A talent recommendation method based on deep learning

Abstract

The application provides a talent recommendation method based on deep learning, and belongs to the field of deep learning; solves the problem of low talent recommendation efficiency; specifically as follows: resume information and recruitment requirements are acquired, a text model is trained based on a knowledge distillation mechanism combined with the recruitment requirements; the trained text model is used to convert the resume information into a feature vector; a standard feature vector is acquired, and the feature vector of a job seeker is processed by clustering, and then a Gaussian mixture model is used for screening to obtain an interviewee; the interview results of each interviewee are recorded, and the job seekers are recommended to each recruitment party; the interview results of each job seeker are recorded in real time, new job seeker resume information is acquired, and the job seekers are recommended to each recruitment party; the application acquires, analyzes and processes the resume information of the job seekers and the recruitment information of the recruitment parties, recommends the job seekers to each recruitment party, and improves the efficiency of talent recruitment.

Description

Technical Field

[0001] This invention relates to a talent recommendation method based on deep learning, and pertains to the field of deep learning. Background Technology

[0002] Existing methods for talent recommendation have the following shortcomings:

[0003] The shortcomings of content-based recommendation systems: Some existing recommendation methods match talents and positions by analyzing the content features of the job list. The main drawback of this method is the limitation of feature extraction: First, the processing of text information faces the challenge of semantic understanding, and simple keyword matching cannot capture the true meaning of complex job requirements and talents' abilities; Second, this method is prone to over-specialization, tending to recommend positions that are highly similar to the user's past choices.

[0004] The problem of combining multiple recommendation methods: Some existing recommendation methods achieve talent recommendation by combining multiple recommendation techniques (such as weighting, switching, cascading, and feature combination). However, the main problem with this approach is increased system complexity. The integration of multiple algorithms significantly increases the difficulty of system design, implementation, and maintenance. Interactions between different recommendation methods may produce unexpected side effects, such as amplifying the bias of a particular method or introducing new noise, consuming more computational resources, and requiring frequent parameter tuning, which is a considerable burden for many organizations.

[0005] Limitations of deep learning methods: There are few existing deep learning methods applied to the field of talent recommendation, and some existing computer models have poor interpretability. The black-box nature of neural networks makes it difficult for human resource experts and job seekers to understand the reasons for the recommendation results, which reduces the credibility and practicality of the system. Summary of the Invention

[0006] In view of the shortcomings of existing technologies, the purpose of this invention is to provide a talent recommendation method based on deep learning, which aims to solve the problem of low efficiency in talent recommendation.

[0007] To achieve the above objectives, the present invention provides a talent recommendation method based on deep learning, comprising:

[0008] The process involves obtaining the number of job seekers and their resumes; the number of employers and the number of openings and requirements for each employer; classifying job seekers based on their applied positions and the openings of each employer; training a text model based on a knowledge distillation mechanism and the requirements of each employer; and using the trained text model to convert each job seeker's resume information into a feature vector.

[0009] The process involves acquiring the resumes of full-time employees for each position from each recruiter and converting them into standard feature vectors; using these standard feature vectors to cluster the job seekers' feature vectors to obtain stable vector clusters; constructing the probability density function for each stable vector cluster using a Gaussian mixture model and filtering job seekers to obtain interviewees; recording the interview scores of each interviewee and recommending applicants to each recruiter.

[0010] It records the interview scores of each job seeker in real time, obtains new job seeker resume information, and recommends candidates to various recruiters.

[0011] Furthermore, the specific steps for a single division are as follows:

[0012] Extract the number of job postings from the 1st to the retrieval position. (1) ~rc (re) ;

[0013] Extract the job positions applied for by each job seeker; based on the job positions applied for by each job seeker and the job positions offered by each recruiter, divide the job seekers into sets A(1,1)~A(re,rc). (re) );

[0014] Extract the job requirements of the first job posting from the first recruiter and use it as text information tex(1,1); remove punctuation marks and text information that are not related to the job requirements from the text information tex(1,1); train a text model mod(1,1) based on the text information tex(1,1); and use this model mod(1,1) to convert the resume information in set A(1,1) into a feature vector.

[0015] Extract the second job posting from the first recruiter to the rc position of the re-th recruiter. (re) The job requirements for each recruitment position are used to train a text model, resulting in text model mod(1,2) to text model mod(re,rc). (re) ), and set A(1,2)~A(re,rc) (re) The resume information in the document is converted into a feature vector.

[0016] Furthermore, the training steps for the text model mod(1,1) are as follows:

[0017] Perform paragraph and sentence segmentation on the text information tex(1,1) to obtain the number of paragraphs pa and the number of sentences se in each paragraph. (1) ~se (pa) ;

[0018] Generate a token for each sentence in the text information tex(1,1), resulting in tok(1,1)~tok(pa,se)(pa) );

[0019] Based on the BERT model, construct a teacher model:

[0020] Let tok(1,1) ~ tok(pa,se) (pa) These are fed into the token embedding layer, segment embedding layer, and position embedding layer of the BERT model, respectively, to obtain the vector code, position code, and sentence code for each sentence;

[0021] Add the vector code, position code, and sentence code corresponding to each sentence to obtain the vector ve(1,1)~ve(pa,se). (pa) );

[0022] Let vectors ve(1, 1) ~ ve(pa, se) be used. (pa) As input to the two-layer Transformer encoder of the BERT model, it learns the relationship between each phrase corresponding to each sentence in the text information tex(1,1) and performs fill-in-the-blank learning on the BERT model;

[0023] A student model is constructed based on a multi-layer BILSTM model.

[0024] Furthermore, the specific steps for performing fill-in-the-blank learning on the BERT model are as follows:

[0025] Step 1: Create fill-in-the-blank questions:

[0026] The BERT model randomly selects multiple sentences and labels one original sentence among the selected sentences for fill-in-the-blank learning; the BERT model also masks a phrase in the original sentence.

[0027] Step 2: Do the fill-in-the-blank questions:

[0028] The BERT model uses the masked phrases in the original sentence to randomly select some phrases from the selected non-original sentences to form options, and guesses the probability of each option being correct.

[0029] The vector corresponding to the original sentence after being masked is vh (mask) The probability of guessing the correct answer is poc(vm):

[0030] poc(vm) = softmax(W (cl) ·vh (mask) +b (cl) );

[0031] Among them, W (cl) and b (cl) These represent the weights and biases of the classification layer in the BERT model, respectively.

[0032] The BERT model selects the option with the highest probability of being correct as the guessed answer and extracts the phrase vector vm of the phrase corresponding to the guessed answer;

[0033] Step 3: Compare with the reference answer:

[0034] Extract the phrase vector vt from the reference answer, and use the softmax() function to calculate the normalized value softmax(vt) of the vector vt;

[0035] Calculate the cosine similarity between vt and vm, the probability of guessing the correct answer, and the cross-entropy loss of the normalized value of vector vt;

[0036] The cosine similarity between vt and vm is: cosine_similarity(vt, vm);

[0037] Let the cross-entropy loss be Lt(poc(vm), softmax(vtt)):

[0038] Lt(poc(vm), softmax(vtt))=poc(vm)×softmax(vt);

[0039] The loss function Ltt for constructing the teacher model's fill-in-the-blank learning:

[0040]

[0041] Step 4: Repeat steps 1 to 3 to train the teacher model, and optimize W using the gradient descent algorithm. (cl) and b (cl) This continues until the loss function reaches its minimum value, thus completing the construction of the teacher model.

[0042] Furthermore, the specific steps for constructing the student model are as follows:

[0043] Processing the embedding layer: Tok(1, 1) ~ Tok(pa, se (pa) The inputs are respectively fed into the embedding layers of the teacher model and the student model; the output of the embedding layer of the teacher model is denoted as te, and the output of the embedding layer of the student model is denoted as st, and the loss function Lf of the embedding layer is constructed. (emb) :

[0044] Lf (emb) =MSE((st*We),te); where MSE represents the mean squared error; We represents the weight matrix of the student model's embedding layer;

[0045] The output of the student model's embedding layer is used as the input to the student model's hidden layer. The same steps are repeated for the embedding layer to construct the loss function Lf for the student model's hidden layer. (hid) :

[0046]

[0047] Where si represents the number of hidden layers in the student model, and hst (si) Wh represents the output of the si-th hidden layer of the student model. (si) dh represents the weight matrix of the si-th hidden layer of the student model. (st) This represents the total number of hidden layers in the student model;

[0048] ti represents the number of hidden layers in the teacher model, hte (ti) This represents the output of the ti-th hidden layer of the student model;

[0049] The correspondence between si and ti is as follows:

[0050]

[0051] Where round represents the floor function; dh (te) This represents the total number of hidden layers in the teacher model;

[0052] In the prediction layer, the student model learns to fill in the blanks.

[0053] Furthermore, the process of obtaining We is as follows:

[0054] Obtain the embedding layer dimension d of the teacher model (te) The embedding layer dimension d of the student model (st) ;

[0055] Construct a (d (st) ×d (te) Let the matrix be denoted as We. Initialize We using Xavier or Glorot so that each element in We has a mean of zero and a standard deviation of . The normal distribution;

[0056] The gradient of the embedding layer loss of the student model with respect to the weight matrix is ​​calculated using the chain rule.

[0057] Where T denotes the transpose of the matrix; N (emb) This indicates the number of tokens input to the student model's embedding layer in a single input.

[0058] Let the learning rate be η, and let the updated weight matrix be Wee. The update matrix We is:

[0059]

[0060] Furthermore, the steps for fill-in-the-blank learning on the student model are as follows:

[0061] Step 1: Teacher model creates fill-in-the-blank questions:

[0062] The teacher model randomly selects multiple sentences and marks one original sentence among the selected sentences for fill-in-the-blank learning; the teacher model also masks a phrase in the original sentence.

[0063] Step 2: Student models complete fill-in-the-blank questions:

[0064] The teacher model, based on the masked phrases in the original sentence, randomly selects some phrases from the selected non-original sentences to form ot options; the teacher model extracts the phrase vector (VR) of the reference answer. (te) The probability pt of the teacher model generating the correct reference answer (true) ;

[0065] The student model inputs each option into the hidden layer and outputs the hidden layer's output or... (1) ~or (ot) Use the softmax function to convert or (1) ~or (ot) Normalize to the interval [0,1] and correct using the temperature parameter to obtain the correct probability pst of the 1st to otth options. (1) pst (2) ~pst (ot) ;

[0066] The student model selects the option with the highest probability of being correct as the guessed answer and extracts the probability pst of the guessed answer. (st) and phrase vector vo (st) ;

[0067] Step 3: Compare with the reference answer:

[0068] Calculate VR (te) With vo (st) The cosine similarity and the probability of guessing the correct answer are compared with the cross-entropy loss of the normalized value of vector vt;

[0069] VR (te) With vo (st) The cosine similarity is: cosine_similarity(vr (te) vo (st) );

[0070] Let the cross-entropy loss be Lt(pst) (st) pt (true) ):

[0071] Lt(pst (st) pt (true) ) = pst (st) ×pt (true) ;

[0072] The loss function Lts for constructing the student model's fill-in-the-blank learning:

[0073]

[0074] Furthermore, the steps for fill-in-the-blank learning on the student model also include:

[0075] Step 4: The teacher model records the probability of the student model's correct answer being true in a fill-in-the-blank learning session (ppt).

[0076] Without changing the original sentences or the reference answer, the teacher model recomposes the options from other unselected non-original sentences, and then conducts a second fill-in-the-blank training for the student model; the teacher model records the probability of the student model correctly answering the reference answer during the second fill-in-the-blank training. (true) ;

[0077] Judge PPT (true) With PTT (true) Are they equal?

[0078] If they are equal, no action is taken; if they are not equal, the PPTs are compared. (true) With PTT (true) Adjust the size and temperature parameters accordingly;

[0079] If PPT (true) >ptt (true) If the value is large, then the temperature parameter value will be increased by a factor of:

[0080] If PPT (true) <ptt (true) If the value is large, then the temperature parameter value will be reduced by a factor of:

[0081] Step 5: Repeat steps 1 to 4 to train the student model, and use the gradient descent algorithm to optimize the weight matrix of the embedded layer and hidden layer of the student model as well as the temperature parameter until the value of the loss function is minimized, thus completing the construction of the student model and obtaining the text model mod(1,1).

[0082] Furthermore, the specific steps to transform the resume information in set A(1,1) into feature vectors are as follows:

[0083] Using the resume information in set A(1,1) as input to the text model mod(1,1), we obtain the word vector sequence we. (1) ~we (an) Where, an represents the number of resumes in set A(1,1);

[0084] The word vector sequence we (1) ~we(an) These are used as inputs to the CNN model and the BILSTM model, respectively.

[0085] In a CNN model, the word vector sequence we (1) we (2) ~we (an) Two convolutions, two activations, and two max pooling operations are performed to obtain a multi-dimensional feature vector; a Flatten layer is then used to transform the multi-dimensional feature vector into a one-dimensional feature vector, yielding the cof(x) value. (1) ,cof (2) ~cof (an) ;

[0086] In the BILSTM model, for the word vector sequence we (1) ~we (an) Bidirectional semantic feature extraction is performed using forward LSTM and backward LSTM to obtain a two-dimensional feature vector; then, a max pooling layer is used to transform the two-dimensional feature vector into a one-dimensional feature vector, resulting in the bof. (1) bof (2) ~bof (an) ;

[0087] The one-dimensional feature vectors cof obtained from the CNN model and the BILSTM model (1) ~cof (an) and bof (1) ~bof (an) The features are concatenated and fused to obtain the mixed feature vector mi. (1) ~mi (an) ;

[0088] Mix the feature vector mi (1) ~mi (an) As input to the fully connected layer of the text model mod(1,1), the feature vector fv of set A(1,1) is obtained. (1) ~fv (an) ;

[0089] Based on feature vector fv (1) ~fv (an) Screen job seekers for the first job opening offered by the first recruiter.

[0090] Furthermore, the specific steps for screening job seekers are as follows:

[0091] Let vec be the standard feature vector of the first job posting provided by the first employer. (1) ; with vec (1) Using the center vector, a clustering algorithm is used to cluster fv (1) ~fv (an)Perform clustering to obtain stable vector clusters; count the number sta of eigenvectors within each stable vector cluster.

[0092] Let the eigenvectors in the stable vector cluster be denoted as vet. (1) ~vet (sta) ;

[0093] Let the eigenvector vet (1) The parameter in is: ev (1，1) ev (1，2) ;

[0094] Similarly, the eigenvector vet (sta) The parameter in is: ev (sta，1) ev (sta，2) ;

[0095] Construct matrix B (1) Calculate the mean vector avet and the centered matrix B. (2) Covariance matrix B (3) ;

[0096] Construct the probability density function p(vet) (j) ), calculate the eigenvector vet (1) ~vet (sta) The probability density p(vet) (1) )~p(vet (sta) );

[0097] Calculate the standard deviation vector sdv, calculate the probability density of vectors avet and sdv, and obtain p(avet) and p(sdv);

[0098] In p(vet) (1) )~p(vet (sta) In the given information, the feature vectors with probability density in the interval [p(avet), p(sdv)] are extracted as the target vectors.

[0099] The job seeker corresponding to the target vector will be the interviewee for the first job offered by the first recruiter.

[0100] Compared with the prior art, the beneficial effects of the present invention are:

[0101] Feature alignment based on recruitment requirements: This invention employs a knowledge distillation mechanism to transform resume information into feature vectors, achieving precise feature alignment with recruitment requirements. Traditional resume analysis methods typically use general feature extraction pipelines, making it difficult to adaptively adjust to the specific needs of different positions. However, this invention can explicitly encode the specific requirements of the recruiter into the feature representation learning by designing specific teacher model training objectives. During the training phase, the teacher model not only learns to accurately parse resume content but also learns to identify the feature dimensions most relevant to the job requirements through job description data. This makes the feature vectors generated by the student model inherently job-suitable, improving the accuracy of subsequent recommendation processes.

[0102] Cross-domain transfer learning capability: This invention possesses cross-domain transfer learning capability, an advantage that is difficult to achieve with traditional methods. In real-world recruitment scenarios, companies often face the challenge of emerging positions or scarce talent. This invention utilizes knowledge distillation to achieve effective knowledge transfer by separating the two stages of feature learning and recommendation adaptation. Specifically, the teacher model can be pre-trained on large-scale general resume data. When facing specific domains or emerging positions, only a relatively small amount of domain data is needed to fine-tune the teacher model, and then the adjusted knowledge is transferred to the student model through distillation. This method significantly reduces the requirements for the amount of training data in the target domain, enabling the system to quickly adapt to new recruitment scenarios.

[0103] Advantages of probabilistic matching of feature vectors: This invention uses Gaussian Mixture Model (GMM) for talent feature matching in talent recommendation. Compared with traditional threshold-based screening methods, this invention models feature vectors through the multivariate Gaussian distribution of GMM (i.e., the off-diagonal elements of the covariance matrix automatically capture the dependencies between different features). In the matching stage, GMM calculates the probability density function for each job seeker, and job seekers with a high degree of matching with the feature vectors of formal employees are recommended. This has a significant advantage in the recruitment of positions that require a balance of multiple abilities. Attached Figure Description

[0104] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0105] Figure 1 This is a schematic diagram of the method of the present invention;

[0106] Figure 2 This is a schematic diagram of the processing flow of the present invention;

[0107] Figure 3 This is a schematic diagram of the feature vector transformation of the present invention. Detailed Implementation

[0108] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0109] Please see Figure 1 and Figure 2 One talent recommendation method based on deep learning includes:

[0110] Step S1: Obtain the number of job seekers and their resume information; obtain the number of recruiters and their recruitment quotas and requirements; divide the job seekers into groups based on their applied positions and the recruitment positions of each recruiter, and then train a text model based on the knowledge distillation mechanism and the recruitment requirements of each recruiter; use the trained text model to convert the resume information of each job seeker into a feature vector.

[0111] The specific steps of step S1 are as follows:

[0112] Step S11: Count the number of job seekers (jo) and the number of recruiters (re);

[0113] For each recruiter's recruitment requirements, extract the number of job openings for the 1st, 2nd, and so on, up to the reth recruiter, to obtain rc. (1) rc (2) ~rc (re) ;

[0114] Extract the job postings applied for by each job seeker from their resume information; based on the job postings applied for by each job seeker and the job postings of each recruiter, divide the job seekers into sets A(1,1), A(1,2)~A(1,rc). (1) )~A(re, rc (re) );

[0115] The set A(1,1) of job seekers represents the set of resumes of job seekers applying for the first job offered by the first recruiter.

[0116] The set A(1,2) of job seekers represents the set of resumes of job seekers applying for the second job offered by the first recruiter.

[0117] And so on, the set of job seekers is A(1, rc). (1) This indicates that the applicant applied for the first job offered by the recruiter. (1) A set of resumes of job seekers for each job opening; A(re, rc) set of job seekers. (re) This indicates that the applicant is applying for the position provided by the recruiter at position re. (re) A collection of resumes from job seekers for each advertised position;

[0118] It should be noted that if there are two or more job postings from different employers, the set of (job seekers) for each job posting must be the same; for example, "the first job posting from the first employer" and "the job posting from the re-th employer" are the same. (re) If the number of job postings is the same, then the set of job seekers A(1, 1) and the set of job seekers A(re, rc) are identical. (re) The content is consistent;

[0119] Step S12: Train the text model based on the knowledge distillation mechanism and the recruitment requirements of each recruiter;

[0120] Step S121: Extract the recruitment requirements of the first job posting from the first recruiter as text information tex(1,1); remove punctuation marks and text information that are not related to the recruitment requirements from the text information tex(1,1), and train the text model mod(1,1) based on the text information tex(1,1);

[0121] Perform paragraph and sentence segmentation on the text information tex(1,1) to obtain the number of paragraphs pa and the number of sentences in each paragraph (i.e., the number of sentences in the 1st, 2nd, up to the path paragraph) se. (1) ,se (2) ~se (pa) ;

[0122] Generate a token for each sentence in the text information tex(1,1), resulting in tok(1,1), tok(1,2) to tok(1,se) (1) )~tok(pa,se (pa) );

[0123] Where tok(1,1) represents the token of the first sentence of the first paragraph in the text information tex(1,1);

[0124] tok(1,2) represents the token of the second sentence in the first paragraph of the text information tex(1,1);

[0125] And so on, tok(1, se (1) ) represents the first segment and the second element in the text information tex(1,1). (1) The token of a sentence; tok(pa, se (pa) ) represents the pa-th segment and se-th sub-segment in the text information tex(1,1). (pa) The token of the sentence;

[0126] Step S122: Construct a teacher model based on the BERT model (i.e., a pre-trained text model based on the Transformer architecture);

[0127] (Embedding processing of BERT model) tok(1,1)~tok(pa,se) (pa) The data are fed into the Token Embedding, Segment Embeddings, and Position Embeddings layers of the BERT model, respectively, to obtain the vector code, position code, and sentence code for each sentence.

[0128] Vector encoding: Token embedding for the BERT model, representing the semantics of each sentence.

[0129] Location encoding: For the segmented embedding of the BERT model, it represents the paragraph to which each sentence belongs;

[0130] Sentence encoding: For the BERT model, it represents the position of each sentence in the corresponding paragraph;

[0131] Add the vector code, position code, and sentence code corresponding to each sentence to obtain tok(1,1)~tok(pa,se) (pa) The corresponding vectors ve(1, 1) ~ ve(pa, se) (pa) );

[0132] Let vectors ve(1, 1) ~ ve(pa, se) be used. (pa) As input to the two-layer Transformer encoder of the BERT model, it learns the relationship between each phrase corresponding to each sentence in the text information tex(1,1), preparing for the BERT model's fill-in-the-blank learning;

[0133] The BERT model undergoes fill-in-the-blank learning, and the steps are as follows:

[0134] Step 1: Create fill-in-the-blank questions:

[0135] The BERT model randomly selects multiple sentences and labels one original sentence among the selected sentences for fill-in-the-blank learning; the BERT model also masks a phrase in the original sentence.

[0136] Step 2: Do the fill-in-the-blank questions:

[0137] The BERT model uses the masked phrases in the original sentence to randomly select some phrases from the selected non-original sentences to form options, and guesses the probability of each option being correct.

[0138] The vector corresponding to the original sentence after being masked is vh (mask) The probability of guessing the correct answer is poc(vm):

[0139] poc(vm) = softmax(W (cl) ·vh (mask) +b (cl) );

[0140] Where · represents scalar multiplication; W (cl) and b (cl) These represent the weights and biases of the classification layer in the BERT model, respectively.

[0141] The BERT model selects the option with the highest probability of being correct as the guessed answer and extracts the phrase vector vm of the phrase corresponding to the guessed answer;

[0142] Step 3: Compare with the reference answer:

[0143] Extract the phrase vector vt from the reference answer (i.e. the covered part of the original sentence), and use the softmax() function to calculate the normalization value softmax(vt) of vector vt (that is, normalize vector vt to the interval [0, 1]);

[0144] Calculate the cosine similarity between vt and vm, the probability of guessing the correct answer, and the cross-entropy loss of the normalized value of vector vt;

[0145] The cosine similarity between vt and vm is: cosine_similarity(vt, vm);

[0146] Let the cross-entropy loss be Lt(poc(vm), softmax(vtt)):

[0147] Lt(poc(vm), softmax(vtt))=poc(vm)×softmax(vt);

[0148] The loss function Ltt for constructing the teacher model's fill-in-the-blank learning:

[0149]

[0150] Step 4: Repeat steps 1 to 3 to train the teacher model, and optimize W using the gradient descent algorithm. (cl) and b (cl) This continues until the loss function reaches its minimum value, thus completing the construction of the teacher model;

[0151] Here is an example of fill-in-the-blank learning:

[0152] Step 1: Create fill-in-the-blank questions:

[0153] (Sentence A) Our company requires applicants to be proficient in office software such as WPS, Excel, and Word (original sentence);

[0154] (Sentence B) Our company allows interns to use video software when required for work; our company does not allow interns to use entertainment software when meeting with clients or attending meetings with leaders.

[0155] Mark sentence A as the original sentence;

[0156] The phrase "office software" in the original sentence is masked out. For example: Our company requires job applicants to be proficient in software such as WPS, Excel, Word, etc.; where "XXX" represents the masked part.

[0157] Step 2: Do the fill-in-the-blank questions:

[0158] In sentence B, randomly select "video software" and "entertainment software" and combine them with "office software" to form options. Guess the probability of each option being correct, for example:

[0159] Option A: Office software (reference answer); Probability of correctness: 30%;

[0160] Option B: Video software; Probability of correctness: 20%;

[0161] Option C: Entertainment software; Probability of correctness: 50%;

[0162] Use "entertainment software" as the guessed answer, and use the phrase vector corresponding to "entertainment software" as the vector vmm;

[0163] It should be noted that the probability of the correct answer corresponding to the non-reference answer is random. It is not necessary for the correct answer to have the highest probability of being correct, only that the sum of the correct probabilities of all options is 100%.

[0164] Step 3: Compare with the reference answer:

[0165] The vector corresponding to the phrase "office software" is used as the vector vtt, and the normalized value softmax(vtt) of the vector vtt is calculated using the softmax() function.

[0166] Calculate the cosine similarity between vtt and vmm: cosine_similarity(vtt, vmm);

[0167] Calculate the cross-entropy loss Lt(50%, softmax(vtt)) between the correct probability of "entertainment software" (50%) and the normalized value of the vector vtt.

[0168] Calculate the loss of the teacher model: [(Lt(50%, softmax(vtt))] -1 ]×cosine_similarity(vtt, vmm));

[0169] Step 4: Repeat the training process from Step 1 to Step 3 until the teacher model is completed;

[0170] Step S123: Construct a student model based on the multi-layer BILSTM model;

[0171] Processing the embedding layer: Tok(1, 1) ~ Tok(pa, se (pa) The inputs are respectively fed into the embedding layers of the teacher model (i.e., the token embedding layer, the position embedding layer, and the segment embedding layer) and the student model; the output of the embedding layer of the teacher model is denoted as te, and the output of the embedding layer of the student model is denoted as st, and the loss function Lf of the embedding layer is constructed. (emb) :

[0172] Lf (emb) =MSE((st*We),te); where MSE represents the mean squared error; * represents matrix multiplication;

[0173] We represents the weight matrix (i.e., the learnable linear transformation matrix) of the student model's embedding layer, which is used to map the student model's embedding layer output to the teacher model's embedding layer space.

[0174] The process of obtaining We is as follows:

[0175] Obtain the embedding layer dimension d of the teacher model (te) The embedding layer dimension d of the student model (st) ;

[0176] Construct a (d (st) ×d (te) Let the matrix be denoted as We. Initialize We using Xavier or Glorot so that each element in We has a mean of zero and a standard deviation of . The normal distribution;

[0177] The gradient of the embedding layer loss of the student model with respect to the weight matrix is ​​calculated using the chain rule.

[0178] Where T denotes the transpose of the matrix; N (emb) This indicates the number of tokens input to the student model's embedding layer in a single input.

[0179] Let the learning rate be η, and let the updated weight matrix be Wee. The update matrix We is:

[0180]

[0181] The output of the student model's embedding layer is used as the input to the student model's hidden layer. The same steps are repeated for the embedding layer to construct the loss function Lf for the student model's hidden layer. (hid) :

[0182]

[0183] Where si represents the number of hidden layers in the student model, and hst (si) Wh represents the output of the si-th hidden layer of the student model. (si) The weight matrix (Wh) of the si-th hidden layer of the student model. (si) The process of obtaining dh is similar to that of obtaining We, i.e., the calculation method is the same, but the parameters are different; (st) The value of si represents the total number of hidden layers in the student model, and its range is 1 to dh. (st) ;

[0184] ti represents the number of hidden layers in the teacher model, hte (ti) This represents the output of the ti-th hidden layer of the student model;

[0185] The correspondence between si and ti is as follows:

[0186]

[0187] Where, round represents the rounding function; dh (te) The total number of hidden layers in the teacher model, denoted by ti, ranges from 1 to dh. (te) ;

[0188] In the prediction layer, the student model undergoes fill-in-the-blank learning. The steps of fill-in-the-blank learning are as follows:

[0189] Step 1: Teacher model creates fill-in-the-blank questions:

[0190] The teacher model randomly selects multiple sentences and marks one original sentence among the selected sentences for fill-in-the-blank learning; the teacher model also masks a phrase in the original sentence.

[0191] Step 2: Student models complete fill-in-the-blank questions:

[0192] The teacher model, based on the masked phrases in the original sentence, randomly selects some phrases from the selected non-original sentences to form ot options; the teacher model extracts the phrase vector (VR) of the reference answer. (te) The teacher model is based on the optimized W. (cl) and b (cl) The probability of generating the correct reference answer (pt) (true) Where ot represents the number of options in the teacher model;

[0193] The student model inputs each option into the hidden layer and outputs the hidden layer's output or... (1) or (2) ~or (ot)(Student model) uses the softmax function to convert or (1) or (2) ~or (ot) Normalized to the interval [0,1] and corrected using temperature parameters, we obtain pst. (1) pst (2) ~pst (ot) (Student model) will use pst (1) pst (2) ~pst (ot) As the probability of the correct answer for the 1st, 2nd, and so on up to the otth option;

[0194] The student model selects the option with the highest probability of being correct as the guessed answer and extracts the probability pst of the guessed answer. (st) and phrase vector vo (st) ;

[0195] It should be noted that the "temperature parameter" in this invention refers to a hyperparameter used to adjust the smoothness of the probability distribution output by the Softmax function in knowledge distillation. The temperature parameter affects the knowledge distribution learned by the student model from the teacher model by adjusting the smoothness of the Softmax output. The initial value of the temperature parameter is generally set to 2.5, and the value range of the temperature parameter is usually [2, 5]. Users or relevant technical personnel can adjust the initial value and value range of the temperature parameter according to actual needs.

[0196] Step 3: Compare with the reference answer:

[0197] Calculate VR (te) With vo (st) The cosine similarity and the probability of guessing the correct answer are compared with the cross-entropy loss of the normalized value of vector vt;

[0198] VR (te) With vo (st) The cosine similarity is: cosine_similarity(vr (te) vo (st) );

[0199] Let the cross-entropy loss be Lt(pst) (st) pt (true) ):

[0200] Lt(pst (st) pt (true) ) = pst (st) ×pt (true) ;

[0201] The loss function Lts for constructing the student model's fill-in-the-blank learning:

[0202]

[0203] Step 4: The teacher model records the probability of the student model's correct answer being true in a fill-in-the-blank learning session (ppt).

[0204] Without changing the original sentences or the reference answer, the teacher model recomposes the options from other unselected non-original sentences, and then conducts a second fill-in-the-blank training for the student model; the teacher model records the probability of the student model correctly answering the reference answer during the second fill-in-the-blank training. (true) ;

[0205] Judge PPT (true) With PTT (true) Are they equal?

[0206] If they are equal, no action is taken; if they are not equal, the PPTs are compared. (true) With PTT (true) Adjust the size and temperature parameters accordingly;

[0207] If PPT (true) >ptt (true) If the value is large, then the temperature parameter value will be increased by a factor of:

[0208] If PPT (true) <ptt (true) If the value is large, then the temperature parameter value will be reduced by a factor of:

[0209] Step 5: Repeat steps 1 to 4 to train the student model, and use the gradient descent algorithm to optimize the weight matrix of the embedding layer and hidden layer of the student model as well as the temperature parameter until the value of the loss function is minimized, thus completing the construction of the student model and obtaining the text model mod(1,1).

[0210] Step S124: Repeat the training process of the text model mod(1,1), extracting the second job posting from the first recruiter to the re-th recruiter and the rc-th job posting. (re) The job requirements for each recruitment position are used to train a text model, resulting in text model mod(1,2) to text model mod(re,rc). (re) );

[0211] Step S13: Please refer to Figure 3 Using the resume information from the (job seeker) set A(1,1) as input to the text model mod(1,1), we obtain the word vector sequence we. (1) we (2) ~we (an)Where, an represents the number of resumes in set A(1,1) (i.e., the number of job seekers applying for the first job offered by the first recruiter); the word vector sequence we (1) we (2) ~we (an) Transform into feature vector fv (1) 、fv (2) ~fv (an) ;

[0212] The word vector sequence we (1) we (2) ~we (an) These are used as inputs to the CNN model and the BILSTM model, respectively.

[0213] In a CNN model, the word vector sequence we (1) we (2) ~we (an) Perform two convolutions, two activations, and two max pooling operations to obtain (the word vector sequence we). (1) we (2) ~we (an) (of) multidimensional feature vectors cmf (1) ,cmf (2) ~cmf (an) The Flatten layer of the (CNN model) is used to process the multidimensional feature vector cmf. (1) ,cmf (2) ~cmf (an) Transforming it into a one-dimensional feature vector yields (a multi-dimensional feature vector cmf) (1) ,cmf (2) ~cmf (an) (of) one-dimensional eigenvectors cof (1) ,cof (2) ~cof (an) ;

[0214] In the BILSTM model, for the word vector sequence we (1) we (2) ~we (an) Bidirectional semantic feature extraction is performed using forward LSTM and backward LSTM to obtain (word vector sequence we). (1) we (2) ~we (an) (btf) two-dimensional feature vector (1) btf (2) ~btf (an) Using the max pooling layer of the (BILSTM model) to convert the two-dimensional feature vector into a BTF (BitTorrent Array) (1) btf (2) ~btf (an)Transforming it into a one-dimensional feature vector yields (a two-dimensional feature vector btf) (1) btf (2) ~btf (an) (of) one-dimensional eigenvector bof (1) bof (2) ~bof (an) ;

[0215] The one-dimensional feature vectors cof obtained from the CNN model and the BILSTM model (1) ,cof (2) ~cof (an) and bof (1) bof (2) ~bof (an) The features are concatenated and fused to obtain the mixed feature vector mi. (1) mi (2) ~mi (an) ;

[0216] Mix the feature vector mi (1) mi (2) ~mi (an) As input to the fully connected layer of the text model mod(1,1), the feature vector fv of set A(1,1) is obtained. (1) 、fv (2) ~fv (an) ;

[0217] Step S14: Repeat the feature vector fv (1) 、fv (2) ~fv (an) The transformation process transforms the set of job seekers A(1,2) into the set of job seekers A(re,rc). (re) The resume information in the document is converted into a feature vector.

[0218] Step S2: Obtain the resume information of full-time employees for each position from each recruiter and convert it into a standard feature vector; use the standard feature vector to perform clustering processing on the feature vectors of job seekers to obtain stable vector clusters; use a Gaussian mixture model to construct the probability density function of each stable vector cluster and screen job seekers to obtain interviewees; record the interview scores of each interviewee and recommend applicants to each recruiter.

[0219] The specific steps of step S2 are as follows:

[0220] Step S21: Obtain the first job posting from the first recruiter to the rc job posting from the re-th recruiter. (re)The resume information of each full-time employee corresponding to each job posting is used as the standard resume information; the standard resume information of each job posting from each recruiter is sequentially substituted into the text model mod(1,1) to the text model mod(re,rc). (re) In ), and repeat the feature vector fv (1) 、fv (2) ~fv (an) The transformation process yields the standard feature vector for each job opening by each recruiter;

[0221] Based on the first job posting provided by the first recruiter up to the rc job posting provided by the re-th recruiter. (re) The standard feature vector of each job posting is used to represent the set of job seekers A(1, 1) ~ A(re, rc). (re) The corresponding feature vectors are clustered, and job seekers are screened using a Gaussian mixture model.

[0222] Step S211: Denote the standard feature vector of the first job posting provided by the first recruiter as vec. (1) vec (2) ~vec (reg) Where reg represents the number of full-time employees for the first job opening by the first recruiter;

[0223] Extract the feature vector fv corresponding to the set of (job seekers) A(1,1). (1) 、fv (2) ~fv (an) Screening job seekers for the first job opening offered by the first recruiter;

[0224] With vec (1) Using the center vector, a clustering algorithm is used to cluster fv (1) 、fv (2) ~fv (an) Perform clustering to obtain stable vector clusters; (if using vec (1) If a stable vector cluster cannot be obtained, then use the other standard feature vectors for the first job posting provided by the first recruiter, i.e., vec. (2) ~vec (reg) , in vec (2) ~vec (reg) (Find the first standard eigenvector that can yield a stable vector cluster as the center vector for clustering).

[0225] The number of eigenvectors sta within a stable vector cluster is calculated; where sta ≤ an.

[0226] Let the eigenvectors in the stable vector cluster be denoted as vet. (1) vet (2) ~vet (sta);

[0227] Let the eigenvector vet (1) The parameter in is: ev (1，1) ev (1，2) ;

[0228] eigenvector vet (2) The parameter in is: ev (2，1) ev (2，2) ;

[0229] And so on, the feature vector vet (sta) The parameter in is: ev (sta，1) ev (sta，2) ;

[0230] Construct a zero matrix of size (sta×2) and fill it with ev. (1，1) ~ev (sta，2) Resulting in matrix B (1) :

[0231]

[0232] Wherein, matrix B (1) The first row in the table represents the eigenvector vet. (1) The second row represents the eigenvector vet. (2) And so on, the sta-th row represents the eigenvector vet. (sta) ;

[0233] Calculate matrix B (1) The centralized matrix B (2) Covariance matrix B (3) eigenvectors vet (1) ~vet (sta) The mean vector avet;

[0234] Let vet be the j-th eigenvector in the first stable vector cluster. (j) The value range of j is 1 to sta;

[0235] vet construction (j) probability density function p(vet) (j) ):

[0236]

[0237] in, Representation matrix B (3) The square root of matrix B (3) Treat it as a determinant and calculate it using a recursive algorithm (such as Laplace expansion);

[0238] exp represents the exponential function, * represents matrix multiplication, T represents the transpose of a matrix, and -1 represents the inverse of a matrix;

[0239] eigenvector vet (1) ~vet (sta) Substitute into function p(vet) (j) In ), the probability density p(vet) is obtained. (1) )~p(vet (sta) );

[0240] Calculate the eigenvector vet (1) ~vet (sta) Standard deviation vector sdv:

[0241] Where, diag represents the diagonal element of the matrix;

[0242] Substitute avet and sdv into the probability density function p(vet) (j) In the above, calculate the probability density of vectors avet and sdv to obtain p(avet) and p(sdv);

[0243] In p(vet) (1) )~p(vet (sta) In the given information, the feature vectors with probability density in the interval [p(avet), p(sdv)] are extracted as the target vectors.

[0244] The job seeker corresponding to the target vector will be selected as the interviewee for the first job offered by the first recruiter.

[0245] Step S212: Repeat the same steps for screening job seekers for the first job posting provided by the first recruiter, and for the second job posting provided by the first recruiter up to the rc job posting provided by the re recruiter. (re) Each job posting screens job applicants;

[0246] Step S23: Record the interview scores of each interviewee for each position, and obtain a fixed percentage of the interviewees with the highest interview scores for each position as recommended candidates for each position.

[0247] It should be noted that the fixed ratio in "step S23" of the present invention is generally 20%, and users or relevant technical personnel can adjust the value of the fixed ratio according to actual needs.

[0248] Step S3: Record the interview scores of each job seeker in real time, and obtain new job seeker resume information (on recruitment websites or recruitment applications) to recommend candidates to various recruiters.

[0249] The above formulas are all dimensionless calculations. The formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation. For example, there are weighting coefficients and proportional coefficients. The values ​​set are to quantify each parameter to obtain a specific value, which is convenient for subsequent comparison. The values ​​of the weighting coefficients and proportional coefficients are only required to not affect the proportional relationship between the parameters and the quantified values.

[0250] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A talent recommendation method based on deep learning, characterized in that, The method includes: The process involves obtaining the number of job seekers and their resumes; the number of employers and the number of openings and requirements for each employer; classifying job seekers based on their applied positions and the openings of each employer; training a text model based on a knowledge distillation mechanism and the requirements of each employer; and using the trained text model to convert each job seeker's resume information into a feature vector. The specific steps for a single partition are as follows: Extract the number of job postings from the 1st to the retrieval position. （1） ~rc （re） ; Extract the job positions applied for by each job seeker; based on the job positions applied for by each job seeker and the job positions offered by each recruiter, divide the job seekers into sets A(1,1) to A(re,rc). （re） ); Extract the job requirements of the first job posting from the first recruiter and use it as text information tex(1,1); remove punctuation marks and text information that are not related to the job requirements from the text information tex(1,1); train a text model mod(1,1) based on the text information tex(1,1); and use this model mod(1,1) to transform the resume information in set A(1,1) into a feature vector. Extract the second job posting from the first recruiter to the rc position of the re-th recruiter. （re） The job requirements for each recruitment position are used to train a text model, resulting in text model mod(1,2) ~ text model mod(re,rc). （re） ), and set A(1,2)~A(re,rc) （re） The resume information in the document is converted into a feature vector; The training steps for the text model mod(1,1) are as follows: Perform paragraph and sentence segmentation on the text information tex(1,1) to obtain the number of paragraphs pa and the number of sentences se in each paragraph. （1） ~se （pa） ; Generate a token for each sentence in the text information tex(1,1), resulting in tok(1,1) ~ tok(pa,se) （pa） ); Based on the BERT model, construct a teacher model: Set tok(1,1) to tok(pa,se) （pa） These are fed into the token embedding layer, segment embedding layer, and position embedding layer of the BERT model, respectively, to obtain the vector code, position code, and sentence code for each sentence; Add the vector code, position code, and sentence code corresponding to each sentence to obtain the vector ve(1,1)~ve(pa,se). （pa） ); Let vectors ve(1,1) ~ve(pa,se) be used. （pa） As the input to the two-layer Transformer encoder of the BERT model, it learns the relationship between each phrase corresponding to each sentence in the text information tex(1,1) and performs fill-in-the-blank learning on the BERT model; A student model is constructed based on a multi-layer BILSTM model. The process involves acquiring the resumes of full-time employees for each position from each recruiter and converting them into standard feature vectors; using these standard feature vectors to cluster the job seekers' feature vectors to obtain stable vector clusters; constructing the probability density function for each stable vector cluster using a Gaussian mixture model and filtering job seekers to obtain interviewees; recording the interview scores of each interviewee and recommending applicants to each recruiter. It records the interview scores of each job seeker in real time, obtains new job seeker resume information, and recommends candidates to various recruiters.

2. The talent recommendation method based on deep learning according to claim 1, characterized in that, The specific steps for performing fill-in-the-blank learning on the BERT model are as follows: Step 1: Create fill-in-the-blank questions: The BERT model randomly selects multiple sentences and labels one original sentence among the selected sentences for fill-in-the-blank learning; the BERT model also masks a phrase in the original sentence. Step 2: Do the fill-in-the-blank questions: The BERT model uses the masked phrases in the original sentence to randomly select some phrases from the selected non-original sentences to form options, and guesses the probability of each option being correct. The vector corresponding to the original sentence after being masked is vh （mask） The probability of guessing the correct answer is poc(vm): ； Among them, W （cl） and b （cl） These represent the weights and biases of the classification layer in the BERT model, respectively. The BERT model selects the option with the highest probability of being correct as the guessed answer and extracts the phrase vector vm of the phrase corresponding to the guessed answer; Step 3: Compare with the reference answer: Extract the phrase vector vt from the reference answer, and use the softmax() function to calculate the normalized value softmax(vt) of the vector vt. Calculate the cosine similarity between vt and vm, the probability of guessing the correct answer, and the cross-entropy loss of the normalized value of vector vt; The cosine similarity between vt and vm is: cosine_similarity(vt, vm). Let the cross-entropy loss be Lt(poc(vm), softmax(vtt)): Lt(poc(vm),softmax(vtt))=poc(vm)×softmax(vt); The loss function Ltt for constructing the teacher model's fill-in-the-blank learning: ； Step 4: Repeat steps 1 to 3 to train the teacher model, and optimize W using the gradient descent algorithm. （cl） and b （cl） This continues until the loss function reaches its minimum value, thus completing the construction of the teacher model.

3. The talent recommendation method based on deep learning according to claim 1, characterized in that, The specific steps for constructing the student model are as follows: Processing the embedding layer: Connect tok(1,1) to tok(pa,se) （pa） The inputs are respectively fed into the embedding layers of the teacher model and the student model; the output of the embedding layer of the teacher model is denoted as te, and the output of the embedding layer of the student model is denoted as st, and the loss function Lf of the embedding layer is constructed. （emb） : Where MSE represents mean squared error; We represents the weight matrix of the student model embedding layer; The output of the student model's embedding layer is used as the input to the student model's hidden layer. The same steps are repeated for the embedding layer to construct the loss function Lf for the student model's hidden layer. （hid） : ； Where si represents the number of hidden layers in the student model, and hst （si） Wh represents the output of the si-th hidden layer of the student model. （si） dh represents the weight matrix of the si-th hidden layer of the student model. （st） This represents the total number of hidden layers in the student model; ti represents the number of hidden layers in the teacher model, hte （ti） This represents the output of the ti-th hidden layer of the student model; The correspondence between si and ti is as follows: ； Where round represents the floor function; dh （te） This represents the total number of hidden layers in the teacher model; In the prediction layer, the student model learns to fill in the blanks.

4. The talent recommendation method based on deep learning according to claim 3, characterized in that, The process of obtaining We is as follows: Obtain the embedding layer dimension d of the teacher model （te） The embedding layer dimension d of the student model （st） ; Construct a (d （st） ×d （te） Let the matrix be We. Initialize matrix We using Xavier or Glorot so that each element in matrix We has a mean of zero and a standard deviation of . The normal distribution; The gradient of the embedding layer loss of the student model with respect to the weight matrix is ​​calculated using the chain rule. : Where T represents the transpose of the matrix; N （emb） This indicates the number of tokens input to the student model's embedding layer in a single input. Let the learning rate be η, and let the updated weight matrix be Wee. The update matrix We is: 。 5. The talent recommendation method based on deep learning according to claim 3, characterized in that, The steps for fill-in-the-blank learning on the student model are as follows: Step 1: Teacher model creates fill-in-the-blank questions: The teacher model randomly selects multiple sentences and marks one original sentence among the selected sentences for fill-in-the-blank learning; the teacher model also masks a phrase in the original sentence. Step 2: Student models complete fill-in-the-blank questions: The teacher model, based on the masked phrases in the original sentence, randomly selects some phrases from the selected non-original sentences to form ot options; the teacher model extracts the phrase vector (VR) of the reference answer. （te） The probability pt of the teacher model generating the correct reference answer （true） ; The student model inputs each option into the hidden layer and outputs the hidden layer's output or... （1） ~or （ot） Use the softmax function to convert or （1） ~or （ot） Normalize to the interval [0,1] and correct using the temperature parameter to obtain the correct probability pst of the 1st to otth options. （1） pst （2） ~pst （ot） ; The student model selects the option with the highest probability of being correct as the guessed answer and extracts the probability pst of the guessed answer. （st） and phrase vector vo （st） ; Step 3: Compare with the reference answer: Calculate VR （te） With vo （st） The cosine similarity and the probability of guessing the correct answer are compared with the cross-entropy loss of the normalized value of vector vt; VR （te） With vo （st） The cosine similarity is: cosine_similarity(vr （te） vo （st） ); Let the cross-entropy loss be Lt(pst). （st） pt （true） ): Lt(pst （st） pt （true） )=pst （st） ×pt （true） ; The loss function Lts for constructing the student model's fill-in-the-blank learning: 。 6. The talent recommendation method based on deep learning according to claim 5, characterized in that, The steps for fill-in-the-blank learning for student models also include: Step 4: The teacher model records the probability of the student model's correct answer in a fill-in-the-blank exercise (PPT). （true） ; Without changing the original sentences or the reference answer, the teacher model recomposes the options from other unselected non-original sentences, and then conducts a second fill-in-the-blank training for the student model; the teacher model records the probability of the student model correctly answering the reference answer during the second fill-in-the-blank training. （true） ; Judge PPT （true） With PTT （true） Are they equal? If they are equal, no action is taken; if they are not equal, the PPTs are compared. （true） With PTT （true） Adjust the size and temperature parameters accordingly; If PPT （true） >ptt （true） If the value is large, then the temperature parameter value will be increased by a factor of: ; If PPT （true） <ptt （true） If the value is large, then the temperature parameter value will be reduced by a factor of: ; Step 5: Repeat steps 1 to 4 to train the student model, and use the gradient descent algorithm to optimize the weight matrix of the embedded layer and hidden layer of the student model as well as the temperature parameter until the value of the loss function is minimized, thus completing the construction of the student model and obtaining the text model mod(1,1).

7. The talent recommendation method based on deep learning according to claim 1, characterized in that, The specific steps to transform the resume information in set A(1,1) into feature vectors are as follows: Using the resume information in set A(1,1) as input to the text model mod(1,1), we obtain the word vector sequence we. （1） ~we （an） Where, an represents the number of resumes in set A(1,1); The word vector sequence we （1） ~we （an） These are used as inputs to the CNN model and the BILSTM model, respectively. In a CNN model, the word vector sequence we （1） we （2） ~we （an） Two convolutions, two activations, and two max pooling operations are performed to obtain a multi-dimensional feature vector; a Flatten layer is then used to transform the multi-dimensional feature vector into a one-dimensional feature vector, yielding the cof(x) value. （1） ,cof （2） ~cof （an） ; In the BILSTM model, for the word vector sequence we （1） ~we （an） Bidirectional semantic feature extraction is performed using forward LSTM and backward LSTM to obtain a two-dimensional feature vector; then, a max pooling layer is used to transform the two-dimensional feature vector into a one-dimensional feature vector, resulting in the bof. （1） bof （2） ~bof （an） ; The one-dimensional feature vectors cof obtained from the CNN model and the BILSTM model （1） ~cof （an） and bof （1） ~bof （an） The features are concatenated and fused to obtain the mixed feature vector mi. （1） ~mi （an） ; Mix the feature vector mi （1） ~mi （an） As input to the fully connected layer of the text model mod(1,1), the feature vector fv of set A(1,1) is obtained. （1） ~fv （an） ; Based on feature vector fv （1） ~fv （an） Screen job seekers for the first job opening offered by the first recruiter.

8. The talent recommendation method based on deep learning according to claim 7, characterized in that, The specific steps for screening job seekers are as follows: Let vec be the standard feature vector of the first job posting provided by the first employer. （1） ; with vec （1） Using the center vector, a clustering algorithm is used to cluster fv （1） ~fv （an） Perform clustering to obtain stable vector clusters; count the number sta of eigenvectors within each stable vector cluster. Let the eigenvectors in the stable vector cluster be denoted as vet. （1） ~vet （sta） ; Let the eigenvector vet （1） The parameter in is: ev （1，1） ev （1，2） ; Similarly, the eigenvector vet （sta） The parameter in is: ev （sta，1） ev （sta，2） ; Construct matrix B （1） Calculate the mean vector avet and the centered matrix B. （2） Covariance matrix B （3） ; Constructing the probability density function Calculate the eigenvector vet （1） ~vet （sta） The probability density p(vet) （1） )~p(vet （sta） ); Calculate the standard deviation vector sdv, calculate the probability density of vectors avet and sdv, and obtain p(avet) and p(sdv). In p(vet) （1） )~p(vet （sta） In the given information, extract the feature vectors whose probability density lies in the interval [p(avet), p(sdv)] as the target vector; The job seeker corresponding to the target vector will be the interviewee for the first job offered by the first recruiter.

Images