People and post matching method based on deep learning

Abstract

The invention discloses a person and post matching method based on deep learning, which comprises the following steps: firstly, constructing a post set and a staff set, carrying out text description, and calculating text matching scores of posts and staff by adopting a text matching model; meanwhile, constructing a relation graph of the posts and the employees, and calculating relation matching scores of the posts and the employees based on a relation matching model of a relation graph convolutional network; the text matching model and the relation matching model are jointly trained through a co-teaching mechanism, then the trained text matching model and relation matching model are adopted to predict and obtain a text matching score and a relation matching score, and finally the text matching score and the relation matching score are subjected to weighted calculation to obtain a final post and employee matching score. According to the method, deep text semantic analysis and the man-sentry relation graph network are combined, training optimization is carried out through a co-teaching mechanism, the matching accuracy and robustness are remarkably improved, and the method is particularly suitable for an enterprise environment with sparse data and more noise.

Description

Technical Field

[0001] The present invention relates to the technical field of computer data processing, and in particular to a person-job matching method based on deep learning. Background Art

[0002] At present, in traditional enterprises, the employee-job matching method often uses keyword matching or simple text similarity calculation to judge the suitability of employees and jobs. This method often ignores the deep semantic information between job descriptions and employee skills, and cannot accurately capture the degree of match between employees' actual abilities and job requirements, resulting in unsatisfactory matching results.

[0003] As the demand for accurate employee competency assessment and job suitability in enterprise human resource management increases, the limitations of traditional methods become more obvious. In particular, when there is sparse data within the enterprise, complex and changing job requirements, and inconsistent employee skill descriptions, existing job matching methods are difficult to handle these problems, resulting in low matching efficiency and accuracy. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a person-job matching method based on deep learning, which combines deep text semantic analysis and a person-job relationship graph network, and performs training optimization through a co-teaching mechanism, thereby significantly improving the accuracy and robustness of the matching, and is particularly suitable for enterprise environments with sparse data and high noise.

[0005] The technical solution of the present invention is:

[0006] A person-job matching method based on deep learning specifically includes the following steps:

[0007] (1) Construct a set of positions and a set of employees. Each position in the position set and each employee in the employee set corresponds to a text description. Construct a text matching model based on the BERT model and the dual attention mechanism. Use the text matching model to calculate the text matching scores of positions and employees.

[0008] (2) Construct a relationship graph between positions and employees, and calculate the relationship matching score between positions and employees based on the relationship matching model of the relationship graph convolutional network;

[0009] (3) The text matching model and the relationship matching model are jointly trained through a co-teaching mechanism to obtain a trained text matching model and a relationship matching model;

[0010] (4) Use the trained text matching model and relationship matching model to predict the text matching score and relationship matching score, and then perform weighted calculation on the text matching score and relationship matching score to obtain the final position and employee matching score.

[0011] The specific steps of constructing a set of positions and a set of employees and using a text matching model to calculate text matching scores of positions and employees are as follows:

[0012] S11. Construct a set of positions J = {j1, j2, ..., j i ,…,j n}、Employee set R=

[0013] {r1,r2,…,r k ,…,r m}, each position j i Corresponding to a job description X ji , each employee k Corresponding to an employee resume Job Description and employee resumes All are vectors of text descriptions;

[0014] S12. Use BERT model to describe the job and employee resumes Encoding is performed to generate the semantic representation of the job text and the semantic representation of the employee text, as shown in the following formula (1):

[0015]

[0016] In formula (1), The semantic representation matrix representing the job text, It represents the semantic representation matrix of employee text, where d is the encoding dimension and L is the text length;

[0017] S13, adopts a dual attention mechanism consisting of affinity attention mechanism and difference attention mechanism to capture the similarities and differences in text;

[0018] The affinity attention mechanism is used to capture the similarity information between job text and employee text. For each job j i and employees k , the similarity of the text is calculated by attention weight, as shown in the following formula (2):

[0019]

[0020] In formula (2), The affinity attention matrix representing the similarity between the job text and the employee text, T represents the transpose of the matrix, and softmax represents the softmax activation function;

[0021] The difference attention mechanism captures the differences between positions and employees through vector subtraction, as shown in the following formula (3):

[0022]

[0023] In formula (3), The difference attention matrix represents the difference between the job text and the employee text, and softmax represents the softmax activation function;

[0024] Then the affinity attention matrix and the difference attention matrix are weighted fused, as shown in the following formula (4):

[0025] A final =λ×A aff +(1-λ)×A diff (4);

[0026] In formula (4), λ∈[0,1] represents the hyperparameter for adjusting similarity and difference; A final represents the dual attention matrix;

[0027] S14. Calculate the matching score between the position and the employee through the semantic representation after text matching. The specific calculation process is shown in the following formula (5):

[0028]

[0029] In formula (5), s text (j i ,r k ) indicates position j i and employees k The matching score represents the semantic similarity between the position and the employee.

[0030] The specific steps of constructing the relationship graph between positions and employees and calculating the relationship matching scores between positions and employees based on the relationship matching model of the relationship graph convolution network are as follows:

[0031] S21. Construct a relationship graph between positions and employees as G = (V, E), where V = J ∪ R is the combination of nodes in the relationship graph, and E is the set of edges representing the relationship between positions and employees;

[0032] S22, using the relationship graph convolution network R-GCN to represent each node v∈V in the relationship graph It is updated through the aggregation of messages from neighboring nodes, as shown in the following formula (6):

[0033]

[0034] In formula (6), N r (v) is the neighbor set of node v under relationship type r, W r is the parameter matrix of relation type r, is the representation of neighbor nodes, W0 represents the weight matrix of node v’s own features, and σ is the activation function;

[0035] After each node is updated through aggregation, the node representation of the updated position or employee is obtained. When node v represents a job node, That is When node v represents an employee node, That is

[0036]

[0037] S23. Calculate the relationship matching score between positions and employees through the relationship representation learned by the graph neural network, as shown in the following formula (7):

[0038]

[0039] In formula (7), s relation (j i ,r k ) indicates position j i and employees k The relationship matching score, is the node representation of any updated position node, yes The transpose of It is the node representation of any employee node after the update.

[0040] The specific steps of jointly training the text matching model and the relationship matching model through the co-teaching mechanism are as follows: the first batch of training data of the text matching model is first predicted by the relationship matching model, and the first batch of high-quality samples are screened out; the first batch of training data of the relationship matching model is first predicted by the text matching model, and the first batch of high-quality samples are screened out; the text matching model and the relationship matching model are respectively trained based on the first batch of high-quality samples after screening, and then multiple batches of cross-training are repeated, so that the text matching model and the relationship matching model share parameters with each other, and the updated results are finally obtained, and the trained text matching model and relationship matching model are finally obtained.

[0041] The text matching score and the relationship matching score are weighted and calculated, as shown in the following formula (8):

[0042] s final (j i ,r k )=α×s text (j i ,r k )+(1-α)×s relation (j i ,r k) (8);

[0043] In formula (8), α is a hyperparameter used to control the weights of text matching score and relationship matching score.

[0044] Advantages of the present invention:

[0045] The text matching model of the present invention combines the BERT model and the dual attention mechanism. The BERT model can generate deep semantic representations, but the semantic complexity of job and employee descriptions, that is, the fine-grained differences that may be contained in the text, are difficult to evaluate through simple vector similarity, and the correlation between employees and jobs cannot be fully considered based on text semantics. Therefore, the dual attention mechanism is introduced to capture the similarities and differences in the text.

[0046] The relationship matching model of the present invention performs neural network learning based on the constructed position-employee relationship graph, thereby constructing the implicit association between employees and positions according to the skill keywords in the position description and the employee resume in the employee resume, capturing the skill association between positions and employees and the similarity between positions.

[0047] The present invention performs joint training of the two models through a co-teaching mechanism, so that the two models can verify each other during the training process, ensuring that the two models can still maintain efficient training effects when the data is sparse and noisy.

[0048] The present invention solves the shortcomings of traditional methods in semantic capture and data noise processing by introducing a dual attention mechanism and a relational graph convolutional network, significantly improving the accuracy and robustness of employee-position matching within an enterprise. It can not only effectively process the explicit semantic information between employees and positions, but also mine implicit associations through relational graphs, providing more comprehensive decision support for the enterprise's human resource management. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] See Figure 1 , a person-job matching method based on deep learning, specifically including the following steps:

[0052] (1) Construct a job set and an employee set. Each job in the job set and each employee in the employee set corresponds to a text description. Construct a text matching model based on the BERT model and the dual attention mechanism. Use the text matching model to calculate the text matching scores of the job and the employee. The specific steps are as follows:

[0053] S11. Construct a set of positions J = {j1, j2, ..., j i ,…,j n}、Employee set R=

[0054] {r1,r2,…,r k ,…,r m}, each position j i Corresponding to a job description X ji , each employee k Corresponding to an employee resume Job Description and employee resumes All are vectors of text descriptions;

[0055] S12. Use BERT model to describe the job and employee resumes Encoding is performed to generate the semantic representation of the job text and the semantic representation of the employee text, as shown in the following formula (1):

[0056]

[0057] In formula (1), The semantic representation matrix representing the job text, It represents the semantic representation matrix of employee text, where d is the encoding dimension and L is the text length;

[0058] S13, adopts a dual attention mechanism consisting of affinity attention mechanism and difference attention mechanism to capture the similarities and differences in text;

[0059] The affinity attention mechanism is used to capture the similarity information between job text and employee text. For each job j i and employees k , the similarity of the text is calculated by attention weight, as shown in the following formula (2):

[0060]

[0061] In formula (2), The affinity attention matrix representing the similarity between the job text and the employee text, T represents the transpose of the matrix, and softmax represents the softmax activation function;

[0062] The difference attention mechanism captures the differences between positions and employees through vector subtraction, as shown in the following formula (3):

[0063]

[0064] In formula (3), The difference attention matrix represents the difference between the job text and the employee text, and softmax represents the softmax activation function;

[0065] Then the affinity attention matrix and the difference attention matrix are weighted fused, as shown in the following formula (4):

[0066] A final =λ×A aff +(1-λ)×A diff (4);

[0067] In formula (4), λ∈[0,1] represents the hyperparameter for adjusting similarity and difference; A final represents the dual attention matrix;

[0068] S14. Calculate the matching score between the position and the employee through the semantic representation after text matching. The specific calculation process is shown in the following formula (5):

[0069]

[0070] In formula (5), s text (j i ,r k ) indicates position j i and employees k The matching score represents the semantic similarity between the position and the employee;

[0071] (2) Construct a relationship graph between positions and employees, and calculate the relationship matching score between positions and employees based on the relationship matching model of the relationship graph convolutional network. The specific steps are as follows:

[0072] S21. Construct a relationship graph between positions and employees as G = (V, E), where V = J ∪ R is the combination of nodes in the relationship graph, and E is the set of edges representing the relationship between positions and employees;

[0073] S22, using the relationship graph convolution network R-GCN to represent each node v∈V in the relationship graph It is updated through the aggregation of messages from neighboring nodes, as shown in the following formula (6):

[0074]

[0075] In formula (6), N r (v) is the neighbor set of node v under relationship type r, Wr is the parameter matrix of relation type r, is the representation of neighbor nodes, W0 represents the weight matrix of node v’s own features, and σ is the activation function;

[0076] After each node is updated through aggregation, the node representation of the updated position or employee is obtained. When node v represents a job node, That is When node v represents an employee node, That is

[0077]

[0078] S23. Calculate the relationship matching score between positions and employees through the relationship representation learned by the graph neural network, as shown in the following formula (7):

[0079]

[0080] In formula (7), s relation (j i ,r k ) indicates position j i and employees k The relationship matching score, is the node representation of any updated position node, yes The transpose of is the node representation of any employee node after the update;

[0081] (3) The text matching model and the relationship matching model are jointly trained through a co-teaching mechanism, that is, the first batch of training data of the text matching model is first predicted by the relationship matching model to screen out the first batch of high-quality samples, and the first batch of training data of the relationship matching model is first predicted by the text matching model to screen out the first batch of high-quality samples. The text matching model and the relationship matching model are trained based on the first batch of high-quality samples after screening, and then multiple batches of cross-training are repeated, so that the text matching model and the relationship matching model share parameters with each other, and the updated results are finally obtained. The trained text matching model and relationship matching model are finally obtained;

[0082] (4) The trained text matching model and relationship matching model are used to predict the text matching score and relationship matching score, and then the text matching score and relationship matching score are weighted to obtain the final position and employee matching score, as shown in the following formula (8):

[0083] s final (j i ,r k )=α×s text(j i ,r k )+(1-α)×s relation (j i ,r k ) (8);

[0084] In formula (8), α is a hyperparameter used to control the weights of text matching score and relationship matching score.

[0085] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A person-job matching method based on deep learning, characterized in that: The specific steps include: (1) Construct a set of positions and a set of employees. Each position in the position set and each employee in the employee set corresponds to a text description. Construct a text matching model based on the BERT model and the dual attention mechanism. Use the text matching model to calculate the text matching scores of positions and employees. (2) Construct a relationship graph between positions and employees, and calculate the relationship matching score between positions and employees based on the relationship matching model of the relationship graph convolutional network; (3) The text matching model and the relationship matching model are jointly trained through a co-teaching mechanism to obtain a trained text matching model and a relationship matching model; (4) Use the trained text matching model and relationship matching model to predict the text matching score and relationship matching score, and then perform weighted calculation on the text matching score and relationship matching score to obtain the final position and employee matching score.

2. The method for matching people with jobs based on deep learning according to claim 1, characterized in that: The specific steps of constructing a set of positions and a set of employees and using a text matching model to calculate text matching scores of positions and employees are as follows: S11. Construct a job set J = {j1, j2, ..., j i , ..., j n }、Employee set R={r1,r2,...,r k , ..., r m }, each position ji corresponds to a position description Each employee k Corresponding to an employee resume Job Description and employee resumes All are vectors of text descriptions; S12. Use BERT model to describe the job and employee resumes Encoding is performed to generate the semantic representation of the job text and the semantic representation of the employee text, as shown in the following formula (1): In formula (1), The semantic representation matrix representing the job text, It represents the semantic representation matrix of employee text, where d is the encoding dimension and L is the text length; S13, adopts a dual attention mechanism consisting of affinity attention mechanism and difference attention mechanism to capture the similarities and differences in text; The affinity attention mechanism is used to capture the similarity information between job text and employee text. For each job j i and employees k , the similarity of the text is calculated by attention weight, as shown in the following formula (2): In formula (2), The affinity attention matrix representing the similarity between the job text and the employee text, T represents the transpose of the matrix, and softmax represents the softmax activation function; The difference attention mechanism captures the differences between positions and employees through vector subtraction, as shown in the following formula (3): In formula (3), The difference attention matrix represents the difference between the job text and the employee text, and softmax represents the softmax activation function; Then the affinity attention matrix and the difference attention matrix are weighted fused, as shown in the following formula (4): A final =λ×A aff +(1-λ)×A diff (4); In formula (4), λ∈[0,1] represents the hyperparameter for adjusting similarity and difference; A final represents the dual attention matrix; S14. Calculate the matching score between the position and the employee through the semantic representation after text matching. The specific calculation process is shown in the following formula (5): In formula (5), s text (j i , r k ) indicates position j i and employees k The matching score represents the semantic similarity between the position and the employee.

3. The method for matching people with jobs based on deep learning according to claim 2, characterized in that: The specific steps of constructing the relationship graph between positions and employees and calculating the relationship matching scores between positions and employees based on the relationship matching model of the relationship graph convolution network are as follows: S21. Construct a relationship graph between positions and employees as G=(V, E), where V=J∪R is the combination of nodes in the relationship graph, and E is the set of edges representing the relationship between positions and employees; S22, using the relationship graph convolution network R-GCN to represent each node v∈V in the relationship graph It is updated through the aggregation of messages from neighboring nodes, as shown in the following formula (6): In formula (6), N r (v) is the neighbor set of node v under relationship type r, W r is the parameter matrix of relation type r, is the representation of neighbor nodes, W0 represents the weight matrix of node v’s own features, and σ is the activation function; After each node is updated through aggregation, the node representation of the updated position or employee is obtained. When node v represents a job node, That is When node v represents an employee node, That is S23. Calculate the relationship matching score between positions and employees through the relationship representation learned by the graph neural network, as shown in the following formula (7): In formula (7), s relation (j I , r k ) indicates position j i and employees k The relationship matching score, is the node representation of any updated position node, yes The transpose of It is the node representation of any employee node after the update.

4. The method for matching people with jobs based on deep learning according to claim 3, characterized in that: The specific steps of jointly training the text matching model and the relationship matching model through the co-teaching mechanism are as follows: the first batch of training data of the text matching model is first predicted by the relationship matching model, and the first batch of high-quality samples are screened out; the first batch of training data of the relationship matching model is first predicted by the text matching model, and the first batch of high-quality samples are screened out; the text matching model and the relationship matching model are respectively trained based on the first batch of high-quality samples after screening, and then multiple batches of cross-training are repeated, so that the text matching model and the relationship matching model share parameters with each other, and the updated results are finally obtained, and the trained text matching model and relationship matching model are finally obtained.

5. The method for matching people with jobs based on deep learning according to claim 4, characterized in that: The text matching score and the relationship matching score are weighted and calculated, as shown in the following formula (8): s final (j i ,r k )=α×s text (j i ,r k )+(1-α)×s relation (j i ,r k ) (8); In formula (8), α is a hyperparameter used to control the weights of text matching score and relationship matching score.

Images