Business organization tree-based label model resume recommendation method

Abstract

The invention provides a resume recommendation method based on a label model of a business organization tree, which comprises the following steps of: constructing the label model of the business organization tree, fusing post inherent characteristics and father node business attributes through label inheritance logic, and generating a full-dimensional business vector; performing structured analysis on the resume data, performing time sequence weighting calculation on the skill entity according to the current time, and generating a resume feature vector; calculating a service capability semantic matching result and a service scene integrating degree matching result between the full-dimensional service vector and the resume feature vector; and performing weighted fusion on the matching result according to the weight parameter to generate a recommendation list. The invention further relates to a dynamic weight optimization mechanism based on reinforcement learning. The post service context can be effectively complemented, the timeliness of the skills of the candidates can be accurately evaluated, the adaptive evolution of the recommendation strategy is realized, and the accuracy of person and post matching is improved.

Description

Technical Field

[0001] This disclosure relates to the fields of internet data processing and artificial intelligence technology, and in particular to a resume recommendation method based on a tag model of a business organization tree. Background Technology

[0002] With the rapid development of internet technology and the deepening of enterprise digital transformation, online recruitment platforms and internal talent pools have accumulated massive amounts of resume data. How to accurately and efficiently identify candidates who meet the specific job requirements from this vast data resource has become a core problem that urgently needs to be solved in the field of intelligent recruitment. Current mainstream solutions typically utilize natural language processing technology to extract keywords or semantic feature vectors from job descriptions and resume texts, and calculate the similarity between the two to achieve person-job matching.

[0003] However, existing resume recommendation technologies still have significant limitations in practical applications. Traditional matching algorithms mostly treat job descriptions as independent textual evidence, ignoring the business context of the job within the complex organizational structure of an enterprise. Because job descriptions are often brief, heavily templated, or outdated, relying solely on their literal content cannot fully reflect the specific technology stack preferences of the business unit to which the job belongs or the specific stage requirements of the current project. As a result, while the recommendation results may match at the keyword level, they fail to align with the team's actual needs at a deeper business level.

[0004] Meanwhile, existing resume parsing technologies typically focus on static entity recognition, lacking consideration of the temporal dimension of skill mastery. When extracting candidate skill tags, current technologies often fail to effectively distinguish whether a candidate has frequently used a skill recently or only used it in historical projects from many years ago, ignoring the forgetting curve and time decay characteristics of technical skills. This static feature extraction approach can easily lead to the system recommending candidates with rusty skills or outdated technology stacks to companies, reducing the effectiveness of person-job matching.

[0005] Furthermore, existing job matching models often employ preset fixed weights or static rules to fuse multi-dimensional scores, making it difficult to adapt to dynamically changing recruitment strategies. In real-world recruitment scenarios, different business stages or interviewers may emphasize different aspects of a candidate's abilities. Fixed-weight models cannot effectively utilize real-time feedback from recruiters regarding recommendations, interviews, or rejections to self-correct, resulting in rigid recommendation strategies that fail to adapt and continuously evolve over time. Summary of the Invention

[0006] This disclosure aims to at least partially address one of the technical problems of existing resume recommendation technologies, which suffer from low accuracy in matching people to jobs and poor business fit due to a lack of in-depth consideration of job context and the timeliness of skills, and the inability of recommendation strategies to adaptively adjust based on user feedback.

[0007] Therefore, this disclosure proposes a resume recommendation method based on a tag model of a business organization tree.

[0008] This invention provides a resume recommendation method based on a tag model of a business organization tree, the method comprising the following steps: Construct a business organization tree label model, which maps the enterprise organizational structure into a multi-level tree structure and configures label inheritance logic. By integrating the inherent characteristics of job nodes with the business attributes of nodes on the parent path of job nodes, a full-dimensional business vector of the job to be recruited is generated. The collected resume data is structured and parsed to identify entities and relationships in the resume text. The identified skill entities are then time-weighted and calculated based on the current system time to generate a resume feature vector containing skill intensity information. Perform multi-dimensional adaptation calculations, at least calculate the semantic matching result of business capabilities between the full-dimensional business vector and the resume feature vector, and calculate the business scenario fit matching result between the business scenario entities in the resume and the scenario tag set of the business node to which the position belongs; Based on the preset weight parameters, the semantic matching results of business capabilities and the matching results of business scenario fit are weighted and fused to obtain a comprehensive matching score, and a resume recommendation list is generated based on the comprehensive matching score.

[0009] In the above method, the process of constructing the business organization tree label model specifically involves defining the hierarchy of the tree structure, which includes at least a root node, business unit nodes, project nodes, and job node. For job nodes, explicit keywords from the job description are extracted as inherent feature vectors. Subsequently, a vector synthesis operation is performed, traversing upwards along the tree structure to obtain the attribute label vectors of all parent nodes on the parent path to which the job node belongs. Using a weighted summation algorithm, the inherent feature vectors are synthesized with the attribute label vectors of each parent node to generate a full-dimensional business vector. During the synthesis process, parent nodes closer to the job node hierarchy are assigned a larger hierarchy weight coefficient, thereby ensuring that the job matching benchmark includes both the specific requirements of the job itself and inherits the macro-level technology stack and business background of the project and business unit to which it belongs.

[0010] In the above method, the process of structured parsing of the collected resume data includes denoising and normalization preprocessing of the original resume data to obtain a standard text sequence. This standard text sequence is then input into a pre-trained deep learning model, which employs a BERT-BiLSTM-CRF network architecture. The deep learning model outputs serialized entity labels, and the identified entity label types include at least skill entities, job title entities, business scenario entities, and time entities.

[0011] Furthermore, the process of performing time-weighted calculations on the identified skill entities based on the current system time includes: for each identified skill entity, determining the project end time of the last occurrence of the skill entity based on the associated time entity; calculating the time difference between the project end time and the current system time; determining the base score of the skill entity; and using an exponential decay function to process the base score and the time difference to calculate the skill intensity value of the skill entity. The exponential decay function is configured with a time decay coefficient to control the decay rate, thereby reducing the weight of outdated skills in the feature vector and highlighting the candidate's recently active skills.

[0012] The multi-dimensional adaptation calculation process also includes a hard-condition filtering step. Before calculating the matching result between the full-dimensional business vector and the resume feature vector, the structured fields and hard constraints of the job position are extracted from the resume data. A hard indicator function containing a series of Boolean logic judgments is constructed. The hard indicator function is used to determine whether the structured fields meet the hard constraints. If not, the subsequent calculation is terminated and the comprehensive matching score is set to zero, thereby improving the screening efficiency.

[0013] Specifically, calculating the semantic matching result of business capabilities between the full-dimensional business vector and the resume feature vector involves mapping the full-dimensional business vector and the resume feature vector to the same vector space. A cosine similarity algorithm is then used to calculate the cosine value of the angle between the two vectors in the vector space. This cosine value characterizes the semantic similarity between the candidate's skill distribution and the job's business and technical requirements. Calculating the business scenario fit matching result between the business scenario entities in the resume and the scenario label set of the business nodes to which the job belongs involves extracting the business scenario labels associated with the business unit nodes and project nodes from which the full-dimensional business vector originates to form a job scenario set. The business scenario entities contained in the resume feature vector are then extracted to form a resume scenario set. The Jaccard similarity coefficient algorithm is used to calculate the ratio of the number of intersection elements to the number of union elements between the job scenario set and the resume scenario set.

[0014] Furthermore, the method also includes weighted fusion of the matching results according to a preset weight parameter configuration. The current weight parameter configuration is obtained, which includes a first weight parameter corresponding to the semantic matching result of the business capability and a second weight parameter corresponding to the matching result of the business scenario fit. The semantic matching result of the business capability is multiplied by the first weight parameter, the matching result of the business scenario fit is multiplied by the second weight parameter, and the two are added together to obtain the comprehensive matching score.

[0015] To achieve adaptive optimization of the recommendation strategy, this method further includes a dynamic weight optimization step based on reinforcement learning. A state space is defined, containing at least the current job type features and the current weight parameter configuration. A reward function is defined, mapping user actions on the resume recommendation list to numerical reward values; these actions include at least browsing, downloading, inviting interviews, and rejecting. A state-action value function is constructed to evaluate the long-term expected return from adjusting the weight parameter configuration in a specific state. The system captures user actions on the recommendation results in real time, calculates the cumulative reward value based on the reward function, updates the state-action value function using a Q-learning algorithm or a deep Q-network strategy based on the cumulative reward value, and determines the optimal weight adjustment action for the next time step based on the updated function. This generates a new weight parameter configuration for application in subsequent weighted fusion steps.

[0016] This invention provides a resume recommendation method based on a tag model of a business organization tree. It has the following beneficial effects: 1. This invention constructs a business organization tree label model and configures label inheritance logic, placing job nodes in the context of the enterprise organizational structure. It uses the attribute labels of business units and project nodes on the parent path to complete the implicit information of the job description, solving the feature sparsity problem caused by the simplified or templated job descriptions in traditional methods. This ensures that the generated full-dimensional business vector can take into account both the inherent skill requirements of the job and the macro business technology stack of the team, thereby improving the business fit of the person-job match.

[0017] 2. This invention adopts a time decay-based skill intensity calculation mechanism, which combines skill entities and time entities identified by natural language processing. It uses an exponential decay function to dynamically weight skill scores, which can quantify the freshness of a skill based on the project end time of its last appearance. This effectively distinguishes between the obsolescence and current proficiency of a candidate's skills, avoiding the technical defects of traditional static matching that cannot distinguish between historical and active skills, and improving the accuracy of assessing the candidate's current competence.

[0018] 3. This invention introduces a dynamic weight optimization mechanism based on reinforcement learning. By constructing a state-action value function and defining a reward function, it captures user feedback behaviors such as browsing, interviewing, or being eliminated from the recommendation list in real time, realizing adaptive closed-loop adjustment of weight parameters. It can automatically optimize the weight allocation of business capabilities and scenario fit according to the actual screening preferences of different recruitment stages and job types, solving the problem of rigidity of traditional fixed weight models and realizing the continuous evolution of recommendation strategies. Attached Figure Description

[0019] The above and / or additional aspects and advantages of this disclosure will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, in which: Figure 1 This is a schematic diagram of the system architecture and module interaction provided in the embodiments of this disclosure; Figure 2 This is a main flowchart of the method provided in the embodiments of this disclosure; Figure 3 This is a schematic diagram of the tag inheritance and multidimensional matching logic provided in the embodiments of this disclosure.

[0020] Among them, 101, Business Organization Tree Tag Model Construction Module; 102, Resume Data Structure Parsing Module; 103, Multi-Dimensional Adaptation Calculation Module; 104, Dynamic Weight Optimization Module. Detailed Implementation

[0021] Embodiments of this disclosure are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this disclosure, and should not be construed as limiting this disclosure.

[0022] It should be noted that all information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.), and signals involved in this disclosure are authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. The acquisition, transmission, storage, use, and processing of data in the technical solution of this disclosure all comply with the relevant provisions of national laws and regulations.

[0023] It should be noted that in the embodiments disclosed herein, certain software, components, models, and other existing solutions in the industry are mentioned. These should be considered as exemplary and are intended only to illustrate the feasibility of implementing the technical solutions disclosed herein. However, they do not mean that the applicant has used or necessarily used such solutions.

[0024] See attached document Figure 1This invention provides a resume recommendation method based on a business organization tree tag model, which runs on a computer device including a processor, memory, and communication interface. The system mainly includes: a business organization tree tag model construction module 101, a resume data structured parsing module 102, a multi-dimensional adaptation calculation module 103, and a dynamic weight optimization module 104.

[0025] The business organization tree label model construction module 101 is used to establish a label system reflecting the hierarchical relationships of an enterprise's business operations. This module reads the enterprise's human resources organizational structure data and constructs a tree-like data structure containing root nodes, business unit nodes, project nodes, and job node nodes. The business organization tree label model construction module 101 is configured with label inheritance logic to pass the business attribute labels of parent nodes to child nodes. For any job node in the tree... This module generates full-dimensional business vectors through a vector synthesis algorithm. Full-dimensional business vector The calculation formula is as follows: ; in, This represents the inherent feature vector extracted from the job description; Indicates job node The set of parent path nodes in the organization tree; Indicates the parent node Attribute tag vector; Indicates the parent node Hierarchical weight coefficients; This represents the inherent characteristic adjustment coefficient, with a value ranging from 0 to 1.

[0026] The simplified data structuring and parsing module 102 connects to an external data source, collects simplified data from ten meters away, and transforms it into structured feature vectors. This module includes a natural language processing unit configured with a BERT-BiLSTM-CRF deep learning model. The BERT layer generates word vector representations of the text, the BiLSTM layer extracts bidirectional semantic features of the text's context, and the CRF layer predicts entity labels in the text sequence. The entity labels include skill entities, job entities, business scenario entities, and time entities. The resume data structuring and parsing module 102 also includes time-series weighted logic for calculating skill intensity values ​​based on the time dimension. This is used to analyze the skills identified in the resume. This module calculates skill strength. The formula is as follows: ; in, This represents the base score extracted based on the resume text description; λ represents the difference between the current time and the time when the skill last appeared in the resume; λ represents the time decay coefficient. The resume data structuring and parsing module 102 generates resume feature vectors based on the calculated skill intensity values ​​and extracted scene entities. .

[0027] The multi-dimensional adaptation calculation module 103 is connected to the business organization tree label model construction module 101 and the resume data structure parsing module 102, and is used to receive full-dimensional business vectors. and resume feature vector This module sequentially performs hard-condition filtering, business capability semantic matching, and business scenario fit matching. In the business capability semantic matching, this module calculates a full-dimensional business vector. With resume feature vector The cosine similarity between them. In business scenario fit matching, this module compares the set of business scenario entities in the candidate's resume with the set of business scenario tags of the business unit node and project node to which the job node belongs, and calculates the overlap between the sets. The multi-dimensional adaptation calculation module 103 generates a comprehensive matching score based on the calculation results of each dimension. The calculation formula is as follows: ; in, The score represents the semantic matching score of business capabilities; The score represents the degree of fit between the business scenario and the target audience. This indicates the score for soft skills matching. , , Let each represent the weight parameter of the corresponding dimension, and satisfy the following conditions: .

[0028] The dynamic weight optimization module 104 is connected to the multi-dimensional adaptation calculation module 103 and is used to adjust the weight parameters based on feedback data. , and This module is configured with a reinforcement learning agent, defining the current job type and weight configuration as a state, and defining the weight adjustment amount as an action. This module receives operation feedback signals from the user terminal and calculates a reward value based on the type of operation feedback signal. The operation feedback signal types include browsing, downloading, interviewing, and elimination. The dynamic weight optimization module 104 updates the state-action value function based on the accumulated reward value and determines the weight parameter configuration for the next time step based on the updated value function.

[0029] The above modules communicate via the system bus. The business organization tree label model construction module 101 outputs a full-dimensional business vector. The resume feature vector is used as the input benchmark for the multi-dimensional adaptation calculation module 103. The resume data structuring and parsing module 102 outputs the resume feature vector. As the matching object of the multi-dimensional adaptation calculation module 103. The weight parameters output by the dynamic weight optimization module 104. , , The calculation logic of the multi-dimensional adaptation calculation module 103 is updated in real time. By executing the functions of the above modules, the computer device generates and outputs the resume recommendation list.

[0030] See attached document Figure 2 The resume recommendation method provided by this invention includes a complete data processing flow from business model construction to closed-loop feedback optimization.

[0031] First, the system executes the initialization construction process of the business organization tree label model. Upon startup, this process reads organizational structure metadata from the enterprise resource planning system or human resource management database. The system maps the read metadata to nodes in a tree-like topology, defining a four-level hierarchy: root node, business unit node, project node, and job node. For each leaf node, i.e., the job node, the system not only extracts explicit keywords from the job description but also triggers label inheritance logic. This logic traverses upwards along the tree structure, obtaining the business attribute labels of all nodes on the parent path of the job node. Based on a preset hierarchical weight configuration, the system weightedly synthesizes the parent node's business attribute label vector with the job's own inherent feature vector, generating a full-dimensional business vector for subsequent matching.

[0032] While completing or updating the business organization tree construction, the system executes the structured parsing process of resume data in parallel or sequentially. The system receives the unstructured resume text data stream to be processed from a preset data source interface. After preprocessing the input text data by denoising, word segmentation, and normalization, it is input into a pre-trained BERT-BiLSTM-CRF deep learning model. The model outputs serialized entity labels, identifying skill entities, job entities, business scenario entities, and time entities in the text. Based on the identified time entities, the system determines the most recent usage time point corresponding to each skill entity. Subsequently, the system calls the time-series decay calculation logic, using the current system time as a benchmark, to calculate the time interval of each skill, and applies an exponential decay function to adjust the basic score of the skill, generating a resume feature vector that includes timeliness weights.

[0033] After obtaining the full-dimensional business vector and resume feature vector, the system enters a multi-dimensional adaptation calculation process. This process first performs Boolean logic filtering based on hard conditions, comparing structured fields such as education and years of work experience with the job's hard indicators, and eliminating resumes that do not meet the threshold conditions. For the filtered resumes, the system performs a two-channel soft matching calculation. The first channel calculates the cosine similarity between the full-dimensional business vector and the resume feature vector in the vector space to quantify the semantic matching degree of business capabilities. The second channel performs a business scenario fit calculation, where the system extracts the set of business scenario entities from the resume and performs intersection or path distance calculations with the scene label sets of corresponding business unit nodes and project nodes in the business organization tree to quantify the overlap between the candidate's past experience and current business background. Based on the current weight parameter configuration, the system linearly weights and fuses the business capability semantic matching result, the business scenario fit calculation result, and the soft skill score to obtain a comprehensive matching score, and generates a recommendation list based on this score, outputting it to the user terminal.

[0034] After the recommendation list is output, the system initiates a dynamic weight optimization process. The system monitors user actions on the recommendation list in real time. When discrete action events such as browsing, downloading, receiving interview invitations, or rejecting rejections are detected, the system maps them to corresponding numerical reward signals. The system uses reinforcement learning algorithms to process these reward signals and update the state-action value function. Based on the updated value function, the system calculates new weight parameter configurations and stores them in the system configuration database. These new weight parameter configurations will be applied to the next round of multi-dimensional adaptation calculations, thereby enabling the recommendation logic to adaptively adjust for specific business nodes.

[0035] The business organization tree label model construction module 101 executes the construction process described in this embodiment, aiming to transform the static organizational structure of an enterprise into a dynamic and computable business semantic space.

[0036] The construction process begins with the initial access to organizational structure metadata. The business organization tree tag model construction module 101 establishes a connection with the enterprise's human resource management system through an application programming interface (API) to read the organizational unit list and hierarchical relationship data in real time. Based on the read data, this module instantiates a tree-like topology structure in memory, which contains four logical levels: root node level (L1), business unit node level (L2), project node level (L3), and position node level (L4).

[0037] At the root node level (L1), the system configures an enterprise-level set of general tags. This set contains tags that reflect the company's core values ​​and general talent quality requirements, such as integrity, teamwork, and innovation-driven development. These tags serve as global baseline attributes, with a default inheritance effect on all lower-level nodes in the tree, and are primarily used for subsequent soft skills matching.

[0038] At the business unit node level (L2), the system extracts business domain tags and technology stack preference tags based on the functional descriptions of each business unit or department. For example, for the Fintech Business Unit node, the system associates business attribute tags such as high concurrency processing, fund security, and data consistency; for the Mobile R&D Department node, the system associates technology stack tags such as iOS / Android ecosystem and cross-platform technologies. These tags define the macro-business context in which all positions within that branch exist.

[0039] At the project node level (L3), the system associates specific project execution environment tags. This module allows administrators to input or automatically parse project management documents to extract tags reflecting project characteristics, including project scale (e.g., tens of millions of users, early stage), development model (e.g., Agile development Scrum, Waterfall), technical architecture (e.g., microservices, Serverless), and lifecycle stage. These tags are used to accurately match candidates' past project experience scenarios.

[0040] For the job node level (L4), i.e. the leaf nodes of the tree, the system parses the specific job description (JD). The business organization tree label model construction module 101 uses natural language processing algorithms to extract explicit hard skill requirements (such as Java, Python), job level requirements (such as P6, senior), and specific job descriptions to form a set of inherent characteristics of the job itself.

[0041] After completing the label association for each level of nodes, the business organization tree label model construction module 101 performs label inheritance and vector synthesis operations. For any job position node to be recruited, the system first uses a pre-trained word vector model (such as Word2Vec or BERT) to convert the text labels on the above nodes into high-dimensional numerical vectors. Subsequently, the system executes a path backtracking algorithm, starting from the job position node, traversing upwards along the tree structure to the root node, and collecting the attribute label vectors of the parent nodes (L3 project node, L2 business unit node, L1 root node) passed along the path.

[0042] The system performs weighted synthesis based on the full-dimensional business vector calculation formula described in Example 1. During the calculation process, the system assigns differentiated hierarchical weight coefficients to parent nodes at different levels. The specific weight allocation strategy follows the distance decay principle, that is, the closer the parent node is to the job node (such as the L3 project node), the greater the impact of its business attributes on the actual work content of the job, and therefore it is assigned a relatively large weight value; while nodes that are farther away (such as the L1 root node) have relatively small weight values. At the same time, the system balances the ratio between the job's own JD description and the inherited business context features by adjusting the inherent feature adjustment coefficient. If the job JD description is extremely detailed, the system automatically increases the adjustment coefficient; if the job JD description is brief, the system automatically decreases the coefficient, relying more on inherited business attributes to complete the job profile.

[0043] Finally, the business organization tree label model construction module 101 generates a full-dimensional business vector for each job node, containing its own characteristics and inherited characteristics. This vector is persistently stored in a graph database or vector database, serving as the benchmark data object for the subsequent multi-dimensional adaptation calculation module 103 to perform person-job matching. When the enterprise's organizational structure changes, such as adding a new business line or adjusting project affiliation, this module automatically triggers the recalculation logic of the local subtree, updating the full-dimensional business vector of the affected job nodes.

[0044] The resume data structuring and parsing module 102 executes the processing flow described in this embodiment. Its core is to transform unstructured, multi-format raw resume data into a high-dimensional vector representation with timeliness features and business semantics.

[0045] The parsing process begins with the access and preprocessing of multi-source heterogeneous data. The resume data structuring and parsing module 102 receives raw resume files from different channels through a standardized data interface, covering formats such as PDF, Word documents, images, and HTML text from web pages. This module first calls an Optical Character Recognition (OCR) engine or document parsing library to extract the plain text content and performs cleaning operations to remove headers, footers, special symbols, and garbled text. Subsequently, the module performs text normalization processing, using a pre-built thesaurus to map non-standard skill descriptions to unified terms. For example, JS and Node.js are uniformly mapped to the JavaScript skill cluster, and AWS and Alibaba Cloud are normalized to standard tags related to cloud computing platforms, thereby eliminating the impact of expression differences on subsequent matching.

[0046] The preprocessed standard text sequence is input into a pre-trained BERT-BiLSTM-CRF deep learning model for entity recognition and relation extraction. In this model architecture, the BERT layer acts as an embedding layer, utilizing its multi-layer attention mechanism to process the input text sequence and generate dynamic word vectors at the character level that contain deep syntactic and semantic information. These word vectors not only encode the meaning of the characters themselves but also dynamically fuse their contextual information within the sentence, effectively distinguishing polysemous words (e.g., distinguishing between "Go" as a programming language and "Go" as an action).

[0047] The dynamic word vector sequence output by the BERT layer is then fed into a Bidirectional Long Short-Term Memory (BiLSTM) layer. The BiLSTM layer contains two LSTM units, one forward and one backward, scanning from the beginning and end of the text, respectively. This layer captures long-range contextual dependencies through gating mechanisms (input gate, forget gate, output gate), outputting the hidden state vector for each character. This step is crucial for understanding complex project experience descriptions, such as determining whether a skill is corely used in a project or only marginally involved in projects one is familiar with.

[0048] The output of the BiLSTM layer is fed into a Conditional Random Field (CRF) layer. The CRF layer introduces a state transition matrix to globally constrain and optimize the output label sequence, ensuring that the predicted labels conform to grammatical rules (e.g., the label I-Skill must immediately follow B-Skill or I-Skill). Through this layer of processing, the resume data structure parsing module 102 accurately identifies key entities in the text, including skill entities (such as specific programming languages ​​or tools), job entity entities (such as job titles), business scenario entities (such as high concurrency, cold start, and SaaS), and time entities (such as project start and end times, and work start and end times). Simultaneously, the model extracts the relationships between entities through dependency parsing, establishing triples linking skills, projects, and times to clarify the specific time period in which a particular skill was used in a project.

[0049] After completing the structured extraction of entities and relationships, the data structure parsing module 102 executes the core time-weighted calculation logic. The system does not consider all skills possessed by the candidate to be of equal importance, but rather calculates the freshness of a skill based on the time of its last use. For each identified skill entity... The module first determines the project end time of the last use of the skill based on its associated time entity, and then calculates the time difference between that end time and the current system time. (Usually in months).

[0050] Subsequently, the module corrects the skill score according to the skill intensity calculation formula given in Example 1. During this calculation process, the base score... The determination of the score depends on the modifiers used for the skill in the resume or the role played in the project. For example, if the resume describes the skill as proficient or the project role as leading, the system assigns a higher base score (e.g., 1.0); if it describes the skill as having some knowledge or the role as assisting, a lower base score (e.g., 0.5) is assigned. Regarding the time decay coefficient... The system supports differentiated configurations based on different technical fields. For skills like front-end frameworks that are updated and iterated very quickly, larger configurations are recommended. The value is designed to decay rapidly over time; for skills related to fundamental theory or low-level algorithms, a smaller configuration is required. Values ​​are preserved to maintain their long-term validity.

[0051] Finally, the resume data structuring and parsing module 102 combines all the time-weighted skill intensity values ​​with the extracted business scenario entities and job entities, and generates a fixed-dimensional resume feature vector through vector concatenation or attention-weighted pooling operations. This vector not only quantifies the candidate's skill level, but also comprehensively portrays the candidate's current business competence through time and scenario dimensions, and transmits the vector to the multi-dimensional adaptation calculation module 103 for subsequent matching.

[0052] See attached document Figure 1 - Appendix Figure 3 The multi-dimensional adaptation calculation module 103 executes the calculation process described in this embodiment. This process combines hard threshold filtering with soft multi-dimensional four-matching algorithm through a strategy that combines serial and parallel processing to calculate the comprehensive fit between the candidate and the job requirements that include business context.

[0053] The calculation process begins by initiating a hard-condition filtering subroutine. The multi-dimensional adaptation calculation module 103 receives structured fields (such as education level, years of work experience, expected salary range, and work location) output by the resume data structure parsing module 102, as well as the job-related hard constraints output by the business organization tree label model construction module 101. To improve the processing efficiency of massive amounts of resumes, the module employs a fast retrieval mechanism based on inverted indexes or bitmap indexing technology. The system defines hard indicator function. Used to determine candidates Do you meet the job requirements? All hard constraint sets Indicator functions The calculation logic is as follows: ; in, For the characteristic function, when the candidate's attribute value Meet the threshold requirements for the position Returns 1 if the condition is met, otherwise returns 0. For non-numerical indicators (such as education level), the system uses a pre-defined level mapping table to convert them into a comparable numerical sequence (e.g., Associate's Degree = 1, Bachelor's Degree = 2, Master's Degree = 3). If the calculation result is 0, the module will directly terminate the subsequent calculation of the current resume and set the comprehensive matching score to 0; if the result is 1, the subsequent soft matching channel will be activated.

[0054] The filtered resume data enters the semantic matching stage for business capabilities. The core of this stage is to measure the vector space distance between the candidate's skills and the job's business and technical requirements. The multi-dimensional adaptation calculation module 103 obtains the full-dimensional business vector for the job generated in Example 1. and resume feature vector Since both vectors are constructed based on the same pre-trained language model embedding space, the module directly calculates their cosine similarity. This calculation not only considers the directional consistency of the vectors, i.e., whether the distribution patterns of skill combinations are similar, but also effectively avoids the interference of resume text length on the matching results. The calculated similarity is... The numerical range is normalized to the [0,1] interval. The higher the value, the more the candidate's skill set matches the job and the business and technical atmosphere of the team.

[0055] In parallel, the module performs a business scenario fit matching step. This step is independent of the skill vector matching mentioned above and is specifically used to assess the alignment between the candidate's past experience and the company's specific business areas. The multi-dimensional adaptation calculation module 103 extracts the set of business scenario tags associated with the business unit node (L2) and project node (L3) to which the current position belongs from the business organization tree tag model. (For example, the collection contains {cross-border e-commerce, high concurrency, payment clearing}). Simultaneously, extract the business scenario entity set from the resume parsing results. The module uses the Jaccard similarity coefficient algorithm to calculate the overlap between two sets, which is then used as the scene fit score. The specific calculation formula for this algorithm is as follows: ; in, This represents the intersection operation of sets. This represents the union operation of sets. The cardinality of a set (the number of elements) is represented. To prevent smoothing terms with a denominator of zero, in an advanced implementation, if the tags do not perfectly match literally, the module also introduces semantic distance calculation based on a knowledge graph. This treats scenario tags belonging to the same domain but with different expressions (such as B2C and retail) as having a certain degree of similarity, thereby expanding the intersection. The calculation range.

[0056] In addition, the module also calculates soft skills matching scores. The calculation is based on the degree of matching between self-evaluation and soft skills keywords extracted from the resume and the general competency tags defined by the root node (L1) of the business organization tree, or by weighting the structured scores of external psychological assessment data.

[0057] Finally, the multi-dimensional adaptation calculation module 103 performs comprehensive weighted fusion. This module reads the weight parameter configuration sent in real time by the dynamic weight optimization module 104. According to the comprehensive matching score formula given in Example 1 The system then synthesizes the final scores. It sorts all candidates' overall scores in descending order and generates a recommendation list. This list not only includes the ranking but also outputs sub-scores for each dimension (ability score, scenario score, and soft skills score) to the user terminal, providing interpretable reasons for the recommendations.

[0058] The dynamic weight optimization module 104 executes the closed-loop feedback adjustment process described in this embodiment, aiming to solve the technical problem that static weight configuration is difficult to adapt to the dynamic recruitment strategies and differentiated job preferences of enterprises, and to achieve adaptive evolution of recommendation logic.

[0059] This process is based on the mathematical model of Markov Decision Process (MDP). The dynamic weight optimization module 104 first defines the three core elements of reinforcement learning: state space, action space, and reward function.

[0060] In terms of state space definition, the module maps the current recruitment context to a state vector. This vector contains not only the current job type characteristics (such as the business unit node ID to which the job belongs and the job function classification), but also the current weight parameter configuration. In addition, it includes feedback statistics from the previous recommendation cycle (such as interview conversion rate). By incorporating weight configuration information, the system can perceive the current strategy's position in the parameter space.

[0061] Regarding action space definition, modules define actions. This refers to the fine-tuning operation on the weight parameter vector. To ensure training convergence, the actions are designed as discrete step adjustments; for example, the action set can be defined as {increase}. reduce ,Increase reduce , remain unchanged, ...}, the adjustment step size corresponding to each action This is a preset fixed value (e.g., 0.05). After the module executes the action, the new weight parameters will be normalized to ensure that they meet the requirements. The constraints.

[0062] In terms of reward function design, this is a key signal guiding model evolution. The dynamic weight optimization module 104 captures in real-time the interaction behavior of recruiters with recommended candidates on the user terminal through the event tracking log interface. The system quantifies these discrete interaction behaviors into numerical, instant rewards. The feedback reward function for a single candidate recommendation result is defined as follows: ; in, , , All are positive real numbers, and satisfy the following conditions: This is to reflect the differences in the level of approval represented by different behaviors; Let be a positive real number, representing the penalty term. After a recommendation batch (e.g., recommending N candidates) ends, the module calculates the cumulative reward for that batch. .

[0063] Based on the above definition, the dynamic weight optimization module 104 updates the state-action value function using the Q-learning algorithm or the Deep Q-Network (DQN) strategy. This function is used to evaluate the state in a specific condition. Take action below The expected long-term returns that can be obtained. At the end of each decision cycle (such as daily or weekly), the module calculates the accumulated rewards based on the data collected. Update the Q value according to the following rules: ; in, This represents the learning rate, used to control the speed at which new and old knowledge are integrated. This represents the discount factor, used to balance the weight of immediate rewards and long-term rewards; This represents the maximum expected value that can be obtained in the next state.

[0064] Through continuous iteration and updates, the dynamic weight optimization module 104 can learn the optimal weight strategy for specific business scenarios. For example, for delivery and implementation positions that emphasize practical experience, if candidates with high academic qualifications but low scenario scores recommended by the system in the early stages are frequently rejected (receiving negative rewards), the Q-learning algorithm will gradually reduce its weighting. (Capability semantic weight) and improve The action Q-value corresponds to the (scene fit weight). In subsequent recommendations, the module will utilize... -greedy strategy (i.e., using probability) Choose the action with the highest Q value, based on probability. (Random exploration is conducted), and the system automatically switches to a weight configuration that prioritizes scenario relevance, thereby outputting a recommendation list that better matches the actual hiring preferences for the position. The updated weight parameters are synchronized in real time to the multi-dimensional adaptation calculation module 103, completing closed-loop optimization.

Claims

1. A resume recommendation method based on a tag model of business organization tree, characterized in that, Includes the following steps: Construct a business organization tree label model, and generate a full-dimensional business vector for the job to be recruited by integrating the inherent characteristics of job nodes with the business attributes of nodes on the parent path of the job node. The collected resume data is parsed, and the identified skill entities are time-weighted according to the current system time to generate a resume feature vector containing skill intensity value information. Perform adaptation calculations to obtain the business capability semantic matching result between the full-dimensional business vector and the resume feature vector, as well as the business scenario fit matching result between the business scenario entities in the resume and the scenario tag set of the business node to which the position belongs; Based on the weight parameter configuration, the semantic matching results of the business capabilities and the matching results of the business scenario are weighted and fused to obtain a comprehensive matching score and generate a resume recommendation list.

2. The resume recommendation method based on a tag model of business organization tree according to claim 1, characterized in that, The steps for constructing the business organization tree label model include: Define the hierarchy of the tree structure, which includes at least a root node, a business unit node, a project node, and a job position node; For the aforementioned job node, explicit keywords from the job description are extracted as inherent feature vectors; Perform vector composition operation, traverse upwards along the tree structure, and obtain the attribute label vectors of all parent nodes on the parent path to which the job node belongs; Using a weighted summation algorithm, the inherent feature vector is synthesized with the attribute label vector of each parent node to generate the full-dimensional business vector; Among them, the closer the parent node is to the job node level, the greater the level weight coefficient it is assigned.

3. The resume recommendation method based on a tag model of business organization tree according to claim 1, characterized in that, The steps for performing structured parsing on the collected resume data include: The original resume data was preprocessed by denoising and normalization to obtain a standard text sequence; The standard text sequence is input into a pre-trained deep learning model, which employs a BERT-BiLSTM-CRF network architecture. The deep learning model outputs serialized entity tags, and the types of the entity tags include at least skill entities, job entities, business scenario entities, and time entities.

4. The resume recommendation method based on a tag model of business organization tree according to claim 3, characterized in that, The step of performing time-weighted calculation on the identified skill entities based on the current system time includes: For each identified skill entity, the project end time of the last occurrence of the skill entity is determined based on the associated time entity; Calculate the time difference between the project's end time and the current system time; The base score of the skill entity is determined, and the difference between the base score and the time is processed using an exponential decay function to calculate the skill intensity value of the skill entity. The exponential decay function has a time decay coefficient that controls the decay rate.

5. The resume recommendation method based on a tag model of business organization tree according to claim 1, characterized in that, The adaptation calculation also includes a hard condition filtering step: Before calculating the matching result between the full-dimensional business vector and the resume feature vector, extract the structured fields and hard constraints of the job position from the resume data; Construct a hard indicator function that contains a series of Boolean logic judgments; The hard index indicator function is used to determine whether the structured field meets the hard constraint conditions. If it does not meet the conditions, the subsequent calculation is terminated and the comprehensive matching score is set to zero.

6. The resume recommendation method based on a tag model of business organization tree according to claim 1, characterized in that, The steps for calculating the semantic matching result of business capabilities between the full-dimensional business vector and the resume feature vector include: Map the full-dimensional business vector and the resume feature vector to the same vector space; The cosine similarity algorithm is used to calculate the cosine value of the angle between two vectors in the vector space. The calculated cosine value of the included angle is used as the semantic matching result of the business capability, which is used to characterize the semantic similarity between the candidate's skill distribution and the job's business and technical requirements.

7. The resume recommendation method based on a tag model of business organization tree according to claim 1, characterized in that, The steps for calculating the business scenario fit matching result between the business scenario entities in the resume and the scenario tag set of the business node to which the job belongs include: Extract the business unit nodes and project nodes associated with the source of the full-dimensional business vector to form a set of job scenarios; Extract the business scenario entities contained in the resume feature vector to form a resume scenario set; The Jaccard similarity coefficient algorithm is used to calculate the ratio of the number of intersection elements to the number of union elements of the job scenario set and the resume scenario set, and the ratio is used as the business scenario fit matching result.

8. The resume recommendation method based on a tag model of business organization tree according to claim 1, characterized in that, The step of weightedly fusing the semantic matching result of the business capability and the matching result of the business scenario according to the preset weight parameters includes: Get the current weight parameter configuration, which includes a first weight parameter corresponding to the semantic matching result of business capabilities, and a second weight parameter corresponding to the matching result of business scenario fit. The semantic matching result of the business capability is multiplied by the first weight parameter, the matching result of the business scenario fit is multiplied by the second weight parameter, and the two are added together to obtain the comprehensive matching score.

9. A resume recommendation method based on a tag model of a business organization tree according to claim 8, characterized in that, It also includes a dynamic weight optimization step based on reinforcement learning: Define a state space, which at least includes the current job type characteristics and the current weight parameter configuration; Define a reward function that maps user actions on the resume recommendation list to numerical reward values. The actions include at least browsing, downloading, interview invitations, and rejection. Construct a state-action value function to evaluate the long-term expected return brought about by adjusting the weight parameter configuration under a preset state.

10. A resume recommendation method based on a tag model of a business organization tree according to claim 9, characterized in that, The reinforcement learning-based dynamic weight optimization step further includes: The system captures user actions in response to recommendation results in real time and calculates cumulative reward values ​​based on the reward function. The state-action value function is updated based on the cumulative reward value using a Q-learning algorithm or a deep Q-network strategy. The optimal weight adjustment action for the next moment is determined based on the updated state-action value function, a new weight parameter configuration is generated, and the new weight parameter configuration is applied to the subsequent weighted fusion step.

Images