Based on AI-powered recruitment interview screening and evaluation management methods and SaaS platform

Abstract

This invention discloses a recruitment interview AI-based screening, evaluation, and management method and SaaS platform. The method includes collecting candidate resume data and target job requirement data; using a multi-head attention neural network model to perform a first screening of resume features; collecting and analyzing candidates' browsing information on the recruitment platform to perform a second screening; conducting video interviews with candidates who pass the second screening, capturing their images using a camera, and outputting their interview emotion features; preprocessing the candidates' resume features, browsing information features, and interview emotion features after the interview to obtain a candidate application vector array; and inputting the candidate application vector array into an interview evaluation model to output the candidates' application results. This invention uses a neural network model to screen candidates multiple times, thereby accurately selecting applicants, improving the efficiency and accuracy of intelligent recruitment, and saving recruitment human resource costs.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence recruitment technology, and more specifically, to a management method and SaaS platform for AI-based recruitment interview screening and evaluation. Background Technology

[0002] With the rapid development of information technology and the widespread adoption of the internet, the recruitment process has gradually shifted from traditional human resource management to a digital and intelligent model. The rise of artificial intelligence (AI) technology has provided technical support and possibilities for the development of intelligent recruitment methods. The current recruitment market faces numerous challenges, and companies have an increasingly urgent need to improve recruitment efficiency, reduce costs, and optimize talent matching, creating market demand for intelligent recruitment methods.

[0003] Related technologies utilize natural language processing to parse resume information and match the parsing results with the corresponding job postings of target companies to achieve intelligent recruitment. However, these technologies can only perform direct matching and screening when matching resumes, making it difficult to obtain hidden related information from the resume information. At the same time, they do not incorporate multi-level artificial intelligence analysis of candidates' browsing information characteristics on recruitment platforms and their emotional characteristics during interviews, nor do they conduct AI analysis of recruitment data throughout the entire recruitment process. This may lead to inaccurate analysis during the recruitment process, thereby reducing the efficiency and accuracy of intelligent recruitment. Summary of the Invention

[0004] The purpose of this invention is to provide a management method and SaaS platform for AI-based screening and evaluation in recruitment interviews, in order to solve the above-mentioned problems existing in the prior art.

[0005] The application is as follows:

[0006] The management method based on AI-based screening and evaluation in recruitment interviews includes the following steps:

[0007] S1. Collect candidate resume data and target job requirement data, use natural language processing (NLP) algorithms to parse resumes and target job requirements, extract key information, generate resume features and target job requirement features, and use a multi-head attention neural network model to conduct the first screening of candidates. Resume features include educational background, work experience, knowledge and skills, and abilities and qualities.

[0008] S2. Collect candidates' browsing information on target positions on recruitment platforms, analyze the candidates' browsing information, perform data preprocessing and feature extraction, generate browsing information features, and establish candidate profiles; train a candidate screening model based on the candidate profiles and target position features, with the goal of training the model to perform a second screening of candidates; the browsing information includes browsing preferences, dwell time, click popularity, favorites, and recruitment interaction behavior on the target position recruitment page.

[0009] S3. Conduct video interviews with candidates who have passed the second screening. Use a camera to capture images of the candidates during the interviews and transmit the captured images to the API interface of the LibreFace pre-trained interview expression recognition model. The API automatically analyzes the candidates' language expression, logical thinking, appearance and clothing, and interview expressions in the images and outputs the interview emotion features.

[0010] S4. Preprocess the resume features, browsing information features, and interview emotion features of the candidates after the interview to obtain the preprocessed candidate application data for the target position. Perform vector feature extraction processing on the preprocessed candidate application data to obtain a candidate application vector array.

[0011] S5. Construct an interview evaluation model, input the candidate application vector array into the model, and output the candidate's application performance results;

[0012] S6. Provide the candidate's performance results to the recruiters for final decision-making.

[0013] Furthermore, in step S1, a Natural Language Processing (NLP) algorithm is used to parse the resume and target job requirements, extract key information, and generate resume features including:

[0014] Natural Language Processing (NLP) algorithms parse resumes to analyze educational background, work experience, knowledge and skills, and competencies. Through preprocessing, they perform resume text cleaning, word segmentation and part-of-speech tagging, stop word filtering, word form restoration, and stemming.

[0015] Natural Language Processing (NLP) algorithms analyze the job requirements of a target position, including job responsibilities, skills, education, work experience, certifications, language proficiency, personal qualities, work location, and hours. Preprocessing involves text cleaning, word segmentation, and part-of-speech tagging of the target job requirements. A pre-trained BERT model is then used, inputting the text data into the BERT model, which outputs a feature vector M representing the target job requirements. f ;

[0016] Based on educational background and work experience, a pre-trained NER model is used to identify and extract work time, work location, and candidate information from the resume text. A Singular Value Decomposition-based LSA model is used to analyze the text, extract the main themes and keywords, and analyze the job competency in the text. An iceberg model is used to obtain the job competency scores of the candidate's educational background and work experience. The educational background includes major, academic qualifications, foreign language proficiency, and awards. The work experience includes work time, work location, company name, company industry background, and job responsibilities.

[0017] For knowledge skills and competencies, the Word2Vec algorithm is used to extract keywords describing knowledge skills and competencies, and the NLTK syntactic analysis tool is used to parse the sentence structure of the resume text to obtain the relationships in the description of knowledge skills and competencies. The knowledge skills include subject-specific knowledge, scientific research achievements, related professional knowledge and professional training, and the competencies include job experience, teamwork ability and innovation ability.

[0018] The extracted candidate objects, topics, keywords, and sentence structures are transformed into resume features. The first BERT word embedding model is used to transform the resume features into feature vectors. The feature vectors are concatenated and weighted to form a resume feature vector.

[0019] Based on existing recruitment history datasets, collect data on the correlation between resumes' educational background, work experience, knowledge and skills, abilities and qualities, and the requirements of the target position.

[0020] The educational background, work experience, knowledge, skills, and competencies information in the resume are used as feature nodes to construct a graph structure. The graph neural network algorithm graph2vec is used to construct the feature vector of each resume. The embedding vector of the resume feature nodes is used as the second resume feature vector N. f ;

[0021] Based on multi-head attention, a weighted sum V vector of the target job requirement feature vector and the second resume feature vector is calculated. The weighted sum vector of the multi-head attention is fused to output the fused feature vector, and the candidate screening probability is calculated.

[0022] Candidate resume scores are calculated based on fused feature vectors. A two-dimensional comparison matrix is ​​constructed to analyze the fit between the target job requirements and the candidate resumes. Candidates are then screened for the first time based on their scores.

[0023] Furthermore, the step of calculating the weighted sum V vector of the target job requirement feature vector and the second resume feature vector based on multi-head attention includes:

[0024] Based on multi-head attention, Q vector, K vector, and V vector are defined according to the feature vector of the target job requirements and the feature vector of the second resume, respectively, as follows:

[0025]

[0026] in and Let Q be the weight matrix of the i-th Q-vector, the i-th K-vector, and the i-th V-vector, respectively. i K i and V i Let M represent the i-th Q-vector, the i-th K-vector, and the i-th V-vector, respectively.f and N f These represent the feature vector of the target job requirements and the feature vector of the second resume, respectively.

[0027] And calculate the weighted sum of vector V based on vectors Q, K, and V.

[0028] Furthermore, the weighted summation vector of the fused multi-head data outputs a fused feature vector, and the candidate selection probability is calculated by:

[0029] The calculation based on the weighted summation of the V vector concatenates the outputs of all weighted summation V vectors, represented as: GH(Q,K,V)=concat(Att1,Att2,...Att) h ).G O Where GH(Q,K,V) represents the fused feature vector, and Att h Let G represent the h-th weighted summation vector V. O This represents the concatenated vector output from the fused feature vectors, where concat represents the concatenation function;

[0030] The output after linear layer mapping is determined based on the fused feature vectors, and is expressed as follows:

[0031] Y i =(GH(Q,K,V).G O W+b i ,

[0032] Where Y i Let W represent the fused feature vector output after the i-th linear layer mapping, and let b represent the linear weights for the resume. i Indicates the bias term;

[0033] The sigmoid activation function is used to transform the matching feature vector of the candidate's target job requirements and the second resume's feature vector into the probability that the candidate's resume will be selected, expressed as:

[0034] Where p represents the probability that a candidate's resume is selected;

[0035] The loss L is calculated for each weight matrix using the Adagrad optimizer with an adaptive learning rate method. and And W and b i The learning rate is adaptively adjusted, and parameters are iteratively optimized. When the loss L calculated during continuous iteration meets the preset threshold, the iteration stops and the parameters are output.

[0036] Furthermore, based on the fused feature vectors, candidate resume scores are calculated, and a two-dimensional comparison matrix is ​​constructed to analyze the fit between the target job requirements and the candidate resumes. The first screening of candidates is then conducted based on their scores, including:

[0037] The fused feature vector is calculated based on the output of multi-head attention, constructing a vector based on the outputs of h heads, and then using the second resume feature vector N. f For each candidate's resume, a feature vector is constructed, which is represented as follows:

[0038] A a =[Att1,Att2,...Att] h ],

[0039] N f =[k1,k2,...k n ];

[0040] Where A a k represents the final output of multi-head attention. n This represents the characteristics of the nth candidate's resume;

[0041] Cosine similarity is introduced to adapt the feature vector by fusing the output of each head of the feature vector with the features of each candidate's resume, as follows:

[0042]

[0043] Where R m0 This represents the resume fit score of the m0th candidate. This represents the transpose of the resume features of the m0th candidate;

[0044] The requirements of the target position are compared sequentially with the characteristics of the candidates' resumes;

[0045] Construct a two-dimensional comparison matrix M, where M(m1,n1) represents the fit score between the m1-th component of the feature vector of the target job requirements and the n1-th component of the feature vector of the candidate's second resume, expressed as:

[0046] M(m1,n1)=max(M(m1-1,n1-1))+Sim(e m1 ,d n1 ).M(m1-1,n1).M(m1,n1-1),

[0047]

[0048] Where e m1 Let d represent the m1-th component of the feature vector representing the requirements of the target position. n1 Sim(e) represents the n1-th component of the second resume feature vector.m1 ,d n1 Let M(m1-1,n1) represent the cosine similarity between the m1-th component of the feature vector of the target job requirements and the n1-th component of the feature vector of the candidate's second resume. Let M(m1-1,n1-1) represent the fit score when only the first m1-1 components of the feature vector of the target job requirements are considered, without considering the n1-th component of the second resume feature vector, and the first n1-1 components of the feature vector of the second resume feature vector. Let M(m1,n1-1) represent the fit score when only the first m1 components of the feature vector of the target job requirements are considered, without considering the n1-1 component of the feature vector of the target job requirements, and the first n1-1 components of the feature vector of the second resume feature vector. Let max(M(m1-1,n1-1)) represent the maximum score of the previous fit. The total number of fits is represented as m1*n1. This represents the transpose of the n1th component of the second resume's eigenvector;

[0049] The fitting score of the two-dimensional alignment matrix M is obtained based on the calculation of M(m1,n1), and the cumulative score in the matrix is ​​selected as the final fitting score of the candidate.

[0050] Based on the fit score, the number of resumes that meet the requirements of the target position is selected.

[0051] Further, step S4 includes:

[0052] S41. The preprocessing of the resume features, browsing information features, and interview emotion features of the candidates after the interview is as follows: low-dimensional word vectors are generated for the words in all features using the word embedding technology GloVe;

[0053] S42. Generate low-dimensional word vectors g1 corresponding to the resume features:

[0054]

[0055] Where exp(.) represents an exponential function with the base of natural numbers, * represents the convolution operator, G1 represents the resume feature convolution matrix, and b1 represents the data sequence in the resume feature preprocessing result where all words are replaced with low-dimensional word vectors;

[0056] S43. Generate low-dimensional word vectors g2 corresponding to the browsing information features:

[0057]

[0058] Where G2 represents the convolution matrix of browsing information features, and b2 represents the data sequence in the browsing information feature preprocessing result where all words are replaced with low-dimensional word vectors. This represents the m-th data element in data sequence b2;

[0059] S44. Generate low-dimensional word vectors g3 corresponding to the emotion features:

[0060]

[0061] Where G3 represents the emotion feature convolution matrix, and b3 represents the data sequence in the emotion feature preprocessing result where all words are replaced with low-dimensional word vectors. This represents the m-th data element in the data sequence b3;

[0062] S45. Construct a candidate application vector array g = [g1, g2, g3].

[0063] Further, step S5 includes:

[0064] S51. Constructing an interview assessment model:

[0065] The interview assessment model includes an input layer, a job suitability initialization configuration layer, an attention extraction layer, a job suitability enhancement layer, and an output layer.

[0066] The input layer is used to receive an array of candidate application vectors.

[0067] The job suitability initialization configuration layer is used to convert the candidate application vector array into the candidate's initial application data. The job suitability initialization configuration layer is set up with three sub-layers, namely the first sub-layer, the second sub-layer, and the third sub-layer.

[0068] The first sub-layer is used to process the low-dimensional word vectors of the resumes in the candidate application vector array;

[0069] The second sub-layer is used to process the browsing information of the candidate application vector array, specifically the low-dimensional word vectors.

[0070] The third sub-layer is used to process the low-dimensional word vectors of sentiment in the candidate application vector array;

[0071] The attention extraction layer is the attention module structure in the second BERT model, used to obtain the attention data of the candidate application vector array;

[0072] The job suitability enhancement layer is used to enhance the initial application data based on attention data to obtain the candidate's job suitability.

[0073] The output layer is used to output the job suitability obtained from the evaluation, and the job suitability is expressed by a score;

[0074] S52. Use an interview assessment model to evaluate the candidate's job suitability.

[0075] Further, step S52 includes:

[0076] S521, The input layer receives the candidate application vector array g;

[0077] S522, The job suitability initialization configuration layer converts the candidate application vector array g into the candidate's initial application data Q:

[0078] Where σ represents the activation function;

[0079] S523, Attention extraction layer obtains attention data from the candidate application vector array:

[0080] B = (B1, B2, B3)

[0081]

[0082] Where B represents the attention data of the candidate application vector array g, and W5 represents the attention convolution matrix. i Represents vector g i Attention data;

[0083] S524, Job Suitability Enhancement Layer: Enhances initial job application data based on attention data.

[0084] Where Y represents the candidate's job suitability, P1 represents the attention matrix of the initial application data, δ represents the activation function, and B T Indicates the transpose of B;

[0085] S525, Output layer outputs the job suitability Y obtained from the evaluation.

[0086] The AI-based recruitment interview screening and evaluation management SaaS platform includes: a SaaS platform, a memory, and a processor, wherein the SaaS platform runs on the processor, the memory stores the SaaS platform, and when the SaaS platform is run by the processor, the processor executes the method described in any one of the recruitment interview AI screening and evaluation management methods.

[0087] Compared with the prior art, the embodiments of the present invention achieve the following beneficial effects:

[0088] This invention provides a method for collecting candidate resume data and target job requirement data, using a multi-head attention neural network model to perform an initial screening of resume features; collecting and analyzing candidates' browsing information on recruitment platforms to perform a second screening; conducting video interviews with candidates who pass the second screening, capturing their images using a camera, and outputting their interview emotion features; preprocessing the candidates' resume features, browsing information features, and interview emotion features after the interview to obtain a candidate application vector array; inputting the candidate application vector array into an interview evaluation model to output the candidates' application results; this invention uses a neural network model to screen candidates multiple times, thereby accurately selecting applicants, improving the efficiency and accuracy of intelligent recruitment, and saving recruitment human resource costs. Attached Figure Description

[0089] Figure 1 This is a flowchart illustrating the AI-based screening, evaluation, and management method for recruitment interviews provided in this embodiment of the invention. Detailed Implementation

[0090] The present invention will now be described in detail with reference to the accompanying drawings.

[0091] Example 1

[0092] First, the technical terms involved in the embodiments of this application will be explained.

[0093] (1) BERT model

[0094] BERT stands for Bidirectional Encoder Representations from Transformers. BERT's core innovation lies in its bidirectional encoder structure, which simultaneously considers left and right contextual information, thereby generating more accurate and semantically richer word representations. Through its bidirectional encoder structure and pre-training tasks, the BERT model has achieved remarkable results in the field of natural language processing. Its emergence has not only improved the performance of various NLP tasks but also driven the development of pre-trained language models. The flexibility and scalability of the BERT model enable it to adapt to a variety of complex NLP tasks, making it an important tool in the current field of natural language processing.

[0095] This invention provides an AI-based screening, evaluation, and management method for recruitment interviews, such as... Figure 1 This includes the following steps:

[0096] S1. Collect candidate resume data and target job requirement data, use natural language processing (NLP) algorithms to parse resumes and target job requirements, extract key information, generate resume features and target job requirement features, and use a multi-head attention neural network model to conduct the first screening of candidates. Resume features include educational background, work experience, knowledge and skills, and abilities and qualities.

[0097] S2. Collect candidates' browsing information on target positions on recruitment platforms, analyze the candidates' browsing information, perform data preprocessing and feature extraction, generate browsing information features, and establish candidate profiles; train a candidate screening model based on the candidate profiles and target position features, with the goal of training the model to perform a second screening of candidates; the browsing information includes browsing preferences, dwell time, click popularity, favorites, and recruitment interaction behavior on the target position recruitment page.

[0098] S3. Conduct video interviews with candidates who have passed the second screening. Use a camera to capture images of the candidates during the interviews and transmit the captured images to the API interface of the LibreFace pre-trained interview expression recognition model. The API automatically analyzes the candidates' language expression, logical thinking, appearance and clothing, and interview expressions in the images and outputs the interview emotion features. LibreFace is an open-source facial expression analysis tool.

[0099] S4. Preprocess the resume features, browsing information features, and interview emotion features of the candidates after the interview to obtain the preprocessed candidate application data for the target position. Perform vector feature extraction processing on the preprocessed candidate application data to obtain a candidate application vector array.

[0100] S5. Construct an interview evaluation model, input the candidate application vector array into the model, and output the candidate's application performance results;

[0101] S6. Provide the candidate's performance results to the recruiters for final decision-making.

[0102] Specifically, this method involves collecting candidate resume data and target job requirement data, using a multi-head attention neural network model for the first screening of resume features; collecting and analyzing candidate browsing information on recruitment platforms for the second screening; conducting video interviews with candidates who pass the second screening, capturing their images using a camera, and outputting their interview emotion characteristics; preprocessing the candidate's resume features, browsing information features, and interview emotion characteristics after the interview to obtain a candidate application vector array; inputting the candidate application vector array into an interview evaluation model to output the candidate's application results; this invention uses a neural network model to screen candidates multiple times, thereby accurately selecting applicants, improving the efficiency and accuracy of intelligent recruitment, and saving recruitment human resource costs.

[0103] In the above embodiments, specifically, step S1 uses a Natural Language Processing (NLP) algorithm to parse the resume and target job requirements, extract key information, and generate resume features including:

[0104] Natural Language Processing (NLP) algorithms parse resumes to analyze educational background, work experience, knowledge and skills, and competencies. Through preprocessing, they perform resume text cleaning, word segmentation and part-of-speech tagging, stop word filtering, word form restoration, and stemming.

[0105] Natural Language Processing (NLP) algorithms analyze the job requirements of a target position, including job responsibilities, skills, education, work experience, certifications, language proficiency, personal qualities, work location, and hours. Preprocessing involves text cleaning, word segmentation, and part-of-speech tagging of the target job requirements. A pre-trained BERT model is then used, inputting the text data into the BERT model, which outputs a feature vector M representing the target job requirements. f ;

[0106] Based on educational background and work experience, a pre-trained NER model is used to identify and extract work time, work location, and candidate information from the resume text. A Singular Value Decomposition-based LSA model is used to analyze the text, extract the main themes and keywords, and analyze the job competency in the text. An iceberg model is used to obtain the job competency scores of the candidate's educational background and work experience. The educational background includes major, academic qualifications, foreign language proficiency, and awards. The work experience includes work time, work location, company name, company industry background, and job responsibilities.

[0107] For knowledge skills and competencies, the Word2Vec algorithm is used to extract keywords describing knowledge skills and competencies, and the NLTK syntactic analysis tool is used to parse the sentence structure of the resume text to obtain the relationships in the description of knowledge skills and competencies. The knowledge skills include subject-specific knowledge, scientific research achievements, related professional knowledge and professional training, and the competencies include job experience, teamwork ability and innovation ability.

[0108] The extracted candidate objects, topics, keywords, and sentence structures are transformed into resume features. The first BERT word embedding model is used to transform the resume features into feature vectors. The feature vectors are concatenated and weighted to form a resume feature vector.

[0109] Based on existing recruitment history datasets, collect data on the correlation between resumes' educational background, work experience, knowledge and skills, abilities and qualities, and the requirements of the target position.

[0110] The educational background, work experience, knowledge, skills, and competencies information in the resume are used as feature nodes to construct a graph structure. The graph neural network algorithm graph2vec is used to construct the feature vector of each resume. The embedding vector of the resume feature nodes is used as the second resume feature vector N. f ;

[0111] Based on multi-head attention, a weighted sum V vector of the target job requirement feature vector and the second resume feature vector is calculated. The weighted sum vector of the multi-head attention is fused to output the fused feature vector, and the candidate screening probability is calculated.

[0112] Candidate resume scores are calculated based on fused feature vectors. A two-dimensional comparison matrix is ​​constructed to analyze the fit between the target job requirements and the candidate resumes. Candidates are then screened for the first time based on their scores.

[0113] It should be noted that by using different methods to extract features from the target job requirements and resume features, the accuracy of each feature extraction can be effectively improved. Using the pre-trained first BERT model to perform semantic analysis on the target job requirements text data can accurately capture subtle semantic differences in the expression of the target job requirements and form a feature vector of the target job requirements. By using the graph neural network algorithm graph2vec to extract resume features, the multi-dimensional attribute information of the resume (such as educational background, work experience, knowledge and skills, and competence) can be comprehensively considered, providing a more comprehensive feature representation for analyzing the candidate's resume. By embedding the resume feature node information into the vector space, a multi-dimensional feature representation can be constructed for each resume. Compared with the traditional resume adaptation method based on keywords or rules, the graph neural network algorithm can more accurately express the relationship between the semantic features of the resume, significantly improving the expressive power and generalization ability of feature extraction.

[0114] In the above embodiments, specifically, the step of calculating the weighted sum V vector of the target job requirement feature vector and the second resume feature vector based on multi-head attention includes:

[0115] Based on multi-head attention, Q vector, K vector, and V vector are defined according to the feature vector of the target job requirements and the feature vector of the second resume, respectively, as follows:

[0116]

[0117] in and Let Q be the weight matrix of the i-th Q-vector, the i-th K-vector, and the i-th V-vector, respectively. i K i and V i Let M represent the i-th Q-vector, the i-th K-vector, and the i-th V-vector, respectively. f and N f These represent the feature vector of the target job requirements and the feature vector of the second resume, respectively.

[0118] And calculate the weighted sum of vector V based on vectors Q, K, and V.

[0119] In the above embodiments, specifically, the weighted summation vector of the fused multi-head data outputs a fused feature vector, and the calculation of the candidate selection probability includes:

[0120] The calculation based on the weighted summation of the V vector concatenates the outputs of all weighted summation V vectors, represented as: GH(Q,K,V)=concat(Att1,Attt2,...Attt) h ).G OWhere GH(Q,K,V) represents the fused feature vector, Atth represents the h-th weighted summation V vector, and G... O This represents the concatenated vector output from the fused feature vectors;

[0121] The output after linear layer mapping is determined based on the fused feature vectors, and is expressed as follows:

[0122] Y i =(GH(Q,K,V).G O W+b i ,

[0123] Where Y i Let W represent the fused feature vector output after the i-th linear layer mapping, and let b represent the linear weights for the resume. i Indicates the bias term;

[0124] The sigmoid activation function is used to transform the matching feature vector of the candidate's target job requirements and the second resume's feature vector into the probability that the candidate's resume will be selected, expressed as:

[0125] Where p represents the probability that a candidate's resume is selected;

[0126] The loss L is calculated for each weight matrix using the Adagrad optimizer with an adaptive learning rate method. and And W and b i The learning rate is adaptively adjusted, and parameters are iteratively optimized. When the loss L calculated during continuous iteration meets the preset threshold, the iteration stops and the parameters are output.

[0127] In the above embodiments, specifically, candidate resume scores are calculated based on fused feature vectors, a two-dimensional comparison matrix is ​​constructed to analyze the fit score between the target job requirements and the candidate resumes, and the first screening of candidates is conducted based on their scores, including:

[0128] The fused feature vector is calculated based on the output of multi-head attention, constructing a vector based on the outputs of h heads, and then using the second resume feature vector N. f For each candidate's resume, a feature vector is constructed, which is represented as follows:

[0129] A a =[Att1,Att2,...Att] h ],

[0130] N f =[k1,k2,...k n ];

[0131] Where A ak represents the final output of multi-head attention. n This represents the characteristics of the nth candidate's resume;

[0132] Cosine similarity is introduced to adapt the feature vector by fusing the output of each head of the feature vector with the features of each candidate's resume, as follows:

[0133]

[0134] Where R m0 This represents the resume fit score of the m0th candidate. This represents the transpose of the resume features of the m0th candidate;

[0135] The requirements of the target position are compared sequentially with the characteristics of the candidates' resumes;

[0136] Construct a two-dimensional comparison matrix M, where M(m1,n1) represents the fit score between the m1-th component of the feature vector of the target job requirements and the n1-th component of the feature vector of the candidate's second resume, expressed as:

[0137] M(m1,n1)=max(M(m1-1,n1-1))+Sim(e m1 ,d n1 ).M(m1-1,n1).M(m1,n1-1),

[0138]

[0139] Where e m1 Let d represent the m1-th component of the feature vector representing the requirements of the target position. n1 Sim(e) represents the n1-th component of the second resume feature vector. m1 ,d n1 Let M(m1-1,n1) represent the cosine similarity between the m1-th component of the feature vector of the target job requirements and the n1-th component of the feature vector of the candidate's second resume. Let M(m1-1,n1-1) represent the fit score when only the first m1-1 components of the feature vector of the target job requirements are considered, without considering the n1-th component of the second resume feature vector, and the first n1-1 components of the feature vector of the second resume feature vector. Let M(m1,n1-1) represent the fit score when only the first m1 components of the feature vector of the target job requirements are considered, without considering the n1-1 component of the feature vector of the target job requirements, and the first n1-1 components of the feature vector of the second resume feature vector. Let max(M(m1-1,n1-1)) represent the maximum score of the previous fit. The total number of fits is represented as m1*n1. This represents the transpose of the n1th component of the second resume's eigenvector;

[0140] The fitting score of the two-dimensional alignment matrix M is obtained based on the calculation of M(m1,n1), and the cumulative score in the matrix is ​​selected as the final fitting score of the candidate.

[0141] Based on the fit score, the number of resumes that meet the requirements of the target position is selected.

[0142] It should be noted that by using attention-weighted summation of the V vector, the different dimensions of the target job requirements and resume features can be weighted and summed, thereby improving the overall information recognition capability. By defining the target job requirement feature vector and the second resume feature vector as Q vector, K vector and V vector respectively, and using attention for weighted summation calculation, the accuracy of candidate resume screening can be effectively improved.

[0143] In the above embodiments, step S2 specifically includes:

[0144] S21. Collect candidates' browsing information, analyze candidates' browsing information, perform data preprocessing and feature extraction, generate candidate browsing information feature vectors, and establish candidate profiles.

[0145] S22. Construct a candidate job preference matrix, where each element in the matrix corresponds to the candidate's rating of their preference for the target job.

[0146] S23. By matrix decomposition, the candidate job intention matrix is ​​decomposed into two full-rank matrices, the two full-rank matrices including the candidate job intention feature matrix and the target job scoring feature matrix;

[0147] S24. Solve for the candidate job preference feature matrix and the target job rating feature matrix by minimizing the loss function;

[0148] S25. Using the trained candidate job intention feature matrix and target job rating feature matrix, predict the candidate's job intention rating for the target job.

[0149] S26. Use the K-fold cross-validation method to evaluate the model's performance. The evaluation metric used is the mean squared error (MSE). Based on the prediction score, candidates are screened a second time according to their scores.

[0150] It should be noted that, in this embodiment, the construction of the candidate job intention matrix is ​​illustrated as follows: Number of candidates: 3 (Candidate 1, Candidate 2, Candidate 3), ratings (browsing bias A, browsing bias B, click popularity C, favorites D, recruitment interaction behavior E), according to step S22, the candidate job intention matrix R is constructed, and the matrix is ​​represented as follows: The scores are calculated by weighting the feature vectors of browsing bias, dwell time, click popularity, favorites, and recruitment interaction behavior in the candidate's profile. Specifically, Candidate 1 scores 5 for browsing bias A, 3 for browsing bias B, 1 for click popularity C, 1 for favorites D, and 1 for recruitment interaction behavior E; Candidate 2 scores 4 for browsing bias A, 3 for browsing bias B, 3 for click popularity C, 2 for favorites D, and 1 for recruitment interaction behavior E; Candidate 3 scores 4 for browsing bias A, 4 for browsing bias B, 2 for click popularity C, 2 for favorites D, and 5 for recruitment interaction behavior E.

[0151] According to step S23, the candidate job intention matrix is ​​decomposed into two full-rank matrices through matrix factorization. The two full-rank matrices include the candidate job intention feature matrix F and the target job rating feature matrix D, as shown in the following formula:

[0152] R = FD T , where F∈R n*k , D∈R m*k k is the dimension of the eigenvector, T represents the transpose of the matrix, n represents the number of rows in R, and m represents the number of columns in R.

[0153] According to step S24, F and D are solved by minimizing the loss function, as shown in the following formula:

[0154]

[0155] in, This means finding the values ​​of feature matrices F and D that minimize the loss function. u d is the browsing information feature vector of candidate u. i X is the feature vector of target position i, X is the set of candidate u and target position i, and θ is the regularization parameter used to balance the prediction error in the loss function.

[0156] Based on step S25, using the trained F and D matrices, predict the candidate's job intention score for the target position, as shown in the following formula:

[0157] DD=F T D, where DD represents the candidate's job preference score for the target position.

[0158] In the above embodiments, specifically, step S4 includes:

[0159] S41. The preprocessing of the resume features, browsing information features, and interview emotion features of the candidates after the interview is as follows: low-dimensional word vectors are generated for the words in all features using the word embedding technology GloVe;

[0160] S42. Generate low-dimensional word vectors g1 corresponding to the resume features:

[0161]

[0162] Where exp(.) represents an exponential function with the base of natural numbers, * represents the convolution operator, G1 represents the resume feature convolution matrix, and b1 represents the data sequence in the resume feature preprocessing result where all words are replaced with low-dimensional word vectors;

[0163] S43. Generate low-dimensional word vectors g2 corresponding to the browsing information features:

[0164]

[0165] Where G2 represents the convolution matrix of browsing information features, and b2 represents the data sequence in the browsing information feature preprocessing result where all words are replaced with low-dimensional word vectors. This represents the m-th data element in data sequence b2;

[0166] S44. Generate low-dimensional word vectors g3 corresponding to the emotion features:

[0167]

[0168] Where G3 represents the emotion feature convolution matrix, and b3 represents the data sequence in the emotion feature preprocessing result where all words are replaced with low-dimensional word vectors. This represents the m-th data element in the data sequence b3;

[0169] S45. Construct a candidate application vector array g = [g1, g2, g3].

[0170] In the above embodiments, specifically, step S5 includes:

[0171] S51. Constructing an interview assessment model:

[0172] The interview assessment model includes an input layer, a job suitability initialization configuration layer, an attention extraction layer, a job suitability enhancement layer, and an output layer.

[0173] The input layer is used to receive an array of candidate application vectors.

[0174] The job suitability initialization configuration layer is used to convert the candidate application vector array into the candidate's initial application data. The job suitability initialization configuration layer is set up with three sub-layers, namely the first sub-layer, the second sub-layer, and the third sub-layer.

[0175] The first sub-layer is used to process the low-dimensional word vectors of the resumes in the candidate application vector array;

[0176] The second sub-layer is used to process the browsing information of the candidate application vector array, specifically the low-dimensional word vectors.

[0177] The third sub-layer is used to process the low-dimensional word vectors of sentiment in the candidate application vector array;

[0178] The attention extraction layer is the attention module structure in the second BERT model, used to obtain the attention data of the candidate application vector array;

[0179] The job suitability enhancement layer is used to enhance the initial application data based on attention data to obtain the candidate's job suitability.

[0180] The output layer is used to output the job suitability obtained from the evaluation, and the job suitability is expressed by a score;

[0181] S52. Use an interview assessment model to evaluate the candidate's job suitability.

[0182] In the above embodiments, specifically, step S52 includes:

[0183] S521, The input layer receives the candidate application vector array g;

[0184] S522, The job suitability initialization configuration layer converts the candidate application vector array g into the candidate's initial application data Q:

[0185] Where σ represents the activation function;

[0186] S523, Attention extraction layer obtains attention data from the candidate application vector array:

[0187] B = (B1, B2, B3)

[0188]

[0189] Where B represents the attention data of the candidate application vector array g, and W5 represents the attention convolution matrix. i Represents vector g i Attention data;

[0190] S524, Job Suitability Enhancement Layer: Enhances initial job application data based on attention data.

[0191] Where Y represents the candidate's job suitability, P1 represents the attention matrix of the initial application data, δ represents the activation function, and B T Indicates the transpose of B;

[0192] S525, Output layer outputs the job suitability Y obtained from the evaluation.

[0193] Example 2

[0194] A recruitment interview AI-based screening, evaluation, and management SaaS platform includes: a SaaS platform, a memory, and a processor. The processor runs the SaaS platform, and the memory stores the SaaS platform. When the processor runs the SaaS platform, it causes the processor to execute the methods described in any one of the recruitment interview AI-based screening, evaluation, and management methods. It should be understood that the above embodiments are one or more embodiments of the present invention, and there are many other embodiments and variations based on the present invention. Any modifications and variations made by those skilled in the art using the present invention without groundbreaking innovation are within the protection scope of the present invention.

Claims

1. A management method for screening, evaluating, and assessing candidates based on AI in recruitment interviews, characterized in that: Includes the following steps: S1. Collect candidate resume data and target job requirement data, use Natural Language Processing (NLP) algorithms to parse the resumes and target job requirements, extract key information, and generate a second resume feature vector. and the feature vector of the target job requirements The first screening of candidates was conducted using a multi-head attention neural network model. Resume features included educational background, work experience, knowledge and skills, and abilities and qualities. The initial screening of candidates using a multi-head attention neural network model includes: Based on multi-head attention, Q vector, K vector, and V vector are defined according to the feature vector of the target job requirements and the feature vector of the second resume, respectively, as follows: in These are the weight matrices for the i-th Q-vector, the i-th K-vector, and the i-th V-vector, respectively. Let these represent the i-th Q-vector, the i-th K-vector, and the i-th V-vector, respectively. and These represent the feature vector of the target job requirements and the feature vector of the second resume, respectively. And calculate the weighted sum V vector based on the Q vector, K vector, and V vector; The weighted summation vector of the multiple heads is fused to output a fused feature vector, and the candidate selection probability is calculated, including: The calculation based on the weighted summation of the V vector concatenates the outputs of all weighted summation V vectors, represented as: Where GH(Q,K,V) represents the fused feature vector. Let V represent the h-th weighted summation vector. This represents the concatenated vector output from the fused feature vectors, where concat represents the concatenation function; The output after linear layer mapping is determined based on the fused feature vectors, and is expressed as follows: in Let represent the fused feature vector output after the i-th linear layer mapping, and W represent the linear weights for the resume. Indicates the bias term; The sigmoid activation function is used to transform the matching feature vector of the candidate's target job requirements and the second resume's feature vector into the probability that the candidate's resume will be selected, expressed as: , where p represents the probability that the candidate's resume is selected; The loss L is calculated for each weight matrix using the Adagrad optimizer with an adaptive learning rate method. And W and The learning rate is adaptively adjusted and the parameters are iteratively optimized. When the loss L calculated during continuous iteration meets the preset threshold, the iteration stops and the parameters are output. S2. Collect candidates' browsing information on target positions on recruitment platforms, analyze the candidates' browsing information, perform data preprocessing and feature extraction, generate browsing information features, and establish candidate profiles; train a candidate screening model based on the candidate profiles and target position features, with the goal of training the model to perform a second screening of candidates; the browsing information includes browsing preferences, dwell time, click popularity, favorites, and recruitment interaction behavior on the target position recruitment page. S3. Conduct video interviews with candidates who have passed the second screening. Use a camera to capture images of the candidates during the interviews and transmit the captured images to the API interface of the LibreFace pre-trained interview expression recognition model. The API automatically analyzes the candidates' language expression, logical thinking, appearance and clothing, and interview expressions in the images and outputs the interview emotion features. S4. Preprocess the resume features, browsing information features, and interview emotion features of the candidates after the interview to obtain the preprocessed candidate application data for the target position. Perform vector feature extraction processing on the preprocessed candidate application data to obtain a candidate application vector array. S5. Construct an interview evaluation model, input the candidate application vector array into the model, and output the candidate's application performance results; S6. Provide the candidate's performance results to the recruiters for final decision-making.

2. The AI-based screening and evaluation management method for recruitment interviews as described in claim 1, characterized in that, Step S1 uses a Natural Language Processing (NLP) algorithm to parse the resume and target job requirements, extract key information, and generate resume features including: Natural Language Processing (NLP) algorithms parse resumes to analyze educational background, work experience, knowledge and skills, and competencies. Through preprocessing, resume text cleaning, word segmentation and part-of-speech tagging, stop word filtering, word form restoration, and stemming are performed. Natural Language Processing (NLP) algorithms analyze the job requirements of a target position, including job responsibilities, skills, education, work experience, certifications, language proficiency, personal qualities, and work location and schedule. Preprocessing involves text cleaning, word segmentation, and part-of-speech tagging of the target job requirements. A pre-trained BERT model is then used, inputting the text data into the BERT model, which outputs a feature vector of the target job requirements. ; For educational background and work experience, a pre-trained NER model is used to identify and extract work time, work location, and candidate information from the resume text. A Singular Value Decomposition-based LSA model is used to analyze the text, extract themes and keywords, and analyze the job competency in the text. An iceberg model is used to obtain the job competency scores of the candidate's educational background and work experience. The educational background includes major, academic qualifications, foreign language proficiency, and awards. The work experience includes work time, work location, company name, company industry background, and job responsibilities. For knowledge skills and competencies, the Word2Vec algorithm is used to extract keywords describing knowledge skills and competencies, and the NLTK syntactic analysis tool is used to parse the sentence structure of the resume text to obtain the relationships in the description of knowledge skills and competencies. The knowledge skills include subject-specific knowledge, scientific research achievements, related professional knowledge and professional training, and the competencies include job experience, teamwork ability and innovation ability. The extracted candidate objects, topics, keywords, and sentence structures are transformed into resume features. The first BERT word embedding model is used to transform the resume features into feature vectors. The feature vectors are concatenated and weighted to form a resume feature vector. Based on existing recruitment history datasets, collect data on the correlation between resumes' educational background, work experience, knowledge and skills, abilities and qualities, and the requirements of the target position. The educational background, work experience, knowledge, skills, and competencies information in the resume are used as feature nodes to construct a graph structure. The graph neural network algorithm graph2vec is used to construct the feature vector for each resume. The embedding vector of the resume feature nodes is used as the second resume feature vector. ; Based on multi-head attention, a weighted sum V vector of the target job requirement feature vector and the second resume feature vector is calculated. The weighted sum vector of the multi-head attention is fused to output the fused feature vector, and the candidate screening probability is calculated. Candidate resume scores are calculated based on fused feature vectors. A two-dimensional comparison matrix is ​​constructed to analyze the fit between the target job requirements and the candidate resumes. Candidates are then screened for the first time based on their scores.

3. The AI-based screening and evaluation management method for recruitment interviews as described in claim 2, characterized in that, Candidate resume scores are calculated based on fused feature vectors. A two-dimensional comparison matrix is ​​constructed to analyze the fit between the target job requirements and the candidate resumes. Candidates are then screened based on their scores, including: The fused feature vector is calculated based on the output of multi-head attention, and a vector is constructed according to the outputs of h heads. The second resume feature vector is then used. For each candidate's resume, a feature vector is constructed, which is represented as follows: in This represents the final output of the bullish attention. This represents the characteristics of the nth candidate's resume; Cosine similarity is introduced to adapt the feature vector by fusing the output of each head of the feature vector with the features of each candidate's resume, as follows: in This represents the resume fit score of the m0th candidate. This represents the transpose of the resume features of the m0th candidate; The requirements of the target position are compared sequentially with the characteristics of the candidates' resumes; Construct a two-dimensional comparison matrix M, where M(m1,n1) represents the fit score between the m1-th component of the feature vector of the target job requirements and the n1-th component of the feature vector of the candidate's second resume, expressed as: in This represents the m1-th component of the feature vector representing the requirements of the target position. This represents the n1th component of the second resume's feature vector. Let M(m1-1,n1) represent the cosine similarity between the m1-th component of the feature vector of the target job requirements and the n1-th component of the feature vector of the candidate's second resume. Let M(m1-1,n1) represent the fit score when only the first m1-1 components of the target job requirements feature vector and the first n1 components of the second resume feature vector are considered, without considering the n1-th component of the second resume feature vector. Let M(m1,n1-1) represent the fit score when only the first m1 components of the target job requirements feature vector and the first n1-1 components of the second resume feature vector are considered, without considering the n1-1 component of the target job requirements feature vector. Let max(M(m1-1,n1-1)) represent the maximum score of the previous fit. The total number of fits is represented as m1*n1. This represents the transpose of the n1th component of the second resume's eigenvector; The fitting score of the two-dimensional alignment matrix M is obtained based on the calculation of M(m1,n1), and the cumulative score in the matrix is ​​selected as the final fitting score of the candidate. Based on the fit score, the number of resumes that meet the requirements of the target position is selected.

4. The AI-based screening and evaluation management method for recruitment interviews as described in claim 1, characterized in that, Step S4 includes: S41. The preprocessing of the resume features, browsing information features, and interview emotion features of the candidates after the interview is as follows: low-dimensional word vectors are generated for the words in all features using the word embedding technology GloVe; S42. Generate low-dimensional word vectors of resume features. : Where exp(.) represents an exponential function to the base natural numbers, and * represents the convolution operator. Represents the convolution matrix of resume features; This represents a data sequence in which all words in the resume feature preprocessing results are replaced with low-dimensional word vectors. S43. Generate low-dimensional word vectors of browsing information corresponding to browsing information features. : in, Represents the convolution matrix of browsing information features. This indicates a data sequence in which all words in the browsing information feature preprocessing results are replaced with low-dimensional word vectors. Represents a data sequence The m-th data in; S44. Generate low-dimensional word vectors of emotion corresponding to emotion features. : in, Represents the convolution matrix of emotion features. This indicates a data sequence in which all words in the emotion feature preprocessing results are replaced with low-dimensional word vectors. Represents a data sequence The m-th data in; S45. Constructing a candidate application vector array .

5. The AI-based screening and evaluation management method for recruitment interviews as described in claim 4, characterized in that, Step S5 includes: S51. Constructing an interview assessment model: The interview assessment model includes an input layer, a job suitability initialization configuration layer, an attention extraction layer, a job suitability enhancement layer, and an output layer. The input layer is used to receive an array of candidate application vectors. The job suitability initialization configuration layer is used to convert the candidate application vector array into the candidate's initial application data. The job suitability initialization configuration layer is set up with three sub-layers, namely the first sub-layer, the second sub-layer, and the third sub-layer. The first sub-layer is used to process the low-dimensional word vectors of the resumes in the candidate application vector array; The second sub-layer is used to process the browsing information of the candidate application vector array, specifically the low-dimensional word vectors. The third sub-layer is used to process the low-dimensional word vectors of sentiment in the candidate application vector array; The attention extraction layer is the attention module structure in the second BERT model, used to obtain the attention data of the candidate application vector array; The job suitability enhancement layer is used to enhance the initial application data based on attention data to obtain the candidate's job suitability. The output layer is used to output the job suitability obtained from the evaluation, and the job suitability is expressed by a score; S52. Use an interview assessment model to evaluate the candidate's job suitability.

6. The management method for AI-based screening and evaluation in recruitment interviews according to claim 5, characterized in that, Step S52 includes: S521, The input layer receives the candidate application vector array g; S522, The job suitability initialization configuration layer converts the candidate application vector array g into the candidate's initial application data Q: ,in, Indicates the activation function; S523, Attention extraction layer obtains attention data from the candidate application vector array: Where B represents the attention data of the candidate application vector array g. Represents the attention convolution matrix. Representing vectors Attention data; S524, Job Suitability Enhancement Layer: Enhances initial job application data based on attention data. Where Y represents the candidate's job suitability, and P1 represents the attention matrix of the initial application data. This represents the activation function. Indicates the transpose of B; S525, Output layer outputs the job suitability Y obtained from the evaluation.

7. A recruitment interview AI-based screening, evaluation, and management SaaS platform, characterized by: include: A SaaS platform, a memory, and a processor, wherein the processor runs the SaaS platform, the memory stores the SaaS platform, and the SaaS platform, when run by the processor, causes the processor to perform the method as described in any one of claims 1 to 6.

Images