Pseudo tag guidance-based comparative learning man-post matching recommendation method

Abstract

The invention discloses a comparative learning post matching recommendation method based on pseudo tag guidance, and particularly relates to the technical field of post matching, and the method comprises the steps: obtaining a job seeker resume text, post description information and each pair of resume-post matching conditions, i.e., whether a job seeker meets an applied post requirement or not and obtains an interview opportunity; carrying out desensitization processing on the data; performing data preprocessing on the resume text and the post description information, constructing a sample matrix required by comparative learning, and performing division to obtain a training set, a verification set and a test set; a two-way resume-post matching network is constructed, a layered semantic coding framework is adopted, matching of resumes and posts can be automatically and accurately carried out, the problem of information asymmetry in a traditional recruitment scene is effectively solved, matched posts are intelligently recommended for job seekers, meanwhile, target talents are accurately screened for enterprises, and an efficient personnel and post adaptation channel is constructed.

Description

Technical Field

[0001] The present invention relates to the technical field of person-job matching, and particularly to a person-job matching recommendation method based on contrastive learning guided by pseudo labels. Background Art

[0002] With the acceleration of the digitalization process, online recruitment has become the mainstream trend. An intelligent person-job matching system based on AI and big data can improve the recruitment efficiency in both directions by analyzing job requirements and job seeker ability profiles, and has significant social value. For enterprises, the system can quickly screen resumes that meet the requirements, shorten the time cost of traditional manual screening, and at the same time reduce the bias caused by subjective judgment. The system can also dynamically match enterprise needs with talent skills to improve the recruitment accuracy. For job seekers, the algorithm can deeply analyze personal skills and career interests, recommend more suitable positions, and help them clarify their career directions. From a social perspective, such systems can provide data references for college talent cultivation by analyzing the gap between enterprise employment needs and talent skills, which helps to alleviate the problem of mismatched talent supply and demand structures and promote the deep connection between the education system and industrial needs. With the continuous development of technology, intelligent person-job matching systems will play a more important role in human resource allocation.

[0003] Early person-job matching research mainly focused on the optimization of neural network structures, aiming to extract hidden semantic features of jobs and resumes through architectures such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and convolutional neural networks (CNNs). However, the performance of these traditional networks is significantly weaker than that of encoder models represented by BERT. As a multi-layer encoder based on bidirectional Transformer, BERT realizes semantic representation through a special token "CLS": when processing job descriptions or resume texts, the system superimposes token embeddings (representing semantics) and position embeddings (representing sequence positions) to form an input vector, and finally extracts the global semantic vector corresponding to "CLS" as the representation result after multi-layer encoding. Currently, mainstream systems generally adopt such encoder architectures for hidden semantic modeling of jobs and resumes.

[0004] Limited by the sensitivity of industry data, existing systems mostly adopt closed-source implementation and customized optimization. MV-CoN dynamically adjusts gradient updates through a collaborative teaching network, InEXIT uses a hierarchical attention mechanism to model job-resume interactions, and DPGNN innovatively combines graph networks and BPR loss functions to optimize the ranking effect. However, these methods have obvious limitations: some studies assume that there are no unknown jobs / resumes in the test scenario, or rely on internal platform data such as "recruiter private messages". The former does not conform to the reality of the continuous emergence of new jobs / job seekers in the actual scenario, and the latter is limited by the operating mechanism of a specific platform.

[0005] In response to the pain point of insufficient algorithm generalization ability, the proposed ConFit algorithm is of breakthrough significance. This solution introduces a contrastive learning framework, enhances data by performing semantically equivalent replacements on resumes or job positions using a pre-trained language model, constructs a highly generalizable job-candidate matching model, and effectively improves the model's adaptability to new platforms and emerging job positions / resumes. Despite the cumbersome data enhancement operation, the model demonstrates excellent performance in both classification and ranking tasks, especially in the ranking task, significantly outperforming previously proposed models.

[0006] The current job-candidate matching system faces three core challenges. First is data sparsity. In the vast amount of job-resume data, the proportion of clearly matched / mismatched sample pairs is extremely low. Second is the real-time requirement. It is necessary to complete massive data retrieval and precise ranking within a millisecond-level response time. Third is cross-domain generalization. It is necessary to ensure the stability of the recommendation quality of the algorithm in different recruitment platforms and emerging job scenarios. Summary of the Invention

[0007] The object of the present invention is to solve the problems of low accuracy, poor real-time performance, and insufficient generalization ability caused by factors such as data sparsity in the prior art, and proposes a job-candidate matching recommendation method based on pseudo-label-guided contrastive learning. In practical applications, this method only needs to collect the resume information of job seekers and job description information, and can generate a matching job recommendation list or a high-quality candidate resume recommendation list in real time, not only with high recommendation accuracy but also extremely fast response speed, so as to solve the problems raised in the background art.

[0008] To achieve the above object, the present invention provides the following technical solution: A job-candidate matching recommendation method based on pseudo-label-guided contrastive learning, comprising the following steps:

[0009] (1) Obtain the resume text of job seekers, job description information, and the matching situation of each resume-job pair, that is, whether the job seeker meets the requirements of the applied job and obtains an interview opportunity, and perform data desensitization processing on the data;

[0010] (2) Perform data preprocessing on the resume text and job description information. By traversing the resume-job matching pairs in the dataset, match positive samples (jobs that match the resume) and negative samples (jobs that do not match the resume) for each resume respectively, and match positive samples (candidate resumes that match the job) and negative samples (candidate resumes that do not match the job) for each job, so as to construct a sample matrix required for contrastive learning and divide it into a training set, a validation set, and a test set;

[0011] (3) Construct a two-way resume-job matching framework. Based on the domain pre-trained language model, perform fine-grained semantic encoding on each field of the resume and job, and obtain a structured semantic representation. Implement field-level dynamic weight allocation through the multi-head attention mechanism, and analyze the relevance between each field within the resume and job texts. Finally, perform spatial mapping through a linear layer to generate the global semantic vectors of the resume and job. Calculate the similarity between the resume vector and the job vector, which is the matching degree between the two. Feed the training set and the complete information of the resume and job into the matching network for classification prediction. By calculating the contrastive loss of the matching network and iteratively updating the parameters, after training for a certain number of rounds, the model reaches a certain confidence level. When the model confidence level reaches the preset convergence condition, automatically trigger the Top-K nearest neighbor retrieval algorithm to construct pseudo-labels and form a contrastive learning supervision signal with self-enhancing characteristics.

[0012] (4) Select the optimal model weights and calculate various evaluation metrics of the model on the test set.

[0013] Preferably, the data desensitization process described in step (1) is specifically as follows: To protect the privacy of job seekers, all records are anonymized after deleting sensitive identity information, and all resume and job information has been parsed into a series of text fields for further analysis and processing.

[0014] Preferably, the data preprocessing in step (2) is specifically as follows: For each successfully matched resume-job combination, there are four contrast relationships: the resume positive example of the matching job, the resume negative example of the non-matching job, the job positive example of the adapted resume, and the job negative example of the non-matching resume, forming a contrastive learning sample matrix with strong discrimination, and dividing the training set, validation set, and test set according to 8:1:1.

[0015] Preferably, the constructed two-way resume-job matching framework in step (3) includes three parts: a multi-granularity semantic representation module, a contrastive learning training module, and a dynamic pseudo-label generation module. Adopt a hierarchical semantic encoding architecture to obtain multi-granularity semantic representations.

[0016] Multi-granularity semantic representation module: Input the resume text and job description into the multilingual-E5 encoder with shared parameters respectively. Perform fine-grained semantic encoding on each field in the resume text and job description, fuse the features of each field through the self-attention mechanism, generate a context-aware global semantic vector, and construct a 1024-dimensional semantic representation space.

[0017] Contrastive learning training module: Adopt a phased optimization strategy, including:

[0018] (c) Coarse-grained contrastive learning, at the initial stage of training, by calculating the cosine similarity of positive and negative sample pairs and minimizing the contrastive loss function, the vector space distance of matching pairs is reduced, and the distance of non-matching pairs is enlarged;

[0019] (d) Fine-grained joint optimization, introducing fine-grained contrastive loss after the initial stage of training, setting an adjustable parameter α to balance global and local features, and achieving multi-granularity semantic alignment through the gradient accumulation strategy:

[0020] L contrasive_loss = L coarse_loss + α·L fine_loss ;

[0021] Among them, L contrasive_loss represents the contrastive loss, which consists of the coarse-grained loss L coarse_loss and the fine-grained loss L fine_loss ;

[0022] The improved contrastive learning training module based on the InfoNCE loss function introduces an explicit negative example penalty mechanism, which is as follows:

[0023] In a batch of training data, the model processes N groups of effectively matching resume-job positive example pairs For each positive example pair Among them represents the resume in the i-th resume-job positive example pair, then represents the job in the i-th resume-job positive example pair; the system automatically uses the remaining N - 1 positive example pairs as implicit negative examples to participate in contrastive learning; for each uses the remaining N - 1 as implicit negative examples for contrastive learning;

[0024] The training data of this framework includes three types of special samples:

[0025] 1) Explicit negative example samples (non-matching resumes / jobs, such as non-matching jobs );

[0026] 4) Enhanced positive example samples (extra matching samples, such as another matching job of );

[0027] 5) Basic positive example samples (original matching pairs);

[0028] Among them, the enhanced positive examples will also participate in the contrast loss calculation as implicit negative examples; the framework adds an explicit negative example contrast loss term, and balances the penalty intensity of explicit negative examples by adjusting the coefficient β; by adding a weighted contrast term for explicit negative examples, the discrimination ability of the model for positive and negative samples is enhanced, and the matching accuracy of job recommendations is significantly improved. The loss function is as follows:

[0029] L = L R + L J

[0030]

[0031] Among them, L represents the overall loss, which is composed of the resume matching loss L R and the job matching loss L J The two parts are composed; N represents the total number of resume-job matching pairs, pos represents the set of positive examples, neg represents the set of negative examples, and S θ represents the cosine similarity, that is, the matching degree;

[0032] Dynamic pseudo-label generation module: When the prediction confidence of the framework reaches the preset threshold, the pseudo-label generation process is automatically triggered to achieve dynamic expansion of data annotation. The specific process: Calculate the Top-K similar samples based on the current semantic space and K = 1, that is, by matching the resume or job with the highest similarity, and reuse the corresponding positive and negative sample data to generate high-quality synthetic samples, realizing dynamic expansion of the training data scale and effectively alleviating the data sparsity problem.

[0033] Preferably, each evaluation index in step (4) includes two parts, namely the evaluation index of the recommendation result sorting ability and the evaluation index of the framework classification ability; the evaluation index of the recommendation result sorting ability includes MAP and nDCG; the evaluation index of the framework classification ability includes F1, Prc, and Rcl.

[0034] The present invention has the following advantages:

[0035] The present invention constructs a two-way resume-job matching framework, and adopts a multi-granularity semantic representation module, a contrast learning training module, and a dynamic pseudo-label generation module, so that the present invention can automatically and accurately match resumes and jobs, effectively solve the information asymmetry problem (data sparsity) in traditional recruitment scenarios, intelligently recommend suitable jobs for job seekers (real-time), and at the same time accurately screen target talents for enterprises (cross-domain generalization), and build an efficient person-job matching channel;

[0036] The hierarchical semantic encoding architecture performs fine-grained semantic encoding on each field of the resume-position based on a domain pre-trained language model to obtain a structured semantic representation; realizes field-level dynamic weight allocation through a multi-head attention mechanism, and models cross-modal association analysis within the document; finally, generates a global semantic vector of the resume-position through spatial mapping by a linear layer; calculates the similarity between the resume vector and the position vector, that is, their matching degree; sends the training set and the complete information of the resume and the position into the matching network for classification prediction, calculates the contrast loss of the matching network, iteratively updates the parameters, and after training for a certain number of rounds, the model reaches a certain confidence level. When the model confidence level reaches the preset convergence condition, the Top-K nearest neighbor retrieval algorithm is automatically triggered to construct pseudo-labels, forming a contrast learning supervision signal with self-enhancing characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a framework diagram of a resume-position matching method based on contrast learning guided by pseudo-labels provided by an embodiment of the present invention;

[0038] Figure 2 It is a structural diagram of a resume-position matching network based on contrast learning guided by pseudo-labels provided by an embodiment of the present invention;

[0039] Figure 3 It is an effect diagram of a resume-position matching algorithm based on contrast learning guided by pseudo-labels provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0040] The following specific embodiments illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0041] Embodiment: The resume-position matching method based on contrast learning guided by pseudo-labels provided by the present invention first performs desensitization processing on the input resume and position information, and generates positive and negative sample pairs required for contrast learning. Then, the data and labels in the training set are input into the matching network for training. After training for a certain number of iterations, the average accuracy of the network reaches fitting, and the network weights are saved. Input the resumes and positions in the test set to obtain a matching resume / position recommendation list for each position and resume. The resumes / positions with higher matching degrees are ranked higher, completing the two-way matching recommendation task of resumes and positions. Refer to Figure 1 As shown, the resume-position matching method based on contrast learning guided by pseudo-labels mainly includes the following steps:

[0042] Step 1: Taking the dataset of a resume-job matching competition of a certain company publicly available in 2019 as an example, the dataset includes three parts: resume information, job information, and resume-job matching information. The above information has been desensitized to hide privacy information, and the parts of the resume involving specific experiences have been processed into keyword fields. After data cleaning, the resume information includes a total of 12 fields, namely: user_id, city of residence, expected work city, expected work industry, expected work type, expected salary, current work industry, current work type, current salary, education level, age, start working time, work experience. The job information includes a total of 11 fields: jd_no, job name, work city, work type, number of recruits, salary, recruitment start time, recruitment end time, whether business trips are required, work years, minimum education level, job description. The resume-job matching information includes a total of three fields: user_id, jd_no, satisfied. To address the problem of imbalanced positive and negative sample ratios in the original data (the proportion of negative samples is significantly higher), we adopted the stratified random sampling technique to adjust the scale of negative samples to a 1:1 balanced state with positive samples, effectively avoiding the problem of sample bias in model training. After removing invalid entries through data cleaning, it is divided into a training set, a validation set, and a test set according to a ratio of 8:1:1. Finally, a high-quality training set of 22,307 entries, and 2,780 entries each for the validation set and the test set are constructed.

[0043] Step 2: Construct contrastive learning positive and negative sample pairs. For each successfully matched resume-job combination, systematically mine four-dimensional contrast relationships: resume positive example (matching job), resume negative example (non-matching job), job positive example (fitting resume), job negative example (non-matching resume), to form a contrastive learning sample matrix with strong discrimination.

[0044] Step 3: Construct a network. The proposed bidirectional resume-job matching network of the present invention includes three parts, namely a multi-granularity semantic representation module, a contrastive learning training module, and a dynamic pseudo-label generation module, as specifically referred to Figure 2 shown. [[ID=??]]

[0045] Multi-granularity Semantic Representation Module: Input the resume text and job description into the multilingual-E5 encoder with shared parameters respectively, and perform fine-grained semantic encoding for core fields such as "Professional Skills" and "Project Experience". Through the self-attention mechanism, fuse the features of each field to generate a context-aware global semantic vector and construct a 1024-dimensional semantic representation space. Contrastive Learning Training Module: Adopt a phased optimization strategy. Coarse-grained contrastive learning, in the initial stage of training (warm-up stage), by calculating the cosine similarity of positive and negative sample pairs, minimize the contrastive loss function, so that the vector space distance of matching pairs is reduced and the distance of non-matching pairs is increased. Fine-grained joint optimization, set an adjustable parameter α to balance global and local features, and achieve multi-granularity semantic alignment through the gradient accumulation strategy:

[0046] L contrasive_loss = L coarse_loss + α·L fine_loss ;

[0047] Add an explicit negative example contrast loss term, and set an adjustable coefficient β to balance the penalty strength of explicit negative examples. By adding a weighted contrast term for explicit negative examples, effectively enhance the network's ability to distinguish positive and negative samples, and significantly improve the matching accuracy of the job recommendation system. The loss function is as follows:

[0048] L = L R + L J

[0049]

[0050] Dynamic Pseudo-label Generation Module: When the network prediction confidence exceeds 0.7, start the pseudo-label generation mechanism. Calculate the Top-K similar samples (K = 1) based on the current semantic space to generate high-quality synthetic samples, dynamically expand the training data scale, and effectively alleviate the data sparsity problem.

[0051] Step 4: Train the network. The present invention is carried out on a workstation equipped with an NVIDIA-RTX4070 GPU, the programming language is Python, the PyTorch deep learning framework is adopted, AdamW is used as the optimizer for network training, the batch size is 16, and the learning rate is 1e-5. Input the training data into the matching network for training. After 50 rounds of training, the network gradually fits, and the network parameters with the best training effect are saved during each iteration of training.

[0052] Step 5: Evaluate the network. To verify the performance of the network, test on the test set. The test set data is sent into the optimal weight network and the predicted detection results are output. The resume and job matching results are as Figure 3 shown.

[0053] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made thereto based on the present invention, which will be obvious to those skilled in the art. Therefore, these modifications or improvements made without departing from the spirit of the present invention all fall within the scope of the present invention claimed.

Claims

1. A method for person-job matching recommendation based on pseudo-label-guided contrast learning, comprising the following steps: (1) Obtain the resume text of job seekers, job description information, and the matching situation of each pair of resume-job, that is, whether the job seeker meets the requirements of the applied position and obtains an interview opportunity, and perform data desensitization processing on the data; (2) Perform data preprocessing on the resume text and job description information, construct a sample matrix required for contrast learning, and divide it into a training set, a validation set, and a test set; characterized in that: (3) Construct a bidirectional resume-job matching framework, perform fine-grained semantic encoding on each field of the resume-job based on a domain pre-trained language model to obtain a structured semantic representation; achieve field-level dynamic weight allocation through a multi-head attention mechanism, and analyze the relevance between each field inside the resume and job texts; finally, perform spatial mapping through a linear layer to generate global semantic vectors of the resume and the job; calculate the similarity between the resume vector and the job vector, that is, their matching degree; send the training set and the complete information of the resume and the job into the matching network for classification prediction, calculate the contrast loss of the matching network, and iteratively update the parameters. After training for a certain number of rounds, the model reaches a certain confidence level. When the model confidence level reaches the preset convergence condition, the Top-K nearest neighbor retrieval algorithm is automatically triggered to construct pseudo-labels and form a contrast learning supervision signal; (4) Select the optimal model weights and calculate various evaluation indicators of the model under the test set.

2. The method for recommending person-job matching based on pseudo-label-guided contrastive learning according to claim 1, wherein: The data desensitization processing described in step (1) is specifically: to protect the privacy of job seekers, all records are anonymized after deleting sensitive identity information, and all resume and job information has been parsed into a series of text fields for further analysis and processing.

3. The method for recommending person-job matching based on pseudo-label-guided contrastive learning according to claim 1, characterized in that: The data preprocessing in step (2) is specifically: for each successfully matched resume-job combination, it includes four contrast relationships: the positive example of the resume that matches the job, the negative example of the resume that does not match the job, the positive example of the job that fits the resume, and the negative example of the job that does not match the resume, to form a contrast learning sample matrix with strong discrimination.

4. The method for recommending person-job matching based on pseudo-label-guided contrastive learning according to claim 1, wherein: The construction of the bidirectional resume-job matching framework described in step (3) includes three parts: a multi-granularity semantic representation module, a contrast learning training module, and a dynamic pseudo-label generation module, and adopts a hierarchical semantic encoding architecture to obtain multi-granularity semantic representations; Multi-granularity semantic representation module: Input the resume text and job description into the multilingual-E5 encoder with shared parameters respectively, perform fine-grained semantic encoding on each field in the resume text and job description, fuse the features of each field through the self-attention mechanism, generate a context-aware global semantic vector, and construct a 1024-dimensional semantic representation space; Contrast learning training module: Adopt a phased optimization strategy, including: (a) Coarse-grained contrast learning, at the initial stage of training, by calculating the cosine similarity of positive and negative sample pairs, minimize the contrast loss function, so that the vector space distance of the matching pairs is reduced and the distance of the non-matching pairs is increased; (b) Fine-grained joint optimization. After the initial stage of training, introduce fine-grained contrastive loss, set an adjustable parameter α to balance global and local features, and achieve multi-granularity semantic alignment through the gradient accumulation strategy: L contrasive_loss = L coarse_loss + α·L fine_loss ; Among them, L contrasive_loss represents the contrastive loss, which consists of the coarse-grained loss L coarse_loss and the fine-grained loss L fine_loss ; Improved contrastive learning training module based on the InfoNCE loss function, introducing an explicit negative example penalty mechanism, specifically as follows: In a batch of training data, the model processes N groups of valid matching resume-job positive pairs For each positive pair where represents the resume in the i-th resume-job positive pair, and represents the job in the i-th resume-job positive pair; the system automatically uses the remaining N - 1 positive pairs as implicit negative examples for contrastive learning; for each the remaining N - 1 are used as implicit negative examples for contrastive learning; The training data of this framework includes three types of samples: 1) Explicit negative example samples, i.e., resumes / jobs that do not match, denote jobs that do not match; 2) Enhance positive example samples, i.e., additional matching samples, denote another matching position; 3) Basic positive example samples, that is, the original matching pairs; Among them, the enhanced positive example samples also participate in the contrastive loss calculation as implicit negative examples; this framework adds an explicit negative example sample contrastive loss term, and adjusts the penalty intensity of the explicit negative examples through the adjustment coefficient β; by adding a weighted contrast term for the newly added explicit negative example samples, to enhance the model's ability to distinguish positive and negative samples, and the loss function is as follows: L = L R + L J Among them, L represents the overall loss, which consists of the resume matching loss L R and the job matching loss L J and is composed of two parts; N represents the total number of resume-job matching pairs, pos represents the set of positive example samples, neg represents the set of negative example samples, and S θ represents the cosine similarity, that is, the matching degree; Dynamic pseudo-label generation module: When the prediction confidence of the framework reaches the preset threshold, automatically trigger the pseudo-label generation process to achieve dynamic expansion of data annotation. The specific process: Calculate the Top-K similar samples based on the current semantic space and K = 1, that is, by matching the resume or position with the highest similarity, and reuse the corresponding positive and negative sample data to generate synthetic samples, realizing dynamic expansion of the training data scale.

5. The method for recommending person-job matching based on pseudo-label-guided contrastive learning according to claim 1, wherein: Each evaluation index in step (4) includes two parts, namely the evaluation index for the recommendation result ranking ability and the evaluation index for the framework classification ability; the evaluation index for the recommendation result ranking ability includes MAP and nDCG; the evaluation index for the framework classification ability includes F1, Prc, and Rcl.

Images