Method and system for intelligently identifying resume counterfeiting through human resources

Abstract

The invention relates to the technical field of human resources, in particular to a method and system for intelligently identifying resume counterfeiting through human resources. The method comprises the following steps: acquiring a modified resume submitted to a platform by an applicant every time; recording the resume, the modification degree and the modification time of the applicant after each modification by adopting a block chain, and encrypting the resume, the modification degree and the modification time; verifying the identity of the recruiter, obtaining a decryption key after verification succeeds, and extracting a historical resume stored in the block chain; scoring the modification intention according to the modification degree, the overall modification frequency and the key area modification frequency; extracting a first facial expression and a first interview audio when the applicant answers the related resume questions during interview, and a second facial expression and a second interview audio when the applicant does not answer the related resume questions, and adopting an interview scoring model for scoring; and in combination with the modification intention scoring result and the interview scoring result, outputting a resume counterfeiting identification result. According to the method, the resume modification intention and the real interview are comprehensively considered, and the accuracy of resume counterfeiting identification is improved.

Description

Technical Field

[0001] This invention relates to the field of human resources technology, specifically to a method and system for intelligent identification of resume fraud in human resources. Background Technology

[0002] A resume is an essential tool for job seekers to showcase their work experience, educational background, and skills to employers. Employers rely on resumes to screen and recruit employees. However, over time, resume fraud has increased, making it difficult for employers to distinguish between genuine and false information. Resume fraud not only wastes employers' time and resources but can also lead to the hiring of unqualified or dishonest employees, posing potential risks to the business.

[0003] Existing methods typically rely on manual resume screening and verification, a time-consuming and labor-intensive task prone to errors and omissions. With technological advancements, electronic tools and automated systems have been introduced for resume screening. For example, Chinese patent application CN201910022438.2 discloses a method and system for identifying resume modification intent based on blockchain storage. The method includes: acquiring user resume-related data stored in blockchain format; determining the user's resume modification history information stored in the blockchain based on the user resume-related data; inputting the resume modification history information into a user intent judgment model for calculation; and identifying the user's intent to modify the resume based on the result of the user intent judgment model's calculation. However, this method only identifies the intent to modify the resume and does not analyze the actual interview scenario, thus remaining limited in detecting false information in resumes. Furthermore, Chinese patent application number CN201810596236.4 discloses a method for quickly determining the authenticity of resume content. This method involves listing common suspicious points in resumes to form an analysis point list for resume authenticity analysis; setting resume authenticity analysis rules for each suspicious point; then, for any resume, extracting information from each dimension listed in the analysis point list and comparing it with the resume authenticity analysis rules to conduct authenticity analysis; when suspicious points are found, they are recorded and stored in a resume authenticity analysis data table to form a resume suspicious point list; finally, the resume suspicious point list is displayed, along with a detailed description of the reasons for the suspicious points, for recruiters' reference. However, this method only relies on text to determine whether a resume is falsified, and the logical verification rules in the authenticity analysis rules only judge based on time points, without considering the actual interview scenario, which can easily lead to resume identification errors.

[0004] To improve the accuracy of resume fraud detection, a method and system for intelligent identification of resume fraud by human resources is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for intelligent identification of resume fraud by human resources. The method uses blockchain to record and encrypt each revised resume, the extent of revision, and the revision time. The intention to revise is scored based on the extent of revision, the overall frequency of revisions, and the frequency of revisions in key areas. At the same time, an interview scoring model is used to score the interview of the applicant. The final result of resume fraud identification is obtained by combining the revision intention score and the interview score.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for intelligent identification of resume fraud by human resources includes:

[0008] The system uses blockchain to record and encrypt each revision of an applicant's resume, including the extent of the revisions and the time of revision.

[0009] Furthermore, the framework of the resume is fixed, including a basic information area, a career goal area, an education background area, a work experience area, a skills and language ability area, a project experience area, an awards and punishments area, an interests and hobbies area, and a self-evaluation area; among which, the key areas of the resume include the education background area, the work experience area, the skills and language ability area, the project experience area, and the awards and punishments area.

[0010] Furthermore, the step of recording and encrypting each revised resume and revision time by the applicant using blockchain includes:

[0011] The initial resume and time of the applicant's first submission are stored in the block node;

[0012] When the applicant submits their resume again, the resubmitted resume will be compared with the previous resume to determine the extent of the modifications.

[0013] The previous resume is the one in the blockchain that is closest in time to the time of the resubmission;

[0014] The degree of modification includes an overall modification threshold and a key area modification threshold;

[0015] Determine whether the resume has been modified based on the overall modification threshold;

[0016] If modifications are made, a new block node is generated, and the resubmitted resume, the extent of modification, and the modification time are stored in the new block node; otherwise, the resubmitted resume and time are not stored.

[0017] Furthermore, the calculation process for the overall modification threshold includes:

[0018] Compare the resubmitted resume with the previous resume;

[0019] The resume comparison is performed using the Doc2Vec model, and the specific steps include:

[0020] Input the resubmitted resume and the previous resume into the comparison Doc2Vec model;

[0021] The comparison Doc2Vec model outputs the overall text vectors of the two resumes as the first overall text vector and the second overall text vector, respectively.

[0022] The first overall text vector and the second overall text vector are compared using an overall similarity comparison.

[0023] The expression for the overall similarity Z is:

[0024]

[0025] Where N represents the number of elements in the two overall text vectors; X i This represents the i-th element in the first overall text vector; The first element represents the i-th element in the second overall text vector; C represents the number of elements that are the same at corresponding positions in the first and second overall text vectors; the similarity Z ranges from 0 to 1; when Z = 1, it means that the resume has not been modified; when Z < 1, it means that the resume has been modified; the overall comparison similarity Z is used as the overall modification threshold.

[0026] Extract the key areas of the resubmitted resume and the key areas of the previous resume, and denote them as the first resume key area and the second resume key area, respectively.

[0027] The comparison Doc2Vec model outputs text vectors for the first key area and the second key area of ​​the resume, respectively.

[0028] The text vectors of the first key region and the text vectors of the second key region are compared using key region similarity.

[0029] The expression for the key region similarity z is:

[0030]

[0031] Where n represents the number of elements in the two key region text vectors; x i This represents the i-th element in the text vector of the first key region; The i-th element in the second key region text vector is represented; c represents the number of elements that are the same at corresponding positions in the first and second key region text vectors; the key region similarity z ranges from 0 to 1; when z = 1, it means that the key region has not been modified; when z < 1, it means that the key region has been modified; the key region similarity z is used as the key region modification threshold.

[0032] Verify the recruiter's identity, and retrieve historical resume information stored in the blockchain after successful verification.

[0033] Furthermore, the identity verification process for the recruiter includes:

[0034] The recruiter's facial image has been stored in the recruiter image set;

[0035] For each face image in the recruiter image set, DenseNet is used to extract the recruiter's face feature vector to construct a recruiter face feature vector set;

[0036] The recruiter is authenticated using DenseNet, and the specific steps include:

[0037] Collect facial images of recruiters to be identified;

[0038] The feature vector of the face to be identified is extracted from the face image to be identified using DenseNet;

[0039] The similarity of the facial feature vector to be identified is compared with that of each recruiter's facial feature vector in the set of recruiter facial feature vectors.

[0040] If the similarity score is greater than the set threshold, the authentication is successful and the decryption key is obtained; otherwise, the authentication fails.

[0041] The recruiter uses the decryption key to retrieve historical resumes stored in the blockchain.

[0042] The intention to modify is scored based on the degree of modification, the overall frequency of modifications, and the frequency of modifications in key areas.

[0043] Furthermore, all historical resumes stored in the blockchain are extracted, and the number of historical resumes is taken as the overall modification frequency; the total number of resumes in the blockchain whose modification threshold for all key areas is not 1 is taken as the key area modification frequency.

[0044] The first facial expression and first interview audio of the applicant answering questions related to their resume during the interview, as well as the second facial expression and second interview audio when not answering resume-related questions, were extracted and scored using an interview scoring model.

[0045] Furthermore, the interview scoring model includes an interview emotion scoring module and an interview content scoring module; the interview emotion scoring module includes an input layer, a facial expression feature extraction layer, an interview audio feature extraction layer, a feature fusion module, and an emotion scoring layer;

[0046] The facial expression features of the two facial expressions were extracted using the NFNet expression model.

[0047] Audio features of the two interview audios were extracted using the Mel spectral feature extractor;

[0048] The feature fusion layer is used to perform cross-modal fusion of the corresponding facial expression features and audio features, and output the first interview feature and the second interview feature;

[0049] The emotion scoring layer uses the VAD model to compare the two interview features to obtain an interview emotion score;

[0050] The interview content will be scored based on the interview audio.

[0051] The interview score is obtained based on the interview emotion rating and the interview content rating.

[0052] Furthermore, the facial expression NFNet model includes an expression feature extraction layer and an expression feature pooling layer; the expression feature extraction layer contains multiple feature extraction modules, a spatial attention mechanism is added between each feature extraction module, and residual connections are used between different feature extraction modules; each feature extraction module outputs a facial expression feature map; the feature pooling layer is used to fuse the expression feature maps and output facial expression features.

[0053] Furthermore, the interview content scoring module includes:

[0054] The interview audio will be used to output the interview text using an automatic language recognition model.

[0055] The Doc2Vec model was used to extract feature vectors of the interview text and the preset answer text from the interview text.

[0056] The similarity D of the interview content is obtained by comparing the feature vector of the interview text and the feature vector of the preset answer text.

[0057] The formula for calculating the similarity of the interview content is as follows:

[0058]

[0059] Where 'a' represents the number of elements in the interview text feature vector and the preset answer text feature vector; 'd' represents the number of elements in the interview text feature vector and the preset answer text feature vector. μThis represents the μ-th element in the feature vector of the interview text; This represents the μ-th element in the preset answer text feature vector; This indicates the number of identical elements at corresponding positions between the interview text feature vector and the preset answer text feature vector; the similarity D of the interview content comparison ranges from 0 to 1.

[0060] The interview content score is output based on the similarity comparison of the interview content.

[0061] Based on the modification intent scoring results and the interview scoring results, the resume fraud detection results are output.

[0062] Furthermore, resume fraud is identified based on the probability of resume fraud, and the formula for calculating the probability of resume fraud is as follows:

[0063] P = P1 * F J +(1-P1)*F M ;

[0064] Among them, F J Indicates the intention to modify the score; F M P1 represents the interview score; P1 represents the weight of the score for the intention to modify.

[0065] The F M The calculation formula is:

[0066] F M =p1F q +(1-p1)F w ;

[0067] Among them, F q Indicates the interview's emotional state rating; F w p1 represents the interview audio score; p1 represents the weight of the interview emotion score.

[0068] A human resources intelligent system for identifying resume fraud includes:

[0069] Resume Acquisition Unit: Used to acquire resumes submitted by recruiters;

[0070] Resume storage unit: Used to store and encrypt modified resumes, the extent of modification, and the modification time;

[0071] Authentication Unit: Used to authenticate the recruiter; upon successful authentication, a decryption key is obtained.

[0072] Resume extraction unit: used by recruiters to extract historical resume information stored in the blockchain using the decryption key;

[0073] Interview recording unit: used to record the applicant's facial expressions and interview audio during the interview;

[0074] Scoring Unit: Used for scoring resume revision intent and interview performance;

[0075] Output unit: Used to output resume fraud identification results based on the modification intent scoring results and the interview scoring results.

[0076] Furthermore, the scoring unit includes a modification intent scoring module and an interview scoring module: the modification intent scoring module is used to score based on the degree of modification; the interview scoring module is used to score based on the facial expressions and the interview audio; the degree of modification includes an overall modification threshold and a key area modification threshold; the calculation of the overall modification threshold and the key area modification threshold includes:

[0077] Compare the resubmitted resume with the previous resume;

[0078] The previous resume is the one in the blockchain that is closest in time to the time of the resubmission;

[0079] The Doc2Vec model is used for resume comparison. The specific steps include:

[0080] Input the resubmitted resume and the previous resume into the comparison Doc2Vec model;

[0081] The comparison Doc2Vec model outputs the overall text vectors of the two resumes as the first overall text vector and the second overall text vector, respectively.

[0082] The first overall text vector and the second overall text vector are compared using an overall similarity comparison.

[0083] The expression for the overall similarity Z is:

[0084]

[0085] Where N represents the number of elements in the two overall text vectors; X i This represents the i-th element in the first overall text vector; The first element represents the i-th element in the second overall text vector; C represents the number of elements that are the same at corresponding positions in the first and second overall text vectors; the similarity Z ranges from 0 to 1; the overall comparison similarity Z is used as the overall modification threshold.

[0086] Extract the key areas of the resubmitted resume and the key areas of the previous resume, and denote them as the first resume key area and the second resume key area, respectively.

[0087] The comparison Doc2Vec model outputs text vectors for the first key area of ​​the resume and the second key area of ​​the resume, respectively, which are the text vectors for the first key area and the text vectors for the second key area.

[0088] The text vectors of the first key region and the text vectors of the second key region are compared using key region similarity.

[0089] The expression for the key region similarity z is:

[0090]

[0091] Where n represents the number of elements in the two key region text vectors; x i This represents the i-th element in the text vector of the first key region; The i-th element in the second key region text vector is represented; c represents the number of elements that are the same at corresponding positions in the first and second key region text vectors; the key region similarity z ranges from 0 to 1; the key region similarity z is used as the key region modification threshold.

[0092] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0093] 1. This invention uses blockchain to record each revision of the applicant's resume, the extent of the revisions, and the revision time when identifying resume fraud. Then, based on the resumes and revision details recorded in the blockchain, it obtains the overall revision frequency and the frequency of revisions to key areas of the resume. The revision intent is scored by combining the revision extent, overall revision frequency, and key area revision frequency. Furthermore, it compares the applicant's facial expressions and interview audio when answering resume-related questions with their facial expressions and interview audio when not answering resume-related questions. An interview scoring model is used to score the applicant's state when answering resume-related questions. The probability of resume fraud is output by combining the revision intent score and the interview score. This method comprehensively considers the resume revision intent and the actual interview, improving the accuracy of resume fraud detection.

[0094] 2. When calculating the degree of resume modification, this invention uses a Doc2Vec model to extract the overall text vector and key region text vectors of the resubmitted resume and the previous resume, respectively. Overall comparison similarity and key region comparison similarity are used to measure the overall modification threshold and the key region modification threshold. The overall comparison similarity and key region comparison similarity used in this invention not only focus on the overall similarity of the vectors but also check the matching of each element in the vectors, making the similarity calculation more comprehensive, capturing more information, and thus improving the accuracy of resume fraud detection.

[0095] 3. This invention adds a spatial attention mechanism between each feature extraction module in the facial expression feature extraction layer of NFNet, and constructs an facial expression NFNet model using residual connections between each feature extraction module. Simultaneously, the facial expression NFNet model is used to extract facial expression features of job applicants during interviews. The facial expression NFNet model can focus on different facial expression features at each level, and the use of residual connections allows the model to more easily capture and propagate gradients, thereby improving the accuracy of facial expression feature extraction, enhancing the accuracy of interview scoring, and ultimately improving the accuracy of resume fraud detection. Attached Figure Description

[0096] Figure 1 This is a flowchart illustrating a method for intelligent identification of resume fraud in human resources, provided by an embodiment of the present invention.

[0097] Figure 2 This is a schematic diagram of the Doc2Vec model structure provided in an embodiment of the present invention;

[0098] Figure 3 This is a schematic diagram illustrating the overall alignment success rate verification of the Doc2Vec model provided in an embodiment of the present invention.

[0099] Figure 4 This is a schematic diagram illustrating the verification of the success rate of the Doc2Vec model for key regions in an embodiment of the present invention.

[0100] Figure 5 This is a schematic diagram of the interview scoring model structure provided in an embodiment of the present invention;

[0101] Figure 6 This is a schematic diagram illustrating the effectiveness verification of the interview scoring model provided in an embodiment of the present invention;

[0102] Figure 7 This is a schematic diagram of a system process for intelligent identification of resume fraud in human resources, provided as an embodiment of the present invention. Detailed Implementation

[0103] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0104] Please see Figures 1 to 7 This invention provides a method and system for intelligent identification of resume fraud in human resources, the technical solution of which is as follows:

[0105] As one embodiment of the present invention, refer to Figure 1A method for intelligent identification of resume fraud by human resources includes: S10. Obtaining each modified resume submitted by the applicant to the platform; S20. Recording and encrypting each modified resume, the degree of modification, and the modification time using blockchain; S30. Verifying the recruiter's identity, and retrieving historical resumes stored in the blockchain after successful verification; S40. Scoring the modification intent based on the degree of modification, overall modification frequency, and modification frequency of key areas; S50. Extracting the applicant's first facial expression and first interview audio when answering resume-related questions during the interview, as well as the second facial expression and second interview audio when not answering resume-related questions, and scoring them using an interview scoring model; S60. Combining the modification intent scoring result and the interview scoring result, outputting the resume fraud identification result.

[0106] S10. Obtain the revised resumes submitted by applicants to the platform each time.

[0107] Furthermore, the system obtains each resume submitted by the applicant. The resume has a fixed framework, including a basic information area, a career goal area, an education background area, a work experience area, a skills and language ability area, a project experience area, an awards and punishments area, an interest and hobby area, and a self-evaluation area. The key areas of the resume include the education background area, the work experience area, the skills and language ability area, the project experience area, and the awards and punishments area.

[0108] Furthermore, in this embodiment, 1000 resumes with varying degrees of revision and editing submitted by applicants, along with their interview videos, were collected. All data collection was done with the applicants' consent, and 215 of these resumes were found to contain falsified information. The 600 employees were divided into a training set and a test set in a 7:3 ratio. The training set contained 2186 resumes from 420 employees with varying degrees of revision, including 420 initial resumes. The test set contained 842 resumes from 180 employees with varying degrees of revision, including 180 initial resumes.

[0109] S20. Use blockchain to record and encrypt each revision of the applicant's resume, the extent of the revision, and the revision time.

[0110] Furthermore, the step of recording and encrypting each revised resume and revision time by the applicant using blockchain includes:

[0111] The initial resume and time of the applicant's first submission are stored in the block node;

[0112] When the applicant submits their resume again, the resubmitted resume will be compared with the previous resume to determine the extent of the modifications.

[0113] The previous resume is the one in the blockchain that is closest in time to the time of the resubmission;

[0114] The degree of modification includes an overall modification threshold and a key area modification threshold;

[0115] Determine whether the resume has been modified based on the overall modification threshold;

[0116] If modifications are made, a new block node is generated, and the resubmitted resume, the extent of modification, and the modification time are stored in the new block node; otherwise, the resubmitted resume and time are not stored.

[0117] Furthermore, the calculation process for the overall modification threshold includes:

[0118] Compare the resubmitted resume with the previous resume;

[0119] Reference Figure 2 The resume comparison is performed using the Doc2Vec model, and the specific steps include:

[0120] Input the resubmitted resume and the previous resume into the comparison Doc2Vec model;

[0121] The comparison Doc2Vec model outputs the overall text vectors of the two resumes as the first overall text vector and the second overall text vector, respectively.

[0122] The first overall text vector and the second overall text vector are compared using an overall similarity comparison.

[0123] The expression for the overall similarity Z is:

[0124]

[0125] Where N represents the number of elements in the two overall text vectors; X i This represents the i-th element in the first overall text vector; Let Z represent the i-th element in the second overall text vector; C represents the number of elements that are the same at corresponding positions in the first overall text vector and the second overall text vector; the similarity Z ranges from 0 to 1; when Z = 1, it means that the resume has not been modified; when Z < 1, it means that the resume has been modified; Z is used as the overall modification threshold.

[0126] Extract the key areas of the resubmitted resume and the key areas of the previous resume, and denote them as the first resume key area and the second resume key area, respectively.

[0127] The comparison Doc2Vec model outputs text vectors for the first key area of ​​the resume and the second key area of ​​the resume, respectively, which are the text vectors for the first key area and the text vectors for the second key area.

[0128] The text vectors of the first key region and the text vectors of the second key region are compared using the key region volume similarity comparison.

[0129] The expression for the key region similarity z is:

[0130]

[0131] Where n represents the number of elements in the two key region text vectors; x i This represents the i-th element in the text vector of the first key region; The i-th element in the second key region text vector is represented; c represents the number of elements that are the same at corresponding positions in the first and second key region text vectors; the key region similarity z ranges from 0 to 1; when z = 1, it means that the key region has not been modified; when z < 1, it means that the key region has been modified; the key region similarity z is used as the key region modification threshold.

[0132] Furthermore, this embodiment uses resumes from the training set to train the Doc2Vec comparison model, and uses a test set to test the effectiveness of the Doc2Vec comparison model. In this embodiment, comparisons with an error within 0.05 are considered successful. Besides using the Doc2Vec comparison model, the Doc2Vec model and the Word2Vec model are also used to verify the effectiveness of the model provided in this embodiment. Specific results can be found in [reference needed]. Figure 3 and Figure 4 Model 1 is a comparison with the Doc2Vec model, Model 2 is the Doc2Vec model, and Model 3 is the Word2Vec model. Twenty experiments were conducted on each model. To more intuitively demonstrate the experimental results of each model, the results of the 20 experiments were averaged, as shown in Table 1.

[0133] Table 1. Validation Table of Doc2Vec Model Performance Comparison

[0134] Model Model 1 Model 2 Model 3 Overall comparison success rate 97.49％ 93.06％ 91.30％ Key area comparison success rate 97.60％ 92.81％ 91.43％

[0135] Furthermore, from Figure 3 , Figure 4 As can be seen from Table 1, the Doc2Vec matching model provided in this embodiment has a higher overall matching success rate and a higher success rate for matching key regions than Model 2 and Model 3. In other words, the Doc2Vec matching model provided in this embodiment can improve the text matching effect.

[0136] In this embodiment, when calculating the degree of resume modification, the Doc2Vec model is used to extract the overall text vector and key region text vector of the resubmitted resume and the resume closest to the resubmission time. Overall comparison similarity and key region comparison similarity are used to measure the overall modification threshold and the key region modification threshold. The overall comparison similarity and key region comparison similarity used in this invention not only focus on the overall similarity of the vectors but also check the matching of each element in the vectors, making the similarity calculation more comprehensive, capturing more information, and thus improving the accuracy of resume fraud detection.

[0137] S30. Verify the recruiter's identity. After successful verification, retrieve the historical resumes stored in the blockchain.

[0138] Furthermore, the identity verification process for the recruiter includes:

[0139] The recruiter's facial image has been stored in the recruiter image set;

[0140] For each face image in the recruiter image set, DenseNet is used to extract the recruiter's face feature vector to construct a recruiter face feature vector set;

[0141] The recruiter is authenticated using DenseNet, and the specific steps include:

[0142] Collect facial images of recruiters to be identified;

[0143] The feature vector of the face to be identified is extracted from the face image to be identified using DenseNet;

[0144] The similarity of the facial feature vector to be identified is compared with that of each recruiter's facial feature vector in the set of recruiter facial feature vectors.

[0145] If the similarity score is greater than the set threshold, the authentication is successful and the decryption key is obtained; otherwise, the authentication fails.

[0146] The recruiter uses the decryption key to retrieve historical resumes stored in the blockchain.

[0147] S40. The modification intent is scored based on the degree of modification, the overall frequency of modifications, and the frequency of modifications in key areas.

[0148] Furthermore, all historical resumes stored in the blockchain are extracted, and the number of historical resumes is taken as the overall modification frequency; the total number of resumes in the blockchain whose modification threshold for all key areas is not 1 is taken as the key area modification frequency.

[0149] S50. Extract the first facial expression and first interview audio of the applicant when answering questions related to the resume, and the second facial expression and second interview audio when not answering questions related to the resume, and score them using an interview scoring model.

[0150] Furthermore, referring to Figure 5 The interview scoring model includes an interview emotion scoring module and an interview content scoring module; the interview emotion scoring module includes a facial expression feature extraction layer, an interview audio feature extraction layer, a feature fusion layer, and an emotion scoring layer;

[0151] The facial expression features of the two facial expressions were extracted using the NFNet expression model.

[0152] Audio features of the two interview audios were extracted using the Mel spectral feature extractor;

[0153] The feature fusion layer is used to perform cross-modal fusion of the corresponding facial expression features and audio features, and output the first interview feature and the second interview feature;

[0154] The emotion scoring layer uses the VAD model to compare the two interview features to obtain an interview emotion score;

[0155] The interview content will be scored based on the first interview audio.

[0156] The interview score is obtained based on the rating results of the interview emotion and the rating results of the interview content.

[0157] Furthermore, the facial expression NFNet model includes an expression feature extraction layer and an expression feature pooling layer; the expression feature extraction layer contains multiple feature extraction modules, a spatial attention mechanism is added between each feature extraction module, and residual connections are used between different feature extraction modules; each feature extraction module outputs a facial expression feature map; the feature pooling layer is used to fuse the expression feature maps and output facial expression features.

[0158] Furthermore, the interview content scoring module includes:

[0159] The interview audio will be used to output the interview text using an automatic language recognition model.

[0160] The Doc2Vec model was used to extract feature vectors from the interview text and the preset answer text.

[0161] The similarity D of the interview content is obtained by comparing the feature vector of the interview text and the feature vector of the preset answer text.

[0162] The formula for calculating the similarity of the interview content is as follows:

[0163]

[0164] Where 'a' represents the number of elements in the interview text feature vector and the preset answer text feature vector; 'd' represents the number of elements in the interview text feature vector and the preset answer text feature vector. μ This represents the μ-th element in the feature vector of the interview text; This represents the μ-th element in the preset answer text feature vector; The number of identical elements at corresponding positions between the interview text feature vector and the preset answer text feature vector; the similarity D of the interview content comparison ranges from 0 to 1.

[0165] The interview content score is output based on the similarity of the interview content, and the interview score result is obtained by combining the interview content score and the interview emotion score.

[0166] Furthermore, the formula for calculating the interview content scoring result is: Interview Score F M The calculation formula is:

[0167] F M =p1F q +(1-p1)F w

[0168] Among them, F q Indicates the interview's emotional state rating; F w p1 represents the interview audio score; p1 represents the weight of the interview emotion score.

[0169] Furthermore, for each interview video, 10 senior interviewers scored the interview emotion and the interview itself. The average score of the 10 interviewers was used as the actual score to train the model. The interview videos in the training set were input into the interview scoring model for training, and the weight of the interview emotion score was found to be 0.42. The model effect was verified using a test set. When the interview score output by the interview scoring model differs from the actual interviewer score by less than 5 points, it indicates that the model output is accurate.

[0170] Furthermore, in addition to the interview scoring model used in this embodiment, three other models were also used for verification. For specific results, please refer to [link / reference needed]. Figure 6Model 1 is the interview scoring model provided in this embodiment; Model 2 is based on the interview scoring model by replacing the facial expression NFNet model with an NFNet model; Model 3 is based on the interview scoring model by replacing the comparison Doc2Vec model with a Doc2Vec model; Model 4 is based on the interview scoring model by replacing the facial expression NFNet model with an NFNet model and the comparison Doc2Vec model with a Doc2Vec model. Each model underwent 20 repeated experiments. To more intuitively demonstrate the experimental results of each model, the results of the 20 experiments were averaged, as shown in Table 2.

[0171] Table 2. Validation of the Interview Scoring Model

[0172] Model Model 1 Model 2 Model 3 Model 4 Interview scoring accuracy 97.67％ 90.11％ 89.89％ 87.69％

[0173] Furthermore, according to Figure 6 As shown in Table 2, the scoring accuracy of the interview scoring model proposed in this embodiment is significantly higher than that of Model 2, Model 3 and Model 4. That is, the interview scoring model provided in this embodiment can effectively improve the accuracy of interview scoring.

[0174] In this embodiment, a spatial attention mechanism is added between each feature extraction module of the NFNet facial expression feature extraction layer, and a residual connection is used to construct the facial expression NFNet model between each feature extraction module. Simultaneously, the facial expression NFNet model is used to extract the facial expression features of job applicants during interviews. The facial expression NFNet model can focus on different facial expression features at each level, and by using residual connections, the model can more easily capture and propagate gradients, thereby improving the accuracy of facial expression feature extraction, which in turn improves the accuracy of interview scoring and ultimately enhances the accuracy of resume fraud detection.

[0175] S60. Combining the modification intent scoring results and the interview scoring results, output the resume fraud detection results.

[0176] Furthermore, resume fraud is identified based on the probability of resume fraud, and the formula for calculating the probability of resume fraud is as follows:

[0177]

[0178] Among them, F J Indicates the intention to modify the score; F M P1 represents the interview score; P1 represents the weight of the score for the intention to modify.

[0179] Furthermore, in this embodiment, based on the modification intent scores and interview scores in the training set, the modification intent score weight and threshold K of the formula for the probability of resume fraud are trained. The result is that the modification intent score weight is 0.36 and K is 0.6. Using this formula to identify resumes, the accuracy rate of resume fraud detection is 98.33%.

[0180] As one embodiment of the present invention, refer to Figure 7 A system for intelligently identifying resume fraud in human resources includes: S210. Resume Acquisition Unit: for acquiring resumes submitted by recruiters; S220. Resume Storage Unit: for storing and encrypting modified resumes, the extent of modification, and the modification time; S230. Identity Verification Unit: for verifying the identity of recruiters and obtaining a decryption key upon successful verification; S240. Resume Extraction Unit: for recruiters to extract historical resumes and modification intent scores stored in the blockchain based on the decryption key; S250. Interview Recording Unit: for recording the facial expressions and audio of applicants during interviews; S260. Scoring Unit: for scoring resume modification intent and interview performance; S270. Output Unit: for outputting resume fraud identification results based on the modification intent score and interview performance.

[0181] Furthermore, the scoring unit includes a modification intent scoring module and an interview scoring module: the modification intent scoring module is used to score based on the degree of modification; the interview scoring module is used to score based on the facial expressions and the interview audio; the degree of modification includes an overall modification threshold and a key area modification threshold; the calculation of the overall modification threshold and the key area modification threshold includes:

[0182] Compare the resubmitted resume with the previous resume;

[0183] The previous resume is the one in the blockchain that is closest in time to the time of the resubmission;

[0184] The Doc2Vec model is used for resume comparison. The specific steps include:

[0185] Input the resubmitted resume and the previous resume into the comparison Doc2Vec model;

[0186] The comparison Doc2Vec model outputs the overall text vectors of the two resumes as the first overall text vector and the second overall text vector, respectively.

[0187] The first overall text vector and the second overall text vector are compared using an overall similarity comparison.

[0188] The expression for the overall similarity Z is:

[0189]

[0190] Where N represents the number of elements in the two overall text vectors; X i This represents the i-th element in the first overall text vector; The i-th element in the second overall text vector is represented; C represents the number of elements that are the same at corresponding positions in the first and second overall text vectors; the similarity Z ranges from 0 to 1; the overall comparison similarity Z is used as the overall modification threshold.

[0191] Extract the key areas of the resubmitted resume and the key areas of the previous resume, and denote them as the first resume key area and the second resume key area, respectively.

[0192] The comparison Doc2Vec model outputs text vectors for the first key area of ​​the resume and the second key area of ​​the resume, respectively, which are the text vectors for the first key area and the text vectors for the second key area.

[0193] The text vectors of the first key region and the text vectors of the second key region are compared using an overall similarity comparison.

[0194] The expression for the key region similarity z is:

[0195]

[0196] Where n represents the number of elements in the two key region text vectors; x i This represents the i-th element in the text vector of the first key region; The i-th element in the second key region text vector is represented; c represents the number of elements that are the same at corresponding positions in the first and second key region text vectors; the key region similarity z ranges from 0 to 1; the key region comparison similarity z is used as the key region modification threshold.

[0197] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0198] One or more embodiments of this specification are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments of this specification should be included within the scope of protection of this disclosure.

Claims

1. A method for intelligent identification of resume fraud by human resources, characterized in that, include: Blockchain is used to record and encrypt each revision of an applicant's resume, the extent of the revisions, and the revision time. Verify the recruiter's identity, and retrieve historical resume information stored in the blockchain after successful verification; The modification intent is scored based on the degree of modification, the overall frequency of modifications, and the frequency of modifications in key areas; The first facial expression and first interview audio of the applicant answering questions related to the resume during the interview, as well as the second facial expression and second interview audio when not answering questions related to the resume, were extracted and scored using an interview scoring model; The interview scoring model includes an interview emotion scoring module and an interview content scoring module; The interview emotion scoring module includes an input layer, a facial expression feature extraction layer, an interview audio feature extraction layer, a feature fusion module, and an emotion scoring layer; The facial expression features of the two facial expressions were extracted using the NFNet expression model. Audio features of the two interview audios were extracted using the Mel spectral feature extractor; The feature fusion layer is used to perform cross-modal fusion of the corresponding facial expression features and audio features, and output the first interview feature and the second interview feature; The emotion scoring layer uses the VAD model to compare the two interview features to obtain an interview emotion score; The interview content will be scored based on the first interview audio. Based on the rating results of the interview emotions and the rating results of the interview content, an interview rating result is obtained; Based on the modification intent scoring results and the interview scoring results, the resume fraud detection results are output.

2. The method for intelligent identification of resume fraud by human resources as described in claim 1, characterized in that, The resume includes a basic information area, a career goal area, an education background area, a work experience area, a skills and language ability area, a project experience area, an awards and punishments area, an interest and hobby area, and a self-evaluation area. The key areas include the education background area, work experience area, skills and language ability area, project experience area, and rewards and punishments area.

3. The method for intelligent identification of resume fraud by human resources as described in claim 1, characterized in that, The steps involved in recording and encrypting each revised resume and the revision time using blockchain include: The initial resume and time of the applicant's first submission are stored in the block node; When the applicant submits their resume again, the resubmitted resume will be compared with the previous resume to determine the extent of the modifications. The previous resume is the one in the blockchain that is closest in time to the time of the resubmission; The degree of modification includes an overall modification threshold and a key area modification threshold; Determine whether the resume has been modified based on the overall modification threshold; If modifications are made, a new block node is generated, and the resubmitted resume, the extent of modification, and the modification time are stored in the new block node; otherwise, the resubmitted resume and time are not stored.

4. The method for intelligent identification of resume fraud by human resources as described in claim 3, characterized in that, The degree of modification includes: The calculation process for the overall modification threshold includes: The resubmitted resume and the previous resume are compared using the Doc2Vec comparison model. The specific steps include: Input the resubmitted resume and the previous resume into the comparison Doc2Vec model; The comparison Doc2Vec model outputs the overall text vectors of the two resumes as the first overall text vector and the second overall text vector, respectively. The first overall text vector and the second overall text vector are compared using an overall similarity comparison. The expression for the overall similarity Z is: Where N represents the number of elements in the two overall text vectors; X i This represents the i-th element in the first overall text vector; Let Z represent the i-th element in the second overall text vector; C represents the number of elements that are the same at corresponding positions in the first and second overall text vectors; the similarity Z ranges from 0 to 1; when Z = 1, it means the resume has not been modified; when Z < 1, it means the resume has been modified; Z is used as the overall modification threshold. Extract the key areas of the resubmitted resume and the key areas of the previous resume, and denote them as the first resume key area and the second resume key area, respectively. The comparison Doc2Vec model outputs text vectors for the first key area of ​​the resume and the second key area of ​​the resume, respectively, which are the text vectors for the first key area and the text vectors for the second key area. The text vectors of the first key region and the text vectors of the second key region are compared using key region similarity. The expression for the key region similarity z is: Where n represents the number of elements in the two key region text vectors; x j This represents the j-th element in the text vector of the first key region; The value of z represents the j-th element in the text vector of the second key region; c represents the number of elements that are the same at corresponding positions in the text vectors of the first and second key regions; z ranges from 0 to 1; when z = 1, it means that the key region has not been modified; when z < 1, it means that the key region has been modified. The similarity z of the key region comparison is used as the key region modification threshold.

5. The method for intelligent identification of resume fraud by human resources as described in claim 1, characterized in that, The identity verification process for recruiters includes: The recruiter's facial image has been stored in the recruiter image set; For each face image in the recruiter image set, DenseNet is used to extract the recruiter's face feature vector to construct a recruiter face feature vector set; The recruiter is authenticated using DenseNet, and the specific steps include: Collect facial images of recruiters to be identified; The feature vector of the face to be identified is extracted from the face image to be identified using DenseNet; The similarity of the facial feature vector to be identified is compared with that of each recruiter's facial feature vector in the set of recruiter facial feature vectors. If the similarity score is greater than the set threshold, the authentication is successful and the decryption key is obtained; otherwise, the authentication fails. The recruiter uses the decryption key to extract historical resume information stored in the blockchain; The historical resume information includes all historical resumes stored in the blockchain and the degree of modification corresponding to each historical resume.

6. The method for intelligent identification of resume fraud by human resources as described in claim 1, characterized in that, The facial expression NFNet model includes: The facial expression NFNet model includes an facial expression feature extraction layer and a facial expression feature pooling layer; The facial expression feature extraction layer includes multiple feature extraction modules, a spatial attention mechanism is added between each feature extraction module, and residual connections are used between different feature extraction modules. Each of the feature extraction modules outputs a facial expression feature map; The feature pooling layer is used to fuse the facial expression feature map and output facial expression features.

7. The method for intelligent identification of resume fraud by human resources as described in claim 1, characterized in that, The interview content scoring module includes: The first interview audio is used to output the interview text using an automatic language recognition model; The Doc2Vec model was used to extract feature vectors from the interview text and the preset answer text. The similarity D of the interview content is obtained by comparing the feature vector of the interview text and the feature vector of the preset answer text. The formula for calculating the similarity of the interview content is as follows: Where 'a' represents the number of elements in the interview text feature vector and the preset answer text feature vector; 'd' represents the number of elements in the interview text feature vector and the preset answer text feature vector. μ This represents the μ-th element in the feature vector of the interview text; This represents the μ-th element in the preset answer text feature vector; This indicates the number of identical elements at corresponding positions between the interview text feature vector and the preset answer text feature vector; the similarity D of the interview content comparison ranges from 0 to 1. The interview content score is output based on the similarity comparison of the interview content.

8. The method for intelligent identification of resume fraud by human resources as described in claim 1, characterized in that, The results of the resume fraud detection include: The identification of resume fraud is based on the probability of resume fraud, and the formula for calculating the probability of resume fraud is as follows: Among them, F J Indicates the intention to modify the score; F M P1 represents the interview score; P1 represents the weight of the score for the intention to modify. When the probability of the resume being fake is greater than a set threshold K, it indicates that the resume is a fake resume. The F M The calculation formula is: F M =p1F q +(1-p1)F w ; Among them, F q Indicates the interview's emotional state rating; F w p1 represents the score for the interview content; p1 represents the weight of the interview emotion score.

9. A human resources intelligent system for identifying resume fraud, characterized in that, include: Resume Acquisition Unit: Used to acquire resumes submitted by recruiters; Resume storage unit: Used to store and encrypt modified resumes, the extent of modification, and the modification time; Authentication Unit: Used to verify the recruiter's identity; upon successful verification, a decryption key is obtained. Resume extraction unit: used by recruiters to extract historical resume information stored in the blockchain using the decryption key; Interview recording unit: used to record the applicant's facial expressions and interview audio during the interview; Scoring Unit: Used for scoring resume revision intent and interview performance; Output unit: Used to output resume fraud identification results based on the modification intent scoring results and the interview scoring results.

10. A system for intelligent identification of resume fraud in human resources according to claim 9, characterized in that, The scoring unit includes a modification intent scoring module and an interview scoring module: The modification intent scoring module is used to score based on the degree of modification; the interview scoring is used to score based on facial expressions and the interview audio. The degree of modification includes an overall modification threshold and a key area modification threshold; The calculation of the overall modification threshold and the key region modification threshold includes: Compare the resubmitted resume with the previous one; The previous resume is the one in the blockchain that is closest in time to the time of the resubmission; The Doc2Vec model is used for resume comparison. The specific steps include: Input the resubmitted resume and the previous resume into the comparison Doc2Vec model; The comparison Doc2Vec model outputs the overall text vectors of the two resumes as the first overall text vector and the second overall text vector, respectively. The first overall text vector and the second overall text vector are compared using an overall similarity comparison. The expression for the overall similarity Z is: Where N represents the number of elements in the two overall text vectors; X i This represents the i-th element in the first overall text vector; Let represent the i-th element in the second overall text vector; C represents the number of elements that are the same at corresponding positions in the first and second overall text vectors; the similarity Z ranges from 0 to 1. The overall similarity Z is used as the overall modification threshold. Extract the key regions of the resubmitted resume and the key regions of the previous resume, and denote them as the first resume key region and the second resume key region, respectively; the comparison Doc2Vec model outputs the text vectors of the first resume key region and the second resume key region as the first key region text vector and the second key region text vector, respectively; compare the first key region text vector and the second key region text vector using key region similarity comparison. The expression for the key region similarity z is: Where n represents the number of elements in the two key region text vectors; x i This represents the i-th element in the text vector of the first key region; represents the i-th element in the second key region text vector; c represents the number of elements that are the same at corresponding positions in the first key region text vector and the second key region text vector; the key region similarity z ranges from 0 to 1; The similarity z of the key region comparison is used as the key region modification threshold.

Images