, breaking the disconnect between the retrieval and generation links.

[0024] Step 4: Task Alignment and Incremental Training Specific training objectives are designed for human resources-related tasks, incremental training is employed, and multiple rounds of iterative optimization of model parameters are used. Specifically, human resources-related tasks include resume screening and evaluation, interview question generation, performance report writing, payroll calculation, and employee turnover risk prediction; incremental training uses efficient parameter fine-tuning technology, freezing the underlying parameters of the model and only adjusting the parameters of the upper adaptation layer.

[0025] First, differentiated and actionable training objectives were developed for each human resources task. The resume screening and evaluation task focused on the precise matching of job-skill requirements with employee-skill mastery. The training objective was to extract key information such as employee skills and work experience from resumes, compare it with job qualifications, and output a fit score and core matching / difference points. The interview question generation task centered on job skill requirements. The training objective was to generate structured interview questions covering core skills, business scenario applications, and professional qualities, with standardized and targeted question wording. The performance report writing task revolved around performance indicators and employee performance. The training objective was to generate objective, detailed, and compliant performance evaluation reports based on performance data and job responsibilities, including highlights and areas for improvement. The payroll calculation task closely followed performance indicators and pay-related content. The training objective was to input employee performance data and pay standards (e.g., base salary, performance bonus) to complete detailed calculations of payable salary, subsidies, deductions, etc., and output accurate results. The employee turnover risk prediction task was based on employee data, historical turnover cases, and correlations. The training objective was to identify risk warnings and output risk levels (e.g., high / medium / low) and potential cause analysis. Then, incremental training is performed: first, the model's pre-training parameters are locked, and only the adaptation parameters are opened for adjustment to avoid damage to the general semantic understanding ability and reduce the computational cost of training. The training data is based on the fused data after association encoding in step 3, covering the structured knowledge corresponding to each task (such as salary system clauses and performance indicator definitions), real business data (such as historical resumes, past performance reports, and salary calculation records), and synchronously accessing the updated knowledge in the dynamic knowledge graph of the HR vertical (such as newly revised salary items and newly added job skill requirements), to ensure that the training data is both close to the actual business and keeps up with knowledge iteration. During the training process, the focus is on strengthening the mapping learning of core entities and relationships. For example, for the salary calculation task, the focus is on training the model's ability to understand and apply the relationship between "performance indicator score - salary coefficient" and "salary item - legal constraint". Finally, the model parameters are optimized through multiple rounds of iteration to ensure continuous improvement in training effectiveness. After each round of training, the core evaluation indicators for each task (resume screening accuracy, interview question relevance, performance report compliance, salary calculation accuracy, and turnover prediction accuracy) are evaluated. If a task indicator fails to meet the preset standard, the root cause of the problem is analyzed (e.g., deviation in skill matching logic, or incomplete understanding of salary rules). The adaptation parameters and training data weights are adjusted accordingly. For example, if the salary calculation accuracy is low, the proportion of training samples with similar salary rules is increased. During the iteration process, the dynamic knowledge graph update mechanism of the HR vertical category is linked. When new relevant regulations are introduced, corporate systems are revised, or positions are adjusted, incremental training iteration is triggered to quickly integrate the new knowledge into the model parameters, ensuring that the model always adapts to the latest business scenarios and knowledge requirements, and ultimately achieving efficient and accurate execution of various human resources tasks.

[0026] Step 5: Adversarial Verification and Optimization The knowledge set and retrieval strategy are optimized through adversarial sample validation and knowledge consistency verification. Specifically, adversarial samples are generated for high-risk HR scenarios; knowledge consistency verification compares the model-generated content with authoritative knowledge in the knowledge set; and retrieval strategy optimization adjusts weight values ​​based on retrieval mismatch cases.

[0027] Firstly, adversarial samples generated for high-risk HR scenarios cover risk points such as payroll compliance, regulatory applicability, performance evaluation fairness, and misjudgment of employee turnover risk. For example, for payroll scenarios, interfering samples such as confusing individual income tax deduction standards and performance bonus ratios are generated. For regulatory applicability scenarios, fuzzy samples are generated to address the intersection of old and new regulations, conflicts between regulations in different regions, and regulatory adaptation for special employee groups (such as probationary periods and pregnant women). For performance evaluation scenarios, misleading samples such as ambiguous performance indicator definitions, ambiguous descriptions of employee performance, and inconsistent evaluation standards are generated. These adversarial samples are input into the trained model to observe whether there are compliance errors, logical deviations, or risk omissions in the model's output results. At the same time, the types and frequencies of misjudgments by the model in high-risk scenarios are recorded to assess the model's robustness shortcomings. Then, knowledge consistency verification is carried out to establish a two-way comparison mechanism between "model output and authoritative knowledge": the content generated by the model in various human resources tasks is extracted, including resume fit scores, interview question lists, performance report texts, salary calculation details, turnover risk analysis reports, etc., and these contents are compared one by one with the authoritative knowledge in the dynamic knowledge graph of HR verticals. The salary calculation results are compared with salary item standards, regulations and corporate salary systems, the performance reports are compared with the definitions of performance indicators, evaluation standards and job responsibilities, and the interview questions are compared with job skill requirements and compliance requirements. Inconsistent content is marked, the reasons for the deviation are analyzed, and it is distinguished whether it is due to information loss caused by inaccurate knowledge retrieval or errors in the model's understanding and application of knowledge. Finally, optimizations were made based on the verification results to improve the knowledge set and retrieval strategy. For cases of mismatched retrieval, such as unreasonable weighting of job level leading to insufficient priority for core job knowledge retrieval, or imbalance of business relevance weighting leading to the filtering of task-irrelevant knowledge, the values ​​of the three weights of job level, knowledge timeliness, and business relevance were adjusted, and the fusion and sorting logic was optimized. For knowledge gaps, outdated or conflicting issues found in the knowledge consistency verification, the dynamic knowledge graph of the HR vertical category was updated, and knowledge such as newly added clauses, revised corporate systems, and newly added job skill requirements was added. Expired old knowledge was deleted, and incorrect mappings in entity relationships were corrected. After optimization, adversarial samples and business data were input again for verification. The "verification-analysis-optimization" process was repeated until the model's misjudgment rate in high-risk scenarios was lower than the preset threshold, and the knowledge consistency standard was met, ensuring that the authority, accuracy, and compliance of the model output continued to improve.

[0028] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A training method for a large-scale human resources vertical model based on HR-RAG knowledge enhancement, characterized in that... : S1. Collect relevant knowledge in the field of human resources, define the core entities and relationships in the knowledge, and form a structured knowledge set; S2. Based on the basic similarity calculation method, multi-dimensional human resources scenario-related weights are introduced to construct a multi-dimensional retrieval strategy for searching and filtering the knowledge set; S3. The retrieved and filtered knowledge is associated with the model input requirements, and a hierarchical coding method is used to achieve deep integration between the two; S4. Design specific training objectives for human resources-related tasks, adopt incremental training, and combine multiple rounds of iterative optimization of model parameters; S5. Optimize the knowledge set and retrieval strategy through adversarial sample verification and knowledge consistency verification.

2. The method for training a large-scale human resource vertical model based on HR-RAG knowledge enhancement according to claim 1, characterized in that... In S1, the core entities include positions, employees, skills, performance indicators, compensation items, relevant laws and regulations, and corporate systems; the relationships include positions-skill requirements, employees-skill mastery, performance indicators-compensation linkage, and relevant laws and regulations-system constraints; the knowledge collection sources include structured data from the human resources information system, semi-structured data from employee handbooks and related documents, and unstructured data from interview records and exit interviews.

3. The method for training a large-scale human resource vertical model based on HR-RAG knowledge enhancement according to claim 1, characterized in that... In S2, the weight dimensions of the multi-dimensional retrieval strategy include job level weight, knowledge timeliness weight, and business relevance weight; the retrieval method adopts a combined sorting method of semantic retrieval, structured retrieval, and timeliness retrieval.

4. The method for training a large-scale human resource vertical model based on HR-RAG knowledge enhancement according to claim 1, characterized in that... In S3, the hierarchical encoding adopts a dual-path encoding structure, including a conventional semantic encoding path for questions and a structured encoding path for retrieved knowledge entities and relations.

5. The method for training a large-scale human resource vertical model based on HR-RAG knowledge enhancement according to claim 1, characterized in that... In S4, human resources-related tasks include resume screening and evaluation, interview question generation, performance report writing, salary calculation, and employee turnover risk prediction; incremental training adopts efficient parameter fine-tuning technology, freezing the underlying parameters of the model and only adjusting the parameters of the upper adaptation layer.

6. The method for training a large-scale human resource vertical model based on HR-RAG knowledge enhancement according to claim 1, characterized in that... In S5, adversarial samples are generated for high-risk HR scenarios; knowledge consistency verification compares the model-generated content with authoritative knowledge in the knowledge set; and retrieval strategy optimization adjusts weight values ​​based on retrieval mismatch cases.

7. The method for training a large-scale human resource vertical model based on HR-RAG knowledge enhancement according to claim 1, characterized in that... The structured knowledge set is stored using a dynamic knowledge graph for HR verticals.

Images