A big data-based recruitment registrant information query system and method

Abstract

The application relates to the technical field of information query, in particular to a recruitment registration personnel information query system and method based on big data, which comprises a data acquisition module, a data verification module and a semantic detection module. The data acquisition module extracts resume data from a registration resume document and outputs structured resume fields. The data verification module verifies the authenticity of the resume fields in association with an external database and outputs a verification result. In use, the semantic detection module jointly participates in judgment through fuzziness detection, exaggeration word recognition, semantic similarity and abnormal probability distribution, stably identifies risks in different contexts, performs background cross verification and multi-dimensional fusion based on a big data model, is favorable for improving system robustness, is favorable for reducing human subjective judgment errors, effectively identifies potential false information such as exaggeration and fuzzy description, reduces the employment cost of 'good-looking resume but poor ability', the system automatically completes initial screening trust degree scoring, does not need staff to repeatedly check, and is favorable for improving recruitment efficiency.

Description

Technical Field

[0001] The present invention relates to the field of information query technology, and in particular to a recruitment registration information query system and method based on big data. Background Art

[0002] Recruitment registration information query based on big data refers to the use of big data technology and analysis methods to process and analyze various data collected during the recruitment process in order to screen and evaluate applicants more efficiently and accurately.

[0003] The patent publication number is CN107315836A, which states in its specification that "the present invention discloses a personnel information query system, which relates to the field of information query technology. It includes an information query system, which includes a database, a Web server, a personal terminal and an administrator; the database is connected to an intelligent query unit through a data transmission module, the intelligent query unit is electrically connected to a processor, and the processor is connected to the Web server through a data transmission unit; the personal terminal is connected to the Web server through a voice conversion transmission unit, the voice conversion transmission unit includes a converter and a communication module, the voice recognizer is connected to the communication module through the converter, the personal terminal is connected to the Web server through a text conversion transmission unit, the text conversion transmission unit includes a decoder and a network Network transmission unit Network transmission unit, the Web server is connected to the voice receptionist terminal through the intelligent access module. The present invention can use text query and voice consultation to query the database according to customer needs. Although the above technology provides a comprehensive platform that supports text query and voice query, realizes convenient query, real-time data update and centralized authority management, and prevents tampering and fraud, so as to solve the problems of difficulty in enterprise recruitment and untimely information acquisition. However, when responding to the information query of recruitment applicants, the above technology has false information and abnormal data in the resumes of some applicants, making it difficult to identify accurate resume information. At the same time, information verification relies on manual labor, which has a long cycle and high cost. There is a lack of scoring and marking mechanism for the credibility of applicant data, and background cross-verification based on big data models cannot be performed.

[0004] In summary, developing a recruitment registration information query system and method based on big data is still a key issue that needs to be solved urgently in the field of information query technology. Summary of the Invention

[0005] The purpose of the present invention is to solve the problems existing in the prior art. Although the above-mentioned technology provides a comprehensive platform supporting text query and voice query, realizes convenient query, real-time data update and centralized authority management, and prevents tampering and fraud, so as to solve the problems of difficulty in enterprise recruitment and untimely information acquisition, when responding to the information query of recruitment applicants, the resume information of some applicants is false and the data is abnormal, which makes it difficult to identify accurate resume information. At the same time, information verification relies on manual labor, with a long cycle and high cost. There is a lack of a scoring and marking mechanism for the credibility of the applicant's data, and background cross-verification cannot be performed based on a big data model. The present invention provides a recruitment applicant information query system and method based on big data.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The present invention provides a recruitment registration personnel information query system based on big data, comprising: a data acquisition module, which extracts resume data according to the registration resume document and outputs structured resume fields;

[0008] A data verification module, which associates with an external database to verify the authenticity of the resume fields and outputs the verification results;

[0009] The semantic detection module uses the trained BERT semantic model and the fraudulent resume corpus to compare the resume fields and output an anomaly risk score;

[0010] The trust scoring module uses a multi-factor scoring algorithm to perform comprehensive calculations based on the verification results and the abnormal risk score to obtain a credibility score, and feeds it back to the semantic detection module.

[0011] Furthermore, the workflow of the data acquisition module is as follows:

[0012] Assume that the application resume document is , where is the document space. Obtain the original text through the OCR parser and clean the original text. Use the bidirectional language model (BiLSTM-CRF) to segment the cleaned resume document to obtain the word sequence. Then, obtain the word vector matrix through the word embedding function. Then, use the multi-head attention mechanism to weightedly combine the context information to obtain the context-enhanced resume document. Build a named entity recognition model, input the context-enhanced resume document into the BiLSTM-CRF layer, and predict the corresponding label sequence. The expression formula is: where Q i is the optimal label sequence finally selected, It means to traverse all possible tag sequences and select the one with the highest score. is the sum of the time step t of the entire sequence, which has a total of W words and fields, R i tis the feature vector at the t-th position belonging to the i-th sample, E tg It is used to convert vector The weight matrix mapped to the score of each label, Trans(T t-1 ,T t ) represents the transfer score between labels, T t-1 ,T t is the label at the t-1th, tth position in the sequence, y is the entire search space, the corresponding entity field is extracted from the label sequence, and the structured resume field is output. The expression formula is:

[0013] where Y i is the set of structured fields extracted from the i-th resume, α i is the set of all “education experience” fields extracted from the i-th resume, I j is the school description in the jth education information, O i The time period of the jth educational information, δ is the complete set of all academic qualifications entities in the educational experience, ε α is the complete set of all time periods in the educational experience, β i is the set of all “work experience” fields extracted from the i-th resume, P k is the job description in the kth job information, O k The time period of the kth job information, φ is the complete set of all companies in the work history, ε β is the complete set of all time periods in the work experience, χ i is the set of all “project experience” fields extracted from the i-th resume, U m is the mth project experience entry, It is a complete collection of all project descriptions.

[0014] Furthermore, the workflow of the data verification module is:

[0015] Receive the resume field, map the content of the resume field into a structured request vector, construct a heterogeneous query matrix, and output a query vector. The query vector initiates a request to the external database through the API, sets the entity set to be returned from the external database, and performs matching calculations and time intersection matching calculations for each resume field entity. At the same time, all matching degrees of each resume field entity are normalized and fused into a confidence score, expressed as follows:

[0016] Among them D η =log(1+RankScore(η)), where A x is the overall A of candidate resume x x ∈[0,1]

[0017] Trust score, S x is the set of all verification dimensions associated with resume x, η∈S x is a dimension item in the verification dimension, γ(x,η) is the score result of resume x on the verification dimension η, D η is the importance weight of the verification dimension η, RankScore(η) is the authority rating of the verification dimension, and finally, the verification result is output, expressed as: VR i is the field verification result set corresponding to the i-th resume, express Is the set of fields to be verified in the i-th resume F i A field of Represented as fields Map a verification result label, ι is the trust threshold of the trust score, ι l is the suspicious lower limit threshold of the trust score, otherwise A x In ι l ≤A x <ι indicates uncertainty and requires manual review. F is a mapping function for verifying resume fields. RS i is the set of structured fields extracted from the i-th resume, → represents a "mapping relationship" A x is the overall trust score of candidate resume x.

[0018] Furthermore, the workflow of the semantic detection module is as follows:

[0019] Input the resume field into the trained BERT semantic model to extract the sentence vector G t , thereby constructing a semantic vector matrix, using the annotated fraudulent resume corpus to generate an abnormal semantic vector set, and using the Gaussian mixture model (GMM) to model the distribution of abnormal semantic vectors to obtain the probability density function, expressed as: in It is a sample The overall probability density value of is the sum of all j″th Gaussian distributions, J is the number of Gaussian distributions, π j″ is the mixing coefficient of the j″th Gaussian distribution, is the probability density of the j″th Gaussian distribution at the point, κ j″ is the mean, ∑ j″ is the covariance, and the language ambiguity detection score based on the sentence vector is: The semantic similarity comparison score based on the sentence vector is: The probability density function is used to calculate the sentence vector and perform log-likelihood scoring: An exaggeration detector is constructed based on a set of common fictitious keywords, and the sentence vector is scored for exaggeration detection:

[0020] Furthermore, the workflow of the semantic detection module is as follows:

[0021] The comparison of the resume fields uses an adjustable weight combination to fuse the scores of each dimension, expressed as follows: Among them L t′ is the total abnormal risk score of the t′th resume, is the language ambiguity detection score, is the semantic similarity comparison score, is the log-likelihood score, is the exaggerated detection score, λ1,λ2,λ3,λ4 are the weighted coefficients of the abnormality score of each dimension, The sum of all weight coefficients must be 1 for normalization, and the output anomaly risk score is: Among them L i It is L i The final trust risk score of the resume, is the risk score L of this resume in the t′th round in history t′ Accumulate, L t′ is the abnormal risk score obtained by the t′th detection, μ is the time impact weight factor, is the current scoring time t′ relative to the maximum number of rounds The standardized time weight of Z is the total number of rounds of cumulative scoring, Indicates the largest time point number among all time periods.

[0022] Furthermore, the workflow of the trust scoring module is as follows:

[0023] The multi-factor scoring algorithm sets the trust score of each field to Expression formula: in It is The trust score of the k′th field in the resume, It is The external verification credibility score of the k′th field in the resume, It is The abnormal risk score of the k′th field in the resume, is the "reversal score" of abnormal risk, are weighted coefficients for the credibility of external verification results and the anti-fraud tendency score of semantic detection, respectively. Dynamic field weights are introduced to perform weighted adjustment on the trust score of each field. Field weights are obtained through expert experience, historical data, and adaptive learning. The trust scores of all fields are combined using a weighted average method to obtain the overall credibility score of the resume field. The expression formula is: Among them B i″′ is the comprehensive trust score of the i′′th resume, k′ is the field number representing the k′th structured field, N″ is the total number of structured fields contained in the current resume, is the importance weight coefficient of the k′th field, It is is the trust score of the k′th field in a resume, and adk′ is the adjustment factor.

[0024] Furthermore, the workflow of the trust scoring module is as follows:

[0025] The credibility score of each calculation is fed back to the semantic detection module, and the BERT semantic model is updated with real recruitment data. The goal of the BERT semantic model update is to maximize the accuracy of the post-recruitment verification data, which is defined as a loss function based on real recruitment data, expressed as: where ξ fdbk is the loss function, Indicates the summation of the resume data from the 1st to the N′′′′′′′′′′′′′′, where N′′′′ is the total number of resumes used for calibration. i″ TrueStatus i″ )} is the sum of the logarithmic values ​​of the conditional probability, H(B i″ TrueStatus i″ ) is the conditional probability, B i″ is the trust score of the i″th resume, TrueStatus i″ is the true and credible state label of the i″th resume after actual verification after recruitment, and is obtained by optimizing the loss function ξ fdbk , and gradually adjust the model parameters.

[0026] On the other hand, the present invention also provides a recruitment registration personnel information query method based on big data, which includes the following steps:

[0027] S1. Extract resume data based on the application resume document and output structured resume fields;

[0028] S2. Connecting to an external database to verify the authenticity of the resume fields and outputting the verification results;

[0029] S3. Use the trained BERT semantic model and the fraudulent resume corpus to compare the resume fields and output an abnormality risk score;

[0030] S4. A multi-factor scoring algorithm is used to perform a comprehensive calculation based on the verification results and the abnormal risk score to obtain a credibility score, and the credibility score is fed back to the semantic detection module.

[0031] Furthermore, in step S1, the method of extracting resume data from the registration resume document and outputting structured resume fields is as follows:

[0032] Assume that the application resume document is , where is the document space. Obtain the original text through the OCR parser and clean the original text. Use the bidirectional language model (BiLSTM-CRF) to segment the cleaned resume document to obtain the word sequence. Then, obtain the word vector matrix through the word embedding function. Then, use the multi-head attention mechanism to weightedly combine the context information to obtain the context-enhanced resume document. Build a named entity recognition model, input the context-enhanced resume document into the BiLSTM-CRF layer, and predict the corresponding label sequence. The expression formula is: where Q i is the optimal label sequence finally selected, It means to traverse all possible tag sequences and select the one with the highest score. The sum of the time steps t of the entire sequence has a total of W words and fields, is the feature vector at the t-th position belonging to the i-th sample, E tg It is used to convert vector The weight matrix mapped to the score of each label, Trans(T t-1 ,T t ) represents the transfer score between labels, T t-1 ,T t is the label at the t-1th, tth position in the sequence, y is the entire search space, the corresponding entity field is extracted from the label sequence, and the structured resume field is output. The expression formula is:

[0033] where Y i is the set of structured fields extracted from the i-th resume, α i is the set of all “education experience” fields extracted from the i-th resume, I j is the school description in the jth education information, O i The time period of the jth educational information, δ is the complete set of all academic qualifications entities in the educational experience, ε α is the complete set of all time periods in the educational experience, β i is the set of all “work experience” fields extracted from the i-th resume, Pk is the job description in the kth job information, O k The time period of the kth job information, φ is the complete set of all companies in the work history, ε β is the complete set of all time periods in the work experience, χ i is the set of all “project experience” fields extracted from the i-th resume, U m is the mth project experience entry, It is the complete set of all project descriptions;

[0034] Furthermore, in step S2, the external database is associated with the resume fields to verify the authenticity and output the verification result in the following manner:

[0035] Receive the resume field, map the content of the resume field into a structured request vector, construct a heterogeneous query matrix, and output a query vector. The query vector initiates a request to the external database through the API, sets the entity set to be returned from the external database, and performs matching calculations and time intersection matching calculations for each resume field entity. At the same time, all matching degrees of each resume field entity are normalized and fused into a confidence score, expressed as follows: Among them D η =log(1+RankScore(η)), where A x is the overall A of candidate resume x x ∈[0,1]

[0036] Trust score, S x is the set of all verification dimensions associated with resume x, η∈S x is a dimension item in the verification dimension, γ(x,η) is the score result of resume x on the verification dimension η, D η is the importance weight of the verification dimension η, RankScore(η) is the authority rating of the verification dimension, and finally, the verification result is output, expressed as: VR i is the field verification result set corresponding to the i-th resume, express Is the set of fields to be verified in the i-th resume F i A field of Represented as fields Map a verification result label, ι is the trust threshold of the trust score, ι l is the suspicious lower limit threshold of the trust score, otherwise A x In ι l ≤A x <ι indicates uncertainty and requires manual review. F is a mapping function for verifying resume fields. RSi is the set of structured fields extracted from the i-th resume, → represents a "mapping relationship" A x is the overall trust score of candidate resume x.

[0037] Furthermore, in step S3, the trained BERT semantic model is used with the fraudulent resume corpus to compare the resume fields and output an abnormality risk score as follows:

[0038] Input the resume field into the trained BERT semantic model to extract the sentence vector G t , thereby constructing a semantic vector matrix, using the annotated fraudulent resume corpus to generate an abnormal semantic vector set, and using the Gaussian mixture model (GMM) to model the distribution of abnormal semantic vectors to obtain the probability density function, expressed as: in It is a sample The overall probability density value of is the sum of all j″th Gaussian distributions, J is the number of Gaussian distributions, π j″ is the mixing coefficient of the j″th Gaussian distribution, is the probability density of the j″th Gaussian distribution at the point, κ j″ is the mean, ∑ j″ is the covariance, and the language ambiguity detection score based on the sentence vector is: The semantic similarity comparison score based on the sentence vector is: The probability density function is used to calculate the sentence vector and perform log-likelihood scoring: An exaggeration detector is constructed based on a set of common fictitious keywords, and the sentence vector is scored for exaggeration detection: The comparison of the resume fields uses an adjustable weight combination to fuse the scores of each dimension, expressed as follows: Among them L t′ is the total abnormal risk score of the t′th resume, is the language ambiguity detection score, is the semantic similarity comparison score, is the log-likelihood score, is the exaggerated detection score, λ1,λ2,λ3,λ4 are the weighted coefficients of the abnormality score of each dimension, The sum of all weight coefficients must be 1 for normalization, and the output anomaly risk score is: Among them L i It is L i The final trust risk score of the resume, is the risk score L of this resume in the t′th round in history t′ Accumulate, L t′is the abnormal risk score obtained by the t′th detection, μ is the time impact weight factor, is the current scoring time t′ relative to the maximum number of rounds The standardized time weight of Z is the total number of rounds of cumulative scoring, Indicates the largest time point number among all time periods;

[0039] Furthermore, in step S4, a multi-factor scoring algorithm is used to perform a comprehensive calculation based on the verification result and the abnormal risk score to obtain a credibility score, and the credibility score is fed back to the semantic detection module as follows:

[0040] The multi-factor scoring algorithm sets the trust score of each field to Expression formula: in It is The trust score of the k′th field in the resume, It is The external verification credibility score of the k′th field in the resume, It is The abnormal risk score of the k′th field in the resume, is the "reversal score" of abnormal risk, are weighted coefficients for the credibility of external verification results and the anti-fraud tendency score of semantic detection, respectively. Dynamic field weights are introduced to perform weighted adjustment on the trust score of each field. Field weights are obtained through expert experience, historical data, and adaptive learning. The trust scores of all fields are combined using a weighted average method to obtain the overall credibility score of the resume field. The expression formula is: Among them B i″′ This is the i″′th copy

[0041] B i″′ =[0,1]

[0042] The comprehensive trust score of the resume, k′ is the field number representing the k′th structured field, N″ is the total number of structured fields contained in the current resume, is the importance weight coefficient of the k′th field, It is The trust score of the k′th field in the resume is calculated, and adk′ is the adjustment factor. The credibility score calculated each time is fed back to the semantic detection module, and the BERT semantic model is updated with real recruitment data. The goal of the BERT semantic model update is to maximize the accuracy of the post-recruitment verification data, which is defined as a loss function based on real recruitment data, expressed as: where ξ fdbk is the loss function, Indicates the summation of the resume data from the 1st to the N′′′′′′′′′′′′′′, where N′′′′ is the total number of resumes used for calibration. i″ TrueStatus i″ )} is the sum of the logarithmic values ​​of the conditional probability, H(B i″ TrueStatus i″ ) is the conditional probability, B i″ is the trust score of the i″th resume, TrueStatus i″ is the true and credible state label of the i″th resume after actual verification after recruitment, and is obtained by optimizing the loss function ξ fdbk , and gradually adjust the model parameters.

[0043] Beneficial effects

[0044] Compared with the known public technology, the technical solution provided by the present invention has the following advantages:

[0045] Beneficial effects:

[0046] When the present invention is in use, fuzzy detection, exaggerated word recognition, semantic similarity, and abnormal probability distribution jointly participate in judgment, stably identify risks in different contexts, and perform background cross-validation and multi-dimensional fusion based on big data models, which is conducive to improving system robustness, reducing human subjective judgment errors, and effectively identifying potential false information such as exaggerated packaging and vague descriptions, reducing the employment cost of "good-looking resumes but mismatched abilities". The system automatically completes the initial screening trust score, without the need for staff to repeatedly check, which is conducive to improving recruitment efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a system diagram of a recruitment registration information query system based on big data of the present invention;

[0048] Figure 2 This is a flow chart of a recruitment registration information query method based on big data of the present invention. DETAILED DESCRIPTION

[0049] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions 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 embodiments described 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 making creative efforts should fall within the scope of protection of the present invention.

[0050] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or are inherent to these processes, methods, products or devices.

[0051] The present invention is described in further detail below with reference to the accompanying drawings:

[0052] Example 1:

[0053] like Figure 1 As shown, the present invention provides a recruitment registration personnel information query system based on big data, including: a data acquisition module, which extracts resume data according to the registration resume document and outputs structured resume fields;

[0054] Furthermore, the workflow of the data acquisition module is as follows:

[0055] Assume that the application resume document is , where is the document space. Obtain the original text through the OCR parser and clean the original text. Use the bidirectional language model (BiLSTM-CRF) to segment the cleaned resume document to obtain the word sequence. Then, obtain the word vector matrix through the word embedding function. Then, use the multi-head attention mechanism to weightedly combine the context information to obtain the context-enhanced resume document. Build a named entity recognition model, input the context-enhanced resume document into the BiLSTM-CRF layer, and predict the corresponding label sequence. The expression formula is: where Q i is the optimal label sequence finally selected, It means to traverse all possible tag sequences and select the one with the highest score. The sum of the time steps t of the entire sequence has a total of W words and fields, is the feature vector at the t-th position belonging to the i-th sample, E tg It is used to convert vector The weight matrix mapped to the score of each label, Trans(T t-1 ,T t ) represents the transfer score between labels, T t-1 ,T tis the label at the t-1th, tth position in the sequence, y is the entire search space, the corresponding entity field is extracted from the label sequence, and the structured resume field is output. The expression formula is:

[0056] where Y i is the set of structured fields extracted from the i-th resume, α i is the set of all “education experience” fields extracted from the i-th resume, I j is the school description in the jth education information, O i The time period of the jth educational information, δ is the complete set of all academic qualifications entities in the educational experience, ε α is the complete set of all time periods in the educational experience, β i is the set of all “work experience” fields extracted from the i-th resume, P k is the job description in the kth job information, O k The time period of the kth job information, φ is the complete set of all companies in the work history, ε β is the complete set of all time periods in the work experience, χ i is the set of all “project experience” fields extracted from the i-th resume, U m is the mth project experience entry, It is the complete set of all project descriptions;

[0057] Specifically, the data acquisition module realizes automatic extraction of the entire process from resume documents to structured fields, which is convenient for reducing the manual input and classification work of staff. At the same time, information is accurately extracted from PDF, pictures and Word resumes, which is convenient for improving the adaptability of the present invention. The introduction of BiLSTM+CRF combined with the multi-head attention mechanism is conducive to improving the entity boundary recognition effect in Chinese resumes.

[0058] A data verification module, which associates with an external database to verify the authenticity of the resume fields and outputs the verification results;

[0059] Furthermore, the workflow of the data verification module is:

[0060] Receive the resume field, map the content of the resume field into a structured request vector, construct a heterogeneous query matrix, and output a query vector. The query vector initiates a request to the external database through the API, sets the entity set to be returned from the external database, and performs matching calculations and time intersection matching calculations for each resume field entity. At the same time, all matching degrees of each resume field entity are normalized and fused into a confidence score, expressed as follows: Among them D η =log(1+RankScore(η)), where A xis the overall A of candidate resume x x ∈[0,1] trust score, S x is the set of all verification dimensions associated with resume x, η∈S x is a dimension item in the verification dimension, γ(x,η) is the score result of resume x on the verification dimension η, D η is the importance weight of the verification dimension η, RankScore(η) is the authority rating of the verification dimension, and finally, the verification result is output, expressed as: VR i is the field verification result set corresponding to the i-th resume, express Is the set of fields to be verified in the i-th resume F i A field of Represented as fields Map a verification result label, ι is the trust threshold of the trust score, ι l is the suspicious lower limit threshold of the trust score, otherwise A x In ι l ≤A x <ι indicates uncertainty and requires manual review. F is a mapping function for verifying resume fields. RS i is the set of structured fields extracted from the i-th resume, → represents a "mapping relationship" A x is the overall trust score of candidate resume x;

[0061] Specifically, the data verification module is used to efficiently connect to the authoritative platform to realize the authenticity identification of core fields such as academic qualifications and experience, which is convenient for reducing the probability of false resumes passing. The data verification module not only matches the field values, but also introduces comprehensive judgment of factors such as time, authority, and weight, making the verification results more credible. By setting the scoring interval, a flexible strategy of "automatic pass - automatic rejection - pending manual review" is realized, which is conducive to avoiding accidental harm to outstanding talents.

[0062] The semantic detection module uses the trained BERT semantic model and the fraudulent resume corpus to compare the resume fields and output an anomaly risk score;

[0063] Furthermore, the workflow of the semantic detection module is as follows:

[0064] Input the resume field into the trained BERT semantic model to extract the sentence vector G t , thereby constructing a semantic vector matrix, using the annotated fraudulent resume corpus to generate an abnormal semantic vector set, and using the Gaussian mixture model (GMM) to model the distribution of abnormal semantic vectors to obtain the probability density function, expressed as: in It is a sample The overall probability density value of is the sum of all j″th Gaussian distributions, J is the number of Gaussian distributions, π j″ is the mixing coefficient of the j″th Gaussian distribution, is the probability density of the j″th Gaussian distribution at the point, κ j″ is the mean, ∑ j″ is the covariance, and the language ambiguity detection score based on the sentence vector is: The semantic similarity comparison score based on the sentence vector is: The probability density function is used to calculate the sentence vector and perform log-likelihood scoring: An exaggeration detector is constructed based on a set of common fictitious keywords, and the sentence vector is scored for exaggeration detection:

[0065] Furthermore, the workflow of the semantic detection module is as follows:

[0066] The comparison of the resume fields uses an adjustable weight combination to fuse the scores of each dimension, expressed as follows: Among them L t′ is the total abnormal risk score of the t′th resume, is the language ambiguity detection score, is the semantic similarity comparison score, is the log-likelihood score, is the exaggerated detection score, λ1,λ2,λ3,λ4 are the weighted coefficients of the abnormality score of each dimension, The sum of all weight coefficients must be 1 for normalization, and the output anomaly risk score is: Among them L i It is L i The final trust risk score of the resume, is the risk score L of this resume in the t′th round in history t′ Accumulate, L t′ is the abnormal risk score obtained by the t′th detection, μ is the time impact weight factor, is the current scoring time t′ relative to the maximum number of rounds The standardized time weight of Z is the total number of rounds of cumulative scoring, Indicates the largest time point number among all time periods;

[0067] Specifically, by combining deep semantic modeling with abnormal corpus, the authenticity of content can be accurately judged from the linguistic level, and resumes with "strong-looking content but no substance" can be identified, facilitating the precise identification of false information. Through semantic similarity comparison, a large amount of plagiarized and template-based resume content can be discovered, thus avoiding misjudgments in talent assessments and improving anti-template capabilities. Semantic models and abnormal corpora are continuously updated based on newly annotated resumes, facilitating continuous improvement in recognition capabilities, and facilitating dynamic adaptation and continuous optimization. Through fuzzy detection, exaggeration recognition, semantic similarity, and abnormal probability distribution, risks can be stably identified in different contexts, facilitating multi-dimensional integration and improving system robustness.

[0068] A trust scoring module uses a multi-factor scoring algorithm to perform a comprehensive calculation based on the verification results and the abnormal risk score to obtain a credibility score, and feeds it back to the semantic detection module;

[0069] Furthermore, the workflow of the trust scoring module is as follows:

[0070] The multi-factor scoring algorithm sets the trust score of each field to Expression formula: in It is The trust score of the k′th field in the resume, It is The external verification credibility score of the k′th field in the resume, It is The abnormal risk score of the k′th field in the resume, is the "reversal score" of abnormal risk, are weighted coefficients for the credibility of external verification results and the anti-fraud tendency score of semantic detection, respectively. Dynamic field weights are introduced to perform weighted adjustment on the trust score of each field. Field weights are obtained through expert experience, historical data, and adaptive learning. The trust scores of all fields are combined using a weighted average method to obtain the overall credibility score of the resume field. The expression formula is: Among them B i″′ is the comprehensive trust score of the i′′th resume, k′ is the field number representing the k′th structured field, N″ is the total number of structured fields contained in the current resume, is the importance weight coefficient of the k′th field, It is is the trust score of the k′th field in a resume, and adk′ is the adjustment factor.

[0071] Furthermore, the workflow of the trust scoring module is as follows:

[0072] The credibility score of each calculation is fed back to the semantic detection module, and the BERT semantic model is updated with real recruitment data. The goal of the BERT semantic model update is to maximize the accuracy of the post-recruitment verification data, which is defined as a loss function based on real recruitment data, expressed as: where ξ fdbk is the loss function, Indicates the summation of the resume data from the 1st to the N′′′′′′′′′′′′′′, where N′′′′ is the total number of resumes used for calibration. i″ TrueStatus i″ )} is the sum of the logarithmic values ​​of the conditional probability, H(B i″ TrueStatus i″ ) is the conditional probability, B i″ is the trust score of the i″th resume, TrueStatus i″ is the true and credible state label of the i″th resume after actual verification after recruitment, and is obtained by optimizing the loss function ξ fdbk , gradually adjust the model parameters;

[0073] Specifically, a comprehensive calculation is performed through the trust scoring module to obtain the credibility score of the entire resume document. After the recruitment is completed, the staff confirms whether the candidate has actually passed the probation period and the degree of job matching. This is used as the real label to reversely update the parameters of the BERT semantic model, and continuously optimize the detection accuracy of exaggerations and false descriptions in the future. This is conducive to reducing human subjective judgment errors, effectively identifying potential false information such as exaggerated packaging and vague descriptions, and reducing the employment cost of "good-looking resumes but not matching the capabilities". The system automatically completes the initial screening trust score, without the need for staff to repeatedly check, which is conducive to improving recruitment efficiency.

[0074] Example 2:

[0075] like Figure 2 As shown, Example 2 provides a recruitment registration personnel information query method based on big data, which includes the following steps:

[0076] S1. Extract resume data based on the application resume document and output structured resume fields;

[0077] S2. Connecting to an external database to verify the authenticity of the resume fields and outputting the verification results;

[0078] S3. Use the trained BERT semantic model and the fraudulent resume corpus to compare the resume fields and output an abnormality risk score;

[0079] S4. A multi-factor scoring algorithm is used to perform a comprehensive calculation based on the verification results and the abnormal risk score to obtain a credibility score, and the credibility score is fed back to the semantic detection module.

[0080] Furthermore, in step S1, the method of extracting resume data from the registration resume document and outputting structured resume fields is as follows:

[0081] Assume that the application resume document is , where is the document space. Obtain the original text through the OCR parser and clean the original text. Use the bidirectional language model (BiLSTM-CRF) to segment the cleaned resume document to obtain the word sequence. Then, obtain the word vector matrix through the word embedding function. Then, use the multi-head attention mechanism to weightedly combine the context information to obtain the context-enhanced resume document. Build a named entity recognition model, input the context-enhanced resume document into the BiLSTM-CRF layer, and predict the corresponding label sequence. The expression formula is: where Q i is the optimal label sequence finally selected, It means to traverse all possible tag sequences and select the one with the highest score. The sum of the time steps t of the entire sequence has a total of W words and fields, is the feature vector at the t-th position belonging to the i-th sample, E tg is used to convert the vector R i t The weight matrix mapped to the score of each label, Trans(T t-1 ,T t ) represents the transfer score between labels, T t-1 ,T t is the label at the t-1th, tth position in the sequence, y is the entire search space, the corresponding entity field is extracted from the label sequence, and the structured resume field is output. The expression formula is:

[0082] where Y i is the set of structured fields extracted from the i-th resume, α i is the set of all “education experience” fields extracted from the i-th resume, I j is the school description in the jth education information, O i The time period of the jth educational information, δ is the complete set of all academic qualifications entities in the educational experience, ε α is the complete set of all time periods in the educational experience, β i is the set of all “work experience” fields extracted from the i-th resume, P k is the job description in the kth job information, O kThe time period of the kth job information, φ is the complete set of all companies in the work history, ε β is the complete set of all time periods in the work experience, χ i is the set of all “project experience” fields extracted from the i-th resume, U m is the mth project experience entry, It is the complete set of all project descriptions;

[0083] Furthermore, in step S2, the external database is associated with the resume fields to verify the authenticity and output the verification result in the following manner:

[0084] Receive the resume field, map the content of the resume field into a structured request vector, construct a heterogeneous query matrix, and output a query vector. The query vector initiates a request to the external database through the API, sets the entity set to be returned from the external database, and performs matching calculations and time intersection matching calculations for each resume field entity. At the same time, all matching degrees of each resume field entity are normalized and fused into a confidence score, expressed as follows:

[0085] Among them D η =log(1+RankScore(η)), where A x is the overall A of candidate resume x x ∈[0,1]

[0086] Trust score, S x is the set of all verification dimensions associated with resume x, η∈S x is a dimension item in the verification dimension, γ(x,η) is the score result of resume x on the verification dimension η, D η is the importance weight of the verification dimension η, RankScore(η) is the authority rating of the verification dimension, and finally, the verification result is output, expressed as: VR i is the field verification result set corresponding to the i-th resume, express Is the set of fields to be verified in the i-th resume F i A field of Represented as fields Map a verification result label, ι is the trust threshold of the trust score, ι l is the suspicious lower limit threshold of the trust score, otherwise A x In ι l ≤A x <ι indicates uncertainty and requires manual review. F is a mapping function for verifying resume fields. RS iis the set of structured fields extracted from the i-th resume, → represents a "mapping relationship" A x is the overall trust score of candidate resume x.

[0087] Furthermore, in step S3, the trained BERT semantic model is used with the fraudulent resume corpus to compare the resume fields and output an abnormality risk score as follows:

[0088] Input the resume field into the trained BERT semantic model to extract the sentence vector G t , thereby constructing a semantic vector matrix, using the annotated fraudulent resume corpus to generate an abnormal semantic vector set, and using the Gaussian mixture model (GMM) to model the distribution of abnormal semantic vectors to obtain the probability density function, expressed as: in It is a sample The overall probability density value of is the sum of all j″th Gaussian distributions, J is the number of Gaussian distributions, π j″ is the mixing coefficient of the j″th Gaussian distribution, is the probability density of the j″th Gaussian distribution at the point, κ j″ is the mean, ∑ j″ is the covariance, and the language ambiguity detection score based on the sentence vector is: The semantic similarity comparison score based on the sentence vector is: The probability density function is used to calculate the sentence vector and perform log-likelihood scoring: An exaggeration detector is constructed based on a set of common fictitious keywords, and the sentence vector is scored for exaggeration detection: The comparison of the resume fields uses an adjustable weight combination to fuse the scores of each dimension, expressed as follows: Among them L t′ is the total abnormal risk score of the t′th resume, is the language ambiguity detection score, is the semantic similarity comparison score, is the log-likelihood score, is the exaggerated detection score, λ1,λ2,λ3,λ4 are the weighted coefficients of the abnormality score of each dimension, The sum of all weight coefficients must be 1 for normalization, and the output anomaly risk score is: Among them L i It is L i The final trust risk score of the resume, is the risk score L of this resume in the t′th round in history t′ Accumulate, L t′is the abnormal risk score obtained by the t′th detection, μ is the time impact weight factor, is the current scoring time t′ relative to the maximum number of rounds The standardized time weight of Z is the total number of rounds of cumulative scoring, Indicates the largest time point number among all time periods;

[0089] Furthermore, in step S4, a multi-factor scoring algorithm is used to perform a comprehensive calculation based on the verification result and the abnormal risk score to obtain a credibility score, and the credibility score is fed back to the semantic detection module as follows:

[0090] The multi-factor scoring algorithm sets the trust score of each field to Expression formula: in It is The trust score of the k′th field in the resume, It is The external verification credibility score of the k′th field in the resume, It is The abnormal risk score of the k′th field in the resume, is the "reversal score" of abnormal risk, are weighted coefficients for the credibility of external verification results and the anti-fraud tendency score of semantic detection, respectively. Dynamic field weights are introduced to perform weighted adjustment on the trust score of each field. Field weights are obtained through expert experience, historical data, and adaptive learning. The trust scores of all fields are combined using a weighted average method to obtain the overall credibility score of the resume field. The expression formula is: Among them B i″′ is the comprehensive trust score of the i′′th resume, k′ is the field number representing the k′th structured field, N″ is the total number of structured fields contained in the current resume, is the importance weight coefficient of the k′th field, It is The trust score of the k′th field in the resume is calculated, and adk′ is the adjustment factor. The credibility score calculated each time is fed back to the semantic detection module, and the BERT semantic model is updated with real recruitment data. The goal of the BERT semantic model update is to maximize the accuracy of the post-recruitment verification data, which is defined as a loss function based on real recruitment data, expressed as: where ξ fdbk is the loss function, Indicates the summation of the resume data from the 1st to the N′′′′′′′′′′′′′′, where N′′′′ is the total number of resumes used for calibration. i″ TrueStatus i″)} is the sum of the logarithmic values ​​of the conditional probability, H(B i″ TrueStatus i″ ) is the conditional probability, B i″ is the trust score of the i″th resume, TrueStatus i″ is the true and credible state label of the i″th resume after actual verification after recruitment, and is obtained by optimizing the loss function ξ fdbk , and gradually adjust the model parameters.

[0091] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. 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 make equivalent replacements for some of the technical features therein. However, these modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A recruitment registration information query system based on big data, characterized by: include: The data acquisition module extracts resume data based on the application resume document and outputs structured resume fields; A data verification module, linked to an external database, verifies the authenticity of the resume fields and outputs the verification results; The semantic detection module uses the trained BERT semantic model and the fraudulent resume corpus to compare the resume fields and output an anomaly risk score; A trust scoring module uses a multi-factor scoring algorithm to perform a comprehensive calculation based on the verification results and the abnormal risk score to obtain a credibility score, and feeds it back to the semantic detection module; The original text is obtained through the OCR parser and cleaned. The bidirectional language model (BiLSTM-CRF) is used to segment the cleaned resume document to obtain a word sequence. The word vector matrix is ​​then obtained through the word embedding function. Next, a multi-head attention mechanism is used to weight the context information to obtain a context-enhanced resume document. A named entity recognition model is constructed and the context-enhanced resume document is input into the BiLSTM-CRF layer to predict the corresponding label sequence. The expression formula is: where Q i is the optimal label sequence finally selected, It means to traverse all possible tag sequences and select the one with the highest score. is the sum of the time step t of the entire sequence, which has a total of W words and fields, R i t is the feature vector at the t-th position belonging to the i-th sample, E tg It is used to convert vector The weight matrix mapped to the score of each label, Trans(T t-1 ,T t ) represents the transfer score between labels, T t-1 ,T t is the label at the t-1th, tth position in the sequence, y is the entire search space, the corresponding entity field is extracted from the label sequence, and the structured resume field is output. The expression formula is: where Y i is the set of structured fields extracted from the i-th resume, α i is the set of all "education experience" fields extracted from the i-th resume, I j is the school description in the jth education information, O i The time period of the jth educational information, δ is the complete set of all academic qualifications entities in the educational experience, ε α is the complete set of all time periods in the educational experience, β i is the set of all "work experience" fields extracted from the i-th resume, P k is the job description in the kth job information, O k The time period of the kth job information, φ is the complete set of all companies in the work history, ε β is the complete set of all time periods in the work experience, χ i is the set of all "project experience" fields extracted from the i-th resume, U m is the mth project experience entry, It is a complete collection of all project descriptions.

2. A recruitment registration information query system based on big data according to claim 1, characterized in that: The workflow of the data validation module is: Receive the resume field, map the content of the resume field into a structured request vector, construct a heterogeneous query matrix, and output a query vector. The query vector initiates a request to the external database through the API, sets the entity set to be returned from the external database, and performs matching calculations and time intersection matching calculations for each resume field entity. At the same time, all matching degrees of each resume field entity are normalized and fused into a confidence score, expressed as follows: Among them D η =log(1+RankScore(η)), where A x is the overall trust score of candidate resume x, S x is the set of all verification dimensions associated with resume x, η∈S x is a dimension item in the verification dimension, γ(x,η) is the score result of resume x on the verification dimension η, D η is the importance weight of the verification dimension η, RankScore(η) is the authority rating of the verification dimension, and finally, the verification result is output, expressed as: VR i is the field verification result set corresponding to the i-th resume, express Is the set of fields to be verified in the i-th resume F i A field of Represented as fields Map a verification result label, ι is the trust threshold of the trust score, ι l is the suspicious lower limit threshold of the trust score, otherwise A x In ι l ≤A x <ι indicates uncertainty and requires manual review. F is a mapping function for verifying resume fields. RS i is the set of structured fields extracted from the i-th resume, → represents a "mapping relationship" A x is the overall trust score of candidate resume x.

3. A recruitment registration information query system based on big data according to claim 2, characterized in that: The workflow of the semantic detection module is: Input the resume field into the trained BERT semantic model to extract the sentence vector G t , thereby constructing a semantic vector matrix, using the annotated fraudulent resume corpus to generate an abnormal semantic vector set, and using the Gaussian mixture model (GMM) to model the distribution of abnormal semantic vectors to obtain the probability density function, expressed as: in It is a sample The overall probability density value of is the sum of all j″th Gaussian distributions, J is the number of Gaussian distributions, π j″ is the mixing coefficient of the j″th Gaussian distribution, is the probability density of the j″th Gaussian distribution at the point, κ j″ is the mean, ∑ j″ is the covariance, and the language ambiguity detection score based on the sentence vector is: The semantic similarity comparison score based on the sentence vector is: The probability density function is used to calculate the sentence vector and perform log-likelihood scoring: An exaggeration detector is constructed based on a set of common fictitious keywords, and the sentence vector is scored for exaggeration detection:

4. A recruitment registration information query system based on big data according to claim 3, characterized in that: The workflow of the semantic detection module is: The comparison of the resume fields uses an adjustable weight combination to fuse the scores of each dimension, expressed as follows: Among them L t′ is the total abnormal risk score of the t′th resume, is the language ambiguity detection score, is the semantic similarity comparison score, is the log-likelihood score, is the exaggerated detection score, λ1,λ2,λ3,λ4 are the weighted coefficients of the abnormality score of each dimension, The sum of all weight coefficients must be 1 for normalization, and the output anomaly risk score is: Among them L i It is L i The final trust risk score of the resume, is the risk score L of this resume in the t′th round in history t′ Accumulate, L t′ is the abnormal risk score obtained by the t′th detection, μ is the time impact weight factor, is the current scoring time t′ relative to the maximum number of rounds The standardized time weight of Z is the total number of rounds of cumulative scoring, Indicates the largest time point number among all time periods.

5. The big data-based recruitment registration information query system according to claim 4 is characterized in that: The workflow of the trust scoring module is: The multi-factor scoring algorithm sets the trust score of each field to Expression formula: in It is The trust score of the k′th field in the resume, It is The external verification credibility score of the k′th field in the resume, It is The abnormal risk score of the k′th field in the resume, is the "reversal score" of abnormal risk, are weighted coefficients for the credibility of external verification results and the anti-fraud tendency score of semantic detection, respectively. Dynamic field weights are introduced to perform weighted adjustment on the trust score of each field. Field weights are obtained through expert experience, historical data, and adaptive learning. The trust scores of all fields are combined using a weighted average method to obtain the overall credibility score of the resume field. The expression formula is: Among them B i″′ is the comprehensive trust score of the i′′th resume, k′ is the field number representing the k′th structured field, N″ is the total number of structured fields contained in the current resume, is the importance weight coefficient of the k′th field, It is is the trust score of the k′th field in a resume, and adk′ is the adjustment factor.

6. A recruitment registration information query system based on big data according to claim 5, characterized in that: The workflow of the trust scoring module is: The credibility score of each calculation is fed back to the semantic detection module, and the BERT semantic model is updated with real recruitment data. The goal of the BERT semantic model update is to maximize the accuracy of the post-recruitment verification data, which is defined as a loss function based on real recruitment data, expressed as: where ξ fdbk is the loss function, Indicates the summation of the resume data from the 1st to the N′′′′′′′′′′′′′′, where N′′′′ is the total number of resumes used for calibration. i″ TrueStatus i″ )} is the sum of the logarithmic values ​​of the conditional probability, H(B i″ TrueStatus i″ ) is the conditional probability, B i″ is the trust score of the i″th resume, TrueStatus i″ is the true and credible state label of the i″th resume after actual verification after recruitment, and is obtained by optimizing the loss function ξ fdbk , and gradually adjust the model parameters.

7. A recruitment applicant information query method based on big data, based on a recruitment applicant information query system based on big data according to any one of claims 1 to 6, characterized in that: The following steps are involved: S1. Extract resume data based on the application resume document and output structured resume fields; S2. Connect to an external database, verify the authenticity of the resume fields, and output the verification result; S3. Use the trained BERT semantic model and the fraudulent resume corpus to compare the resume fields and output an abnormality risk score; S4. Using a multi-factor scoring algorithm, a comprehensive calculation is performed based on the verification results and the abnormal risk score to obtain a credibility score, and the credibility score is fed back to the semantic detection module; The original text is obtained through the OCR parser and cleaned. The bidirectional language model (BiLSTM-CRF) is used to segment the cleaned resume document to obtain a word sequence. The word vector matrix is ​​then obtained through the word embedding function. Next, a multi-head attention mechanism is used to weight the context information to obtain a context-enhanced resume document. A named entity recognition model is constructed and the context-enhanced resume document is input into the BiLSTM-CRF layer to predict the corresponding label sequence. The expression formula is: where Q i is the optimal label sequence finally selected, It means to traverse all possible tag sequences and select the one with the highest score. The sum of the time steps t of the entire sequence has a total of W words and fields, is the feature vector at the t-th position belonging to the i-th sample, E tg It is used to convert vector The weight matrix mapped to the score of each label, Trans(T t-1 ,T t ) represents the transfer score between labels, T t-1 ,T t is the label at the t-1th, tth position in the sequence, y is the entire search space, the corresponding entity field is extracted from the label sequence, and the structured resume field is output. The expression formula is: where Y i is the set of structured fields extracted from the i-th resume, α i is the set of all "education experience" fields extracted from the i-th resume, I j is the school description in the jth education information, O i The time period of the jth educational information, δ is the complete set of all academic qualifications entities in the educational experience, ε α is the complete set of all time periods in the educational experience, β i is the set of all "work experience" fields extracted from the i-th resume, P k is the job description in the kth job information, O k The time period of the kth job information, φ is the complete set of all companies in the work history, ε β is the complete set of all time periods in the work experience, χ i is the set of all "project experience" fields extracted from the i-th resume, U m is the mth project experience entry, It is a complete collection of all project descriptions.

8. The method for querying recruitment registration information based on big data according to claim 7 is characterized in that: In step S2, the external database is associated with the resume fields to verify the authenticity and output the verification result in the following manner: Receive the resume field, map the content of the resume field into a structured request vector, construct a heterogeneous query matrix, and output a query vector. The query vector initiates a request to the external database through the API, sets the entity set to be returned from the external database, and performs matching calculations and time intersection matching calculations for each resume field entity. At the same time, all matching degrees of each resume field entity are normalized and fused into a confidence score, expressed as follows: Among them D η =log(1+RankScore(η)), where A x is the overall trust score of candidate resume x, S x is the set of all verification dimensions associated with resume x, η∈S x is a dimension item in the verification dimension, γ(x,η) is the score result of resume x on the verification dimension η, D η is the importance weight of the verification dimension η, RankScore(η) is the authority rating of the verification dimension, and finally, the verification result is output, expressed as: VR i is the field verification result set corresponding to the i-th resume, express Is the set of fields to be verified in the i-th resume F i A field of Represented as fields Map a verification result label, ι is the trust threshold of the trust score, ι l is the suspicious lower limit threshold of the trust score, otherwise A x In ι l ≤A x <ι indicates uncertainty and requires manual review. F is a mapping function for verifying resume fields. RS i is the set of structured fields extracted from the i-th resume, → represents a "mapping relationship" A x is the overall trust score of candidate resume x.

9. A recruitment applicant information query method based on big data according to claim 8, characterized in that: In step S3, the trained BERT semantic model is used with the fraudulent resume corpus to compare the resume fields and output an abnormality risk score as follows: Input the resume field into the trained BERT semantic model to extract the sentence vector G t , thereby constructing a semantic vector matrix, using the annotated fraudulent resume corpus to generate an abnormal semantic vector set, and using the Gaussian mixture model (GMM) to model the distribution of abnormal semantic vectors to obtain the probability density function, expressed as: in It is a sample The overall probability density value of is the sum of all j″th Gaussian distributions, J is the number of Gaussian distributions, π j″ is the mixing coefficient of the j″th Gaussian distribution, is the probability density of the j″th Gaussian distribution at the point, κ j″ is the mean, ∑ j″ is the covariance, and the language ambiguity detection score based on the sentence vector is: The semantic similarity comparison score based on the sentence vector is: The probability density function is used to calculate the sentence vector and perform log-likelihood scoring: An exaggeration detector is constructed based on a set of common fictitious keywords, and the sentence vector is scored for exaggeration detection: The comparison of the resume fields uses an adjustable weight combination to fuse the scores of each dimension, expressed as follows: Among them L t′ is the total abnormal risk score of the t′th resume, is the language ambiguity detection score, is the semantic similarity comparison score, is the log-likelihood score, is the exaggerated detection score, λ1,λ2,λ3,λ4 are the weighted coefficients of the abnormality score of each dimension, The sum of all weight coefficients must be 1 for normalization, and the output anomaly risk score is: Among them L i It is L i The final trust risk score of the resume, is the risk score L of this resume in the t′th round in history t′ Accumulate, L t′ is the abnormal risk score obtained by the t′th detection, μ is the time impact weight factor, is the current scoring time t′ relative to the maximum number of rounds The standardized time weight of Z is the total number of rounds of cumulative scoring, Indicates the largest time point number among all time periods; In step S4, a multi-factor scoring algorithm is used to perform a comprehensive calculation based on the verification results and the abnormal risk score to obtain a credibility score, and the credibility score is fed back to the semantic detection module as follows: The multi-factor scoring algorithm sets the trust score of each field to Expression formula: in It is The trust score of the k′th field in the resume, It is The external verification credibility score of the k′th field in the resume, It is The abnormal risk score of the k′th field in the resume, is the "reversal score" of abnormal risk, are weighted coefficients for the credibility of external verification results and the anti-fraud tendency score of semantic detection, respectively. Dynamic field weights are introduced to perform weighted adjustment on the trust score of each field. Field weights are obtained through expert experience, historical data, and adaptive learning. The trust scores of all fields are combined using a weighted average method to obtain the overall credibility score of the resume field. The expression formula is: Among them B i″′ is the comprehensive trust score of the i′′th resume, k′ is the field number representing the k′th structured field, N″ is the total number of structured fields contained in the current resume, is the importance weight coefficient of the k′th field, It is The trust score of the k′th field in the resume is calculated, and adk′ is the adjustment factor. The credibility score calculated each time is fed back to the semantic detection module, and the BERT semantic model is updated with real recruitment data. The goal of the BERT semantic model update is to maximize the accuracy of the post-recruitment verification data, which is defined as a loss function based on real recruitment data, expressed as: where ξ fdbk is the loss function, Indicates the summation of the resume data from the 1st to the N′′′′′′′′′′′′′′, where N′′′′ is the total number of resumes used for calibration. i″ TrueStatus i″ )} is the sum of the logarithmic values ​​of the conditional probability, H(B i″ TrueStatus i″ ) is the conditional probability, B i″ is the trust score of the i″th resume, TrueStatus i″ is the true and credible state label of the i″th resume after actual verification after recruitment, and is obtained by optimizing the loss function ξ fdbk , and gradually adjust the model parameters.

Images