Intelligent matching method of human and post based on two-way hierarchical adaptation and double closed-loop adjustment

Abstract

The application provides a post-personnel intelligent matching method based on bidirectional hierarchical adaptation and double closed-loop adjustment, relates to the technical field of human resources, and comprises the following steps: constructing a hierarchical structure of post requirements and candidate ability through hierarchical division and performing bidirectional mapping to obtain a multi-dimensional score. Through the double closed-loop mechanism of real-time adjustment and offline optimization, the dimension weight is dynamically adjusted and the comprehensive matching score is calculated, and when the core dimension does not meet the constraint, the attenuation weight is applied to trigger the elimination judgment. The method realizes the accurate quantification and dynamic optimization of post-personnel matching, and improves the recruitment efficiency and quality.

Description

Technical Field

[0001] This invention relates to the field of human resources technology, and in particular to an intelligent matching method for people and positions based on bidirectional hierarchical adaptation and dual closed-loop weight adjustment. Background Technology

[0002] In the field of digital human resource management, job-person matching is a core element in improving recruitment efficiency and quality. Current technologies typically rely on keyword extraction and static rule matching between job requirements and candidate resumes. The conventional approach involves pre-setting a series of competency tags or keywords and generating a matching score by calculating the frequency of related words or simple semantic similarity in the resume text. Some more advanced systems introduce weighted models based on historical data, assigning fixed or adjustable weight coefficients to different skills or experience items, and finally obtaining a comprehensive matching score through weighted summation. This type of method constitutes the mainstream technical path for current automated screening and preliminary assessment.

[0003] However, conventional technical approaches still have several inherent flaws. The matching process is often unidirectional and flat, lacking in-depth analysis of the inherent hierarchical structure of needs and capabilities. Job requirements typically include multiple levels of intensity, such as core prerequisites, important reference items, and additional expectations, while candidates' capabilities also encompass different dimensions, including specific actions, outputs, and influence. Existing methods struggle to effectively distinguish these levels, leading to matching results that may overlook missing core qualifications due to over-matching of secondary features, or miss high-potential unconventional candidates due to strict matching of core qualifications, resulting in insufficient accuracy and flexibility in matching. The weight configuration mechanisms of existing systems are usually quite rigid, with weights largely based on expert experience or obtained through limited offline training, making it difficult to adapt to dynamically changing recruitment scenarios, the specific preferences of different business departments, and real-time feedback information during the recruitment process. This static weighting cannot effectively handle scenario-specific conflicts, nor can it perform closed-loop optimization and dynamic adjustment based on subsequent recruitment results and employee performance data, thus limiting the model's long-term adaptability and predictive accuracy. Summary of the Invention

[0004] This invention provides a human-job intelligent matching method based on bidirectional hierarchical adaptation and dual closed-loop weight adjustment, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a human-job intelligent matching method based on bidirectional hierarchical adaptation and dual closed-loop weighting, comprising: Obtain the job requirements text and candidate competency text to be matched; The job requirement text is hierarchically divided based on the semantic features of requirement intensity to obtain the requirement hierarchy structure; action features and result features are extracted from the candidate's ability text, an action-result association matrix is ​​constructed and the responsibility dominance index is calculated, and the ability hierarchy leap is determined based on the responsibility dominance index and the result quantification value to obtain the ability hierarchy structure. The requirement hierarchy structure and the capability hierarchy structure are bidirectionally adapted and mapped to obtain the core dimension score, secondary dimension score and bonus dimension score. Through the real-time weight adjustment channel, the target weight configuration is retrieved from the weight template library based on the conflict detection results of the recruitment process state vector and the scenario weight configuration; through the offline optimization channel, historical recruitment result data and performance data are collected to calculate the prediction deviation and generate weight correction amount to be sent back to the weight template library; based on the scenario weight configuration, the core dimension score, secondary dimension score and bonus dimension score are calculated to generate a comprehensive matching score. When the core dimension score is lower than the preset core constraint threshold, a decay weight is applied to the scores of other dimensions to bring the comprehensive matching score to the zero range and trigger the elimination decision; otherwise, the comprehensive matching score is output.

[0006] In one optional embodiment, the job requirement text is hierarchically divided based on the semantic features of requirement intensity to obtain a requirement hierarchy structure; action features and outcome features are extracted from the candidate's ability text, an action-outcome correlation matrix is ​​constructed, and a responsibility dominance index is calculated. Based on the responsibility dominance index and the quantitative value of the outcome, a capability hierarchy transition is determined, resulting in a capability hierarchy structure including: Semantic parsing is performed on the job requirement text to extract requirement nodes and modification nodes, construct a requirement semantic graph, and calculate edge weights based on the syntactic features of modification nodes to determine the semantic features of requirement intensity. A requirement propagation attenuation factor is introduced to reduce the requirement intensity on multi-hop paths. Based on the adjusted semantic features of requirement intensity, the requirement nodes are divided into a requirement hierarchy structure. Verb phrases are extracted from the candidate's ability text as action features, and quantitative expressions and achievement nouns are extracted as achievement features. A two-layer structure containing intra-sentence dependency distance matrix and cross-sentence semantic similarity matrix is ​​established to obtain the action-achievement association matrix. The dominant action pattern vector is extracted by performing singular value decomposition on the intra-sentence dependency distance matrix. The contribution rate of the dominant action pattern vector to the total variance is calculated to determine the responsibility dominance index. The quantitative descriptions in the outcome characteristics are normalized to obtain the outcome quantification value. A capability level leap decision tree is constructed with the responsibility dominance index as the first branch node and the outcome quantification value as the second branch node. The output determines whether the capability description should leap to a higher capability level, forming a capability level structure, including the core necessary requirement level, the secondary adaptation requirement level, and the bonus requirement level.

[0007] In one alternative embodiment, introducing a demand propagation attenuation factor to adjust the demand intensity on multi-hop paths includes: In the demand semantic graph, a set of directed propagation paths from modifier nodes to demand nodes is constructed. For each propagation path, the intermediate node type sequence is extracted. The semantic decay mode is identified based on the intermediate node type sequence. The semantic decay mode includes the negation word blocking mode, the condition word weakening mode, and the conjunction word transmission mode. For the identified semantic decay patterns, a first decay factor is set for the negation word blocking pattern, a second decay factor is set for the condition word weakening pattern, and a third decay factor is set for the connector word propagation pattern. The first decay factor is less than the second decay factor, and the second decay factor is less than the third decay factor. The corresponding decay factors are applied to the edge weights on the propagation path. When there are multiple propagation paths for the same demand node, the propagation paths with different semantic decay modes are prioritized. The decay result of the path with the negative word blocking mode is used first to adjust the demand intensity. When there is no negative word blocking mode, the decay results of the condition word weakening mode and the connector word propagation mode are combined to obtain the adjusted demand intensity semantic features of the demand node.

[0008] In one optional embodiment, retrieving the target weight configuration from the weight template library based on the conflict detection results between the recruitment process state vector and the scenario weight configuration via a real-time weight adjustment channel includes: Extract process stage type identifiers, candidate pool size characteristics, and job urgency characteristics from the current recruitment process stage to construct a recruitment process state vector. Cross-validate the recruitment process state vector with multiple scenario weight configurations in the weight template library. For each scenario weight configuration, backtrack the weight application results of similar recruitment process states in history and extract abnormal samples and corresponding abnormal feature labels that cause matching failures in the weight application results. Construct a conflict detection matrix between the recruitment process state vector and the abnormal feature label. When the recruitment process state vector has a feature component that overlaps with the abnormal feature label, mark the corresponding scenario weight configuration with risk and remove it from the candidate range. Select the configuration with the highest matching degree with the recruitment process state vector feature from the scenario weight configurations that have not been marked with risk as the target weight configuration. The target weight configuration is retrieved from the weight template library through the real-time weight adjustment channel, thus completing the dynamic retrieval of the scene weight configuration.

[0009] In one optional embodiment, the process of collecting historical recruitment results data and performance data through an offline optimization channel, calculating the prediction deviation, and generating a weight correction amount to be sent back to the weight template library includes: Extract matching prediction records and actual hiring records from historical recruitment results data, calculate prediction deviation, and construct a delayed correlation graph by linking the prediction deviation with the candidate's performance data after joining the company according to timestamps. In the delayed correlation graph, count the time interval between the occurrence of prediction deviation and the performance fluctuation, identify the lag effect period, extract the weight value sequence of weight allocation for each job competency dimension within the lag effect period to form a weight allocation trajectory, extract the score change sequence of each job competency dimension in the performance data to form a performance fluctuation curve, and compare the change trend of the weight allocation trajectory and the performance fluctuation curve to verify synchronization. When the weight allocation trajectory and the performance fluctuation curve show an inverse synchronous relationship, a contribution decomposition tree is constructed for the weight parameters of each dimension in the weight configuration. The transmission path of the weight parameters of each dimension to the prediction deviation is tracked in the contribution decomposition tree. The intersection nodes of weight superposition interference in the transmission path are identified. A weight decoupling factor is introduced for the intersection nodes, and the weight parameter proportion at the intersection nodes is redistributed to generate the weight correction amount. The weight correction amount and weight decoupling factor are sent back to the weight template library through the offline optimization channel to update the weight configuration.

[0010] In one optional embodiment, when the core dimension score is lower than a preset core constraint threshold, a decay weight is applied to the scores of other dimensions to converge the comprehensive matching score to the zero range and trigger an elimination decision; otherwise, the comprehensive matching score is output, including: The core dimension score is compared with the preset core constraint threshold. When the core dimension score is lower than the preset core constraint threshold, the calculation steps of other dimensions that have been executed are traced back. The related calculation nodes that depend on the core dimension score in the calculation step chain are identified. Blocking flags are injected into the related calculation nodes to stop the data flow of the calculation steps. The coupling degree of the other dimension scores that have been calculated is detected with the core dimension score. Based on the coupling degree detection results, the attenuation weight is applied to the other dimension scores to generate invalid score data. Construct a set of constraint judgment conditions for the comprehensive matching score, and use the judgment result that the core dimension score is lower than the preset core constraint threshold as the trigger condition to inject into the set of constraint judgment conditions. Perform constraint operation on the failure score data and the core dimension score based on the set of constraint judgment conditions. Force the comprehensive matching score to converge to the zero value interval through constraint operation. Generate and output the elimination judgment result based on the comprehensive matching score in the zero value interval. When the score of the core dimension is greater than or equal to the preset core constraint threshold, the score calculation process of other dimensions continues, and the scores of each dimension are summarized to form a comprehensive matching score and output.

[0011] In one optional embodiment, coupling degree detection is performed between the calculated scores of other dimensions and the core dimension scores. Based on the coupling degree detection results, a decay weight is applied to the scores of other dimensions to generate failure score data, including: Extract the job competency tags for other dimensions corresponding to the calculated scores of other dimensions, extract the job competency tags for core dimensions corresponding to the scores of core dimensions, perform tag intersection matching between job competency tags for other dimensions and job competency tags for core dimensions, count the proportion of overlapping tags in the tag intersection matching, calculate the dependence strength of other dimension scores on core dimension scores based on the proportion of overlapping tags, and use the dependence strength as the coupling degree value. The scores of other dimensions are grouped into layers according to the coupling degree. For the scores of other dimensions whose coupling degree exceeds the preset coupling threshold, the attenuation weight coefficient is calculated. The attenuation weight coefficient is set to have an inverse relationship with the coupling degree. The scores of other dimensions are weighted and calculated with the corresponding attenuation weight coefficient to generate failure score data.

[0012] A second aspect of this invention provides a human-job intelligent matching system based on bidirectional hierarchical adaptation and dual closed-loop weighting, comprising: The text acquisition unit is used to acquire the job requirements text and candidate competency text to be matched; The demand and capability assessment unit is used to hierarchically divide the job demand text based on the semantic features of demand intensity to obtain the demand hierarchy structure; extract action features and result features from the candidate capability text, construct an action-result association matrix and calculate the responsibility dominance index; and determine the capability level jump based on the responsibility dominance index and the quantitative value of the result to obtain the capability hierarchy structure. The structural mapping unit is used to perform bidirectional adaptation mapping between the demand hierarchy and the capability hierarchy to obtain core dimension scores, secondary dimension scores and bonus dimension scores. The comprehensive calculation unit is used to retrieve the target weight configuration from the weight template library through the real-time weight adjustment channel based on the conflict detection results of the recruitment process state vector and the scenario weight configuration; through the offline optimization channel, it collects historical recruitment result data and performance data to calculate the prediction deviation and generates weight correction amount to be sent back to the weight template library; and calculates the comprehensive matching score based on the scenario weight configuration for the core dimension score, secondary dimension score and bonus dimension score. The elimination decision unit is used to apply a decay weight to the scores of other dimensions when the core dimension score is lower than the preset core constraint threshold, so as to converge the comprehensive matching score to the zero value range and trigger the elimination decision; otherwise, the comprehensive matching score is output.

[0013] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] In this embodiment of the invention, a structured hierarchy of requirements is formed by hierarchically dividing the job requirement text based on semantic features of requirement intensity. This avoids misjudgments caused by semantic ambiguity in traditional keyword matching. Action features and outcome features are extracted from the candidate's ability text, and an action-outcome correlation matrix is ​​constructed to calculate the degree of responsibility dominance. This enables a quantitative assessment of the candidate's ability depth and true contribution. Combining the quantitative outcome value with the ability level leap judgment can more objectively reflect the candidate's actual ability level and effectively distinguish between superficial experience and core competencies. The introduction of a dual closed-loop weight adjustment mechanism ensures the dynamic adaptability and scenario-based accuracy of the matching model. The real-time weight adjustment channel can intelligently call the most suitable weight template according to the specific status and scenario configuration of the current recruitment process, and quickly respond to temporary adjustments in recruitment strategies. The offline optimization channel continuously calculates the prediction deviation and generates weight correction by analyzing historical recruitment results and subsequent performance data, and iteratively optimizes the weight template library, enabling the model to learn from actual employment results and continuously improve long-term prediction accuracy. A core dimension-based elimination judgment mechanism is set up to strengthen the bottom-line thinking and decision-making efficiency of matching. When a candidate's core dimension score is below the standard, the system automatically applies a decay weight to the scores of other dimensions, so that the overall score quickly converges to the elimination range. This avoids secondary or bonus advantages masking core deficiencies and ensures that all selected candidates meet the most basic and critical requirements of the position. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the intelligent matching method for personnel and positions based on bidirectional hierarchical adaptation and dual closed-loop weight adjustment in an embodiment of the present invention.

[0017] Figure 2 This is a flowchart illustrating the semantic decay and path propagation logic of demand. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0020] Figure 1 This is a flowchart illustrating the intelligent matching method for personnel and positions based on bidirectional hierarchical adaptation and dual closed-loop weight adjustment according to an embodiment of the present invention. Figure 1 As shown, the method includes: Obtain the job requirements text and candidate competency text to be matched; The job requirement text is hierarchically divided based on the semantic features of requirement intensity to obtain the requirement hierarchy structure; action features and result features are extracted from the candidate's ability text, an action-result association matrix is ​​constructed and the responsibility dominance index is calculated, and the ability hierarchy leap is determined based on the responsibility dominance index and the result quantification value to obtain the ability hierarchy structure. The requirement hierarchy structure and the capability hierarchy structure are bidirectionally adapted and mapped to obtain the core dimension score, secondary dimension score and bonus dimension score. Through the real-time weight adjustment channel, the target weight configuration is retrieved from the weight template library based on the conflict detection results of the recruitment process state vector and the scenario weight configuration; through the offline optimization channel, historical recruitment result data and performance data are collected to calculate the prediction deviation and generate weight correction amount to be sent back to the weight template library; based on the scenario weight configuration, the core dimension score, secondary dimension score and bonus dimension score are calculated to generate a comprehensive matching score. When the core dimension score is lower than the preset core constraint threshold, a decay weight is applied to the scores of other dimensions to bring the comprehensive matching score to the zero range and trigger the elimination decision; otherwise, the comprehensive matching score is output.

[0021] In one specific implementation, the text of job requirements to be matched and the text of candidate competency are obtained. The job requirements text includes skill requirements, qualifications, and job descriptions, such as "Proficient in Python programming, skilled in data analysis, and has more than 3 years of project management experience." The candidate competency text includes work experience, project experience, and skill descriptions, such as "Leaded to achieve sales exceeding 5 million, organized team training to improve team performance by 20%."

[0022] When processing job requirement text, a dependency parser is used to extract requirement nodes and modifier nodes. Taking "proficient in Python" as an example, "proficient in" is identified as the core verb and used as the requirement node, while "proficient" is used as the modifier node. Connection edges are labeled using syntactic dependency relations, and the adverb strength level, verb voice, and qualifier type of the modifier node are extracted as syntactic features. When calculating edge weights, mandatory modifiers like "must" are assigned a weight of 0.9, deliberate modifiers like "priority" are assigned a weight of 0.6, and deliberate modifiers like "understand" are assigned a weight of 0.3. The edge weights are numerical representations of the semantic features of requirement strength.

[0023] When constructing the demand semantic graph, all demand nodes are treated as vertices in the graph, and edge connections are established by calculating the semantic similarity between nodes using a skill ontology knowledge base. A demand propagation attenuation factor of 0.7 is introduced. When there is a three-hop path between two demand nodes, the demand intensity is attenuated to the original value multiplied by the cube of the attenuation factor after propagation. Based on the adjusted demand intensity values, demand nodes with a value greater than 0.7 are classified into the core essential demand level, those in the range of 0.4 to 0.7 are classified into the secondary adaptive demand level, and those below 0.4 are classified into the bonus demand level.

[0024] When processing candidate competency texts, verb phrases are extracted as action features, and quantitative expressions and achievement nouns are extracted as achievement features. Taking "led the completion of sales exceeding 5 million" as an example, "led the completion of" is identified as a verb phrase and extracted as an action feature; "5 million" is a quantitative expression, and "sales exceeding" is an achievement noun, together constituting the achievement feature. An intra-sentence dependency distance matrix is ​​constructed, and the shortest path length between action features and achievement features in the dependency tree is calculated. For cross-sentence descriptions, semantic vectors are extracted, cosine similarity is calculated, and the cosine similarity is filled into the cross-sentence semantic similarity matrix, forming a two-layer structure of the action-achievement association matrix.

[0025] Singular value decomposition is performed on the intra-sentence dependency distance matrix, and the first column of the left singular matrix is ​​extracted as the dominant action pattern vector. The ratio of the square of the corresponding singular value to the sum of the squares of all singular values ​​is calculated. When the contribution rate reaches 0.65, this value is determined as the responsibility dominance index. The quantitative descriptions in the outcome characteristics are normalized, and a capability hierarchy leap decision tree is constructed to determine whether the capability description should leap to a higher capability level, thus forming a capability hierarchy structure.

[0026] A bidirectional mapping between the requirement hierarchy and the capability hierarchy is implemented. Specifically, the requirements in the core essential requirement hierarchy are precisely matched with the corresponding levels in the candidate's capability hierarchy to obtain a core dimension score. For example, matching the core job requirement "proficient in machine learning algorithms" with the candidate's capability "independently designing and optimizing neural network models, improving prediction accuracy by 35%" yields a semantic similarity score of 0.82. Similarly, scores for secondary and bonus dimensions are calculated to complete the bidirectional hierarchical mapping.

[0027] When retrieving target weight configurations through the real-time weight adjustment channel, features are extracted from the current recruitment process stage to construct a recruitment process state vector. The process stage type is identified by enumerating coded records: the initial screening stage is assigned a value of 1, the professional interview stage a value of 2, and the final interview stage a value of 3. The candidate pool size is represented by calculating the ratio of the current remaining candidates to the number of positions available; a ratio greater than 10 is marked as sufficient, between 3 and 10 as moderate, and less than 3 as scarce. The job urgency is calculated based on the time difference between the posting duration and the deadline; a time difference of less than 7 days is assigned a high urgency flag.

[0028] The weight template library stores various scenario weight configuration schemes. In the initial screening scenario configuration, the core dimension weight is set to 0.7, the secondary dimension weight to 0.25, and the bonus dimension weight to 0.05. In the final interview scenario configuration, the core dimension weight is adjusted to 0.5, the secondary dimension weight is increased to 0.35, and the bonus dimension weight is increased to 0.15. A conflict detection mechanism is used to filter out abnormal samples based on historical recruitment data—those with satisfactory matching scores but substandard performance after joining the company—marking them as risky configurations and removing them from the candidate pool. The remaining configurations are optimized through feature vector similarity calculation, selecting the configuration that best matches the current state as the target weight configuration.

[0029] The offline optimization channel collects historical recruitment results and performance data to calculate prediction deviations. Candidate matching prediction records and actual hiring results records are extracted from the historical recruitment database. The absolute value of the difference between the predicted pass probability and the actual hiring label is used as a deviation quantification indicator; values ​​exceeding 0.3 are marked as significant deviation events. Post-employment performance appraisal data is collected to construct a delayed correlation graph, identifying the moment when performance scores first fluctuate by more than 20%, and recording the lag period, typically ranging from 90 to 180 days.

[0030] Within the lag effect period, the weight configuration change sequence is extracted to form a weight allocation trajectory, and the performance data is simultaneously extracted to form a performance fluctuation curve by extracting the assessment score change sequence of each capability dimension. The Pearson correlation coefficient of the two curves is calculated, and a value less than -0.5 is considered an inverse synchronous relationship. A contribution decomposition tree is constructed to trace the transmission path, and the intersection nodes of weight superposition interference are identified. A weight decoupling factor is introduced to redistribute the weight parameters, and the weight correction amount is generated and fed back to the weight template library.

[0031] A comprehensive matching score is generated by calculating the scores of each dimension based on the scenario weight configuration. The core dimension score is compared with a preset core constraint threshold of 0.65. If the score is lower than the threshold, the calculation process is traced back, and a blocking flag is injected to stop the data flow. During coupling degree detection, the job competency tags corresponding to each dimension are extracted and their intersection is matched. The percentage of tag overlap is calculated as the coupling degree value, and a decay weight is applied to dimensions with a coupling degree exceeding 0.5. When the coupling degree is 0.7, the decay coefficient is 0.76; when the coupling degree is 0.55, the decay coefficient is 0.94. Constraint operations are performed through the constraint judgment condition set to force the comprehensive matching score to converge to the zero value range of 0 to 0.05, triggering the elimination judgment.

[0032] When the core dimension score is greater than or equal to the preset threshold, the score calculation for each dimension continues, and the scores are weighted and aggregated according to the weight configuration. Taking a weight configuration of 0.5 for the core dimension, 0.3 for the secondary dimension, and 0.2 for the bonus dimension as an example, the comprehensive matching score is calculated and the result is output for recruitment decision-making.

[0033] In one optional embodiment, the job requirement text is hierarchically divided based on the semantic features of requirement intensity to obtain a requirement hierarchy structure; action features and outcome features are extracted from the candidate's ability text, an action-outcome correlation matrix is ​​constructed, and a responsibility dominance index is calculated. Based on the responsibility dominance index and the quantitative value of the outcome, a capability hierarchy transition is determined, resulting in a capability hierarchy structure including:

[0034] Semantic parsing is performed on the job requirement text to extract requirement nodes and modification nodes, construct a requirement semantic graph, and calculate edge weights based on the syntactic features of modification nodes to determine the semantic features of requirement intensity. A requirement propagation attenuation factor is introduced to reduce the requirement intensity on multi-hop paths. Based on the adjusted semantic features of requirement intensity, the requirement nodes are divided into a requirement hierarchy structure.

[0035] Verb phrases are extracted from the candidate's ability text as action features, and quantitative expressions and achievement nouns are extracted as achievement features. A two-layer structure containing intra-sentence dependency distance matrix and cross-sentence semantic similarity matrix is ​​established to obtain the action-achievement association matrix. The dominant action pattern vector is extracted by performing singular value decomposition on the intra-sentence dependency distance matrix. The contribution rate of the dominant action pattern vector to the total variance is calculated to determine the responsibility dominance index.

[0036] The quantitative descriptions in the outcome characteristics are normalized to obtain the outcome quantification value. A capability level leap decision tree is constructed with the responsibility dominance index as the first branch node and the outcome quantification value as the second branch node. The output determines whether the capability description should leap to a higher capability level, forming a capability level structure, including the core necessary requirement level, the secondary adaptation requirement level, and the bonus requirement level.

[0037] In one specific implementation, after obtaining the job requirement text, a dependency parser is used to parse the text. The phrase "proficient in Python" is identified as the core verb and extracted as a requirement node, while "proficient" is used as a modifier node. The connection edges between the modifier node and the requirement node are labeled using syntactic dependency relations. The adverb strength level, verb voice, and qualifier type of the modifier node are extracted as syntactic features. When calculating the edge weights, mandatory modifiers like "must" are assigned a weight of 0.9, deliberate modifiers like "priority" are assigned a weight of 0.6, and deliberate modifiers like "understand" are assigned a weight of 0.3. The edge weights represent the numerical representation of the semantic features of the requirement strength.

[0038] When constructing the demand semantic graph, all demand nodes are treated as vertices in the graph, and edge connections are established by calculating the semantic similarity between nodes using a skill ontology knowledge base. When the demand nodes "Data Analysis" and "Machine Learning" in the job description have a three-hop path in the knowledge graph, a demand propagation attenuation factor of 0.7 is set. Therefore, the initial demand intensity of the "Data Analysis" node, 0.9, attenuates to 0.9 × 0.7 after three hops of propagation. 3 = 0.31. Based on the adjusted demand intensity values, demand nodes with values ​​greater than 0.7 are assigned to the core essential demand level, those with values ​​between 0.4 and 0.7 are assigned to the secondary adaptive demand level, and those with values ​​less than 0.4 are assigned to the bonus demand level, forming a three-tier demand hierarchy structure.

[0039] When processing candidate competency texts, the phrase "led the achievement of sales exceeding 5 million" is analyzed. Part-of-speech tagging is used to identify "led the achievement" as a verb phrase, extracting it as an action feature; "5 million" is a quantitative expression, and "sales exceeding" is an achievement noun, together constituting the achievement feature. When constructing the intra-sentence dependency distance matrix, the shortest path length between the action feature "led the achievement" and the achievement feature "sales exceeding" in the dependency tree is calculated, and this distance is recorded at the corresponding position in the matrix. For cross-sentence description scenarios, "organized team training" and the subsequent sentence "team performance improved by 20%" are extracted and semantically encoded, and a cosine similarity of 0.78 is calculated and filled into the cross-sentence semantic similarity matrix, forming a two-layer structure of the action-achievement association matrix.

[0040] Singular value decomposition (SVD) is performed on the intra-sentence dependency distance matrix, yielding a left singular matrix, a diagonal singular value matrix, and a right singular matrix. The first column of the left singular matrix is ​​extracted as the dominant action pattern vector, where the values ​​of each dimension reflect the weight distribution of different action types. The ratio of the square of the corresponding singular value to the sum of the squares of all singular values ​​is calculated to obtain the contribution rate of the dominant action pattern to the total variance. When the contribution rate reaches 0.65, this value is determined as the responsibility dominance index, indicating that the candidate exhibits a clear core responsibility pattern.

[0041] The quantitative descriptions in the performance characteristics, such as "sales revenue of 5 million" and "user growth of 30%", are processed. Absolute numerical results are converted into relative multiples using industry benchmark data, and percentage results are directly mapped to the 0-1 range, uniformly normalized to obtain quantified values. When constructing the capability level transition decision tree, the first branch node determines whether the responsibility dominance index is greater than 0.6. If it is, the second branch node determines whether the quantified value of the performance is greater than 0.5. When both conditions are met, a transition signal is output, elevating the capability description from the secondary adaptation level to the core essential level; when only the responsibility dominance condition is met, it remains at the secondary adaptation level; when neither condition is met, it is classified into the bonus requirement level. After traversing all capability descriptions of candidates to complete the transition determination, a capability hierarchy structure containing three levels is formed, ensuring that capability assessment and job requirements are corresponding in the hierarchical dimension.

[0042] In one alternative embodiment, introducing a demand propagation attenuation factor to adjust the demand intensity on multi-hop paths includes: In the demand semantic graph, a set of directed propagation paths from modifier nodes to demand nodes is constructed. For each propagation path, the intermediate node type sequence is extracted. The semantic decay mode is identified based on the intermediate node type sequence. The semantic decay mode includes the negation word blocking mode, the condition word weakening mode, and the conjunction word transmission mode. For the identified semantic decay patterns, a first decay factor is set for the negation word blocking pattern, a second decay factor is set for the condition word weakening pattern, and a third decay factor is set for the connector word propagation pattern. The first decay factor is less than the second decay factor, and the second decay factor is less than the third decay factor. The corresponding decay factors are applied to the edge weights on the propagation path. When there are multiple propagation paths for the same demand node, the propagation paths with different semantic decay modes are prioritized. The decay result of the path with the negative word blocking mode is used first to adjust the demand intensity. When there is no negative word blocking mode, the decay results of the condition word weakening mode and the connector word propagation mode are combined to obtain the adjusted demand intensity semantic features of the demand node.

[0043] In one specific implementation, when parsing job requirement text, there are often multi-hop connections between modifier nodes (such as adjectives, adverbs, and degree words) and requirement nodes (such as skill names and qualification requirements) in the requirement semantic graph. A set of directed propagation paths from modifier nodes to requirement nodes is constructed, and the sequence of intermediate nodes on the paths is extracted through dependency parsing. For example, in the path from the modifier "not needed" to the requirement word "overtime," if the intermediate node contains the negative word "no," this path is marked as a specific semantic decay pattern.

[0044] The identification of semantic decay patterns relies on matching rules for intermediate node types. The negation word blocking pattern detects the presence of negation words such as "not," "none," and "non-" in the path; when a negation word is detected, it indicates that the modifying relationship has a negative effect on the strength of the demand. The conditional word weakening pattern identifies conditional words such as "if," "preferential," and "best," which reduce the rigidity of the demand. The conjunction word propagation pattern corresponds to parallel conjunctions such as "and," "as well as," and "simultaneously," where the demand strength propagation effect is relatively stable.

[0045] Different attenuation factors are set for different semantic attenuation patterns. The first attenuation factor for the negation word blocking pattern is set to 0.1, indicating that the demand intensity is almost completely suppressed. The second attenuation factor for the conditional word weakening pattern is set to 0.5, reducing the demand intensity to half of its original value. The third attenuation factor for the connector word propagation pattern is set to 0.85, only slightly attenuating the demand intensity. These attenuation factors are applied to the edge weights on the propagation path, and the effective intensity reaching the demand node is calculated by hop-by-hop multiplication.

[0046] When multiple propagation paths of different types exist simultaneously for the same demand node, a priority ranking mechanism is used to resolve conflicts. It checks for paths with a negation word blocking pattern; if such a path exists, its attenuation result is directly adopted, ignoring the influence of other paths. This is because the semantic effect of negation words has absolute priority; for example, the negation word in "not requiring proficiency in Python" directly negates the skill requirement. When no negation word blocking pattern exists, the attenuation results of both the conditional word weakening pattern and the conjunction word propagation pattern are calculated comprehensively. Specifically, a weighted average method is used, calculating a weighting coefficient based on the original strength values ​​of each path to obtain the adjusted demand strength semantic feature of the demand node.

[0047] The adjusted semantic features of demand intensity are reassigned to the corresponding nodes in the demand semantic graph for subsequent hierarchical division operations. This mechanism effectively solves the problem of insufficient understanding of modifying semantics in traditional text matching, enabling accurate quantification of the rigidity and flexibility of job requirements. It avoids misjudging conditional requirements such as "priority consideration" as hard requirements, or incorrectly identifying negative expressions such as "not required" as positive requirements, thereby improving the accuracy of bidirectional hierarchical adaptation.

[0048] like Figure 2 The diagram shown illustrates the logical flowchart of demand semantic decay and path propagation.

[0049] In one optional embodiment, retrieving the target weight configuration from the weight template library based on the conflict detection results between the recruitment process state vector and the scenario weight configuration via a real-time weight adjustment channel includes: Extract process stage type identifiers, candidate pool size characteristics, and job urgency characteristics from the current recruitment process stage to construct a recruitment process state vector. Cross-validate the recruitment process state vector with multiple scenario weight configurations in the weight template library. For each scenario weight configuration, backtrack the weight application results of similar recruitment process states in history and extract abnormal samples and corresponding abnormal feature labels that cause matching failures in the weight application results. Construct a conflict detection matrix between the recruitment process state vector and the abnormal feature label. When the recruitment process state vector has a feature component that overlaps with the abnormal feature label, mark the corresponding scenario weight configuration with risk and remove it from the candidate range. Select the configuration with the highest matching degree with the recruitment process state vector feature from the scenario weight configurations that have not been marked with risk as the target weight configuration. The target weight configuration is retrieved from the weight template library through the real-time weight adjustment channel, thus completing the dynamic retrieval of the scene weight configuration.

[0050] In one specific implementation, constructing the recruitment process state vector requires capturing the dynamic characteristics of the current recruitment environment from multiple dimensions. The process stage type identifier records the current stage through enumeration encoding; for example, the initial screening stage is assigned a value of 1, the professional interview stage a value of 2, and the final interview stage a value of 3. The candidate pool size characteristic is represented by calculating the ratio of the remaining candidates to the number of positions available at the current stage. This ratio reflects the intensity of competition; a ratio greater than 10 is marked as sufficient, a ratio between 3 and 10 as moderate, and a ratio less than 3 as scarce. The job urgency characteristic is calculated based on the time difference between the job posting duration and the recruitment deadline; a time difference of less than 7 days is assigned a high urgency identifier, 7 to 30 days a medium urgency identifier, and more than 30 days a low urgency identifier. These three types of characteristics combine to form a three-dimensional recruitment process state vector.

[0051] The weight template library pre-stores various scenario weight configuration schemes, each containing specific numerical allocations for core dimension weights, secondary dimension weights, and bonus dimension weights. For example, in the initial screening scenario configuration, the core dimension weight is set to 0.7, the secondary dimension weight is 0.25, and the bonus dimension weight is 0.05, emphasizing the screening of basic abilities; while in the final interview scenario configuration, the core dimension weight is adjusted to 0.5, the secondary dimension weight is increased to 0.35, and the bonus dimension weight is increased to 0.15, focusing more on comprehensive quality assessment.

[0052] The conflict detection mechanism relies on in-depth mining of historical data. For each scenario's weight configuration, the system backtracks all recruitment cases that applied that configuration within the past six months, filtering out anomalous samples where the matching score was qualified but the performance evaluation after hiring failed. Feature tags that led to the failure are extracted from these anomalous samples. For example, if a technical position uses a lenient weight configuration in a high-urgency scenario, and it is found that neglecting in-depth verification of core skills leads to insufficient actual ability of the hired personnel, then the combination of "high urgency + technical position + lenient configuration" is marked as an anomalous feature tag.

[0053] The conflict detection matrix uses a two-dimensional Boolean array structure. The row index corresponds to each feature component of the recruitment process state vector, and the column index corresponds to the number of the scenario weight configuration. A value of 1 in the matrix cell indicates that there is a historical conflict record between the state feature and the weight configuration, while a value of 0 indicates no conflict. During conflict detection, each feature component of the current recruitment process state vector is compared with one of the matrix rows. When the process stage in the state vector is 3, the candidate pool size is in a shortage state, and the job urgency is high, the matrix search finds that this feature combination conflicts with the relaxed weight configuration numbered 5. Configuration numbered 5 is immediately marked as risk and removed from the candidate range.

[0054] The remaining unlabeled scenario weight configurations are optimized through feature matching degree calculation. The matching degree calculation adopts the feature vector similarity measurement method, which calculates the cosine similarity between the recruitment process state vector and the applicable scenario feature vector of each candidate configuration. For example, if the current state vector is [3, scarce, high], the applicable scenario vector of candidate configuration A is [3, moderate, high], and the applicable scenario vector of configuration B is [3, scarce, high]. Through dimensional comparison, it is found that configuration B is a perfect match in all three dimensions, with a similarity of 1.0, while configuration A has a difference in the candidate pool size dimension, and the similarity drops to 0.85. Therefore, configuration B is selected as the target weight configuration.

[0055] The real-time weight adjustment channel reads the specific parameter values ​​of the target weight configuration from the weight template library via an API interface, including the weight coefficients for each dimension, core constraint thresholds, and decay weight function parameters. A version verification mechanism is added during the retrieval process to ensure that the weight configuration read is the latest version corrected by the offline optimization channel, avoiding matching errors caused by using outdated parameters.

[0056] In one optional embodiment, the process of collecting historical recruitment results data and performance data through an offline optimization channel, calculating the prediction deviation, and generating a weight correction amount to be sent back to the weight template library includes: Extract matching prediction records and actual hiring records from historical recruitment results data, calculate prediction deviation, and construct a delayed correlation graph by linking the prediction deviation with the candidate's performance data after joining the company according to timestamps. In the delayed correlation graph, count the time interval between the occurrence of prediction deviation and the performance fluctuation, identify the lag effect period, extract the weight value sequence of weight allocation for each job competency dimension within the lag effect period to form a weight allocation trajectory, extract the score change sequence of each job competency dimension in the performance data to form a performance fluctuation curve, and compare the change trend of the weight allocation trajectory and the performance fluctuation curve to verify synchronization. When the weight allocation trajectory and the performance fluctuation curve show an inverse synchronous relationship, a contribution decomposition tree is constructed for the weight parameters of each dimension in the weight configuration. The transmission path of the weight parameters of each dimension to the prediction deviation is tracked in the contribution decomposition tree. The intersection nodes of weight superposition interference in the transmission path are identified. A weight decoupling factor is introduced for the intersection nodes, and the weight parameter proportion at the intersection nodes is redistributed to generate the weight correction amount. The weight correction amount and weight decoupling factor are sent back to the weight template library through the offline optimization channel to update the weight configuration.

[0057] In one specific implementation, the offline optimization channel continuously adjusts the configuration parameters in the weight template library by establishing a correlation verification mechanism between prediction results and performance. Matching prediction records for candidates are extracted from the historical recruitment database, including scores for each dimension, a comprehensive matching score, and the predicted pass probability. Simultaneously, actual hiring results records are extracted, noting the final hiring decision and the basis for the hiring decision. When calculating prediction deviation, the absolute value of the difference between the predicted pass probability and the actual hiring label is used as a deviation quantification indicator; when this value exceeds 0.3, it is marked as a significant deviation event.

[0058] For hired personnel, performance appraisal data was collected from the human resources management system for 3, 6, and 12 months after onboarding, including indicators such as job-specific competency scores, teamwork scores, and learning and growth rate. The prediction deviation data and performance data were timestamped according to the candidate's unique identifier to construct a delayed correlation graph. This graph starts at the moment the prediction deviation occurred and ends at the moment of performance fluctuation, unfolding the correlation along a time axis. When calculating time intervals, the moment when the performance score first fluctuated by more than 20% was identified, and the number of days between this moment and the prediction moment was recorded. The time window in which the fluctuations occurred was defined as the lag effect period, with a typical period range of 90 to 180 days.

[0059] Within the lag period, historical versions of weight configurations for the corresponding job type are extracted, and the changes in the weight values ​​of core, secondary, and bonus dimensions over time are recorded. When the weight of a certain dimension undergoes more than three adjustments within the period and the adjustment magnitude exceeds 0.05, that dimension is marked as an active adjustment dimension, forming a weight allocation trajectory curve. Simultaneously, the time-series changes in assessment scores for each competency dimension are extracted from performance data, and a performance fluctuation curve is generated after smoothing with a moving average.

[0060] When performing synchronization verification, the Pearson correlation coefficient of the two curves within the lag period is calculated. When the correlation coefficient is less than -0.5, it is determined to be a reverse synchronization relationship, indicating that the performance of the dimension corresponding to the weight allocation improvement has decreased, suggesting a bias in weight allocation. At this time, a contribution decomposition tree is constructed, with the root node representing the total prediction bias, and the first-level child nodes corresponding to the bias contribution of each capability dimension, calculated by multiplying the bias amount by the weight of that dimension.

[0061] When tracing the transmission path, scenarios where multiple dimensions jointly affect the same prediction result are identified, and the nodes where these dimensions intersect are marked as intersection nodes. When the sum of the weights of the core and secondary dimensions at an intersection node exceeds 0.85, it is determined that there is weight superposition interference. A weight decoupling factor α is introduced, calculated as the reciprocal of the variance of the weights of each dimension at the intersection node, with a value ranging from 0.1 to 0.3. The weight proportions are redistributed according to the decoupling factor. The original weight values ​​are multiplied by (1-α) and then redistributed according to the proportion of dimension contribution to ensure that the total weight sum after adjustment is still 1.0.

[0062] When generating weight corrections, for dimensions undergoing reverse synchronization, the correction is set as a negative adjustment to the current weight value, with an adjustment magnitude of β × the absolute value of the correlation coefficient, where β is the correction strength coefficient, defaulting to 0.15. For dimensions undergoing forward synchronization, the correction is a positive enhancement, with an magnitude of 5% to 10% of the original weight. The correction and decoupling factors are encapsulated into update instructions, written to the weight template library via an offline optimization channel, replacing the corresponding weight configuration version, and the update timestamp and triggering reason are recorded for subsequent auditing and traceability.

[0063] In one optional embodiment, when the core dimension score is lower than a preset core constraint threshold, a decay weight is applied to the scores of other dimensions to converge the comprehensive matching score to the zero range and trigger an elimination decision; otherwise, the comprehensive matching score is output, including: The core dimension score is compared with the preset core constraint threshold. When the core dimension score is lower than the preset core constraint threshold, the calculation steps of other dimensions that have been executed are traced back. The related calculation nodes that depend on the core dimension score in the calculation step chain are identified. Blocking flags are injected into the related calculation nodes to stop the data flow of the calculation steps. The coupling degree of the other dimension scores that have been calculated is detected with the core dimension score. Based on the coupling degree detection results, the attenuation weight is applied to the other dimension scores to generate invalid score data. Construct a set of constraint judgment conditions for the comprehensive matching score, and use the judgment result that the core dimension score is lower than the preset core constraint threshold as the trigger condition to inject into the set of constraint judgment conditions. Perform constraint operation on the failure score data and the core dimension score based on the set of constraint judgment conditions. Force the comprehensive matching score to converge to the zero value interval through constraint operation. Generate and output the elimination judgment result based on the comprehensive matching score in the zero value interval. When the score of the core dimension is greater than or equal to the preset core constraint threshold, the score calculation process of other dimensions continues, and the scores of each dimension are summarized to form a comprehensive matching score and output.

[0064] In one specific implementation, constraint checks need to be performed on the core dimension scores during the person-job matching calculation process. Specifically, a core constraint threshold of 0.65 is set, representing the minimum standard that candidates must meet in core dimensions such as key skills, educational requirements, and years of work experience. Once the core dimension scores are obtained, a numerical comparison operation is immediately performed to determine whether the score is below 0.65. If the comparison result is true, a failure handling process is initiated.

[0065] When initiating the failure handling process, it is necessary to trace the executed calculation steps. This path records the calculation process of secondary dimension scores and bonus dimension scores, including each intermediate calculation node and its input-output relationships. Through dependency analysis, related calculation nodes in the path that depend on the core dimension score are identified. For example, the calculation of some bonus items may use the core dimension score as a base value for multiplication; such nodes are related calculation nodes. A blocking flag is injected into these nodes, specifically by writing the string "BLOCKED" to the node's status flag, stopping subsequent data flow and preventing invalid data from continuing to propagate.

[0066] For the calculated secondary and bonus dimension scores, a coupling degree test is performed. The coupling degree test calculates the correlation coefficient between each dimension score and the core dimension score; a higher correlation coefficient indicates stronger coupling. Differential attenuation weights are applied for different coupling degree levels: an attenuation coefficient of 0.2 is applied when the correlation coefficient is greater than 0.7; an attenuation coefficient of 0.5 is applied when the correlation coefficient is between 0.4 and 0.7; and an attenuation coefficient of 0.8 is applied when the correlation coefficient is less than 0.4. Each dimension score is multiplied by its corresponding attenuation coefficient to generate failure score data.

[0067] When constructing the constraint judgment condition set, this set contains multiple judgment rules. The judgment result of a core dimension score below 0.65 is injected into the set as the primary trigger condition. Auxiliary constraints are also added, such as setting the weight coefficients of secondary dimension scores and bonus dimension scores to 0. Constraint operations are performed on the failed score data based on the constraint judgment condition set. The constraint operation uses a weighted summation method, but because the weight coefficients are set to zero, the contribution value of the failed score data is zero. Only the core dimension score is retained in the calculation, and a penalty factor of 0.1 is applied to it, forcing the comprehensive matching score to converge to the zero value range, specifically 0 to 0.05. Based on this converged comprehensive matching score, an elimination judgment result is generated, indicating the elimination reason as "core dimension not meeting the standard," and a detailed dimension score breakdown is output for subsequent analysis.

[0068] When the core dimension score is greater than or equal to 0.65, the scoring process for other dimensions continues. Secondary dimension scores and bonus dimension scores are calculated sequentially. After each dimension is calculated, a weighted sum is performed according to the preset weight configuration. Assuming the core dimension weight is 0.5, the secondary dimension weight is 0.3, and the bonus dimension weight is 0.2, the overall matching score is equal to the sum of the products of each dimension score and its corresponding weight. After calculation, the overall matching score is directly output, along with detailed data for each dimension score, for further processing by the recruitment decision system.

[0069] In one optional embodiment, coupling degree detection is performed between the calculated scores of other dimensions and the core dimension scores. Based on the coupling degree detection results, a decay weight is applied to the scores of other dimensions to generate failure score data, including: Extract the job competency tags for other dimensions corresponding to the calculated scores of other dimensions, extract the job competency tags for core dimensions corresponding to the scores of core dimensions, perform tag intersection matching between job competency tags for other dimensions and job competency tags for core dimensions, count the proportion of overlapping tags in the tag intersection matching, calculate the dependence strength of other dimension scores on core dimension scores based on the proportion of overlapping tags, and use the dependence strength as the coupling degree value. The scores of other dimensions are grouped into layers according to the coupling degree. For the scores of other dimensions whose coupling degree exceeds the preset coupling threshold, the attenuation weight coefficient is calculated. The attenuation weight coefficient is set to have an inverse relationship with the coupling degree. The scores of other dimensions are weighted and calculated with the corresponding attenuation weight coefficient to generate failure score data.

[0070] In one specific implementation, when the core dimension score falls below the core constraint threshold, a coupling degree detection mechanism is activated to invalidate the scores of other already calculated dimensions. The set of job competency tags associated with the secondary and bonus dimension scores is extracted from the database, along with the set of core job competency tags corresponding to the core dimension score. Job competency tags include technical skill tags, business domain tags, and work scenario tags. For example, secondary dimensions might include tags such as "data analysis" and "project coordination," while core dimensions might include tags such as "Python programming" and "machine learning algorithms."

[0071] A tag intersection matching operation is performed, identifying overlapping tags through a combination of exact string matching and semantic similarity calculation. When calculating the percentage of overlapping tags, the number of intersection elements between the tag sets of other dimensions and the core dimension tag set is counted, and this intersection number is divided by the total number of tags in other dimensions to obtain the percentage. Based on this percentage, the dependency strength is calculated, reflecting the subordinate relationship between capabilities in other dimensions and capabilities in the core dimension. When the percentage is 0.6, it indicates that 60% of the capability tags in that dimension depend on the existence of the core capability; 0.6 is directly used as the coupling degree value.

[0072] All other dimension scores are stratified and grouped according to their coupling degree, with a coupling threshold of 0.5. When the coupling degree exceeds 0.5, it is determined that the score of that dimension has a strong dependency on the core dimension, and a decay weight needs to be applied. The decay weight coefficient is calculated using an inverse mapping function, specifically: the decay weight coefficient equals 1 minus the difference between the coupling degree and the threshold, multiplied by the decay factor. The decay factor is set to 1.2 to ensure that high coupling corresponds to a stronger decay effect.

[0073] The failure score data is generated by weighting the scores of other dimensions with their corresponding attenuation weight coefficients. A secondary dimension score of 85 corresponds to a coupling degree of 0.7, resulting in a failure score of 85 × 0.76 = 64.6. A bonus dimension score of 20 corresponds to a coupling degree of 0.55, resulting in a failure score of 20 × (1 - (0.55 - 0.5) × 1.2) = 18.8. For dimension scores with coupling degrees below the threshold, the original values ​​remain unchanged. All failure score data is then aggregated and entered into the comprehensive matching score calculation formula. Since the core dimension score is already below the constraint threshold, the final comprehensive score will converge to the zero range, triggering an elimination decision. This mechanism ensures that when a candidate's core competency is insufficient, other auxiliary skills that rely on core competencies cannot compensate for the deficiency, avoiding the risk of misjudgment due to high scores in secondary dimensions masking core deficiencies. After the failure score data is generated, it is passed to the result output module to complete the termination decision of the matching process.

[0074] The present invention provides an intelligent human-job matching system based on bidirectional hierarchical adaptation and dual closed-loop weight adjustment, comprising: The text acquisition unit is used to acquire the job requirements text and candidate competency text to be matched; The demand and capability assessment unit is used to hierarchically divide the job demand text based on the semantic features of demand intensity to obtain the demand hierarchy structure; extract action features and result features from the candidate capability text, construct an action-result association matrix and calculate the responsibility dominance index; and determine the capability level jump based on the responsibility dominance index and the quantitative value of the result to obtain the capability hierarchy structure. The structural mapping unit is used to perform bidirectional adaptation mapping between the demand hierarchy and the capability hierarchy to obtain core dimension scores, secondary dimension scores and bonus dimension scores. The comprehensive calculation unit is used to retrieve the target weight configuration from the weight template library through the real-time weight adjustment channel based on the conflict detection results of the recruitment process state vector and the scenario weight configuration; through the offline optimization channel, it collects historical recruitment result data and performance data to calculate the prediction deviation and generates weight correction amount to be sent back to the weight template library; and calculates the comprehensive matching score based on the scenario weight configuration for the core dimension score, secondary dimension score and bonus dimension score. The elimination decision unit is used to apply a decay weight to the scores of other dimensions when the core dimension score is lower than the preset core constraint threshold, so as to converge the comprehensive matching score to the zero value range and trigger the elimination decision; otherwise, the comprehensive matching score is output.

[0075] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0076] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0077] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A human-job intelligent matching method based on bidirectional hierarchical adaptation and dual closed-loop weight adjustment, characterized in that, include: Obtain the job requirements text and candidate competency text to be matched; The job requirement text is hierarchically divided based on the semantic features of requirement intensity to obtain the requirement hierarchy structure; Action features and outcome features are extracted from the candidate's ability text, an action-outcome correlation matrix is ​​constructed and a responsibility dominance index is calculated. Based on the responsibility dominance index and the quantitative value of the outcome, the ability level leap is determined, and the ability level structure is obtained. The requirement hierarchy structure and the capability hierarchy structure are bidirectionally adapted and mapped to obtain the core dimension score, secondary dimension score and bonus dimension score. Through the real-time weight adjustment channel, the target weight configuration is retrieved from the weight template library based on the conflict detection results of the recruitment process state vector and the scenario weight configuration; through the offline optimization channel, historical recruitment result data and performance data are collected to calculate the prediction deviation, and the weight correction amount is generated and sent back to the weight template library. A comprehensive matching score is generated by calculating the scores of the core dimension, secondary dimension, and bonus dimension based on the scenario weight configuration. When the core dimension score is lower than the preset core constraint threshold, a decay weight is applied to the scores of other dimensions to bring the comprehensive matching score to the zero range and trigger the elimination decision; otherwise, the comprehensive matching score is output. The step of retrieving the target weight configuration from the weight template library through a real-time weight adjustment channel based on the conflict detection results between the recruitment process state vector and the scenario weight configuration includes: Extract process stage type identifiers, candidate pool size characteristics, and job urgency characteristics from the current recruitment process stage to construct a recruitment process state vector. Cross-validate the recruitment process state vector with multiple scenario weight configurations in the weight template library. For each scenario weight configuration, backtrack the weight application results of similar recruitment process states in history and extract abnormal samples and corresponding abnormal feature labels that cause matching failures in the weight application results. Construct a conflict detection matrix between the recruitment process state vector and the abnormal feature label. When the recruitment process state vector has a feature component that overlaps with the abnormal feature label, mark the corresponding scenario weight configuration with risk and remove it from the candidate range. Select the configuration with the highest matching degree with the recruitment process state vector feature from the scenario weight configurations that have not been marked with risk as the target weight configuration. The target weight configuration is retrieved from the weight template library through the real-time weight adjustment channel to complete the dynamic retrieval of the scene weight configuration; The step of collecting historical recruitment results data and performance data through an offline optimization channel, calculating the prediction deviation, and generating a weight correction amount to be sent back to the weight template library includes: Extract matching prediction records and actual hiring records from historical recruitment results data, calculate prediction deviation, and construct a delayed correlation graph by linking the prediction deviation with the candidate's performance data after joining the company according to timestamps. In the delayed correlation graph, count the time interval between the time when the prediction deviation occurs and the time when the performance fluctuates, identify the lag effect period, extract the weight value sequence of the weight allocation for each job competency dimension within the lag effect period to form a weight allocation trajectory, extract the sequence of changes in the assessment scores of each job competency dimension in the performance data to form a performance fluctuation curve, and compare the changing trends of the weight allocation trajectory and the performance fluctuation curve to verify synchronization. When the weight allocation trajectory and the performance fluctuation curve show an inverse synchronous relationship, a contribution decomposition tree is constructed for the weight parameters of each dimension in the weight configuration. The transmission path of the weight parameters of each dimension to the prediction deviation is tracked in the contribution decomposition tree. The intersection nodes of weight superposition interference in the transmission path are identified. A weight decoupling factor is introduced for the intersection nodes, and the weight parameter proportion at the intersection nodes is redistributed to generate the weight correction amount. The weight correction amount and weight decoupling factor are sent back to the weight template library through the offline optimization channel to update the weight configuration.

2. The method according to claim 1, characterized in that, The job requirement text is hierarchically divided based on the semantic features of requirement intensity to obtain a requirement hierarchy structure; action features and outcome features are extracted from the candidate's ability text, an action-outcome correlation matrix is ​​constructed, and a responsibility dominance index is calculated. Based on the responsibility dominance index and the quantitative value of the outcome, ability hierarchy transitions are determined, resulting in an ability hierarchy structure including: Semantic parsing is performed on the job requirement text to extract requirement nodes and modification nodes, construct a requirement semantic graph, and calculate edge weights based on the syntactic features of modification nodes to determine the semantic features of requirement intensity. A requirement propagation attenuation factor is introduced to reduce the requirement intensity on multi-hop paths. Based on the adjusted semantic features of requirement intensity, the requirement nodes are divided into a requirement hierarchy structure. Verb phrases are extracted from the candidate's ability text as action features, and quantitative expressions and achievement nouns are extracted as achievement features. A two-layer structure containing intra-sentence dependency distance matrix and cross-sentence semantic similarity matrix is ​​established to obtain the action-achievement association matrix. The dominant action pattern vector is extracted by performing singular value decomposition on the intra-sentence dependency distance matrix. The contribution rate of the dominant action pattern vector to the total variance is calculated to determine the responsibility dominance index. The quantitative descriptions in the outcome characteristics are normalized to obtain the outcome quantification value. A capability level leap decision tree is constructed with the responsibility dominance index as the first branch node and the outcome quantification value as the second branch node. The output determines whether the capability description should leap to a higher capability level, forming a capability level structure, including the core necessary requirement level, the secondary adaptation requirement level, and the bonus requirement level.

3. The method according to claim 1, characterized in that, Introducing a demand propagation attenuation factor to reduce the demand intensity on multi-hop paths includes: In the demand semantic graph, a set of directed propagation paths from modifier nodes to demand nodes is constructed. For each propagation path, the intermediate node type sequence is extracted. The semantic decay mode is identified based on the intermediate node type sequence. The semantic decay mode includes the negation word blocking mode, the condition word weakening mode, and the conjunction word transmission mode. For the identified semantic decay patterns, a first decay factor is set for the negation word blocking pattern, a second decay factor is set for the condition word weakening pattern, and a third decay factor is set for the connector word propagation pattern. The first decay factor is less than the second decay factor, and the second decay factor is less than the third decay factor. The corresponding decay factors are applied to the edge weights on the propagation path. When there are multiple propagation paths for the same demand node, the propagation paths with different semantic decay modes are prioritized. The decay result of the path with the negative word blocking mode is used first to adjust the demand intensity. When there is no negative word blocking mode, the decay results of the condition word weakening mode and the connector word propagation mode are combined to obtain the adjusted demand intensity semantic features of the demand node.

4. The method according to claim 1, characterized in that, When the core dimension score is lower than the preset core constraint threshold, a decay weight is applied to the scores of other dimensions to converge the comprehensive matching score to the zero range and trigger an elimination decision; otherwise, the comprehensive matching score is output, including: The core dimension score is compared with the preset core constraint threshold. When the core dimension score is lower than the preset core constraint threshold, the calculation steps of other dimensions that have been executed are traced back. The related calculation nodes that depend on the core dimension score in the calculation step chain are identified. Blocking flags are injected into the related calculation nodes to stop the data flow of the calculation steps. The coupling degree of the other dimension scores that have been calculated is detected with the core dimension score. Based on the coupling degree detection results, the attenuation weight is applied to the other dimension scores to generate invalid score data. Construct a set of constraint judgment conditions for the comprehensive matching score, and use the judgment result that the core dimension score is lower than the preset core constraint threshold as the trigger condition to inject into the set of constraint judgment conditions. Perform constraint operation on the failure score data and the core dimension score based on the set of constraint judgment conditions. Force the comprehensive matching score to converge to the zero value interval through constraint operation. Generate and output the elimination judgment result based on the comprehensive matching score in the zero value interval. When the score of the core dimension is greater than or equal to the preset core constraint threshold, the score calculation process of other dimensions continues, and the scores of each dimension are summarized to form a comprehensive matching score and output.

5. The method according to claim 4, characterized in that, The calculated scores for other dimensions are coupled with the core dimension scores. Based on the coupling detection results, attenuation weights are applied to the scores for other dimensions to generate failure score data, including: Extract the job competency tags for other dimensions corresponding to the calculated scores of other dimensions, extract the job competency tags for core dimensions corresponding to the scores of core dimensions, perform tag intersection matching between job competency tags for other dimensions and job competency tags for core dimensions, count the proportion of overlapping tags in the tag intersection matching, calculate the dependence strength of other dimension scores on core dimension scores based on the proportion of overlapping tags, and use the dependence strength as the coupling degree value. The scores of other dimensions are grouped into layers according to the coupling degree. For the scores of other dimensions whose coupling degree exceeds the preset coupling threshold, the attenuation weight coefficient is calculated. The attenuation weight coefficient is set to have an inverse relationship with the coupling degree. The scores of other dimensions are weighted and calculated with the corresponding attenuation weight coefficient to generate failure score data.

6. A human-job intelligent matching system based on bidirectional hierarchical adaptation and dual closed-loop weight adjustment, used to implement the method of any one of claims 1-5, characterized in that, include: The text acquisition unit is used to acquire the job requirements text and candidate ability text to be matched; The demand and capability assessment unit is used to hierarchically divide the job demand text based on the semantic features of demand intensity to obtain the demand hierarchy structure; extract action features and result features from the candidate capability text, construct an action-result association matrix and calculate the responsibility dominance index; and determine the capability level jump based on the responsibility dominance index and the quantitative value of the result to obtain the capability hierarchy structure. The structural mapping unit is used to perform bidirectional adaptation mapping between the demand hierarchy and the capability hierarchy to obtain core dimension scores, secondary dimension scores, and bonus dimension scores. The integrated calculation unit is used to retrieve the target weight configuration from the weight template library through the real-time weight adjustment channel based on the conflict detection results of the recruitment process state vector and the scenario weight configuration; and to collect historical recruitment result data and performance data through the offline optimization channel to calculate the prediction deviation and generate weight correction amount to be sent back to the weight template library. A comprehensive matching score is generated by calculating the scores of the core dimension, secondary dimension, and bonus dimension based on the scenario weight configuration. The elimination decision unit is used to apply a decay weight to the scores of other dimensions when the core dimension score is lower than the preset core constraint threshold, so as to converge the comprehensive matching score to the zero value range and trigger the elimination decision; otherwise, the comprehensive matching score is output.

7. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 5.

Images