Person-job matching method and system

Abstract

This invention discloses a deep learning-based intelligent person-job matching method and system based on a large language model and multiple prompts, belonging to the field of big data technology. It involves collecting resume and job datasets, constructing a dual-tower model that takes the resume and job datasets as input, outputs resume and job embedding vectors, and uses the matching similarity of the embedding vectors as the prediction target. The dual-tower model is then trained to construct a matching model that takes the resume and job datasets as input and outputs matching scores. This matching model is also trained. Person-job matching requests are collected, and the dual-tower model provides an initial screening set for these requests. The matching model then filters the final matching results from this initial screening set. Through the collaborative work of the two models, efficient and accurate person-job matching is achieved.

Description

Technical Field

[0001] This invention relates to the field of big data technology, specifically to a deep learning-based intelligent human-job matching method and system based on large language models and multiple prompts. Background Technology

[0002] With rapid global economic growth and continuous technological advancements, the human resources market is also developing rapidly, and enterprises are increasingly demanding efficient and accurate job matching. Traditional recruitment processes often rely on manual screening and interviews, which is not only time-consuming and labor-intensive but also difficult to process large volumes of resumes and job requirements in a short period. Therefore, how to utilize advanced technologies to improve the efficiency and accuracy of job matching has become an important research topic.

[0003] Traditional matching methods have the following limitations:

[0004] Traditional matching methods mainly rely on keyword matching or simple text similarity calculation. While these methods are simple and easy to use, they have the following significant limitations:

[0005] Lack of semantic understanding: Traditional keyword matching methods cannot capture deep semantic information in the text, resulting in inaccurate matching results.

[0006] Limited information processing capacity: Resumes and job requirements usually contain a lot of information, and traditional matching methods are unable to effectively process this complex information and perform efficient matching.

[0007] Chinese Patent No. CN111144723B discloses a method and system for recommending job-person matching, and a storage medium. The method includes the following steps: obtaining job posting text information entered by employers and resume text information entered by job seekers; the job posting text information is relevant to the job being advertised, and the resume text information is the job seeker's resume; parsing the job posting text information and resume text information using a text parsing algorithm to obtain tag information; and processing the tag information using a job similarity algorithm and a job-resume matching algorithm to recommend the most suitable resume for the current job. However, this method fails to utilize multiple prompts, making it difficult to grasp the correlation between the resume and the job posting.

[0008] To this end, this invention proposes a deep learning-based intelligent human-job matching method and system based on a large language model and multiple prompts. Summary of the Invention

[0009] This invention aims to address at least one of the technical problems existing in the prior art. To this end, this invention proposes a deep learning-based intelligent person-job matching method and system based on a large language model and multiple prompts. Through the collaborative work of the two models, efficient and accurate person-job matching is achieved.

[0010] To achieve the above objectives, a deep learning-based intelligent person-job matching method based on a large language model and multiple prompts is proposed, including the following steps:

[0011] Step 1: Collect resume datasets and job posting datasets;

[0012] Step 2: Construct a dual-tower model that takes the resume dataset and job dataset as input, the resume embedding vector and the job embedding vector as output, and the matching similarity of the embedding vectors as the prediction target, and train the dual-tower model.

[0013] Step 3: Construct a matching model that takes the resume dataset and job dataset as input and the matching score as output, and train the matching model;

[0014] Step 4: Collect job matching requests, and use the dual-tower model to provide an initial screening set of job matching requests. Then, use the matching model to filter out the final matching results from the initial screening set.

[0015] The method for collecting resume datasets and job posting datasets is as follows:

[0016] The resume dataset is composed of all resumes uploaded by job seekers and all job postings uploaded by companies.

[0017] The method for constructing a dual-tower model that takes resume datasets and job datasets as input and outputs resume embedding vectors and job embedding vectors is as follows:

[0018] The dual-tower model includes a job tower and a resume tower;

[0019] The job tower includes an input layer, a slice layer, an embedded model, a fully connected layer, an attention pooling layer, and an output layer;

[0020] The resume tower includes an input layer, a slice layer, an embedding model, a Transformer encoding layer, an attention pooling layer, and an output layer;

[0021] The input layer of the job tower takes the job description of each job in the job dataset as input;

[0022] The job tower's slicing layer and embedding model are used to segment and vectorize job descriptions to obtain initial job embedding vectors.

[0023] The fully connected layer and attention pooling layer of the job tower are conventional techniques in this field. The initial job embedding vector is transformed into a job embedding vector of length d through the fully connected layer and attention pooling layer; d is the preset embedding vector length.

[0024] The input layer of the resume tower takes the resume text of each job seeker in the resume dataset as input;

[0025] The slicing layer and embedding model of the resume tower are used to segment and vectorize the resume text to obtain the initial embedding vector of the resume.

[0026] The method for training the dual-tower model is as follows:

[0027] Job positions and resumes are matched manually, and each match is labeled.

[0028] Use labeled matching pairs to train the dual-tower model;

[0029] The method for constructing a matching model that takes resume and job datasets as input and matches the degree of matching as output is as follows:

[0030] The matching model includes a job description processing part, a resume text processing part, and a merging and matching part;

[0031] The processing procedure for the job description section includes:

[0032] The job description input for each job in the job dataset is segmented into layers to obtain M core content items;

[0033] Each core element is input into the embedding model to obtain the embedding vector corresponding to the core element.

[0034] The embedding vector of each core content item is input into the fully connected layer to obtain the corresponding hint embedding vector of the core content.

[0035] By stacking the prompts for various core contents, job prompts are obtained;

[0036] The processing steps for the resume text processing section include:

[0037] Input the resume text of each job seeker in the resume dataset into the segmentation and embedding model in sequence to obtain the initial embedding vector of the resume output by the embedding model;

[0038] The initial resume embedding vector is then input into the Transformer encoding layer and the attention pooling layer in sequence to obtain the resume embedding vector.

[0039] The processing steps for the merging and matching part include:

[0040] Input the job prompt embedding and resume embedding vectors into the multi-head attention layer to obtain the merged embedding vector;

[0041] The merged embedding vector is input into the fully connected layer to obtain an intermediate result vector;

[0042] The intermediate result vector is input into the sigmoid function layer, and the output is a real value in the range of [0, 1], which is the classification probability value, and is used as the confidence of the classification result, as well as the matching degree.

[0043] Set a classification threshold. Samples with values ​​greater than or equal to the classification threshold are classified as positive, meaning the resume matches the job posting; samples with values ​​less than the classification threshold are classified as negative, meaning the resume does not match the job posting.

[0044] The method for training the matching model is as follows:

[0045] Use labeled matching pairs to train the matching model;

[0046] Due to the imbalance of sample classes, Focal Loss is used to calculate the training loss and update the parameter weights of the matching model;

[0047] The process of collecting job matching requests involves collecting job seekers' resumes through the backend of the recruitment site and generating job matching requests for them, as well as collecting job descriptions from companies and generating resume matching requests for them.

[0048] The method of providing an initial screening set for job matching requests using the dual-tower model is as follows:

[0049] For job matching requests, the dual-tower model is used to obtain the resume embedding vector corresponding to the job seeker's resume and the job embedding vector of all jobs. Then, the cosine similarity between the resume embedding vector and the embedding vector of each job is calculated, and the jobs with a cosine similarity greater than the preset similarity threshold are formed into an initial screening set.

[0050] For resume matching requests, the job embedding vector corresponding to the job description of the enterprise and the resume embedding vector of all resumes are obtained through the dual-tower model. Then, the cosine similarity between the job embedding vector and the embedding vector of each resume is calculated. Resumes with a cosine similarity greater than the preset similarity threshold are formed into an initial screening set.

[0051] The method for using a matching model to select the final matching results from the initial screening set is as follows:

[0052] For job matching requests, the job seeker's resume and the job descriptions in the initial screening set are input into the matching model to obtain the matching degree between the resume and each job in the initial screening set. The job with the highest matching degree is the final matching result.

[0053] For a resume matching request, the job description of the company and each resume in the initial screening set are input into the matching model to obtain the matching degree between the job description and each resume in the initial screening set. The resume with the highest matching degree is the final matching result.

[0054] A deep learning-based intelligent human-job matching system based on a large language model and multiple prompts is proposed, comprising a dataset collection module, a dual-tower model training module, a matching model training module, and a matching module; wherein the modules are electrically connected to each other.

[0055] The dataset collection module collects resume datasets and job posting datasets, and sends them to the dual-tower model training module and the matching model training module.

[0056] The dual-tower model training module constructs a dual-tower model that takes resume datasets and job datasets as inputs, resume embedding vectors and job embedding vectors as outputs, and the matching similarity of embedding vectors as the prediction target. The module trains the dual-tower model and sends the trained dual-tower model to the matching module.

[0057] The matching model training module constructs a matching model that takes the resume dataset and job dataset as input and the matching degree as output, trains the matching model, and sends the trained matching model to the matching module.

[0058] The matching module collects job matching requests and provides an initial screening set of these requests using a dual-tower model. Then, the matching model is used to filter the final matching results from the initial screening set.

[0059] An electronic device is proposed, comprising: a processor and a memory, wherein the memory stores a computer program that can be called by the processor;

[0060] The processor executes the aforementioned deep learning-based intelligent human-job matching method based on a large language model and multiple prompts by calling the computer program stored in the memory.

[0061] A computer-readable storage medium is proposed, on which an erasable and rewritable computer program is stored;

[0062] When the computer program is run on a computer device, the computer device executes the aforementioned deep learning-based intelligent human-job matching method based on a large language model and multiple prompts.

[0063] Compared with the prior art, the beneficial effects of the present invention are:

[0064] This invention collects resume and job datasets, constructs a dual-tower model that takes these datasets as input, outputs resume and job embedding vectors, and uses the matching similarity of the embedding vectors as the prediction target. The dual-tower model is trained to construct a matching model that takes the resume and job datasets as input and outputs matching scores. This matching model is also trained. The invention collects job-person matching requests and uses the dual-tower model to provide an initial screening set. The matching model then filters the final matching results from this initial set. By training the dual-tower and matching models, the invention enables the generation of context vectors for job positions and resumes, which are stored in a database to support rapid resume and job retrieval. This allows for flexible retrieval in various ways, such as "resume vs. job," "resume vs. resume," and "job vs. job."

[0065] The matching model predicts the match between job postings and resumes. Segmenting the resume and job posting text ensures that detailed information is preserved even in long resumes. Merging the job posting as a prompt with the resume enhances the model's semantic understanding and predictive performance, making it more suitable for complex and diverse job and resume matching scenarios. Through the collaborative work of these two models, efficient and accurate job-person matching is achieved. Attached Figure Description

[0066] Figure 1 This is a flowchart of the deep learning-based intelligent human-job matching method based on a large language model and multiple prompts in Embodiment 1 of the present invention;

[0067] Figure 2 This is a schematic diagram of the twin-tower model in Embodiment 1 of the present invention;

[0068] Figure 3 This is a schematic diagram of the matching model in Embodiment 1 of the present invention;

[0069] Figure 4 This is a module connection diagram of the deep learning intelligent human-job matching system based on a large language model and multiple prompts in Embodiment 2 of the present invention. Detailed Implementation

[0070] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] Example 1

[0072] like Figure 1As shown, the deep learning-based intelligent person-job matching method based on a large language model and multiple prompts includes the following steps:

[0073] Step 1: Collect resume datasets and job posting datasets;

[0074] Step 2: Construct a dual-tower model that takes the resume dataset and job dataset as input, the resume embedding vector and the job embedding vector as output, and the matching similarity of the embedding vectors as the prediction target, and train the dual-tower model.

[0075] Step 3: Construct a matching model that takes the resume dataset and job dataset as input and the matching score as output, and train the matching model;

[0076] Step 4: Collect job matching requests, and use the dual-tower model to provide an initial screening set of job matching requests. Then, use the matching model to filter out the final matching results from the initial screening set.

[0077] The method for collecting the resume dataset and job dataset is as follows:

[0078] The resume dataset is composed of all resumes uploaded by job seekers and all job postings uploaded by companies.

[0079] Specifically, the recruitment sites include, but are not limited to, offline or online sites that provide job matching services, such as human resources markets and recruitment websites;

[0080] The resume data includes, but is not limited to:

[0081] Job seeker's personal information: name, gender, age, education, etc.;

[0082] Work experience: past positions, years of service, job responsibilities, project experience, etc.;

[0083] Skills and expertise: Proficiency in programming languages, frameworks, tools, etc.

[0084] Awards and certificates: including educational and professional certificates and awards;

[0085] Personal description: A statement about one's abilities, career plans, etc.;

[0086] The job dataset includes, but is not limited to, company descriptions, job descriptions, and job requirements.

[0087] It should be noted that the job postings and resumes involved in this application were obtained with full consent and authorization, and the collection, use and processing of such information must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0088] Furthermore, the method for constructing the dual-tower model, which takes the resume dataset and job dataset as input and outputs the resume embedding vector and job embedding vector, is as follows:

[0089] like Figure 2 The diagram shows a dual-tower model, a model training method that maps data from different modalities to the same vector space. Resumes and job descriptions are treated as two different modalities; by ensuring that the embedding vectors of resumes and job descriptions have the same dimension, the model can map them to the same vector space.

[0090] The dual-tower model includes a job tower and a resume tower;

[0091] The job tower includes an input layer, a slice layer, an embedded model, a fully connected layer, an attention pooling layer, and an output layer;

[0092] The resume tower includes an input layer, a slice layer, an embedding model, a Transformer encoding layer, an attention pooling layer, and an output layer;

[0093] The input layer of the job tower takes the job description of each job in the job dataset as input.

[0094] The job tower's slicing layer and embedding model are used to segment and vectorize job descriptions to obtain initial job embedding vectors.

[0095] The fully connected layer and attention pooling layer of the job tower are conventional techniques in this field. The initial job embedding vector is transformed into a job embedding vector of length d through the fully connected layer and attention pooling layer; d is the preset embedding vector length.

[0096] Specifically, the method for segmenting and vectorizing the job description to obtain the initial embedding vector is as follows:

[0097] The job tower's slice layer reads M core content items for each job from the job dataset; M is the preset number of core content items.

[0098] Generally, M is set to 3, and the core content described in M ​​items corresponds to the company description, job description, and job requirements for each position.

[0099] Input the M core contents into the job tower embedding model to obtain a description embedding sequence of length M output by the embedding model, which is used as the initial job embedding vector.

[0100] The input layer of the resume tower takes the resume text of each job seeker in the resume dataset as input;

[0101] The slicing layer and embedding model of the resume tower are used to segment and vectorize the resume text to obtain the initial embedding vector of the resume.

[0102] Specifically, the method for segmenting and vectorizing the resume text to obtain the initial embedding vector of the resume is as follows:

[0103] The slicing layer of Resume Tower divides the resumes of all job seekers in the resume dataset into a resume slice sequence according to the preset resume text segmentation rules.

[0104] Input the resume slice sequence into the embedding model of Resume Tower to obtain the initial resume embedding vector output by the embedding model;

[0105] It should be noted that because resumes vary greatly in length and structure, it's not easy to segment them according to a fixed structure. Therefore, a set of segmentation rules needs to be pre-defined for segmentation, such as simple fixed-length text segmentation or more complex segmentation using natural language processing techniques and machine learning models. The specific approach needs to be evaluated based on factors such as project resource investment and experimental results. Since each resume varies in length, the final length of the resulting resume segment sequence will also differ.

[0106] In machine learning and natural language processing, embedding models refer to the process of mapping high-dimensional data (such as text, images, and videos) to a low-dimensional space. Simply put, an embedding vector is a multi-dimensional real-valued vector that represents the input data as points in a continuous numerical space. Mainstream embedding vector construction methods include, but are not limited to, BERT models and CimCSE models.

[0107] The Transformer encoding layer and attention pooling layer of the resume tower are conventional techniques in the field. The Transformer Encoder is used to process the non-fixed-length resume vector sequence, and the attention pooling is used to compress the variable-length sequence into a fixed-length embedding vector d.

[0108] Furthermore, the method for training the dual-tower model is as follows:

[0109] Job positions and resumes are matched manually, and each match is labeled.

[0110] It is understandable that the method of labeling each matching pair can be to manually select matching pairs that are more consistent between the job posting and the resume and label them as 1, while labeling the remaining matching pairs of job posting and resume as 0.

[0111] Use labeled matching pairs to train the dual-tower model;

[0112] Specifically, distance metrics (such as dot product, cosine similarity, and Euclidean distance) are used to calculate the matching similarity between the resume embedding vector and the job embedding vector;

[0113] Considering that only a small portion of the resumes received are screened and ultimately offer positions, this reflects an imbalanced sample problem in the dataset. Therefore, the following mixture loss function is used to calculate the model loss and update the model weights accordingly:

[0114] ;

[0115] in, For model weights, To compare the losses, For FocalLoss, and All are preset weighting coefficients;

[0116] The contrastive loss ensures that matched sample pairs (positive class) have smaller distances, while mismatched sample pairs (negative class) have larger distances, as expressed in the formula below:

[0117] ;

[0118] Where N is the number of all matching pairs. These are the resume embedding vector and the job title embedding vector in the matching pair, respectively. To match the labeled values, margin is a set boundary value that controls the minimum distance between negative samples; max is a maximum value function used to take the maximum value among several values; d is a distance metric function, which can be Euclidean distance.

[0119] Focal Loss addresses the class imbalance problem in resume matching by reducing the loss contribution of simple samples and increasing the weight of hard-to-classify samples. The formula is as follows:

[0120] ;

[0121] in, It is the weight of the category, usually the inverse ratio of the number of samples in the positive or negative class (labeled 1 or 0);

[0122] This is an adjustment parameter that controls the Focal Loss's focus on difficult samples; it is usually set to 2.

[0123] It is the matching probability output by the dual-tower model. It is the actual label value of each matching pair;

[0124] The two-tower model described above is a common architecture, particularly suitable for handling two different types of input (such as resumes and job requirements, which, although both are text, have very different content structures). Each tower independently encodes one type of input, generating its own embedding vector.

[0125] This structure has the following advantages:

[0126] - Independent optimization: Each tower can be optimized for a specific type of input, thereby better capturing the unique properties of its respective input.

[0127] - High-efficiency computation: Once the embedding vectors are generated, matching can be performed through simple vector operations (such as cosine similarity), which greatly improves the response speed of online services.

[0128] - Scalability: Embedded vectors can be pre-computed and stored offline, enabling the system to handle large-scale datasets.

[0129] Furthermore, the multi-prompt technique refers to using multiple prompts when processing text to enhance the model's semantic understanding and context awareness. By integrating job requirements as prompts with resumes for matching and prediction, the correlation between the two can be better captured, improving the accuracy and robustness of the matching.

[0130] Therefore, the method for constructing a matching model that takes resume datasets and job datasets as input and the matching degree as output is as follows:

[0131] like Figure 3 The diagram shows a matching model, which includes a job description processing part, a resume text processing part, and a merging and matching part.

[0132] The processing procedure for the job description section includes:

[0133] The job description input for each job in the job dataset is segmented into layers to obtain M core content items;

[0134] In this embodiment, M is 3, and the corresponding core contents are company description, job description and job requirements;

[0135] Each core element is input into the embedding model to obtain the embedding vector corresponding to the core element.

[0136] In this embodiment, that is Figure 3 Embedded company description, embedded job description, and embedded job requirements;

[0137] It is understandable that the layering and embedding model in the job description processing section is consistent with the layering and embedding model in the job pyramid of the dual-tower model.

[0138] The embedding vector of each core content item is input into the fully connected layer to obtain the corresponding hint embedding vector of the core content.

[0139] In this embodiment, that is Figure 3 Embedded company description prompts, embedded job description prompts, and embedded job requirement prompts;

[0140] By embedding and stacking the prompts for various core contents, job prompts can be obtained.

[0141] Furthermore, the processing procedure of the resume text processing section includes:

[0142] Input the resume text of each job seeker in the resume dataset into the segmentation and embedding model in sequence to obtain the initial embedding vector of the resume output by the embedding model;

[0143] The initial resume embedding vector is then input into the Transformer encoding layer and the attention pooling layer in sequence to obtain the resume embedding vector.

[0144] It is understood that the functions of the segmentation layer, embedding model, Transformer encoding layer, and attention pooling layer in the resume text processing part are consistent with those of the segmentation layer, embedding model, Transformer encoding layer, and attention pooling layer in the resume tower of the dual-tower model.

[0145] Furthermore, the processing of the merging and matching part includes:

[0146] Input the job prompt embedding and resume embedding vectors into the multi-head attention layer to obtain the merged embedding vector;

[0147] The merged embedding vector is input into the fully connected layer to obtain an intermediate result vector;

[0148] The intermediate result vector is input into the sigmoid function layer, and the output is a real value in the range of [0, 1], which is the classification probability value, and is used as the confidence of the classification result, as well as the matching degree.

[0149] Set a classification threshold. Samples with values ​​greater than or equal to the threshold are classified as positive, meaning the resume matches the job posting; samples with values ​​less than the threshold are classified as negative, meaning the resume does not match the job posting.

[0150] Furthermore, the method for training the matching model is as follows:

[0151] Use labeled matching pairs to train the matching model;

[0152] Due to the imbalance of sample classes, Focal Loss is used to calculate the training loss and update the parameter weights of the matching model;

[0153] The process of collecting job matching requests involves collecting job seekers' resumes through the backend of recruitment websites and generating job matching requests for them, as well as collecting job descriptions from companies and generating resume matching requests for them.

[0154] The method of providing an initial screening set for job matching requests through the dual-tower model is as follows:

[0155] For job matching requests, the dual-tower model is used to obtain the resume embedding vector corresponding to the job seeker's resume and the job embedding vector of all jobs. Then, the cosine similarity between the resume embedding vector and the embedding vector of each job is calculated, and the jobs with a cosine similarity greater than the preset similarity threshold are formed into an initial screening set.

[0156] For resume matching requests, the job embedding vector corresponding to the job description of the enterprise and the resume embedding vector of all resumes are obtained through the dual-tower model. Then, the cosine similarity between the job embedding vector and the embedding vector of each resume is calculated. Resumes with a cosine similarity greater than the preset similarity threshold are formed into an initial screening set.

[0157] In a further preferred embodiment, for the dual-tower model, various retrieval methods can be supported by calculating cosine similarity, including:

[0158] Use the job search function to find resumes and list all potential job candidates;

[0159] Use your resume to search for jobs and find all available positions to apply for.

[0160] Use the job search function to find other similar jobs;

[0161] Use the resume search function to find resumes and view similar applicants.

[0162] Furthermore, the method for using a matching model to filter the final matching results from the initial screening set is as follows:

[0163] For job matching requests, the job seeker's resume and the job descriptions in the initial screening set are input into the matching model to obtain the matching degree between the resume and each job in the initial screening set. The job with the highest matching degree is the final matching result.

[0164] For a resume matching request, the job description of the company and each resume in the initial screening set are input into the matching model to obtain the matching degree between the job description and each resume in the initial screening set. The resume with the highest matching degree is the final matching result.

[0165] The following are some practical application scenarios of this invention, demonstrating how the dual-tower model and matching model can provide convenient and quick assistance to job seekers and companies:

[0166] Scenario 1: Online Recruitment Platform - Recommended Jobs

[0167] Application scenario: Matching job seekers' resumes with job descriptions posted by companies, and recommending jobs to job seekers;

[0168] Specific implementation:

[0169] The platform collects a large amount of resume and job description data, uses a dual-tower model to generate context vectors for resumes and job descriptions, and stores them in the database.

[0170] When job seekers upload their resumes, the platform obtains a resume embedding vector through the resume pyramid of the dual-pyramid model. The platform can then quickly retrieve a preliminary matching job description using this resume embedding vector.

[0171] The matching model is then used to predict the matching results of the retrieved job descriptions and resumes, and the prediction results and matching confidence scores are given.

[0172] Filter out job descriptions that are not a good match or do not meet the confidence threshold, and recommend suitable positions to job seekers.

[0173] Scenario 2: Online Recruitment Platform - Recommended Resumes

[0174] Application scenario: Matching job descriptions posted by companies with job seekers' resumes and recommending resumes to companies;

[0175] Specific implementation:

[0176] The platform collects a large amount of resume and job description data, uses a dual-tower model to generate context vectors for resumes and job descriptions, and stores them in the database.

[0177] When a company posts a job, it obtains a job embedding vector through the job pyramid of the dual-pyramid model. The platform then uses this job embedding vector to quickly find a preliminary match in the resume.

[0178] The matching model is then used to predict the matching results between the retrieved resumes and job descriptions, and the prediction results and matching confidence scores are given.

[0179] Filter out resumes that are not a good match or do not meet the confidence threshold, and recommend suitable resumes to companies.

[0180] Scenario 3: Human Resource Management System

[0181] Application scenario: The method of this invention can be used in an enterprise's internal human resource management system to optimize the allocation of internal talent;

[0182] Specific implementation:

[0183] Companies collect employee resumes and internal job descriptions, use a dual-tower model to generate context vectors for employee resumes, and store them in a database.

[0184] When there are new job openings, the job embedding vector is obtained through the job pyramid of the dual-pyramid model. The platform can then quickly retrieve and find a preliminary match of employee resumes using this job embedding vector.

[0185] Further, the matching model is used to make matching predictions, and the prediction results and prediction confidence are output.

[0186] The system displays matching results and confidence levels to help the HR department make decisions.

[0187] Similarly, employee resumes can be used to match vacant internal positions, thereby facilitating internal talent transfer and quickly finding suitable internal positions for employees who are idle after the project ends.

[0188] Scenario 4: Career Counseling and Planning

[0189] Application scenarios: Career counseling agencies can use the method of this invention to help job seekers with career planning and job recommendations;

[0190] Specific implementation:

[0191] Consulting firms collect job descriptions from the job market, use a two-tower model to generate job description context vectors, and store them in a database.

[0192] When job seekers seek career advice, the system can quickly find preliminary matching positions by searching through the job seeker's resume.

[0193] Further, the matching model is used to make matching predictions, and the prediction results and prediction confidence are output.

[0194] The system displays matching results and confidence levels, helping job seekers better understand their career development direction.

[0195] Scenario 5: Education and Training Industry

[0196] Application scenarios: The education and training industry can use the method of this invention to recommend suitable career paths and courses for students;

[0197] Specific implementation:

[0198] Educational institutions collect students' resumes and job descriptions based on market demand, use a dual-tower model to generate context vectors, and store them in a database.

[0199] Once trainees complete their training, the system can quickly find preliminary matching job postings for them.

[0200] Further, the matching model is used to make matching predictions, and the prediction results and prediction confidence are output.

[0201] The system displays the matching results and confidence levels, and recommends the most suitable career paths and related courses for students.

[0202] Example 2

[0203] like Figure 4 As shown, the deep learning-based intelligent human-job matching system based on a large language model and multiple prompts includes a dataset collection module, a dual-tower model training module, a matching model training module, and a matching module; wherein, the modules are connected to each other electrically.

[0204] The dataset collection module collects resume datasets and job posting datasets, and sends them to the dual-tower model training module and the matching model training module.

[0205] The dual-tower model training module constructs a dual-tower model that takes resume datasets and job datasets as inputs, resume embedding vectors and job embedding vectors as outputs, and the matching similarity of embedding vectors as the prediction target. The module trains the dual-tower model and sends the trained dual-tower model to the matching module.

[0206] The matching model training module constructs a matching model that takes the resume dataset and job dataset as input and the matching degree as output, trains the matching model, and sends the trained matching model to the matching module.

[0207] The matching module collects job matching requests and provides an initial screening set of these requests using a dual-tower model. Then, the matching model is used to filter the final matching results from the initial screening set.

[0208] Example 3

[0209] According to another aspect of this application, an electronic device is also provided. The electronic device may include one or more processors and one or more memories. The memories store computer-readable code that, when executed by the one or more processors, can perform the deep learning-based intelligent human-job matching method based on a large language model and multiple prompts as described above.

[0210] Electronic devices may include buses, one or more CPUs, ROMs, RAMs, communication ports connected to a network, input / output components, hard disks, etc. Storage devices within the electronic device, such as ROMs or hard disks, may store the deep learning-based intelligent human-job matching method based on large language models and multiple prompts provided in this application.

[0211] Furthermore, the electronic device may also include a user interface. Of course, this architecture is merely exemplary, and one or more components in the electronic device may be omitted as needed when implementing different devices.

[0212] Example 4

[0213] A computer-readable storage medium according to one embodiment of this application. The computer-readable storage medium stores computer-readable instructions. When the computer-readable instructions are executed by a processor, the deep learning-based intelligent human-job matching method based on a large language model and multiple prompts according to an embodiment of this application, as described with reference to the above figures, can be executed. The computer-readable storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0214] Furthermore, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, this application provides a non-transitory machine-readable storage medium storing machine-readable instructions that can be executed by a processor to perform instructions corresponding to the method steps provided in this application. When the computer program is executed by a central processing unit (CPU), it performs the functions defined in the method of this application.

[0215] The methods, apparatus, and devices of this application may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the method is for illustrative purposes only, and the steps of the method of this application are not limited to the order specifically described above, unless otherwise specifically stated. Furthermore, in some embodiments, this application may also be implemented as a program recorded on a recording medium, the program including machine-readable instructions for implementing the method according to this application. Thus, this application also covers recording media storing programs for performing the method according to this application.

[0216] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0217] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0218] The preset parameters or preset thresholds mentioned above are all set by those skilled in the art based on actual conditions or obtained through large-scale data simulation.

[0219] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A deep learning-based intelligent person-job matching method based on large language models and multiple prompts, characterized in that, Includes the following steps: Step 1: Collect resume datasets and job posting datasets; Step 2: Construct a dual-tower model that takes the resume dataset and job dataset as input, the resume embedding vector and the job embedding vector as output, and the matching similarity of the embedding vectors as the prediction target, and train the dual-tower model. Step 3: Construct a matching model that takes the resume dataset and job dataset as input and the matching score as output, and train the matching model; Step 4: Collect job matching requests, and use the dual-tower model to provide an initial screening set of job matching requests. Then, use the matching model to filter out the final matching results from the initial screening set. The dual-tower model includes a job tower and a resume tower; The job tower includes an input layer, a slice layer, an embedded model, a fully connected layer, an attention pooling layer, and an output layer; The resume tower includes an input layer, a slice layer, an embedding model, a Transformer encoding layer, an attention pooling layer, and an output layer; The method for training the dual-tower model is as follows: Job positions and resumes are matched manually, and each match is labeled. Use labeled matching pairs to train the dual-tower model; Use a distance metric to calculate the matching similarity between resume embedding vectors and job embedding vectors; The model loss is calculated using the following mixture loss function, and the model weights are updated accordingly: ； in, For model weights, To compare the losses, For Focal Loss, and All are preset weighting coefficients; Among them, the contrastive loss ensures that the distance between matched samples is smaller, while the distance between unmatched sample pairs is larger; FocalLoss addresses the class imbalance problem in resume matching by reducing the loss contribution of simple samples and increasing the weight of hard-to-classify samples. The method for constructing a matching model that takes resume and job datasets as input and matches the degree of matching as output is as follows: The matching model includes a job description processing part, a resume text processing part, and a merging and matching part; The processing procedure for the job description section includes: The job description input for each job in the job dataset is segmented into layers to obtain M core content items; Each core element is input into the embedding model to obtain the embedding vector corresponding to the core element. The embedding vector of each core content item is input into the fully connected layer to obtain the corresponding hint embedding vector of the core content. By stacking the prompts for various core contents, job prompts are obtained; The processing steps for the resume text processing section include: Input the resume text of each job seeker in the resume dataset into the segmentation and embedding model in sequence to obtain the initial embedding vector of the resume output by the embedding model; The initial resume embedding vector is then input into the Transformer encoding layer and the attention pooling layer in sequence to obtain the resume embedding vector. The processing steps for the merging and matching part include: Input the job prompt embedding and resume embedding vectors into the multi-head attention layer to obtain the merged embedding vector; The merged embedding vector is input into the fully connected layer to obtain an intermediate result vector; The intermediate result vector is input into the sigmoid function layer, and the output is a real value in the range of [0, 1], which is the classification probability value, and is used as the confidence of the classification result, as well as the matching degree. Set a classification threshold. Samples with values ​​greater than or equal to the threshold are classified as positive, meaning the resume matches the job posting; samples with values ​​less than the threshold are classified as negative, meaning the resume does not match the job posting.

2. The deep learning-based intelligent person-job matching method based on a large language model and multiple prompts as described in claim 1, characterized in that, The method for collecting resume datasets and job posting datasets is as follows: The database retrieves resume data uploaded by all job seekers from the recruitment site's backend database to form a resume dataset, and combines this with job posting data uploaded by all companies to form a job posting dataset.

3. The deep learning-based intelligent person-job matching method based on a large language model and multiple prompts as described in claim 2, is characterized in that... The method for constructing a dual-tower model that takes resume datasets and job datasets as input and outputs resume embedding vectors and job embedding vectors is as follows: The input layer of the job tower takes the job description of each job in the job dataset as input; The job tower's slicing layer and embedding model are used to segment and vectorize job descriptions to obtain initial job embedding vectors. The input layer of the resume tower takes the resume text of each job seeker in the resume dataset as input; The slicing layer and embedding model of the resume tower are used to segment and vectorize the resume text to obtain the initial embedding vector of the resume.

4. The deep learning-based intelligent person-job matching method based on a large language model and multiple prompts as described in claim 3, is characterized in that... The process of collecting job matching requests involves collecting job seekers' resumes through the backend of recruitment websites and generating job matching requests for them, as well as collecting job descriptions from companies and generating resume matching requests for them.

5. The deep learning-based intelligent person-job matching method based on a large language model and multiple prompts as described in claim 4, characterized in that, The method of providing an initial screening set for job matching requests using the dual-tower model is as follows: For job matching requests, the dual-tower model is used to obtain the resume embedding vector corresponding to the job seeker's resume and the job embedding vector of all jobs. Then, the cosine similarity between the resume embedding vector and the embedding vector of each job is calculated, and the jobs with a cosine similarity greater than the preset similarity threshold are formed into an initial screening set. For resume matching requests, the job embedding vector corresponding to the job description of the enterprise and the resume embedding vector of all resumes are obtained through the dual-tower model. Then, the cosine similarity between the job embedding vector and the embedding vector of each resume is calculated. Resumes with a cosine similarity greater than the preset similarity threshold are formed into an initial screening set.

6. The deep learning-based intelligent person-job matching method based on a large language model and multiple prompts as described in claim 5, is characterized in that... The method for using a matching model to select the final matching results from the initial screening set is as follows: For job matching requests, the job seeker's resume and the job descriptions in the initial screening set are input into the matching model to obtain the matching degree between the resume and each job in the initial screening set. The job with the highest matching degree is the final matching result. For a resume matching request, the job description of the company and each resume in the initial screening set are input into the matching model to obtain the matching degree between the job description and each resume in the initial screening set. The resume with the highest matching degree is the final matching result.

7. A deep learning-based intelligent job matching system based on a large language model and multiple prompts, used to implement the deep learning-based intelligent job matching method based on a large language model and multiple prompts as described in any one of claims 1-6, characterized in that, It includes a dataset collection module, a dual-tower model training module, a matching model training module, and a matching module; the modules are connected electrically. The dataset collection module collects resume datasets and job posting datasets, and sends them to the dual-tower model training module and the matching model training module. The dual-tower model training module constructs a dual-tower model that takes resume datasets and job datasets as inputs, resume embedding vectors and job embedding vectors as outputs, and the matching similarity of embedding vectors as the prediction target. The module trains the dual-tower model and sends the trained dual-tower model to the matching module. The matching model training module constructs a matching model that takes the resume dataset and job dataset as input and the matching degree as output, trains the matching model, and sends the trained matching model to the matching module. The matching module collects job matching requests and provides an initial screening set of these requests using a dual-tower model. Then, the matching model is used to filter the final matching results from the initial screening set.

8. An electronic device, characterized in that, include: Processor and memory, wherein: The memory stores computer programs that can be called by the processor; The processor executes the deep learning-based intelligent human-job matching method based on a large language model and multiple prompts as described in any one of claims 1-6 in the background by calling the computer program stored in the memory.

9. A computer-readable storage medium, characterized in that, It contains erasable and rewritable computer programs; When the computer program is run on a computer device, the computer device executes the deep learning-based intelligent human-job matching method based on a large language model and multiple prompts as described in any one of claims 1-6 in the background.

Images