Intelligent processing method and platform for recruitment information based on cloud service

Abstract

The invention provides a cloud service-based recruitment information intelligent processing method and platform, and relates to the technical field of information processing, and the method comprises the steps: receiving recruitment information sent by a plurality of recruitment information sources, storing the recruitment information to a cloud server, carrying out the filtering of the normalized position information based on a knowledge graph construction rule, and obtaining the recruitment information of a plurality of recruitment information sources; according to the effective position information, constructing a position knowledge graph containing node features and edge relationships, and generating an initial recommendation list; an intelligent question and answer model is constructed for each position in the initial recommendation list, a final recommendation list is generated, the final recommendation list and the intelligent question and answer model are deployed to the cloud server, and personalized position information is pushed to a job seeker terminal through the cloud server.

Description

Technical Field

[0001] The present invention relates to information processing technology, and in particular to an intelligent processing method and platform for recruitment information based on cloud services. Background Art

[0002] Existing technologies usually require manual review of a large amount of recruitment information one by one, which is time-consuming and labor-intensive and cannot meet the rapidly changing market needs. At the same time, relying on manual methods can easily lead to information omissions or errors, reducing the accuracy and efficiency of recruitment.

[0003] Traditional recruitment information matching methods are often based on simple keyword searches, which fail to deeply understand the semantics of job descriptions and the actual capabilities of job seekers. This method easily leads to inaccurate matching between positions and job seekers, thus affecting recruitment results and causing waste of resources for both companies and job seekers.

[0004] The existing technology is relatively simple in the interaction between job seekers and recruitment information, and lacks the support of intelligent question-answering systems. When job seekers query job information, they can often only get static information display and cannot answer their questions immediately, resulting in poor user experience and affecting job seekers' trust in recruitment platforms and their enthusiasm for using them.

[0005] In summary, existing recruitment information processing technology urgently needs innovation to improve processing efficiency, enhance matching accuracy, and enhance intelligent interaction capabilities. Summary of the invention

[0006] The embodiments of the present invention provide a method and platform for intelligently processing recruitment information based on cloud services, which can solve the problems in the prior art.

[0007] According to a first aspect of the embodiments of the present invention, Provided is a method for intelligently processing recruitment information based on cloud services, including: Receive recruitment information sent by multiple recruitment information sources, and store the recruitment information in a cloud server, use a natural language processing model to perform word segmentation and part-of-speech tagging on the recruitment information, extract the job title, salary range, work location, job requirements, job responsibilities and company information in the recruitment information to form initial job information, use a bidirectional long short-term memory neural network to perform semantic understanding and association analysis on the initial job information, generate a job semantic vector, match the job semantic vector with a preset industry dictionary, normalize the job description, generate normalized job information, and filter the normalized job information based on the knowledge graph construction rules to obtain valid job information; Based on the valid job information, a job knowledge graph including node features and edge relationships is constructed, a graph neural network is used to perform convolution operations on the job knowledge graph, correlation features between jobs are extracted, the correlation features are input into a pre-trained deep neural network, a job feature vector is generated, a multi-dimensional job profile is constructed based on the job feature vector, a matching score between the multi-dimensional job profile and the job seeker resume profile in each dimension is calculated, a job-talent matching score is obtained, the valid job information is sorted based on the job-talent matching score, and an initial recommendation list is generated; An intelligent question-and-answer model is constructed for each position in the initial recommendation list, and the intelligent question-and-answer model includes a question understanding module and an answer generation module, wherein the question understanding module receives a question text input by a job seeker and converts the question text into a question vector; the answer generation module retrieves relevant content in the valid position information based on the question vector, generates an answer text, records the question text and the answer text to form a question-and-answer record, adjusts the order of the positions in the initial recommendation list based on the question-and-answer record, generates a final recommendation list, deploys the final recommendation list and the intelligent question-and-answer model to the cloud server, and pushes personalized position information to the job seeker terminal through the cloud server.

[0008] In an optional embodiment, A bidirectional long short-term memory neural network is used to perform semantic understanding and association analysis on the initial job information to generate a job semantic vector, match the job semantic vector with a preset industry dictionary, normalize the job description, generate normalized job information, and filter the normalized job information based on the knowledge graph construction rules to obtain valid job information including: Input the initial position information into a hierarchical attention encoder, the hierarchical attention encoder comprises a character-level encoding layer, a word-level encoding layer and a sentence-level encoding layer; the character-level encoding layer extracts character features through a convolutional neural network, and the word-level encoding layer and the sentence-level encoding layer extract word-level features and sentence-level features through a multi-head self-attention structure; input the character features, word-level features and sentence-level features into a cross-layer interaction module, and the gated update unit of the cross-layer interaction module selects and fuses the input features to generate an enhanced word vector sequence; The enhanced word vector sequence is input into a bidirectional long short-term memory neural network, wherein the bidirectional long short-term memory neural network comprises a memory enhancement layer and a semantic modeling layer; the memory enhancement layer stores long-term dependency features through an external memory matrix, and updates the external memory matrix based on an attention reading and writing mechanism; the semantic modeling layer processes features through a gating mechanism and residual connection and performs adversarial training to generate a position semantic vector; The position semantic vector is input into an industry dictionary matching module, and the industry dictionary matching module includes a multi-dimensional similarity calculation unit and a weight allocation unit; the multi-dimensional similarity calculation unit calculates the matching scores of the position semantic vector with standard terms at the lexical level, syntactic level and semantic level respectively; the weight allocation unit allocates weight coefficients to the matching scores at the three levels through an attention mechanism, selects corresponding standard terms based on the weighted matching scores, generates standardized position information, and filters the standardized position information based on knowledge graph construction rules to obtain valid position information.

[0009] In an optional embodiment, The standardized job information is filtered based on the knowledge graph construction rules to obtain valid job information including: Constructing an integrity rule set, a consistency rule set and an authenticity rule set, wherein the integrity rule set includes necessary field existence constraints and value range constraints, the consistency rule set includes field dependency constraints and field logic constraints, and the authenticity rule set includes knowledge matching constraints and value rationality constraints; extracting the trigger condition features and constraint condition features of each rule in the integrity rule set, the consistency rule set and the authenticity rule set to generate a rule feature set; Input the rule feature set into the word vector model to obtain a rule feature vector, perform attention weighted fusion on the rule feature vector, and generate a rule feature matrix; Inputting the standardized position information into the convolutional neural network, extracting the field local features of the standardized position information, calculating the similarity between the field local features and the rule feature matrix, and obtaining the rule matching features; The rule matching feature is input into a conditional random field model, wherein the conditional random field model comprises a state feature extraction unit and a transfer feature extraction unit, wherein the state feature extraction unit performs state modeling on the rule matching feature, and the transfer feature extraction unit models the label sequence through a recurrent neural network to identify the information position in the standardized job information that does not meet the rule constraints and the corresponding rule type; Based on the information position that does not meet the rule constraints and the corresponding rule type, the fields of the information that does not meet the integrity rules in the standardized position information are supplemented and the value ranges are corrected, the conflicts of the information that does not meet the consistency rules are adjusted, and the information that does not meet the authenticity rules is deleted to generate valid position information that meets the constraints of the integrity rule set, the consistency rule set and the authenticity rule set.

[0010] In an optional embodiment, According to the valid job information, a job knowledge graph including node features and edge relationships is constructed, a graph neural network is used to perform convolution operation on the job knowledge graph, and correlation features between jobs are extracted. The correlation features are input into a pre-trained deep neural network to generate a job feature vector, including: Constructing a job knowledge graph based on valid job information, the job knowledge graph includes a node feature set and an edge relationship set; the node feature set includes job attribute features, skill requirement features and experience requirement features, wherein the job attribute features include salary range and work location, the skill requirement features include professional skills and general skills, and the experience requirement features include years of work and project experience; the edge relationship set includes job hierarchical relationships, skill correlation relationships and job conversion relationships; constructing a graph structure matrix based on the node feature set and the edge relationship set; The graph structure matrix is ​​input into a multi-task graph neural network, which includes a contrastive learning module, an attribute reconstruction module and a relationship prediction module. The contrastive learning module generates a node representation by maximizing the mutual information of adjacent position nodes; the attribute reconstruction module optimizes the feature representation by reconstructing the masked position node attributes; the relationship prediction module enhances the structure representation by predicting the edge relationship between nodes; multi-task joint optimization is performed based on the node representation, feature representation and structure representation to generate a fusion representation of the position node; The fusion representation is input into a graph convolutional network, edge relationships are aggregated on features of adjacent position nodes through a feature propagation layer, nonlinear transformation is performed on aggregated features through a feature update layer, and position feature information of each layer is maintained through a residual connection to generate position association features; The associated features are input into a pre-trained deep neural network, which includes an encoding layer and a decoding layer. The encoding layer maps the associated features to a low-dimensional space through a multi-layer perceptron, and the decoding layer obtains an initial position feature vector through reconstruction optimization, and the initial position feature vector is normalized to generate a final position feature vector.

[0011] In an optional embodiment, Constructing a multi-dimensional job profile based on the job feature vector, calculating the matching score between the multi-dimensional job profile and the job seeker resume profile in each dimension to obtain a job-talent matching score, sorting the valid job information based on the job-talent matching score, and generating an initial recommendation list includes: The position feature vector is input into a multi-dimensional mapping network, including a professional skill dimension mapping network, a management capability dimension mapping network, a business experience dimension mapping network, and a general quality dimension mapping network. Each dimension mapping network performs feature transformation through a multi-layer perceptron and performs feature enhancement through residual connection and layer normalization to obtain initial dimension features; Performing self-attention calculation on the initial dimensional features to obtain an intra-dimensional feature association matrix, performing cross-attention calculation based on the feature association matrix to obtain inter-dimensional feature dependencies, and combining the features with inter-dimensional dependencies through learnable weights to obtain a multi-dimensional profile of the position; At the same time, the resume features of the job seeker are input into the multi-dimensional mapping network to perform dimension mapping and self-attention calculation processes to obtain a multi-dimensional portrait of the resume, the multi-dimensional portrait of the position and the multi-dimensional portrait of the resume are calculated in each dimension to match the importance weight and cosine similarity to obtain local dimensional matching features, and the multi-dimensional portrait of the position and the multi-dimensional portrait of the resume are input into the deep interaction network to obtain global matching features; The local dimension matching features and the global matching features are filtered through feature selection gating to generate filtered matching features, dynamic fusion weights are calculated for the filtered matching features, and weighted combinations are performed to obtain fused matching features; The gradient of the fused matching feature is calculated, and perturbations are added in the gradient direction to generate adversarial samples. The adversarial samples are input into a pre-built matching score prediction model for training to obtain a robust matching score prediction model. The position-talent matching pairs are scored according to the matching score prediction model to obtain the position-talent matching scores. The valid position information is sorted according to the position-talent matching scores to generate an initial recommendation list.

[0012] In an optional embodiment, Calculating the gradient of the fused matching feature, adding perturbations in the gradient direction to generate adversarial samples, inputting the adversarial samples into a pre-built matching score prediction model for training, and obtaining a robust matching score prediction model includes: Inputting the fused matching feature into a pre-built matching score prediction model to obtain a predicted matching score, and calculating an original loss function value based on the predicted matching score and the true matching score; Calculating the gradient of the original loss function value with respect to the fusion matching feature, normalizing the gradient to obtain a normalized gradient; constructing a feature correlation matrix based on the normalized gradient, extracting the main eigenvector of the feature correlation matrix; projecting the normalized gradient to the orthogonal complement space of the main eigenvector to obtain a projected gradient; Adding a disturbance amount to the fused matching feature along the direction of the projected gradient to obtain an initial adversarial sample, and truncating the initial adversarial sample based on a disturbance threshold to obtain an optimized adversarial sample; Inputting the optimized adversarial sample into a matching score prediction model to obtain an adversarial predicted matching score, and calculating an adversarial loss function value based on the adversarial predicted matching score and the true matching score; The original loss function value and the adversarial loss function value are weightedly summed to obtain a total loss function value. The model parameters of the matching score prediction model are updated based on the total loss function value to obtain updated model parameters. Repeat the iteration until the total loss function value is less than the preset convergence threshold to obtain a trained matching score prediction model.

[0013] In an optional embodiment, Adjusting the order of positions in the initial recommendation list based on the question and answer records to generate a final recommendation list, deploying the final recommendation list and the intelligent question and answer model to the cloud server, and pushing personalized position information to the job seeker terminal through the cloud server includes: The question text in the question and answer record is converted into a question word vector sequence through a word embedding layer, and the question word vector sequence is input into a bidirectional long short-term memory network to obtain a question semantic vector; at the same time, the answer text in the question and answer record is input into a pre-trained language model to obtain an answer feature vector, and the question semantic vector and the answer feature vector are input into a cross attention network to obtain a question and answer interaction vector; The question-answer interaction vector is used as a query matrix, and a multi-head attention calculation is performed on the feature association matrix of the positions in the initial recommendation list to obtain a feature attention score, and the feature attention score is multiplied by the corresponding professional weight coefficient and the sum is obtained to obtain a combined attention score; Calculating a temporal decay coefficient based on the timestamp of the question and answer record, adaptively fusing the product of the combined attention score and the temporal decay coefficient with the pre-obtained position-talent matching score to obtain a position ranking score, and reordering the positions in the initial recommendation list based on the position ranking score to generate a final recommendation list, wherein each list item in the final recommendation list includes position information, ranking score, timestamp, access count, and intelligent question and answer model; The intelligent question-answering model corresponding to the position whose access count exceeds the preset threshold is stored in a distributed cache cluster, and the final recommendation list is deployed to a cloud server master-slave cluster, where the master server maintains a global index and the slave servers form a processing cluster based on a consistent hashing algorithm; When receiving a job seeker's access request, extracting a user ID from the access request and mapping it to a designated slave server, the slave server obtaining the job seeker's historical question and answer records and calculating the behavioral similarity with the positions in the final recommendation list; A personalized recommendation score is calculated based on the behavior similarity and the job ranking score, the job information with the highest personalized recommendation score is selected, and the job information is pushed to the job seeker terminal through a long connection protocol.

[0014] According to a second aspect of the embodiments of the present invention, Provides an intelligent processing platform for recruitment information based on cloud services, including: The first unit is used to receive recruitment information sent by multiple recruitment information sources, store the recruitment information in a cloud server, perform word segmentation and part-of-speech tagging on the recruitment information using a natural language processing model, extract the job title, salary range, work location, job requirements, job responsibilities and company information in the recruitment information to form initial job information, use a bidirectional long short-term memory neural network to perform semantic understanding and association analysis on the initial job information, generate a job semantic vector, match the job semantic vector with a preset industry dictionary, normalize the job description, generate normalized job information, and filter the normalized job information based on knowledge graph construction rules to obtain valid job information; The second unit is used to construct a job knowledge graph including node features and edge relationships according to the valid job information, perform convolution operation on the job knowledge graph using a graph neural network, extract correlation features between jobs, input the correlation features into a pre-trained deep neural network, generate a job feature vector, construct a multi-dimensional job profile based on the job feature vector, calculate the matching score between the multi-dimensional job profile and the job seeker resume profile in each dimension, obtain a job-talent matching score, sort the valid job information based on the job-talent matching score, and generate an initial recommendation list; The third unit is used to build an intelligent question and answer model for each position in the initial recommendation list, and the intelligent question and answer model includes a question understanding module and an answer generation module, wherein the question understanding module receives a question text input by a job seeker, and converts the question text into a question vector; the answer generation module retrieves relevant content in the valid position information based on the question vector, generates an answer text, records the question text and the answer text to form a question and answer record, adjusts the order of the positions in the initial recommendation list based on the question and answer record, generates a final recommendation list, deploys the final recommendation list and the intelligent question and answer model to the cloud server, and pushes personalized position information to the job seeker terminal through the cloud server.

[0015] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0016] According to a fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0017] In this embodiment, through natural language processing and deep learning technology, a large amount of recruitment information can be efficiently processed and analyzed, key position information can be extracted, and multi-dimensional matching can be performed with the job applicant's resume portrait through the position feature vector, thereby improving the matching accuracy between positions and talents and helping job applicants find suitable positions faster. By building an intelligent question-and-answer model, it is possible to respond to job applicants' questions in real time and provide relevant position information. This personalized interaction method not only improves the user experience, but also makes the recommendation list more in line with the needs of job applicants, ensuring that job applicants can obtain more accurate and expected position information. Using knowledge graph and graph neural network technology, recruitment information is filtered and associated analyzed, redundant and invalid information is effectively removed, and the quality of recommended position information is guaranteed. This efficient information management mechanism makes the recruitment process smoother, reduces the time cost of job applicants in information screening, and improves the overall recruitment efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A flowchart of an intelligent processing method for recruitment information based on cloud services according to an embodiment of the present invention; Figure 2 The present invention is a schematic diagram of the structure of an intelligent processing platform for recruitment information based on cloud services according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0021] Figure 1 FIG. 1 is a flow chart of an intelligent processing method for recruitment information based on cloud services according to an embodiment of the present invention. Figure 1 As shown, the method includes: S101. Receive recruitment information sent by multiple recruitment information sources, and store the recruitment information in a cloud server, use a natural language processing model to segment and tag the recruitment information, extract the job title, salary range, work location, job requirements, job responsibilities and company information in the recruitment information to form initial job information, use a bidirectional long short-term memory neural network to perform semantic understanding and association analysis on the initial job information, generate a job semantic vector, match the job semantic vector with a preset industry dictionary, normalize the job description, generate standardized job information, filter the standardized job information based on the knowledge graph construction rules, and obtain valid job information.

[0022] The system needs to receive various recruitment information by integrating different recruitment information sources (such as recruitment websites, company databases, third-party recruitment platforms, etc.). Each information source transmits the recruitment information to the system through API, data stream interface or batch import. The system should have multiple ways to collect and integrate data, such as obtaining recruitment information in JSON format through HTTP protocol, or parsing data from XML files. This information may include job title, salary range, work location, job requirements, job responsibilities, and company information.

[0023] Once the job posting is received, the system will store it in the database of the cloud server. Cloud servers usually use distributed storage to ensure high availability and scalability of data. The information storage format can be structured (e.g. SQL database) or semi-structured (e.g. NoSQL database). Each field of the job posting (such as job title, salary range, etc.) will be stored as a record in the database, and the content of each record is the specific field value extracted from the job posting.

[0024] The natural language processing model is a type of computing model used to process and analyze human language, including tasks such as text preprocessing, word segmentation, part-of-speech tagging, and information extraction. In the present invention, the natural language processing model is used to perform text analysis on recruitment information to help the system understand and extract key information (such as job titles, job requirements, etc.).

[0025] Once the information is stored, the system will then perform text preprocessing on the recruitment information, including tokenization and part-of-speech tagging. This step uses a natural language processing (NLP) model to perform text analysis on the recruitment information and convert information such as job titles and salary ranges into words and their grammatical functions. In the tokenization process, the long text in the recruitment information will be divided into words with independent meanings; in part-of-speech tagging, each word will be marked with its grammatical role in the sentence, such as noun, verb, adjective, etc.

[0026] After word segmentation and part-of-speech tagging, the system automatically extracts key information fields from the recruitment information through rules or models, including job title, salary range, job location, job requirements, job responsibilities, and company information. This step uses information extraction technology, which can be based on machine learning models (such as CRF, BERT, etc.) or rule-based template matching methods.

[0027] Next, the system inputs the initial job information into a bidirectional long short-term memory network (BiLSTM). Through its bidirectional propagation mechanism, BiLSTM can deeply understand the contextual information in the recruitment information. This step generates a job semantic vector by learning the sequence relationship of the job information (such as the relationship between the job title and the job requirements, the connection between the work location and the salary range, etc.), expressing the semantic and structural characteristics of the job information.

[0028] After semantic understanding, the system will match the generated job semantic vector with a preset industry dictionary. This industry dictionary includes common industry positions, skills, job responsibilities and other related terms, with the purpose of standardizing and normalizing job information through matching rules. Dictionary matching helps to further unify the terms and expressions in job information, making job descriptions more standardized.

[0029] After matching with the industry dictionary, the system normalizes the job description, unifies the terminology and format, and eliminates inconsistencies in the information. After normalization, the content of the job description will meet industry standards, making it easier to process and analyze later.

[0030] Next, the system uses the knowledge graph and industry rules to filter the standardized job information. The knowledge graph is a graphical database that records the relationships and dependencies between various positions and requirements. The rules in the graph are used to determine which job information is valid and which is redundant or irrelevant. The filtering rules can be based on the logical relationship of job requirements, skill matching, and the compatibility of location and position.

[0031] Through the above filtering steps, we finally get a valid job information that has been normalized, semantically understood and filtered by rules. At this point, the job information has been cleaned, standardized and complies with industry standards, and can be used for further analysis, recommendation or display.

[0032] In an optional implementation, a bidirectional long short-term memory neural network is used to perform semantic understanding and association analysis on the initial job information to generate a job semantic vector, the job semantic vector is matched with a preset industry dictionary, the job description is normalized to generate normalized job information, and the normalized job information is filtered based on knowledge graph construction rules to obtain valid job information including: Input the initial position information into a hierarchical attention encoder, the hierarchical attention encoder comprises a character-level encoding layer, a word-level encoding layer and a sentence-level encoding layer; the character-level encoding layer extracts character features through a convolutional neural network, and the word-level encoding layer and the sentence-level encoding layer extract word-level features and sentence-level features through a multi-head self-attention structure; input the character features, word-level features and sentence-level features into a cross-layer interaction module, and the gated update unit of the cross-layer interaction module selects and fuses the input features to generate an enhanced word vector sequence; The enhanced word vector sequence is input into a bidirectional long short-term memory neural network, wherein the bidirectional long short-term memory neural network comprises a memory enhancement layer and a semantic modeling layer; the memory enhancement layer stores long-term dependency features through an external memory matrix, and updates the external memory matrix based on an attention reading and writing mechanism; the semantic modeling layer processes features through a gating mechanism and residual connection and performs adversarial training to generate a position semantic vector; The position semantic vector is input into an industry dictionary matching module, and the industry dictionary matching module includes a multi-dimensional similarity calculation unit and a weight allocation unit; the multi-dimensional similarity calculation unit calculates the matching scores of the position semantic vector with standard terms at the lexical level, syntactic level and semantic level respectively; the weight allocation unit allocates weight coefficients to the matching scores at the three levels through an attention mechanism, selects corresponding standard terms based on the weighted matching scores, generates standardized position information, and filters the standardized position information based on knowledge graph construction rules to obtain valid position information.

[0033] For example, the initial job information is input into the hierarchical attention encoder, and the character-level encoding layer is first designed to extract character features using a convolutional neural network. The convolutional neural network can effectively capture local features and contextual information between characters through the combination of multiple convolutional layers and pooling layers. By training on a large amount of job description data, a deep understanding of characters can be formed.

[0034] In the word-level encoding layer and sentence-level encoding layer, a multi-head self-attention mechanism is used to extract word-level features and sentence-level features. This mechanism allows the model to focus on multiple parts of the input sequence while processing each word, thereby capturing richer contextual information. By performing multiple self-attention calculations on the input sequence, the contextual representation of each word is generated, further enhancing the expressiveness of the information. After this series of processing, the character features, word-level features, and sentence-level features obtained will be input into the cross-layer interaction module.

[0035] The gated update unit of the cross-layer interaction module is responsible for selecting and fusing input features. This unit sets a gating mechanism to determine which features to retain and which features to suppress, thereby generating an enhanced word vector sequence. This step ensures that the model can effectively integrate information at different levels and improve the accuracy of subsequent processing.

[0036] The enhanced word vector sequence is input into the bidirectional long short-term memory neural network. The network contains a memory enhancement layer and a semantic modeling layer. The memory enhancement layer stores long-term dependency features through an external memory matrix and uses the attention mechanism to read and write the external memory matrix to dynamically update the stored information. This can effectively capture the long-term dependencies in the job description.

[0037] The semantic modeling layer processes features through gating mechanisms and residual connections. The gating mechanism can control the flow of information, while the residual connection helps alleviate the gradient vanishing problem in deep networks. Through adversarial training, the model can continuously optimize the expression of features and generate high-quality job semantic vectors.

[0038] The position semantic vector is input into the industry dictionary matching module. First, in the multi-dimensional similarity calculation unit, the matching scores of the position semantic vector and the standard terminology at the lexical level, syntactic level and semantic level are calculated respectively. The matching at the lexical level focuses on the similarity of vocabulary, the matching at the syntactic level focuses on the similarity of grammatical structure, and the matching at the semantic level focuses on the similarity of overall meaning.

[0039] The weight assignment unit assigns weight coefficients to the matching scores at the three levels through the attention mechanism. According to the weighted matching scores, the most suitable standard terms are selected to generate standardized job information. This step ensures that the generated job information can accurately reflect industry standards.

[0040] Based on the knowledge graph, rules are constructed to filter standardized job information to ensure that the final valid job information meets the preset standards and requirements. Through the knowledge graph, the job information can be contextually understood and relevance analyzed to further improve the accuracy and practicality of the information.

[0041] In this embodiment, the semantic understanding ability of the job description is improved, and the core content of the job information can be captured more accurately through hierarchical feature extraction and enhancement. The generation of job information is standardized, the matching degree between job information and industry standards is improved, and the consistency and availability of information are enhanced. Through the filtering of the knowledge graph, it is ensured that the job information finally output is effective and meets actual needs, reducing the risk of information redundancy and errors.

[0042] In an optional implementation, the standardized job information is filtered based on the knowledge graph construction rules to obtain valid job information including: Constructing an integrity rule set, a consistency rule set and an authenticity rule set, wherein the integrity rule set includes necessary field existence constraints and value range constraints, the consistency rule set includes field dependency constraints and field logic constraints, and the authenticity rule set includes knowledge matching constraints and value rationality constraints; extracting the trigger condition features and constraint condition features of each rule in the integrity rule set, the consistency rule set and the authenticity rule set to generate a rule feature set; Input the rule feature set into the word vector model to obtain a rule feature vector, perform attention weighted fusion on the rule feature vector, and generate a rule feature matrix; Inputting the standardized position information into the convolutional neural network, extracting the field local features of the standardized position information, calculating the similarity between the field local features and the rule feature matrix, and obtaining the rule matching features; The rule matching feature is input into a conditional random field model, wherein the conditional random field model comprises a state feature extraction unit and a transfer feature extraction unit, wherein the state feature extraction unit performs state modeling on the rule matching feature, and the transfer feature extraction unit models the label sequence through a recurrent neural network to identify the information position in the standardized job information that does not meet the rule constraints and the corresponding rule type; Based on the information position that does not meet the rule constraints and the corresponding rule type, the fields of the information that does not meet the integrity rules in the standardized position information are supplemented and the value ranges are corrected, the conflicts of the information that does not meet the consistency rules are adjusted, and the information that does not meet the authenticity rules is deleted to generate valid position information that meets the constraints of the integrity rule set, the consistency rule set and the authenticity rule set.

[0043] Exemplarily, building integrity rule sets, consistency rule sets, and authenticity rule sets are important steps to achieve effective job information filtering. The integrity rule set focuses on the existence of required fields and the range of field values, ensuring that all job information contains necessary fields and that the values ​​of these fields are within a reasonable range. For example, for the "work experience" field in job information, the integrity rule requires that the field must exist and the value should be between 0 and 30 years.

[0044] The consistency rule set focuses on the dependencies and logical relationships between fields. For example, if the value of the "education" field is "Master", then the "major" field should be related to it, and no unreasonable combination can appear. By defining these rules, you can effectively identify contradictory information between fields.

[0045] The authenticity rule set is used to verify the authenticity of the information, including knowledge matching and numerical rationality. For example, the "salary" field in the job information should be consistent with the market situation, and values ​​that are too high or too low need to be marked as unreasonable.

[0046] When extracting rule features, we first need to identify the triggering conditions and constraints of each rule. Triggering condition features can be the value, type, or existence of a field, while constraint condition features include the logical relationship and dependency between fields. The generated rule feature set will serve as the basis for subsequent processing.

[0047] After the rule feature set is input into the word vector model, the features can be converted into vector form, which is convenient for subsequent calculation and processing. By performing attention weighted fusion on the rule feature vector, important features can be highlighted and a rule feature matrix can be generated. This process ensures that important features can be given greater weight in subsequent steps.

[0048] Next, the normalized job information is input into the convolutional neural network to extract the local features of the field. The convolutional neural network can automatically identify important patterns and features in the job information, such as job title, work location, etc. By calculating the similarity between the local features of the field and the rule feature matrix, the rule matching feature can be obtained. This feature represents the degree of fit between the job information and the rule, providing a basis for subsequent state modeling.

[0049] The rule matching features are input into the conditional random field model, and the state feature extraction unit will model the rule matching features and identify the information locations that do not meet the rule constraints. The transfer feature extraction unit analyzes the label sequence through a recurrent neural network to further confirm the specific types that do not meet the rules. This process can accurately locate the problems in the job information and provide data support for subsequent adjustments.

[0050] Based on the identified information location and rule type that does not meet the rule constraints, corresponding processing is performed. For information that does not meet the integrity rules, field supplementation and value range correction are performed to ensure that all necessary fields exist and the values ​​are within a reasonable range. For information that does not meet the consistency rules, conflict adjustment is performed to ensure consistency between fields. For information that does not meet the authenticity rules, it is deleted to ensure that the final position information is authentic and reliable.

[0051] Through the above steps, the final generated valid job information will meet all the rules and constraints of completeness, consistency and authenticity. This process not only improves the quality of job information, but also provides a reliable basis for subsequent data analysis and decision-making.

[0052] In this embodiment, the accuracy and reliability of job information are improved to ensure that the information obtained by users is authentic and valid. The automation of data processing is enhanced to reduce manual intervention and improve work efficiency. Through the systematic management of rules, subsequent data maintenance and updating are facilitated, and the flexibility of information management is improved.

[0053] S102. Based on the valid job information, a job knowledge graph including node features and edge relationships is constructed, and a graph neural network is used to perform convolution operations on the job knowledge graph to extract correlation features between jobs. The correlation features are input into a pre-trained deep neural network to generate a job feature vector, and a multi-dimensional job profile is constructed based on the job feature vector. The matching scores between the multi-dimensional job profile and the job applicant's resume profile in each dimension are calculated to obtain a job-talent matching score. The valid job information is sorted based on the job-talent matching score to generate an initial recommendation list.

[0054] Among them, the graph neural network (GNN) is used to perform convolution operations on the job knowledge graph to extract the correlation features between jobs. These features reflect the similarities and correlations between different jobs. The extracted correlation features are input into the pre-trained deep neural network to generate a job feature vector. Based on this feature vector, a multi-dimensional job profile is constructed, covering all aspects of the job (such as salary, work location, skill requirements, etc.).

[0055] Calculate the matching scores of the multi-dimensional job profile and the job seeker resume profile in each dimension. The job seeker resume profile is also generated through natural language processing and machine learning models to ensure that the dimensions are consistent with the job profile. After obtaining the job-talent matching score, sort the valid job information based on the score to generate an initial recommendation list.

[0056] In an optional implementation, based on the valid job information, a job knowledge graph including node features and edge relationships is constructed, a graph neural network is used to perform a convolution operation on the job knowledge graph, and correlation features between jobs are extracted. The correlation features are input into a pre-trained deep neural network to generate a job feature vector, including: Constructing a job knowledge graph based on valid job information, the job knowledge graph includes a node feature set and an edge relationship set; the node feature set includes job attribute features, skill requirement features and experience requirement features, wherein the job attribute features include salary range and work location, the skill requirement features include professional skills and general skills, and the experience requirement features include years of work and project experience; the edge relationship set includes job hierarchical relationships, skill correlation relationships and job conversion relationships; constructing a graph structure matrix based on the node feature set and the edge relationship set; The graph structure matrix is ​​input into a multi-task graph neural network, which includes a contrastive learning module, an attribute reconstruction module and a relationship prediction module. The contrastive learning module generates a node representation by maximizing the mutual information of adjacent position nodes; the attribute reconstruction module optimizes the feature representation by reconstructing the masked position node attributes; the relationship prediction module enhances the structure representation by predicting the edge relationship between nodes; multi-task joint optimization is performed based on the node representation, feature representation and structure representation to generate a fusion representation of the position node; The fusion representation is input into a graph convolutional network, edge relationships are aggregated on features of adjacent position nodes through a feature propagation layer, nonlinear transformation is performed on aggregated features through a feature update layer, and position feature information of each layer is maintained through a residual connection to generate position association features; The associated features are input into a pre-trained deep neural network, which includes an encoding layer and a decoding layer. The encoding layer maps the associated features to a low-dimensional space through a multi-layer perceptron, and the decoding layer obtains an initial position feature vector through reconstruction optimization, and the initial position feature vector is normalized to generate a final position feature vector.

[0057] For example, the first step in building a job knowledge graph is to collect valid job information. This information should include data on job attributes, skill requirements, and experience requirements. Job attribute features should cover salary ranges and work locations, skill requirement features should include professional skills and general skills, and experience requirement features should include years of work experience and project experience. These data will serve as a node feature set to form the basis of the job knowledge graph.

[0058] On the basis of the node feature set, an edge relationship set is constructed. The edge relationship set should include position hierarchy relationships, skill correlation relationships, and position conversion relationships. Position hierarchy relationships can reflect the superior-subordinate relationship between different positions, skill correlation relationships indicate the similarity of skills required for different positions, and position conversion relationships describe the conversion path between positions. By integrating the node feature set and the edge relationship set, a graph structure matrix is ​​constructed to form the overall structure of the position knowledge graph.

[0059] Next, the graph structure matrix is ​​input into the multi-task graph neural network. The network contains multiple modules, including contrastive learning module, attribute reconstruction module and relationship prediction module. The goal of the contrastive learning module is to generate node representation by maximizing information sharing between adjacent position nodes. The attribute reconstruction module further optimizes the feature representation by reconstructing the masked position node attributes. The relationship prediction module is responsible for predicting the edge relationship between nodes, thereby enhancing the structural representation. Through the joint optimization of these three modules, a fused representation of the position node is generated to ensure that the feature, attribute and structural information of the node are fully utilized.

[0060] After obtaining the fused representation, it is input into the graph convolutional network. The graph convolutional network aggregates the features of adjacent position nodes through the feature propagation layer and integrates the features of adjacent nodes using the characteristics of edge relationships. In the feature update layer, the aggregated features are transformed nonlinearly to enhance the expressive power of the model. At the same time, the integrity of the position feature information at each layer is maintained through residual connections. This process ultimately generates the associated features between positions.

[0061] Finally, the obtained association features are input into the pre-trained deep neural network. The network consists of an encoding layer and a decoding layer. The encoding layer maps the association features to a low-dimensional space through a multi-layer perceptron to extract a more refined feature representation. The decoding layer generates an initial position feature vector through reconstruction optimization. The initial feature vector is normalized to ensure the stability of the feature vector, and finally the final position feature vector is generated.

[0062] In this embodiment, by constructing a job knowledge graph, the features and relationships between different jobs can be effectively integrated, thereby improving the relevance and availability of job information. Using graph neural networks for multi-task learning can extract richer and more meaningful features from complex job data and improve the performance of subsequent models. By generating high-quality job feature vectors, more accurate job matching can be achieved during the recruitment and job search process, improving recruitment efficiency and job search success rate.

[0063] In an optional implementation, constructing a multi-dimensional job profile based on the job feature vector, calculating the matching score between the multi-dimensional job profile and the job seeker resume profile in each dimension to obtain a job-talent matching score, sorting the valid job information based on the job-talent matching score, and generating an initial recommendation list includes: The position feature vector is input into a multi-dimensional mapping network, including a professional skill dimension mapping network, a management capability dimension mapping network, a business experience dimension mapping network, and a general quality dimension mapping network. Each dimension mapping network performs feature transformation through a multi-layer perceptron and performs feature enhancement through residual connection and layer normalization to obtain initial dimension features; Performing self-attention calculation on the initial dimensional features to obtain an intra-dimensional feature association matrix, performing cross-attention calculation based on the feature association matrix to obtain inter-dimensional feature dependencies, and combining the features with inter-dimensional dependencies through learnable weights to obtain a multi-dimensional profile of the position; At the same time, the resume features of the job seeker are input into the multi-dimensional mapping network to perform dimension mapping and self-attention calculation processes to obtain a multi-dimensional portrait of the resume, the multi-dimensional portrait of the position and the multi-dimensional portrait of the resume are calculated in each dimension to match the importance weight and cosine similarity to obtain local dimensional matching features, and the multi-dimensional portrait of the position and the multi-dimensional portrait of the resume are input into the deep interaction network to obtain global matching features; The local dimension matching features and the global matching features are filtered through feature selection gating to generate filtered matching features, dynamic fusion weights are calculated for the filtered matching features, and weighted combinations are performed to obtain fused matching features; The gradient of the fused matching feature is calculated, and perturbations are added in the gradient direction to generate adversarial samples. The adversarial samples are input into a pre-built matching score prediction model for training to obtain a robust matching score prediction model. The position-talent matching pairs are scored according to the matching score prediction model to obtain the position-talent matching scores. The valid position information is sorted according to the position-talent matching scores to generate an initial recommendation list.

[0064] Exemplarily, the process of constructing a multi-dimensional job profile begins with the input of a job feature vector. The job feature vector contains information on multiple dimensions, such as professional skills, management capabilities, business experience, and general qualities of the job. These feature vectors will be input into a multi-dimensional mapping network, which consists of multiple dimension mapping sub-networks, each of which focuses on a specific dimension.

[0065] In each dimension mapping subnetwork, the input job features are first transformed by a multi-layer perceptron. The purpose of these transformations is to extract the core features of each dimension. Subsequently, residual connections and layer normalization techniques are used to enhance the feature representation to ensure the training stability of the network and the effectiveness of the features. After these processes, the initial dimensional features obtained will serve as the basis for subsequent steps.

[0066] Next, self-attention calculation is performed on the initial dimension features to generate an intra-dimensional feature correlation matrix. This matrix can reflect the relationship between features in the same dimension. Based on this feature correlation matrix, cross-attention calculation is performed to capture the feature dependencies between different dimensions. Finally, the features with inter-dimensional dependencies are combined through learnable weights to obtain a comprehensive multi-dimensional profile of the position.

[0067] When processing job applicant resumes, resume features are also input into the multi-dimensional mapping network, and the same dimension mapping and self-attention calculation process as the position features is performed to generate a multi-dimensional portrait of the resume. Next, the multi-dimensional portrait of the position is matched with the multi-dimensional portrait of the resume in each dimension, and the importance weight and cosine similarity of the match are calculated to obtain the local dimension matching features.

[0068] After obtaining the local dimension matching features, they are combined with the global matching features. The global matching features are obtained by inputting the multi-dimensional profile of the position and the multi-dimensional profile of the resume into the deep interaction network. The feature selection gating filters the local and global matching features to generate filtered matching features. Subsequently, the dynamic fusion weights are calculated for these features and weighted combination is performed to obtain the optimized fusion matching features.

[0069] On this basis, the gradient of the fused matching features is calculated, and perturbations are added in the gradient direction to generate adversarial samples. The adversarial samples will be input into the pre-built matching score prediction model for training to improve the robustness of the model. Finally, the trained matching score prediction model is used to score the job-talent matching pairs to obtain the job-talent matching scores. Based on these scores, the valid job information is sorted to generate an initial recommendation list.

[0070] In this embodiment, the matching accuracy between positions and job seekers is improved, so that the recommendation results are more in line with user needs. Through comprehensive consideration of multi-dimensional features, the expressiveness of position portraits and resume portraits is enhanced, which can more comprehensively reflect the capabilities of candidates. The robustness of the matching score prediction model is improved through adversarial sample training, ensuring stability and reliability under different data conditions.

[0071] In an optional implementation, calculating the gradient of the fused matching feature, adding perturbations in the gradient direction to generate adversarial samples, inputting the adversarial samples into a pre-built matching score prediction model for training, and obtaining a robust matching score prediction model includes: Inputting the fused matching feature into a pre-built matching score prediction model to obtain a predicted matching score, and calculating an original loss function value based on the predicted matching score and the true matching score; Calculating the gradient of the original loss function value with respect to the fusion matching feature, normalizing the gradient to obtain a normalized gradient; constructing a feature correlation matrix based on the normalized gradient, extracting the main eigenvector of the feature correlation matrix; projecting the normalized gradient to the orthogonal complement space of the main eigenvector to obtain a projected gradient; Adding a disturbance amount to the fused matching feature along the direction of the projected gradient to obtain an initial adversarial sample, and truncating the initial adversarial sample based on a disturbance threshold to obtain an optimized adversarial sample; Inputting the optimized adversarial sample into a matching score prediction model to obtain an adversarial predicted matching score, and calculating an adversarial loss function value based on the adversarial predicted matching score and the true matching score; The original loss function value and the adversarial loss function value are weightedly summed to obtain a total loss function value. The model parameters of the matching score prediction model are updated based on the total loss function value to obtain updated model parameters. Repeat the iteration until the total loss function value is less than the preset convergence threshold to obtain a trained matching score prediction model.

[0072] Exemplarily, first, the fused matching features are processed and input into a pre-built matching score prediction model to obtain a predicted matching score. This step involves extracting features from the dataset and fusing them to form a comprehensive feature representation. By inputting these features into the model, a preliminary matching score can be obtained, which will be used for subsequent loss calculations.

[0073] Next, the original loss function value is calculated based on the difference between the predicted match score and the actual match score. This process aims to evaluate the performance of the model under the current feature representation. The smaller the value of the loss function, the more accurate the model's prediction. Among them, the true match score is a known standard value that indicates the degree of match between the position and the talent. This true match score can be obtained through historical recruitment data, manual annotation, expert evaluation, or external evaluation tools. In historical data, if a job seeker is successfully hired and performs well, this position-talent match is usually assigned a higher score, while a mismatch is assigned a lower score.

[0074] After obtaining the original loss function value, its gradient needs to be calculated with respect to the fused matching features. This step is to determine the optimization direction of the model in the feature space. The calculated gradient will be normalized to ensure its effectiveness and comparability in subsequent steps. The process of normalizing the gradient aims to eliminate the dimensional differences between different features so that the direction of the gradient can accurately reflect the impact of the feature on the loss.

[0075] By normalizing the gradient, constructing the feature correlation matrix, and extracting the main eigenvectors. The feature correlation matrix can reveal the relationship between different features, while the main eigenvectors represent the main direction of change in the feature space. Next, the normalized gradient is projected into the orthogonal complement space of the main eigenvector to obtain the projected gradient. The purpose of this step is to reduce unnecessary interference in the feature space, making subsequent perturbations more accurate.

[0076] After obtaining the projected gradient, add a perturbation to the fused matching feature along its direction to generate an initial adversarial sample. The addition of the perturbation needs to be controlled within a certain range to ensure that the adversarial sample is still representative. Next, the initial adversarial sample is truncated according to the set perturbation threshold to ensure that the adversarial sample does not deviate from the distribution of the original features, and the optimized adversarial sample is obtained.

[0077] The optimized adversarial sample is input into the matching score prediction model to obtain the adversarial predicted matching score. The adversarial loss function value is calculated by comparing the difference between the adversarial predicted matching score and the true matching score. This process can help evaluate the robustness of the model when facing adversarial samples.

[0078] Finally, the original loss function value and the adversarial loss function value are weighted and summed to obtain the total loss function value. Based on the total loss function value, the parameters of the matching score prediction model are updated. By repeatedly iterating this process until the total loss function value is less than the preset convergence threshold, a trained and robust matching score prediction model is finally obtained.

[0079] In this embodiment, by introducing adversarial samples, the model can better adapt to potential input disturbances, thereby improving stability and reliability in practical applications. By constructing a feature correlation matrix, the relationship between features can be better understood, thereby improving the model's ability to handle complex data. By dynamically adjusting the calculation method of the loss function, it is possible to converge to the optimal model parameters more quickly, improving the efficiency of the training process.

[0080] S103. Construct an intelligent question-and-answer model for each position in the initial recommendation list, the intelligent question-and-answer model includes a question understanding module and an answer generation module, wherein the question understanding module receives a question text input by a job seeker and converts the question text into a question vector; the answer generation module retrieves relevant content in the valid position information based on the question vector, generates an answer text, records the question text and the answer text to form a question-and-answer record, adjusts the order of the positions in the initial recommendation list based on the question-and-answer record, generates a final recommendation list, deploys the final recommendation list and the intelligent question-and-answer model to the cloud server, and pushes personalized position information to the job seeker terminal through the cloud server.

[0081] First, the task of the question understanding module is to process the text questions input by job seekers. This module converts the received question text into a form that can be processed by computers, namely the question vector. The question vector is a mathematical representation of the text semantics, which can reflect the intention and key information of the question. In actual operation, the question understanding module usually uses natural language processing (NLP) technology to implement functions such as word segmentation, part-of-speech tagging, and syntactic parsing of the text to help understand the specific content of the question.

[0082] Next, the answer generation module retrieves relevant content from the job information based on the question vector. The core of this module is to find the job information most relevant to the question by indexing and matching the job data. In this process, the system needs to ensure that the matching content is not just a superficial keyword match, but to ensure the accuracy and pertinence of the answer through in-depth semantic understanding. For example, if a job seeker asks "Are there any Python development positions in Beijing?", the system needs to accurately extract positions related to "Python development" from the job information and confirm whether the position is in the "Beijing" area.

[0083] The generated answer text will be saved in the question and answer record for subsequent use. Every time a user asks a question and gets an answer, the system will record this pair of questions and answers and use it to dynamically adjust the ranking of job recommendations. The goal of this adjustment is to make the recommendation list better reflect the personalized needs of job seekers. For example, if a job seeker frequently asks about technical positions, the system can analyze his or her historical questions, infer his or her interests and needs, and optimize the job list accordingly, so that these positions will be ranked higher in the recommendation list.

[0084] In an optional implementation, adjusting the order of positions in the initial recommendation list based on the question and answer records to generate a final recommendation list, deploying the final recommendation list and the intelligent question and answer model to the cloud server, and pushing personalized position information to the job seeker terminal through the cloud server includes: The question text in the question and answer record is converted into a question word vector sequence through a word embedding layer, and the question word vector sequence is input into a bidirectional long short-term memory network to obtain a question semantic vector; at the same time, the answer text in the question and answer record is input into a pre-trained language model to obtain an answer feature vector, and the question semantic vector and the answer feature vector are input into a cross attention network to obtain a question and answer interaction vector; The question-answer interaction vector is used as a query matrix, and a multi-head attention calculation is performed on the feature association matrix of the positions in the initial recommendation list to obtain a feature attention score, and the feature attention score is multiplied by the corresponding professional weight coefficient and the sum is obtained to obtain a combined attention score; Calculating a temporal decay coefficient based on the timestamp of the question and answer record, adaptively fusing the product of the combined attention score and the temporal decay coefficient with the pre-obtained position-talent matching score to obtain a position ranking score, and reordering the positions in the initial recommendation list based on the position ranking score to generate a final recommendation list, wherein each list item in the final recommendation list includes position information, ranking score, timestamp, access count, and intelligent question and answer model; The intelligent question-answering model corresponding to the position whose access count exceeds the preset threshold is stored in a distributed cache cluster, and the final recommendation list is deployed to a cloud server master-slave cluster, where the master server maintains a global index and the slave servers form a processing cluster based on a consistent hashing algorithm; When receiving a job seeker's access request, extracting a user ID from the access request and mapping it to a designated slave server, the slave server obtaining the job seeker's historical question and answer records and calculating the behavioral similarity with the positions in the final recommendation list; A personalized recommendation score is calculated based on the behavior similarity and the job ranking score, the job information with the highest personalized recommendation score is selected, and the job information is pushed to the job seeker terminal through a long connection protocol.

[0085] For example, by analyzing the question and answer records of job seekers, adjusting the order of positions in the initial recommendation list, generating a final recommendation list, and deploying the list together with the intelligent question and answer model to the cloud server, the technical solution for implementing personalized job information push is as follows: First, we collect the candidates’ Q&A records, including the text content of the questions and answers. We convert the question text into a sequence of question word vectors through a word embedding layer. This process involves converting each word into a vector of fixed dimension, so that the semantic relationship between words can be captured. Next, we input the question word vector sequence into a bidirectional long short-term memory network (Bi-LSTM), which extracts the context information of the question and obtains the semantic vector of the question.

[0086] At the same time, the answer text in the question and answer record is input into a pre-trained language model, such as BERT or GPT, to generate an answer feature vector. This feature vector can effectively represent the semantic information of the answer. Next, the obtained question semantic vector and answer feature vector are input into the cross attention network. This network generates a question-answer interaction vector by focusing on the relationship between the question and the answer, which contains the comprehensive semantic information of the question and the answer.

[0087] The generated question-answer interaction vector is used as the query matrix to perform multi-head attention calculation with the feature association matrix of the positions in the initial recommendation list. The purpose of this calculation is to evaluate the correlation between the question-answer interaction and the features of each position to obtain the feature attention score. Then, the feature attention score is multiplied and summed with the corresponding professional weight coefficient to obtain the combined attention score. The professional weight coefficient is set based on the professional requirements of the position and the background information of the job seeker to ensure that the recommended position matches the professional skills of the job seeker.

[0088] Based on the timestamp of the question and answer record, the time decay coefficient is calculated. This coefficient is used to adjust the combined attention score to reflect the timeliness of the information. The product of the combined attention score and the time decay coefficient is adaptively fused with the pre-obtained position-talent matching score to obtain the position ranking score. This ranking score is used to re-rank the positions in the initial recommendation list to generate the final recommendation list. Each list item in the final recommendation list contains position information, ranking score, timestamp, access count, and intelligent question and answer model.

[0089] For positions whose access counts exceed the preset threshold, the corresponding intelligent question-answering model is stored in the distributed cache cluster to improve the response speed of subsequent requests. The final recommendation list is deployed to the master-slave cluster of the cloud server. The master server is responsible for maintaining the global index, while the slave servers form a processing cluster based on the consistent hashing algorithm to achieve load balancing and high availability.

[0090] When receiving a job seeker's access request, the user ID is extracted from the request and mapped to the specified slave server. The slave server obtains the job seeker's historical question and answer records and calculates the behavioral similarity with the positions in the final recommendation list. The calculation of behavioral similarity is based on the degree of match between the job seeker's historical behavior (such as viewing and applying for positions, etc.) and the recommended positions.

[0091] Based on the behavior similarity and the job ranking score, the personalized recommendation score is calculated. Job information with high personalized recommendation scores is selected and pushed to the job seeker's terminal through the persistent connection protocol. In this way, job seekers can get the job recommendations that best meet their needs and interests in a timely manner.

[0092] In this embodiment, by comprehensively considering the job seeker's question and answer records and job characteristics, the generated recommendation list is more in line with the actual needs of the job seeker. By storing the intelligent question and answer model of popular positions in the cache, the efficiency of responding to job seekers' requests is improved, and the user experience is enhanced. Based on the job seeker's historical behavior and question and answer records, personalized job information is provided to increase the chance of successful job search.

[0093] Figure 2 FIG. 1 is a schematic diagram of a structure of an intelligent processing platform for recruitment information based on cloud services according to an embodiment of the present invention. Figure 2 As shown, the platform includes: The first unit is used to receive recruitment information sent by multiple recruitment information sources, store the recruitment information in a cloud server, perform word segmentation and part-of-speech tagging on the recruitment information using a natural language processing model, extract the job title, salary range, work location, job requirements, job responsibilities and company information in the recruitment information to form initial job information, use a bidirectional long short-term memory neural network to perform semantic understanding and association analysis on the initial job information, generate a job semantic vector, match the job semantic vector with a preset industry dictionary, normalize the job description, generate normalized job information, and filter the normalized job information based on knowledge graph construction rules to obtain valid job information; The second unit is used to construct a job knowledge graph including node features and edge relationships according to the valid job information, perform convolution operation on the job knowledge graph using a graph neural network, extract correlation features between jobs, input the correlation features into a pre-trained deep neural network, generate a job feature vector, construct a multi-dimensional job profile based on the job feature vector, calculate the matching score between the multi-dimensional job profile and the job seeker resume profile in each dimension, obtain a job-talent matching score, sort the valid job information based on the job-talent matching score, and generate an initial recommendation list; The third unit is used to build an intelligent question and answer model for each position in the initial recommendation list, and the intelligent question and answer model includes a question understanding module and an answer generation module, wherein the question understanding module receives a question text input by a job seeker, and converts the question text into a question vector; the answer generation module retrieves relevant content in the valid position information based on the question vector, generates an answer text, records the question text and the answer text to form a question and answer record, adjusts the order of the positions in the initial recommendation list based on the question and answer record, generates a final recommendation list, deploys the final recommendation list and the intelligent question and answer model to the cloud server, and pushes personalized position information to the job seeker terminal through the cloud server.

[0094] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0095] According to a fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0096] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent processing method for recruitment information based on cloud services, characterized in that: include: Receive recruitment information sent by multiple recruitment information sources, and store the recruitment information in a cloud server, use a natural language processing model to perform word segmentation and part-of-speech tagging on the recruitment information, extract the job title, salary range, work location, job requirements, job responsibilities and company information in the recruitment information to form initial job information, use a bidirectional long short-term memory neural network to perform semantic understanding and association analysis on the initial job information, generate a job semantic vector, match the job semantic vector with a preset industry dictionary, normalize the job description, generate normalized job information, and filter the normalized job information based on the knowledge graph construction rules to obtain valid job information; Based on the valid job information, a job knowledge graph including node features and edge relationships is constructed, a graph neural network is used to perform convolution operations on the job knowledge graph, correlation features between jobs are extracted, the correlation features are input into a pre-trained deep neural network, a job feature vector is generated, a multi-dimensional job profile is constructed based on the job feature vector, a matching score between the multi-dimensional job profile and the job seeker resume profile in each dimension is calculated, a job-talent matching score is obtained, the valid job information is sorted based on the job-talent matching score, and an initial recommendation list is generated; An intelligent question-and-answer model is constructed for each position in the initial recommendation list, and the intelligent question-and-answer model includes a question understanding module and an answer generation module, wherein the question understanding module receives a question text input by a job seeker and converts the question text into a question vector; the answer generation module retrieves relevant content in the valid position information based on the question vector, generates an answer text, records the question text and the answer text to form a question-and-answer record, adjusts the order of the positions in the initial recommendation list based on the question-and-answer record, generates a final recommendation list, deploys the final recommendation list and the intelligent question-and-answer model to the cloud server, and pushes personalized position information to the job seeker terminal through the cloud server.

2. The method according to claim 1, characterized in that A bidirectional long short-term memory neural network is used to perform semantic understanding and association analysis on the initial job information to generate a job semantic vector, match the job semantic vector with a preset industry dictionary, normalize the job description, generate normalized job information, and filter the normalized job information based on knowledge graph construction rules to obtain valid job information including: Input the initial position information into a hierarchical attention encoder, the hierarchical attention encoder comprises a character-level encoding layer, a word-level encoding layer and a sentence-level encoding layer; the character-level encoding layer extracts character features through a convolutional neural network, and the word-level encoding layer and the sentence-level encoding layer extract word-level features and sentence-level features through a multi-head self-attention structure; input the character features, word-level features and sentence-level features into a cross-layer interaction module, and the gated update unit of the cross-layer interaction module selects and fuses the input features to generate an enhanced word vector sequence; The enhanced word vector sequence is input into a bidirectional long short-term memory neural network, wherein the bidirectional long short-term memory neural network comprises a memory enhancement layer and a semantic modeling layer; the memory enhancement layer stores long-term dependency features through an external memory matrix, and updates the external memory matrix based on an attention reading and writing mechanism; the semantic modeling layer processes features through a gating mechanism and residual connection and performs adversarial training to generate a position semantic vector; The position semantic vector is input into an industry dictionary matching module, and the industry dictionary matching module includes a multi-dimensional similarity calculation unit and a weight allocation unit; the multi-dimensional similarity calculation unit calculates the matching scores of the position semantic vector with standard terms at the lexical level, syntactic level and semantic level respectively; the weight allocation unit allocates weight coefficients to the matching scores at the three levels through an attention mechanism, selects corresponding standard terms based on the weighted matching scores, generates standardized position information, and filters the standardized position information based on knowledge graph construction rules to obtain valid position information.

3. The method according to claim 2, characterized in that The standardized job information is filtered based on the knowledge graph construction rules to obtain valid job information including: Constructing an integrity rule set, a consistency rule set and an authenticity rule set, wherein the integrity rule set includes necessary field existence constraints and value range constraints, the consistency rule set includes field dependency constraints and field logic constraints, and the authenticity rule set includes knowledge matching constraints and value rationality constraints; extracting the trigger condition features and constraint condition features of each rule in the integrity rule set, the consistency rule set and the authenticity rule set to generate a rule feature set; Input the rule feature set into the word vector model to obtain a rule feature vector, perform attention weighted fusion on the rule feature vector, and generate a rule feature matrix; Inputting the standardized position information into the convolutional neural network, extracting the field local features of the standardized position information, calculating the similarity between the field local features and the rule feature matrix, and obtaining the rule matching features; The rule matching feature is input into a conditional random field model, wherein the conditional random field model comprises a state feature extraction unit and a transfer feature extraction unit, wherein the state feature extraction unit performs state modeling on the rule matching feature, and the transfer feature extraction unit models the label sequence through a recurrent neural network to identify the information position in the standardized job information that does not meet the rule constraints and the corresponding rule type; Based on the information position that does not meet the rule constraints and the corresponding rule type, the fields of the information that does not meet the integrity rules in the standardized position information are supplemented and the value ranges are corrected, the conflicts of the information that does not meet the consistency rules are adjusted, and the information that does not meet the authenticity rules is deleted to generate valid position information that meets the constraints of the integrity rule set, the consistency rule set and the authenticity rule set.

4. The method according to claim 1, characterized in that: According to the valid job information, a job knowledge graph including node features and edge relationships is constructed, a graph neural network is used to perform convolution operation on the job knowledge graph, and correlation features between jobs are extracted. The correlation features are input into a pre-trained deep neural network to generate a job feature vector, including: Constructing a job knowledge graph based on valid job information, the job knowledge graph includes a node feature set and an edge relationship set; the node feature set includes job attribute features, skill requirement features and experience requirement features, wherein the job attribute features include salary range and work location, the skill requirement features include professional skills and general skills, and the experience requirement features include years of work and project experience; the edge relationship set includes job hierarchical relationships, skill correlation relationships and job conversion relationships; constructing a graph structure matrix based on the node feature set and the edge relationship set; The graph structure matrix is ​​input into a multi-task graph neural network, which includes a contrastive learning module, an attribute reconstruction module and a relationship prediction module. The contrastive learning module generates a node representation by maximizing the mutual information of adjacent position nodes; the attribute reconstruction module optimizes the feature representation by reconstructing the masked position node attributes; the relationship prediction module enhances the structure representation by predicting the edge relationship between nodes; multi-task joint optimization is performed based on the node representation, feature representation and structure representation to generate a fusion representation of the position node; The fusion representation is input into a graph convolutional network, edge relationships are aggregated on features of adjacent position nodes through a feature propagation layer, nonlinear transformation is performed on aggregated features through a feature update layer, and position feature information of each layer is maintained through a residual connection to generate position association features; The associated features are input into a pre-trained deep neural network, which includes an encoding layer and a decoding layer. The encoding layer maps the associated features to a low-dimensional space through a multi-layer perceptron, and the decoding layer obtains an initial position feature vector through reconstruction optimization, and the initial position feature vector is normalized to generate a final position feature vector.

5. The method according to claim 1, characterized in that Constructing a multi-dimensional job profile based on the job feature vector, calculating the matching score between the multi-dimensional job profile and the job seeker resume profile in each dimension to obtain a job-talent matching score, sorting the valid job information based on the job-talent matching score, and generating an initial recommendation list includes: The position feature vector is input into a multi-dimensional mapping network, including a professional skill dimension mapping network, a management capability dimension mapping network, a business experience dimension mapping network, and a general quality dimension mapping network. Each dimension mapping network performs feature transformation through a multi-layer perceptron and performs feature enhancement through residual connection and layer normalization to obtain initial dimension features; Performing self-attention calculation on the initial dimensional features to obtain an intra-dimensional feature association matrix, performing cross-attention calculation based on the feature association matrix to obtain inter-dimensional feature dependencies, and combining the features with inter-dimensional dependencies through learnable weights to obtain a multi-dimensional profile of the position; At the same time, the resume features of the job seeker are input into the multi-dimensional mapping network to perform dimension mapping and self-attention calculation processes to obtain a multi-dimensional portrait of the resume, the multi-dimensional portrait of the position and the multi-dimensional portrait of the resume are calculated in each dimension to match the importance weight and cosine similarity to obtain local dimensional matching features, and the multi-dimensional portrait of the position and the multi-dimensional portrait of the resume are input into the deep interaction network to obtain global matching features; The local dimension matching features and the global matching features are filtered through feature selection gating to generate filtered matching features, dynamic fusion weights are calculated for the filtered matching features, and weighted combinations are performed to obtain fused matching features; The gradient of the fused matching feature is calculated, and perturbations are added in the gradient direction to generate adversarial samples. The adversarial samples are input into a pre-built matching score prediction model for training to obtain a robust matching score prediction model. The position-talent matching pairs are scored according to the matching score prediction model to obtain the position-talent matching scores. The valid position information is sorted according to the position-talent matching scores to generate an initial recommendation list.

6. The method according to claim 5, characterized in that Calculating the gradient of the fused matching feature, adding perturbations in the gradient direction to generate adversarial samples, inputting the adversarial samples into a pre-built matching score prediction model for training, and obtaining a robust matching score prediction model includes: Inputting the fused matching feature into a pre-built matching score prediction model to obtain a predicted matching score, and calculating an original loss function value based on the predicted matching score and the true matching score; Calculating the gradient of the original loss function value with respect to the fusion matching feature, normalizing the gradient to obtain a normalized gradient; constructing a feature correlation matrix based on the normalized gradient, extracting the main eigenvector of the feature correlation matrix; projecting the normalized gradient to the orthogonal complement space of the main eigenvector to obtain a projected gradient; Adding a disturbance amount to the fused matching feature along the direction of the projected gradient to obtain an initial adversarial sample, and truncating the initial adversarial sample based on a disturbance threshold to obtain an optimized adversarial sample; Inputting the optimized adversarial sample into a matching score prediction model to obtain an adversarial predicted matching score, and calculating an adversarial loss function value based on the adversarial predicted matching score and the true matching score; The original loss function value and the adversarial loss function value are weightedly summed to obtain a total loss function value. The model parameters of the matching score prediction model are updated based on the total loss function value to obtain updated model parameters. Repeat the iteration until the total loss function value is less than the preset convergence threshold to obtain a trained matching score prediction model.

7. The method according to claim 1, characterized in that Adjusting the order of positions in the initial recommendation list based on the question and answer records to generate a final recommendation list, deploying the final recommendation list and the intelligent question and answer model to the cloud server, and pushing personalized position information to the job seeker terminal through the cloud server includes: The question text in the question and answer record is converted into a question word vector sequence through a word embedding layer, and the question word vector sequence is input into a bidirectional long short-term memory network to obtain a question semantic vector; at the same time, the answer text in the question and answer record is input into a pre-trained language model to obtain an answer feature vector, and the question semantic vector and the answer feature vector are input into a cross attention network to obtain a question and answer interaction vector; The question-answer interaction vector is used as a query matrix, and a multi-head attention calculation is performed on the feature association matrix of the positions in the initial recommendation list to obtain a feature attention score, and the feature attention score is multiplied by the corresponding professional weight coefficient and the sum is obtained to obtain a combined attention score; Calculating a temporal decay coefficient based on the timestamp of the question and answer record, adaptively fusing the product of the combined attention score and the temporal decay coefficient with the pre-obtained position-talent matching score to obtain a position ranking score, and reordering the positions in the initial recommendation list based on the position ranking score to generate a final recommendation list, wherein each list item in the final recommendation list includes position information, ranking score, timestamp, access count, and intelligent question and answer model; The intelligent question-answering model corresponding to the position whose access count exceeds the preset threshold is stored in a distributed cache cluster, and the final recommendation list is deployed to a cloud server master-slave cluster, where the master server maintains a global index and the slave servers form a processing cluster based on a consistent hashing algorithm; When receiving a job seeker's access request, extracting a user ID from the access request and mapping it to a designated slave server, the slave server obtaining the job seeker's historical question and answer records and calculating the behavioral similarity with the positions in the final recommendation list; A personalized recommendation score is calculated based on the behavior similarity and the job ranking score, the job information with the highest personalized recommendation score is selected, and the job information is pushed to the job seeker terminal through a long connection protocol.

8. An intelligent processing platform for recruitment information based on cloud services, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to receive recruitment information sent by multiple recruitment information sources, store the recruitment information in a cloud server, perform word segmentation and part-of-speech tagging on the recruitment information using a natural language processing model, extract the job title, salary range, work location, job requirements, job responsibilities and company information in the recruitment information to form initial job information, use a bidirectional long short-term memory neural network to perform semantic understanding and association analysis on the initial job information, generate a job semantic vector, match the job semantic vector with a preset industry dictionary, normalize the job description, generate normalized job information, and filter the normalized job information based on knowledge graph construction rules to obtain valid job information; The second unit is used to construct a job knowledge graph including node features and edge relationships according to the valid job information, perform convolution operation on the job knowledge graph using a graph neural network, extract correlation features between jobs, input the correlation features into a pre-trained deep neural network, generate a job feature vector, construct a multi-dimensional job profile based on the job feature vector, calculate the matching score between the multi-dimensional job profile and the job seeker resume profile in each dimension, obtain a job-talent matching score, sort the valid job information based on the job-talent matching score, and generate an initial recommendation list; The third unit is used to build an intelligent question and answer model for each position in the initial recommendation list, and the intelligent question and answer model includes a question understanding module and an answer generation module, wherein the question understanding module receives a question text input by a job seeker, and converts the question text into a question vector; the answer generation module retrieves relevant content in the valid position information based on the question vector, generates an answer text, records the question text and the answer text to form a question and answer record, adjusts the order of the positions in the initial recommendation list based on the question and answer record, generates a final recommendation list, deploys the final recommendation list and the intelligent question and answer model to the cloud server, and pushes personalized position information to the job seeker terminal through the cloud server.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images