Intelligent post-job matching method based on post demand analysis

Abstract

This application relates to the fields of artificial intelligence and human resource management technology, specifically disclosing an intelligent job matching method based on job requirement analysis. It constructs a job requirement tree by semantically parsing and hierarchically extracting job requirement data using a pre-trained language model, and introduces a cross-branch collaborative coupling quantification mechanism to enhance the edge weights of skill nodes with functional interdependencies. Simultaneously, it extracts skill entities from candidate resumes, assigns dynamic weights reflecting current proficiency to each node based on temporal decay and reuse transition functions, and constructs a dynamic skill evolution tree. Subsequently, it encodes the node features of the two trees using a graph attention network, constructs an asymmetric cost matrix using cross-granularity structure penalty distance, performs hierarchical optimal transmission solution, and integrates the bulldozer distance scalar and topological alignment feature vector to obtain a matching degree score. Finally, it generates an interpretable diagnostic report based on the transmission path matrix, ensuring the accuracy and interpretability of matching composite jobs.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and human resource management technology, and more specifically, to an intelligent person-job matching method based on job requirement analysis. Background Technology

[0002] As enterprises deepen their digital transformation, the complexity of job requirements and the diversity of candidate resumes continue to rise. Traditional recruitment models that rely on keyword matching or manual screening are no longer able to meet the needs of efficient and accurate talent allocation. There is an urgent need to build an intelligent job-person matching method based on job requirement analysis in order to achieve a deep-level structured alignment between job requirement semantics and candidate ability characteristics.

[0003] However, existing intelligent job matching technologies generally employ dual-tower vector retrieval or shallow graph neural network matching architectures. Their core limitation lies in their lack of effective modeling capabilities for the inherent cross-granularity semantic asymmetry between job requirements and candidate resumes. Job descriptions are typically presented as highly generalized macro-level competencies, while candidate resumes consist of fragmented, specific project experiences. A natural gap exists between the two in terms of semantic granularity. Traditional methods forcibly map both ends to the same vector space for cosine similarity calculations, easily leading to semantic representation collapse and degraded matching accuracy. Simultaneously, existing methods model candidate skills as static snapshots, ignoring the ability degradation effect caused by long-term disuse and the proficiency leap resulting from project reuse. This causes the matching results to fail to reflect the candidate's current true competency profile. Furthermore, in the context of composite positions, there is a common cross-branch collaborative coupling relationship between the underlying technology stacks belonging to different macro-level capability branches. That is, two skills must appear together and reinforce each other to effectively support the core responsibilities of the position. However, the existing tree structure construction method only models the vertical parent-child relationship. The technology stack nodes under different branches are in a completely orthogonal isolated state, and the horizontal collaborative semantics are completely lost. This results in an artificially high matching score for candidates who only excel in a single branch skill but lack cross-branch collaborative ability, which seriously affects the accuracy and interpretability of the person-job matching for composite positions.

[0004] Therefore, an optimized intelligent method for matching people to jobs based on job requirements analysis is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide an intelligent person-job matching method based on job requirement analysis, comprising: S1, through a pre-trained language model, performs semantic parsing and hierarchical knowledge extraction on the acquired job requirement data to obtain a job requirement tree; S2, extract skill entities and dynamically measure proficiency from the acquired candidate resume data to obtain a dynamic skill evolution tree; S3, feature encoding and cross-granularity structure penalty distance quantization are performed on all nodes in the job requirement tree and dynamic skill evolution tree to obtain the asymmetric cost matrix; S4 uses the asymmetric cost matrix as the cost function constraint to perform hierarchical optimal transmission solution and pooling dimensionality reduction on the node probability distribution of the job requirement tree and the dynamic skill evolution tree to obtain the bulldozer distance scalar, transmission path matrix and topology alignment feature vector. S5, perform a comprehensive matching degree estimation on the bulldozer distance scalar and topological alignment feature vector to obtain the human-position matching degree score; S6. Perform node mapping and source tracing analysis on the transmission path matrix and the person-job matching score to obtain an interpretable matching diagnostic report.

[0006] Compared with existing technologies, this application proposes an intelligent job matching method based on job requirement analysis. First, it uses a pre-trained language model to perform semantic parsing and hierarchical knowledge extraction on job requirement data, constructing a job requirement tree from macro-level capabilities to the underlying specific technology stack. During tree construction, a cross-branch collaborative coupling strength quantification mechanism is introduced, enabling cross-branch skill nodes with functional interdependencies to receive synergistically enhanced edge weights, thereby breaking the orthogonal isolation between different capability branches. Simultaneously, skill entities are extracted from candidate resume data, and dynamic weights reflecting the current true proficiency are assigned to each skill node based on a temporal decay function and a reuse transition activation mechanism, constructing a dynamic skill evolution tree. Then, the method is further refined. Then, graph attention network encoding is performed on all nodes of the two trees, and an asymmetric cost matrix incorporating hierarchical structure penalties is calculated. Using this as a cost function constraint, the bulldozer distance scalar and transmission path matrix are obtained through entropy regularization-based optimal transmission iteration. This achieves topological-level structured alignment across granular semantics rather than simple point-to-point vector matching. Finally, node mapping and source tracing are performed based on the transmission path matrix to generate an interpretable matching diagnostic report with fine-grained attribution evidence. This eliminates the semantic granularity gap, captures the temporal evolution of skills and cross-branch collaborative constraints, and provides transparent and reliable matching basis for business decisions. Attached Figure Description

[0007] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0008] Figure 1This is a flowchart of an intelligent person-job matching method based on job requirement analysis according to an embodiment of this application; Figure 2 This is a data flow diagram illustrating an intelligent human-job matching method based on job requirement analysis according to an embodiment of this application; Figure 3 The flowchart illustrates an intelligent job matching method based on job requirement analysis according to an embodiment of this application, which uses a pre-trained language model to perform semantic parsing and hierarchical knowledge extraction on the acquired job requirement data to obtain a job requirement tree. Figure 4 To illustrate the process of obtaining a job requirement tree based on a business scenario identification method for an intelligent person-job matching method based on job requirement analysis according to an embodiment of this application, a flowchart is provided that performs topological mapping and mask pruning on the macro-capability dimension node set and the underlying specific technology stack node set to obtain a job requirement tree. Figure 5 This document presents a flowchart illustrating the process of extracting skill entities and dynamically quantifying proficiency from acquired candidate resume data to obtain a dynamic skill evolution tree, based on an intelligent person-job matching method according to job requirement analysis, as described in an embodiment of this application. Figure 6 This is a flowchart illustrating the process of performing feature encoding and cross-granularity structure penalty distance quantization on all nodes in a job requirement tree and a dynamic skill evolution tree to obtain an asymmetric cost matrix, according to an embodiment of this application of an intelligent human-job matching method based on job requirement analysis. Figure 7 This document describes a flowchart illustrating an intelligent human-job matching method based on job requirement analysis, which uses an asymmetric cost matrix as a cost function constraint. The flowchart describes a process for performing hierarchical optimal transmission solution and pooling dimensionality reduction on the node probability distributions of the job requirement tree and the dynamic skill evolution tree to obtain the bulldozer distance scalar, transmission path matrix, and topology alignment feature vector. Detailed Implementation

[0009] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0010] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0011] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0012] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0013] Existing intelligent job matching technologies generally employ dual-tower vector retrieval or shallow graph network architectures, forcibly mapping job requirements and candidate resumes to the same vector space to calculate similarity. This fails to effectively address the semantic gap between the macro-level generality of job descriptions and the fine-grained fragmentation of project experience in resumes. Furthermore, it treats candidate skills as static snapshots, ignoring the decline and reuse of skill proficiency over time. In complex job scenarios, it fails to capture the synergistic coupling relationships between different skill branches, resulting in insufficient matching accuracy and interpretability. Therefore, this application proposes an intelligent job matching method based on job requirement analysis. This method first constructs a job requirement tree by semantically parsing and extracting hierarchical knowledge from job requirement data using a pre-trained language model. During tree assembly, a cross-branch collaborative coupling strength quantification mechanism is introduced to enhance the edge weights of skill nodes with functional interdependencies, breaking down orthogonal isolation between different ability branches. Simultaneously, skill entities are extracted from candidate resumes, and dynamic weights reflecting current proficiency are assigned to each node based on temporal decay and reuse transition functions, constructing a dynamic skill evolution tree. Subsequently, a graph attention network is used to encode features of all nodes in both trees. An asymmetric cost matrix is ​​constructed using cross-granularity structure penalty distance, and this matrix serves as a cost function constraint for hierarchical optimal transmission. The bulldozer distance scalar and topological alignment feature vector are fused to calculate the person-job matching score. Finally, node mapping and source tracing are performed based on the transmission path matrix to generate an interpretable matching diagnostic report. This approach effectively bridges the cross-granularity semantic gap while ensuring the accuracy and interpretability of composite job matching results.

[0014] Figure 1 This is a flowchart of an intelligent human-job matching method based on job requirement analysis, according to an embodiment of this application. Figure 2 This is a data flow diagram illustrating an intelligent person-job matching method based on job requirement analysis according to an embodiment of this application. Figure 1 and Figure 2As shown, an intelligent job matching method based on job requirement analysis according to an embodiment of this application includes: S1, using a pre-trained language model to perform semantic parsing and hierarchical knowledge extraction on the acquired job requirement data to obtain a job requirement tree; S2, extracting skill entities and dynamically quantifying proficiency from the acquired candidate resume data to obtain a dynamic skill evolution tree; S3, performing feature encoding and cross-granularity structure penalty distance quantization on all nodes in the job requirement tree and the dynamic skill evolution tree to obtain an asymmetric cost matrix; S4, using the asymmetric cost matrix as a cost function constraint, performing hierarchical optimal transmission solution and pooling dimensionality reduction on the node probability distributions of the job requirement tree and the dynamic skill evolution tree to obtain a bulldozer distance scalar, a transmission path matrix, and a topology alignment feature vector; S5, performing comprehensive matching degree estimation on the bulldozer distance scalar and the topology alignment feature vector to obtain a job matching degree score; S6, performing node mapping and source tracing analysis on the transmission path matrix and the job matching degree score to obtain an interpretable matching diagnostic report.

[0015] Specifically, in step S1, a pre-trained language model is used to perform semantic parsing and hierarchical knowledge extraction on the acquired job requirement data to obtain a job requirement tree. It should be noted that, given that the power industry job system encompasses various categories such as operation and maintenance, marketing, R&D, and management, job requirement data is typically described using highly generalized industry terminology and abstract competency requirements. Its semantic granularity is far coarser than the specific project experience descriptions in candidate resumes. If unstructured job text is directly matched with candidate resumes using keywords or shallow semantic similarity calculations, the granularity gap between macro-level competency dimensions and the underlying specific technology stack cannot be bridged. This leads to a systematic omission of high-value talent with genuine practical experience who do not directly match the terminology. Therefore, the technical solution of this application first uses a pre-trained language model to perform semantic parsing and hierarchical knowledge extraction on the acquired job requirement data to obtain a job requirement tree. Through the above processing, unstructured job descriptions and qualification requirements can be effectively transformed into structured tree representations with parent-child hierarchical topological relationships. This provides a target measurement basis with a clear semantic hierarchy for subsequent cross-granularity node alignment based on optimal transport theory, thereby eliminating the loss of matching accuracy caused by the asymmetry of expression granularity between job requirements and candidate resumes.

[0016] Figure 3 This document describes a flowchart illustrating an intelligent job-person matching method based on job requirement analysis, according to an embodiment of this application. It describes how a pre-trained language model performs semantic parsing and hierarchical knowledge extraction on acquired job requirement data to obtain a job requirement tree. For example... Figure 3As shown, step S1 includes: S11, using the encoder of a pre-trained language model to perform semantic parsing and entity span probability filtering on the job description text and the qualification requirement text to obtain a set of macro-capability dimension nodes; S12, using the set of macro-capability dimension nodes as contextual constraints, using the pre-trained language model to perform zero-shot drill-down relation extraction on the skill tag set to obtain a set of underlying specific technology stack nodes; S13, based on the business scenario identifier, performing topological mapping and mask pruning topological connection on the set of macro-capability dimension nodes and the set of underlying specific technology stack nodes to obtain a job requirement tree.

[0017] In step S11, the encoder of a pre-trained language model performs semantic parsing and entity span probability filtering on the job description text and the qualification requirement text to obtain a set of macro-level capability dimension nodes. It should be noted that because the capability requirements contained in the job description text and the qualification requirement text are nested in long sentences in natural language form, there is a lack of explicit boundary labeling between different capability dimensions. Relying solely on rule matching or word frequency statistics cannot accurately identify macro-level capability entities with complete semantic boundaries. Therefore, the technical solution of this application first uses the encoder of a pre-trained language model to perform semantic parsing and entity span probability filtering on the job description text and the qualification requirement text to obtain a set of macro-level capability dimension nodes. Through the above processing, semantically complete and clearly defined macro-level capability dimension nodes can be extracted from unstructured text, providing contextual constraint anchors for subsequent drill-down extraction of the underlying technology stack.

[0018] More specifically, in a concrete example of this application, the job description text and the qualification requirements text are first concatenated to form a contextual global text sequence. Taking the position of intelligent scheduling algorithm engineer as an example, the job description text includes responsibilities such as developing power load prediction models and deploying distributed computing platforms, while the qualification requirements text includes qualifications such as the ability to apply deep learning frameworks and experience in large-scale data processing. The concatenation of the two forms a complete input sequence covering all the semantic capabilities of the position. The contextual global text sequence is then input into the encoder of a pre-trained language model. After context modeling using a multi-layer self-attention mechanism, a contextual hidden state feature matrix corresponding one-to-one with the input lexical units is output. The vector at each lexical unit position in this matrix incorporates global contextual information. Based on this, the contextual hidden state feature matrix is ​​input into a feedforward neural network classifier. For each candidate lexical unit span, the extraction probability of it belonging to a macroscopic capability entity is calculated. The calculation logic for this extraction probability is that the hidden layer feature vector is linearly transformed by the weight matrix and a bias term is added, then mapped to the 0-1 interval through a Sigmoid activation function. in, For a given job description text Job Qualification Requirements Text Candidate nodes under certain conditions The probability value of becoming a macroscopic capability entity. It is the Sigmoid activation function. This refers to the pre-trained weight matrix in a feedforward neural network classifier. Candidate nodes The corresponding contextual hidden state feature vector, This is the bias vector. Finally, entities with extraction probabilities exceeding the preset boundary threshold are filtered and deduplicated. In the above scenario of intelligent scheduling algorithm engineer positions, the macro-capability dimension node set obtained after filtering includes several semantically complete first-level capability dimension nodes such as power AI algorithm capabilities, engineering development capabilities, and power grid operation and maintenance capabilities.

[0019] Specifically, before applying the aforementioned pre-trained language model to semantic parsing of job requirement data, it needs to undergo domain-adaptive training for the power industry's job-person matching task. First, a large-scale unlabeled corpus is collected from publicly available job posting texts, qualification specification documents, and skill standard systems within the power industry. This corpus is then used to further pre-train the general pre-trained language model using a masked language modeling task, enabling the model to learn the power industry's unique terminology and hierarchical concept associations. Subsequently, based on a manually labeled power industry job competency entity annotation dataset, the pre-trained model undergoes supervised fine-tuning. The model is trained using a sequence labeling task to identify the start and end boundaries and category labels of macro-level competency entities in job posting texts. During the fine-tuning phase, the cross-entropy loss function is used to measure the deviation between the model's predicted entity span probability distribution and the manually labeled true labels. The model parameters are iteratively updated through backpropagation and gradient descent until the entity recognition precision and recall on the validation set reach a preset convergence threshold. This completes the model training and solidifies the parameters for subsequent inference.

[0020] In step S12, using the set of macro-capability dimension nodes as contextual constraints, a pre-trained language model is used to perform zero-shot drill-down relationship extraction on the skill tag set to obtain the underlying specific technology stack node set. It should be noted that since the skill terms in the skill tag set exist as flat, isolated tags, lacking subordinate classification information with the upper-level macro-capability dimensions, without introducing contextual constraints, it is impossible to determine which macro-capability branch each skill tag belongs to. Based on this, the technical solution of this application further uses the set of macro-capability dimension nodes as contextual constraints and a pre-trained language model to perform zero-shot drill-down relationship extraction on the skill tag set to obtain the underlying specific technology stack node set. Through the above processing, a clear subordinate relationship can be established between isolated skill tags and their corresponding macro-capability dimensions, providing a set of child nodes with hierarchical belonging information for the subsequent assembly of the tree-like topology.

[0021] More specifically, in a concrete example of this application, the skill tag set is first segmented and standardized to construct a candidate skill word array. In the scenario of an intelligent scheduling algorithm engineer, the skill tag set contains several isolated skill tag terms such as Python programming, TensorFlow framework, Spark distributed computing, power load forecasting, and relay protection setting. After standardization, a candidate skill word array with a unified format is formed. Subsequently, using each node in the macro-capability dimension node set as contextual constraints and the candidate skill word array as the candidate answer domain, a prompt word injection template is constructed. This template arranges the semantic information of the macro-capability dimension nodes and the candidate skill terms into a structured input format that can be understood by the pre-trained language model. Based on this, the pre-trained language model is called to perform zero-shot relation extraction prediction. Utilizing the domain knowledge learned by the model during the pre-training stage, it is determined whether there is a subordinate relationship between each candidate skill term and each macro-capability dimension, thereby aligning and classifying isolated skill terms to their corresponding macro-capability dimensions. In the aforementioned job scenarios, after zero-sample drill-down relationship extraction, Python programming and Spark distributed computing are categorized under the Engineering Development Capability branch, TensorFlow framework and power load forecasting are categorized under the Power AI Algorithm Capability branch, and relay protection setting is categorized under the Power Grid Operation and Maintenance Capability branch. The aforementioned categorized skill entities together constitute the underlying specific technology stack node set.

[0022] In step S13, based on the business scenario identifier, topological mapping and masked pruning are performed on the macro-capability dimension node set and the underlying specific technology stack node set to obtain the job requirement tree. It should be noted that although a preliminary subordinate relationship has been established between the macro-capability dimension node set and the underlying specific technology stack node set, this relationship is only inferred from the general knowledge of the language model and has not been validated for the rationality of specific business scenarios. The correlation strength between the same skill tag and capability dimension varies in different job scenarios within the power industry. If all correlations are retained indiscriminately, noisy connection edges unrelated to the current job's business scenario will be introduced, interfering with the accuracy of the cost matrix in the subsequent optimal transmission solution. Therefore, the technical solution of this application further performs topological mapping and masked pruning on the macro-capability dimension node set and the underlying specific technology stack node set based on the business scenario identifier to obtain the job requirement tree. Through the above processing, unreasonable node connection edges in specific business scenarios can be eliminated according to the pre-set scenario constraint rules in the industry knowledge graph. The parent-child node relationships that have been screened for scenario adaptation can be assembled into a job requirement tree with a clear hierarchical topology, providing a structured and scenario-related tree input for feature encoding and target metric extraction of optimal transmission in the downstream graph attention network.

[0023] Figure 4 This document describes a flowchart illustrating a business scenario-based intelligent person-job matching method based on job requirement analysis, which involves topological mapping and mask pruning of a set of macro-level capability nodes and a set of underlying specific technology stack nodes to obtain a job requirement tree. For example... Figure 4 As shown, step S13 includes: S131, performing mask pruning and branch attribution determination on the macro capability dimension node set and the underlying specific technology stack node set based on business scenario identifiers and industry knowledge graphs to obtain a pruned adjacency matrix and a branch attribution label vector; S132, filtering cross-branch node pairs in the underlying specific technology stack node set based on the branch attribution label vector, and quantifying the co-occurrence collaboration strength of the cross-branch node pairs to obtain a collaboration coupling strength matrix; S133, based on the collaboration enhancement coefficients of each technology stack node calculated from the collaboration coupling strength matrix, performing enhancement injection and hierarchical topology assembly on the corresponding edge weights in the pruned adjacency matrix to obtain a job requirement tree.

[0024] In step S131, mask pruning and branch attribution determination are performed on the macro-capability dimension node set and the underlying specific technology stack node set based on business scenario identifiers and industry knowledge graphs to obtain a pruned adjacency matrix and a branch attribution label vector. It should be noted that, given that the cosine similarity calculation and scenario mask pruning operations in existing methods only generate an adjacency matrix describing vertical membership relationships, without explicitly labeling which macro-capability branch each technology stack node belongs to, subsequent processes cannot determine whether any two technology stack nodes constitute a cross-branch pairing, thus hindering the identification and quantification of collaborative coupling relationships. Therefore, the technical solution of this application further performs mask pruning and branch attribution determination on the macro-capability dimension node set and the underlying specific technology stack node set based on business scenario identifiers and industry knowledge graphs to obtain a pruned adjacency matrix and a branch attribution label vector. Through the above processing, explicit branch affiliation coding information can be generated while completing the construction of basic adjacency relationships. This provides subsequent steps with explicit evidence to determine which skill pairs are cross-branch, filling the blind spot of unidentifiable relationships between different branches and laying a computational data foundation for the quantitative mining of cross-branch collaborative coupling relationships.

[0025] More specifically, in a concrete example of this application, the scenario relationship mask matrix corresponding to the business scenario is first retrieved from a pre-defined industry graph rule base based on the business scenario identifier of the job position. Then, high-dimensional word vectors of all nodes in the macro-capability dimension node set and high-dimensional word vectors of all nodes in the underlying specific technology stack node set are extracted, and a bipartite graph structure original relevance adjacency matrix is ​​generated through cosine similarity calculation. Next, the original relevance adjacency matrix and the scenario relationship mask matrix are subjected to a Hadamard product operation to prune and remove unreasonable node connections in specific business scenarios, outputting a pruned adjacency matrix.

[0026] Based on this, each row of the pruned adjacency matrix is ​​traversed, with each row corresponding to a low-level technology stack node. The column index with the largest connection weight value in that row is found, and this column index corresponds to a macro-capability dimension node. This index is recorded as the macro-capability branch label to which the current technology stack node belongs. After repeating the above operation for all technology stack nodes, all the labels are integrated and encoded to output a branch affiliation label vector. In the scenario of an intelligent scheduling algorithm engineer, this process is specifically manifested as follows: Python programming is labeled as belonging to the engineering development capability branch, power load prediction is labeled as belonging to the power AI algorithm capability branch, and relay protection setting is labeled as belonging to the power grid operation and maintenance capability branch. Thus, subsequent steps obtain explicit branch affiliation encoding information, providing a computable data foundation for the quantitative mining of cross-branch collaborative coupling relationships.

[0027] In step S132, cross-branch node pairs are selected from the underlying specific technology stack node set based on the branch affiliation label vector, and the co-occurrence synergy strength of the cross-branch node pairs is quantified to obtain a synergy coupling strength matrix. It should be noted that simply knowing two nodes belong to different branches does not equate to actual synergy coupling between them. It is also necessary to mine the actual co-occurrence tightness of the two nodes in the business scenario from real job requirement data, transforming the vague qualitative judgment of cross-branch into a calculable quantitative indicator of synergy coupling strength. Based on this, the technical solution of this application further selects cross-branch node pairs from the underlying specific technology stack node set based on the branch affiliation label vector and quantifies the co-occurrence synergy strength of the cross-branch node pairs to obtain a synergy coupling strength matrix. Through the above processing, the implicit functional interdependence patterns between skills can be explicitly mined from real job requirement corpora in a data-driven manner, compensating for the structural defect of complete orthogonal isolation between different macro-level capability branches.

[0028] More specifically, in a concrete example of this application, firstly, based on the branch affiliation label vector, all nodes in the underlying specific technology stack node set are paired and enumerated. Only node pairs with different branch affiliation labels are retained, while node pairs belonging to the same macro-capability branch are removed, thus constructing a cross-branch candidate node pair set. This filtering operation ensures that the synergistic coupling analysis focuses only on genuine cross-disciplinary skill combinations, avoiding meaningless synergy calculations between skills within the same branch, such as Java programming and Go programming. This reduces computational complexity while improving the scenario relevance of subsequent analyses.

[0029] Subsequently, for each pair of nodes in the cross-branch candidate node pair set, the co-occurrence probability and their respective marginal occurrence probability are calculated in a pre-built industry job description corpus, and the statistical association strength between the two is measured using point mutual information. For technology stack nodes... With technology stack nodes The calculation logic for the original point mutual information value of the cross-branch node pair is to take the base-2 logarithm of the ratio of the product of the joint occurrence probability and the individual independent occurrence probability of the two nodes. in, For nodes With nodes The original point mutual information values ​​of the cross-branch node pairs are formed. For nodes With nodes The joint probability of two or more job descriptions appearing simultaneously in the same job description text within an industry job description corpus. For nodes The marginal probability of occurrence in the corpus For nodes The marginal probability of occurrence in the corpus. When the point mutual information value is positive and large, it means that the frequency of co-occurrence of the two skills in real job descriptions is higher than the product of their individual random occurrence frequencies, representing a strong functional interdependence.

[0030] Because the original point mutual information values ​​have an inflationary bias for low-frequency co-occurrence events, even accidental co-occurrence of two skills with extremely low individual probabilities can produce abnormally high point mutual information values. To eliminate this statistical bias, the original point mutual information values ​​need to be self-information normalized. in, For nodes With nodes The normalized point mutual information values ​​between them are constrained to a range of values. The interval is defined as follows: 1 indicates perfect co-occurrence, meaning the two skills always appear together in the corpus; 0 indicates statistical independence, meaning the co-occurrence of the two skills perfectly matches random expectations. Indicates complete mutual exclusion. For nodes With nodes The original point mutual information values ​​of the cross-branch node pairs are formed. The self-information of the joint probability is used as the normalization denominator to eliminate the inflationary effect of low-frequency events on point mutual information. After normalization, the synergy strength between different skill pairs is compressed to a unified metric, ensuring comparability across positions and scenarios.

[0031] Finally, the normalized point mutual information values ​​of all cross-branch candidate node pairs are non-negatively truncated and cross-branch indicator constraints are superimposed. The results are then assembled into a symmetric matrix to output the cooperative coupling strength matrix. The matrix contains the [missing information - likely a specific matrix or instruction]. Line 1 The calculation logic for the element values ​​of a column is as follows: in, Technology stack nodes in the cooperative coupling strength matrix With technology stack nodes The cross-branch cooperative coupling strength scalar between them As a non-negative truncation function, it sets the negative normalized point mutual information value to zero, retaining only the positive cooperative coupling signal. For nodes With nodes Normalized point mutual information values ​​between them For cross-branch indicator functions, when node Branch affiliation tag With nodes Branch affiliation tag The value is 1 when they are not equal, and 0 otherwise. The combined effect of non-negative truncation and indicator functions constitutes a double-filtering gating. The first gating eliminates node pairs that do not constitute positive cooperation in a statistical sense, and the second gating ensures that node pairs that only cross branches are given non-zero coupling strength.

[0032] For example, in the power industry corpus, the point mutual information value between distributed computing frameworks and load forecasting modeling is much higher than that between distributed computing frameworks and transformer maintenance. This result confirms from a data perspective that the former does indeed constitute a strong collaborative pair for intelligent dispatching positions, while the latter lacks a business collaborative relationship. Distributed computing frameworks belong to the engineering development capability branch, while load forecasting modeling belongs to the power AI algorithm capability branch. The two form a typical strong collaborative coupling pair in job responsibilities. The former is the necessary operational carrier for the engineering implementation of the latter, and the latter is the core enabling outlet for the former to generate business value. When a candidate only possesses one of these skills, the actual business value of that single skill will be diminished due to the lack of collaborative partners.

[0033] In step S133, based on the collaborative enhancement coefficients of each technology stack node calculated from the collaborative coupling strength matrix, the corresponding edge weights in the pruned adjacency matrix are enhanced and injected with hierarchical topology to obtain the job requirement tree. It should be noted that, given that the horizontal collaborative information in the collaborative coupling strength matrix exists as an independent matrix and is not integrated with the vertical hierarchical topology, if the assembly logic of simply mapping the pruned adjacency matrix to parent-child connection edges is directly adopted, the cross-branch collaborative information will still be discarded. The edge weights of each technology stack node in the tree will be determined solely by the cosine similarity between it and its parent node, failing to reflect the structural importance gain that the node should receive due to participating in cross-branch collaboration. Therefore, the technical solution of this application further enhances the corresponding edge weights in the pruned adjacency matrix based on the collaborative enhancement coefficients of each technology stack node calculated from the collaborative coupling strength matrix, and then performs hierarchical topology assembly to obtain the job requirement tree. Through the above processing, the structural information of horizontal collaborative coupling can be encoded into the vertical edge weights of the tree, so that the key technology stack nodes that play the role of multi-branch bridging hubs can obtain higher edge weights and obtain a greater probability quality share in the extraction of downstream optimal transmission metric distribution.

[0034] More specifically, in a concrete example of this application, the non-zero connection weights in the pruned adjacency matrix are first mapped to parent-child node connection edges with a hierarchical topology. Using each node in the macro-capability dimension node set as a first-level parent node, the data structure is instantiated and assembled from top to bottom to construct a basic hierarchical topology tree. Subsequently, based on the cooperative coupling strength matrix, a cooperative enhancement coefficient is calculated for each technology stack node. For the... Technology stack nodes Its synergistic enhancement coefficient is equal to the base value plus the product of the synergistic amplification factor and the sum of all elements in the corresponding row of the synergistic coupling strength matrix for that node: in, For the first Technology stack nodes The synergistic enhancement coefficient scalar has a baseline value of 1, meaning that the original weights remain unchanged when there is no synergistic gain. The preset hyperparameter for the cooperative amplification factor controls the sensitivity of the cooperative coupling information to the modulation of node weights. This represents the total number of valid technology stack nodes contained in the underlying specific technology stack node set. For nodes in the cooperative coupling strength matrix With nodes The scalar of cross-branch cooperative coupling strength between them. Summation term. The mathematical meaning of is to connect nodes The strength of the cooperative coupling with all other nodes is globally aggregated. The more strong cooperative relationships a node participates in and the stronger the coupling, the higher the cumulative value and the greater the cooperative enhancement coefficient of the node.

[0035] Finally, the synergy enhancement coefficient is multiplied and added to the original weight of the edge connecting the corresponding technology stack node and its parent node in the basic hierarchical topology tree to obtain the enhanced edge weight injected with cross-branch synergy coupling information: in, Parent node for macro-level capabilities in the job requirement tree With technology stack child nodes The final connection edge weights after collaborative enhancement The macroscopic capability parent node in the pruned adjacency matrix With technology stack child nodes The original edge weights between the links are the cosine similarity values ​​after mask pruning. For technology stack child nodes The synergistic enhancement coefficient. This multiplicative injection operation essentially encodes the horizontal structural information of cross-branch synergistic coupling into the vertical edge weights of the tree. The larger the synergistic enhancement coefficient of a node, the higher its edge weight in the tree, and the greater its probability quality share in the optimal transmission metric distribution extraction in the downstream step four. All enhanced edge weights and topology are solidified, ultimately generating and outputting the job requirement tree.

[0036] In the context of intelligent scheduling algorithm engineers, if the distributed computing framework nodes and multiple nodes from different macro-level capability branches, such as load forecasting modeling, new energy output forecasting, and scheduling decision optimization, all have high cooperative coupling strength, then its A value greater than 1 reflects the crucial role this skill plays as a bridging hub between multiple branches in intelligent scheduling and other complex roles. Conversely, nodes that are only associated with skills within their own branch and do not participate in cross-branch collaboration, such as code style management, will have zero elements in the corresponding rows of their collaboration coupling strength matrix. The value is always equal to 1, keeping the original weight unchanged.

[0037] Specifically, in step S2, skill entity extraction and dynamic proficiency quantification are performed on the acquired candidate resume data to obtain a dynamic skill evolution tree. It should be noted that, given that skill information in candidate resume data is scattered across multiple work records with varying time spans in the form of unstructured project experience descriptions, and that the actual proficiency level of the same skill varies depending on usage duration, frequency, and time interval, if all candidate skills are treated as a static, equally weighted set of flat labels for matching, it would be impossible to perceive the increase in proficiency over time or the decline in ability due to long-term inactivity. This would lead to inflated or inflated matching judgments for candidates who have recently changed positions or have been away from frontline practical work for a long time. Based on this, the technical solution of this application further performs skill entity extraction and dynamic proficiency quantification on the acquired candidate resume data to obtain a dynamic skill evolution tree. Through the above processing, the candidate's historical resume can be transformed into a dynamic tree representation with temporal weight distribution and hierarchical topology, accurately depicting the candidate's current true ability profile and providing a timely source metric distribution basis for optimal transmission alignment with the job requirement tree.

[0038] Figure 5 This document describes a flowchart illustrating the process of extracting skill entities and dynamically quantifying proficiency from acquired candidate resume data to obtain a dynamic skill evolution tree, based on an intelligent person-job matching method according to job requirement analysis, as described in an embodiment of this application. Figure 5 As shown, step S2 includes: S21, extracting skill entities from the project experience description text sequence using a named entity recognition model to obtain a set of candidate skill entity nodes, and calculating the logarithmic growth proficiency of each node based on the work time interval sequence to obtain a set of initial weight scalars for the nodes; S22, performing temporal decay and reuse transition weighting on the set of candidate skill entity nodes and the set of initial weight scalars for the nodes to obtain a set of current weight scalars for dynamic skills; S23, parsing the hierarchical relationship between nodes in the set of candidate skill entity nodes based on the technology ontology knowledge base and constructing directed connection edges, and assigning the set of current weight scalars for dynamic skills to the corresponding nodes to obtain a dynamic skill evolution tree.

[0039] In step S21, a named entity recognition model is used to extract skill entities from the project experience description text sequence to obtain a set of candidate skill entity nodes. Then, the logarithmic growth proficiency of each node is calculated based on the work time interval sequence to obtain an initial set of node weight scalars. It should be noted that since the skill information in the project experience description text sequence is embedded in mixed text such as project background and job description in natural language narrative form, and each skill has a differentiated level of usage accumulation due to different project durations, without precise entity-level extraction and initial quantification in the time dimension, a calculable node basis cannot be provided for subsequent time-series decay and reuse transition weighting. Based on this, the technical solution of this application further uses a named entity recognition model to extract skill entities from the project experience description text sequence to obtain a set of candidate skill entity nodes, and calculates the logarithmic growth proficiency of each node based on the work time interval sequence to obtain an initial set of node weight scalars. Through the above processing, discrete skill entities can be extracted from unstructured resume text and assigned initial proficiency quantification values ​​based on usage duration, providing a starting data basis for subsequent dynamic weight evolution.

[0040] More specifically, in a concrete example of this application, a pre-trained named entity recognition sequence labeling model is first used to decode and extract the project experience description text sequence segment by segment to obtain the skill entity elements that the candidate has actually applied in past projects, forming a candidate skill entity node set. Taking a candidate applying for the position of intelligent scheduling algorithm engineer as an example, their project experience description text sequence contains 3 project records. After extraction by named entity recognition, the candidate skill entity node set obtained includes skill entities such as Spark distributed computing, power load forecasting, Python programming, and TensorFlow framework. Subsequently, the candidate skill entity node set is mapped one-to-one with the work time interval sequence according to the sequence index to calculate the project duration span of the candidate's use of each skill entity. Based on this, according to the project duration span, a logarithmic growth model is used to calculate the initial proficiency of each skill entity formed by the accumulation of usage time. The calculation logic is: the basic proficiency gain coefficient multiplied by the log smooth growth term corresponding to the project duration span. in, For the first The initial weight scalar of each skill entity, This is a preset base proficiency gain coefficient, used to adjust the marginal contribution rate of usage time to skill proficiency. For the first The duration of each skill entity's corresponding project within the work time interval sequence is calculated in months. A time constant threshold is used to smooth out outlier fluctuations caused by extremely short time spans. A logarithmic function is used instead of a nonlinear function because skill proficiency growth follows a diminishing marginal return principle; that is, initial practice contributes significantly to proficiency improvement, while the marginal gain gradually slows down as the duration increases. In the candidate scenarios described above, the Spark distributed computing skill project lasted 18 months, and the power load forecasting skill project lasted 12 months. After logarithmic growth calculations, their respective initial node weight scalars were obtained, collectively forming the initial node weight scalar set.

[0041] In step S22, the candidate skill entity node set and the initial node weight scalar set are subjected to temporal decay and reuse transition weighting to obtain the dynamic skill current weight scalar set. It should be noted that since the initial node weight scalar set only reflects the cumulative time contribution of the candidate's skill usage in various projects, it does not consider the ability decay caused by the skill's inactivity interval and the proficiency enhancement resulting from repeated use of the same skill in multiple projects. If the initial weight is directly used as the final weight, it cannot accurately depict the candidate's current true ability profile. Therefore, the technical solution of this application further applies temporal decay and reuse transition weighting to the candidate skill entity node set and the initial node weight scalar set to obtain the dynamic skill current weight scalar set. Through the above processing, the time forgetting effect and the reinforcement effect of multiple reuses of skills can be simultaneously incorporated into the weight calculation, making the final weight accurately reflect the candidate's actual skill proficiency status at the current time point.

[0042] More specifically, in a concrete example of this application, the set of candidate skill entity nodes is first traversed, and nodes of the same skill entity are merged to identify the number of times the skill has been reused across multiple historical projects. Taking the candidate mentioned above as an example, their Spark distributed computing skill appeared twice in three project records, their power load forecasting skill appeared once, and their Python programming skill appeared three times. Then, the idle time interval from the most recent use end time to the current evaluation time is calculated for each identical skill entity. Based on this, a negative exponential time decay function containing a forgetting constant is applied to the corresponding values ​​in the initial weight scalar set of the nodes. Simultaneously, for skills reused in subsequent projects, a transition activation factor is applied based on their reuse sequence position for compensation enhancement. The final value of the same skill entity after multiple project stacking, decay, and transition activation is summarized. For a specific skill entity... The calculation logic for its dynamic current weight scalar is based on the skill's performance across all historical data. The decayed sum of the initial weights generated in each project: in, For the first The dynamic current weight scalar of a class-specific skill entity. The first in the candidate's resume Total number of reuses for a class-specific skill entity For the first Skill-type entities in the first The initial weight scalar is calculated from the previous sub-step during the next reuse. The skill forgetting decay constant is based on the Ebbinghaus forgetting curve. The base timestamp for the current matching run. For the first Class of skills in Use the end timestamp in each project. Indicates the time interval between skill idle periods. To reuse the gain rate of the transition activation factor, This represents the weighted transition amplification factor that mitigates the decay of a skill each time it is reused. In the candidate scenarios above, the Python programming skill was used in all three projects, and the most recent use was only two months ago; therefore, its decay term is close to 1, and its transition amplification factor is [value missing]. The final dynamic current weight is relatively high; however, the power load forecasting skill has only been used in one project and has been in use for 24 months. Its attenuation term is close to a small value due to the long idle time and it lacks reuse transition gain, resulting in a relatively low final dynamic current weight. After completing the above calculations for all skill entities, the set of dynamic skill current weight scalars is output.

[0043] In step S23, the hierarchical relationships between nodes in the candidate skill entity node set are analyzed based on the technology ontology knowledge base, and directed connections are constructed. The current weight scalar set of dynamic skills is then assigned to the corresponding nodes to obtain the dynamic skill evolution tree. It should be noted that since the candidate skill entity node set and the current weight scalar set of dynamic skills still exist in a flat list format at this point, there is a lack of hierarchical topological relationships between the skill nodes. If they are not organized into a tree structure isomorphic to the job requirement tree, the graph attention network cannot perform neighborhood aggregation encoding in subsequent steps, and the optimal transmission solution cannot calculate cross-granularity costs under structurally equivalent conditions. Therefore, the technical solution of this application further analyzes the hierarchical relationships between nodes in the candidate skill entity node set based on the technology ontology knowledge base and constructs directed connections, assigning the current weight scalar set of dynamic skills to the corresponding nodes to obtain the dynamic skill evolution tree. Through the above processing, isolated skill nodes can be assembled into a structured tree diagram with topological hierarchical depth, making the candidate's ability representation and the job requirement tree equivalent in structural dimension, providing a source input with hierarchical semantics for subsequent cross-granularity feature alignment and optimal transmission solution.

[0044] More specifically, in a concrete example of this application, a pre-defined ontology knowledge base for computer and power engineering technology is first invoked to parse the hierarchical relationship between entities in the candidate skill entity node set. In the aforementioned candidate scenario, the ontology knowledge base defines distributed computing as a subordinate concept of engineering development, Spark distributed computing as a subordinate concept of distributed computing, power load forecasting as a subordinate concept of power AI algorithms, and the TensorFlow framework as a subordinate concept of deep learning tools, thereby determining the parent-child hierarchy between nodes. Subsequently, based on the parsed subordinate relationship, directed connections between parent and child nodes are constructed, assembling isolated skill entity nodes into a structured tree diagram with topological hierarchy depth. On this basis, the values ​​corresponding to the current weight scalar set of dynamic skills are assigned as numerical feature attributes to each corresponding leaf node in the tree structure. Python programming receives a higher weight value due to high frequency of reuse and recent use, while power load forecasting receives a lower weight value due to long-term inactivity. The topological edge relationships and node weight distribution of the entire tree are solidified, ultimately generating a dynamic skill evolution tree.

[0045] Specifically, in step S3, feature encoding and cross-granularity structure penalty distance quantization are performed on all nodes in the job requirement tree and the dynamic skill evolution tree to obtain an asymmetric cost matrix. It should be noted that, given that the nodes in the job requirement tree and the dynamic skill evolution tree are at different semantic granularity levels, there is a natural gap in abstraction between macro-level capability nodes and underlying specific technology stack nodes. Furthermore, the cost of transferring features from a candidate's underlying skills to the macro-level job requirements is not equal to the cost of transferring features in the reverse direction in terms of business logic. Matching low-level coding requirements with high-level architectural experience is less costly, while forcibly matching macro-level architectural requirements with low-level coding experience is extremely costly. If a symmetric distance metric is used, it cannot reflect this directional difference, leading to an inflated matching judgment for candidates with only fragmented underlying skills in the optimal transfer solution. Based on this, the technical solution of this application further performs feature encoding and cross-granularity structure penalty distance quantization on all nodes in the job requirement tree and the dynamic skill evolution tree to obtain an asymmetric cost matrix. Through the above processing, the semantic features and hierarchical depth information of all nodes in the two trees can be fused into a direction-sensitive cost function, providing an asymmetric transport cost constraint that simultaneously encodes semantic distance and structural misalignment penalty for subsequent optimal transport solution, ensuring that the matching process has the ability to perceive cross-granularity directional differences.

[0046] Figure 6 This is a flowchart illustrating the process of feature encoding and cross-granularity structure penalty distance quantization for all nodes in a job requirement tree and a dynamic skill evolution tree, according to an embodiment of this application, to obtain an asymmetric cost matrix. (See flowchart for example.) Figure 6As shown, step S3 includes: S31, using a graph attention network to perform multi-hop neighborhood aggregation encoding on all nodes in the job requirement tree and the dynamic skill evolution tree to obtain the set of feature vectors for job nodes and the set of feature vectors for candidate skill nodes; S32, calculating the cross-granularity semantic distance matrix between the set of feature vectors for job nodes and the set of feature vectors for candidate skill nodes; S33, based on the cross-granularity semantic distance matrix, applying penalty correction to the job requirement tree and the dynamic skill evolution tree to obtain an asymmetric cost matrix.

[0047] In step S31, a graph attention network is used to perform multi-hop neighborhood aggregation encoding on all nodes in the job requirement tree and the dynamic skill evolution tree to obtain the set of feature vectors for job nodes and the set of feature vectors for candidate skill nodes. It should be noted that since the initial word vectors of each node in the job requirement tree and the dynamic skill evolution tree only contain isolated semantic information of the node itself, without incorporating the contextual semantic relationships between the node and its parent, child, and sibling nodes in the tree topology, directly calculating the distance between nodes using isolated word vectors cannot reflect the influence of the node's structural position in the tree on its semantic expression. Based on this, the technical solution of this application further utilizes a graph attention network to perform multi-hop neighborhood aggregation encoding on all nodes in the job requirement tree and the dynamic skill evolution tree to obtain the set of feature vectors for job nodes and the set of feature vectors for candidate skill nodes. Through the above processing, the feature representation of each node can be integrated with the contextual semantic information of its local topological neighborhood, providing a structure-aware, high-quality feature representation for subsequent cross-granularity semantic distance calculation.

[0048] More specifically, in a concrete example of this application, the topological structures of the job requirement tree and the dynamic skill evolution tree are first traversed separately, extracting the tree hierarchy depth values ​​of all nodes, and simultaneously obtaining the initial word vector features of each node as input to the graph attention network. Subsequently, a multi-layer graph attention network is constructed, performing multi-hop information transmission within the local neighborhood of each tree structure, calculating the importance weights between adjacent nodes through an attention mechanism, and aggregating the features of neighboring nodes to update the representation of the central node. For the target node... In the The hidden state feature vector of the layer is updated by calculating the logic of the first layer's update of the target node and its local neighbor node set. The layer features are linearly transformed, multiplied by the attention weight scalar, summed, and finally nonlinearly mapped using the exponential linear unit activation function. in, For nodes After the first The feature representation obtained after processing by the layered graph attention network It is the activation function of the exponential linear unit. For nodes The set of first-order adjacent nodes in a tree graph topology. For nodes With nodes The normalized attention weight scalar between nodes represents the contribution of neighbor node information to the feature reconstruction of the central node. For the first in the graph attention network The learnable weight transformation matrix of the layer, For nodes In the The hidden state feature vectors of the layer. Taking the hidden state output of the last layer of the graph attention network, we obtain the feature vector sets of job node features and candidate skill node features, which incorporate the contextual information. In the scenario of the intelligent scheduling algorithm engineer job, the feature vector of the parent node "Power AI Algorithm Capability" in the job requirement tree, after multi-hop aggregation, incorporates the semantic information of its child nodes such as power load prediction and TensorFlow framework. Similarly, the feature vector of the Spark distributed computing node in the candidate dynamic skill evolution tree, after aggregation, incorporates the contextual semantics of its parent node "Distributed Computing" and its sibling nodes.

[0049] In a preferred embodiment of this application, the graph attention network adopts a 2-layer GAT structure, each layer contains 8 attention heads, the hidden dimension is 128, the multi-head attention output is aggregated by splicing, the dropout rate is set to 0.6, and the Adam optimizer is used for end-to-end training with a learning rate of 0.005.

[0050] In step S32, the cross-granularity semantic distance matrix between the set of feature vectors of job nodes and the set of feature vectors of candidate skill nodes is calculated. It should be noted that since the set of feature vectors of job nodes and the set of feature vectors of candidate skill nodes originate from two independent tree structures, no quantified distance relationship has been established between them. Without calculating the basic semantic distance between nodes, subsequent structural penalty correction and optimal transport solution lack an operational cost basis. Therefore, the technical solution of this application further calculates the cross-granularity semantic distance matrix between the set of feature vectors of job nodes and the set of feature vectors of candidate skill nodes. Through the above processing, a basic transport cost based on feature space metric can be established for all node pairs in the two trees, providing a distance basis for subsequent superposition of hierarchical structural penalties.

[0051] More specifically, in a concrete example of this application, a fully connected matching mesh is first constructed, pairing each feature vector in the set of feature vectors for job nodes with each feature vector in the set of feature vectors for candidate skill nodes. Then, based on a multidimensional vector space metric, the Euclidean distance between each pair of vectors in the corresponding dimension is calculated. The calculation logic involves summing the squared differences of the corresponding components of the two feature vectors in each dimension and then taking the square root. in, Job requirement nodes With candidate skill nodes The basic semantic distance scalar between them This is the preset maximum dimension of the feature vector. For nodes in the feature vector set of job nodes The eigenvectors in the th Numerical components of each dimension For the node in the candidate skill node feature vector set The eigenvectors in the th Numerical components in each dimension. The scalar distance values ​​calculated for all node pairs are aggregated into a dense matrix, and the cross-granularity semantic distance matrix is ​​output.

[0052] In step S33, based on the cross-granularity semantic distance matrix, a penalty correction is applied to the job requirement tree and the dynamic skill evolution tree to obtain an asymmetric cost matrix. It should be noted that since the cross-granularity semantic distance matrix only measures the basic Euclidean distance between nodes in the feature space, it does not consider the directional impact of the difference in the hierarchical depth of the nodes in the two trees on the transport cost. The cost of transporting features from the bottom-level specific skill nodes to the macro-level capability requirement nodes should be higher than the cost of reverse transport. If symmetric distance is used as the cost function for optimal transport, this directional difference in hierarchical misalignment cannot be distinguished. Therefore, the technical solution of this application further applies a penalty correction to the job requirement tree and the dynamic skill evolution tree based on the cross-granularity semantic distance matrix to obtain an asymmetric cost matrix. Through the above processing, the difference in hierarchical depth can be introduced into the cost function to form a direction-sensitive transport cost constraint, ensuring that the optimal transport solution process has the ability to perceive the penalty for cross-granularity directional misalignment.

[0053] More specifically, in a concrete example of this application, the topological structures of the job requirement tree and the dynamic skill evolution tree are first traversed separately to extract the level depth value of each node; the smaller the value, the more macroscopic the level. Then, the level depth values ​​of each node in both trees are jointly called. For each pair of job nodes and candidate nodes that have a distance calculation relationship in the cross-granularity semantic distance matrix, an asymmetric exponential level penalty term for the depth difference between the two nodes is calculated based on directional asymmetric logic. This penalty term is then applied as a coefficient multiplier to the corresponding basic semantic distance. The calculation logic for a certain element value in the asymmetric cost matrix is ​​the basic semantic distance scalar between nodes multiplied by an exponential correction structure penalty function incorporating the asymmetric depth difference term. in, For the asymmetric cost matrix from the candidate skill nodes Towards job requirements The mapping value of the comprehensive cost required to transport features. This is the basic semantic distance scalar between node pairs in the cross-granularity semantic distance matrix. The preset structural penalty base weight coefficient is used to adjust the overall response magnitude of the depth difference penalty. The preset asymmetric sensitivity hyperparameter, Job requirement nodes The depth value of the level it is in. For candidate skill nodes The depth value of the hierarchy. When the job requirement node is at the macro level while the candidate skill node is at the bottom specific level, the depth difference is... A negative value indicates a lighter penalty, meaning the cost of matching specific experience with macro-level requirements is relatively acceptable. Conversely, when the candidate node is at the macro level while the job requirement node is at the bottom specific level, the depth difference is positive, the exponent is greater than 1, and the penalty is heavier, meaning the cost of forcibly matching specific technical requirements with generalization ability is high.

[0054] Specifically, in step S4, using the asymmetric cost matrix as a cost function constraint, hierarchical optimal transport solution and pooling dimensionality reduction are performed on the node probability distributions of the job requirement tree and the dynamic skill evolution tree to obtain the bulldozer distance scalar, transport path matrix, and topology alignment feature vector. It should be noted that, given that the asymmetric cost matrix has encoded the direction-sensitive transport cost between all node pairs in the two trees, but has not yet determined the optimal path scheme for feature transport from the candidate skill distribution to the job requirement distribution—that is, it has not yet answered the core operational problem of what allocation strategy to transport the limited ability quality of candidates to the various job requirement nodes to minimize the total cost—relying solely on node-by-node local similarity comparison cannot capture the overall structural matching degree and local coverage gap distribution between the two trees from a global perspective. Based on this, the technical solution of this application further uses the asymmetric cost matrix as a cost function constraint to perform hierarchical optimal transport solution and pooling dimensionality reduction on the node probability distributions of the job requirement tree and the dynamic skill evolution tree to obtain the bulldozer distance scalar, transport path matrix, and topology alignment feature vector. Through the above processing, the job matching problem can be transformed into an optimal transport problem of transferring quality from the candidate's ability distribution to the job requirement distribution. At the same time, we obtain a bulldozer distance scalar representing the total cost of global matching, a transport path matrix recording the specific flow schemes between nodes, and a topology alignment feature vector representing the coverage of the candidate's local optimal experience to the job topology structure. These provide multi-granular data support for subsequent comprehensive matching score and interpretability report generation.

[0055] Figure 7 This document describes a flowchart illustrating an intelligent job-person matching method based on job requirement analysis, constrained by an asymmetric cost matrix. It describes the hierarchical optimal transmission solution and pooling dimensionality reduction of the node probability distributions of the job requirement tree and dynamic skill evolution tree to obtain the bulldozer distance scalar, transmission path matrix, and topology alignment feature vector. Figure 7 As shown, step S4 includes: S41, performing probability quality mapping on the weight values ​​of each node in the job requirement tree and the dynamic skill evolution tree to obtain the target measure probability distribution vector and the source measure probability distribution vector; S42, using the asymmetric cost matrix as the cost function and introducing an information entropy regularization term, performing alternating scaling and iterative solving on the target measure probability distribution vector and the source measure probability distribution vector to obtain the bulldozer distance scalar and the transmission path matrix; S43, performing max pooling on the transmission path matrix row by row along the dimension of the job requirement nodes, extracting the maximum transmission value of each job requirement node among all candidate skill nodes, and concatenating and reducing the dimension to obtain the topology-aligned feature vector.

[0056] In step S41, probability quality mapping is performed on the weight values ​​of each node in the job requirement tree and the dynamic skill evolution tree to obtain the target measure probability distribution vector and the source measure probability distribution vector, respectively. It should be noted that since the optimal transmission solution requires that the inputs at both the source and target ends be discrete measure distributions that satisfy probability normalization constraints, and the weight values ​​of each node in the job requirement tree and the dynamic skill evolution tree are unnormalized raw scalars, they cannot be directly used as marginal constraints for the transmission problem. Therefore, the technical solution of this application further performs probability quality mapping on the weight values ​​of each node in the job requirement tree and the dynamic skill evolution tree to obtain the target measure probability distribution vector and the source measure probability distribution vector, respectively. Through the above processing, the raw weights of each node in the two trees can be transformed into discrete measure distributions that satisfy probability axioms, providing a marginal distribution input that conforms to mathematical constraints for subsequent Sinkhorn iterative solutions.

[0057] More specifically, in a concrete example of this application, the job requirement tree is first traversed, and the edge weight values ​​of each job requirement node after collaboration enhancement are extracted. These weight values ​​are then processed... Norm normalization is applied to make the sum equal to 1, thus constructing a discrete target measure representing the distribution of job requirements and outputting a probability distribution vector of the target measure. In the scenario of intelligent scheduling algorithm engineer positions, distributed computing framework nodes with high collaboration enhancement coefficients obtain a larger probability quality share in the target measure due to their higher edge weights, while code specification management nodes obtain only a smaller probability quality share because their collaboration enhancement coefficient is 1. Subsequently, the dynamic skill evolution tree is traversed to extract the dynamic current weight scalar of each candidate skill node, and the process is repeated similarly. Norm-normalized mapping outputs the source measure probability distribution vector. The nth source measure distribution... The probability quality calculation logic for a candidate skill node is to divide the node's dynamic current weight scalar by the sum of the dynamic current weight scalars of all nodes in the tree: in, For the source measure probability distribution vector corresponding to the first The normalized probability quality scalar of each candidate skill node For the first in the dynamic skill evolution tree The dynamic current weight scalar of each node. This represents the total number of valid nodes in the candidate skill evolution tree. In the candidate scenario described above, Python programming nodes occupy a higher probability quality in the source metric due to their high frequency of reuse and recent continued use, while power load forecasting nodes occupy only a lower probability quality due to their low dynamic current weight caused by long-term inactivity.

[0058] In step S42, using the asymmetric cost matrix as the cost function and introducing an information entropy regularization term, the target measure probability distribution vector and the source measure probability distribution vector are iteratively scaled and solved to obtain the bulldozer distance scalar and the transmission path matrix. It should be noted that since the original optimal transmission problem is a linear programming problem, its computational complexity increases hypercubicly with the number of nodes. When the number of nodes in the job requirement tree and dynamic skill evolution tree is large, directly solving for the exact solution is computationally inefficient. Entropy regularization is introduced to transform it into a smooth convex optimization problem that can be efficiently solved iteratively by alternating matrix scaling. Based on this, the technical solution of this application further uses the asymmetric cost matrix as the cost function and introduces an information entropy regularization term to iteratively scale and solve the target measure probability distribution vector and the source measure probability distribution vector to obtain the bulldozer distance scalar and the transmission path matrix. Through the above processing, the globally optimal transmission scheme can be obtained with a controllable computational cost while ensuring solution accuracy. Simultaneously, the bulldozer distance scalar representing the total matching cost and the transmission path matrix recording the specific mass flow paths between nodes are output.

[0059] More specifically, in a concrete example of this application, a joint probability distribution space is first constructed using the source measure probability distribution vector as the starting point constraint for quality transport and the target measure probability distribution vector as the ending point constraint for quality transport. Then, using the asymmetric cost matrix as the cost objective function, an information entropy regularization term is introduced, transforming the originally non-differentiable linear programming problem into a smooth, strongly convex optimization problem. The computational logic of the optimal transport objective function including the entropy regularization term is as follows: subtract the transport matrix information entropy weighted by the smoothing coefficient from the Frobenius inner product of the joint probability transport matrix and the cost matrix. in, To obtain the bulldozer distance scalar, The obtained transmission path matrix belongs to the joint probability distribution space that satisfies the marginal distribution constraints. , and These are the source measure probability distribution vector and the target measure probability distribution vector, respectively. It is an asymmetric cost matrix. The Frobenius inner product of two matrices is the sum of the corresponding element-wise multiplications. The preset entropy regularization smoothing hyperparameters are used. The information entropy of the transmission path matrix is... Based on this, the Sinkhorn-Knopp matrix alternating scaling iterative algorithm is used to solve the problem. This algorithm alternately normalizes and scales the rows and columns of the transmission matrix to gradually approximate the optimal solution that simultaneously satisfies the source measure edge constraints and the target measure edge constraints, until the convergence condition is met, at which point the bulldozer distance scalar and the transmission path matrix are output.

[0060] In step S43, the transmission path matrix is ​​max-pooled row-by-row along the dimension of job requirement nodes, and the maximum transmission value of each job requirement node among all candidate skill nodes is extracted and concatenated to reduce dimensionality to obtain a topology-aligned feature vector. It should be noted that since the transmission path matrix is ​​a two-dimensional matrix, its dimension is related to the number of nodes in the two trees, and it cannot be directly used as a fixed-dimensional input feature for the subsequent multilayer perceptron scoring network. It needs to be compressed into a one-dimensional compact vector that can represent the degree of coverage of each job requirement node by candidate skills. Based on this, the technical solution of this application further max-pools the transmission path matrix row-by-row along the dimension of job requirement nodes, extracts the maximum transmission value of each job requirement node among all candidate skill nodes, and concatenates and reduces dimensionality to obtain a topology-aligned feature vector. Through the above processing, the local coverage information contained in the transmission path matrix can be compressed into a fixed-dimensional one-dimensional feature representation, providing a dimension-compatible input for subsequent feature concatenation and scoring calculation with the bulldozer distance scalar.

[0061] More specifically, in a concrete example of this application, the transmission path matrix is ​​first processed by row-wise feature extraction, with each row corresponding to the quality transmission allocation of a job requirement node across all candidate skill nodes. Then, max pooling is applied to each row, iterating through all column elements to find the maximum quality transmission value obtained by the current job requirement node from all candidate skill nodes. This value represents the strongest mapping strength where the job requirement node is covered by a candidate's most matching skill. The topology alignment feature vector... The logic for calculating the eigenvalue of the dimension is as follows: in the transmission path matrix... Iterate through all columns in the row and find the maximum value: in, The first eigenvector in the topological alignment feature vector The value of the i-th element represents the i-th element. The mapping strength of a job requirement node being most fully covered by a candidate's specific skill. The transmission path matrix is ​​located at the th Line 1 Column matrix element scalar, This represents the total number of valid nodes in the candidate's skill evolution tree. Finally, the maximum transmission values ​​of all job requirement nodes are concatenated into a one-dimensional array in row index order, and the topology-aligned feature vector is output.

[0062] Specifically, in step S5, a comprehensive matching degree estimation is performed on the bulldozer distance scalar and the topology alignment feature vector to obtain a person-job matching degree score. It should be noted that the bulldozer distance scalar only reflects a single numerical information of the total global matching cost, while the topology alignment feature vector encodes fine-grained distribution information of how each job requirement node is locally covered by the candidate's skills. The two represent the matching status from two complementary perspectives, global and local. Relying solely on either indicator cannot comprehensively consider both the overall cost and local coverage. Therefore, the technical solution of this application further performs a comprehensive matching degree estimation on the bulldozer distance scalar and the topology alignment feature vector to obtain a person-job matching degree score. Through the above processing, the global transportation cost and local coverage intensity can be integrated into a unified matching metric, providing a comprehensive quantitative basis for person-job matching decisions.

[0063] More specifically, in a concrete example of this application, the zero-dimensional bulldozer distance scalar and the one-dimensional topology alignment feature vector are first concatenated along the feature channel dimension to construct a fused feature vector that comprehensively represents global cost and local coverage. The fused feature vector is then input into a pre-trained multilayer perceptron neural network, sequentially undergoing linear weighted summation and ReLU nonlinear activation operations in the hidden layers for high-dimensional nonlinear feature combination. The scalar result from the multilayer perceptron output layer is then subjected to probabilistic constraints using the Sigmoid activation function, mapping to obtain a person-job matching score within the range of 0 to 1. The calculation logic is as follows: the bulldozer distance scalar and the topology alignment feature vector are concatenated using a connector, transformed through two network layers, and finally mapped using the Sigmoid function. in, The output is the person-job fit score. It is the Sigmoid activation function. For topologically aligned feature vectors, For bulldozer distance scalar, This is the vector concatenation operator. and Here are the learnable weight matrices and bias vectors for the hidden layers of a multilayer perceptron. It is a linear rectified activation function. and These are the learnable weight vectors and bias scalars for the output layer of the multilayer perceptron.

[0064] Specifically, the pre-trained multilayer perceptron neural network mentioned above needs to undergo supervised training based on historical job-person matching data before being applied to comprehensive matching degree estimation. First, job-candidate pairings with confirmed hiring results from historical recruitment scenarios are collected. Steps S1 to S4 are executed for each pairing sample to calculate the corresponding bulldozer distance scalar and topology alignment feature vector, and the actual hiring results are used as binary labels to construct the training dataset. Then, the bulldozer distance scalar and topology alignment feature vector of each sample are concatenated and input into the multilayer perceptron network. The binary cross-entropy between the predicted matching degree score output by the network and the true label is used as the loss function. Backpropagation and gradient descent are used to iteratively update the weight matrix and bias parameters of each layer in the network. During training, a mini-batch stochastic gradient descent strategy combined with a learning rate decay mechanism is adopted until the prediction accuracy and loss value on the validation set tend to stabilize, completing the network training and fixing the parameters for subsequent inference.

[0065] Specifically, in step S6, node mapping and source tracing analysis are performed on the transmission path matrix and the person-job matching score to obtain an interpretable matching diagnostic report. It should be noted that since the person-job matching score is only a scalar value, it cannot explain the specific attribution basis of the matching conclusion to business decision-makers. However, the transmission path matrix records the complete allocation scheme of the transmission quality from each skill node of the candidate to each requirement node of the job, containing traceable node-level mapping evidence. Based on this, the technical solution of this application further performs node mapping and source tracing analysis on the transmission path matrix and the person-job matching score to obtain an interpretable matching diagnostic report. Through the above processing, the numerical matching results can be transformed into natural language diagnostic text with node-level attribution evidence, providing fine-grained interpretability support for talent allocation decisions.

[0066] More specifically, in a concrete example of this application, firstly, based on a preset interpretation threshold, all elements in the transmission path matrix are traversed. Each element value is compared with the interpretation threshold, and the row and column indices of all matrix elements strictly greater than the threshold are extracted. The row indices correspond to job requirement nodes, and the column indices correspond to candidate skill nodes. The matching indices are then paired to construct a core resume traceability pair set. This filtering operation removes weak and meaningless noisy connections, retaining only strong mapping relationships with sufficient transmission strength to support attribution explanation. Subsequently, based on the original text dictionary mapping table, each numerical index in the core resume traceability pair set is reverse-engineered into natural language vocabulary. The restored natural language vocabulary pairs, along with the job-person matching score, are injected into a preset interpretation template. After grammatical correction and word order polishing by a natural language generation engine, an interpretable matching diagnostic report is finally generated.

[0067] In a preferred embodiment of this application, the natural language generation engine adopts a combination of template filling and rule post-processing. The preset explanation templates include three types: overall matching score template, core matching basis template, and capability gap prompt template. After filling the natural language words in the core resume tracing set into the corresponding template slots, the subject-verb-object word order is adjusted and conjunctions are added through a preset grammar rule library to generate diagnostic text that conforms to the natural language expression habits.

[0068] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. An intelligent person-job matching method based on job requirements analysis, characterized in that, include: S1, through a pre-trained language model, performs semantic parsing and hierarchical knowledge extraction on the acquired job requirement data to obtain a job requirement tree; S2, extract skill entities and dynamically measure proficiency from the acquired candidate resume data to obtain a dynamic skill evolution tree; S3, feature encoding and cross-granularity structure penalty distance quantization are performed on all nodes in the job requirement tree and dynamic skill evolution tree to obtain the asymmetric cost matrix; S4 uses the asymmetric cost matrix as the cost function constraint to perform hierarchical optimal transmission solution and pooling dimensionality reduction on the node probability distribution of the job requirement tree and the dynamic skill evolution tree to obtain the bulldozer distance scalar, transmission path matrix and topology alignment feature vector. S5, perform a comprehensive matching degree estimation on the bulldozer distance scalar and topological alignment feature vector to obtain the human-position matching degree score; S6. Perform node mapping and source tracing analysis on the transmission path matrix and the person-job matching score to obtain an interpretable matching diagnostic report.

2. The intelligent person-job matching method based on job requirement analysis according to claim 1, characterized in that, Job requirement data includes job description text, qualification requirements text, skill tag set and business scenario identifier; candidate resume data includes project experience description text sequence and work time interval sequence.

3. The intelligent person-job matching method based on job requirement analysis according to claim 2, characterized in that, Step S1 includes: By using the encoder of a pre-trained language model, semantic parsing and entity span probability filtering are performed on the job description text and the qualification requirement text to obtain a set of macro-level capability dimension nodes. Using the set of nodes in the macro capability dimension as contextual constraints, a pre-trained language model is used to extract relationships from the skill tag set with zero-sample drill-down to obtain the set of nodes in the underlying specific technology stack. Based on business scenario identification, a topology mapping and mask pruning topology connection are performed on the macro capability dimension node set and the underlying specific technology stack node set to obtain the job requirement tree.

4. The intelligent person-job matching method based on job requirement analysis according to claim 1, characterized in that, Step S2 includes: Skill entities are extracted from the text sequence describing project experience using a named entity recognition model to obtain a set of candidate skill entity nodes. The logarithmic growth proficiency of each node is calculated based on the work time interval sequence to obtain a set of initial weight scalars for the nodes. Temporal decay and reuse transition weighting are applied to the candidate skill entity node set and the initial node weight scalar set to obtain the dynamic skill current weight scalar set. Based on the technology ontology knowledge base, the hierarchical relationship between nodes in the candidate skill entity node set is analyzed and directed connection edges are constructed. The current weight scalar set of dynamic skills is assigned to the corresponding nodes to obtain the dynamic skill evolution tree.

5. The intelligent person-job matching method based on job requirement analysis according to claim 1, characterized in that, Step S3 includes: A graph attention network is used to perform multi-hop neighborhood aggregation encoding on all nodes in the job requirement tree and the dynamic skill evolution tree to obtain the set of feature vectors of job nodes and the set of feature vectors of candidate skill nodes. Calculate the cross-granularity semantic distance matrix between the set of feature vectors for job nodes and the set of feature vectors for candidate skill nodes; Based on the cross-granularity semantic distance matrix, a penalty correction is applied to the job requirement tree and the dynamic skill evolution tree to obtain an asymmetric cost matrix.

6. The intelligent person-job matching method based on job requirement analysis according to claim 1, characterized in that, Step S4 includes: Probabilistic quality mapping is performed on the weight values ​​of each node in the job requirement tree and the dynamic skill evolution tree to obtain the target measure probability distribution vector and the source measure probability distribution vector. Using the asymmetric cost matrix as the cost function and introducing an information entropy regularization term, the target measure probability distribution vector and the source measure probability distribution vector are alternately scaled and iteratively solved to obtain the bulldozer distance scalar and the transmission path matrix. Max pooling is performed row by row along the dimension of job requirement nodes in the transmission path matrix. The maximum transmission value of each job requirement node among all candidate skill nodes is extracted and concatenated to reduce the dimensionality and obtain the topology-aligned feature vector.

7. The intelligent person-job matching method based on job requirement analysis according to claim 1, characterized in that, Step S5 includes: performing feature splicing and scoring calculation on the bulldozer distance scalar and topology alignment feature vector to obtain the human-job matching score.

8. The intelligent person-job matching method based on job requirement analysis according to claim 1, characterized in that, Step S6 includes: Based on a preset interpretation and judgment threshold, all elements in the transmission path matrix are filtered, and the job node and skill node indexes corresponding to non-zero weight connections exceeding the threshold are paired to obtain a core resume traceability pair set. Using a natural language generation engine, the core resume traceability indexes in the set are reversed and converted into natural language vocabulary. This vocabulary, along with the person-job matching score, is then injected into a preset explanation template for polishing to generate an interpretable matching diagnostic report.

9. The intelligent person-job matching method based on job requirement analysis according to claim 3, characterized in that, Based on business scenario identifiers, a topology mapping and masked pruning topology connection are performed on the macro-capability dimension node set and the underlying specific technology stack node set to obtain the job requirement tree, including: Based on business scenario identifiers and industry knowledge graphs, mask pruning and branch attribution determination are performed on the macro capability dimension node set and the underlying specific technology stack node set to obtain the pruned adjacency matrix and branch attribution label vector. Based on the branch affiliation label vector, cross-branch node pairs in the set of underlying specific technology stack nodes are selected, and the co-occurrence coordination strength of the cross-branch node pairs is quantified to obtain the coordination coupling strength matrix. Based on the collaborative enhancement coefficients of each technology stack node calculated by the collaborative coupling strength matrix, the corresponding edge weights in the pruned adjacency matrix are enhanced and injected with hierarchical topology to obtain the job requirement tree.

Images