Talent matching method and device based on dynamic post portrait, equipment and storage medium

Abstract

The invention relates to the technical field of artificial intelligence and human resource management, and discloses an intelligent talent matching method and device based on a dynamic post portrait, equipment and a storage medium, which are used for improving the accuracy of talent matching and the system adaptability. The intelligent talent matching method based on the dynamic post portraits comprises the steps of obtaining post description and candidate skill information, performing semantic enhancement processing on post texts through local and global attention mechanisms, and constructing the dynamic post portraits; performing similarity comparison on the post portraits and candidate skills, and screening out primary candidate subsets; then, precise matching degree prediction is achieved through deep interactive calculation, and a recommendation list is generated; user feedback data and industry trend information are continuously collected, a closed-loop learning mechanism is established, and dynamic adjustment and continuous optimization of post portraits are achieved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and human resource management technology, and in particular to a talent matching method, apparatus, equipment and storage medium based on dynamic job profiles. Background Technology

[0002] Currently, existing talent matching technologies mainly rely on static job descriptions. They typically build job models based on fixed content such as job titles, job descriptions, and qualifications, and use methods such as keyword matching, rule filtering, or traditional semantic similarity calculation to match people with jobs.

[0003] The static job descriptions of existing technologies are unable to respond in a timely manner to changes in business and the iteration of skill requirements, resulting in a disconnect between job profiles and actual needs. The descriptions are mostly limited to surface information and lack a three-dimensional presentation of the core competencies required for the job. The matching process relies on literal overlap calculations, making it difficult to deeply understand the intrinsic relationship between job requirements and talent capabilities. At the same time, existing technical solutions are insufficient in cross-domain adaptability, which limits their application value in industry-level talent analysis. Summary of the Invention

[0004] This invention provides a talent matching method, apparatus, device, and storage medium based on dynamic job profiles to solve the problems of low matching accuracy and poor adaptability caused by static job profiles, insufficient semantic understanding, and lack of evolutionary capabilities in the prior art.

[0005] The first aspect of this invention provides a talent matching method based on dynamic job profiles, comprising: acquiring job description text of a target job and skill information of candidates; parsing the job description text and performing semantic enhancement processing based on local attention and global attention mechanisms to generate a dynamic job profile; comparing the similarity between the dynamic job profile and the skill information of all candidates, and selecting a target candidate subset whose similarity exceeds a preset similarity threshold; calculating the target matching degree between each target candidate and the job by allowing the dynamic job profile to interact deeply with the skill information of each target candidate in the target candidate subset, and generating a recommendation list based on the target matching degree; collecting user feedback data on the recommendation list, and adjusting and updating the dynamic job profile based on the user feedback data and the external industry trend data.

[0006] In one feasible implementation, parsing the job description text and performing semantic enhancement processing based on local attention and global attention mechanisms to generate a dynamic job profile includes: encoding each responsibility text in the job description text using an item-level encoder to obtain multiple item-level representation vectors; extracting local features from the multiple item-level representation vectors based on a local self-attention mechanism to obtain a local attention representation; performing semantic enhancement based on the multiple item-level representation vectors and the neighboring job information of the target job through a global attention mechanism to obtain a global attention representation; fusing the local attention representation and the global attention representation, and processing them through a feedforward neural network to generate a dynamic job profile.

[0007] In one feasible implementation, the step of extracting local features from the plurality of item-level representation vectors based on a local self-attention mechanism to obtain a local attention representation includes: combining the plurality of item-level representation vectors to form an item-level representation vector matrix; calculating a query matrix, a key matrix, and a value matrix based on the item-level representation vector matrix; calculating the local attention weights between the query matrix and the key matrix using scaled dot product attention; and performing a weighted summation of the value matrix according to the local attention weights to obtain the local attention representation.

[0008] In one feasible implementation, the step of semantic enhancement based on the plurality of item-level representation vectors and the neighboring job information of the target job to obtain a global attention representation through a global attention mechanism includes: obtaining the item-level representation vectors of at least one neighboring job related to the target job; concatenating the plurality of item-level representation vectors of the target job with the item-level representation vectors of the at least one neighboring job to construct a global input matrix; generating a global query matrix, a global key matrix, and a global value matrix based on the global input matrix through trainable linear transformations; calculating the scaled dot product between the global query matrix and the transpose of the global key matrix, and applying the Softmax function for normalization to obtain a global attention weight matrix; and performing a weighted summation of the global value matrix according to the global attention weight matrix to obtain the global attention representation.

[0009] In one feasible implementation, fusing the local attention representation with the global attention representation and processing it through a feedforward neural network to generate a dynamic job profile includes: concatenating the local attention representation with the global attention representation to form a fused representation vector; inputting the fused representation vector into a feedforward neural network for nonlinear transformation and feature dimensionality reduction; and encoding the output of the feedforward neural network to generate the dynamic job profile in vector form.

[0010] In one feasible implementation, the step of comparing the dynamic job profile with the skill information of all candidates to filter out a target candidate subset whose similarity exceeds a preset similarity threshold includes: calculating the cosine similarity between the skill distribution vector predicted by the dynamic job profile and the skill distribution vector of each candidate; comparing the cosine similarity with a preset similarity threshold; and filtering out candidates whose similarity is greater than the preset similarity threshold to form a target candidate subset.

[0011] In one feasible implementation, the step of calculating the target matching degree between each target candidate and the job by deeply interacting with the skill information of each target candidate in the target candidate subset, and generating a recommendation list based on the target matching degree, includes: taking the vector representation of the dynamic job profile and the skill vector representation of the target candidate as input; calculating the interaction representation between the two using a cross-attention mechanism; inputting the interaction representation into a fully connected layer for matching degree prediction; outputting a set of target matching degree probability values ​​through a Sigmoid function; and sorting all target candidates in descending order of the target matching degree probability value set to generate a recommendation list.

[0012] A second aspect of the present invention provides a talent matching device based on dynamic job profiles, comprising: an acquisition module for acquiring job description text of a target job and skill information of candidates; a first generation module for parsing the job description text and performing semantic enhancement processing based on local attention and global attention mechanisms to generate a dynamic job profile; a filtering module for comparing the similarity between the dynamic job profile and the skill information of all candidates, and filtering out a target candidate subset whose similarity exceeds a preset similarity threshold; a second generation module for calculating the target matching degree between each target candidate and the job by allowing the dynamic job profile to interact deeply with the skill information of each target candidate in the target candidate subset, and generating a recommendation list based on the target matching degree; and a processing module for collecting user feedback data on the recommendation list, and adjusting and updating the dynamic job profile based on the user feedback data and the external industry trend data.

[0013] In one feasible implementation, the first generation module includes: an encoding unit, configured to encode each responsibility text in the job description text using an item-level encoder to obtain multiple item-level representation vectors; an extraction unit, configured to extract local features from the multiple item-level representation vectors based on a local self-attention mechanism to obtain a local attention representation; an enhancement unit, configured to perform semantic enhancement based on the multiple item-level representation vectors and the neighboring job information of the target job through a global attention mechanism to obtain a global attention representation; and a generation unit, configured to fuse the local attention representation and the global attention representation, and process them through a feedforward neural network to generate a dynamic job profile.

[0014] In one feasible implementation, the extraction unit is specifically used to: combine the plurality of item-level representation vectors to form an item-level representation vector matrix; calculate the query matrix, key matrix, and value matrix based on the item-level representation vector matrix; calculate the local attention weights between the query matrix and the key matrix by scaling dot product attention; and perform weighted summation on the value matrix according to the local attention weights to obtain the local attention representation.

[0015] In one feasible implementation, the enhancement unit is specifically used to: obtain the item-level representation vectors of at least one neighboring job related to the target job; concatenate multiple item-level representation vectors of the target job with the item-level representation vectors of the at least one neighboring job to construct a global input matrix; based on the global input matrix, generate a global query matrix, a global key matrix, and a global value matrix through trainable linear transformations; calculate the scaled dot product between the global query matrix and the transpose of the global key matrix, and normalize it using the Softmax function to obtain a global attention weight matrix; and perform a weighted summation of the global value matrix according to the global attention weight matrix to obtain the global attention representation.

[0016] In one feasible implementation, the generation unit is specifically used to: concatenate the local attention representation and the global attention representation to form a fused representation vector; input the fused representation vector into a feedforward neural network for nonlinear transformation and feature dimensionality reduction; and encode the output of the feedforward neural network to generate the dynamic job profile in vector form.

[0017] In one feasible implementation, the screening module is specifically used to: calculate the cosine similarity between the skill distribution vector predicted by the dynamic job profile and the skill distribution vector of each candidate; compare the cosine similarity with a preset similarity threshold; and screen out candidates whose similarity is greater than the preset similarity threshold to form a target candidate subset.

[0018] In one feasible implementation, the second generation module is specifically used to: take the vector representation of the dynamic job profile and the skill vector representation of the target candidate as input; calculate the interaction representation between the two using a cross-attention mechanism; input the interaction representation into a fully connected layer for matching degree prediction; output a set of target matching degree probability values ​​through a Sigmoid function; and sort all target candidates in descending order of the target matching degree probability value set to generate a recommendation list.

[0019] A third aspect of the present invention provides an electronic device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the electronic device to perform the electronic method described above.

[0020] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the aforementioned talent matching method based on dynamic job profiles.

[0021] The technical solution provided by this invention involves: acquiring the job description text of the target position and the skill information of the candidates; parsing the job description text and performing semantic enhancement processing based on local attention and global attention mechanisms to generate a dynamic job profile; comparing the similarity between the dynamic job profile and the skill information of all candidates to filter out a target candidate subset whose similarity exceeds a preset similarity threshold; calculating the target matching degree between each target candidate and the job by allowing the dynamic job profile to interact deeply with the skill information of each target candidate in the target candidate subset, and generating a recommendation list based on the target matching degree; collecting user feedback data on the recommendation list and combining it with external industry trend data to adjust and update the dynamic job profile based on the user feedback data and the external industry trend data.

[0022] In this embodiment of the invention, by introducing local and global attention mechanisms to perform deep semantic parsing and enhancement of job descriptions, a job profile that can accurately depict multi-dimensional responsibilities and dynamic needs is constructed. A two-stage matching strategy of "preliminary screening + in-depth interaction" is adopted, which significantly improves the accuracy and efficiency of person-job matching. At the same time, the profile model can be continuously and automatically updated based on user feedback and external industry trend data, effectively overcoming the shortcomings of traditional static job descriptions, such as information lag and single dimension. This provides a quantifiable, adaptive, and efficient solution for enterprise talent management and industry intelligent analysis. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of an embodiment of the talent matching method based on dynamic job profiles in this invention; Figure 2 This is a schematic diagram of another embodiment of the talent matching method based on dynamic job profiles in this invention; Figure 3 This is a schematic diagram of one embodiment of the talent matching device based on dynamic job profiles in this invention. Figure 4 This is a schematic diagram of another embodiment of the talent matching device based on dynamic job profiles in this invention. Figure 5 This is a schematic diagram of one embodiment of the electronic device in this invention. Detailed Implementation

[0024] This invention provides a talent matching method, apparatus, device, and storage medium based on dynamic job profiles. By constructing dynamic job profiles that integrate local semantic features and global contextual information, and establishing a continuous evolution mechanism based on multi-source feedback, the accuracy of job matching, the system's adaptability to changes in business needs, and the generalization ability of cross-domain applications are improved.

[0025] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0026] It is understood that the executing entity of this invention can be a talent matching device based on dynamic job profiles, or it can be a terminal or a server; the specific implementation is not limited here. This embodiment of the invention will be described using a server as an example.

[0027] It should be noted that all information acquisition under this plan shall be conducted only with the explicit authorization of the user and in strict compliance with data privacy and security regulations.

[0028] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the talent matching method based on dynamic job profiles in this invention includes: 101. Obtain the job description text for the target position and the candidate's skill information; We obtain job description texts for target positions from internal or external recruitment platforms, and collect candidates' skill information, including skill tags, years of experience, project experience, etc. The job description texts usually contain structured or unstructured content such as job descriptions, skill requirements, and experience requirements. Candidate skill information is obtained through resume parsing, skill tag databases, etc., and is uniformly represented as a skill distribution vector.

[0029] 102. Parse the job description text and perform semantic enhancement processing based on local attention mechanism and global attention mechanism to generate dynamic job profile; Each responsibility or requirement in the job description text is encoded using a project-level encoder (such as TextCNN) to obtain an item-level vector representation. Then, a local self-attention mechanism is used, specifically applied to the target job's own set of item-level vectors. By calculating the association strength between each item and all other items in the set, the internal dependencies and local semantic structure among the responsibility requirements are captured, and a locally enhanced item-level representation is output. A global attention mechanism is then introduced. Its input is a hybrid set of item-level vectors, directly concatenated from the original item-level vectors obtained after project-level encoding of the target job and the item-level vectors obtained from the same encoding of its semantically neighboring jobs. Global attention calculates attention on this hybrid set, enabling each item of the target job to interact with items of all neighboring jobs, thereby capturing macroscopic semantic relationships in the cross-job context. The outputs of the local self-attention mechanism and the global attention mechanism are then concatenated, and this fused vector is fed into a feedforward neural network for nonlinear transformation and feature compression. Finally, a global job representation vector with both deep local details and broad contextual information is generated—a dynamic job profile.

[0030] Project-level coding is performed according to the following formula: m=1,2,...,M

[0031] in, This refers to the m-th responsibility or requirement in the job description text. Let M be the d-dimensional vector representation of the m-th item; M is the total number of items in the job description text; R is the set of real numbers. Let be a d-dimensional real vector space; d is the fixed dimension of the word vector.

[0032] Local self-attention mechanism: , ,

[0033]

[0034] , is a local input matrix, obtained by concatenating all term-level vectors; Query matrix; The key matrix; It is a value matrix; A trainable weight matrix used to generate query vectors; A trainable weight matrix used to generate key vectors; A trainable weight matrix used to generate value vectors; This represents the number of heads for local attention.

[0035] In this local self-attention mechanism, the query matrix Key matrix Value matrix All originate from the same input matrix This is achieved through different linear transformations. This mechanism allows the model to dynamically aggregate information based on the degree of association between each item and all other items in the set (including itself), thereby capturing local dependencies between responsibility items.

[0036] Global attention mechanism formula: , ,

[0037]

[0038] The global input matrix is ​​composed of the M item-level representation vectors of the target job and the item-level representation vectors of all k neighboring jobs, with dimension M. ; Query matrix; The key matrix; It is a value matrix; A trainable weight matrix used to generate query vectors; A trainable weight matrix used to generate key vectors; A trainable weight matrix used to generate value vectors; This represents the number of global attention heads.

[0039] 103. Compare the dynamic job profile with the skill information of all candidates to select a target candidate subset whose similarity exceeds the preset similarity threshold; The global representation vector of the job position is mapped to a job skill distribution vector through a fully connected layer and a Softmax function. This job skill distribution vector represents the probability of the job requiring all predefined skills. At the same time, the skill information of each candidate is standardized into a skill distribution vector, where each dimension represents the candidate's mastery of the corresponding skill. The cosine similarity between the job skill distribution vector and the skill distribution vectors of all candidates is calculated to form a similarity list. The similarity list is filtered according to a preset similarity threshold, and all candidates with similarity below the similarity threshold are removed. A target candidate subset is generated based on all remaining candidates.

[0040] 104. By enabling deep interaction between the dynamic job profile and the skill information of each target candidate in the target candidate subset, the target matching degree between each target candidate and the job is calculated, and a recommendation list is generated based on the target matching degree. A deep interactive representation of dynamic job profiles and candidate profiles is constructed using a cross-attention mechanism. The global representation vector of the job is used as the query vector, and the skill distribution vector of the candidate is used as the key vector and value vector, respectively. The joint representation vector is calculated and then input into a deep ranking model containing a fully connected layer and a sigmoid activation function to predict the final matching probability between each target candidate and the job. Based on the calculated matching probability, all target candidates are sorted in descending order to generate the final job recommendation list.

[0041] The formula for representing cross-attention interaction is: , ,

[0042]

[0043] in, This is the global representation vector for the job position; Let be the skill distribution vector of the i-th candidate. , and These represent the query vector, key vector, and value vector, respectively. , and It is a trainable linear transformation matrix; This is a temperature coefficient used to adjust the smoothness of attention distribution; This is the joint representation vector of the job position and the candidate.

[0044] The matching degree prediction formula is:

[0045] in, Indicates a fully connected layer. Use the Sigmoid activation function; This represents the predicted probability of a candidate matching a job position.

[0046] The formula for the ranking loss function is:

[0047] This represents the total loss during the sorting phase. For positive sample combinations, include known successfully matched job-candidate pairs; For negative sample combinations, there are known unsuccessful job-candidate pairs; j represents the job; u represents the candidate.

[0048] 105. Collect user feedback data on the recommended list and combine it with external industry trend data to adjust and update the dynamic job profile based on the user feedback data and external industry trend data.

[0049] By collecting real-time user interaction feedback data on the recommendation list, including multi-dimensional behavioral signals such as clicks, applications, interviews, and onboarding, and integrating external industry trend data, a feedback vector based on time decay weight is constructed. An incremental update strategy is adopted to weight and fuse the feedback vector with the industry trend vector. The dynamic job profile vector is iteratively optimized through a dynamic adjustment formula to achieve real-time evolution of the dynamic job profile.

[0050] The formula for updating dynamic job profiles is:

[0051] in, The profile vector of job j at time t; This is the amount of correction based on user feedback; This is a correction based on industry trends; and These are the update weight coefficients for feedback and trend, respectively.

[0052] The formula for calculating the feedback correction amount is:

[0053] in, For individual user feedback data (such as clicks, applications, interviews, onboarding, etc.); For feedback encoders; The weights are based on the feedback type and timestamp; N is the total number of feedback data.

[0054] The formula for calculating industry trend correction is:

[0055] in, Industry trend data (such as industry skills demand reports, technology development information, and market recruitment hotspots) collected within a time period t. For industry trend encoders; This is the global adjustment coefficient for the impact of industry trends.

[0056] In this embodiment of the invention, by introducing local and global attention mechanisms to perform deep semantic parsing and enhancement of job descriptions, a job profile that can accurately depict multi-dimensional responsibilities and dynamic needs is constructed. A two-stage matching strategy of "preliminary screening + in-depth interaction" is adopted, which significantly improves the accuracy and efficiency of person-job matching. At the same time, the profile model can be continuously and automatically updated based on user feedback and external industry trend data, effectively overcoming the shortcomings of traditional static job descriptions, such as information lag and single dimension. This provides a quantifiable, adaptive, and efficient solution for enterprise talent management and industry intelligent analysis.

[0057] Please see Figure 2 Another embodiment of the talent matching method based on dynamic job profiles in this invention includes: 201. Obtain the job description text for the target position and the candidate's skill information; To obtain the job description text for the target position, the system primarily accesses the company's internal recruitment management system or human resources information system via an Application Programming Interface (API). This API automatically extracts structured job data, including core job responsibilities, hard skill requirements (such as programming languages ​​and certifications), soft skill requirements (such as communication, collaboration, and project management), and educational and experience thresholds. Simultaneously, to acquire candidate skill information, a standardized resume parsing service processes the resumes submitted by candidates. This service utilizes natural language processing technologies such as named entity recognition and semantic role labeling to extract key information from the document, including a list of specific technical skills, proficiency levels, project experience using those skills, and relevant professional certifications. This information is then uniformly converted into a structured skill vector.

[0058] 202. Use a project-level encoder to encode each responsibility text in the job description text to obtain multiple item-level representation vectors; A hierarchical encoding method based on semantic unit separation is employed to vectorize and represent diverse information such as core job responsibilities, hard skills, soft skills, and educational background and experience. Multi-granularity text parsing technology automatically breaks down complete job descriptions into responsibility units with independent semantics. For example, descriptions of different dimensions, such as system architecture design requirements and technology stack requirements, are separated into discrete text fragments. Each semantic unit is then input into a pre-trained language model based on the Transformer architecture for encoding. This model preferably uses BERT (Bidirectional Encoder Representations from Transformers) or its variants (e.g., RoBERTa, ERNIE) as its basic architecture and undergoes further domain-adaptive pre-training on large-scale professional domain text corpora (e.g., job descriptions, technical documents, professional resumes). This training process uses a masked language model and a next-sentence prediction task, enabling the model to deeply understand the context of professional terms such as "microservice architecture" and "concurrent programming." The model uses a deep bidirectional attention mechanism to parse the semantic relationships between technical terms and competency requirements in the text, and particularly strengthens the semantic representation ability of professional vocabulary during the encoding process. For hard skills descriptions, the model focuses on capturing technical features such as programming languages ​​and tool certifications; for soft skills requirements, it enhances the vector representation of abstract concepts such as communication and collaboration through sentiment semantic analysis. After feature extraction and fusion of each semantic unit through a twelve-layer neural network, a corresponding 768-dimensional dense vector is finally generated. These vectors not only preserve the semantic integrity of the original text but also incorporate professional knowledge from the occupational field, ultimately forming a set of item-level representation vectors covering all the requirements of the job.

[0059] 203. Based on the local self-attention mechanism, local features are extracted from multiple item-level representation vectors to obtain local attention representations; Multiple item-level representation vectors are combined to form an item-level representation vector matrix; based on the item-level representation vector matrix, the query matrix, key matrix, and value matrix are calculated respectively; the local attention weights between the query matrix and the key matrix are calculated by scaling dot product attention; the value matrix is ​​weighted and summed according to the local attention weights to obtain the local attention representation.

[0060] The n item-level representation vectors of dimension 768 are stacked row-wise to form an n×768-dimensional item-level representation vector matrix X. Then, X is projected onto the query space, key space, and value space through three independent linear transformation layers, respectively, to obtain the query matrix Q1, key matrix K1, and value matrix V1. The output dimension of each transformation layer remains n×64, forming the low-dimensional representation required for the eight attention heads. When calculating the attention weights, a scaled dot product attention mechanism is used. The query matrix Q1 is multiplied by the transpose of the key matrix K1, and the result is divided by the square root of 8 for numerical stability adjustment. Then, the matrix is ​​normalized using the Softmax function to obtain an n×n local attention weight matrix, which accurately reflects the semantic association strength between each responsibility item. Finally, the value matrix V1 is weighted and fused based on the attention weights. The weighted attention output is obtained by multiplying the weight matrix and the value matrix V1. The outputs of the eight attention heads are then concatenated and linearly transformed to restore the n×768-dimensional matrix representation. The final output local attention representation not only preserves the semantic information of the original item-level vector, but more importantly, it captures the intrinsic relationship between different responsibility requirements through the self-attention mechanism, so that the originally independent item-level representations such as architectural design requirements and programming skill requirements form semantic synergy and enhancement in the vector space.

[0061] 204. Based on multiple item-level representation vectors and neighboring job information of the target job, semantic enhancement is performed through a global attention mechanism to obtain a global attention representation; Obtain the item-level representation vectors of at least one neighboring job related to the target job; concatenate multiple item-level representation vectors of the target job with the item-level representation vectors of at least one neighboring job to construct a global input matrix; based on the global input matrix, generate a global query matrix, a global key matrix, and a global value matrix through trainable linear transformations; calculate the scaled dot product between the transpose of the global query matrix and the global key matrix, and apply the Softmax function for normalization to obtain the global attention weight matrix; perform a weighted summation of the global value matrix based on the global attention weight matrix to obtain the global attention representation.

[0062] A multi-dimensional feature space is constructed based on job responsibility vectors, skill requirement vectors, and departmental affiliation. An approximate nearest neighbor search algorithm is used to retrieve the k most similar neighboring jobs from the company's full job database. These neighboring jobs typically share core skill requirements or operate in the same technical field as the target job. Next, the item-level representation vectors of these neighboring jobs are concatenated with the item-level representation vector of the target job itself to construct a global input matrix. Assuming the target job has m responsibility items and each neighboring job has an average of n responsibility items, the system will form a (m + k * n) × 768-dimensional global input matrix, which fully preserves the fine-grained requirements of the target job and the contextual information of related external jobs. Subsequently, the global input matrix is ​​projected onto the query, key, and value spaces respectively through three independent trainable linear transformation layers, generating a global query matrix Q2, a global key matrix K2, and a global value matrix V2. During the projection process, a dimensionality reduction strategy is used to map the original 768-dimensional vector to a 64-dimensional attention subspace, reducing computational complexity while preserving core semantic features. In the attention calculation phase, the product of the global query matrix and the transpose of the global key matrix is ​​first calculated, then the product is scaled by dividing by the square root of 8, and then normalized using the Softmax function to obtain a (m + k * n) × (m + k * n) global attention weight matrix. This weight matrix precisely quantifies the semantic association strength between each responsibility item of the target job and all neighboring job responsibilities. For example, system architecture design requirements may have a high-weight association with the "distributed system development" responsibility of a neighboring job, while the weight with the "user interface design" responsibility is relatively low. Finally, the global value matrix is ​​weighted and summed based on the global attention weight matrix to generate the final global attention representation. This output matrix not only strengthens the features of the target job responsibilities that are highly relevant to the external environment, but also introduces cross-job semantic information, enabling the generated job profile to reflect broader industry practices and technological trends. The entire calculation process uses a multi-head attention mechanism for parallel processing. Eight independent attention heads capture different types of cross-job dependencies, and finally, the outputs of each head are concatenated and linearly transformed to form a unified global attention representation, providing rich contextual semantic information for subsequent profile fusion.

[0063] 205. Fuse local attention representations with global attention representations and process them through a feedforward neural network to generate dynamic job profiles; The local attention representation and the global attention representation are concatenated to form a fused representation vector; the fused representation vector is input into a feedforward neural network for nonlinear transformation and feature dimensionality reduction; the output of the feedforward neural network is encoded to generate a dynamic job profile in vector form.

[0064] To effectively integrate with the local attention representation (n × 768 dimensions) derived from the target job itself, a dimension alignment operation is required. This operation is based on a learnable attention pooling layer: First, a learnable query vector corresponding to the number of responsibilities n of the target job is generated; then, this query vector is computed with the global attention representation matrix, and a (m + k * n) × 1 dimensional attention weight vector is generated through the Softmax function; finally, the original (m + k * n) × 768 dimensional global representation is weighted and summed according to this weight vector, compressing it into a compact n × 768 dimensional global semantic summary.

[0065] The local attention representation (n×768-dimensional) and the global attention representation (n×768-dimensional) are concatenated along the channel dimension to form an n×1536-dimensional fusion matrix, where each responsibility item simultaneously contains local fine-grained features and global contextual features. This fusion matrix is ​​then fed into a two-layer feedforward neural network with residual connections: the first layer uses a fully connected transformation from 1536 to 1024 dimensions combined with the GELU activation function to capture high-order nonlinear interactions between responsibility items; the second layer performs feature compression from 1024 to 768 dimensions and uses LayerNorm layers to stabilize the training process. The network output is processed by an attention pooling layer, and the weights of each responsibility item are calculated using a learnable query vector. The 768-dimensional item-level representation is then weighted and summed to finally generate a 768-dimensional dynamic profile vector representing the entire job position, i.e., a dynamic job profile. This dynamic job profile retains specific skill requirements such as "Java development" and "system architecture" while also incorporating emerging capabilities such as "cloud-native architecture" brought about by industry technology trends, forming a semantic representation that combines job characteristics with industry adaptability.

[0066] 206. Compare the dynamic job profile with the skill information of all candidates to select a target candidate subset whose similarity exceeds the preset similarity threshold. Calculate the cosine similarity between the skill distribution vector predicted by the dynamic job profile and the skill distribution vector of each candidate; compare the cosine similarity with a preset similarity threshold; select candidates with similarity greater than the preset similarity threshold to form a target candidate subset.

[0067] The dynamic job profile vector and the pre-processed skill vector of each candidate are standardized to ensure that all vectors are in the same metric space. The cosine similarity algorithm is used to calculate the pairwise matching degree. This algorithm evaluates the semantic similarity by calculating the consistency of the two vectors in direction, effectively avoiding the bias caused by vector length. In the specific calculation process, the dot product of the two vectors is divided by the product of their magnitudes to obtain a similarity score ranging from negative one to positive one. The similarity threshold is set according to the actual situation, for example, set to 0.6. When the cosine similarity between the candidate's skill vector and the job profile is greater than the similarity threshold, the candidate will be automatically marked as a potential match. Finally, all marked candidates are sorted from high to low according to their similarity scores to form a target candidate subset.

[0068] 207. By enabling deep interaction between the dynamic job profile and the skill information of each target candidate in the target candidate subset, the target matching degree between each target candidate and the job is calculated, and a recommendation list is generated based on the target matching degree. The dynamic job profile vector representation and the target candidate skill vector representation are used as inputs; a cross-attention mechanism is used to calculate the interaction representation between the two; the interaction representation is input to a fully connected layer for matching degree prediction; the target matching degree probability value set is output through the Sigmoid function; all target candidates are sorted in descending order of the target matching degree probability value set to generate a recommendation list.

[0069] Using dynamic job profile vectors and target candidate skill vectors as input, a deep interaction model is established through a cross-attention mechanism. First, the job profile vector is used as the query vector, and the target candidate's skill vector is used as the key-value pair. A weighted interaction representation is calculated through a multi-head attention layer, which integrates the correlation features between job requirements and target candidate abilities. This interaction representation is then input into a three-layer fully connected neural network. Each layer uses the ReLU activation function for non-linear transformation. The final output node is converted into a matching probability value between 0 and 1 using the Sigmoid function. This process is repeated for all target candidates to generate a corresponding matching probability set. Finally, the candidates are sorted from highest to lowest probability value to form the final talent recommendation list.

[0070] 208. Collect user feedback data on the recommended list and combine it with external industry trend data to adjust and update the dynamic job profile based on the user feedback data and external industry trend data.

[0071] The system collects implicit and explicit user feedback data on the recommendation list, including behavioral data such as hiring managers' clicks on candidates, interview invitations, and hiring decisions. It also obtains industry skill demand trends from authoritative talent market reports via API. This multi-source data, after standardization, is input into the incremental learning module along with the original dynamic job profile. This module employs a contrastive learning objective function based on negative sampling, fine-tuning the encoder parameters to bring the job profile vector closer to the skill distribution of candidates receiving positive feedback, and simultaneously adjusting to reflect the market weight of emerging skills. The entire update process is executed automatically while protecting data privacy, ensuring that the dynamic job profile continuously adapts to changes in organizational hiring needs and the market environment.

[0072] In this embodiment of the invention, a dual semantic enhancement strategy combining local and global attention is employed. This strategy not only deeply analyzes the inherent relationships between job responsibilities but also incorporates cross-job industry contextual information, thereby constructing a dynamic job profile that accurately depicts the core competency requirements of the job while also possessing an industry perspective. Based on this, a two-stage matching process of "recall-ranking" and a deep interaction model based on cross-attention are used to accurately assess the match between talent and job at the semantic level. Finally, a closed-loop learning mechanism based on online feedback and industry trend data enables the job profile to continuously adapt to changes in business needs and skill iterations, significantly improving the accuracy, timeliness, and robustness of human-job matching across industries.

[0073] The above describes the talent matching method based on dynamic job profiles in the embodiments of the present invention. The following describes the talent matching device based on dynamic job profiles in the embodiments of the present invention. Please refer to [link / reference]. Figure 3 One embodiment of the talent matching device based on dynamic job profiles in this invention includes: The acquisition module 301 is used to acquire the job description text of the target position and the candidate's skill information; The first generation module 302 is used to parse the job description text and perform semantic enhancement processing based on local attention mechanism and global attention mechanism to generate dynamic job profile; The filtering module 303 is used to compare the similarity between the dynamic job profile and the skill information of all candidates, and filter out the target candidate subset whose similarity exceeds the preset similarity threshold. The second generation module 304 is used to calculate the target matching degree between each target candidate and the job by allowing the dynamic job profile to interact deeply with the skill information of each target candidate in the target candidate subset, and to generate a recommendation list based on the target matching degree. The processing module 305 is used to collect user feedback data on the recommendation list and combine it with external industry trend data to adjust and update the dynamic job profile based on the user feedback data and external industry trend data.

[0074] In this embodiment of the invention, by introducing local and global attention mechanisms to perform deep semantic parsing and enhancement of job descriptions, a job profile that can accurately depict multi-dimensional responsibilities and dynamic needs is constructed. A two-stage matching strategy of "preliminary screening + in-depth interaction" is adopted, which significantly improves the accuracy and efficiency of person-job matching. At the same time, the profile model can be continuously and automatically updated based on user feedback and external industry trend data, effectively overcoming the shortcomings of traditional static job descriptions, such as information lag and single dimension. This provides a quantifiable, adaptive, and efficient solution for enterprise talent management and industry intelligent analysis.

[0075] Please see Figure 4 Another embodiment of the talent matching device based on dynamic job profiles in this invention includes: The acquisition module 301 is used to acquire the job description text of the target position and the candidate's skill information; The first generation module 302 is used to parse the job description text and perform semantic enhancement processing based on local attention mechanism and global attention mechanism to generate dynamic job profile; The filtering module 303 is used to compare the similarity between the dynamic job profile and the skill information of all candidates, and filter out the target candidate subset whose similarity exceeds the preset similarity threshold. The second generation module 304 is used to calculate the target matching degree between each target candidate and the job by allowing the dynamic job profile to interact deeply with the skill information of each target candidate in the target candidate subset, and to generate a recommendation list based on the target matching degree. The processing module 305 is used to collect user feedback data on the recommendation list and combine it with external industry trend data to adjust and update the dynamic job profile based on the user feedback data and external industry trend data.

[0076] Optionally, the first generation module 302 includes: Encoding unit 3021 is used to encode each responsibility text in the job description text using a project-level encoder to obtain multiple item-level representation vectors; Extraction unit 3022 is used to extract local features from multiple item-level representation vectors based on a local self-attention mechanism to obtain local attention representations; Enhancement unit 3023 is used to perform semantic enhancement based on multiple item-level representation vectors and neighboring job information of the target job through a global attention mechanism to obtain a global attention representation; The generation unit 3024 is used to fuse the local attention representation with the global attention representation and process it through a feedforward neural network to generate a dynamic job profile.

[0077] Optionally, the extraction unit 3022 can be specifically used for: Multiple item-level representation vectors are combined to form an item-level representation vector matrix; based on the item-level representation vector matrix, the query matrix, key matrix, and value matrix are calculated respectively; the local attention weights between the query matrix and the key matrix are calculated by scaling dot product attention; the value matrix is ​​weighted and summed according to the local attention weights to obtain the local attention representation.

[0078] Optionally, the enhancement unit 3023 can be specifically used for: Obtain the item-level representation vectors of at least one neighboring job related to the target job; concatenate the multiple item-level representation vectors of the target job with the item-level representation vectors of at least one neighboring job to construct a global input matrix; based on the global input matrix, generate a global query matrix, a global key matrix, and a global value matrix through trainable linear transformations; calculate the scaled dot product between the transpose of the global query matrix and the global key matrix, and apply the Softmax function for normalization to obtain the global attention weight matrix; The global attention representation is obtained by weighted summation of the global value matrix based on the global attention weight matrix. Optionally, the generation unit 3024 can be specifically used for: The local attention representation and the global attention representation are concatenated to form a fused representation vector; the fused representation vector is input into a feedforward neural network for nonlinear transformation and feature dimensionality reduction; the output of the feedforward neural network is encoded to generate a dynamic job profile in vector form.

[0079] Optionally, the filtering module 303 can be specifically used for: Calculate the cosine similarity between the skill distribution vector predicted by the dynamic job profile and the skill distribution vector of each candidate; compare the cosine similarity with a preset similarity threshold; select candidates with similarity greater than the preset similarity threshold to form a target candidate subset.

[0080] Optionally, the second generation module can be specifically used for: The dynamic job profile vector representation and the target candidate skill vector representation are used as inputs; a cross-attention mechanism is used to calculate the interaction representation between the two; the interaction representation is input to a fully connected layer for matching degree prediction; the target matching degree probability value set is output through the Sigmoid function; all target candidates are sorted in descending order of the target matching degree probability value set to generate a recommendation list.

[0081] In this embodiment of the invention, a dual semantic enhancement strategy combining local and global attention is employed. This strategy not only deeply analyzes the inherent relationships between job responsibilities but also incorporates cross-job industry contextual information, thereby constructing a dynamic job profile that accurately depicts the core competency requirements of the job while also possessing an industry perspective. Based on this, a two-stage matching process of "recall-ranking" and a deep interaction model based on cross-attention are used to accurately assess the match between talent and job at the semantic level. Finally, a closed-loop learning mechanism based on online feedback and industry trend data enables the job profile to continuously adapt to changes in business needs and skill iterations, significantly improving the accuracy, timeliness, and robustness of human-job matching across industries.

[0082] above Figure 3 and Figure 4 The talent matching device based on dynamic job profiles in this embodiment of the invention will be described in detail from the perspective of modular functional entities. The electronic equipment in this embodiment of the invention will be described in detail from the perspective of hardware processing.

[0083] See Figure 5 As shown, the electronic device includes a processor 500 and a memory 501. The memory 501 stores machine-executable instructions that can be executed by the processor 500. The processor 500 executes the machine-executable instructions to implement the aforementioned talent matching method based on dynamic job profiles.

[0084] Furthermore, Figure 5 The electronic device shown also includes a bus 502 and a communication interface 503. The processor 500, the communication interface 503 and the memory 501 are connected via the bus 502.

[0085] The memory 501 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 503 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 502 may be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0086] The processor 500 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 500 or by instructions in software form. The processor 500 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this disclosure. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this disclosure can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules may reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory 501. The processor 500 reads the information in memory 501 and, in conjunction with its hardware, completes the method steps of the aforementioned embodiment.

[0087] The present invention also provides an electronic device, the computer device including a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor performs the steps of the talent matching method based on dynamic job profiles described in the above embodiments. The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions, and when the instructions are executed on a computer, the computer performs the steps of the talent matching method based on dynamic job profiles.

[0088] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0089] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0090] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A talent matching method based on dynamic job profiles, characterized in that, The talent matching method based on dynamic job profiles includes: Obtain the job description text for the target position and the candidate's skills information; The job description text is parsed, and semantic enhancement processing is performed based on local attention mechanism and global attention mechanism to generate dynamic job profile; The dynamic job profile is compared with the skill information of all candidates to filter out the target candidate subset whose similarity exceeds a preset similarity threshold. By enabling the dynamic job profile to interact deeply with the skill information of each target candidate in the target candidate subset, the target matching degree between each target candidate and the job is calculated, and a recommendation list is generated based on the target matching degree. Collect user feedback data on the recommended list and combine it with external industry trend data to adjust and update the dynamic job profile.

2. The talent matching method based on dynamic job profiles according to claim 1, characterized in that, The process of parsing the job description text and performing semantic enhancement based on local and global attention mechanisms to generate a dynamic job profile includes: Each responsibility text in the job description text is encoded using a project-level encoder to obtain multiple item-level representation vectors; Local feature extraction is performed on the multiple term-level representation vectors based on a local self-attention mechanism to obtain a local attention representation; Based on the multiple item-level representation vectors and the neighboring job information of the target job, semantic enhancement is performed through a global attention mechanism to obtain a global attention representation; The local attention representation is fused with the global attention representation and processed by a feedforward neural network to generate a dynamic job profile.

3. The talent matching method based on dynamic job profiles according to claim 2, characterized in that, The local attention representation is obtained by extracting local features from the multiple term-level representation vectors based on the local self-attention mechanism, including: The multiple term-level representation vectors are combined to form a term-level representation vector matrix; Based on the item-level representation vector matrix, calculate the query matrix, key matrix, and value matrix respectively; Calculate the local attention weights between the query matrix and the key matrix by scaling the dot product attention; The local attention representation is obtained by weighting and summing the value matrix according to the local attention weights.

4. The talent matching method based on dynamic job profiles according to claim 2, characterized in that, The semantic enhancement based on the multiple item-level representation vectors and the neighboring job information of the target job is performed through a global attention mechanism to obtain a global attention representation, including: Obtain the item-level representation vector of at least one neighboring job related to the target job; The multiple item-level representation vectors of the target position are concatenated with the item-level representation vectors of at least one neighboring position to construct a global input matrix; Based on the global input matrix, a global query matrix, a global key matrix, and a global value matrix are generated through trainable linear transformations. Calculate the scaled dot product between the global query matrix and the transpose of the global key matrix, and normalize it using the Softmax function to obtain the global attention weight matrix; The global attention representation is obtained by weighted summation of the global value matrix based on the global attention weight matrix.

5. The talent matching method based on dynamic job profiles according to claim 1, characterized in that, The process of fusing the local attention representation with the global attention representation and processing it through a feedforward neural network to generate a dynamic job profile includes: The local attention representation is concatenated with the global attention representation to form a fused representation vector; The fused representation vector is input into a feedforward neural network for nonlinear transformation and feature dimensionality reduction. The output of the feedforward neural network is encoded to generate the dynamic job profile in vector form.

6. The talent matching method based on dynamic job profiles according to claim 1, characterized in that, The step of comparing the dynamic job profile with the skill information of all candidates to filter out a target candidate subset whose similarity exceeds a preset similarity threshold includes: Calculate the cosine similarity between the skill distribution vector predicted by the dynamic job profile and the skill distribution vector of each candidate; The cosine similarity is compared with a preset similarity threshold; Candidates with a similarity greater than the preset similarity threshold are selected to form a target candidate subset.

7. The talent matching method based on dynamic job profiles according to claim 1, characterized in that, The process involves deep interaction between the dynamic job profile and the skill information of each target candidate in the target candidate subset to calculate the target matching degree between each target candidate and the job, and generating a recommendation list based on the target matching degree, including: The vector representation of the dynamic job profile and the skill vector representation of the target candidate are used as inputs; A cross-attention mechanism is used to compute the interaction representation between the two. The interaction representation is input into a fully connected layer for matching degree prediction; The Sigmoid function outputs a set of target matching probability values. All target candidates are sorted from high to low according to the set of target matching probability values ​​to generate a recommendation list.

8. A talent matching device based on dynamic job profiles, characterized in that, The talent matching device based on dynamic job profiles includes: The acquisition module is used to acquire the job description text of the target position and the candidate's skill information; The first generation module is used to parse the job description text and perform semantic enhancement processing based on local attention mechanism and global attention mechanism to generate dynamic job profile. The filtering module is used to compare the dynamic job profile with the skill information of all candidates and filter out the target candidate subset whose similarity exceeds a preset similarity threshold. The second generation module is used to calculate the target matching degree between each target candidate and the job by deeply interacting the dynamic job profile with the skill information of each target candidate in the target candidate subset, and to generate a recommendation list based on the target matching degree. The processing module is used to collect user feedback data on the recommendation list and, in conjunction with external industry trend data, adjust and update the dynamic job profile based on the user feedback data and the external industry trend data.

9. An electronic device, characterized in that, The electronic device includes: a memory and at least one processor, wherein the memory stores instructions; The at least one processor invokes the instructions in the memory to cause the electronic device to execute the talent matching method based on dynamic job profiles as described in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instruction is executed by the processor, it implements the talent matching method based on dynamic job profiles as described in any one of claims 1-7.

Images