Enterprise post recruitment and resume matching method, system and device and medium

Abstract

The invention discloses an enterprise post recruitment and resume matching method, system and device and a medium, and relates to the technical field of personnel and post matching, and the method comprises the steps: carrying out the modeling of a job seeker resume and enterprise recruitment information through a BERT model, and generating the local semantic representation of the post recruitment and resume; the method comprises the following steps: constructing a graph structure according to a historical successful recruitment record, inputting a constructed resume graph into a graph neural network, introducing an attention mechanism to identify importance of different skills in position requirements and resume experience, obtaining global semantic representation of a resume, fusing local semantics and global semantics of the resume through a KAN network, and obtaining a new resume. Inputting the final semantic representation of the resume and the representation of the position recruitment demand into a person and post matching interaction layer to automatically judge a matching result of the resume and enterprise recruitment; according to the method, resume information of job seekers is deeply mined by using double-view-angle features, the ability of capturing dynamic changes by a historical recruitment data improvement model is introduced, and the matching of position requirements and potential abilities of candidates is further enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of person-job matching, and particularly relates to a method, a system, a device and a medium for matching enterprise job recruitment and resumes. Background Art

[0002] Online recruitment services are rapidly changing the traditional recruitment methods in the employment market. Accurate person-job matching is crucial for intelligent recruitment. In recent years, the problem of person-job matching, especially the matching of recruitment information and resume information, has attracted extensive attention from researchers.

[0003] In this research field, some researchers have drawn inspiration from recommendation systems and studied the person-job matching task, that is, the matching of job positions and resumes. Some scholars have adopted the method of manually extracting features to extract features from job seeker resumes and company job descriptions, and designed a person-job recommendation system to promote the matching of candidates and job positions; some scholars have proposed a job recommendation algorithm based on job seeker preferences and job requirements to match job requirements and resume information; some scholars have used machine learning methods to explore the hierarchical relationship between recruitment texts and resume texts to achieve personalized resume recommendations. Early research regarded the person-job matching task as a recommendation problem, but recent research has explored the person-job matching problem from another perspective. Currently, some scholars have proposed a job-resume matching modeling method based on convolutional neural network (CNN), aiming to match the right person to the right job position. This method is called the person-job fit neural network PJFNN (Person-Job Fit Neural Network), which uses parallel neural networks (such as job neural network and resume neural network) to perform semantic representation on job position and resume texts, and measures the matching degree between the job position and the job seeker by calculating semantic similarity. On this basis, some scholars have achieved better results by using BiLSTM and attention mechanism to replace CNN, and proposed a competence-aware person-job matching method. First, BiLSTM is used to encode the word-level representations of job seeker resumes and job requirements, and then a hierarchical attention mechanism is adopted to further capture the different importance of skills (work experience) in job seeker resumes and job requirements. Some scholars have also proposed an effective and interpretable person-job matching method. First, CNN and co-attention mechanism are used to encode resumes and job requirements, and then the intention scores of job seekers and recruiters are calculated through a bilateral intention network; some scholars have designed a twin SBERT model by fine-tuning, and used job resume pairs to construct text embeddings of job requirements and resumes; some scholars have proposed a person-job matching model that explores the internal and external interactions of multiple attributes of job descriptions and resumes. First, BERT is used to project the keys and values of each attribute of job descriptions and resumes and the corresponding attribute sources into the same semantic space, and then the key embeddings and source embeddings are gradually fused into the value embeddings, and finally a matching layer is used to predict the matching degree.

[0004] However, as job seekers accumulate more experience, the important features in their resumes may gradually change. Most of the existing technologies rely on static features for matching, without considering the persistence of feature learning. They analyze from a single perspective and do not explore job seekers' resume information from multiple perspectives, making it difficult to comprehensively capture job seekers' features and resulting in inaccurate matching results. Summary of the Invention

[0005] Aiming at the problem that the existing technology fails to explore job seekers' resume information from multiple perspectives, making it difficult to comprehensively capture job seekers' features and resulting in a low accuracy rate of matching results, the present invention proposes a matching method, system, device and medium for enterprise job recruitment and resumes, which solves the problems existing in the existing technology through local semantic representation of job resumes based on BERT, semantic representation of resumes based on graph networks, and person-job matching interaction.

[0006] A matching method for enterprise job recruitment and resumes includes the following steps:

[0007] Obtain the resume text data of the current applicant, the enterprise job recruitment requirement data, and the resume data in the enterprise's historical successful recruitment records;

[0008] By inputting the enterprise job recruitment requirement data and the resume text data of the current applicant into the BERT model for sentence embedding, obtain the local semantics of the enterprise job recruitment requirements containing multiple relationships and the local semantics of the resume text of the current applicant; construct a resume graph with the resume text data of the current applicant and the resume data in the enterprise's historical successful recruitment records having a relevant relationship as nodes and the relevant relationships between each resume as edges; obtain a resume embedding matrix and an adjacency matrix according to the resume graph; input the resume embedding matrix and the adjacency matrix into a graph neural network to obtain the resume representations of all resumes in the resume graph; splice each resume representation with the experience representation after interaction between the enterprise job recruitment requirements and the resume text data of the current applicant to obtain the global semantics of the resume text data of the current applicant;

[0009] Fuse the local semantics and the global semantics of the resume text of the current applicant, and judge the matching result between the resume and the enterprise job recruitment according to the final semantics of the resume text after fusion and the local semantics of the enterprise job recruitment requirements.

[0010] Furthermore, the local semantics of the enterprise job recruitment requirements containing multiple relationships and the local semantics of the resume text of the current applicant are respectively expressed as:

[0011]

[0012] Among them, represents the input sequence {j k,1 ,jk,2 ,…j k,u}, represents the experience sequence {r l,1 , r l,2 ,…r l,v} in the input resume text; j k,1 …j k,u and r l,1 …r l,v respectively represent the job requirements of the input neural network and the resume experience description, represents the local semantic information of the enterprise job recruitment requirements containing multiple relationships, represents the local semantic information of the resume text of the current applicant.

[0013] Furthermore, constructing a resume graph with the resume text data of the current applicant and the resume data in the enterprise's historical successful recruitment records related to it as nodes and the relevant relationships between each resume as edges specifically includes the following steps:

[0014] Constructing a resume graph based on an undirected graph; the resume graph G R-R is represented by (V R , E R ); the node V R represents the resume text data R of the current applicant c and the resume data in the enterprise's historical successful recruitment records related to R c , and the edge set E R represents the relevant relationships between each resume;

[0015] For each resume R i (R i ∈R r ), search for the resume set R h that satisfies the condition (J j , R h , 1) ∈ S K ; where, R r refers to the set of all resumes, J h refers to the job requirements that need to be matched and predicted for the resume R i , R j refers to the historical recruitment resume that has been successfully matched with J h , and S h refers to the pair of successfully matched job requirements and historical resumes;

[0016] According to the resume set R K , construct the adjacency matrix of the resume graph and calculate the weights of the edges in the adjacency matrix; where, for each R i , the [R1, R2,..., R p ∈ R that meets the search conditions will be usedr are regarded as different individual nodes, and p refers to the number of resume nodes that meet the conditions; for each resume R i Create a separate graph; obtain the weight w of each edge in the graph through the random number generation method; construct the adjacency matrix A in the resume graph according to the weight of each edge R .

[0017] Furthermore, inputting the resume embedding matrix and the adjacency matrix into the graph neural network to obtain the resume representations of all resumes in the resume graph, the specific process includes the following steps:

[0018] Regard each resume in the resume graph as a single node, and learn the representation of each resume node through the update function of the graph neural network

[0019]

[0020] Among them, represents the node vector of the resume, and A i is the i-th row of the adjacency matrix corresponding to node i, and H t , and b t are all learning parameters, is the update gate, is the reset gate, and ⊙ represents element-wise multiplication; σ is the Sigmoid function; is the activation value of node i at time step t; is the candidate node feature, represents the representation of node i at time step t - 1, is the feature representation of node i at time step t;

[0021] Construct a resume word embedding matrix H GR , and use the resume word embedding matrix to embed the representation of each resume node to obtain the resume embedding matrix g0;

[0022] Input the resume embedding matrix into the graph neural network to obtain the representations of all nodes in the resume graph, denoted as g R .

[0023] Furthermore, it also includes, after obtaining the resume representations of all resumes in the resume graph, using the soft attention mechanism to embed the resume representations into the vector space of the enterprise job recruitment requirements, and estimating the matching degree of each resume with V J ; the V J represents the semantic representation of the historical successfully recruited resumes in the vector space of the enterprise job recruitment requirements, and is expressed as:

[0024]

[0025] Weight coefficient α i It is expressed as:

[0026]

[0027] Wherein, represents the vector of each node in the resume graph, q r and W r are both learning parameters.

[0028] Furthermore, the splicing of each resume representation and the experience representation after the interaction between the enterprise job recruitment requirements and the resume text data of the current applicant is performed to obtain the global semantics of the resume text data of the current applicant, which is expressed as:

[0029]

[0030] Wherein, represents the resume representation after being trained by the graph neural network, W R and b R are learning parameters, and e R represents the experience relationship between the applicant's resume and the job requirements.

[0031] Furthermore, judging the matching result between the resume and the enterprise job recruitment according to the final semantics of the fused resume text and the local semantics of the enterprise job recruitment requirements specifically includes the following steps:

[0032] Furthermore, the local semantics of the resume text of the current applicant is fused through the KAN network and the global semantics Specifically expressed as:

[0033]

[0034] Wherein, H J represents the semantic representation of the job requirements, H R represents the semantic representation of the resume text, represents the local semantic information of the job recruitment requirements;

[0035] The cosine similarity is used to judge the matching result c between the resume and the enterprise recruitment, which is expressed as:

[0036] c = cosine(H J , H R ).

[0037] The present invention also includes a matching system for enterprise job recruitment and resumes, including:

[0038] An acquisition module, configured to acquire the resume text data of the current applicant, the enterprise job recruitment requirement data, and the resume data in the enterprise's historical successful recruitment records.

[0039] A semantic representation module, configured to input the enterprise job recruitment requirement data and the resume text data of the current applicant into the BERT model for sentence embedding, so as to obtain the local semantics of the enterprise job recruitment requirements containing multiple relationships and the local semantics of the resume text of the current applicant; construct a resume graph with the resume text data of the current applicant and the resume data in the enterprise's historical successful recruitment records having a relevant relationship as nodes and the relevant relationships between each resume as edges; obtain a resume embedding matrix and an adjacency matrix according to the resume graph; input the resume embedding matrix and the adjacency matrix into a graph neural network to obtain the resume representations of all resumes in the resume graph; splice each resume representation with the experience representation after the interaction between the enterprise job recruitment requirements and the resume text data of the current applicant to obtain the global semantics of the resume text data of the current applicant.

[0040] A matching module, configured to fuse the local semantics and the global semantics of the resume text of the current applicant, and judge the matching result between the resume and the enterprise job recruitment according to the final semantics of the fused resume text and the local semantics of the enterprise job recruitment requirements.

[0041] The present invention further includes a computer device for matching enterprise job recruitment and resumes, including: a memory, a processor, and a computer program stored in the memory, and when the processor executes the computer program, the steps of the method for matching enterprise job recruitment and resumes are implemented.

[0042] The present invention further includes a readable storage medium, the readable storage medium stores a computer program, the computer program includes program instructions, and when the program instructions are executed by a processor, the steps of the method for matching enterprise job recruitment and resumes are executed.

[0043] The present invention provides a method for matching enterprise job recruitment and resumes, having the following beneficial effects:

[0044] The present invention constructs a resume graph with the resume text data of the current applicant and the resume data in the enterprise's historical successful recruitment records having a relevant relationship as nodes and the relevant relationships between each resume as edges, introduces historical recruitment data to improve the model's ability to capture dynamic changes, and obtains the local semantic representation of job resumes based on BERT and the resume semantic representation based on the graph network; at the same time, according to each resume representation and the experience representation after the interaction between the enterprise job recruitment requirements and the resume text data of the current applicant, mines the resume information of job seekers from multiple perspectives, comprehensively captures the characteristics of job seekers, and improves the accuracy of the matching result. Brief Description of the Drawings

[0045] Figure 1 This is the flowchart of the method for matching enterprise job recruitment and resumes in the embodiments of the present invention;

[0046] Figure 2 This is the box plot of GPJFNNF and five baseline models on four metrics in the embodiments of the present invention;

[0047] Figure 3 This is the flowchart of the method for matching enterprise job recruitment and resumes in the embodiments of the present invention. Detailed implementation manners

[0048] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0049] The present invention proposes a method for matching enterprise job recruitment and resumes. As Figure 3 shown, a Graph-based person-job fit neural network fusion (GPJFNNF) model is proposed. The GPJFNNF model mainly includes three parts, namely, the local semantic representation of job resumes based on BERT, the semantic representation of resumes based on graph networks, and person-job matching interaction; Figure 1 where j k,1 …j k,u and r l,1 …r l,v respectively represent the job requirements and resume experience descriptions input into the neural network; R global represents the global semantic representation of the resume; H J , R local are respectively the local semantic representation of the job recruitment requirements and the local semantic representation of the resume; H R represents the final resume semantic representation after the KAN network fuses R global and R local ; represents the person-job matching result. The method specifically includes the following steps:

[0050] S1. Use the job requirements and resume experience descriptions as the input of the neural network, and the output is the label corresponding to the recruitment result (1 indicates successful recruitment, and 0 indicates failed recruitment).

[0051] The present invention aims to study the person-job matching task, with a focus on studying the matching degree between the requirements of the recruitment position and the experience in the applicant's resume. The job hunting process consists of a job requirement, the applicant's resume, and the job hunting record. Use J to represent the job, and the job contains o job requirements. J is represented by {j1, j2,... j o} is represented by. Each job requirement j k contains u words and can be represented by {j k,1 ,j k,2 ,…j k,u}. The same J may have multiple different job applications. Similarly, let R represent the resume of the applicant. The resume contains p pieces of experience (such as work experience, project experience, or competition experience, etc.). R is represented by {r1, r2, … r p}. Each piece of experience r l contains v words and can be represented by {r l,1 ,r l,2 ,…r l,v}. S h represents the historical successful recruitment records. Use S to represent the set of all job application processes, where each s i is composed of a pair {(J, R)}. s i is represented by s i =(j i ,r i ,y i ). The label y i represents the job application record of this job seeker. y i =1 indicates successful recruitment, while y i =0 indicates recruitment failure. The objective of the present invention is to learn a person-job matching prediction model to calculate the similarity of each pair {(J, R)} to predict the future job application process of applicants.

[0052] S2. In the local semantic representation of job resumes based on BERT, the BERT model is used to model the resumes of job seekers and enterprise recruitment information to generate the local semantic representations of job recruitment requirements and resumes.

[0053] To effectively capture the sentence-embedding semantic information of job recruitment requirements and resume texts, the present invention introduces a BERT model to extract the local semantics in job requirements and resume texts. Specifically, the pre-trained BERT model is used to perform sentence embedding on the input job requirements and resumes, generating corresponding semantic vectors, thereby obtaining a local semantic representation containing multiple relationships. First, the job requirements and resume sequences are passed into the word embedding layer and the position embedding layer. Secondly, through the Linear Transformations layer, the query, key, and value vectors of the job and the resume are obtained respectively. Then, by concatenating multiple attention heads to capture the features in different sentence spaces of the job resume, a semantic representation containing multiple relationships between different positions in the job requirements and the semantic representation of the resume are obtained. Then, Add&Normalize and Feed Forward Neural Network operations are performed. By inputting each word in the job recruitment requirements and resume texts, the local semantic information of the job recruitment requirements and resume containing multiple relationships is finally obtained The semantic representation vectors of the job and the resume are calculated by the following formula:

[0054]

[0055] Where represents the input sequence {j k,1 ,j k,2 ,…j k,u} of job requirements, represents the experience sequence {r l,1 ,r l,2 ,…r l,v} in the input resume text.

[0056] S3. For the resume semantic representation module based on the graph network, a graph structure is constructed according to historical successful recruitment records, and the constructed resume graph is input into the graph neural network. The attention mechanism is introduced to identify the importance of different skills in job requirements and resume experiences, and further obtain the global semantic representation of the resume; the graph neural network is used to model the resume text of the applicant.

[0057] S3.1 Resume graph construction method: In order to use the graph neural network to capture the global semantics of the resume in historical successful recruitment records, an undirected graph needs to be constructed first. The undirected graph is named Graph R-R, G R-R consisting of (V R , E R) It is shown that each node in graph G represents a resume. The initial features of the nodes are embedded through the BERT model to enhance the node representation and ensure that the initial features of each node contain deep semantic information. Node V R represents the resume R that needs to be predicted currently c and the historical successfully recruited resumes E c that have a relevant relationship with R R represents the edge set between these resumes. The edge represents the resume to be predicted and the historical successfully recruited resumes

[0058] Specifically, for each resume R i (R i ∈R r ), we want to predict whether this resume matches J h ; First, we need to search for the resume set R h that satisfies the condition (J j , R h , 1) ∈ S K , and consider that this set contains all resumes that have a relevant relationship with R i ; Then, for these resumes, we construct the adjacency matrix of the graph and calculate the weights of the edges in the adjacency matrix. For each R i , we regard [R1, R2,..., R p ∈ R r that meets the search condition as different individual nodes. For each resume R i , we create a graph and its adjacency matrix separately. We use the method of generating random numbers to obtain the weight w of each edge. After obtaining the weights w of all edges, we construct the adjacency matrix A R .

[0059] S3.2. Resume representation based on graph neural network: After constructing the graph matrix, we obtain the adjacency matrix A R related to resume information, and regard each resume in Graph R - R as an individual node. The representation of each resume node is learned through the update function of the GNN unit, as shown in the formula:

[0060]

[0061] Among them, represents the node vector of the resume, A i is the i-th row of the adjacency matrix corresponding to node i, H t , and b t are all learning parameters, is the update gate; is the reset gate, ⊙ represents element-wise multiplication, and σ is the Sigmoid function; is the candidate node feature, denotes the activation value of node i at time step t, denotes the representation of node i at time step t - 1, represents the hidden state of node i after t rounds (time step t) of propagation, that is, the updated node feature representation.

[0062] In addition, a trainable resume word embedding matrix H is constructed GR . The semantic information of the resume is embedded using the word embedding matrix to obtain g0, where g0 represents the initial representation of the node; then g0 is fed into the graph neural network to obtain the representations of all nodes in Graph R-R, denoted as g R ; the resume representation after training by the graph neural network is To further extract the higher-level semantic representation of the resume (i.e., the global semantics of the resume), after obtaining the vector representations of all nodes in Graph R-R, the resume semantic representation is embedded and mapped to the vector space of the recruitment position requirements through a soft attention mechanism. The calculation formula is as follows:

[0063]

[0064] The weight coefficient α i can be expressed as:

[0065]

[0066] where, represents the vector of each node in Graph R-R, q r and W r are both learnable parameters, V J represents the semantic representation of historical successfully recruited resumes in the position requirement vector space. Through V J the experiences (such as skills) that job seekers were valued for in historical recruitments can be evaluated. The attention mechanism is used to estimate the matching degree of each resume with V J . By designing an adaptive feature fusion mechanism, the continuous learning ability of the model is enhanced, the forgetting of key features of the resume during the learning process is alleviated, and the matching of job requirements and candidates' potential abilities is further enhanced.

[0067] The resume representation of Graph R-R after training by the graph neural network is concatenated with the experience representation e after the interaction between the position requirements and job search information R to obtain the global information of the current resume

[0068] where, e RIt represents the experience relationship between the applicant's resume and the job requirements, which is the relationship representation between the job requirements and the job hunting information. It is calculated through the attention mechanism and is the weighted sum of all node representations; W R and b R are learnable parameters.

[0069] S4. Integrate the local semantics and global semantics of the resume through the KAN network, input the final semantic representation of the resume and the representation of the job recruitment requirements into the human-job matching interaction layer, and automatically judge the matching result between the resume and the enterprise recruitment.

[0070] Obtain the local semantic information of the job requirements through the BERT-based job semantic representation method Obtain the local semantic information of the resume text through the BERT-based resume semantic representation and the graph network-based resume representation method respectively and the global semantic information For job requirements, only use the local semantics as the representation of the job recruitment text; for resume text, concatenate the local semantics and global semantics of the resume, and then use the KAN network to fuse the local and global semantics to represent the final semantics of the resume text.

[0071]

[0072] Each layer of the KAN network maps the input through a set of learnable functions to enhance the fusion of the local semantics and global semantics of the resume text, so that the network has better representation ability and strong adaptability.

[0073] After obtaining the semantic representations of the job requirements and the resume text respectively, use the cosine similarity to calculate the result of human-job matching. The calculation of the cosine similarity is as follows:

[0074] c = cosine(H J , H R ).

[0075] Experimental analysis:

[0076] To evaluate the effectiveness of the proposed GPJFNNF model in the job-person matching task, the GPJFNNF model was compared with some baseline methods. The methods compared in this invention include Random Forest (RF), XGBoost, PJFNN, BPJFNN, APJFNN, IPJF, and conSultantBERT. Table 1 lists the experimental results of accuracy, precision, recall, and F1 value. The first column of Table 1 lists the seven comparison methods and the proposed GPJFNNF model. The best results for each evaluation metric are shown in bold. The experimental results show that the proposed GPJFNNF model is very suitable for modeling the matching degree between the job requirements released by the company and the resumes of job seekers.

[0077] As shown in Table 1, the following observations can be made. In the job-person matching task, the GPJFNNF model achieved an accuracy of 94.63%, a precision of 94.15%, a recall of 95.04%, and an F1 value of 94.59%. Compared with all baselines, the proposed GPJFNNF model outperformed all baselines in terms of accuracy, precision, recall, and F1 value.

[0078] Table 1 Overall performance of the GPJFNNF model and all baselines in the job-person matching task

[0079]

[0080] Comparing traditional machine learning methods (such as RF and XGBoost) with deep learning methods (such as PJFNN, etc.), it can be observed that the overall performance of deep learning methods is better than that of traditional machine learning methods. The main reason is that deep models can learn more efficient semantic features from recruitment data than shallow models. Shallow models use doc2vec pre-trained word vectors as input features, which means that the pre-trained word vectors cannot fully represent the semantic features of job resume texts. The experimental results show that in this invention, deep neural network methods can capture semantic features better than traditional machine learning methods, which helps to improve the final performance.

[0081] Among the deep learning-based job-person matching models, PJFNN performed the worst in terms of accuracy, precision, and F1 value. This means that the CNN-based PJFNN model is not the best choice for matching jobs and job seeker resumes. Compared with PJFNN, BPJFNN further improved the performance of the job-person matching task. This phenomenon shows that parallel BiLSTM helps to learn the semantic representation of each word in job requirements and job seeker experiences.

[0082] The experimental results of conSultantBERT were further observed. The experimental results show that the performance of the fine-tuned Siamese Sentence-BERT model is better, indicating that the BERT-based deep learning model has better performance than the recurrent neural network-based model. The main reason for this result is the different encoding methods used in the job text and resume text. The research results show that fine-tuning the Siamese Sentence-BERT model is more suitable for modeling the text information in job requirements and candidate resumes.

[0083] The proposed GPJFNNF model was compared with the best-performing method conSultantBERT in the job-person matching task. The experimental results show that the GPJFNNF model outperforms this baseline method in all four evaluation metrics. The main reason is that the conSultant BERT method fails to effectively learn the global semantic relationship and non-Euclidean space information between job requirements and resume text. The introduction of graph neural networks is beneficial for specific job-person matching tasks for talent recruitment. The model of the present invention has two main advantages. On the one hand, the GPJFNNF model can effectively learn the semantics of job requirements and the global semantic relationship of job seeker resume text. On the other hand, the dual-perspective features of local and global semantics of resume text are fused using the KAN network, enhancing the semantic representation ability of the applicant's resume. The experimental results show that exploring the job-person matching task from another novel perspective significantly improves the performance.

[0084] Ablation experiment: To evaluate the impact of each component in GPJFNNF on the model performance, ablation experiments were conducted next. Two variants of GPJFNNF were designed: (A) GPJFNNF (w / o KAN) means removing the KAN network; (B) GPJFNNF (w / o GNN) means removing the GNN network. The ablation experiment results are shown in Table 2.

[0085] Table 2 Ablation Experiment Results

[0086]

[0087] The experimental results show that:

[0088] (1) No matter which component is removed, the performance of the model will be weakened. The effectiveness of each component in the model is verified.

[0089] (2) The GPJFNNF (w / o KAN) component decreased by 2.81%, 2.01%, 3.79%, and 2.90% respectively in the evaluation metrics of accuracy, precision, recall, and F1 value. The results show that the adaptive feature fusion mechanism can effectively fuse the local information and global information of the resume.

[0090] (3)The (w / o GNN) component of GPJFNNF decreased by 11.57%, 11.67%, 11.59%, and 11.63% respectively in terms of the evaluation metrics of Accuracy, Precision, Recall, and F1-score. The ablation experiment results indicate that GNN has the ability to efficiently capture the global information of resumes.

[0091] In this invention, repeated experiments are conducted on the proposed model and the top five baseline models in terms of performance. Two data analysis methods, Wilcoxon signed-rank test and Cliff's delta test, are used to analyze the performance of GPJFNNF, PJFNN, BPJFNN, APJFNN, IPJF, and conSultantBERT.

[0092] The Wilcoxon signed-rank test is a non-parametric statistical hypothesis test that does not assume the data follows a normal distribution and is commonly used to compare whether the distributions of two groups of samples are the same. Compared with the Student's t-test, the Wilcoxon signed-rank test is more sensitive in statistical analysis and can more effectively detect significant differences. The p-value is used to evaluate whether the differences between two paired samples are statistically significant. A p-value < 0.05 indicates a significant difference.

[0093] Cliff's delta test is a non-parametric effect size test and serves as a supplementary analysis to the Wilcoxon signed-rank test in this invention. It measures the difference between two populations in numerical form. Table 3 shows the relationship between Cliff's δ value and its corresponding effect size.

[0094] Table 3 Relationship between Cliff's δ value and its corresponding effect size

[0095]

[0096] Table 4 presents the comparison results of the proposed model with the top five performing baselines in four metrics. The first column represents the evaluation metrics of the model; the second to sixth columns respectively show the performance comparison results between the GPJFNNF method and other baseline methods. <0.05: indicates that the P-value is less than 0.05 in the Wilcoxon signed-rank test, indicating that the results are statistically significant. (+Large): indicates that at the effective level of Cliff's δ value, the GPJFNNF method has achieved a significant performance improvement compared to other methods, and the improvement amplitude is large. The "+" and "-" signs represent the direction of the difference in performance between the two compared models. A positive value indicates that the performance of the first model is generally greater than that of the second model, while a negative value is the opposite. The results of all metrics show that the proposed GPJFNNF model is superior to other baselines. The box plots are used to show the performance distribution of these five models in four metrics, as shown in (a), (b), (c), and (d) in Figure 2 as shown.

[0097] Table 4 Significance Test of GPJFNNF and Top 5 Baseline Performances

[0098]

[0099] Based on the same inventive concept, the present invention also proposes an enterprise job recruitment and resume matching system, including:

[0100] An acquisition module, configured to acquire the resume text data of the current applicant, the enterprise job recruitment requirement data, and the resume data in the enterprise's historical successful recruitment records.

[0101] A semantic representation module, configured to perform sentence embedding on the enterprise job recruitment requirement data and the resume text data of the current applicant by inputting them into the BERT model to obtain the local semantics of the enterprise job recruitment requirements containing multiple relationships and the local semantics of the resume text of the current applicant; construct a resume graph with the resume text data of the current applicant and the resume data in the enterprise's historical successful recruitment records having a relevant relationship as nodes and the relevant relationship between each resume as edges, obtain a resume embedding matrix and an adjacency matrix according to the resume graph; input the resume embedding matrix and the adjacency matrix into a graph neural network to obtain the resume representations of all resumes in the resume graph; splice each resume representation with the experience representation after the interaction between the enterprise job recruitment requirements and the resume text data of the current applicant to obtain the global semantics of the resume text data of the current applicant.

[0102] A matching module, configured to fuse the local semantics and the global semantics of the resume text of the current applicant through the KAN network, and judge the matching result between the resume and the enterprise job recruitment according to the final semantics of the fused resume text and the local semantics of the enterprise job recruitment requirements.

[0103] The present invention also provides a computer device for matching enterprise job recruitment and resumes, comprising: a memory, a processor, and a computer program stored in the memory. When the processor executes the computer program, the steps of the method for matching enterprise job recruitment and resumes are implemented.

[0104] The present invention also provides a readable storage medium storing a computer program, the computer program including program instructions which, when executed by a processor, are used to execute the steps of the method for matching enterprise job recruitment and resumes.

[0105] As described above, the above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, shall be covered by the protection scope of the present invention.

Claims

1. A matching method for enterprise job recruitment and resumes, characterized in that It includes the following steps: Obtain the resume text data of the current applicant, the enterprise job recruitment requirement data, and the resume data in the historical successful recruitment records of this enterprise; By inputting the enterprise job recruitment requirement data and the resume text data of the current applicant into the BERT model for sentence embedding, obtain the local semantics of the enterprise job recruitment requirements with multiple relationships and the local semantics of the resume text of the current applicant; construct a resume graph with the resume text data of the current applicant and the resume data in the historical successful recruitment records of the enterprise that has a relevant relationship with it as nodes and the relevant relationships between each resume as edges; Obtain the resume embedding matrix and the adjacency matrix according to the resume graph; input the resume embedding matrix and the adjacency matrix into the graph neural network to obtain the resume representations of all resumes in the resume graph; splice each resume representation with the experience representation after the interaction between the enterprise job recruitment requirements and the resume text data of the current applicant to obtain the global semantics of the resume text data of the current applicant; Fuse the local semantics and the global semantics of the resume text of the current applicant, and judge the matching result between the resume and the enterprise job recruitment according to the final semantics of the fused resume text and the local semantics of the enterprise job recruitment requirements.

2. The matching method of enterprise job recruitment and resume according to claim 1, characterized in that The local semantics of the enterprise job recruitment requirements with multiple relationships and the local semantics of the resume text of the current applicant are respectively expressed as: Among them, denotes the input sequence {j k,1 , j k,2 , … j k,u} representing job requirements, denotes the experience sequence {r l,1 , r l,2 , … r l,v} in the input resume text; j k,1 … j k,u and r l,1 … r l,v respectively represent the job requirements and resume experience descriptions input to the neural network, denotes the local semantic information of the enterprise job recruitment requirements containing multiple relationships, denotes the local semantic information of the current applicant's resume text.

3. A matching method for enterprise job recruitment and resumes according to claim 1, characterized in that The construction of the resume graph with the resume text data of the current applicant and the resume data in the historical successful recruitment records of the enterprise that has a relevant relationship with it as nodes and the relevant relationships between each resume as edges specifically includes the following steps: Construct a resume graph based on an undirected graph; the resume graph G R-R is represented by (V R , E R ). The node V R represents the resume text data R of the current applicant c and the resume data in the historical successful recruitment records of the enterprise that has a relevant relationship with R c . The edge set E R represents the relevant relationship between each resume; For each resume R i (R i ∈R r ), search for the set of resumes R h that satisfy the condition (J j , R h , 1) ∈ S K ; where R r refers to the set of all resumes, J h refers to the job requirements that need to be matched with the resume R i , R j refers to the historical recruitment resumes that have been successfully matched with J h , and S h refers to the pairs of job requirements and historical resumes that have been successfully matched; According to the resume set R K , construct the adjacency matrix of the resume graph and calculate the weights of the edges in the adjacency matrix; among them, for each R i , the [R1, R2,..., R p ∈ R that meets the search conditions r is regarded as different single nodes, and p refers to the number of resume nodes that meet the conditions; for each resume R i , create a separate graph; obtain the weight w of each edge in the graph through the random number generation method; construct the adjacency matrix A in the resume graph according to the weights of each edge R .

4. A method for matching enterprise job recruitment and resumes according to claim 1, characterized in that The input of the resume embedding matrix and the adjacency matrix into the graph neural network to obtain the resume representations of all resumes in the resume graph, and the specific process includes the following steps: Regarding each resume in the resume graph as a single node, the representation of each resume node is learned through the update function of the graph neural network Among them, represents the node vector of the resume, A i is the i-th row of the adjacency matrix corresponding to node i, H t , and b t are all learning parameters, is the update gate, is the reset gate, ⊙ represents element-wise multiplication, and σ is the Sigmoid function; is the activation value of node i at time step t; is the candidate node feature, represents the representation of node i at time step t - 1, is the feature representation of node i at time step t; Construct a resume word embedding matrix H GR , and use the resume word embedding matrix to embed the representation of each resume node to obtain the resume embedding matrix g0; Input the resume embedding matrix into the graph neural network to obtain the representations of all nodes in the resume graph, denoted as gR.

5. The matching method of enterprise job recruitment and resume according to claim 4, characterized in that It also includes, after obtaining the resume representations of all resumes in the resume graph, using a soft attention mechanism to embed and map the resume representations into the vector space of the enterprise job recruitment requirements, and estimating the matching degree of each resume with V J ; where the V J represents the semantic representation of the historical successfully recruited resumes in the vector space of the enterprise job recruitment requirements, and is expressed as: Weight coefficient α i It is expressed as: Among them, represents the vector of each node in the resume graph, q r and W r are both learning parameters, and b is the bias term.

6. The method for matching an enterprise job recruitment and a resume according to claim 4, characterized in that Splicing each resume representation with the experience representation after interacting with the enterprise job recruitment requirements and the resume text data of the current applicant to obtain the global semantics of the resume text data of the current applicant It is expressed as: Among them, represents the resume representation after being trained by the graph neural network, W R and b R are learning parameters, e R represents the experience relationship between the applicant's resume and the job requirements.

7. A method for matching enterprise job recruitment and resumes according to claim 6, characterized in that, The judgment of the matching result between the resume and the enterprise job recruitment according to the final semantics of the fused resume text and the local semantics of the enterprise job recruitment requirements specifically includes the following steps: Fuse the local semantics of the current applicant's resume text through the KAN network and the global semantics Specifically expressed as: Among them, H J represents the semantic representation of job requirements, and H R represents the semantic representation of resume text, representing the local semantic information of job recruitment requirements; Use cosine similarity to judge the matching result c between the resume and the enterprise recruitment, which is expressed as: c = cosine(H J , H R ).

8. An enterprise job recruitment and resume matching system, characterized in that, It includes: An acquisition module for acquiring the resume text data of the current applicant, the enterprise job recruitment requirement data, and the resume data in the historical successful recruitment records of this enterprise; A semantic representation module for obtaining the local semantics of the enterprise job recruitment requirements with multiple relationships and the local semantics of the resume text of the current applicant by inputting the enterprise job recruitment requirement data and the resume text data of the current applicant into the BERT model for sentence embedding; constructing a resume graph with the resume text data of the current applicant and the resume data in the historical successful recruitment records of the enterprise that has a relevant relationship with it as nodes and the relevant relationships between each resume as edges; Obtain the resume embedding matrix and the adjacency matrix according to the resume graph; input the resume embedding matrix and the adjacency matrix into the graph neural network to obtain the resume representations of all resumes in the resume graph; splice each resume representation with the experience representation after the interaction between the enterprise job recruitment requirements and the resume text data of the current applicant to obtain the global semantics of the resume text data of the current applicant. A matching module is used to fuse the local semantics and the global semantics of the resume text of the current applicant, and judge the matching result between the resume and the enterprise job recruitment according to the final semantics of the fused resume text and the local semantics of the enterprise job recruitment requirements.

9. A computer device for matching enterprise job recruitment and resumes, characterized in that, It includes: A memory, a processor, and a computer program stored in the memory. When the processor executes the computer program, the steps of the matching method for enterprise job recruitment and resume described in any one of claims 1-7 are implemented.

10. A readable storage medium, characterized in that, The readable storage medium stores a computer program, and the computer program includes program instructions. When the program instructions are executed by the processor, they are used to execute the steps of the matching method for enterprise job recruitment and resume described in any one of claims 1-7.

Images