Resume and post intelligent matching method and device based on knowledge graph and storage medium

Abstract

This invention relates to the field of job matching technology, and in particular to a knowledge graph-based intelligent resume and job matching method, apparatus, and storage medium. The method includes acquiring resume text data and job requirement text data to be matched, and performing text preprocessing on both. The advantages of this invention are: hierarchical and transferable skill relationships are explicitly expressed; even if there are differences in wording between the resume and the job description, high-quality talent can be effectively recalled through multi-hop association paths between entities, avoiding the missed detection problem caused by word overlap calculation in traditional keyword matching; simultaneously, the mutual exclusion and pre-dependency relationships between entities provide a graph structure foundation for the linkage verification of multi-dimensional constraints, enabling hard constraints such as education level, years of experience, and mandatory qualifications to be uniformly modeled and jointly determined in the graph structure, thus overcoming the shortcomings of existing technologies in the linkage verification capability of multiple constraints.

Description

Technical Field

[0001] This invention relates to the field of job matching technology, and in particular to a knowledge graph-based intelligent resume and job matching method, apparatus, and storage medium. Background Technology

[0002] With the widespread adoption of digital recruitment models, massive amounts of resume texts and job description texts need to be quickly and automatically matched and screened. Traditional manual comparison methods are inefficient and prone to subjective interference. Existing recruitment systems generally use three types of technologies to achieve person-job matching: keyword retrieval, deep learning semantic vector matching, or simple industry dictionary-assisted matching. However, all of these have significant drawbacks. Traditional keyword matching relies solely on word overlap to calculate matching scores, failing to identify synonyms, hierarchical relationships, and transferable skill connections between skills, majors, and job positions. It is highly susceptible to missing excellent candidates due to differences in text wording. Furthermore, it cannot distinguish key weighted information such as skill mastery depth and hard job requirements, is easily influenced by job seekers' keyword stuffing behavior, lacks the ability to verify multi-dimensional constraints, and results in distorted screening results with a large workload for review. While pure semantic vector matching based on pre-trained models alleviates the problem of word ambiguity, the model only outputs a single similarity value, and the matching reasoning process is completely uninterpretable, making it difficult to meet the needs of enterprise recruitment. For recruitment tracing and verification needs, this type of model heavily relies on a large amount of labeled data. When faced with newly added job positions and lagging skill terminology in emerging industries, it can only complete shallow text similarity comparisons and cannot rely on entity associations to carry out multi-layer logical reasoning. Furthermore, in large-scale resume batch matching scenarios, the computational cost is high and the response latency is high, resulting in high implementation costs. A few matching schemes that use flat industry dictionaries can only achieve simple word mapping and do not have a hierarchical and networked knowledge association structure. They cannot connect multi-dimensional information such as industry, projects, certificates, and work experience. Dictionary updates rely entirely on manual operation, resulting in limited knowledge coverage and low iteration efficiency. Existing matching technologies lack a standardized and structured knowledge system for the talent domain. They fail to uniformly model the entities and complex relationships between resumes and job positions, such as professions, skills, qualifications, and industries. These technologies generally suffer from insufficient matching accuracy, lack of interpretability of results, weak ability to handle multiple constraints, slow adaptation to industry knowledge updates, and low efficiency in retrieving massive amounts of data. The industry urgently needs a technical solution that relies on knowledge graphs to build a talent domain association network, performs graph-based transformation of resume and job information, and achieves fine-grained, interpretable, weighted intelligent matching through graph relationship reasoning. This solution should be accompanied by hardware processing devices capable of running the matching method and computer-readable storage media to store the corresponding program, thereby addressing the various technical shortcomings of current intelligent matching of people and jobs. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a knowledge graph-based intelligent matching method, device and storage medium for resumes and job positions, which effectively solves the deficiencies of the prior art.

[0004] To achieve the above objectives, this invention provides a knowledge graph-based intelligent matching method for resumes and job postings, comprising the following steps: S1. Obtain the resume text data and job requirement text data to be matched, and perform text preprocessing on the resume text data and job requirement text data respectively. The text preprocessing includes at least word segmentation, part-of-speech tagging, stop word filtering and preliminary named entity recognition. S2. Based on the pre-built talent domain knowledge graph, perform entity recognition and linking on the pre-processed resume text, extract the resume entity set, the resume entity set includes at least the education entity, major entity, skill entity, project experience entity, certificate and qualification entity and work experience entity, and assign corresponding entity type label and initial confidence to each entity; S3. Based on the talent domain knowledge graph, perform entity recognition and linking on the preprocessed job requirement text, and extract the job entity set. The job entity set includes at least the job name entity, required skills entity, education requirement entity, major requirement entity, experience requirement entity, and qualification requirement entity. S4. Using each entity in the resume entity set and the job entity set as nodes and the semantic relationships between entities as edges, construct a resume knowledge subgraph and a job knowledge subgraph respectively. The semantic relationships include at least hierarchical relationships, synonym relationships, mutual exclusion relationships, prerequisite dependency relationships and transferable skill relationships. S5. Perform graph embedding representation learning on the resume knowledge subgraph and the job knowledge subgraph respectively to generate resume graph embedding vector and job graph embedding vector. The graph embedding representation learning retains at least the first-order neighborhood structure information and the second-order semantic association information of the nodes. S6. Calculate the global semantic similarity between the resume image embedding vector and the job image embedding vector, and use it as the first matching score; S7. Based on the predefined entity weight system and constraint rule system in the talent domain knowledge graph, perform fine-grained weighted matching on a per-entity basis on the resume entity set and the job entity set to generate a second matching score. The entity weight system includes at least the skill mastery depth weight, hard threshold weight and industry scarcity weight, and the constraint rule system includes at least the hard constraint of education level, the constraint of years of experience and the mandatory constraint of qualifications. S8. Based on the resume knowledge subgraph and the job knowledge subgraph, perform graph path reasoning and multi-hop relationship reasoning to identify the implicit association path between the resume entity and the job entity, and generate a third matching score according to the number of hops of the implicit association path, the type of relationship on the path and the path confidence. S9. Input the first matching score, the second matching score and the third matching score into a preset multi-dimensional fusion matching model. After weighted fusion and nonlinear comprehensive judgment, output the final comprehensive matching score and generate an interpretable traceability path for the matching result. The interpretable traceability path includes at least the contribution of each dimension of the matching score and the corresponding relationship chain of key matching entities. S10. Based on the final comprehensive matching score and the preset matching threshold, output the matching conclusion between the job and the resume. The matching conclusion includes at least three states: successful matching, pending manual review, and failed matching.

[0005] Preferably, in step S2, entity recognition and linking of the preprocessed resume text based on a pre-constructed talent domain knowledge graph includes the following steps: 1) Input the preprocessed resume text into a preset entity recognition model, which is built based on the BERT-BiLSTM-CRF architecture and outputs candidate entities in the text and their entity types; 2) For each candidate entity, calculate the semantic similarity between the candidate entity and each entity in the talent domain knowledge graph, and select the graph entity with the highest semantic similarity as the link target of the candidate entity; 3) When the semantic similarity between a candidate entity and a graph entity is lower than the preset link threshold, the candidate entity is marked as an unlinked entity, and the new entity entry process is triggered.

[0006] Preferably, in any of the above schemes, the construction of the resume knowledge subgraph and the job knowledge subgraph in step S4 includes the following steps: 1) Using each entity in the resume entity set as a node, establish a first type of edge between each node of the resume entity based on the pre-stored inter-entity relationship in the talent domain knowledge graph; 2) Using each entity in the set of job entities as a node, and based on the pre-stored inter-entity relationships in the talent domain knowledge graph, establish a second type of edge between the nodes of the job entities; 3) Link the resume knowledge subgraph and the job knowledge subgraph through shared entity nodes to form a unified matching association graph structure.

[0007] Preferably, in step S5, graph embedding representation learning is performed on the resume knowledge subgraph and the job knowledge subgraph respectively, using a graph embedding model based on a graph attention network. When aggregating the features of neighbor nodes, the graph attention network assigns different attention weights to different neighbor nodes, and the attention weights are calculated based on node type and relationship type.

[0008] Preferably, in any of the above schemes, step S7 involves performing fine-grained weighted matching on an entity-by-entity basis based on the predefined entity weight system and constraint rule system in the talent domain knowledge graph. This includes the following steps: 1) For each pair of entities of the same type in the resume entity set and the job entity set, calculate the entity similarity score; 2) Call the corresponding weight function based on the entity type to calculate the matching contribution weight of the entity; 3) When the job requirements corresponding to the entity are hard constraints, check whether the resume entity meets the hard constraints. If it does not meet the constraints, the final matching result will be determined as unsuccessful. 4) Aggregate the weighted similarity scores of all pairs of entities of the same type to generate dimension matching scores for each dimension. Then, sum the dimension matching scores according to the preset dimension weights to generate the second matching score.

[0009] Preferably, in any of the above schemes, step S8 involves performing graph path reasoning and multi-hop relationship reasoning, which includes the following steps: 1) Starting from the resume entity node in the resume knowledge subgraph and ending from the job entity node in the job knowledge subgraph, perform a multi-hop path search in the matching association graph structure. 2) For each path found, calculate the confidence score of the path, which is calculated based on the weight product of each relation edge in the path and the decay factor of the path hop count; 3) Select the top N paths with the highest confidence scores as key implicit association paths, and aggregate the confidence scores of the key implicit association paths to generate a third matching score.

[0010] Preferably, in any of the above schemes, the multi-dimensional fusion matching model in step S9 adopts a multilayer perceptron architecture. The input layer of the multilayer perceptron receives the first matching score, the second matching score, and the third matching score. The hidden layer uses the ReLU activation function for nonlinear transformation, and the output layer uses the Sigmoid activation function to output the final comprehensive matching score in the range of 0 to 1.

[0011] Preferably, as described in any of the above schemes, the method for constructing the talent domain knowledge graph includes: 1) Collect multi-source talent field data, which includes at least job posting data from recruitment websites, professional encyclopedia data, industry skill standard data, and professional curriculum system data; 2) Perform structured processing on the multi-source talent domain data, extract talent domain entities and relational triples between entities. The entities include at least majors, skills, positions, industries, certificates, and institutions. The relations include at least inclusion relations, membership relations, prerequisite dependency relations, and capability mapping relations. 3) Store the entity and relation triples into a graph database to construct a knowledge graph ontology model for the talent domain; 4) Through an incremental learning mechanism, new entities and relationships are periodically extracted from newly collected data to dynamically update the talent domain knowledge graph.

[0012] To achieve the above objectives, the present invention provides a knowledge graph-based intelligent resume and job matching device for implementing the above method, comprising: The data acquisition module is used to acquire resume text data and job requirement text data to be matched; The text preprocessing module is used to perform text preprocessing on the resume text data and the job requirement text data respectively. The text preprocessing includes at least word segmentation, part-of-speech tagging, stop word filtering and preliminary named entity recognition. The resume entity extraction module is used to perform entity recognition and linking on the pre-processed resume text based on a pre-built talent domain knowledge graph, and extract a resume entity set. The resume entity set includes at least education entity, major entity, skill entity, project experience entity, certificate and qualification entity, and work experience entity. The job entity extraction module is used to perform entity recognition and linking on the preprocessed job requirement text based on the talent domain knowledge graph, and extract a set of job entities. The set of job entities includes at least job name entity, required skills entity, education requirement entity, professional requirement entity, experience requirement entity, and qualification requirement entity. The knowledge subgraph construction module is used to construct a resume knowledge subgraph and a job knowledge subgraph respectively, using each entity in the resume entity set and the job entity set as nodes and the semantic relationship between entities as edges. The graph embedding representation module is used to perform graph embedding representation learning on the resume knowledge subgraph and the job knowledge subgraph respectively, and generate resume graph embedding vector and job graph embedding vector. The global semantic matching module is used to calculate the global semantic similarity between the resume image embedding vector and the job image embedding vector, and output the first matching score. The fine-grained weighted matching module is used to perform fine-grained weighted matching on an entity-by-entity basis on the resume entity set and the job entity set based on the predefined entity weight system and constraint rule system in the talent domain knowledge graph, and output a second matching score. The graph reasoning and matching module is used to perform graph path reasoning and multi-hop relationship reasoning based on the resume knowledge subgraph and the job knowledge subgraph, identify the implicit association path between resume entities and job entities, and output a third matching score. The multi-dimensional fusion output module is used to input the first matching score, the second matching score, and the third matching score into a preset multi-dimensional fusion matching model, and output the final comprehensive matching score and matching conclusion after weighted fusion and nonlinear comprehensive judgment.

[0013] To achieve the above objectives, the present invention provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of a knowledge graph-based intelligent matching method for resumes and job positions. The storage medium includes any one of ROM, RAM, magnetic disk, and optical disk.

[0014] The present invention has the following advantages: 1. The knowledge graph-based intelligent resume and job matching method device and storage medium, through steps S2 to S4, perform entity recognition and linking on resume text and job requirement text based on a pre-constructed talent domain knowledge graph, respectively, extracting resume entity sets and job entity sets containing multiple types of entities such as education, major, skills, project experience, certificates and qualifications, and work experience. Then, using each entity as a node and the hierarchical relationships, synonym relationships, mutual exclusion relationships, prerequisite dependencies, and transferable skill associations between entities as edges, resume knowledge subgraphs and job knowledge subgraphs are constructed respectively, so that the various dimensions of information in the resume and job are converted into structured graph data. Under this structure, synonyms, hierarchical relationships, and transferable skill associations among skills, professions, and positions can be explicitly expressed. Even if there are differences in the wording between the resume and the position, high-quality talents can be effectively recalled through multi-hop association paths between entities, avoiding the missed detection problem caused by word overlap calculation in traditional keyword matching. At the same time, the mutual exclusion and pre-dependency relationships between entities provide a graph structure foundation for the linkage verification of multi-dimensional constraints, enabling hard constraints such as education level, years of experience, and mandatory qualifications to be uniformly modeled and jointly judged in the graph structure, making up for the shortcomings of existing technologies in the linkage verification capability of multiple constraints.

[0015] 2. The knowledge graph-based intelligent resume and job matching method device and storage medium, through steps S5 to S9, perform graph embedding representation learning on the resume knowledge subgraph and the job knowledge subgraph respectively to generate graph embedding vectors and calculate global semantic similarity as the first matching score. At the same time, based on the predefined entity weight system and constraint rule system in the talent domain knowledge graph, fine-grained weighted matching is performed on an entity-by-entity basis to generate the second matching score. Then, graph path reasoning and multi-hop relationship reasoning are performed to generate the third matching score. Finally, the three matching scores are input into a multi-dimensional fusion matching model with a multilayer perceptron architecture. After weighted fusion and nonlinear comprehensive judgment, the final comprehensive matching score is output and a solvable result is generated. The interpretable tracing path includes at least the contribution of each dimension of the matching score and the corresponding relationship chain of key matching entities. This makes the matching result no longer an uninterpretable single similarity value, but a transparent decision-making process that can be traced back to specific entity matching pairs, graph path reasoning results, and the contribution of scores in each dimension, thus meeting the needs of enterprises for tracing and verifying recruitment results. At the same time, graph embedding representation learning and graph path reasoning are both based on graph structures. Their computational complexity mainly depends on the scale of graph nodes and edges rather than the length of resume text. In large-scale resume batch matching scenarios, it has lower computational overhead and response latency than pure semantic vector matching models, reducing the cost of implementation.

[0016] 3. The knowledge graph-based intelligent resume and job matching method device and storage medium, through step S5, adopts a graph embedding model based on graph attention network, which assigns different attention weights to different neighbor nodes when aggregating neighbor node features. The attention weights are calculated based on node type and relation type, so that the graph embedding vector can adaptively adjust the emphasis of information aggregation according to entity type and relation type during the generation process. Compared with the traditional graph embedding method that aggregates all neighbor nodes equally, the graph attention mechanism of this invention can more accurately capture the semantic information of entities that play a key role in job matching, improve the semantic representation quality of graph embedding vector, and thus improve the accuracy of global semantic similarity calculation. In addition, in step S7, the corresponding weight function is called to calculate the matching contribution weight according to the entity type, and the final matching result is directly judged as unsuccessful when the hard constraints are not met, so that the differentiated importance of different entity types in matching can be quantified, avoiding the problem of insufficient matching accuracy caused by treating all words equally in shallow text similarity matching.

[0017] 4. The knowledge graph-based intelligent resume and job matching method device and storage medium, in step S8, performs multi-hop path search in the matching association graph structure, starting from resume entity nodes in the resume knowledge subgraph and ending with job entity nodes in the job knowledge subgraph. For each searched path, a confidence score is calculated based on the weight product of each relation edge in the path and the attenuation factor of the path hop count. The top N paths with the highest confidence scores are selected as key implicit association paths. A third matching score is generated by aggregating the confidence scores of these paths, ensuring that resumes and job titles do not directly exist in the same text. Implicit relationships can be explicitly inferred through multi-hop paths in the knowledge graph. For example, a job seeker's project experience entity can be associated with the target position through multi-hop paths of projects, skills, and positions. Or, a work experience entity can achieve cross-domain matching of transferable skills through the path of "industry-ability mapping-position". This mechanism breaks through the limitation of pure text similarity comparison, which can only process explicit co-occurrence information. It enables matching reasoning to carry out multi-level logical reasoning based on entity association, making up for the technical shortcomings of existing technologies that cannot connect multi-dimensional information such as industry, projects, certificates, and work experience for comprehensive reasoning.

[0018] 5. The knowledge graph-based intelligent resume and job matching method device and storage medium collect multi-source talent domain data through a talent domain knowledge graph construction method. The multi-source data undergoes structured processing to extract entities including at least majors, skills, positions, industries, certificates, and institutions, as well as relational triples such as inclusion relationships, membership relationships, prerequisite dependencies, and ability mapping relationships. These are stored in a graph database to construct a talent domain knowledge graph ontology model. An incremental learning mechanism is used to periodically extract new entities and relationships from newly collected data to dynamically update the graph. This gives the talent domain knowledge graph a hierarchical and networked knowledge association structure, enabling it to continuously adapt to newly added job and skill terms in emerging industries. Compared to the flat industry dictionary update method that relies on manual operation, the knowledge graph of this invention can automatically complete knowledge iteration through an incremental learning mechanism, significantly improving the knowledge coverage and update efficiency, thereby ensuring the timeliness and adaptability of the matching method in the face of rapidly changing industry talent needs. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the system framework of the present invention. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0021] like Figures 1 to 2As shown, the present invention provides a knowledge graph-based intelligent resume and job matching method, apparatus and storage medium.

[0022] Embodiment 1: A knowledge graph-based intelligent resume and job matching method This embodiment provides a knowledge graph-based intelligent resume and job matching method. By constructing a knowledge graph in the talent domain, the method converts resume texts and job requirement texts into graph representations, and realizes interpretable and high-precision intelligent job-candidate matching through a multi-dimensional integration strategy of graph embedding representation learning, fine-grained weighted matching and graph path reasoning. The method specifically includes the following steps S1 to S10.

[0023] S1. Data acquisition and text preprocessing To-be-matched resume text data and job requirement text data are acquired. The resume text data comes from resume documents uploaded by job seekers on the recruitment platform or resume information filled in online, and the job requirement text data comes from job description documents published by enterprise users on the recruitment platform or job requirement information filled in online. The acquisition methods include but are not limited to: reading through a platform data interface, importing from a file storage system, or directly inputting through a user interface.

[0024] Text preprocessing is performed on the acquired resume text data and job requirement text data respectively. The specific process of text preprocessing is as follows: first, a dictionary and statistics combined word segmentation algorithm is used to segment the text, so as to split the continuous text sequence into independent word units; second, part-of-speech tagging is performed on the word segmentation results to identify the part-of-speech category of each word, such as noun, verb, adjective, etc., to provide lexical features for subsequent entity recognition; third, a preset stop word list is loaded to filter and remove stop words with no actual semantic contribution in the text (such as function words like "de", "le", "zai" and common words with no practical meaning); finally, preliminary named entity recognition is performed based on a predefined entity type template, to preliminarily mark entity fragments in the text that may belong to categories such as person name, organization name, place name, time, etc., so as to provide candidate entity boundary information for subsequent accurate knowledge graph-based entity recognition and linking.

[0025] S2. Entity recognition and linking of resume texts Based on the pre-constructed talent domain knowledge graph, entity recognition and linking are performed on the preprocessed resume text, and a resume entity set is extracted. This step specifically includes the following sub-steps.

[0026] The first step involves inputting the preprocessed resume text into a predefined entity recognition model. This model is built on a BERT-BiLSTM-CRF architecture. The BERT (Bidirectional Encoder Representations from Transformers) layer converts each word in the input text into a vector representation that incorporates contextual semantic information. The BiLSTM (Bidirectional Long Short-Term Memory) layer captures long-distance dependencies in the text sequence and outputs the probability distribution of each word belonging to each entity category. The CRF (Conditional Random Field) layer performs global optimization on the entire label sequence, ensuring that the output entity label sequence conforms to predefined label transition constraints (e.g., the "I-Skill" label cannot appear before the "B-Skill" label). The entity recognition model outputs candidate entities in the text and their entity types, which include at least education level, major, skills, project experience, certifications, and work experience.

[0027] The second step involves calculating the semantic similarity between each candidate entity and all entities in the talent domain knowledge graph. The semantic similarity is calculated by inputting the text representation of the candidate entity and the name and alias texts of the entities in the knowledge graph into a pre-defined semantic matching model. This model is built on the Sentence-BERT architecture and outputs a cosine similarity score between the two. The graph entity with the highest semantic similarity is selected as the link target for the candidate entity, which means finding the corresponding standard entity node for that candidate entity in the knowledge graph.

[0028] The third step involves identifying entities whose semantic similarity to graph entities is below a preset linking threshold (e.g., 0.75). This indicates that no standard entity in the knowledge graph fully matches the semantics of the candidate entity. The candidate entity is then marked as an unlinked entity, triggering the new entity entry process. This process includes: storing the unlinked entity and its context information in a pending review queue; after confirmation by the system administrator or an automated review mechanism; writing the entity as a new node into the talent domain knowledge graph; and automatically inferring potential association edges for the new entity based on existing entity relationships in the knowledge graph.

[0029] Through the above steps, a set of resume entities is extracted. This set includes at least the following entities: education level (e.g., "Master's degree"), major (e.g., "Computer Science and Technology"), skills (e.g., "Python programming"), project experience (e.g., "Development of a financial risk control system"), certification (e.g., "PMP project management certification"), and work experience (e.g., "5 years of software development experience"). Each entity is assigned a corresponding entity type label and an initial confidence level. The initial confidence level is calculated based on the output probability of the entity recognition model and the link similarity.

[0030] S3. Entity recognition and linking in job postings Based on a talent domain knowledge graph, entity recognition and linking are performed on the preprocessed job requirement text to extract a set of job entities. The specific implementation of this step is the same as step S2, the difference being that the extracted entity types are oriented towards the job requirement dimension. The set of job entities includes at least the job title entity (e.g., "Senior Java Development Engineer"), required skills entity (e.g., "SpringBoot"), educational requirement entity (e.g., "Full-time Bachelor's Degree or Above"), major requirement entity (e.g., "Computer-related Major"), experience requirement entity (e.g., "3+ Years of Backend Development Experience"), and qualification requirement entity (e.g., "OCP Certificate Preference Given"). Each entity is also assigned a corresponding entity type label and initial confidence level.

[0031] S4. Construction of Knowledge Subgraphs Using entities in the resume entity set and the job entity set as nodes, and semantic relationships between entities as edges, construct resume knowledge subgraphs and job knowledge subgraphs respectively. Semantic relationships include at least hierarchical relationships (e.g., "programming language" is a hierarchical concept of "Java"), synonymous relationships (e.g., "K8s" and "Kubernetes" are synonyms), mutually exclusive relationships (e.g., "full-time education" and "part-time education" are mutually exclusive), prerequisite dependencies (e.g., "proficiency in Java" is a prerequisite for "learning Spring Boot"), and transferable skill relationships (e.g., "C++ programming skills" can be transferred to "C# development"). This step specifically includes the following sub-steps.

[0032] The first step is to use each entity in the resume entity set as a node, and establish a first-type edge between the nodes of the resume entities based on the pre-existing inter-entity relationships in the talent domain knowledge graph. For example, if the resume entity set contains both "Java" skill entities and "SpringBoot" skill entities, and the knowledge graph pre-exists a "prerequisite dependency" relationship between "Java" and "SpringBoot", then an edge labeled "prerequisite dependency" is established between the "Java" node and the "SpringBoot" node in the resume knowledge subgraph.

[0033] The second step involves using each entity in the job entity set as a node, and establishing a second type of edge between the nodes of the job entity based on the pre-existing entity relationships in the talent domain knowledge graph. For example, if the job entity set contains both the "Computer Science and Technology" major entity and the "Java" skill entity, and the knowledge graph pre-exists an "inclusion relationship" between "Computer Science and Technology" and "Java" (i.e., the major includes Java courses), then an edge labeled "inclusion" is established between the "Computer Science and Technology" node and the "Java" node in the job knowledge subgraph.

[0034] The third step involves linking the resume knowledge subgraph and the job knowledge subgraph through shared entity nodes, forming a unified matching association graph structure. A shared entity node refers to the same entity appearing simultaneously in both the resume entity set and the job entity set (e.g., the "Java" skill in the resume and the "Java" skill in the job requirements both point to the same "Java" entity node in the knowledge graph). This shared node connects the two subgraphs into a unified matching association graph structure, providing a foundation for cross-subgraph path search in subsequent graph path reasoning.

[0035] S5, Graph Embedding Representation Learning Graph embedding representation learning is performed on the resume knowledge subgraph and the job knowledge subgraph respectively to generate resume graph embedding vectors and job graph embedding vectors. The graph embedding representation learning adopts a graph embedding model based on Graph Attention Network (GAT). When aggregating the features of neighbor nodes, the GAT assigns different attention weights to different neighbor nodes. These attention weights are calculated based on node type and relationship type.

[0036] Specifically, for each target node in the knowledge subgraph, the graph attention network first calculates the attention coefficient between that node and each of its neighboring nodes. The attention coefficient is calculated by concatenating the feature vector of the target node, the feature vectors of its neighboring nodes, and the type encoding vectors of the edges between them. This concatenation is then input into a single-layer feedforward neural network, processed by the LeakyReLU activation function to obtain an unnormalized attention score. Next, the attention scores of all neighboring nodes of the target node are normalized using Softmax to obtain a normalized attention weight. This attention weight considers both the entity type of the neighboring nodes (e.g., the different semantic roles of skill nodes compared to education nodes in matching) and the relation type (e.g., the different roles of hierarchical relations compared to synonymous relations in semantic transmission), enabling the model to adaptively emphasize the semantic information of entities that play a crucial role in job matching during the aggregation process.

[0037] After obtaining the attention weights of each neighboring node, the graph attention network performs a weighted summation of the feature vectors of the neighboring nodes, and then uses the sum after a nonlinear transformation as the new feature vector of the target node. Through multiple rounds of iterative aggregation of all nodes in the knowledge subgraph, the feature vector of each node gradually incorporates the structural and semantic information of its multi-hop neighborhood. Finally, the feature vectors of all nodes are pooled (e.g., average pooling or max pooling) to obtain the graph embedding vector of the entire knowledge subgraph. This graph embedding represents the learning to retain at least the first-order neighborhood structural information (i.e., features of directly adjacent nodes) and the second-order semantic association information (i.e., semantic associations transmitted through two-hop paths).

[0038] S6. Global semantic similarity calculation The global semantic similarity between the resume graph embedding vector and the job graph embedding vector is calculated and used as the first matching score. The global semantic similarity is calculated by using the cosine similarity between the two vectors. The cosine similarity ranges from [-1, 1], and to facilitate subsequent multi-dimensional fusion, it can be mapped to the [0, 1] interval. This first matching score reflects the degree of similarity between the resume and the job at the overall semantic level, and is a preliminary measure of their matching degree from a global graph structure perspective.

[0039] S7, Entity-by-Entity Fine-Grained Weighted Matching Based on the predefined entity weighting and constraint rule system in the talent domain knowledge graph, a fine-grained weighted matching of the resume entity set and the job entity set is performed to generate a second matching score. The entity weighting system includes at least the skill mastery depth weight (reflecting the job seeker's mastery of a certain skill, such as "proficient" having a higher weight than "familiar"), hard threshold weight (reflecting whether a certain entity requirement is a hard entry condition for the job), and industry scarcity weight (reflecting the scarcity of a certain skill or qualification in the current industry; the higher the scarcity, the greater the matching contribution weight). The constraint rule system includes at least the hard educational constraints (e.g., if the job requires "full-time master's degree or above," then a part-time master's degree is not sufficient), the experience years constraints (e.g., if the job requires "5 years or more," then 4 years of experience is not sufficient), and the mandatory qualification constraints (e.g., if the job requires "holding a lawyer's license," then those without a license are not sufficient). This step specifically includes the following sub-steps.

[0040] The first step is to calculate the entity similarity score for each pair of entities of the same type in the resume entity set and the job entity set. The entity similarity score is calculated as follows: For skill-related entities, semantic similarity is calculated based on hierarchical and synonym relationships in the knowledge graph (e.g., "Java" and "J2EE" have a high similarity score based on hierarchical relationships); for education-related entities, hierarchical similarity is calculated based on the education level system (e.g., "Master's" and "Doctorate" have a higher similarity score than "Master's" and "Bachelor's"); for experience-related entities, difference similarity is calculated based on the number of years of experience.

[0041] The second step involves calling the corresponding weighting function based on the entity type to calculate the matching contribution weight for that entity. The weighting function is a predefined mapping function that sets different base weight coefficients for different entity types (such as skills, education, major, qualifications, etc.) and combines them with dynamic adjustment factors such as skill mastery depth weight and industry scarcity weight to calculate the final matching contribution weight. For example, for a skills entity, its matching contribution weight = base weight × skill mastery depth coefficient × industry scarcity coefficient.

[0042] The third step involves checking whether the resume entity meets the hard constraint when the job requirement corresponding to the entity is a hard constraint. If it does not, the final matching result is directly determined as unsuccessful. For example, if the constraint type of the "education requirement entity" in the job requirement is marked as "hard," and the "education entity" in the resume does not meet the requirement after hierarchical comparison (e.g., the job requires "Master's degree" while the resume shows "Bachelor's degree"), then the final result of the entire matching process is directly determined as unsuccessful, and subsequent calculations are terminated.

[0043] The fourth step involves aggregating the weighted similarity scores of all entity pairs of the same type to generate dimension matching scores for each dimension (such as skills, education, profession, qualifications, etc.). These dimension matching scores are then weighted and summed according to preset dimension weights to generate a second matching score. Dimension weights can be dynamically adjusted based on different job types; for example, the weight of the skills dimension can be increased for technical positions, while the weight of the experience dimension can be increased for management positions.

[0044] S8. Graph Path Reasoning and Multi-hop Relationship Reasoning Based on the resume knowledge subgraph and the job knowledge subgraph, graph path reasoning and multi-hop relationship reasoning are performed to identify the implicit association paths between resume entities and job entities. A third matching score is generated based on the number of hops in the implicit association path, the type of relationship on the path, and the path confidence. This step specifically includes the following sub-steps.

[0045] The first step involves performing a multi-hop path search within the matching association graph structure, starting with the resume entity node in the resume knowledge subgraph and ending with the job entity node in the job knowledge subgraph. The multi-hop path search employs either breadth-first search or depth-first search algorithms to find all reachable paths from the resume entity node to the job entity node within the matching association graph structure. The number of hops for each path can be preset to 3 to 5 hops to control search complexity and the semantic relevance of the paths.

[0046] The second step is to calculate the confidence score for each searched path. The confidence score is calculated as the product of the weights of each relation edge in the path multiplied by a hop count decay factor. The weights of the relation edges are predefined based on the relation type (e.g., the weight of "synonymous relation" is higher than that of distant superiors and inferiors in "hypothesis relation"). The hop count decay factor decreases exponentially with the number of hops (e.g., a decay factor of 0.7 raised to the power of hops) to reflect the principle that more hops indicate a more indirect semantic connection.

[0047] The third step involves selecting the top N paths with the highest confidence scores as key implicit connection paths (N can be preset to 5 or 10). A third matching score is generated by aggregating the confidence scores of these key implicit connection paths. The aggregation method involves weighted summation or taking the maximum value of the confidence scores of the selected N paths to obtain the third matching score. This third matching score reflects the degree of implicit matching between the resume and the job posting inferred through multi-hop paths in the knowledge graph, and can capture deep connections between the two that do not directly exist in the same text.

[0048] S9, Multi-dimensional integration and interpretable output The first matching score, the second matching score, and the third matching score are input into a preset multi-dimensional fusion matching model. After weighted fusion and nonlinear comprehensive judgment, the final comprehensive matching score is output, and an interpretable traceability path for the matching result is generated.

[0049] The multi-dimensional fusion matching model employs a Multilayer Perceptron (MLP) architecture. The input layer of the MLP receives three nodes, corresponding to the first, second, and third matching scores, respectively. At least one hidden layer is then placed after the input layer, employing the ReLU activation function for non-linear transformation to capture the non-linear interaction between the three matching scores. The output layer uses the Sigmoid activation function to map the output value to the [0,1] interval, serving as the final comprehensive matching score. The parameters of the MLP model are obtained through supervised training using a pre-labeled person-job matching dataset.

[0050] An interpretable tracing path should include at least the contribution of each dimension of the matching score and the corresponding relationship chain of key matching entities. Specifically, during the forward computation of the multi-dimensional fusion matching model, the contribution weight of each node in the input layer to the output layer result is recorded. Simultaneously, the corresponding relationship chains of the entities with the highest similarity are extracted from the entity-by-entity matching results in step S7, and the implicit association paths with the highest confidence are extracted from the graph path reasoning results in step S8. This information is then integrated into a structured interpretable tracing path, output in visual or text form, for recruiters to verify the reasonableness of the matching results.

[0051] S10, Matching Result Output Based on the final comprehensive matching score and a preset matching threshold, the matching conclusion between the job and the resume is output. The matching conclusion includes at least three states: when the final comprehensive matching score is higher than the first preset threshold (e.g., 0.8), a "match successful" conclusion is output, indicating that the resume and job are highly matched and the candidate can directly proceed to the interview or hiring stage; when the final comprehensive matching score is between the first and second preset thresholds (e.g., 0.5), a "pending manual review" conclusion is output, indicating that the resume and job are partially matched and require further manual review by recruiters; when the final comprehensive matching score is lower than the second preset threshold, a "match failed" conclusion is output, indicating that the resume and job are poorly matched and the candidate can be directly rejected. Simultaneously, the matching conclusion and the interpretable traceability path generated in step S9 are output together to provide recruiters with a basis for decision-making.

[0052] Methods for constructing knowledge graphs in the talent domain The talent domain knowledge graph used in steps S2 and S3 above is constructed using the following steps.

[0053] The first step is to collect multi-source talent field data. Multi-source talent field data should include at least: job posting data from recruitment websites (textual data such as job descriptions and requirements collected from major recruitment platforms), occupational encyclopedia data (such as occupational classification standards and occupational competency standards for various industries), industry skill standard data (such as skill level standards and skill certification systems for various industries), and professional curriculum system data (such as curriculum settings and training programs for various majors in higher education).

[0054] The second step involves structuring the multi-source talent domain data. A combination of rule-based and deep learning-based information extraction methods is used to extract talent domain entities and their relational triples from unstructured text data. Entities include at least categories such as major, skill, position, industry, certificate, and institution. Relationships include at least inclusion relationships (e.g., a major includes a course), membership relationships (e.g., a skill belongs to a skill category), prerequisite dependencies (e.g., learning a skill requires mastering a prerequisite skill), and ability mapping relationships (e.g., a job requirement maps to a specific professional skill).

[0055] The third step is to store the extracted entity and relation triples into a graph database (such as Neo4j) to construct a knowledge graph ontology model for the talent domain. In the ontology model, entities are nodes in the graph database, and relations are edges between nodes. Each node and each edge can be accompanied by attribute information (such as node aliases, relation strength weights, etc.).

[0056] The fourth step involves using an incremental learning mechanism to periodically extract new entities and relationships from newly collected data, dynamically updating the talent domain knowledge graph. This incremental learning mechanism includes: periodically triggering data collection tasks to obtain the latest talent domain data; performing entity recognition and relationship extraction on the new data; performing conflict detection and fusion between the new entities and relationships and existing content in the graph (e.g., disambiguation when entities with the same name but different meanings are detected); and writing the approved new content into the graph database to complete the dynamic update of the knowledge graph.

[0057] Example 2: Knowledge Graph-Based Intelligent Resume and Job Matching Device and Storage Medium This embodiment provides a knowledge graph-based intelligent resume and job matching device to implement the method described in Embodiment 1. The device includes the following modules.

[0058] The data acquisition module is used to acquire resume text data and job requirement text data to be matched. This module obtains raw text data through platform data interfaces, file import interfaces, or user input interfaces, and then transmits the data to the text preprocessing module.

[0059] The text preprocessing module is used to preprocess resume text data and job requirement text data separately. Text preprocessing includes at least word segmentation, part-of-speech tagging, stop word filtering, and preliminary named entity recognition. This module calls preset word segmentation dictionaries, part-of-speech tagging models, stop word lists, and preliminary entity recognition rules to complete the cleaning and structured preprocessing of the original text.

[0060] The resume entity extraction module is used to perform entity recognition and linking on pre-processed resume text based on a pre-built talent domain knowledge graph, extracting a resume entity set. The resume entity set includes at least entities for education level, major, skills, project experience, certifications, and work experience. This module integrates an entity recognition model based on the BERT-BiLSTM-CRF architecture and a semantic similarity calculation model based on Sentence-BERT, enabling the identification of candidate entities and accurate linking to the knowledge graph. When a candidate entity cannot be linked to an existing entity in the knowledge graph, this module triggers a new entity addition process.

[0061] The Job Entity Extraction module, based on a talent domain knowledge graph, performs entity recognition and linking on preprocessed job requirement text to extract a set of job entities. This set includes at least the job title entity, required skills entity, education requirement entity, professional requirement entity, experience requirement entity, and qualification requirement entity. The implementation of this module is the same as the resume entity extraction module, the difference being that it identifies entity types within the job requirement text.

[0062] The knowledge subgraph construction module is used to construct resume knowledge subgraphs and job knowledge subgraphs, respectively, using entities in the resume entity set and job entity set as nodes and semantic relationships between entities as edges. This module reads pre-stored relationships between entities from the talent domain knowledge graph, establishes first-type edges between nodes of the resume entity and second-type edges between nodes of the job entity, and connects the two subgraphs through shared entity nodes to form a unified matching association graph structure.

[0063] The graph embedding representation module is used to learn graph embedding representations for the resume knowledge subgraph and the job knowledge subgraph, respectively, generating resume graph embedding vectors and job graph embedding vectors. This module adopts a graph embedding model based on graph attention networks. When aggregating the features of neighbor nodes, it assigns attention weights to different neighbor nodes based on node type and relationship type, generating graph embedding vectors that integrate first-order neighborhood structure information and second-order semantic association information.

[0064] The global semantic matching module calculates the global semantic similarity between the resume image embedding vector and the job posting image embedding vector, and outputs the first matching score. This module obtains the global semantic similarity score through cosine similarity calculation.

[0065] The fine-grained weighted matching module performs entity-by-entity fine-grained weighted matching on the resume entity set and the job entity set based on a predefined entity weight system and constraint rule system in the talent domain knowledge graph, outputting a second matching score. The entity weight system includes at least skill mastery depth weight, hard threshold weight, and industry scarcity weight, while the constraint rule system includes at least hard constraints on education level, years of experience, and mandatory qualification constraints. This module calculates a similarity score for each pair of entities of the same type, calls the corresponding weight function to calculate the matching contribution weight according to the entity type, and directly determines that the match fails when hard constraints are not met. Finally, it aggregates the weighted similarity scores of all dimensions to generate the second matching score.

[0066] The graph reasoning and matching module performs graph path reasoning and multi-hop relationship reasoning based on the resume knowledge subgraph and the job knowledge subgraph. It identifies implicit relationship paths between resume entities and job entities and outputs a third matching score. This module searches for multi-hop paths starting from resume entity nodes and ending at job entity nodes. It calculates a path confidence score based on the weight product of each relation edge in the path and a decay factor for the number of hops. The module then aggregates the top N paths with the highest confidence scores to generate the third matching score.

[0067] The multi-dimensional fusion output module is used to input the first, second, and third matching scores into a preset multi-dimensional fusion matching model. After weighted fusion and nonlinear comprehensive judgment, it outputs the final comprehensive matching score and matching conclusion. The multi-dimensional fusion matching model adopts a multilayer perceptron architecture. The input layer receives three matching scores, the hidden layer uses the ReLU activation function, and the output layer uses the Sigmoid activation function to output the final comprehensive matching score in the range of 0 to 1. This module also generates an interpretable traceability path for the matching results, including at least the contribution of each dimension of the matching score and the corresponding relationship chain of key matching entities. The matching conclusion includes at least three states: successful matching, pending manual review, and failed matching.

[0068] storage media This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the knowledge graph-based intelligent resume and job matching method described in Embodiment 1. The storage medium includes, but is not limited to, any one of ROM (Read-Only Memory), RAM (Random Access Memory), magnetic disk (such as hard disk, floppy disk), or optical disk (such as CD-ROM, DVD). This storage medium can be installed in a computer, server, or other electronic device with data processing capabilities, whereby a processor loads and executes the computer program stored therein to implement the intelligent job matching method provided by this invention.

[0069] In summary, steps S2 to S4, based on a pre-built talent domain knowledge graph, perform entity recognition and linking on resume text and job requirement text, respectively. This extracts resume entity sets and job entity sets containing various types of entities, including education, major, skills, project experience, certifications, and work experience. Resume knowledge subgraphs and job knowledge subgraphs are constructed using each entity as a node and hierarchical, synonymous, mutually exclusive, pre-dependency, and transferable skill relationships as edges. This transforms the multi-dimensional information in the resume and job requirements into a structured graph data structure. Under this structure, synonyms, hierarchical, and transferable skill relationships between skills, majors, and job requirements are explicitly expressed. Even if there are differences in wording between the resume and job requirements, high-quality talent can be effectively recalled through multi-hop association paths between entities, avoiding the missed detection problem caused by word overlap calculation in traditional keyword matching. Simultaneously, the mutual exclusion and pre-dependency relationships between entities provide a graph structure foundation for the linkage verification of multi-dimensional constraints, enabling the implementation of hard constraints such as education level, years of experience, and qualifications. Quality-driven constraints can be uniformly modeled and jointly determined in the graph structure, making up for the shortcomings of existing technologies in the ability to verify multiple constraints. Through steps S5 to S9, graph embedding representation learning is performed on the resume knowledge subgraph and the job knowledge subgraph to generate graph embedding vectors and calculate global semantic similarity as the first matching score. At the same time, based on the predefined entity weight system and constraint rule system in the talent domain knowledge graph, fine-grained weighted matching is performed on an entity-by-entity basis to generate the second matching score. Then, graph path reasoning and multi-hop relationship reasoning are performed to generate the third matching score. Finally, the three matching scores are input into the multi-dimensional fusion matching model of the multilayer perceptron architecture. After weighted fusion and nonlinear comprehensive determination, the final comprehensive matching score is output and an interpretable traceability path is generated. This interpretable traceability path includes at least the contribution of each dimension of the matching score and the corresponding relationship chain of key matching entities. This makes the matching result no longer an uninterpretable single similarity value, but a transparent decision-making process that can be traced back to specific entity matching pairs, graph path reasoning results, and the contribution of each dimension score, thus meeting the needs of enterprise recruitment traceability and verification.Meanwhile, both graph embedding representation learning and graph path reasoning are based on graph structures. Their computational complexity primarily depends on the scale of graph nodes and edges rather than the resume text length. In large-scale resume batch matching scenarios, they have lower computational overhead and response latency compared to pure semantic vector matching models, reducing deployment costs. Step S5 employs a graph embedding model based on graph attention networks, assigning different attention weights to different neighbor nodes when aggregating their features. These attention weights are calculated based on node type and relation type, allowing the graph embedding vector to adaptively adjust the emphasis on information aggregation according to entity type and relation type during generation. Compared to traditional graph embedding methods that aggregate all neighbor nodes equally, this invention's graph attention mechanism can more accurately capture the semantic information of entities that play a crucial role in job matching, improving the semantic representation quality of the graph embedding vector and thus enhancing the accuracy of global semantic similarity calculation. Furthermore, step S7 calls the corresponding weight function to calculate the matching contribution weight based on the entity type, and directly determines the final matching result as unsuccessful when hard constraints are not met. This quantifies the differentiated importance of different entity types in matching, avoiding shallow text similarity. To address the issue of insufficient matching accuracy caused by treating all words equally in similarity matching, step S8 involves performing a multi-hop path search within the matching association graph structure. Starting with the resume entity node in the resume knowledge subgraph and ending with the job entity node in the job knowledge subgraph, a confidence score is calculated for each searched path based on the weight product of the relation edges and the path hop count decay factor. The top N paths with the highest confidence scores are selected as key implicit association paths. A third matching score is generated by aggregating the confidence scores of these paths, ensuring that resumes and job titles do not directly exist in the same text. Implicit relationships can be explicitly inferred through multi-hop paths in a knowledge graph. For example, a job seeker's project experience entity can be associated with a target position through multi-hop paths of projects, skills, and positions. Alternatively, work experience entity can achieve cross-domain matching of transferable skills through a "industry-ability mapping-position" path. This mechanism overcomes the limitation of pure text similarity comparison, which can only handle explicit co-occurrence information. It enables matching reasoning to conduct multi-layered logical reasoning based on entity relationships, making up for the technical shortcomings of existing technologies that cannot connect multi-dimensional information such as industry, projects, certificates, and work experience for comprehensive reasoning.

[0070] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A knowledge graph-based intelligent resume and job matching method, characterized by: Includes the following steps: S1. Obtain the resume text data and job requirement text data to be matched, and perform text preprocessing on the resume text data and job requirement text data respectively. The text preprocessing includes at least word segmentation, part-of-speech tagging, stop word filtering and preliminary named entity recognition. S2. Based on the pre-built talent domain knowledge graph, perform entity recognition and linking on the pre-processed resume text, extract the resume entity set, the resume entity set includes at least the education entity, major entity, skill entity, project experience entity, certificate and qualification entity and work experience entity, and assign corresponding entity type label and initial confidence to each entity; S3. Based on the talent domain knowledge graph, perform entity recognition and linking on the preprocessed job requirement text, and extract the job entity set. The job entity set includes at least the job name entity, required skills entity, education requirement entity, major requirement entity, experience requirement entity, and qualification requirement entity. S4. Using each entity in the resume entity set and the job entity set as nodes and the semantic relationships between entities as edges, construct a resume knowledge subgraph and a job knowledge subgraph respectively. The semantic relationships include at least hierarchical relationships, synonym relationships, mutual exclusion relationships, prerequisite dependency relationships and transferable skill relationships. S5. Perform graph embedding representation learning on the resume knowledge subgraph and the job knowledge subgraph respectively to generate resume graph embedding vector and job graph embedding vector. The graph embedding representation learning retains at least the first-order neighborhood structure information and the second-order semantic association information of the nodes. S6. Calculate the global semantic similarity between the resume image embedding vector and the job image embedding vector, and use it as the first matching score; S7. Based on the predefined entity weight system and constraint rule system in the talent domain knowledge graph, perform fine-grained weighted matching on a per-entity basis on the resume entity set and the job entity set to generate a second matching score. The entity weight system includes at least the skill mastery depth weight, hard threshold weight and industry scarcity weight, and the constraint rule system includes at least the hard constraint of education level, the constraint of years of experience and the mandatory constraint of qualifications. S8. Based on the resume knowledge subgraph and the job knowledge subgraph, perform graph path reasoning and multi-hop relationship reasoning to identify the implicit association path between the resume entity and the job entity, and generate a third matching score according to the number of hops of the implicit association path, the type of relationship on the path and the path confidence. S9. Input the first matching score, the second matching score and the third matching score into a preset multi-dimensional fusion matching model. After weighted fusion and nonlinear comprehensive judgment, output the final comprehensive matching score and generate an interpretable traceability path for the matching result. The interpretable traceability path includes at least the contribution of each dimension of the matching score and the corresponding relationship chain of key matching entities. S10. Based on the final comprehensive matching score and the preset matching threshold, output the matching conclusion between the job and the resume. The matching conclusion includes at least three states: successful matching, pending manual review, and failed matching.

2. The knowledge graph-based intelligent resume and job matching method according to claim 1, characterized in that: Step S2 involves entity recognition and linking of the preprocessed resume text based on a pre-built talent domain knowledge graph, which includes the following steps: 1) Input the preprocessed resume text into a preset entity recognition model, which is built based on the BERT-BiLSTM-CRF architecture and outputs candidate entities in the text and their entity types; 2) For each candidate entity, calculate the semantic similarity between the candidate entity and each entity in the talent domain knowledge graph, and select the graph entity with the highest semantic similarity as the link target of the candidate entity; 3) When the semantic similarity between a candidate entity and a graph entity is lower than the preset link threshold, the candidate entity is marked as an unlinked entity, and the new entity entry process is triggered.

3. The knowledge graph-based intelligent resume and job matching method according to claim 1, characterized in that: The step S4, which involves constructing the resume knowledge subgraph and the job knowledge subgraph, includes the following steps: 1) Using each entity in the resume entity set as a node, establish a first type of edge between each node of the resume entity based on the pre-stored inter-entity relationship in the talent domain knowledge graph; 2) Using each entity in the set of job entities as a node, and based on the pre-stored inter-entity relationships in the talent domain knowledge graph, establish a second type of edge between the nodes of the job entities; 3) Link the resume knowledge subgraph and the job knowledge subgraph through shared entity nodes to form a unified matching association graph structure.

4. The knowledge graph-based intelligent resume and job matching method according to claim 1, characterized in that: In step S5, graph embedding representation learning is performed on the resume knowledge subgraph and the job knowledge subgraph respectively. A graph embedding model based on graph attention network is adopted. When aggregating the features of neighbor nodes, the graph attention network assigns different attention weights to different neighbor nodes. The attention weights are calculated based on node type and relationship type.

5. The knowledge graph-based intelligent resume and job matching method according to claim 1, characterized in that: Step S7, which involves performing fine-grained weighted matching on a per-entity basis based on the predefined entity weight system and constraint rule system in the talent domain knowledge graph, includes the following steps: 1) For each pair of entities of the same type in the resume entity set and the job entity set, calculate the entity similarity score; 2) Call the corresponding weight function based on the entity type to calculate the matching contribution weight of the entity; 3) When the job requirements corresponding to the entity are hard constraints, check whether the resume entity meets the hard constraints. If it does not meet the constraints, the final matching result will be determined as unsuccessful. 4) Aggregate the weighted similarity scores of all pairs of entities of the same type to generate dimension matching scores for each dimension. Then, sum the dimension matching scores according to the preset dimension weights to generate the second matching score.

6. The knowledge graph-based intelligent resume and job matching method according to claim 1, characterized in that: Step S8 involves performing graph path reasoning and multi-hop relationship reasoning, which includes the following steps: 1) Starting from the resume entity node in the resume knowledge subgraph and ending from the job entity node in the job knowledge subgraph, perform a multi-hop path search in the matching association graph structure. 2) For each path found, calculate the confidence score of the path, which is calculated based on the weight product of each relation edge in the path and the decay factor of the path hop count; 3) Select the top N paths with the highest confidence scores as key implicit association paths, and aggregate the confidence scores of the key implicit association paths to generate a third matching score.

7. The knowledge graph-based intelligent resume and job matching method according to claim 1, characterized in that: The multi-dimensional fusion matching model in step S9 adopts a multilayer perceptron architecture. The input layer of the multilayer perceptron receives the first matching score, the second matching score, and the third matching score. The hidden layer uses the ReLU activation function for nonlinear transformation, and the output layer uses the Sigmoid activation function to output the final comprehensive matching score in the range of 0 to 1.

8. The knowledge graph-based intelligent resume and job matching method according to claim 1, characterized in that: The method for constructing the talent domain knowledge graph includes: 1) Collect multi-source talent field data, which includes at least job posting data from recruitment websites, professional encyclopedia data, industry skill standard data, and professional curriculum system data; 2) Perform structured processing on the multi-source talent domain data, extract talent domain entities and relational triples between entities. The entities include at least majors, skills, positions, industries, certificates, and institutions. The relations include at least inclusion relations, membership relations, prerequisite dependency relations, and capability mapping relations. 3) Store the entity and relation triples into a graph database to construct a knowledge graph ontology model for the talent domain; 4) Through an incremental learning mechanism, new entities and relationships are periodically extracted from newly collected data to dynamically update the talent domain knowledge graph.

9. A knowledge graph-based intelligent resume and job matching device, used to implement the method described in any one of claims 1-8, characterized in that: include: The data acquisition module is used to acquire resume text data and job requirement text data to be matched; The text preprocessing module is used to perform text preprocessing on the resume text data and the job requirement text data respectively. The text preprocessing includes at least word segmentation, part-of-speech tagging, stop word filtering and preliminary named entity recognition. The resume entity extraction module is used to perform entity recognition and linking on the pre-processed resume text based on a pre-built talent domain knowledge graph, and extract a resume entity set. The resume entity set includes at least education entity, major entity, skill entity, project experience entity, certificate and qualification entity, and work experience entity. The job entity extraction module is used to perform entity recognition and linking on the preprocessed job requirement text based on the talent domain knowledge graph, and extract a set of job entities. The set of job entities includes at least job name entity, required skills entity, education requirement entity, professional requirement entity, experience requirement entity, and qualification requirement entity. The knowledge subgraph construction module is used to construct a resume knowledge subgraph and a job knowledge subgraph respectively, using each entity in the resume entity set and the job entity set as nodes and the semantic relationship between entities as edges. The graph embedding representation module is used to perform graph embedding representation learning on the resume knowledge subgraph and the job knowledge subgraph respectively, and generate resume graph embedding vector and job graph embedding vector. The global semantic matching module is used to calculate the global semantic similarity between the resume image embedding vector and the job image embedding vector, and output the first matching score. The fine-grained weighted matching module is used to perform fine-grained weighted matching on an entity-by-entity basis on the resume entity set and the job entity set based on the predefined entity weight system and constraint rule system in the talent domain knowledge graph, and output a second matching score. The graph reasoning and matching module is used to perform graph path reasoning and multi-hop relationship reasoning based on the resume knowledge subgraph and the job knowledge subgraph, identify the implicit association path between resume entities and job entities, and output a third matching score. The multi-dimensional fusion output module is used to input the first matching score, the second matching score, and the third matching score into a preset multi-dimensional fusion matching model, and output the final comprehensive matching score and matching conclusion after weighted fusion and nonlinear comprehensive judgment.

10. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, implements the steps of the knowledge graph-based intelligent matching method for resumes and job positions as described in any one of claims 1 to 8. The storage medium includes any one of ROM, RAM, magnetic disk, and optical disk.

Images