Human-post matching recommendation method based on BERT and latent semantic algorithm model

Abstract

The invention discloses a person and post matching recommendation method based on BERT and a latent semantic algorithm model, and relates to the technical field of person and post matching recommendation, and the method comprises the steps: obtaining the related data of job hunters and the post recruitment text data of employers; performing word segmentation and part-of-speech tagging on the obtained text data, screening out candidate words and noun phrases, obtaining candidate word and document word embedding vectors by adopting a trained BERT model, calculating semantic similarity of the embedding vectors, and creating BERT bitmap data according to a calculation result; a latent semantic model is adopted to analyze semantic correlation in text data, preference prediction scores of job hunters for posts are calculated, corresponding class vectors are calculated through a deep random forest method, the matching probability of the job hunters and the posts is extracted, and post matching recommendation is carried out according to a probability sorting list. By adopting the hybrid recommendation method based on the latent semantic model and the deep random forest, the recommendation is more accurate, and the efficiency is higher.

Description

Technical Field

[0001] The present application relates to the technical field of person-job matching recommendation, and specifically to a person-job matching recommendation method based on BERT and latent semantic algorithm models. Background Art

[0002] With the development of the Internet and AI technology, online recruitment has become a common recruitment method. The rise of online recruitment platforms has made it difficult for corporate human resources departments to sort out and match suitable applicants from massive online data. Matching positions with resumes is the core of online recruitment platforms. Traditional recruitment cannot accurately analyze the individual development potential of job seekers, and a single matching method is difficult to fully reflect the matching degree between job seekers' abilities and positions, resulting in low recruitment efficiency and high employee turnover. Summary of the invention

[0003] In view of this, this application provides a person-job matching recommendation method based on BERT and latent semantic algorithm model, which solves the technical problems that traditional recruitment cannot accurately analyze the individual development potential of job seekers, and a single matching method is difficult to fully reflect the matching degree between job seekers' abilities and positions, resulting in low recruitment efficiency and high employee turnover rate.

[0004] To achieve the above object, the present invention provides the following technical solutions: A person-job matching recommendation method based on BERT and latent semantic algorithm model comprises the following steps: obtaining personnel-related data of job seekers and job recruitment description text data of employers; Sort and annotate the acquired text data, select candidate words and noun phrases, use the trained BERT model to obtain the word embedding vectors of these candidate words and the word embedding vectors of the document, and then calculate the semantic similarity of the word embedding vectors of the candidate words and the word embedding vectors of the document. According to the calculation results, the final candidate word score is obtained and keywords are extracted. User-position profiles and position-position profiles of job seekers based on the BERT bitmap structure are created for data management. The latent semantic model is used to analyze and represent the semantic correlation in text data, and the latent semantic feature vector is obtained. The job seeker's preference prediction score for the position is calculated, and the corresponding class vector is calculated through the deep random forest method. The probability of matching between job seekers and positions is extracted, the matching probabilities are processed and sorted, and job matching recommendations are made based on the sorted list.

[0005] Furthermore, the personnel-related data includes resume text data, job browsing and job collection preference data, and job seekers' interview history text data.

[0006] Furthermore, the semantic similarity of the embedded vectors is calculated, and the final candidate word scores are obtained based on the calculation results and keywords are extracted to create user-position profiles and position-position profiles of job seekers based on the BERT bitmap structure, including: Construct a document-based global graph network, use candidate words as nodes of the graph network, use the cosine similarity calculation method to calculate the semantic similarity of the embedding vector, and judge the jump probability between nodes in the graph network based on the semantic similarity. The weight of the edge is the product of the similarity and co-occurrence of the candidate words. Through iterative calculation, the final candidate word score is determined and keywords are extracted.

[0007] Furthermore, the latent semantic model is used to analyze and represent the semantic relevance in the text data, and the latent semantic feature vector is obtained to calculate the job seeker's preference prediction score for the position, including: Obtain the job seeker's preference information matrix R for the position, decompose the job seeker's preference information matrix R for the position into low-dimensional matrices P and Q through the latent semantic model, so that users and positions are linked through hidden classes, and the fused implicit features are obtained. The job seeker's preference prediction score for the position is calculated. The matrix P is the relationship matrix between users and hidden classes, and the matrix Q is the relationship matrix between positions and hidden classes. Calculate the job seekers' scores on the positions based on the matrix P and matrix Q.

[0008] Furthermore, the corresponding class vector is calculated by the deep random forest method to extract the probability of job seekers matching with positions, and the matching probabilities are processed and sorted, including: Construct a multi-level forest model, learn the data features based on the multi-level forest model, and extract the probability of matching job seekers with positions. Each level is composed of multiple decision trees, and each level between each level is composed of multiple decision trees; Based on the probability of matching job seekers with positions, a fusion factor is introduced for weighted fusion to obtain the final recommendation value, which is then sorted to give a list of recommended positions.

[0009] It can be seen from the above technical solution that the advantages of the present invention are: This application uses the BERT model to more fully mine text, and adopts a hybrid recommendation method based on latent semantic model and deep random forest to make recommendations more accurate and human resource management more efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application.

[0011] Figure 1A schematic diagram of the steps of this application.

[0012] Figure 2 Schematic diagram of the steps of the keyword extraction process of this application.

[0013] Figure 3 This is a schematic diagram of the ordered tag set and the ordered personnel set of this application.

[0014] Figure 4 This is a schematic diagram of the relationship between tags in this application.

[0015] Figure 5 This is a schematic diagram of the relationship between tags and personnel in this application.

[0016] Figure 6 Schematic diagram of matrix decomposition in LFM of this application. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the implementation modes and the accompanying drawings. Here, the illustrative implementation modes and descriptions of the present application are used to explain the present application, but are not intended to limit the present application.

[0018] This example proposes a person-job matching study based on the BERT (Bidirectional Encoder Representations from Transformers) bitmap structure (BS) and LFM (Latent Factor Model, latent semantic model) algorithm, and improves the accuracy and effectiveness of person-job matching by optimizing the text data processing method. Figures 1 to 6 This embodiment provides a person-job matching recommendation method based on BERT and latent semantic algorithm model, and the method specifically includes: Step S1: Obtaining personnel-related data of job seekers and job description text data of employers.

[0019] Specifically, the personnel-related data includes resume text data, job browsing and job collection preference data, and job seekers' interview history text data.

[0020] In the human resource management system of an enterprise, employee data is the core data of the system. The characteristic of employee portraits (UserProfiles, UP) is to provide a set of labels of different sizes for each employee. The current human resource management system usually only stores text-type employee data without mining the potential information therein. In order to extract information from text-type data and form labels, this application uses the BERT model to enrich the UP of employees through bitmap-based UP data management to improve statistical efficiency. UP is general information that models and describes the characteristics and interests of users. In the dataset, each document records information such as the employee's work history and past performance. It contains a large number of keywords that reflect employee characteristics. The purpose of the BERT-based UP generation algorithm is to extract keywords from documents, such as Figure 2 shown.

[0021] Step S2: Sort and annotate the acquired text data, screen out candidate words and noun phrases, use the trained BERT model to obtain the word embedding vectors of these candidate words and the word embedding vectors of the document, and then calculate the semantic similarity between the word embedding vectors of the candidate words and the word embedding vectors of the document, obtain the final candidate word score based on the calculation results and extract keywords, and create user-job profiles and job-job portraits of job seekers based on the BERT bitmap structure.

[0022] Specifically, the semantic similarity of the embedded vectors is calculated, and the final candidate word scores are obtained based on the calculation results and keywords are extracted. The user-position profile and position-position profile of job seekers based on the BERT bitmap structure are created, including: Construct a document-based global graph network, use candidate words as nodes of the graph network, use the cosine similarity calculation method to calculate the semantic similarity of the embedding vector, and judge the jump probability between nodes in the graph network based on the semantic similarity. The weight of the edge is the product of the similarity and co-occurrence of the candidate words. Through iterative calculation, the final candidate word score is determined and keywords are extracted.

[0023] BERT-based UP generation algorithm, such as Figure 2As shown in the figure, the documents are first sorted and lexically annotated to screen out candidate words and noun phrases. Then, the pre-trained BERT model is used to obtain these candidate words and document word embeddings, which in turn calculates their semantic similarity. In the graph network, the jump probability between nodes is based on the similarity of word embeddings, and the weight of the edge is the product of the similarity and co-occurrence of the candidate words. Through iterative calculation, the final candidate word score is determined and the keywords are extracted. The algorithm is divided into three steps: candidate word selection, scoring, and sorting. The selection stage focuses on identifying possible keywords, and the scoring stage uses the BERT model to enrich the semantic representation of the candidate words and calculate their similarity with the document. In the candidate word scoring, the pre-trained BERT model is used to calculate the word embedding representation of each candidate word relative to the entire document. Word embedding is a technique for converting words into numerical vectors that can capture the semantic information of words. In the BERT model, the word embedding representation of the candidate words is richer and can more accurately capture contextual information. These word embedding representations are then used to calculate the semantic similarity between the candidate words and the document. Before calculating the semantic similarity, the candidate words need to be converted into numerical representations. Use BERT to implement text vectorization. For document D, we can get a set of tokens by splitting the words. The candidate word set of document D is , the candidate word consists of one to multiple token sets, and the embedding representation of the token set is calculated as follows:

[0024] in, Represents a word feature set; yes Word embedding representation, Is a candidate word The embedding representation of can be calculated by averaging the embedding representations of the candidate word features. The embedding representation of the document is , and its calculation formula is as follows:

[0025] Among them, the embedded representation of the document contains all the semantic information of the document. After obtaining the embedded representation of the candidate word and the document, the semantic similarity between them can be calculated. In this embodiment, the cosine distance is used as a measure of the semantic similarity between the candidate word and the document. The degree of semantic similarity is expressed by calculating the cosine of the angle between their vectors. The formula for calculating the semantic similarity between the candidate word and the document is as follows:

[0026] According to the above formula, the semantic similarity between the candidate word and the document is obtained to determine whether the candidate word reflects the main information of the document well. Based on the sorting of semantic similarity, the top-ranked candidate words have a stronger correlation with the document content, while the lower-ranked candidate words may have a weaker correlation with the document content. After obtaining the semantic similarity, the weighted PageRank algorithm is used to dynamically sort the keywords. This is an algorithm based on a graph network that can evaluate the importance of nodes as a whole. First, a document-based global graph network is constructed, in which the candidate words are the nodes of the graph network. Whether the candidate words represented by two nodes appear simultaneously in a sliding window reflects whether there is an edge association between the nodes. The formula for calculating the edge weight between candidate words and is as follows:

[0027] in, represents the semantic similarity between candidate words, is the co-occurrence of two nodes, expressed as point mutual information, as shown below:

[0028] in, Represents the mutual information of points between candidate words.

[0029] The score of each candidate word is obtained by calculating the PageRank node score. The formula is as follows:

[0030] in, Indicates that the candidate word is in the document The importance of represents all candidate words that have edge connections with the candidate words and d is the damping factor.

[0031] The candidate words are sorted according to the score of each candidate word. In order to improve statistical efficiency, this embodiment uses BS to organize the relationship data between people and tags, and between tags and tags on the BERT-based UP data, and proposes a bitmap-based data management method. This embodiment uses two types of auxiliary data, namely, the tag sorting set and the person sorting set. The schematic diagram of the ordered tag set and the ordered person set is shown in Figure 3 As shown. Among them, is an ordered set of labels, i.e., a vector representing the labels ; Indicates the number of tags; Indicates The identity of the label; is an ordered set of people, i.e., a vector representing a person ; Indicates the number of people; represents the identity of the i-th person. The relationship model between labels is as follows: Figure 4 This relationship is shown in By a matrix where each row or column represents a label. The order and identification of rows and columns are defined by . Each element in takes the value 0 or 1. If , then it means The label is The parent tag of the tag. Otherwise, it means The label is not The parent tag of the tag.

[0032] The relationship between tags and people is as follows Figure 5 As shown, the relationship between tags and people Using the matrix The rows represent labels. , whose serial numbers and identifiers are defined by, and the columns represent characters , whose sequence number and identity are defined by . Each element in the formula takes a value of 0 or 1. , then it means The label is assigned to Otherwise it means Tags are not assigned to personal.

[0033] This embodiment uses a recommendation algorithm based on the LFM (latent semantic model) model and DF (deep random forest algorithm) on the basis of a constructed human resource management system to achieve the function of person-job matching.

[0034] As in step S3: adopt latent semantic model to analyze and represent semantic correlation in text data, and obtain latent semantic feature vector, calculate job seekers' predicted scores of job preferences, calculate corresponding class vectors through deep random forest method, extract the probability of job seekers matching with jobs, process and sort the matching probabilities, and make job matching recommendations based on the sorted list.

[0035] The latent semantic model is used to analyze and represent the semantic relevance in text data, and the latent semantic feature vector is obtained. The predicted scores of job seekers' preferences for positions are calculated, including: Obtain the job seeker's preference information matrix R for the position, decompose the job seeker's preference information matrix R for the position into low-dimensional matrix P and matrix Q through the latent semantic model, so that the user and the position can be linked through the hidden class to obtain the fused implicit features, and calculate the job seeker's preference prediction score for the position. The matrix P is the relationship matrix between the user and the hidden class, and the matrix Q is the relationship matrix between the position and the hidden class; calculate the job seeker's score for the position based on the matrix P and the matrix Q.

[0036] The corresponding class vector is calculated by the deep random forest method, the probability of matching job seekers with positions is extracted, and the matching probabilities are processed and sorted, including: Construct a multi-level forest model, learn the data features based on the multi-level forest model, and extract the probability of matching job seekers with positions. Each level is composed of multiple decision trees, and each level between each level is composed of multiple decision trees; Based on the probability of matching job seekers with positions, a fusion factor is introduced for weighted fusion to obtain the final recommendation value, which is then sorted to give a list of recommended positions.

[0037] The recommendation algorithm based on LFM and DF can be divided into three parts: LFM based on user behavior, DF integrated with implicit features, and hybrid recommendation. Latent Factor Model (LFM) is a type of matrix decomposition, which is used to analyze and represent the semantic relevance in text data. The schematic diagram of matrix decomposition in LFM is as follows: Figure 6 shown.

[0038] The matrix represents the user's preference information for posts, where The element represents the user Comments Ratings; The matrix represents the relationship between users and hidden classes, where The element represents the user For hidden classes level of interest; The matrix represents the relationship between posts and hidden classes, where The element represents a post In hidden class The weights on LFM are calculated by Decompose into a low-dimensional matrix and The product of the hidden class is the dot product. and posts "Post" refers to a position.

[0039] The formula for calculating the user's rating of a post is as follows:

[0040] Among them, F represents the number of hidden classes, Indicates user For hidden classes level of interest, And said the post In hidden class In order to calculate the matrices P and Q, this embodiment is implemented by optimizing the loss function. The loss function is defined as follows:

[0041] In order to avoid overfitting problems, a penalty factor is added ,in Represents the regularization parameter. According to the loss function, the root mean square error between the actual score and the predicted score of the position by the personnel is minimized by stochastic gradient descent to achieve the optimization of the objective function. Considering that different personnel have large variability in behavioral habits and other aspects, which may have a great impact on the rating results of the position, this embodiment introduces a deviation component to adjust the LFM, which is expressed as follows:

[0042] in, represents the grand average of all ratings, Indicates personnel Rating mean and total mean The deviation, Indicate job title Mean and total mean of ratings Based on this, the calculation formula of LFM prediction score is as follows:

[0043] By using LFM based on user behavior, it is possible to integrate the combined feature vectors of positions and people, leading to fused implicit features. Since the human resources platform is classified according to user behavior, it becomes difficult to directly judge the match between positions and people. The DF algorithm is an integrated learning method. DF learns the representation of data by constructing a multi-level forest, each level consisting of multiple decision trees, similar to the way data representation is learned through multi-layer neural networks in deep learning, but it does not rely on gradient descent and back propagation, and is mainly composed of hierarchical structure and cascade structure.

[0044] Using the hybrid recommendation model, we obtain the implicit semantic feature vector and calculate the predicted score of the personnel for the position. post The interest level is calculated as follows: , Then, the corresponding class vector is calculated through the DF model to extract the probability of matching personnel and positions. Then, the fusion factor is introduced for weighted fusion to obtain the final recommendation value, as shown in the following formula.

[0045]

[0046] in, represents the probability of a person matching a job; represents the fusion factor; Indicates the final recommendation value. When the user's score for a position is higher, the matching degree is higher and the recommendation value increases. Therefore, the JR (Jobre commendation, job recommendation) matching algorithm based on LFM will recommend jobs with high matching degree to users as much as possible.

[0047] The key point of the person-job matching recommendation algorithm lies in text representation and matching technology. The matching task is usually converted into a supervised text matching problem. Supervised learning is powerful when there is enough labeled data. However, on the Internet recruitment platform, the job-resume interaction data is sparse and noisy, which affects the performance of the job-resume matching algorithm. The semantic relevance can be calculated by extracting text concepts and attributes using latent semantic analysis and then mapping them to fuzzy formal concept analysis. It can effectively help users perform semantic analysis. Therefore, the person-job matching recommendation algorithm model based on latent semantics proposed in this application improves the accuracy and performance of text expression, reduces the cost of human recruitment and shortens the recruitment cycle, and can help both parties of recruitment match more suitable candidates and positions.

[0048] This embodiment also includes an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the computer program.

[0049] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the embodiments of the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A person-job matching recommendation method based on BERT and latent semantic algorithm model, characterized in that: The steps include: Obtaining personnel-related data of job seekers and job description text data of employers; Sort and annotate the acquired text data, select candidate words and noun phrases, use the trained BERT model to obtain the word embedding vectors of the candidate words and the word embedding vectors of the document, and then calculate the semantic similarity between the word embedding vectors of the candidate words and the word embedding vectors of the document. According to the calculation results, the final candidate word score is obtained and keywords are extracted. User-position profiles and position-position profiles of job seekers based on the BERT bitmap structure are created for data management. The latent semantic model is used to analyze and represent the semantic correlation in text data, and the latent semantic feature vector is obtained. The job seeker's preference prediction score for the position is calculated, and the corresponding class vector is calculated through the deep random forest method. The probability of matching between job seekers and positions is extracted, the matching probabilities are processed and sorted, and job matching recommendations are made based on the sorted list.

2. The person-job matching recommendation method based on BERT and latent semantic algorithm model according to claim 1 is characterized in that: The personnel-related data includes resume text data, job browsing and job collection preference data, and job seekers' interview history record text data.

3. The person-job matching recommendation method based on BERT and latent semantic algorithm model according to claim 1 is characterized in that: The calculation of the semantic similarity of the embedding vectors, obtaining the final candidate word score according to the calculation results and extracting keywords, and creating the user-position profile and position-position profile of the job seeker based on the BERT bitmap structure include: Construct a document-based global graph network, use candidate words as nodes of the graph network, use the cosine similarity calculation method to calculate the semantic similarity of the embedding vector, and judge the jump probability between nodes in the graph network based on the semantic similarity. The weight of the edge is the product of the similarity and co-occurrence of the candidate words. Through iterative calculation, the final candidate word score is determined and keywords are extracted.

4. The person-job matching recommendation method based on BERT and latent semantic algorithm model according to claim 1 is characterized in that: The method of using a latent semantic model to analyze and represent semantic relevance in text data and obtain a latent semantic feature vector to calculate the job seeker's job preference prediction score includes: Obtain job seekers' preference information matrix R for positions, decompose the job seekers' preference information matrix R for positions into low-dimensional matrices P and Q through latent semantic models, so as to link users and positions through hidden classes, obtain fused implicit features, and calculate job seekers' preference prediction scores for positions, wherein the matrix P is a relationship matrix between users and hidden classes, and the matrix Q is a relationship matrix between positions and hidden classes; Calculate the job seekers' scores on the positions based on the matrix P and matrix Q.

5. The person-job matching recommendation method based on BERT and latent semantic algorithm model according to claim 1 is characterized in that: The corresponding class vector is calculated by the deep random forest method, the probability of matching job seekers with positions is extracted, and the matching probabilities are processed and sorted, including: Construct a multi-level forest model, learn the data features based on the multi-level forest model, and extract the probability of matching job seekers with positions. Each level is composed of multiple decision trees, and each level between each level is composed of multiple decision trees; Based on the probability of matching job seekers with positions, a fusion factor is introduced for weighted fusion to obtain the final recommendation value, which is then sorted to give a list of recommended positions.

Images