Resume and job matching prediction method and related equipment

Abstract

The present application provides a resume and job matching prediction method and related equipment. The resume and job matching prediction method and related equipment first obtain the semi-structured keys (key) and values ​​(value) and their sources (source) in the job information and resume information. Then, based on the keys and values ​​of each job information and resume information attribute and the corresponding source representation, a prediction model including a cascaded pre-trained language model, a Transformer encoder and a single-label classification model is used to predict the matching degree between the resume information and the job information. Thus, by comprehensively exploring the internal and external interactions of semi-structured multi-attributes in person-job matching, the matching results are made more accurate.

Description

Technical Field

[0001] The present application relates to label classification technology, and more particularly to a resume and job matching prediction method and related equipment. Background Art

[0002] Person-job matching is the core of online recruitment services, which is used to predict the matching degree between jobs and resumes. Most related person-job matching methods simplify this task to matching between free text attributes in jobs and resumes, thus achieving good performance. However, they ignore the contribution of semi-structured multivariate attributes such as education and salary, which may lead to unsuccessful prediction. Summary of the invention

[0003] In view of this, the purpose of this application is to propose a resume and job matching prediction method and related equipment.

[0004] Based on the above purpose, the present application provides a method for predicting the matching degree between resume information and job information, which is characterized by comprising:

[0005] Obtaining a first key and a first value of each semi-structured job attribute in the job information, and a second key and a second value of each semi-structured resume attribute in the resume information, wherein the first key, the first value, the second key and the second value are all represented by text data;

[0006] Based on the first key and the first value of each of the job attributes, the first source representation corresponding to the job information, the second key and the second value of each of the resume attributes, and the second source representation corresponding to the resume information, a prediction model including a cascaded pre-trained language model, a Transformer encoder and a single-label classification model is used to predict the matching degree between the resume information and the job information.

[0007] Furthermore, the predicting of the matching degree between the resume information and the job information includes:

[0008] For each of the first keys and the first values ​​of the job attributes, respectively encode the first keys and the first values ​​into the semantic space through the pre-trained language model to obtain a first key embedding and a first value embedding, and fuse the first key embedding and the first value embedding to obtain a first fused embedding of the job attribute;

[0009] encoding the first source representation into the semantic space through the pre-trained language model to obtain a first source embedding;

[0010] For each second key and second value of the resume attribute, respectively encode the second key and the second value into the semantic space through the pre-trained language model to obtain a second key embedding and a second value embedding, and fuse the second key embedding and the second value embedding to obtain a second fused embedding of the resume attribute;

[0011] encoding the second source representation into the semantic space through the pre-trained language model to obtain a second source embedding;

[0012] By means of the Transformer encoder, internal interaction of a first matrix including a first fused embedding of each of the job attributes is performed to obtain a first internal interaction attribute embedding matrix, and internal interaction of a second matrix including a second fused embedding of each of the resume attributes is performed to obtain a second internal interaction attribute embedding matrix;

[0013] fusing each element in the first internal interaction attribute embedding matrix with the first source embedding to obtain a first updated embedding matrix;

[0014] fusing each element in the second internal interaction attribute embedding matrix with the second source embedding to obtain a second updated embedding matrix;

[0015] Performing external interaction between the first updated embedding matrix and the second updated embedding matrix through the Transformer encoder to obtain a first external interaction attribute embedding matrix and a second external interaction attribute embedding matrix;

[0016] The single-label classification model is used to predict the degree of match between the resume information and the job information according to the first external interaction attribute embedding matrix and the second external interaction attribute embedding matrix.

[0017] Furthermore, the fusing of the first key embedding and the first value embedding, the fusing of the second key embedding and the second value embedding, the fusing of each element in the first internal interaction attribute embedding matrix with the first source embedding, and the fusing of each element in the second internal interaction attribute embedding matrix with the second source embedding all include additive fusion operations.

[0018] Furthermore, the fusing of the first key embedding and the first value embedding, the fusing of the second key embedding and the second value embedding, the fusing of each element in the first internal interaction attribute embedding matrix with the first source embedding, and the fusing of each element in the second internal interaction attribute embedding matrix with the second source embedding all include splicing and fusion operations.

[0019] Further, performing internal interaction of the first matrix to obtain the first internal interaction attribute embedding matrix includes: using the multi-head self-attention mechanism in the Transformer encoder to obtain a first multi-head self-attention matrix according to the first matrix; obtaining the first internal interaction attribute embedding matrix according to the first multi-head self-attention matrix through a first fully connected feedforward layer in the Transformer encoder;

[0020] Performing internal interaction of the second matrix to obtain the second internal interaction attribute embedding matrix includes: utilizing the multi-head self-attention mechanism in the Transformer encoder to obtain the second multi-head self-attention matrix according to the second matrix; and obtaining the second internal interaction attribute embedding matrix according to the second multi-head self-attention matrix through the second fully connected feedforward layer in the Transformer encoder.

[0021] Further, performing the external interaction to obtain the first external interaction attribute embedding matrix and the second external interaction attribute embedding matrix includes:

[0022] Concatenating the first updated embedding matrix with the second updated embedding matrix to obtain a joint matrix;

[0023] Utilizing the multi-head self-attention mechanism in the Transformer encoder, a multi-head self-attention joint matrix is ​​obtained according to the joint matrix;

[0024] The first external interaction attribute embedding matrix and the second external interaction attribute embedding matrix are obtained according to the multi-head self-attention joint matrix through the third fully connected feed-forward layer in the Transformer encoder.

[0025] Further, predicting the matching degree according to the first external interaction attribute embedding matrix and the second external interaction attribute embedding matrix includes:

[0026] Pooling the first external interaction attribute embedding matrix and the second external interaction attribute embedding matrix through a first pooling layer and a second pooling layer in the single-label classification model, respectively, to obtain a job vector and a resume vector;

[0027] Aggregate the position vector and the resume vector to obtain an aggregate vector;

[0028] The aggregate vector is semantically fused by a multi-layer perceptron in the single-label classification model, and a matching score is obtained by using a single neuron according to the semantically fused aggregate vector;

[0029] The activation function is used to obtain the matching degree between the resume information and the position information according to the matching score.

[0030] Furthermore, the pooling is performed using an adaptive average pooling function provided by the machine learning framework PyTorch.

[0031] Furthermore, the prediction model is trained by minimizing the binary cross entropy loss.

[0032] Based on the same concept, the present application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements any of the methods described above when executing the program.

[0033] As can be seen from the above, the resume and job matching prediction method and related equipment provided by this application first obtain the semi-structured keys and values ​​and their sources in the job information and resume information. Then, based on the keys and values ​​of each job information and resume information attribute and the corresponding source representation, a prediction model including a cascaded pre-trained language model, a Transformer encoder, and a single-label classification model is used to predict the matching degree between the resume information and the job information. Thus, by comprehensively exploring the internal and external interactions of semi-structured multi-attributes in person-job matching, the matching results are made more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the present application or related technologies, the drawings required for use in the embodiments or related technical descriptions are briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0035] Figure 1 This is a schematic diagram of the position and resume information of the embodiment of this application;

[0036] Figure 2 A schematic diagram of a method flow chart for predicting the matching degree between resumes and positions according to an embodiment of the present application;

[0037] Figure 3 A schematic diagram of a matching degree prediction method flow chart of an embodiment of the present application;

[0038] Figure 4 A schematic diagram of the external interaction attribute embedding matrix generation process of an embodiment of the present application;

[0039] Figure 5 A schematic diagram of a flow chart showing the matching degree between job information and resume information in an embodiment of the present application;

[0040] Figure 6 A schematic diagram of multiple attributes in positions and resumes in an embodiment of the present application;

[0041] Figure 7 A schematic diagram of a device for predicting the matching degree between a resume and a position according to an embodiment of the present application;

[0042] Figure 8 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0043] In order to make the objectives, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below in combination with specific embodiments and with reference to the accompanying drawings.

[0044] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the embodiments of the present application should be the usual meanings understood by people with ordinary skills in the field to which the present application belongs. The "first", "second" and similar words used in the embodiments of the present application do not represent any order, quantity or importance, but are only used to distinguish different components. "Including" or "comprising" and similar words mean that the elements or objects appearing in front of the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0045] As described in the background technology section, existing job matching solutions are still difficult to meet the needs of efficient and accurate job matching. In the process of implementing this application, the applicant found that existing job matching solutions ignore the fact that in the real world, jobs and resumes are usually composed of semi-structured multi-attributes (also called features), such as education, city, salary, etc. Each attribute consists of a key (also called a name) and a value (also called content). The lack of consideration of some key attributes in the matching process will lead to inaccurate prediction results. For example, reference Figure 1, when the person-job matching algorithm only considers the job description and work experience, it may mistakenly believe that there is a high match between the job and the resume, even if the job city in the job ("Shanghai") and the talent's desired job city ("Beijing") obviously do not match. The related solution is to simply extract entities from the job and resume, and then fuse them into the representation of free text to enhance the missing semantics. The source of the extracted entities is not fully considered, and these entities are undoubtedly crucial for matching jobs and resumes. Specifically, for example, in Figure 1 In the example, the entity "Shanghai" in the city attribute of the job position and the entity "Shanghai" in the job seeker's work experience are obviously not treated equally when matching jobs and resumes.

[0046] In addition, unlike in typical text matching tasks, where the sequential features and local interactions of text can be well modeled by LSTM and CNN, the grammatical and contextual features of semi-structured attributes of positions or resumes cannot be well modeled because there is no clear relationship between the multiple attributes. Therefore, it is necessary to model the complex relationships between the multiple attributes in positions or resumes (i.e., internal interactions). At the same time, the external interactions between the free text attributes of positions and resumes are also important factors affecting the matching results.

[0047] In view of this, one or more embodiments of this specification provide a resume and job matching prediction scheme, first obtaining the semi-structured keys (key) and values ​​(value) and their sources (source) in the job information and resume information. Then, based on the keys and values ​​of each job information and resume information attribute and the corresponding source representation, a prediction model including a cascaded pre-trained language model, a Transformer encoder and a single-label classification model is used to predict the matching degree between the resume information and the job information. Thus, by comprehensively exploring the internal and external interactions of semi-structured multi-attributes in person-job matching, the matching results are made more accurate. The technical solutions of each specific embodiment of the present application are described below.

[0048] refer to Figure 2 A resume and job matching prediction method according to an embodiment of the present specification includes the following steps:

[0049] Step S201, obtaining a first key and a first value of each semi-structured job attribute in the job information, and a second key and a second value of each semi-structured resume attribute in the resume information, wherein the first key, the first value, the second key and the second value are all represented by text data;

[0050] In this embodiment, in order to solve the problem of person-job matching, it is necessary to predict the degree of match between the job information released by the recruiter and the resume information uploaded by the job seeker. Generally speaking, in online recruitment services, person-job matching data consists of three parts, namely, the job, the resume, and the interaction record between the job and the resume. A job consists of a job profile (for example, job title, city, and salary) and a job description (for example, requirements for skills or experience). A resume consists of a job seeker's personal profile (such as age, gender, and education), as well as work experience (such as project experience in previous / current companies). An interaction record includes the IDs of a job and a resume and their corresponding matching degrees.

[0051] In this step, the job information and resume information are composed of semi-structured multi-attributes, where each attribute consists of an attribute key and an attribute value. The attribute key is usually described by text data; while the attribute value can be text, category or number, etc. The attribute key and attribute value follow a certain subordinate relationship and can be easily converted into text data. For example, in a job, the attribute key can be "job name", "minimum salary", "whether to go on a business trip", etc., and their corresponding attribute values ​​are given in the form of text, number and category, for example, "marketing manager", "1000" and "yes". We can easily convert them into "marketing manager", "1000RMB" and "need to go on a business trip" through manual operation. The attribute source is used to indicate the source of the attribute, which indicates whether the attribute comes from the job information or the resume information.

[0052] Step S202: Based on the first key and the first value of each of the job attributes, the first source representation corresponding to the job information, the second key and the second value of each of the resume attributes, and the second source representation corresponding to the resume information, a prediction model including a cascaded pre-trained language model, a Transformer encoder, and a single-label classification model is used to predict the degree of match between the resume information and the job information.

[0053] In this embodiment, by processing the semi-structured attribute keys, attribute values, and attribute sources in the job information and resume information, the matching degree between the resume information and the job information is finally obtained. The relationship between the semi-structured attributes in the job information and resume information is considered, making the matching result more accurate.

[0054] In some other embodiments, reference Figure 3 , the prediction of the matching degree between the resume information and the position information in the above embodiment includes:

[0055] Step S301: for each of the first keys and the first values ​​of the job attributes, respectively encode the first keys and the first values ​​into the semantic space through the pre-trained language model to obtain a first key embedding and a first value embedding, and fuse the first key embedding and the first value embedding to obtain a first fused embedding of the job attribute;

[0056] Step S302: encoding the first source representation into the semantic space through the pre-trained language model to obtain a first source embedding;

[0057] Step S303: for each second key and second value of the resume attribute, respectively encode the second key and the second value into the semantic space through the pre-trained language model to obtain a second key embedding and a second value embedding, and fuse the second key embedding and the second value embedding to obtain a second fused embedding of the resume attribute;

[0058] Step S304: encoding the second source representation into the semantic space through the pre-trained language model to obtain a second source embedding;

[0059] In this embodiment, for obtaining job information and resume information in the aforementioned embodiment, it also includes encoding the sub-parts of each attribute in the job information and resume information (i.e., attribute key, attribute value, and attribute source) into the same semantic space to unify the representation of the sub-parts of each attribute.

[0060] As a specific example, we can use and Symbols are used to represent the job information and resume information that appear in an interaction record:

[0061]

[0062]

[0063] Among them, j i Representative position The i-th attribute in , i∈{1, 2, …, m}. j Resume The jth attribute in , j∈{1,2,…,n}; m and n are and The corresponding number of attributes in The i-th attribute j i The keys and values ​​are represented as and CV The properties in and so on. In addition Used to indicate job position and resume The binary label of the true match result between . That is, 0 means no match (negative example) and 1 means match (positive example).

[0064] Based on the above notation, the person-job matching task can be expressed as learning a prediction function from existing job-resume interaction records in order to accurately predict the matching degree of unseen job-resume pairs. The common process is to build a person-job matching algorithm, encode the job and resume into a latent vector, and then apply a classification function to predict the matching degree. The above process can be expressed as:

[0065]

[0066] in It is a job and resume The predicted match between . is the prediction function, whose purpose is to make As close as possible to the true matching tag

[0067] In this embodiment, the sub-parts of each attribute in the job information and resume information are encoded by the pre-trained language model BERT, and the formula is as follows:

[0068]

[0069]

[0070]

[0071]

[0072] in, yes The i-th attribute and The attribute key and attribute value corresponding to the jth attribute in , i∈{1,2,…,m} and j∈{1,2,…,n}. em is the dimension of the pre-set BERT output. Similarly, the source encoding process is as follows:

[0073] j s =BERT(j s ),

[0074] r s =BERT(r s ),

[0075] in, They are and In this embodiment, the hidden state and pooling operation of the last layer of BERT are used to obtain the output attribute key, attribute value and attribute source.

[0076] In this embodiment, the obtained attribute keys and attribute values ​​need to establish corresponding connections, and each attribute key and the corresponding attribute value features are fused to obtain the embedding after all features are fused. The first fused embedding corresponds to the position, and the second fused embedding corresponds to the resume. Each element in the embedding corresponds to a specific attribute.

[0077] Specifically, and The attribute value and attribute key of each attribute in are fused and embedded as follows:

[0078]

[0079]

[0080] in, They are The i-th attribute and The fused embedded expression of the j-th attribute in .

[0081] In the right and After the above fusion embedding operation is performed on all attributes in , we get and The matrix representation of is as follows:

[0082]

[0083]

[0084] in, and They are and Matrix representation of (ie, the first matrix and the second matrix).

[0085] Step S305: performing internal interaction of a first matrix including a first fused embedding of each of the job attributes through the Transformer encoder to obtain a first internal interaction attribute embedding matrix, and performing internal interaction of a second matrix including a second fused embedding of each of the resume attributes to obtain a second internal interaction attribute embedding matrix;

[0086] In this step, in order to capture the relationship between multiple attributes, internal interaction of the first matrix is ​​performed to obtain the first internal interaction attribute embedding matrix, including: utilizing the multi-head self-attention mechanism in the Transformer encoder to obtain the first multi-head self-attention matrix according to the first matrix; and obtaining the first internal interaction attribute embedding matrix according to the first multi-head self-attention matrix through the first fully connected feedforward layer in the Transformer encoder.

[0087] Performing internal interaction of the second matrix to obtain the second internal interaction attribute embedding matrix includes: utilizing the multi-head self-attention mechanism in the Transformer encoder to obtain the second multi-head self-attention matrix according to the second matrix; and obtaining the second internal interaction attribute embedding matrix according to the second multi-head self-attention matrix through the second fully connected feedforward layer in the Transformer encoder.

[0088] As a specific example:

[0089]

[0090]

[0091] M J =concat(M J 1 , M J 2 , …, M J H )·O J ,

[0092] M R =concat(M R 1 , M R 2 ,…,M R H )·O R ,

[0093] in, and is a trainable network parameter; h∈{1,2,…,H}, where H represents the number of heads of the Transformer encoder. In this embodiment, d q =d k =d v =d in / H.M J h and M R h yes and The self-attention matrix representation of the h-th head. M J and M R yes and The multi-head self-attention matrix representation of (i.e. the multi-head self-attention matrix of the position and the multi-head self-attention matrix of the resume).

[0094] After that, the multi-head self-attention matrix of the position and the multi-head self-attention matrix of the resume are input into the feed-forward layer of the Transformer encoder to obtain and The internal interaction representation (i.e., the internal interaction attribute embedding matrix of the position and the internal interaction attribute embedding matrix of the resume) is as follows:

[0095]

[0096]

[0097] in, They are The i-th attribute and The internal interaction attribute embedding of the j-th attribute in , i∈{1,2,…,m} and j∈{1,2,…,n}.

[0098] Step S306: fusing each element in the first internal interaction attribute embedding matrix with the first source embedding to obtain a first updated embedding matrix;

[0099] Step S307: fusing each element in the second internal interaction attribute embedding matrix with the second source embedding to obtain a second updated embedding matrix;

[0100] In this step, we mainly focus on the multi-attribute interaction between positions and resumes. We further merge the attribute sources into the attributes that have undergone internal interaction to indicate the source of the attributes and generate updated matrix representations for positions and resumes. The acquired attributes and attribute sources need to establish corresponding connections, and each attribute source and the corresponding attribute feature are merged and embedded to obtain the first updated embedding matrix corresponding to the position and the second updated embedding matrix corresponding to the resume. Each element in the matrix corresponds to a specific attribute.

[0101] Specifically, it can be expressed as:

[0102]

[0103]

[0104]

[0105]

[0106] in, They are The i-th attribute and The representation of the attribute source features of the j-th attribute in .

[0107] Step S308: performing external interaction between the first updated embedding matrix and the second updated embedding matrix through the Transformer encoder to obtain a first external interaction attribute embedding matrix and a second external interaction attribute embedding matrix;

[0108] In this step, when modeling external interactions, we need to use the multi-head self-attention mechanism and feedforward neural network in the Transformer encoder to obtain the first external interaction attribute embedding matrix of the position and the second external interaction attribute embedding matrix of the resume. For details, refer to Figure 4 , which may also include the following steps:

[0109] Step S401, concatenating the first updated embedding matrix and the second updated embedding matrix to obtain a joint matrix;

[0110] Step S402, using the multi-head self-attention mechanism in the Transformer encoder to obtain a multi-head self-attention joint matrix according to the joint matrix;

[0111] Step S403: obtaining the first external interaction attribute embedding matrix and the second external interaction attribute embedding matrix according to the multi-head self-attention joint matrix through the third fully connected feed-forward layer in the Transformer encoder.

[0112] Specifically, and The second position attribute matrix M J′ and the second resume attribute matrix M R′ Merge into joint matrix X E :

[0113]

[0114] Then a multi-head self-attention layer is applied to capture and The multi-attribute relationship between them is obtained by inputting the joint matrix into the self-attention layer of the Transformer encoder to obtain the multi-head self-attention joint matrix as follows:

[0115]

[0116] M E =concat(M E 1 , ME 2 , …, M E H )·O E ,

[0117] in, and is a trainable network parameter; h∈{1, 2, …, H}, where H represents the number of heads of the Transformer encoder. Set d q =d k =d v =d ex / H. After that, M E h yes and The self-attention joint representation in the h-th head. M E yes and The joint matrix of multi-head self-attention.

[0118] Finally, to M E Using a fully connected feed-forward layer, we get and The external interaction representation is input into the third fully connected feed-forward layer of the Transformer encoder, and the external interaction attribute embedding matrix of the position and the external interaction attribute embedding matrix of the resume are obtained, as shown below:

[0119]

[0120]

[0121]

[0122] in, and yes The i-th attribute and The external interaction attribute embedding representation of the jth attribute in M, i∈{1, 2, …, m} and j∈{1, 2, …, n}. Finally, M J″ and M R″ yes and The respective external interaction attributes are embedded in the matrix.

[0123] Step S309: using the single-label classification model, according to the first external interaction attribute embedding matrix and the second external interaction attribute embedding matrix, predict the matching degree between the resume information and the job information.

[0124] In this step, after exploring the internal and external interactions of semi-structured multivariate attributes, a rich representation of positions and resumes is obtained. By analyzing the external interaction attribute embedding matrix of positions and resumes, the results of person-job matching can be obtained.

[0125] As can be seen from the above, the embodiment of the present application first obtains the keys and values ​​of the job information and resume information and their sources. Then, the internal interactions between the keys and values ​​of the positions and resumes and the external interactions between the sources and keys and values ​​between the positions and resumes are hierarchically modeled. The key embedding and source embedding are respectively integrated into the value embedding to clearly indicate the keys and sources corresponding to the values ​​in the attributes. Finally, the matching degree between the positions and resumes is predicted based on the external interaction attribute embedding matrix of the positions and resumes. Thus, by comprehensively exploring the internal and external interactions of semi-structured multi-attributes in person-job matching, the matching results are made more accurate.

[0126] In some other embodiments, the fusion of the first key embedding and the first value embedding, the fusion of the second key embedding and the second value embedding, the fusing of each element in the first internal interaction attribute embedding matrix with the first source embedding, and the fusing of each element in the second internal interaction attribute embedding matrix with the second source embedding all include additive fusion operations.

[0127] In some other embodiments, the fusion of the first key embedding and the first value embedding, the fusion of the second key embedding and the second value embedding, the fusion of each element in the first internal interaction attribute embedding matrix with the first source embedding, and the fusion of each element in the second internal interaction attribute embedding matrix with the second source embedding all include splicing fusion operations.

[0128] In the present application, the feature fusion embedding of attribute keys and attribute values ​​can be performed by additive fusion or splicing fusion. In additive fusion, the number of elements in the obtained matrix is ​​the sum of the number of attribute keys and the number of attribute values. In splicing fusion, the number of elements in the obtained matrix is ​​the same as the number of attribute keys.

[0129] Specifically, the fusion embedding process can be expressed as:

[0130]

[0131]

[0132] in, You can perform addition or splicing fusion operations based on the actual experimental results. They are The i-th attribute and The fused expression of the jth attribute in . inCan be d em or 2·d em , depends on the fusion strategy adopted. When adding fusion, d in is 2·d em , when splicing and fusion, d in Yes em .

[0133] In some other embodiments, reference Figure 5 In the aforementioned embodiment, predicting the matching degree according to the first external interaction attribute embedding matrix and the second external interaction attribute embedding matrix includes:

[0134] Step S501: pooling the first external interaction attribute embedding matrix and the second external interaction attribute embedding matrix through the first pooling layer and the second pooling layer in the single label classification model respectively to obtain a job vector and a resume vector;

[0135] Step S502, aggregating the position vector and the resume vector to obtain an aggregate vector;

[0136] Step S503: semantically fuse the aggregated vectors through a multi-layer perceptron in the single-label classification model, and obtain a matching score using a single neuron according to the aggregated vectors after semantic fusion;

[0137] Step S504: using an activation function to obtain a matching degree between the resume information and the job information according to the matching score.

[0138] In this embodiment, a pooling layer is first used to convert the matrix representations of positions and resumes into vector representations, aggregate the vectors of positions and resumes, and then send the aggregated vectors to a multi-layer perceptron (MLP) for further semantic fusion. A single neuron can be used to output a matching score, and finally an activation function is used to obtain a matching result based on the matching score. Specifically, the matching score is mapped to a range of [0,1]. According to actual needs, positions and resumes within a preset range can be selected to output a positive or negative result. For example, the matching result within [0,0.5] can be selected as a matching failure, and 0 is output, and the matching result within (0.5,1] can be selected as a matching success, and 1 is output. That is, 0 represents a mismatch (negative example), and 1 represents a match (positive example).

[0139] As a specific example, we first use a pooling layer to and The matrix representations are converted into vector representations as follows:

[0140] J = pooling(M J″ ),

[0141] R=pooling(M R″ ),

[0142] Among them, J and R are and In this embodiment, the adaptive average pooling function provided by the open source machine learning framework PyTorch is used.

[0143] Next, aggregate and The aggregated vector is then fed into a multi-layer perceptron (MLP) for further semantic fusion. Then, a single neuron is used to output the matching score as the predicted label. In this embodiment, the label prediction process is formulated as follows:

[0144] g JR =concat(J,R,JR)

[0145]

[0146] Among them, g JR yes and Aggregate vector; MLP in this embodiment refers to a multilayer perceptron with two nonlinear layers and one ReLu layer; sigmoid is used to map the matching score to the range of [0,1]. Specifically, the matching score is mapped to the range of [0,1] by sigmoid. According to actual needs, you can select positions and resumes within the preset range to output a positive or negative result. Generally speaking, the prediction threshold is set to 0.5 by default. If the predicted match of a sample exceeds 0.5, it can be classified as a positive sample, that is, the current position matches the current resume. Otherwise, it is a negative sample.

[0147] In some other embodiments, the pooling in the aforementioned embodiments is performed using the adaptive average pooling function provided by the machine learning framework PyTorch.

[0148] In some other embodiments, the prediction model of the aforementioned embodiments is trained by minimizing the binary cross entropy loss.

[0149] In some other embodiments, in order to make the output results of the model constructed according to the resume and job matching prediction method of the present application (internal and external interaction (Interaction) model, referred to as InEXIT model) more accurate, the model needs to be trained. In the training stage of the model, each training instance consists of a job A resume and their corresponding matching In this embodiment, a mini-batch strategy is used to randomly extract training instances from the training set. The parameters of the model are trained by minimizing the following binary cross entropy loss to obtain the optimal training round:

[0150]

[0151] in, is the true matching degree of the i-th training instance; is the predicted matching degree of the ith training instance generated by the model; N is the total number of training instances. In this embodiment, the model is trained by using exponential decay to change the learning rate of each training round to avoid overfitting.

[0152] As a specific example, a real person-job matching dataset with semi-structured multivariate attributes was constructed from a public dataset provided by a person-job intelligent matching competition. In all the raw data, the data was first cleaned by removing those positions and resumes with incomplete attributes. Then, the interaction histories that did not clearly signal whether the current position matched the position were filtered out. It is worth noting that in the raw data, there are many more instances of mismatches (negative examples) than matching instances (positive examples). In order to reduce the bias of the data, negative examples with the same number as positive examples were randomly extracted. Finally, 80% of the useful interaction histories were randomly divided into training, 10% for validation, and the remaining 10% for testing. In order to protect the privacy of talents and companies, all interaction histories were anonymized and identity attributes such as gender, name, address, etc. were removed. The statistics of the processed dataset are summarized in Table 1. As shown in Table 1, it is worth mentioning that for a position and resume, there are 11 and 12 attributes, respectively, and each attribute contains 28.20 and 14.74 words on average. In general, a large number of attributes may bring greater challenges to the person-job matching method in predicting the accurate match between positions and resumes. In addition, Figure 6 The multi-attributes in positions and resumes are described in detail, including different attribute sources, namely the attribute key descriptions in positions and resumes, and the corresponding value type of each attribute key, which can be numeric, text, and category.

[0153] Table 1: Dataset statistics

[0154]

[0155] Next, we compare the proposed model with a competitive job matching baseline model to verify the effectiveness of the model constructed by the job matching method proposed in this application. The baseline and the proposed model are summarized as follows:

[0156] PJFNN: A CNN-based model in which two CNNs are used to explore the semantic representation of job descriptions and work experience. The matching degree is calculated by cosine similarity.

[0157] BPJFNN: A RNN-based model that uses BiLSTM instead of CNN as a feature extractor. In addition, it also uses an aggregation layer to predict the match between positions and resumes.

[0158] JRMPM: It adopts a content memory module to explore the preferences of recruiters and job seekers, which are then used to enrich their representations.

[0159] APJFNN: An improved version of BPJFNN, it proposes a hierarchical recursive attention model to comprehensively learn competency-aware representations of positions and resumes.

[0160] MV-CoN: It consists of a text-based and relation-based matching component followed by a co-teaching mechanism, which aims to learn matching from sparse and noisy JRM interaction data.

[0161] InEXIT: The model established by the person-job matching method proposed in this application explores internal and external interactions, and focuses on comprehensive modeling of semi-structured multivariate attributes in person-job matching.

[0162] In this specific example, all validation experiments were implemented in Python and the open source machine learning framework PyTorch and run on a server equipped with an NVIDIA RTX3090 GPU graphics card. The sub-part embeddings of the attributes are initialized with the pre-trained BERT-Base-Chinese language model, with a dimension of d em =768. The hidden layer size and number of heads of the multi-head self-attention mechanism are set to 768 and 8 respectively. In order to reduce the complexity of the model, this embodiment pads / truncates the attribute sub-parts of different lengths. Specifically, the length of the values ​​of job description and work experience is set to 256; the length of the remaining sub-parts is set to 16.

[0163] When optimizing model parameters, the Adaptive Moment (Adam) optimizer is used in this embodiment to train the model for up to 20 rounds. An early stop mechanism is used to determine the best rounds in the training phase. If the ACC indicator on the validation set does not increase for 5 consecutive rounds, the training will be terminated early. Layer normalization and gradient clipping mechanisms are also used to avoid overfitting. The batch size is 16 and the dropout rate is 0.1. The initial learning rate is set to 4*10-6 , the decay rate of the learning rate is 1*10 -9 As for the fusion strategy, it is experimentally observed that the best performance is achieved when the addition operation is adopted in the step-by-step fusion mechanism.

[0164] Refer to Table 2, which shows the performance of the aforementioned existing baseline models and InEXIT in terms of AUC, ACC, Precision, Recall and F1 scores, where the results of the best baseline and best model in each column are underlined and bolded, respectively.

[0165] Table 2

[0166]

[0167] Specifically, InEXIT improves the AUC, ACC, Precision, Recall, and F1 indicators by 4.28%, 4.10%, 4.80%, 2.34%, and 3.56% over the best baseline Mv-CoN, respectively. The difference between the prediction results of InEXIT and the best baseline Mv-CoN of this embodiment was detected to be statistically significant by a two-tailed paired t-test at the 0.01 level. It is worth noting that InEXIT has the highest performance improvement (4.80%) in the Precision indicator, while the improvement over the best baseline Mv-CoN in the Recall indicator is the lowest (2.34%). This difference indicates that more positive examples are mistakenly classified as negative examples than the number of negative examples that are mistakenly classified as positive examples. This may be due to the fact that too many attributes are considered, which introduces additional noise in the matching pattern, making it more difficult to accept a match between a position and a resume than to reject it. The phenomenon observed in the overall performance proves that considering semi-structured multivariate attributes can be beneficial to job matching to a certain extent. The person-job matching method of the present application has certain advantages over the current related methods.

[0168] A specific implementation scenario is given below, and the job information is:

[0169] Job title: Operations Manager Salary: 6001-8000

[0170] Whether on business trip: Yes Working city: C530

[0171] Number of recruits: 3 Minimum working years: 3 years

[0172] Position Type: None Minimum Education Requirement: Bachelor's Degree

[0173] Recruitment start time: 2019.02.19 Recruitment end time: 2019.03.31

[0174] Job Description

[0175] 1. Assist the Operation Manager to complete the department's planning tasks.

[0176] 2. Guide and supervise the operation assistants to carry out various tasks, and conduct regular training and assessment

[0177] 3. Responsible for the daily operations of the area under jurisdiction.

[0178] 4. Maintain good cooperative relations with merchants, and follow up and return visits to repairs and complaints in a timely manner.

[0179] 5. Responsible for inspecting and supervising the service standards and working conditions of staff.

[0180] 6. Inspect stores regularly and resolve problems promptly, such as fire hazards, safety facility failures, etc.

[0181] 7. Responsible for collecting rent, property fees, energy fees and other related fees during the lease period.

[0182] The resume information for the application is:

[0183] Expected salary: 10001-15000 Current salary: 10001-15000

[0184] Expected working city: C530 Current living city: C530

[0185] Started working: 2015 Age: 36

[0186] Education: College Current Position Type: None

[0187] Expected job industry: property management, construction engineering Current job industry: commercial center property management Expected job type: property management, production management, operation

[0188] Work Experience

[0189] 1. Worked in customer service, negotiated business with merchants, handled consumer complaints, and was good at maintaining relationships with the public.

[0190] 2. Worked as a property manager, assisting the operations department in property management.

[0191] 3. Good at employee management, training, assessment and team building.

[0192] 4. Previously responsible for fire inspection, public safety, supervision and inspection of public facilities.

[0193] 5. Was responsible for the daily operation of the mall, presided over various affairs, and kept the documents and resources department.

[0194] In this example, the matching results predicted by the best baseline model Mv-CoN and InEXIT proposed in the embodiment of the present application are: Mv-CoN: 0.58, InEXIT: 0.17. The corresponding real matching result is 0, that is, no match.

[0195] The best baseline Mv-CoN gives a predicted match of 0.58 between the current position and the resume, which indicates a positive match at a generally accepted threshold (i.e., 0.5). After browsing the job description and work experience, which the Mv-CoN model mainly relies on, it can be found that the match between the position and the resume is very high. For example, in the job description, the recruiter requires the job seeker to be responsible for "daily operations." In the work experience, the job seeker shows the corresponding ability to make himself a suitable candidate, that is, "responsible for the daily operation of the mall." In addition, other overlapping parts between the job description and work experience, such as "training," "assessment," "complaints," and "safety," also provide strong support for Mv-CoN to classify the position and resume as a positive match.

[0196] However, compared with the best baseline model, the match between the job and the resume predicted by InEXIT proposed in this application is 0.17, which is close to the true match of 0 under the generally accepted threshold. The difference between Mv-CoN and InEXIT can be attributed to the fact that InEXIT considers the entire semi-structured multivariate attributes, not just the free text attributes, namely, job description and work experience. In this case, the recruiter requires the job seeker to have a minimum education level of "bachelor's degree" and can only offer a salary of 6,001 to 8,000. However, the job seeker requires a salary that is difficult to meet, i.e., from 10,001 to 15,000, and the educational level is not up to standard, i.e., "junior college". Given the above observations, InEXIT is likely to classify the current job and resume as a mismatched negative example. Qualitative analysis of real cases shows that the InEXIT model proposed in this application can make good use of semi-structured multivariate attributes to comprehensively match jobs and resumes in online recruitment services.

[0197] It should be noted that the method of the embodiment of the present application can be performed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario and completed by multiple devices cooperating with each other. In the case of such a distributed scenario, one of the multiple devices can only perform one or more steps in the method of the embodiment of the present application, and the multiple devices will interact with each other to complete the described method.

[0198] It should be noted that the above describes some embodiments of the present application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the above embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0199] Based on the same inventive concept, corresponding to any of the above-mentioned embodiment methods, the present application also provides a resume and job matching prediction device.

[0200] refer to Figure 7 The resume and job matching prediction device comprises:

[0201] The acquisition module 701 is configured to acquire a first key and a first value of each semi-structured job attribute in the job information, and a second key and a second value of each semi-structured resume attribute in the resume information, wherein the first key, the first value, the second key and the second value are all represented by text data;

[0202] The prediction module 702 is configured to predict the matching degree between the resume information and the job information based on the first key and the first value of each of the job attributes, the first source representation corresponding to the job information, the second key and the second value of each of the resume attributes, and the second source representation corresponding to the resume information, through a prediction model including a cascaded pre-trained language model, a Transformer encoder and a single-label classification model.

[0203] In some other embodiments, the prediction module 702 predicts the matching degree between the resume information and the job information specifically including:

[0204] For each of the first keys and the first values ​​of the job attributes, respectively encode the first keys and the first values ​​into the semantic space through the pre-trained language model to obtain a first key embedding and a first value embedding, and fuse the first key embedding and the first value embedding to obtain a first fused embedding of the job attribute;

[0205] encoding the first source representation into the semantic space through the pre-trained language model to obtain a first source embedding;

[0206] For each second key and second value of the resume attribute, respectively encode the second key and the second value into the semantic space through the pre-trained language model to obtain a second key embedding and a second value embedding, and fuse the second key embedding and the second value embedding to obtain a second fused embedding of the resume attribute;

[0207] encoding the second source representation into the semantic space through the pre-trained language model to obtain a second source embedding;

[0208] By means of the Transformer encoder, internal interaction of a first matrix including a first fused embedding of each of the job attributes is performed to obtain a first internal interaction attribute embedding matrix, and internal interaction of a second matrix including a second fused embedding of each of the resume attributes is performed to obtain a second internal interaction attribute embedding matrix;

[0209] fusing each element in the first internal interaction attribute embedding matrix with the first source embedding to obtain a first updated embedding matrix;

[0210] fusing each element in the second internal interaction attribute embedding matrix with the second source embedding to obtain a second updated embedding matrix;

[0211] Performing external interaction between the first updated embedding matrix and the second updated embedding matrix through the Transformer encoder to obtain a first external interaction attribute embedding matrix and a second external interaction attribute embedding matrix;

[0212] The single-label classification model is used to predict the degree of match between the resume information and the job information according to the first external interaction attribute embedding matrix and the second external interaction attribute embedding matrix.

[0213] Furthermore, the fusing of the first key embedding and the first value embedding, the fusing of the second key embedding and the second value embedding, the fusing of each element in the first internal interaction attribute embedding matrix with the first source embedding, and the fusing of each element in the second internal interaction attribute embedding matrix with the second source embedding all include additive fusion operations.

[0214] Furthermore, the fusing of the first key embedding and the first value embedding, the fusing of the second key embedding and the second value embedding, the fusing of each element in the first internal interaction attribute embedding matrix with the first source embedding, and the fusing of each element in the second internal interaction attribute embedding matrix with the second source embedding all include splicing and fusion operations.

[0215] Further, performing internal interaction of the first matrix to obtain the first internal interaction attribute embedding matrix includes: using the multi-head self-attention mechanism in the Transformer encoder to obtain a first multi-head self-attention matrix according to the first matrix; obtaining the first internal interaction attribute embedding matrix according to the first multi-head self-attention matrix through a first fully connected feedforward layer in the Transformer encoder;

[0216] Performing internal interaction of the second matrix to obtain the second internal interaction attribute embedding matrix includes: utilizing the multi-head self-attention mechanism in the Transformer encoder to obtain the second multi-head self-attention matrix according to the second matrix; and obtaining the second internal interaction attribute embedding matrix according to the second multi-head self-attention matrix through the second fully connected feedforward layer in the Transformer encoder.

[0217] Further, performing the external interaction to obtain the first external interaction attribute embedding matrix and the second external interaction attribute embedding matrix includes:

[0218] Concatenating the first updated embedding matrix with the second updated embedding matrix to obtain a joint matrix;

[0219] Utilizing the multi-head self-attention mechanism in the Transformer encoder, a multi-head self-attention joint matrix is ​​obtained according to the joint matrix;

[0220] The first external interaction attribute embedding matrix and the second external interaction attribute embedding matrix are obtained according to the multi-head self-attention joint matrix through the third fully connected feed-forward layer in the Transformer encoder.

[0221] Further, predicting the matching degree according to the first external interaction attribute embedding matrix and the second external interaction attribute embedding matrix includes:

[0222] Pooling the first external interaction attribute embedding matrix and the second external interaction attribute embedding matrix through a first pooling layer and a second pooling layer in the single-label classification model, respectively, to obtain a job vector and a resume vector;

[0223] Aggregate the position vector and the resume vector to obtain an aggregate vector;

[0224] The aggregate vector is semantically fused by a multi-layer perceptron in the single-label classification model, and a matching score is obtained by using a single neuron according to the semantically fused aggregate vector;

[0225] The activation function is used to obtain the matching degree between the resume information and the position information according to the matching score.

[0226] Furthermore, the pooling is performed using an adaptive average pooling function provided by the machine learning framework PyTorch.

[0227] Furthermore, the prediction model is trained by minimizing the binary cross entropy loss.

[0228] For the convenience of description, the above device is described in terms of functions divided into various modules. Of course, when implementing the present application, the functions of each module can be implemented in the same or multiple software and / or hardware.

[0229] The device of the above embodiment is used to implement the corresponding person-job matching method in any of the above embodiments, and has the beneficial effects of the corresponding method embodiment, which will not be repeated here.

[0230] Based on the same inventive concept, corresponding to any of the above-mentioned embodiments, the present application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the resume and job matching prediction method described in any of the above embodiments is implemented.

[0231] Figure 8 A more specific schematic diagram of the hardware structure of an electronic device provided in this embodiment is shown, and the device may include: a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040 are connected to each other through the bus 1050 in the device.

[0232] The processor 1010 can be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0233] The memory 1020 may be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 may store an operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program codes are stored in the memory 1020 and are called and executed by the processor 1010.

[0234] The input / output interface 1030 is used to connect the input / output module to realize information input and output. The input / output module can be configured in the device as a component (not shown in the figure), or it can be externally connected to the device to provide corresponding functions. The input device may include a keyboard, a mouse, a touch screen, a microphone, various sensors, etc., and the output device may include a display, a speaker, a vibrator, an indicator light, etc.

[0235] The communication interface 1040 is used to connect a communication module (not shown) to realize communication interaction between the device and other devices. The communication module can realize communication through a wired mode (such as USB, network cable, etc.) or a wireless mode (such as mobile network, WIFI, Bluetooth, etc.).

[0236] The bus 1050 includes a path that transmits information between the various components of the device (eg, the processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040).

[0237] It should be noted that, although the above device only shows the processor 1010, the memory 1020, the input / output interface 1030, the communication interface 1040 and the bus 1050, in the specific implementation process, the device may also include other components necessary for normal operation. In addition, it can be understood by those skilled in the art that the above device may also only include the components necessary for implementing the embodiments of the present specification, and does not necessarily include all the components shown in the figure.

[0238] The electronic device of the above embodiment is used to implement the corresponding resume and job matching prediction method in any of the above embodiments, and has the beneficial effects of the corresponding method embodiment, which will not be repeated here.

[0239] Based on the same inventive concept, corresponding to any of the above-mentioned embodiment methods, the present application also provides a non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to execute the resume and job matching prediction method as described in any of the above embodiments.

[0240] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, read-only compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device.

[0241] The computer instructions stored in the storage medium of the above embodiment are used to enable the computer to execute the person-job matching method described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0242] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present application (including the claims) is limited to these examples. In line with the concept of the present application, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of the different aspects of the embodiments of the present application as described above, which are not provided in detail for the sake of simplicity.

[0243] In addition, to simplify the description and discussion, and in order not to make the embodiments of the present application difficult to understand, the known power supply / ground connection with the integrated circuit (IC) chip and other components may or may not be shown in the provided drawings. In addition, the device can be shown in the form of a block diagram to avoid making the embodiments of the present application difficult to understand, and this also takes into account the fact that the details of the implementation of these block diagram devices are highly dependent on the platform to be implemented in the embodiments of the present application (that is, these details should be fully within the scope of understanding of those skilled in the art). In the case of elaborating specific details (e.g., circuits) to describe exemplary embodiments of the present application, it is obvious to those skilled in the art that the embodiments of the present application can be implemented without these specific details or when these specific details are changed. Therefore, these descriptions should be considered to be illustrative rather than restrictive.

[0244] Although the present application has been described in conjunction with specific embodiments of the present application, many replacements, modifications and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may use the embodiments discussed.

[0245] The embodiments of the present application are intended to cover all such substitutions, modifications and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present application should be included in the scope of protection of the present application.

Claims

1. A method for predicting the matching degree between resume information and job information, characterized in that: include: Obtaining a first key and a first value of each semi-structured job attribute in the job information, and a second key and a second value of each semi-structured resume attribute in the resume information, wherein the first key, the first value, the second key and the second value are all represented by text data; Based on the first key and the first value of each of the job attributes, the first source representation corresponding to the job information, the second key and the second value of each of the resume attributes, and the second source representation corresponding to the resume information, predicting the matching degree between the resume information and the job information through a prediction model including a cascaded pre-trained language model, a Transformer encoder, and a single-label classification model; The predicting of the matching degree between the resume information and the job information includes: For each of the first keys and the first values ​​of the job attributes, respectively encode the first keys and the first values ​​into the semantic space through the pre-trained language model to obtain a first key embedding and a first value embedding, and fuse the first key embedding and the first value embedding to obtain a first fused embedding of the job attribute; encoding the first source representation into the semantic space through the pre-trained language model to obtain a first source embedding; For each second key and second value of the resume attribute, respectively encode the second key and the second value into the semantic space through the pre-trained language model to obtain a second key embedding and a second value embedding, and fuse the second key embedding and the second value embedding to obtain a second fused embedding of the resume attribute; encoding the second source representation into the semantic space through the pre-trained language model to obtain a second source embedding; By means of the Transformer encoder, internal interaction of a first matrix including a first fused embedding of each of the job attributes is performed to obtain a first internal interaction attribute embedding matrix, and internal interaction of a second matrix including a second fused embedding of each of the resume attributes is performed to obtain a second internal interaction attribute embedding matrix; fusing each element in the first internal interaction attribute embedding matrix with the first source embedding to obtain a first updated embedding matrix; fusing each element in the second internal interaction attribute embedding matrix with the second source embedding to obtain a second updated embedding matrix; Performing external interaction between the first updated embedding matrix and the second updated embedding matrix through the Transformer encoder to obtain a first external interaction attribute embedding matrix and a second external interaction attribute embedding matrix; The single-label classification model is used to predict the matching degree between the resume information and the job information according to the first external interaction attribute embedding matrix and the second external interaction attribute embedding matrix.

2. The method according to claim 1, characterized in that The fusing of the first key embedding and the first value embedding, the fusing of the second key embedding and the second value embedding, the fusing of each element in the first internal interaction attribute embedding matrix with the first source embedding, and the fusing of each element in the second internal interaction attribute embedding matrix with the second source embedding all include additive fusion operations.

3. The method according to claim 1, characterized in that The fusing of the first key embedding and the first value embedding, the fusing of the second key embedding and the second value embedding, the fusing of each element in the first internal interaction attribute embedding matrix with the first source embedding, and the fusing of each element in the second internal interaction attribute embedding matrix with the second source embedding all include splicing and fusion operations.

4. The method according to claim 1, characterized in that: Performing internal interaction of the first matrix to obtain the first internal interaction attribute embedding matrix includes: using a multi-head self-attention mechanism in the Transformer encoder to obtain a first multi-head self-attention matrix according to the first matrix; obtaining the first internal interaction attribute embedding matrix according to the first multi-head self-attention matrix through a first fully connected feedforward layer in the Transformer encoder; Performing internal interaction of the second matrix to obtain the second internal interaction attribute embedding matrix includes: utilizing the multi-head self-attention mechanism in the Transformer encoder to obtain the second multi-head self-attention matrix according to the second matrix; and obtaining the second internal interaction attribute embedding matrix according to the second multi-head self-attention matrix through the second fully connected feedforward layer in the Transformer encoder.

5. The method according to claim 4, characterized in that Performing the external interaction to obtain the first external interaction attribute embedding matrix and the second external interaction attribute embedding matrix includes: Concatenating the first updated embedding matrix with the second updated embedding matrix to obtain a joint matrix; Utilizing the multi-head self-attention mechanism in the Transformer encoder, a multi-head self-attention joint matrix is ​​obtained according to the joint matrix; The first external interaction attribute embedding matrix and the second external interaction attribute embedding matrix are obtained according to the multi-head self-attention joint matrix through the third fully connected feed-forward layer in the Transformer encoder.

6. The method according to claim 1, characterized in that Predicting the matching degree according to the first external interaction attribute embedding matrix and the second external interaction attribute embedding matrix includes: Pooling the first external interaction attribute embedding matrix and the second external interaction attribute embedding matrix through a first pooling layer and a second pooling layer in the single-label classification model, respectively, to obtain a job vector and a resume vector; Aggregate the position vector and the resume vector to obtain an aggregate vector; The aggregate vector is semantically fused by a multi-layer perceptron in the single-label classification model, and a matching score is obtained by using a single neuron according to the semantically fused aggregate vector; The activation function is used to obtain the matching degree between the resume information and the position information according to the matching score.

7. The method according to claim 6, characterized in that The pooling is performed using the adaptive average pooling function provided by the machine learning framework PyTorch.

8. The method according to any one of claims 1 to 7, characterized in that The prediction model is trained by minimizing the binary cross entropy loss.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 8 is implemented.

Images