Intelligent resume screening method and system based on dynamic rule fusion

Abstract

The invention relates to the technical field of natural language processing and human resources, in particular to an intelligent resume screening method and system based on dynamic rule fusion, and the method comprises the following steps: analyzing post description to generate a dynamic scoring rule, and executing large model semantic scoring and rule condition scoring in parallel, constructing a reinforcement learning optimization mechanism based on historical artificial decision feedback; the method has the advantages that through a rule generation module driven by post characteristics, a two-channel scoring engine (large model prediction and artificial rules) and an autonomous optimization system based on reinforcement learning, interpretable scoring rules are automatically generated based on post description; a weighted fusion mechanism is adopted to balance model generalization and rule controllability; and a continuous optimization scoring strategy is fed back through historical manual operation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing and human resources, in particular to an intelligent resume screening method and system based on dynamic rule fusion. BACKGROUND

[0002] Traditional ATS systems rely on fixed keyword matching and cannot adapt to the differences in ability dimensions of positions such as "Java development" and "financial audit".

[0003] Although the pure model solution can process unstructured data, it has compliance risks such as "education discrimination". The HR historical scoring records are not systematically utilized, resulting in the inability to pass on knowledge.

[0004] The contradiction between the efficiency needs of enterprise batch recruitment and the accuracy / compliance of talent screening is increasingly prominent. SUMMARY

[0005] The present application aims to provide an intelligent resume screening method and system based on dynamic rule fusion to solve the problems raised in the background.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solution: an intelligent resume screening method based on dynamic rule fusion, comprising the following steps:

[0007] Analyzing the job description to generate dynamic scoring rules, specifically using a knowledge graph to associate job JDs with industry capability standard libraries and determining the initial weights of each capability dimension through an attention mechanism;

[0008] Parallelly executing large model semantic scoring and rule condition scoring, while dynamically adjusting the weight coefficients of the two channels, wherein beta = 0.8-0.005 x the number of hard conditions;

[0009] Based on historical artificial decision feedback, an reinforcement learning optimization mechanism is constructed to continuously optimize the dynamic scoring rules, large model semantic scoring, and rule condition scoring processes.

[0010] Preferably, in the step of using a knowledge graph to associate job JDs with industry capability standard libraries, first, collect job JDs of different industries and corresponding industry capability standard information to construct a knowledge graph; then, use natural language processing technology to analyze the input job JDs, extract key information, match and associate corresponding industry capability standards in the knowledge graph, and form a mapping relationship between job descriptions and capability standards.

[0011] Preferably, in the step of determining the initial weight of each capability dimension by the attention mechanism, each capability dimension associated with the post JD is taken as input, and the attention mechanism model is used to calculate the attention score of each capability dimension relative to the post demand; the attention score is normalized, and the normalized score is taken as the initial weight of each capability dimension.

[0012] Preferably, in the step of performing the large model semantic scoring and the rule condition scoring in parallel, the large model semantic scoring adopts a pre-trained large language model to perform semantic similarity calculation on the resume text and the post description to obtain a semantic scoring result; the rule condition scoring matches and scores the hard conditions and the key capability information in the resume according to the generated dynamic scoring rule to obtain a rule condition scoring result; and the semantic scoring result and the rule condition scoring result are weighted and fused according to the double-channel weight coefficient β after dynamic adjustment to obtain the final resume score.

[0013] Preferably, in the step of constructing the reinforcement learning optimization mechanism based on historical artificial decision feedback, historical artificial decision data in the resume screening process is collected, including resume information, artificial score and final hiring result; the parameters of the dynamic scoring rule, the parameters of the large model semantic scoring and the parameters of the rule condition scoring are taken as the state space, the adjusted parameters are taken as the action space, and the accuracy and consistency of the artificial decision are taken as the reward function; the state space, the action space and the reward function are trained by using the reinforcement learning algorithm to continuously optimize the parameters, so that the intelligent resume screening method is more in line with the logic and standard of the artificial decision.

[0014] A system for an intelligent resume screening method based on dynamic rule fusion, comprising:

[0015] A dynamic scoring rule generation module for generating dynamic scoring rules by analyzing post descriptions, which includes a knowledge graph association unit and a weight determination unit; the knowledge graph association unit associates the post JD with the industry capability standard library using the knowledge graph; and the weight determination unit determines the initial weight of each capability dimension by the attention mechanism;

[0016] A double-channel scoring execution module for performing large model semantic scoring and rule condition scoring in parallel, and the module includes a weight adjustment unit that dynamically adjusts the double-channel weight coefficient β, where β = 0.8-0.005 x the number of hard conditions;

[0017] A reinforcement learning optimization module for constructing a reinforcement learning optimization mechanism based on historical artificial decision feedback to optimize and adjust the dynamic scoring rule generation module and the double-channel scoring execution module.

[0018] Preferably, the knowledge graph association unit specifically performs the following operations:

[0019] Pre-construct a knowledge graph containing different industry post JDs and corresponding industry capability standard information;

[0020] Receive the input post JD, parse it using natural language processing technology, and extract key information;

[0021] In the knowledge graph, match according to the extracted key information, associate the corresponding industry capability standard, form the mapping relationship between the post description and the capability standard, and provide basic data for dynamic scoring rule generation.

[0022] Preferably, the weight determination unit determines the initial weight of each capability dimension through an attention mechanism as follows: input each capability dimension information associated with the post JD into the attention mechanism model; the attention mechanism model calculates the attention score of each capability dimension relative to the post demand; normalize the calculated attention score, and use the normalized score as the initial weight of each capability dimension for evaluation of different capability dimensions in the dynamic scoring rule.

[0023] Preferably, the double-channel scoring execution module includes: a large model semantic scoring unit that uses a pre-trained large language model to calculate the semantic similarity between the resume text and the post description, obtaining a semantic scoring result;

[0024] A rule condition scoring unit matches and scores the hard conditions and key capability information in the resume according to the dynamic scoring rule generated by the dynamic scoring rule generation module, obtaining a rule condition scoring result;

[0025] A scoring fusion unit receives the double-channel weight coefficient β adjusted by the weight adjustment unit, and performs weighted fusion on the semantic scoring result obtained by the large model semantic scoring unit and the rule condition scoring result obtained by the rule condition scoring unit, obtaining the final resume score.

[0026] Preferably, the process of constructing the reinforcement learning optimization mechanism by the reinforcement learning optimization module is as follows: the data collection unit collects historical artificial decision data in the resume screening process, including resume information, artificial scoring, and final hiring results; the model construction unit constructs a reinforcement learning model using the parameters of the dynamic scoring rule generation module and the parameters of the double-channel scoring execution module as the state space, using the adjusted parameters as the action space, and using the accuracy and consistency of artificial decision as the reward function; the training optimization unit trains the reinforcement learning model using the collected historical artificial decision data, continuously adjusts the parameters in the action space, and makes the system perform better in the state space to meet the requirements of the reward function, thereby optimizing and adjusting the dynamic scoring rule generation module and the double-channel scoring execution module.

[0027] Compared with the prior art, the beneficial effects of the present application are:

[0028] The intelligent resume screening method and system based on dynamic rule fusion provided by the application realize the following through a rule generation module driven by post characteristics, a double-channel scoring engine (large model prediction + artificial rules), and an autonomous optimization system based on reinforcement learning: automatically generating interpretable scoring rules based on post descriptions; balancing model generalization and rule controllability using a weighted fusion mechanism; continuously optimizing scoring strategies through historical manual operation feedback. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 The method flowchart of the application. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical solutions of the application clear, complete and more clear and understandable, the embodiments of the application are further described in detail below in combination with the drawings. It should be understood that the specific embodiments described herein are part of the embodiments of the application, not all embodiments, and are only used to explain the embodiments of the application, and do not limit the embodiments of the application. All other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0031] Embodiment one, please refer to Figure 1 The application provides a technical solution: an intelligent resume screening method based on dynamic rule fusion, comprising the following steps:

[0032] The post description is parsed to generate dynamic scoring rules, specifically, the job description (JD) and the industry capability standard library are associated using a knowledge graph, and the initial weights of each capability dimension are determined through an attention mechanism; in the step of associating the job description (JD) and the industry capability standard library using a knowledge graph, the job descriptions (JD) of different industries and the corresponding industry capability standard information are collected to construct a knowledge graph; then the input job description (JD) is parsed through natural language processing technology to extract key information, and the corresponding industry capability standard is matched and associated in the knowledge graph to form a mapping relationship between the post description and the capability standard. In the step of determining the initial weights of each capability dimension through an attention mechanism, each capability dimension associated with the job description (JD) is input, and the attention score of each capability dimension relative to the job requirements is calculated using an attention mechanism model; the attention score is normalized, and the normalized score is used as the initial weight of each capability dimension.

[0033] The big model semantic score and the rule condition score are executed in parallel, and a double-channel weight coefficient β is dynamically adjusted, wherein β = 0.8-0.005 x the number of hard conditions; the big model semantic score adopts a pre-trained large language model to perform semantic similarity calculation on the resume text and the post description, and a semantic score result is obtained; the rule condition score matches and scores the hard conditions and key ability information in the resume according to the generated dynamic scoring rule, and a rule condition score result is obtained; the semantic score result and the rule condition score result are weighted and fused according to the double-channel weight coefficient β after dynamic adjustment, and a final resume score is obtained.

[0034] An optimization mechanism of reinforcement learning is constructed based on historical artificial decision feedback, and the dynamic scoring rule, the big model semantic score and the rule condition score process are continuously optimized by using the optimization mechanism; the artificial decision data in the historical resume screening process are collected, including the resume information, the artificial score and the final employment result; the parameters of the dynamic scoring rule, the parameters of the big model semantic score and the parameters of the rule condition score are taken as the state space, the adjusted parameters are taken as the action space, and the accuracy and consistency of the artificial decision are taken as the reward function; the state space, the action space and the reward function are trained by using the reinforcement learning algorithm, and the parameters are continuously optimized, so that the intelligent resume screening method is more in line with the logic and standard of the artificial decision

[0035] In the embodiment two, an intelligent resume screening method based on dynamic rule fusion is provided, and the system comprises:

[0036] The dynamic scoring rule generation module is configured to parse the post description to generate a dynamic scoring rule, and the module comprises a knowledge graph association unit and a weight determination unit; the knowledge graph association unit is configured to associate the post JD with an industry capability standard library by using a knowledge graph; and the weight determination unit is configured to determine the initial weight of each capability dimension by using an attention mechanism; the knowledge graph association unit is specifically configured to: pre-construct a knowledge graph comprising different industry post JDs and corresponding industry capability standard information; receive the input post JD, parse the post JD by using a natural language processing technology, and extract key information; match the key information in the knowledge graph, associate the corresponding industry capability standard, form a mapping relationship between the post description and the capability standard, and provide basic data for dynamic scoring rule generation; and the weight determination unit is configured to determine the initial weight of each capability dimension by using an attention mechanism as follows: input the information of each capability dimension associated with the post JD into an attention mechanism model; the attention mechanism model calculates the attention score of each capability dimension relative to the post demand; and normalize the calculated attention score, and take the normalized score as the initial weight of each capability dimension for evaluation of different capability dimensions in the dynamic scoring rule.

[0037] The double-channel scoring execution module is configured to execute the large model semantic scoring and the rule condition scoring in parallel, and the module comprises a weight adjustment unit configured to dynamically adjust a double-channel weight coefficient β, where β = 0.8-0.005 x the number of hard conditions. The double-channel scoring execution module comprises: a large model semantic scoring unit configured to perform semantic similarity calculation on the resume text and the job description by using a pre-trained large language model to obtain a semantic scoring result; a rule condition scoring unit configured to perform matching and scoring on the hard conditions and the key capability information in the resume according to the dynamic scoring rules generated by the dynamic scoring rule generation module to obtain a rule condition scoring result; and a scoring fusion unit configured to receive the double-channel weight coefficient β adjusted by the weight adjustment unit, and perform weighted fusion on the semantic scoring result obtained by the large model semantic scoring unit and the rule condition scoring result obtained by the rule condition scoring unit to obtain a final resume score.

[0038] The reinforcement learning optimization module is configured to construct a reinforcement learning optimization mechanism based on historical artificial decision feedback, and to optimize and adjust the dynamic scoring rule generation module and the double-channel scoring execution module. The process of constructing the reinforcement learning optimization mechanism by the reinforcement learning optimization module comprises: a data collection unit collects artificial decision data in a historical resume screening process, including resume information, artificial scores, and final hiring results; a model construction unit constructs a reinforcement learning model by taking the parameters of the dynamic scoring rule generation module and the parameters of the double-channel scoring execution module as a state space, taking the adjusted parameters as an action space, and taking the accuracy and consistency of artificial decision as a reward function; and a training optimization unit trains the reinforcement learning model by using the collected historical artificial decision data, and continuously adjusts the parameters in the action space to make the system perform better in the state space in accordance with the requirements of the reward function, thereby optimizing and adjusting the dynamic scoring rule generation module and the double-channel scoring execution module.

[0039] In the third embodiment, the following scheme is proposed based on the second embodiment.

[0040] 1. Rule generation:

[0041] Input JD: "require mastery of Transformer architecture, have LLM optimization experience"

[0042] Generated rules:

[0043] must_have = ["BERT / Transformer", "GPU optimization"]

[0044] weighted = {"distributed training": 0.30, "model quantization": 0.25}

[0045] 2. Double-channel scoring:

[0046] Resume A: Model score 0.92 (project experience matches LLM optimization)

[0047] Rule score 76 (lack of "model quantification" experience)

[0048] Combined score: 0.65 x 0.92 + 0.35 x 0.76 = 0.861 → top recommendation

[0049] 3. Autonomous optimization:

[0050] HR manually boosts "distributed training" weight to 0.40

[0051] System automatically reduces model dependency for unmodified dimensions.

[0052] While embodiments of the present application have been shown and described, it is to be understood that the embodiments can be varied, modified, substituted and changed by those skilled in the art without departing from the principles and spirit of the application, the scope of which is defined by the claims and their equivalents.

Claims

1. An intelligent resume screening method based on dynamic rule fusion, characterized in that: Includes the following steps: The job description is analyzed to generate dynamic scoring rules. Specifically, a knowledge graph is used to link the job description with the industry competency standard library, and an attention mechanism is used to determine the initial weight of each competency dimension. The semantic scoring and rule-based condition scoring of the large model are performed in parallel, while the dual-channel weight coefficient β is dynamically adjusted, where β = 0.8 - 0.005 × the number of hard conditions; A reinforcement learning optimization mechanism is constructed based on historical human decision feedback, and this mechanism is used to continuously optimize dynamic scoring rules, large model semantic scoring, and rule-based conditional scoring processes.

2. The intelligent resume screening method based on dynamic rule fusion according to claim 1, characterized in that: The steps of using a knowledge graph to link job descriptions (JDs) with industry competency standards involve first collecting job descriptions and corresponding industry competency standards from different industries to construct a knowledge graph; then, using natural language processing technology to parse the input job descriptions, extract key information, and match and associate them with the corresponding industry competency standards in the knowledge graph to form a mapping relationship between job descriptions and competency standards.

3. The intelligent resume screening method based on dynamic rule fusion according to claim 2, characterized in that: In the step of determining the initial weights of each ability dimension through the attention mechanism, each ability dimension associated with the job description is taken as input, and the attention mechanism model is used to calculate the attention score of each ability dimension relative to the job requirements. The attention scores are normalized, and the normalized scores are used as the initial weights for each ability dimension.

4. The intelligent resume screening method based on dynamic rule fusion according to claim 3, characterized in that: In the parallel execution of large-scale model semantic scoring and rule-based conditional scoring, the large-scale model semantic scoring uses a pre-trained large language model to calculate the semantic similarity between the resume text and the job description, obtaining the semantic scoring result; the rule-based conditional scoring matches and scores the hard conditions and key ability information in the resume according to the generated dynamic scoring rules, obtaining the rule-based conditional scoring result; and the semantic scoring result and the rule-based conditional scoring result are weighted and fused according to the dynamically adjusted dual-channel weight coefficient β to obtain the final resume score.

5. The intelligent resume screening method based on dynamic rule fusion according to claim 4, characterized in that: The construction steps of the reinforcement learning optimization mechanism based on historical human decision feedback involve collecting human decision data from the historical resume screening process, including resume information, human scores, and final hiring results. The parameters of the dynamic scoring rules, the parameters of the large model semantic scoring, and the parameters of the rule-based conditional scoring are used as the state space, the adjusted parameters as the action space, and the accuracy and consistency of human decisions as the reward function. Reinforcement learning algorithms are then used to train the state space, action space, and reward function, continuously optimizing the parameters to make the intelligent resume screening method more consistent with the logic and standards of human decision-making.

6. A system for the intelligent resume screening method based on dynamic rule fusion according to claim 5, characterized in that: include: The dynamic scoring rule generation module is used to parse job descriptions and generate dynamic scoring rules. This module includes a knowledge graph association unit and a weight determination unit. The knowledge graph association unit uses the knowledge graph to associate job descriptions (JDs) with an industry competency standard library; the weight determination unit determines the initial weights of each competency dimension through an attention mechanism. A dual-channel scoring execution module is used to execute semantic scoring and rule-based condition scoring of a large model in parallel. This module includes a weight adjustment unit, which dynamically adjusts the dual-channel weight coefficient β, where β = 0.8 - 0.005 × the number of hard conditions. The reinforcement learning optimization module is used to build a reinforcement learning optimization mechanism based on historical human decision feedback, and to optimize and adjust the dynamic scoring rule generation module and the dual-channel scoring execution module.

7. The system according to claim 6, characterized in that: The knowledge graph association unit performs the following operations: Pre-construct a knowledge graph containing job descriptions (JDs) for different industries and corresponding industry competency standards; The system receives the job description (JD) input and uses natural language processing technology to parse it and extract key information. Based on the extracted key information in the knowledge graph, matching is performed and corresponding industry competency standards are associated to form a mapping relationship between job descriptions and competency standards, providing basic data for the generation of dynamic scoring rules.

8. The system according to claim 7, characterized in that: The process by which the weight determination unit determines the initial weights of each capability dimension through the attention mechanism is as follows: the information of each capability dimension associated with the job description is input into the attention mechanism model; The attention mechanism model calculates the attention score for each competency dimension relative to the job requirements; The calculated attention scores are normalized, and the normalized scores are used as the initial weights for each ability dimension in the dynamic scoring rules for evaluating different ability dimensions.

9. A system according to claim 8, characterized in that: The dual-channel scoring execution module includes: a large-model semantic scoring unit, which uses a pre-trained large language model to calculate the semantic similarity between the resume text and the job description, and obtains the semantic scoring results; The rule-based scoring unit matches and scores the hard conditions and key competency information in the resume based on the dynamic scoring rules generated by the dynamic scoring rule generation module, and obtains the rule-based scoring results. The scoring fusion unit receives the dual-channel weight coefficient β adjusted by the weight adjustment unit, and performs weighted fusion of the semantic scoring result obtained by the large model semantic scoring unit and the rule-condition scoring result obtained by the rule-condition scoring unit to obtain the final resume score.

10. A system according to claim 9, characterized in that: The process of constructing a reinforcement learning optimization mechanism in the reinforcement learning optimization module is as follows: The data collection unit collects human decision data from the historical resume screening process, including resume information, human scores, and final hiring results; The model construction unit uses the parameters of the dynamic scoring rule generation module and the parameters of the dual-channel scoring execution module as the state space, the adjusted parameters as the action space, and the accuracy and consistency of human decisions as the reward function to construct a reinforcement learning model. The training optimization unit uses collected historical human decision-making data to train the reinforcement learning model. By continuously adjusting the parameters in the action space, the system's performance in the state space is made to better meet the requirements of the reward function, thereby optimizing and adjusting the dynamic scoring rule generation module and the dual-channel scoring execution module.

Images