Post matching method and device based on resume quality and structure correlation learning

Abstract

The application discloses a post matching method and device based on resume quality and structure correlation learning, and relates to the field of natural language processing.The method comprises obtaining a skill vector and a work experience vector from a resume, and obtaining a post demand vector from a post demand; fusing word vectors from the same work experience in the work experience vector into a vector; performing attention interaction between multi-scale post features obtained after learning and resume representation to obtain attention graphs of the resume and the post at different scales; taking the output of the last layer of the post sub-network and splicing the same as the final resume-post pair representation; extracting structure information in the constructed structure graph; fusing the extracted structure information and the final resume-post pair representation to obtain fusion features and inputting the fusion features into an MLP layer to obtain a matching score.The application can effectively solve the problem of ignoring resume quality, and considers the extraction of shallow structure information of the resume-post.

Description

Technical Field

[0001] This invention relates to the field of natural language processing, and specifically to a job matching method and apparatus based on learning the relevance between resume quality and structure. Background Technology

[0002] With the rapid development of internet and digital technologies, an increasing number of job seekers and recruiters are posting their resumes and job requirements on online recruitment platforms. Online recruitment is a completely different model from traditional recruitment methods. It is unaffected by geographical location, time, or transportation. Job seekers can search for suitable positions anytime, anywhere based on keywords such as job title, work location, salary, and job content, and submit their resumes. Recruiters can also quickly post job openings and receive satisfactory resumes. Furthermore, the convenience of the internet saves time for both recruiters and job seekers to communicate, bringing convenience to both parties.

[0003] While recruitment websites offer convenience, the sheer number of job seekers has led to a massive accumulation of resumes in recruiters' pools. The influx of graduates submitting resumes to recruitment websites during campus recruitment season forces recruiters to sift through a massive volume of applications in a short period. Previously, human resources departments manually compared resumes for each job opening and selected suitable candidates for interviews. However, given the rapid increase in users, manually screening a large number of resumes and job postings is extremely time-consuming. Therefore, researching how to automatically match resume content with job requirements to predict suitable candidates and accelerate the recruitment process is an important topic.

[0004] Despite the success of the methods mentioned above, the following problems remain: Existing methods often only calculate the similarity between resumes and job titles, while the quality of the resume is ignored. An excellent candidate's resume is usually of high quality, and if the resume is of low quality, it should not be given much consideration. Previous work generated the final matching score by encoding the resume and job title and then generating it through simple attention interaction. This interaction relies on semantic information but lacks utilization of the resume-job title structural information. Structural information refers to the relevance of the resume and job title in sentence structure, such as word relationships and contextual order. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a job matching method and device based on resume quality and structural relevance learning, which can effectively solve the problem of ignoring resume quality and take into account the extraction of shallow structural information of resume-job.

[0006] To achieve the above objectives, this invention provides a job matching method based on resume quality and structural relevance learning, which specifically includes the following steps:

[0007] We obtain resumes and job requirements to build a dataset, and perform word segmentation and filtering on the text data in the dataset. We then vectorize the text data, obtaining skill vectors and work experience vectors from resumes, and job requirement vectors from job requirements.

[0008] The obtained vectors are adjusted based on BiLSTM, and word attention mechanism is used to merge word vectors from the same work experience into a single vector.

[0009] The adjusted skill vector and work experience vector are fed into the constructed fine-grained resume representation learning subnetwork to learn resume consistency features, and the adjusted job requirement vector is fed into the constructed multi-scale job representation learning subnetwork to learn multi-scale job features.

[0010] The multi-scale job features obtained after learning are interacted with the resume representation through attention to obtain attention maps of resumes and job positions at different scales.

[0011] The attention map is applied to the resume representation. Based on the max pooling method, the attention-enhanced resume features are transformed into discriminative features. The output of the last layer of the job sub-network is concatenated with it to obtain the final resume-job pair representation.

[0012] A structure graph is constructed based on skill vectors, work experience vectors, and job requirement vectors, and structural information is extracted from the constructed structure graph using a selective graph attention network.

[0013] An adaptive fusion gate is used to fuse the extracted structural information and the final resume-job pair representation to obtain fused features, which are then fed into an MLP layer to obtain a matching score.

[0014] Based on the above technical solution, the fine-grained CV representation learning subnetwork includes a multi-head, noisy common attention module based on Transformer with multiple iterations, and a residual connection module to prevent gradient vanishing.

[0015] Based on the above technical solutions,

[0016] The Transformer-based multi-head noise-included common attention module includes a Transformer encoder and a multi-head noise-included common attention module.

[0017] The Transformer encoder is used as a feature extractor to extract features from the skill representation and the work experience representation, respectively.

[0018] The multi-headed, noise-included common attention module is used to generate mutual attention maps between skill representations and work experience representations, and adds noise vectors to absorb inconsistent skills and work experience.

[0019] Based on the above technical solution, the multi-scale job representation learning subnetwork is composed of multiple convolutional layers of different sizes stacked together, and the output of the multi-scale job representation learning subnetwork is job features of different scales.

[0020] Based on the above technical solution, the multi-scale job features obtained after learning are interacted with the resume representation through attention to obtain attention maps of resumes and job positions at different scales. The obtained attention maps are represented as follows:

[0021]

[0022] H∈R (N+M)×K

[0023]

[0024] Where f represents the softmax function, H i Let be the similarity matrix between the resume representation and the job posting representation at the i-th scale. Let R represent the attention map at the i-th scale, and let R be the final representation of the resume. Let H represent the transpose of the output of the i-th layer of the multi-scale job description learning subnetwork, H represent the similarity matrix, N represent the number of job skills, M represent the number of job experiences, and K represent the number of words in the job description.

[0025] Based on the above technical solutions,

[0026] The structure diagram constructed based on skill vectors, work experience vectors, and job requirement vectors includes four types of nodes: skill nodes, work experience nodes, job requirement nodes, and global nodes.

[0027] For nodes in the structure graph:

[0028] If nodes i and j have a contextual relationship, and the distance between the two words corresponding to nodes i and j is less than or equal to a set distance c, then the edge connecting these two words is a. ij =1;

[0029] When nodes i and j are skill nodes or job nodes, and the two words corresponding to nodes i and j are the same, then the edge connecting these two words is a.ij =1;

[0030] When nodes i and j are global nodes or work experience nodes, and the connecting edge a between the two words corresponding to nodes i and j is... ij =1.

[0031] Based on the above technical solution, the selective graph attention network extracts structural information from the constructed structural graph, wherein, for the selective graph attention network:

[0032] Selective graph attention network projects the input features of the (k+1)th layer of the graph attention network onto the information selection space through a projection matrix, and uses the sigmoid activation function to map them to the range of 0 to 1, thus obtaining a selection matrix β with selection function;

[0033] The selection matrix β is used as an update gate to selectively fuse information from the current order and previous orders. 1-β controls the degree of retention of the previous order and β controls the degree of retention of the current order.

[0034] Based on the above technical solution, the extracted structural information and the final resume-job pair representation are fused using an adaptive fusion gate. Specifically:

[0035] An adaptive fusion gate is used to combine the representation features h of the resume-job pair representation with the resume-job structure information X. s The features are spliced ​​together to obtain the fused feature X. m The calculation process is as follows:

[0036] X m =h||X s .

[0037] Based on the above technical solutions,

[0038] The activation function of the first layer of the MLP layer is the ReLU function, and the activation function of the second layer is the sigmoid function.

[0039] The matching score ranges from [0,1];

[0040] The specific process of sending the data to the MLP layer to obtain the matching score is as follows:

[0041]

[0042] in, represents the matching score, sigmoid represents the sigmoid activation function, ReLU represents the ReLU function, W4, W3, b3, and b4 represent the learnable parameters, and x represents the fused feature.

[0043] This invention provides a job matching device based on resume quality and structure relevance learning, comprising:

[0044] The acquisition module is used to acquire resumes and job requirements to build a dataset, and to perform word segmentation and filtering on the text data of the dataset, and to vectorize the text data, obtaining skill vectors and work experience vectors from resumes, and job requirement vectors from job requirements.

[0045] The fusion module is used to adjust the obtained vectors based on BiLSTM and fuse word vectors from the same work experience into a single vector through a word attention mechanism.

[0046] The learning module is used to feed the adjusted skill vector and work experience vector into a fine-grained resume representation learning subnetwork to learn resume consistency features, and to feed the adjusted job requirement vector into a multi-scale job representation learning subnetwork to learn multi-scale job features.

[0047] The interaction module is used to interact with the multi-scale job features obtained after learning and the resume representation through attention, so as to obtain attention maps of resumes and job positions at different scales.

[0048] The application module is used to apply the obtained attention map to the resume representation. Based on the max pooling method, the attention-enhanced resume features are transformed into discriminative features. The output of the last layer of the job sub-network is concatenated with it to obtain the final resume-job pair representation.

[0049] The module is used to construct a structure graph based on skill vectors, work experience vectors, and job requirement vectors, and to extract structural information from the constructed structure graph based on a selective graph attention network.

[0050] The execution module is used to fuse the extracted structural information and the final resume-job pair representation through an adaptive fusion gate to obtain fused features, which are then fed into the MLP layer to obtain a matching score.

[0051] Compared with the prior art, the advantages of the present invention are as follows:

[0052] (1) This invention incorporates resume quality into the consideration of job matching tasks, divides resume features into skill features and work experience features. To measure the quality of a resume, it is to measure the consistency between the resume skill features and the resume work experience features, effectively solving the problem of ignoring resume quality, and also considering the extraction of shallow structural information of resume-job.

[0053] (2) This invention takes into account the neglected shallow structural information and extracts the shallow structural information of resume-job position. Structural information refers to the relevance of resume and job position in sentence structure (for example, job positions that match the resume usually contain the same keywords and contextual associations, etc.). The use of this structural information is conducive to achieving more accurate matching score calculation. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a flowchart of a job matching method based on resume quality and structural relevance learning in an embodiment of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of this application, but not all embodiments.

[0057] This invention provides a job matching method based on resume quality and structural relevance learning, which solves the problems of previous methods that only calculated the similarity between resumes and positions while ignoring resume quality and lacking utilization of resume-job structure information.

[0058] See Figure 1 As shown in the figure, the job matching method based on resume quality and structure relevance learning provided by this embodiment of the invention specifically includes the following steps:

[0059] S1: Obtain resumes and job requirements to build a dataset, and perform word segmentation and filtering on the text data in the dataset. Vectorize the text data to obtain skill vectors and work experience vectors from resumes, and job requirement vectors from job requirements.

[0060] This involves removing meaningless stop words to reduce the dimensionality of the text data, as well as performing lexical reconstruction and stemming to convert words into their basic forms, aiming to reduce noise and errors introduced by word morphological variations in text analysis. Finally, BERT was chosen as the pre-trained model for word vector embedding. Compared to methods such as Word2Vec, BERT generates word vectors dynamically based on different sentence contexts, incorporating contextual information.

[0061] S2: The obtained vectors are adjusted based on BiLSTM (Bidirectional Long Short-Term Memory Network), and word attention mechanism is used to merge word vectors from the same work experience into a single vector;

[0062] That is, for the obtained skill vector S emd = [s1, s2, ... s N ]∈R d×N Job requirement vector P emmd = [p1, p2, ... p K ]∈R d×K The obtained vectors are fed into a BiLSTM layer to learn the semantic features of the sequence and obtain the skill representation s = [s1, s2, ... s2]. N ]∈R d×N The job positions are represented as P = [p1, p2, ... p2]. K ]∈R d×K .

[0063] Unlike the skills and job description sections, the work experience section may contain many sentences describing various work experiences, each with many words. Without processing, this would generate numerous word vectors, slowing down inference. To address this issue, a hierarchical attention network (HAN) is used to learn a unified representation for the candidate's work experience. The resulting vector for all work experience is E = [e1, e2, ... e2]. M ]∈R d×M .

[0064] S3: The adjusted skill vector and work experience vector are fed into the constructed fine-grained resume representation learning sub-network to learn resume consistency features, and the adjusted job requirement vector is fed into the constructed multi-scale job representation learning sub-network to learn multi-scale job features.

[0065] In this invention, the fine-grained CV representation learning subnetwork includes a multi-head noisy joint attention module based on Transformer with multiple iterations, and a residual connection module to prevent gradient vanishing. The multi-head noisy joint attention module based on Transformer includes a Transformer encoder and a multi-head noisy joint attention module; the Transformer encoder is used as a feature extractor to extract features from the skill representation and work experience representation respectively; the multi-head noisy joint attention module is used to generate mutual attention maps between the skill representation and work experience representation, and adds noise vectors to absorb inconsistent skills and work experience.

[0066] First, a Transformer encoder is used as a feature extractor to extract features from the skill representation and work experience representation respectively. The skill and work experience features obtained through the Transformer encoder are then fed into a noisy joint attention module. The quality of the resume is crucial for the outcome matching of the talent-job task, but the standard joint attention mechanism cannot handle poor-quality resumes, such as those lacking supporting evidence for skills or work experience. This may lead to the attention map misleading network training. To address this issue, T noise vectors are generated to store these inconsistent skills and experiences. These noise vectors, combined with the skill and work experience features, are used to compute the attention map. The skill and work experience features are concatenated with the noise vectors respectively to obtain noisy skill features S′ and noisy work experience features E′. Simultaneously, a multi-head mechanism is used to improve the robustness of the framework. Let q represent the number of low-dimensional features, and their dimension is represented by d. q = d / q. Therefore, each similarity matrix A i The calculation method is as follows:

[0067]

[0068] in, and These are learnable weight parameters, i = 1, ..., q. Similar to the previous operations, each attention graph can be obtained from this:

[0069]

[0070]

[0071] Here, f is the softmax function. Therefore, the average attention map can be obtained as follows:

[0072]

[0073] Then, a dot product attention operation is performed to obtain the consistent skill feature S. c and consistent work experience E c Meanwhile, since inconsistent skills or experiences are stored in the noise vector, they are ignored in this process. Therefore, the consistency feature is obtained by the following formula:

[0074] S c =S′A S [1:M,:] T

[0075] E c =E′A E [1:N,:] T

[0076] The Transformer encoder and a multi-headed, noisy common attention module are integrated into a module called the MNICAT module. This module is then looped M times to obtain deeper semantic features and interaction information.

[0077] After obtaining the consistent skill characteristic S c and consistent work experience E c Subsequently, fine-grained resume feature matrices are learned using residual connected networks. These consistent features are concatenated with the original skill features S and work experience features E, projected onto a dimensional space equal to the original input using a fully connected layer and a ReLU activation function. The mapped features are then added to the original feature matrix, as follows:

[0078] S p =ReLU(W1(S) c ⊙S)+b1)

[0079] E p =ReLU(W2(E) c ⊙E)+b2)

[0080] S res =S p +S

[0081] E res =E p +E

[0082] Where ⊙ represents the concatenation operation, and W1, W2, b1, b2 are learnable parameters, S res and E res This represents the final skill and work experience representation. Finally, S... res and E res The columns are concatenated into a feature matrix R to obtain a fine-grained resume representation, as follows:

[0083] R = S res ⊙E res .

[0084] The obtained job description features are fed into a multi-scale job description learning sub-network to learn multi-scale job description features. This involves stacking N convolutional layers and processing the first convolutional layer with F convolutional kernels to obtain a new job description feature, where X depends on the kernel size and stride. Stacking multiple convolutional layers and concatenating convolutional layers with kernels of different sizes can capture language features of different lengths. The job descriptions extracted from different levels are as follows:

[0085] J i =f i (Ji-1 ), i = 1, ..., N

[0086] Among them, J i It starts from the i-th convolutional layer f i The output job description, J0, represents the original job description features from the feature extraction module described above.

[0087] S4: The multi-scale job features obtained after learning are interacted with the resume representation through attention to obtain attention maps of resumes and job positions at different scales;

[0088] In this invention, the multi-scale job representation learning subnetwork is composed of multiple convolutional layers of different sizes stacked together, and the output of the multi-scale job representation learning subnetwork is job features of different scales.

[0089] In this invention, the multi-scale job features obtained after learning are interacted with the resume representation through attention, resulting in attention maps of the resume and job at different scales. The resulting attention maps are represented as follows:

[0090]

[0091] H∈R (N+M)×K

[0092]

[0093] Where f represents the softmax function, H i Let be the similarity matrix between the resume representation and the job posting representation at the i-th scale. Let R represent the attention map at the i-th scale, and let R be the final representation of the resume. Let H represent the transpose of the output of the i-th layer of the multi-scale job description learning subnetwork, H represent the similarity matrix, N represent the number of job skills, M represent the number of job experiences, and K represent the number of words in the job description.

[0094] S5: Apply the obtained attention map to the resume representation, and transform the attention-enhanced resume features into discriminative features based on the max pooling method. Take the output of the last layer of the job sub-network and concatenate it with the resume to obtain the final resume-job pair representation.

[0095] The resulting multi-scale attention maps are applied to resumes to obtain multi-level attentional resume features (but not to job title features). Then, max pooling is used to transform the attentional resume features into discriminative features {r1, r2, r3, ..., r...}. N Finally, the output of the last layer of the job subnetwork is obtained and max-pooled into a vector, which serves as the job representation J.N We concatenate all these vectors to obtain the final resume-job pair representation h.

[0096] S6: Construct a structure graph based on skill vectors, work experience vectors, and job requirement vectors, and extract structural information from the constructed structure graph based on a selective graph attention network;

[0097] In this invention, a structure graph is constructed based on skill vectors, work experience vectors, and job requirement vectors. The constructed structure graph includes four types of nodes: skill nodes, work experience nodes, job nodes, and global nodes.

[0098] In this invention, for nodes in the structural diagram:

[0099] If nodes i and j have a contextual relationship, and the distance between the two words corresponding to nodes i and j is less than or equal to a set distance c, then the edge connecting these two words is a. ij =1;

[0100] When nodes i and j are skill nodes or job nodes, and the two words corresponding to nodes i and j are the same, then the edge connecting these two words is a. ij =1;

[0101] When nodes i and j are global nodes or work experience nodes, and the connecting edge a between the two words corresponding to nodes i and j is... ij =1.

[0102] In this invention, structural information is extracted from the constructed structural graph based on a selective graph attention network. Specifically, for the selective graph attention network:

[0103] Selective graph attention network projects the input features of the (k+1)th layer of the graph attention network onto the information selection space through a projection matrix, and uses the sigmoid activation function to map them to the range of 0 to 1, thus obtaining a selection matrix β with selection function;

[0104] The selection matrix β is used as an update gate to selectively fuse information from the current and previous orders. 1-β controls the degree of retention of information from the previous order, while β controls the degree of retention of information from the current order. The closer β is to 1, the more information from the previous order is fused; the closer β is to 0, the more information from the current order is fused.

[0105] In this invention, structural information is extracted from the constructed structural graph using an improved Selective Graph Attention Network (SGAT). Typically, multiple layers of the Graph Attention Network (GAT) are stacked to learn deeper graph structural information. However, with a large number of layers, excessive aggregation can lead to performance degradation. Therefore, to better integrate the multi-level neighborhood information of nodes, it is necessary to filter the information transmitted between different layers. Thus, the Selective Graph Attention Network (SGAT) is designed to selectively integrate multi-level information. Utilizing the attention mechanism of the Graph Attention Network (GAT), for a GAT with K layers and H attention heads, the set of nodes i and their connected nodes is calculated in each layer. The attention weight of each node.

[0106]

[0107] Where k represents the k-th layer of GAT; h represents the h-th tap. Both are learnable weight matrices, and || denotes the concatenation operation. After obtaining the attention weights, the new feature representation of node i can be updated by weighted summation of neighboring nodes, as follows:

[0108]

[0109] in, Let be the learnable weight matrix, and tanh be the activation function. The input features of layer (k+1) are projected onto a matrix. Projecting onto the information selection space and mapping it to the range of 0 to 1 using a sigmoid activation function, we obtain the selection matrix β with selection capabilities as follows:

[0110]

[0111] Among them, W c This is the projection matrix, and σ is the sigmoid function. The value β in the selection matrix β... ij Let β represent the selection coefficient of the i-th node in the j-th dimension, with a value ranging from 0 to 1. This selection coefficient controls the degree of fusion of neighborhood information of the node at different orders. The selection matrix β is used as an update gate to selectively fuse information from the current order and previous orders to obtain the node representation of the next order.

[0112]

[0113] After extracting features using SGAT, the representation of the global nodes in the last layer of SGAT is used as the final representation of the graph structure features.

[0114] S7: The extracted structural information and the final resume-job pair representation are fused through an adaptive fusion gate to obtain fused features, which are then fed into the MLP layer to obtain a matching score.

[0115] In this invention, an adaptive fusion gate is used to fuse the extracted structural information and the final resume-job title pair representation. Specifically:

[0116] An adaptive fusion gate is used to combine the representation features h of the resume-job pair representation with the resume-job structure information X. s The features are spliced ​​together to obtain the fused feature X. m The calculation process is as follows:

[0117] X m =h||X s .

[0118] Fuse feature X m Mapping these features to the global semantic representation feature space and the global structural information feature space respectively is called activation, and this operation yields activation values ​​e for the semantic representation features and structural information features respectively. h and e s as follows:

[0119] e h =σ(Wx m +b)

[0120] e s =σ(Wx m +b)

[0121] Here, σ is the sigmoid function, and the values ​​of W and b from the two activations are not shared; they are trained independently during network training. The obtained activation value e... h and e s The activation values ​​are used as weights for semantic and structural representations, respectively. Weighted features are obtained by performing a Hadamard product between the activation values ​​and the original features. Finally, the weighted features are added to the original features to obtain a semantic representation incorporating structural features and a structural representation incorporating semantic features, X′. s And h′ are as follows:

[0122] X′ s =e s ⊙X s +X s

[0123] h′=e h ⊙h+h

[0124] Where ⊙ represents the Hadamard product, the new semantic representation and structural representation features are concatenated to obtain the final fused representation x as follows:

[0125] x=X′ s ||h′.

[0126] In this invention, the activation function of the first layer of the MLP layer is the ReLU function, and the activation function of the second layer is the sigmoid function;

[0127] The matching score ranges from [0, 1]. By limiting the matching score to [0, 1], the dropout strategy is used to prevent overfitting and improve the model's generalization ability.

[0128] The specific process of sending the data to the MLP layer to obtain the matching score is as follows:

[0129]

[0130] in, represents the matching score, sigmoid represents the sigmoid activation function, ReLU represents the ReLU function, W4, W3, b3, and b4 represent the learnable parameters, and x represents the fused feature.

[0131] This invention provides a job matching device based on resume quality and structure relevance learning, comprising an acquisition module, a fusion module, a learning module, an interaction module, an application module, a construction module, and an execution module.

[0132] The acquisition module acquires resumes and job requirements to build a dataset, performs word segmentation and filtering on the text data, and vectorizes the text data to obtain skill vectors and work experience vectors from resumes, and job requirement vectors from job requirements. The fusion module adjusts the obtained vectors based on BiLSTM and uses a word attention mechanism to fuse word vectors from the same work experience into a single vector. The learning module feeds the adjusted skill and work experience vectors into a fine-grained resume representation learning subnetwork to learn resume consistency features, and feeds the adjusted job requirement vector into a multi-scale job requirement representation learning subnetwork to learn multi-scale job features. The interaction module is used to process the learned... The obtained multi-scale job features are interacted with the resume representation through attention to obtain attention maps of resumes and job positions at different scales. The application module applies the obtained attention maps to the resume representation, transforming the attention-treated resume features into discriminative features based on the max pooling method, and concatenating the output of the last layer of the job position sub-network with it to obtain the final resume-job position pair representation. The construction module constructs a structure graph based on skill vectors, work experience vectors, and job requirement vectors, and extracts structural information from the constructed structure graph based on a selective graph attention network. The execution module fuses the extracted structural information and the final resume-job position pair representation through an adaptive fusion gate to obtain fused features, which are then fed into an MLP layer to obtain a matching score.

[0133] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

[0134] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

Claims

1. A job matching method based on the relevance learning of resume quality and structure, characterized in that, Specifically, the following steps are included: We obtain resumes and job requirements to build a dataset, and perform word segmentation and filtering on the text data in the dataset. We then vectorize the text data, obtaining skill vectors and work experience vectors from resumes, and job requirement vectors from job requirements. The obtained vectors are adjusted based on BiLSTM, and word attention mechanism is used to merge word vectors from the same work experience into a single vector. The adjusted skill vector and work experience vector are fed into the constructed fine-grained resume representation learning subnetwork to learn resume consistency features, and the adjusted job requirement vector is fed into the constructed multi-scale job representation learning subnetwork to learn multi-scale job features. The multi-scale job features obtained after learning are interacted with the resume representation through attention to obtain attention maps of resumes and job positions at different scales. The attention map is applied to the resume representation. Based on the max pooling method, the attention-enhanced resume features are transformed into discriminative features. The output of the last layer of the job sub-network is concatenated with it to obtain the final resume-job pair representation. A structure graph is constructed based on skill vectors, work experience vectors, and job requirement vectors, and structural information is extracted from the constructed structure graph using a selective graph attention network. An adaptive fusion gate is used to fuse the extracted structural information and the final resume-job pair representation to obtain fused features, which are then fed into an MLP layer to obtain a matching score.

2. The job matching method based on resume quality and structural relevance learning as described in claim 1, characterized in that: The fine-grained CV representation learning subnetwork includes a multi-head, noisy, Transformer-based common attention module with multiple repetitions, and a residual connection module to prevent gradient vanishing.

3. The job matching method based on resume quality and structure relevance learning as described in claim 2, characterized in that: The Transformer-based multi-head noise-included common attention module includes a Transformer encoder and a multi-head noise-included common attention module. The Transformer encoder is used as a feature extractor to extract features from the skill representation and the work experience representation, respectively. The multi-headed, noise-included common attention module is used to generate mutual attention maps between skill representations and work experience representations, and adds noise vectors to absorb inconsistent skills and work experience.

4. The job matching method based on resume quality and structural relevance learning as described in claim 1, characterized in that: The multi-scale job representation learning subnetwork is composed of multiple convolutional layers of different sizes stacked together, and the output of the multi-scale job representation learning subnetwork is job features at different scales.

5. The job matching method based on resume quality and structure relevance learning as described in claim 1, characterized in that, The multi-scale job features obtained after learning are interacted with the resume representation through attention to obtain attention maps of the resume and job at different scales. The obtained attention maps are represented as follows: H i =RJ i T H∈R (N+M)×K Where f represents the softmax function, H i Let be the similarity matrix between the resume representation and the job posting representation at the i-th scale. Let R represent the attention map at the i-th scale, and let J be the final representation of the resume. i T Let H represent the transpose of the output of the i-th layer of the multi-scale job description learning subnetwork, H represent the similarity matrix, N represent the number of job skills, M represent the number of job experiences, and K represent the number of words in the job description.

6. The job matching method based on resume quality and structural relevance learning as described in claim 1, characterized in that: The structure diagram constructed based on skill vectors, work experience vectors, and job requirement vectors includes four types of nodes: skill nodes, work experience nodes, job requirement nodes, and global nodes. For nodes in the structure graph: If nodes i and j have a contextual relationship, and the distance between the two words corresponding to nodes i and j is less than or equal to a set distance c, then the edge connecting these two words is a. ij =1; When nodes i and j are skill nodes or job nodes, and the two words corresponding to nodes i and j are the same, then the edge connecting these two words is a. ij =1; When nodes i and j are global nodes or work experience nodes, and the connecting edge a between the two words corresponding to nodes i and j is... ij =1.

7. The job matching method based on resume quality and structural relevance learning as described in claim 1, characterized in that: The selective graph attention network extracts structural information from the constructed structural graph, wherein, for the selective graph attention network: Selective graph attention network projects the input features of the (k+1)th layer of the graph attention network onto the information selection space through a projection matrix, and uses the sigmoid activation function to map them to the range of 0 to 1, thus obtaining a selection matrix β with selection function; The selection matrix β is used as an update gate to selectively fuse information from the current order and previous orders. 1-β controls the degree of retention of the previous order and β controls the degree of retention of the current order.

8. The job matching method based on resume quality and structure relevance learning as described in claim 1, characterized in that, The process involves fusing the extracted structural information and the final resume-job title pair representation using an adaptive fusion gate. Specifically: An adaptive fusion gate is used to combine the representation features h of the resume-job pair representation with the resume-job structure information X. s The features are spliced ​​together to obtain the fused feature X. m The calculation process is as follows: X m =h||X s 。 9. The job matching method based on resume quality and structural relevance learning as described in claim 1, characterized in that: The activation function of the first layer of the MLP layer is the ReLU function, and the activation function of the second layer is the sigmoid function. The matching score ranges from [0,1]; The specific process of sending the data to the MLP layer to obtain the matching score is as follows: in, represents the matching score, sigmoid represents the sigmoid activation function, ReLU represents the ReLU function, W4, W3, b3, and b4 represent the learnable parameters, and x represents the fused feature.

10. A job matching device based on resume quality and structural relevance learning, characterized in that, include: The acquisition module is used to acquire resumes and job requirements to build a dataset, and to perform word segmentation and filtering on the text data of the dataset, and to vectorize the text data, obtaining skill vectors and work experience vectors from resumes, and job requirement vectors from job requirements. The fusion module is used to adjust the obtained vectors based on BiLSTM and fuse word vectors from the same work experience into a single vector through a word attention mechanism. The learning module is used to feed the adjusted skill vector and work experience vector into a fine-grained resume representation learning subnetwork to learn resume consistency features, and to feed the adjusted job requirement vector into a multi-scale job representation learning subnetwork to learn multi-scale job features. The interaction module is used to interact with the multi-scale job features obtained after learning and the resume representation through attention, so as to obtain attention maps of resumes and job positions at different scales. The application module is used to apply the obtained attention map to the resume representation. Based on the max pooling method, the attention-enhanced resume features are transformed into discriminative features. The output of the last layer of the job sub-network is concatenated with it to obtain the final resume-job pair representation. The module is used to construct a structure graph based on skill vectors, work experience vectors, and job requirement vectors, and to extract structural information from the constructed structure graph based on a selective graph attention network. The execution module is used to fuse the extracted structural information and the final resume-job pair representation through an adaptive fusion gate to obtain fused features, which are then fed into the MLP layer to obtain a matching score.

Images