AI screening evaluation management method based on recruitment interview and saas platform

Abstract

The invention discloses a recruitment interview AI-based screening evaluation management method and a saas platform, and the method comprises the steps: collecting resume data of candidates and target post requirement data, and carrying out the first screening of resume features through a multi-head attention neural network model; collecting browsing information of the candidates for the target post on the recruitment platform, analyzing the browsing information of the candidates, and screening the candidates for the second time; video interview is carried out on the candidates screened for the second time, images of the candidates are collected through a camera, and interview emotion features are output; preprocessing the resume features, the browsing information features and the interview emotion features of the candidate after interview to obtain a candidate application vector array; inputting the candidate employment vector array into an interview evaluation model, and outputting employment results of candidates; according to the method, the candidates are screened for multiple times through the neural network model, so that the applicants are accurately screened, the intelligent recruitment efficiency and accuracy are improved, and the recruitment human resource cost is saved.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence recruitment, and more specifically, to a management method and a SaaS platform for AI screening, evaluation, and management based on recruitment interviews. Background Art

[0002] With the rapid development of information technology and the popularization of the Internet, the recruitment process has gradually changed from the traditional human resource management method to a digital and intelligent form. 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 many challenges, and enterprises' demands for improving recruitment efficiency, reducing costs, and optimizing talent matching are becoming increasingly urgent, which provides a market demand for the application of intelligent recruitment methods.

[0003] In related technologies, natural language processing technology is used to parse resume information, and then the parsed results are matched with the recruitment content corresponding to the target company to achieve intelligent recruitment. However, in related technologies, only direct matching and screening can be performed during resume matching and screening, it is difficult to obtain hidden associated information from resume information, and at the same time, the browsing information characteristics and interview emotion characteristics of candidates are not introduced for multi-level artificial intelligence analysis, and AI analysis is not performed on the recruitment data throughout the recruitment process. As a result, the analysis may not be accurate enough during the recruitment process, thereby reducing the efficiency and accuracy of intelligent recruitment. Summary of the Invention

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

[0005] Specifically, this application is as follows:

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

[0007] S1. Collect the resume data of candidates and the data of target position requirements, use natural language processing (NLP) algorithms to parse the resumes and target position requirements, extract key information, generate resume features and target position requirement features, and use a multi-head attention neural network model to perform the first screening on candidates. The resume features include educational background, work experience, knowledge and skills, and ability and accomplishment;

[0008] S2. Collect the browsing information of candidates on the target position on the recruitment platform, analyze the browsing information of candidates, perform data preprocessing and feature extraction, generate browsing information features, and establish candidate portraits; train a candidate screening model based on the candidate portraits and target position features. The goal of model training is to perform the second screening on candidates; the browsing information includes browsing preferences, residence time, click popularity, collection hobbies, and recruitment interaction behaviors of the target position recruitment page.

[0009] S3. Conduct a video interview for the candidates who pass the second screening. Use a camera to capture the images of the candidates during the interview, and transmit the captured images to the API interface of the LibreFace pre-trained interview expression recognition model. The API automatically analyzes the language expression, logical thinking, appearance and dressing, and interview expressions of the candidates' 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 all the candidate application data corresponding to the target position after preprocessing. Perform vector feature extraction processing on the preprocessed candidate application data to obtain a candidate application vector array.

[0011] S5. Build 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 application performance results to the recruiters for the final decision.

[0013] Furthermore, in step S1, use the natural language NLP algorithm to parse the resume and the requirements of the target position, extract key information, and generate resume features including:

[0014] The natural language NLP algorithm parses the educational background, work experience, knowledge and skills, and ability qualities of the resume. Through preprocessing, perform resume text cleaning, word segmentation and part-of-speech tagging, stop word filtering, word form reduction, and stem extraction.

[0015] The natural language NLP algorithm parses the job responsibilities, skill requirements, educational background, work experience, qualification certificates, language ability, personal qualities, work location and time required by the target position. Through preprocessing, perform target position requirement text cleaning, word segmentation and part-of-speech tagging. Use the pre-trained first BERT model, input the text data into the first BERT model, and the first BERT model outputs the target position requirement feature vector M f ;

[0016] For the educational background and work experience, use the pre-trained NER model to identify and extract the working time, work location, and candidates' objects in the resume text. Use the LSA model based on singular value decomposition to analyze the text, extract the main themes and keywords, analyze the job competence tendency in the text, and use the iceberg model to obtain the job competence tendency scores of the candidates' educational background and work experience. The educational background includes major, academic degree, foreign language level, and award-winning situation. The work experience includes working time, work location, company name, company's industry background, and job responsibility content.

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

[0018] The extracted candidate object, the subject, the keywords, and the sentence structure are transformed into the resume features, and the resume features are transformed into feature vectors using the word embedding first BERT model; the feature vectors are concatenated and weighted averaged to form a resume feature vector;

[0019] Based on the existing recruitment history dataset, collect the interrelated data of the educational background, work experience, knowledge, skills, and competencies information of the resume and the requirements of the target position;

[0020] Use the educational background, work experience, knowledge, skills, and competencies information of the resume as resume feature nodes to construct a graph structure. Use the graph neural network algorithm graph2vec to form the feature vector of each resume, and use the embedded vector of the resume feature node as the second resume feature vector N f ;

[0021] Based on multi-head attention, calculate the weighted sum V vector of the target position requirement feature vector and the second resume feature vector, fuse the weighted sum value vectors of multiple heads to output a fused feature vector, and calculate the screening probability of the candidate;

[0022] Calculate the candidate resume score based on the fused feature vector, construct a two-dimensional comparison matrix to analyze the matching score between the target position requirements and the candidate resume, and conduct the first screening of candidates according to the candidate scores.

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

[0024] Based on multi-head attention, define the Q vector, K vector, and V vector according to the target position requirement feature vector and the second resume feature vector respectively, which are expressed as:

[0025]

[0026] where and are the weight matrices of the i-th Q vector, the i-th K vector, and the i-th V vector respectively, Q i 、K i and V i represent the i-th Q vector, the i-th K vector, and the i-th V vector respectively, Mf and N f respectively represent the target position requirement feature vector and the second resume feature vector;

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

[0028] Furthermore, the weighted sum value vector of the fusion multi-head outputs a fusion feature vector, and calculating the screening probability of the candidate includes:

[0029] Based on the calculation of the weighted sum V vector, the outputs of all weighted sum V vectors are concatenated, expressed as: GH(Q, K, V) = concat(Att1, Att2,... Att h ).G O , where GH(Q, K, V) represents the fusion feature vector, Att h represents the h-th weighted sum V vector, and G O represents the concatenated vector output of the fusion feature vector, and concat represents the concatenation function;

[0030] Based on the fusion feature vector, determine the output after linear layer mapping, expressed as:

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

[0032] where Y i represents the output of the i-th fusion feature vector after linear layer mapping, W represents the linear weight for the resume, and b i represents the bias term;

[0033] Use the sigmoid activation function to convert the matching feature vector of the candidate's target position requirement feature vector and the second resume feature vector into the probability that the candidate's resume is selected, expressed as:

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

[0035] Use the Adagrad optimizer to calculate the loss L for each weight matrix and as well as W and b i by the adaptive learning rate method, and perform parameter iteration optimization. When the calculated loss L in the continuous iteration process meets the preset threshold, stop the iteration and output the parameters.

[0036] Further, calculate the candidate resume scores based on the fused feature vectors, construct a two-dimensional comparison matrix to analyze the matching scores between the target position requirements and the candidate resumes, and conduct the first screening of the candidates according to the candidate scores, including:

[0037] Based on the fused feature vectors output by the multi-head attention, construct vectors according to the outputs of the included h heads, and according to the second resume feature vector N f Construct vectors for the characteristics of each candidate resume, which are respectively expressed as:

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

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

[0040] where A a represents the final output of the multi-head attention, and k n represents the characteristics of the resume of the nth candidate;

[0041] Introduce the cosine similarity to match through the output of each head of the fused feature vector and the characteristics of each candidate resume, which is expressed as:

[0042]

[0043] where R m0 represents the matching score of the resume of the m0th candidate, represents the transpose of the characteristics of the resume of the m0th candidate;

[0044] Perform a sequence comparison between the target position requirements and the candidate resume characteristics;

[0045] Construct a two-dimensional comparison matrix M, where M(m1,n1) represents the matching score between the m1th component of the target position requirement feature vector and the n1th component of the second resume feature vector of the candidate, which is 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 represents the m1th component of the target position requirement feature vector, and d n1 represents the n1th component of the second resume feature vector, Sim(em1 ,d n1 ) represents the cosine similarity between the m1-th component of the target position requirement feature vector and the n1-th component of the candidate's second resume feature vector. M(m1 - 1, n1) represents the adaptation score considering only the first m1 - 1 components of the target position requirement feature vector and the first n1 components of the second resume feature vector, excluding the n1-th component of the second resume feature vector. M(m1, n1 - 1) represents the adaptation score considering only the first m1 components of the target position requirement feature vector and the first n1 - 1 components of the second resume feature vector, excluding the n1 - 1-th component of the target position requirement feature vector. max(M(m1 - 1, n1 - 1)) represents the maximum score of the previous adaptation. The total number of adaptation times is expressed as m1 * n1, represents the transpose of the n1-th component of the second resume feature vector;

[0049] The adaptation score of the two-dimensional comparison matrix M is obtained according to the calculation of M(m1, n1), and the cumulative score in the matrix is selected as the final adaptation score of the candidate;

[0050] According to the level of the adaptation score, the number of resumes meeting the requirements of the target position is screened.

[0051] Furthermore, step S4 includes:

[0052] S41. Preprocess the resume features, browsing information features, and interview emotion features of the candidates after the interview as follows: Generate low-dimensional word vectors for the words in all features through the word embedding technology GloVe;

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

[0054]

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

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

[0057]

[0058] where G2 represents the browsing information feature convolution matrix, b2 represents the data sequence in which all words in the browsing information feature preprocessing result are replaced by low-dimensional word vectors, represents the m-th data in the data sequence b2;

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

[0060]

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

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

[0063] Furthermore, step S5 includes:

[0064] S51. Construct an interview evaluation model:

[0065] The interview evaluation model includes an input layer, a job adaptation ability initialization configuration layer, an attention extraction layer, a job adaptation ability enhancement layer, and an output layer;

[0066] The input layer is used to receive the candidate application vector array;

[0067] The job adaptation ability initialization configuration layer is used to convert the candidate application vector array into the candidate's initial application data, and three sub-layers are set in the job adaptation ability initialization configuration layer, and the three sub-layers include a first sub-layer, a second sub-layer, and a third sub-layer;

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

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

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

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

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

[0073] The output layer is used to output the evaluated job adaptation ability, and the job adaptation ability is expressed by a score;

[0074] S52. Use the interview evaluation model to evaluate the candidate's job adaptation ability.

[0075] Furthermore, step S52 includes:

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

[0077] S522. The job adaptation ability initialization and configuration layer converts the candidate application vector array g into the initial application data Q of the candidate:

[0078] where σ represents the activation function;

[0079] S523. The attention extraction layer obtains the attention data of 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, W5 represents the attention convolution matrix, and B i represents the attention data of the vector g i ;

[0083] S524. The job adaptation ability enhancement layer enhances the initial application data based on the attention data:

[0084] where Y represents the job adaptation ability of the candidate, P1 represents the attention matrix of the initial application data, δ represents the activation function, and B T represents the transpose of B;

[0085] S525. The output layer outputs the evaluated job adaptation ability Y.

[0086] Based on the recruitment interview AI screening and evaluation management SaaS platform, including: the SaaS platform, the memory, and the processor, where the SaaS platform runs in the processor, the memory stores the SaaS platform, and when the SaaS platform runs by the processor, the processor is made to execute any one of the methods in the recruitment interview AI screening and evaluation management method.

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

[0088] An embodiment of the present invention provides a method for collecting resume data of candidates and requirement data of target positions, and using a multi-head attention neural network model to perform the first screening on resume features; collecting the browsing information of candidates for the target position on the recruitment platform, analyzing the browsing information of candidates, and performing the second screening on candidates; conducting a video interview on the candidates who pass the second screening, using a camera to collect the images of candidates, and outputting interview emotion features; preprocessing the resume features, browsing information features, and interview emotion features of the candidates after the interview to obtain a candidate application vector array; inputting the candidate application vector array into an interview evaluation model to output the application results of candidates; the present invention screens candidates multiple times through a neural network model, thereby accurately screening applicants, improving the efficiency and accuracy of intelligent recruitment, and saving the recruitment human resource cost. Description of the Drawings

[0089] Figure 1 It is a schematic flowchart of a method for AI screening, evaluation and management based on recruitment interviews provided by an embodiment of the present invention. Detailed Embodiments

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

[0091] Embodiment 1

[0092] First, technical terms related to the embodiments of the present application will be explained.

[0093] (1) BERT model

[0094] BERT, short for Bidirectional Encoder Representations from Transformers. The core innovation of BERT lies in its bidirectional encoder structure, which can consider both the left and right context information simultaneously, 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 not only improves the performance of various NLP tasks but also promotes the development of pre-trained language models. The flexibility and scalability of the BERT model enable it to adapt to various complex NLP tasks and become an important tool in the current field of natural language processing.

[0095] An embodiment of the present invention provides a method for AI screening, evaluation and management based on recruitment interviews, such as Figure 1 , including the following steps:

[0096] S1. Collect the resume data of candidates and the requirement data of the target position. Use natural language NLP algorithms to parse the resumes and the requirements of the target position, extract key information, generate resume features and target position requirement features, and conduct the first screening of candidates using a multi-head attention neural network model. Resume features include educational background, work experience, knowledge and skills, and ability and accomplishment.

[0097] S2. Collect the browsing information of candidates for the target position on the recruitment platform, analyze the browsing information of candidates, conduct data preprocessing and feature extraction, generate browsing information features, and establish candidate portraits. Train a candidate screening model based on candidate portraits and target position features. The goal of model training is to conduct the second screening of candidates. The browsing information includes browsing preferences, residence time, click popularity, collection hobbies, and recruitment interaction behaviors on the recruitment page of the target position.

[0098] S3. Conduct a video interview for candidates who pass the second screening. Use a camera to collect the images of candidates during the interview, transmit the collected images to the API interface of the LibreFace pre-trained interview expression recognition model. The API automatically analyzes the language expression, logical thinking, appearance and dressing, and interview expressions of the candidate's images and outputs interview emotion features. LibreFace is an open-source facial expression analysis tool.

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

[0100] S5. Build 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 application performance results to the recruiters for the final decision.

[0102] Specifically, this method provides for collecting resume data of candidates and job requirement data of target positions, and using a multi-head attention neural network model to conduct the first screening of resume features; collecting the browsing information of candidates for the target position on the recruitment platform, analyzing the browsing information of candidates, and conducting the second screening of candidates; conducting video interviews on candidates who pass the second screening, using a camera to collect the images of candidates, and outputting interview emotion features; preprocessing the resume features, browsing information features, and interview emotion features of candidates after the interview to obtain an array of candidate application vectors; inputting the array of candidate application vectors into an interview evaluation model to output the application results of candidates; the present invention conducts multiple screenings of candidates through a neural network model, thereby accurately screening applicants, improving the efficiency and accuracy of intelligent recruitment, and saving the recruitment human resource cost.

[0103] In the above embodiment, specifically, in step S1, a natural language NLP algorithm is used to parse the resume and the job requirements of the target position, extract key information, and generate resume features including:

[0104] The natural language NLP algorithm parses the educational background, work experience, knowledge and skills, and ability and accomplishment of the resume, and through preprocessing, conducts resume text cleaning, word segmentation and part-of-speech tagging, stop word filtering, word form restoration, and stem extraction;

[0105] The natural language NLP algorithm parses the job responsibilities, skill requirements, educational background, work experience, qualification certificates, language ability, personal accomplishment, work location and time of the target position requirements. Through preprocessing, text cleaning, word segmentation and part-of-speech tagging of the target position requirements are carried out. The pre-trained first BERT model is used to input the text data into the first BERT model, and the first BERT model outputs the target position requirement feature vector M f ;

[0106] For the educational background and work experience, a pre-trained NER model is used to identify and extract the working time, work location, and objects of candidates in the resume text, the LSA model based on singular value decomposition is used to analyze the text, extract the main themes and keywords, analyze the job competence tendency in the text, and the iceberg model is used to obtain the job competence tendency scores of the educational background and work experience of candidates. The educational background includes major, academic degree, foreign language level, and award-winning situation, and the work experience includes working time, work location, company name, company's industry background, and job responsibility content;

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

[0108] The extracted candidate object, the subject, the keywords, and the sentence structure are transformed into the resume features, and the resume features are transformed into feature vectors using the word embedding first BERT model; the feature vectors are concatenated and weighted averaged to form a resume feature vector.

[0109] Based on the existing recruitment history dataset, collect the interrelated data of the educational background, work experience, knowledge, skills, and competencies information of the resume and the requirements of the target position.

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

[0111] Based on multi-head attention, calculate the weighted sum V vector of the target position requirement feature vector and the second resume feature vector, fuse the weighted sum value vectors of multiple heads to output the fused feature vector, and calculate the screening probability of the candidate.

[0112] Calculate the candidate resume score based on the fused feature vector, construct a two-dimensional comparison matrix to analyze the matching score between the target position requirements and the candidate resume, and conduct the first screening of the candidates according to the candidate scores.

[0113] It should be noted that by adopting different methods for feature extraction of the target position requirements features 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 position requirements text data can accurately capture the subtle semantic differences in the expression of the target position requirements, and form the feature vector of the target position 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, ability and accomplishment, etc.) can be comprehensively considered, providing a more comprehensive feature representation for parsing 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 association between the semantic features of the resume, significantly improving the expression ability and generalization ability of feature extraction.

[0114] In the above embodiment, specifically, calculating the weighted sum V vector of the target position requirements feature vector and the second resume feature vector based on multi-head attention includes:

[0115] Based on multi-head attention, define the Q vector, K vector, and V vector according to the target position requirements feature vector and the second resume feature vector respectively, which are expressed as:

[0116]

[0117] Where and are the weight matrices of the i-th Q vector, the i-th K vector, and the i-th V vector respectively, Q i , K i and V i represent the i-th Q vector, the i-th K vector, and the i-th V vector respectively, M f and N f represent the target position requirements feature vector and the second resume feature vector respectively;

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

[0119] In the above embodiment, specifically, the fusion of the weighted sum value vectors of multiple heads outputs the fusion feature vector and calculates the screening probability of the candidate, including:

[0120] Based on the calculation of the weighted sum V vector, splice the outputs of all weighted sum V vectors, which is expressed as: GH(Q,K,V) = concat(Att1,Attt2,...Attt h ).G O, where GH(Q, K, V) represents the fused feature vector, Atth represents the h-th weighted sum of the V vectors, and G O represents the concatenated vector of the fused feature vector output;

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

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

[0123] where Y i represents the output of the fused feature vector after linear layer mapping for the i-th one, W represents the linear weight for the resume, and b i represents the bias term;

[0124] The matching feature vector of the candidate's target position requirement feature vector and the second resume feature vector is transformed into the probability that the candidate's resume is selected using the sigmoid activation function, expressed as:

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

[0126] The Adagrad optimizer is used to calculate the adaptive learning rate of the loss L for each weight matrix and as well as W and b i by the adaptive learning rate method, and parameter iterative optimization is performed. When the calculated loss L in the continuous iteration process meets the preset threshold, the iteration stops and the parameters are output.

[0127] In the above embodiment, specifically, the candidate resume score is calculated based on the fused feature vector, a two-dimensional comparison matrix is constructed to analyze the matching score between the target position requirements and the candidate resume, and the first screening of candidates is performed according to the candidate scores, including:

[0128] Based on the fused feature vector output by the multi-head attention calculation, vectors are constructed according to the outputs of the included h heads, and according to the second resume feature vector N f Vectors are constructed for the features of each candidate resume, respectively expressed as:

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

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

[0131] where A aRepresents the final output of the multi-head attention, k n Represents the features of the resume of the nth candidate;

[0132] Introduce cosine similarity to adapt by fusing the output of each head of the feature vector and the features of each candidate's resume, expressed as:

[0133]

[0134] Where R m0 Represents the adaptation score of the resume of the m0th candidate, Represents the transpose of the features of the resume of the m0th candidate;

[0135] Perform a sequence comparison between the requirements of the target position and the features of the candidate's resume;

[0136] Construct a two-dimensional comparison matrix M, where M(m1,n1) represents the adaptation score between the m1th component of the feature vector of the target position requirements and the n1th component of the second resume feature vector of the candidate, 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 Represents the m1th component of the feature vector of the target position requirements, d n1 Represents the n1th component of the second resume feature vector, Sim(e m1 ,d n1 ) represents the cosine similarity between the m1th component of the feature vector of the target position requirements and the n1th component of the second resume feature vector of the candidate, M(m1 - 1,n1) represents the adaptation score without considering the n1th component of the second resume feature vector, only considering the first m1 - 1 components of the feature vector of the target position requirements and the first n1 components of the second resume feature vector, M(m1,n1 - 1) represents the adaptation score without considering the n1 - 1th component of the feature vector of the target position requirements, only considering the first m1 components of the feature vector of the target position requirements and the first n1 - 1 components of the second resume feature vector, max(M(m1 - 1,n1 - 1)) represents the maximum score of the previous adaptation, and the total number of adaptations is expressed as m1*n1, Represents the transpose of the n1th component of the second resume feature vector;

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

[0141] According to the level of the adaptation score, the number of resumes meeting the requirements of the target position is screened.

[0142] It should be noted that by the method of attention weighted summation of the V vector, different dimensions of the target position requirements and resume characteristics can be weighted and summed, so as to improve the information recognition ability as a whole. By respectively defining the target position requirement feature vector and the second resume feature vector as the Q vector, K vector and V vector, and using attention for weighted summation calculation, the accuracy of candidate resume screening can be effectively improved.

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

[0144] S21. Collect the browsing information of candidates, analyze the browsing information of candidates, perform data preprocessing and feature extraction, generate the browsing information feature vector of candidates, and establish a candidate portrait;

[0145] S22. Construct a candidate position intention matrix, where each element in the matrix corresponds to the score of the candidate's intention for the target position;

[0146] S23. Through matrix decomposition, decompose the candidate position intention matrix into two full-rank matrices, and the two full-rank matrices include a candidate position intention feature matrix and a target position score feature matrix;

[0147] S24. By minimizing the loss function, solve the candidate position intention feature matrix and the target position score feature matrix;

[0148] S25. Use the trained candidate position intention feature matrix and target position score feature matrix to predict the score of the candidate's position intention for the target position;

[0149] S26. Use the K-fold cross-validation method to evaluate the performance of the model. The evaluation index used is the mean square error MSE. According to the predicted score, the candidates are screened for the second time according to the level of the score.

[0150] It should be noted that in this embodiment, an example of constructing a candidate position intention matrix is as follows: the number of candidates: 3 (candidate 1, candidate 2, candidate 3), scores (browsing preference A, browsing preference B, click popularity C, collection hobby D, recruitment interaction behavior E). According to step S22, a candidate position intention matrix R is constructed, and the matrix is expressed as: The scoring is calculated by weighting the eigenvectors of browsing preferences, dwell time, click popularity, collection preferences, and recruitment interaction behaviors in the candidate's portrait. Among them, Candidate 1 scores 5 for browsing preference A, 3 for browsing preference B, 1 for click popularity C, 1 for collection preference D, and 1 for recruitment interaction behavior E; Candidate 2 scores 4 for browsing preference A, 3 for browsing preference B, 3 for click popularity C, 2 for collection preference D, and 1 for recruitment interaction behavior E; Candidate 3 scores 4 for browsing preference A, 4 for browsing preference B, 2 for click popularity C, 2 for collection preference D, and 5 for recruitment interaction behavior E;

[0151] According to step S23, through matrix factorization, the candidate position intention matrix is decomposed into two full-rank matrices, and the two full-rank matrices include the candidate position intention feature matrix F and the target position scoring feature matrix D. The formula is as follows:

[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 of R, and m represents the number of columns of R;

[0153] According to step S24, by minimizing the loss function, F and D are solved. The formula is as follows:

[0154]

[0155] Among them, means to find the values of the feature matrices F and D such that the value of the loss function is minimized. f u is the browsing information eigenvector of candidate u, d i is the eigenvector 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] According to step S25, using the trained F and D matrices, the scoring of the candidate position intention for the target position is predicted. The formula is as follows:

[0157] DD = F T D, where DD represents the scoring of the candidate position intention for the target position.

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

[0159] S41. Preprocess the resume features, browsing information features, and interview emotion features of the candidates after the interview as follows: Generate low-dimensional word vectors for the words in all features through the word embedding technology GloVe;

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

[0161]

[0162] Among them, exp(.) represents the exponential function with the natural number as the base, * represents the convolution operator, G1 represents the resume feature convolution matrix; b1 represents the data sequence in which all words in the resume feature preprocessing result are replaced with low-dimensional word vectors;

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

[0164]

[0165] Among them, G2 represents the browsing information feature convolution matrix, b2 represents the data sequence in which all words in the browsing information feature preprocessing result are replaced with low-dimensional word vectors, represents the m-th data in the data sequence b2;

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

[0167]

[0168] Among them, G3 represents the emotion feature convolution matrix, b3 represents the data sequence in which all words in the emotion feature preprocessing result are replaced with low-dimensional word vectors, represents the m-th data in the data sequence b3;

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

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

[0171] S51. Construct an interview evaluation model:

[0172] The interview evaluation model includes an input layer, a job matching ability initialization configuration layer, an attention extraction layer, a job matching ability enhancement layer, and an output layer;

[0173] The input layer is used to receive the candidate application vector array;

[0174] The job matching ability initialization configuration layer is used to convert the candidate application vector array into the candidate's initial application data, and three sub-layers are set in the job matching ability initialization configuration layer, and the three sub-layers include a first sub-layer, a second sub-layer, and a third sub-layer;

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

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

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

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

[0179] The job adaptation ability enhancement layer is used to enhance the initial application data based on the attention data to obtain the job adaptation ability of the candidate;

[0180] The output layer is used to output the evaluated job adaptation ability, and the job adaptation ability is expressed by a score;

[0181] S52. Use the interview evaluation model to evaluate the job adaptation ability of the candidate.

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

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

[0184] S522. The job adaptation ability initialization configuration layer converts the candidate application vector array g into the initial application data Q of the candidate:

[0185] where σ represents the activation function;

[0186] S523. The attention extraction layer obtains the attention data of 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, W5 represents the attention convolution matrix, and B i represents the attention data of the vector g i ;

[0190] S524. The job adaptation ability enhancement layer enhances the initial application data based on the attention data:

[0191] where Y represents the job adaptation ability of the candidate, P1 represents the attention matrix of the initial application data, δ represents the activation function, and B T represents the transpose of B;

[0192] S525. The output layer outputs the evaluated job adaptation ability Y.

[0193] Embodiment 2

[0194] Based on the recruitment interview AI screening and evaluation management SaaS platform, including: the SaaS platform, a memory, and a processor. The SaaS platform runs in the processor, and the memory stores the SaaS platform. When the SaaS platform runs by the processor, it causes the processor to execute the method described in any one of the recruitment interview AI screening and evaluation management methods. It should be understood that the above embodiments are one or more embodiments of the present invention. Based on the present invention, there are many other embodiments and their variations. When ordinary technicians in this industry do not make pioneering innovations, the variations and modifications made through the present invention all fall within the protection scope of the present invention.

Claims

1. Based on the AI screening and evaluation management method for recruitment interviews, it is characterized in that, It includes the following steps: S1. Collect the resume data of candidates and the requirement data of the target position, use the natural language NLP algorithm to parse the resume and the requirements of the target position, extract key information, generate resume features and target position requirement features, and use the multi-head attention neural network model to conduct the first screening of candidates. The resume features include educational background, work experience, knowledge and skills, and ability and accomplishment; S2. Collect the browsing information of candidates for the target position on the recruitment platform, analyze the browsing information of candidates, conduct data preprocessing and feature extraction, generate browsing information features, and establish candidate portraits; Train a candidate screening model based on the candidate portraits and target position features. The goal of model training is to conduct the second screening of candidates; The browsing information includes browsing preference, stay time, click popularity, collection hobby, and recruitment interaction behavior of the target position recruitment page; S3. Conduct a video interview for the candidates who pass the second screening, use the camera to collect the images of the candidates during the interview, transmit the collected images to the API interface of the LibreFace pre-trained interview expression recognition model, and the API automatically analyzes the language expression, logical thinking, appearance and dress, and interview expression of the candidates' images and outputs interview emotion features; S4. Preprocess the resume features, the browsing information features, and the interview emotion features of the candidates after the interview to obtain all the candidate application data corresponding to the target position after preprocessing, and conduct 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 application performance results to the recruiters for the final decision.

2. The management method for AI screening and evaluation based on recruitment interviews according to claim 1 is characterized in that, In step S1, the natural language NLP algorithm is used to parse the resume and the requirements of the target position, extract key information, and the generated resume features include: The natural language NLP algorithm parses the educational background, work experience, knowledge and skills, and ability and accomplishment of the resume. Through preprocessing, resume text cleaning, word segmentation and part-of-speech tagging, stop word filtering, word form reduction, and stemming are carried out; The natural language NLP algorithm analyzes the job responsibilities, skill requirements, educational background, work experience, qualification certificates, language ability, personal qualities, work location and time required by the target position. Through preprocessing, the text cleaning, word segmentation and part-of-speech tagging of the target position requirements text are carried out. The pre-trained first BERT model is used to input the text data into the first BERT model, and the first BERT model outputs the target position requirement feature vector M f ; For the educational background and work experience, use the pre-trained NER model to identify and extract the working time, working location, and candidate object in the resume text, use the LSA model based on singular value decomposition to analyze the text, extract the theme and keywords, analyze the job competence tendency in the text, and use the iceberg model to obtain the job competence tendency scores of the candidate's educational background and work experience. The educational background includes major, academic degree, foreign language level, and award-winning situation, and the work experience includes working time, working location, company name, company's industry background, and job responsibility content; Regarding knowledge, skills, and competencies, the Word2Vec algorithm is used to extract keywords describing knowledge, skills, and competencies. The syntactic analysis tool NLTK is used to parse the sentence structure of the resume text to obtain the relationships in the descriptions of knowledge, skills, and competencies. The knowledge and skills include subject professional knowledge, research achievements, relevant professional knowledge, and professional training. The competencies include job work experience, teamwork ability, and innovation ability. The extracted candidate object, the subject, the keywords, and the sentence structure are transformed into the resume features. The word embedding first BERT model is used to transform the resume features into feature vectors. The feature vectors are concatenated and weighted averaged to form a resume feature vector. Based on the existing recruitment history dataset, collect the interrelated data of the educational background, work experience, knowledge, skills, competencies information of the resume, and the requirements of the target position. Construct the educational background, work experience, knowledge and skills, and competency information of the resume as resume feature nodes into a graph structure, use the graph neural network algorithm graph2vec to form the feature vector of each resume, and use the embedded vector of the resume feature node as the second resume feature vector N f ; Based on multi-head attention, calculate the weighted sum V vector of the target position requirement feature vector and the second resume feature vector, fuse the weighted sum value vectors of multiple heads to output the fused feature vector, and calculate the screening probability of the candidate. Calculate the candidate resume score based on the fused feature vector, construct a two-dimensional comparison matrix to analyze the matching score between the target position requirements and the candidate resume, and conduct the first screening of candidates according to the candidate scores from high to low.

3. The management method for AI screening and evaluation based on recruitment interviews according to claim 2, wherein The calculating the weighted sum V vector of the target position requirement feature vector and the second resume feature vector based on multi-head attention includes: Based on multi-head attention, define the Q vector, K vector, and V vector according to the target position requirement feature vector and the second resume feature vector, which are expressed as: Among them and are the weight matrices of the i-th Q vector, the i-th K vector, and the i-th V vector respectively, where Q i , K i and V i represent the i-th Q vector, the i-th K vector, and the i-th V vector respectively, M f and N f represent the target job requirement feature vector and the second resume feature vector respectively; And calculate the weighted sum V vector based on the Q vector, K vector, and V vector.

4. The management method for AI screening and evaluation based on recruitment interviews according to claim 2, characterized in that The fusing the weighted sum value vectors of multiple heads to output the fused feature vector and calculating the screening probability of the candidate includes: Based on the calculation of the weighted sum V vector, splice the outputs of all weighted sum V vectors, which is expressed as: GH(Q, K, V) = concat(Att1, Att2,... Att h ). G O , where GH(Q, K, V) represents the fused feature vector, Att h represents the h-th weighted sum V vector, G O represents the concatenated vector output from the fused feature vector, and concat represents the concatenation function; Based on the fused feature vector, determine the output after linear layer mapping, which is expressed as: Y i = (GH(Q, K, V).G O )W + b i , Among them, Y i represents the output of the i-th fused feature vector after being mapped by the linear layer, W represents the linear weight for the resume, and b i represents the bias term; Use the sigmoid activation function to transform the matching feature vector of the candidate target position requirement feature vector and the second resume feature vector into the probability that the candidate resume is selected, which is expressed as: where p represents the probability that the candidate's resume is selected; Calculate the loss L for each weight matrix using the Adagrad optimizer through an adaptive learning rate method and as well as W and b i for adaptive adjustment of the learning rate, and perform parameter iterative optimization. Stop the iteration and output the parameters when the calculated loss L in the continuous iteration process meets the preset threshold.

5. The screening, evaluation and management method based on AI for recruitment interviews according to claim 2, wherein Calculating the candidate resume score based on the fused feature vector, constructing a two-dimensional comparison matrix to analyze the matching score between the target position requirements and the candidate resume, and conducting the first screening of candidates according to the candidate scores from high to low includes: Based on the fused feature vectors calculated from the outputs of multi-head attention, vectors are constructed according to the outputs of h heads included, and according to the second resume feature vector N f Vectors are constructed for the features of each candidate's resume, respectively represented as: A a = [Att1, Att2,... Att h , N f = [k1, k2,... k n ; Among them, A a represents the final output of the multi-head attention, and k n represents the features of the resume of the nth candidate; Introduce cosine similarity to adapt through the output of each head of the fused feature vector and the features of each candidate resume, which is expressed as: where R m0 represents the resume adaptation score of the m0-th candidate, represents the transpose of the resume features of the m0-th candidate; Conduct a sequence comparison between the target position requirements and the candidate resume features. Construct a two-dimensional comparison matrix M, where M(m1,n1) represents the matching score between the m1-th component of the target position requirement feature vector and the n1-th component of the candidate's second resume feature vector, which is expressed as: M(m1,n1) = max(M(m1 - 1,n1 - 1)) + Sim(e m1 ,d n1 ).M(m1 - 1,n1).M(m1,n1 - 1), where e m1 represents the m1-th component of the target job requirement feature vector, d n1 represents the n1-th component of the second resume feature vector, Sim(e m1 , d n1 ) represents the cosine similarity between the m1-th component of the target job requirement feature vector and the n1-th component of the candidate's second resume feature vector, M(m1 - 1, n1) represents the adaptation score considering only the first m1 - 1 components of the target job requirement feature vector and the first n1 components of the second resume feature vector without considering the n1-th component of the second resume feature vector, M(m1, n1 - 1) represents the adaptation score considering only the first m1 components of the target job requirement feature vector and the first n1 - 1 components of the second resume feature vector without considering the n1 - 1-th component of the target job requirement feature vector, max(M(m1 - 1, n1 - 1)) represents the maximum score of the previous adaptation, and the total number of adaptations is expressed as m1 * n1, represents the transpose of the n1-th component of the second resume feature vector; Obtain the matching score of the two-dimensional comparison matrix M according to the calculation of M(m1,n1), and select the cumulative score in the matrix as the final matching score of the candidate. According to the level of the matching score, screen the number of resumes that meet the requirements of the target position.

6. The screening, evaluation and management method based on AI for recruitment interviews according to claim 1, characterized in that, Step S4 includes: S41. Preprocess the resume features, browsing information features, and interview emotion features of the candidates after the interview as follows: Generate low-dimensional word vectors for the words in all features through the word embedding technology GloVe; S42. Generate the resume low-dimensional word vector g1 corresponding to the resume features: where exp(.) represents the exponential function with the natural number as the base, * represents the convolution operator, G1 represents the resume feature convolution matrix; b1 represents the data sequence in which all words in the resume feature preprocessing result are replaced with low-dimensional word vectors; S43. Generate the browsing information low-dimensional word vector g2 corresponding to the browsing information features: Among them, G2 represents the convolutional matrix of browsing information features, and b2 represents the data sequence in which all words in the preprocessing result of browsing information features are replaced with low-dimensional word vectors. represents the m-th data in the data sequence b2; S44. Generate the emotion low-dimensional word vector g3 corresponding to the emotion features: Among them, G3 represents the emotional feature convolution matrix, and b3 represents the data sequence in which all words in the preprocessed result of emotional features are replaced with low-dimensional word vectors. represents the m-th data in the data sequence b3; S45. Construct the candidate application vector array g = [g1, g2, g3].

7. The screening, evaluation and management method based on AI for recruitment interviews according to claim 6, wherein Step S5 includes: S51. Construct an interview evaluation model: The interview evaluation model includes an input layer, a job suitability ability initialization configuration layer, an attention extraction layer, a job suitability ability enhancement layer, and an output layer; The input layer is used to receive the candidate application vector array; The job suitability ability initialization configuration layer is used to convert the candidate application vector array into the candidate's initial application data, and three sub-layers are set in the job suitability ability initialization configuration layer, and the three sub-layers include a first sub-layer, a second sub-layer, and a third sub-layer; The first sub-layer is used to process the resume low-dimensional word vectors of the candidate application vector array; The second sub-layer is used to process the browsing information low-dimensional word vectors of the candidate application vector array; The third sub-layer is used to process the emotion low-dimensional word vectors of the candidate application vector array; The attention extraction layer is the attention module structure in the second BERT model, and is used to obtain the attention data of the candidate application vector array; The job suitability ability enhancement layer is used to enhance the initial application data based on the attention data to obtain the candidate's job suitability ability; The output layer is used to output the evaluated job suitability ability, and the job suitability ability is expressed by a score; S52. Use the interview evaluation model to evaluate the candidate's job suitability ability.

8. The screening, evaluation and management method based on AI for recruitment interviews according to claim 7, wherein The step S52 includes: S521. The input layer receives the candidate application vector array g; S522. The job suitability ability initialization configuration layer converts the candidate application vector array g into the candidate's initial application data Q: Among them, σ represents the activation function; S523. The attention extraction layer obtains the attention data of the candidate application vector array: B = (B1, B2, B3), Among them, B represents the attention data of the candidate application vector array g, W5 represents the attention convolution matrix, and B i represents the vector g i 's attention data; S524. The job suitability ability enhancement layer enhances the initial application data based on the attention data: Among them, Y represents the job adaptation ability of the candidate, P1 represents the attention matrix of the initial application data, δ represents the activation function, and B T represents the transpose of B; S525. The output layer outputs the evaluated job suitability ability Y.

9. The AI screening, evaluation and management SaaS platform for recruitment interviews is characterized in that Includes: A saas platform, a memory, and a processor, wherein the saas platform runs in the processor, the memory stores the saas platform, and when the saas platform runs by the processor, the processor executes the method according to any one of claims 1 to 8.

Images