A multi-dimensional resume intelligent screening and post matching method based on a large model
By using large language models and multi-algorithm fusion technology, we can achieve multi-dimensional feature extraction and global optimal matching between resumes and job positions. This solves the problems of low efficiency and insufficient accuracy in traditional resume screening, improves recruitment efficiency and accuracy, and reduces human resource management costs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Filing Date
- 2026-04-04
- Publication Date
- 2026-07-10
AI Technical Summary
Traditional resume screening and job matching are inefficient and lack accuracy. They cannot fully extract multi-dimensional features and have weak semantic understanding capabilities, resulting in high human resource management costs, long recruitment cycles, and the loss of high-quality talent.
This method employs a large language model combined with hierarchical attention networks and knowledge graph fusion techniques. It extracts multi-dimensional features from resumes and job postings using a BiLSTM-CRF model, achieves global optimal matching using cosine similarity and the Hungarian algorithm, and outputs efficient matching results by combining adaptive threshold filtering.
It significantly improves resume screening efficiency and matching accuracy, reduces human resource management costs, shortens the recruitment cycle, reduces subjective judgment bias, and achieves efficient and intelligent screening and matching.
Smart Images
Figure CN122364442A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence and recruitment technology, specifically to a multi-dimensional intelligent resume screening and job matching method based on a large model.
[0002] More specifically, this invention belongs to the interdisciplinary field of natural language processing, machine learning, and talent recruitment. It utilizes the semantic understanding capabilities of large language models, the association analysis capabilities of knowledge graphs, and multi-algorithm fusion technology to achieve accurate extraction of multi-dimensional features of resumes and efficient matching with job requirements. It is applicable to scenarios such as corporate recruitment and human resource management, and solves the technical problems of low efficiency and insufficient matching accuracy in traditional resume screening. Background Technology
[0003] In modern corporate recruitment, resume screening and job matching are core aspects of human resource management. With the widespread use of online recruitment, companies often receive hundreds, thousands, or even tens of thousands of resumes for each recruitment drive. Traditional manual screening methods are not only time-consuming and labor-intensive, but also suffer from problems such as subjective judgment bias, inconsistent screening standards, and low efficiency. Especially for medium and large enterprises or popular positions, screening a large number of invalid resumes consumes a significant amount of time and energy for the human resources department, leading to extended recruitment cycles and the loss of high-quality talent.
[0004] To address the aforementioned issues, several resume screening and matching systems have emerged in the existing technology field, primarily employing keyword matching, simple rule matching, or traditional machine learning models. However, these existing technologies still have significant limitations: First, keyword matching is too mechanical, failing to understand the semantic relationship between the resume and the job description, easily leading to situations where "keywords are matched but skills do not match" or "skills match but keywords are not matched," resulting in low matching accuracy. Second, traditional machine learning models have limited semantic understanding capabilities of resume text, making it difficult to extract implicit skill features from resumes. They can only match based on basic information and explicit skills, failing to meet the needs of enterprises for multi-dimensional talent evaluation. Third, most existing technologies do not incorporate the strong semantic understanding capabilities of large models, resulting in insufficient parsing accuracy for unstructured resumes and an inability to achieve globally optimal matching between resumes and jobs, easily leading to matching omissions or biases.
[0005] In recent years, large language models have made groundbreaking progress in the field of natural language processing, possessing powerful semantic understanding, text parsing, and logical reasoning capabilities, and accurately capturing deep semantic information in text. Applying large models to resume screening and job matching has become an effective way to address the pain points of existing technologies. However, currently, there is no mature solution in existing technologies that combines large models with multi-dimensional feature extraction and multi-algorithm fusion to achieve accurate resume screening and efficient job matching, failing to fully leverage the technical advantages of large models and still suffering from problems such as incomplete feature extraction, simplistic matching logic, and insufficient adaptability. Therefore, developing a multi-dimensional intelligent resume screening and job matching method based on large models has significant practical significance and application value. Summary of the Invention
[0006] The purpose of this invention is to overcome the shortcomings of existing technologies, such as low resume screening efficiency, insufficient matching accuracy, incomplete feature extraction, and weak semantic understanding. It provides a multi-dimensional intelligent resume screening and job matching method based on a large model. By utilizing the strong semantic understanding capability of the large model and combining it with multi-algorithm fusion technology, it can achieve accurate extraction of multi-dimensional features of resumes and efficient matching with job requirements, thereby improving resume screening efficiency and matching accuracy, reducing human resource management costs, and helping enterprises quickly discover high-quality talents.
[0007] To achieve the above-mentioned objectives, this invention provides a multi-dimensional intelligent resume screening and job matching method based on a large model, specifically including the following steps:
[0008] Resume and Job Information Preprocessing: Collect resume files in various formats and job description texts. Perform format parsing and noise reduction on the resume files to remove invalid and redundant information. Extract key information from the job description text to obtain a standardized set of resumes and a standardized set of job descriptions.
[0009] A multi-dimensional resume feature extraction model is constructed based on a large model. A large language model fine-tuned using SFT in the recruitment domain is adopted as the base model, combined with hierarchical attention network theory, to extract basic information features, professional skills features, work experience features, professional qualities features, and implicit ability features from resumes.
[0010] An improved BERT-Graph knowledge graph fusion theory is adopted to construct a technology stack association network in the recruitment domain. The resume text embedding vector output by the large model is fused with the knowledge graph node embedding vector to obtain the initial feature vector: hr = λ・hbert + (1-λ)・hkg
[0011] The initial feature vector is optimized by a triple attention mechanism at the word, sentence, and paragraph levels. The attention weights are calculated using the Softmax function: αi = exp (wT・hr,i) / Σexp (wT・hr,k).
[0012] By combining the BiLSTM-CRF model to accurately extract time series features, a structured feature vector is obtained. This structured feature vector is then fused with the weighted initial feature vector to generate the final multi-dimensional feature vector, VR.
[0013] Job Feature Extraction. Using the same large model and processing methods as resume feature extraction, multi-dimensional feature extraction is performed on the job description text to obtain the job feature vector vj.
[0014] Multi-dimensional matching model construction and global optimal matching. Based on the cosine similarity algorithm combined with the Hungarian algorithm, a multi-dimensional matching model is constructed to calculate the comprehensive matching degree between resumes and job positions, and to achieve global optimal matching.
[0015] The cosine similarity algorithm is used to calculate the feature matching degree of each dimension: sim (vr,k,vj,k) = (vr,k・vj,k) / (||vr,k||・||vj,k||)
[0016] The weights of each feature dimension are determined using the analytic hierarchy process (AHP), and the overall matching degree is calculated using a weighted average: match(r,j) = Σωk・sim(vr,k,vj,k)
[0017] The Hungarian algorithm is used to perform globally optimal allocation of resumes and job positions, avoiding talent mismatch caused by local optima.
[0018] Screening and Output. An adaptive matching threshold is set, determined based on the F1 maximization principle: F1 = 2PR / (P+R). High-match resume-job pairs are selected, and matching details and overall scores are output to provide a basis for recruitment decisions.
[0019] Compared with the prior art, the present invention has the following beneficial effects:
[0020] Screening efficiency is significantly improved: The large model enables automated parsing, feature extraction, and matching of resumes and job information, replacing traditional manual screening and simple keyword matching, greatly reducing the workload of the human resources department and effectively shortening the recruitment cycle.
[0021] Significantly improved matching accuracy: By integrating the semantic understanding capabilities of large-scale models with the association analysis capabilities of knowledge graphs, extracting explicit and implicit multi-dimensional features, and combining multiple algorithms to fuse matching models, the matching accuracy is significantly improved, reducing the omission of high-quality talents.
[0022] Highly adaptable and flexible: The large model has been fine-tuned specifically for the recruitment field, and can adapt to resumes and job positions in various formats and industries. It supports flexible adjustment of weights according to recruitment needs and has strong versatility.
[0023] High level of intelligence: It realizes fully automated screening, outputs structured matching results, reduces subjective judgment bias, and improves the scientific nature of recruitment decisions. Attached Figure Description
[0024] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.
[0025] Figure 1 is a flowchart of the overall system of the present invention;
[0026] Figure 2 is a schematic diagram of the multi-dimensional resume feature extraction model structure;
[0027] Figure 3 is a diagram illustrating the principle of multi-dimensional matching calculation between resumes and job positions. Detailed Implementation
[0028] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.
[0029] As shown in Figure 1, this embodiment discloses a multi-dimensional intelligent resume screening and job matching method based on a large model. The overall process includes: resume and job text collection, information preprocessing, multi-dimensional feature extraction of the large model, job feature alignment, multi-dimensional matching calculation, global optimal matching, adaptive threshold screening, and result output.
[0030] Step 1: Resume and Job Information Preprocessing. The information preprocessing module 100 performs text extraction, noise reduction, deduplication, and format standardization on PDF, Word, and image format resumes; keyword extraction and structuring are performed on JD job descriptions to obtain a unified format resume dataset and job dataset.
[0031] Step 2: Multi-dimensional resume feature extraction based on a large model. As shown in Figure 2, the recruitment domain large model module 200 includes a knowledge graph fusion unit 201, a hierarchical attention weighting unit 202, and a BiLSTM-CRF sequence feature extraction unit 203.
[0032] First, semantic embedding is performed on the resume text using a large model. Then, the knowledge graph fusion unit 201 fuses the knowledge related to skills, industry, and job position with the text embedding vector to obtain the initial feature vector: hr = λ・hbert + (1-λ)・hkg
[0033] The hierarchical attention weighting unit 202 weights word-level, sentence-level, and paragraph-level features, with the weights normalized using Softmax: αi = exp (wT・hr,i) / Σexp (wT・hr,k)
[0034] Then, the BiLSTM-CRF sequence feature extraction unit 203 is used to extract time-series features such as work experience and education experience in a structured manner, and finally outputs a multi-dimensional feature vector vr.
[0035] Step 3: Job Feature Extraction. Using the Job Feature Extraction module 300, a large model and structure completely consistent with the resume feature extraction are adopted to output a multi-dimensional feature vector vj for the job.
[0036] Step 4: Multi-dimensional matching calculation. As shown in Figure 3, the multi-dimensional matching calculation module 400 includes a single-dimensional cosine similarity calculation unit 401. The calculation formula is: sim (vr,k,vj,k) = (vr,k・vj,k) / (||vr,k||・||vj,k||)
[0037] The AHP weight allocation unit 402 assigns weights to dimensions such as skills, experience, potential, and qualities, satisfying Σωk=1; the comprehensive matching degree calculation unit 403 outputs the final matching score: match (r,j) = Σωk・sim (vr,k,vj,k)
[0038] Step 5: Global Optimal Matching. A matching matrix is constructed using the Hungarian Global Optimal Matching Module 500 to achieve a globally optimal allocation between resumes and job positions, avoiding talent mismatch caused by local optima.
[0039] Step Six: Adaptive Threshold Filtering and Output. The adaptive threshold filtering module 600 dynamically determines the filtering threshold based on the F1 maximization principle: F1 = 2PR / (P+R).
[0040] Finally, the results output and display module 700 outputs the matching ranking, scores for each dimension, matching details, and a list of recommended talents, completing the entire intelligent screening process.
Claims
1. A multi-dimensional intelligent resume screening and job matching method based on a large model, characterized in that, Includes the following steps: Step 1: Resume and Job Information Preprocessing. Collect resume files in various formats and job description texts. Perform format parsing and noise reduction on the resume files to remove invalid and redundant information. Extract key information from the job description text to obtain a standardized resume set R = {r1, r2, ..., rN} (N is the total number of resumes) and a standardized job set J = {j1, j2, ..., jM} (M is the total number of jobs). Step 2: Construct a multi-dimensional resume feature extraction model based on a large model. Use a large language model fine-tuned by SFT in the recruitment domain as the basic model. Combine it with hierarchical attention network theory to extract multi-dimensional features of the resume. The multi-dimensional features include basic information features, professional skills features, work experience features, professional qualities features and implicit ability features. An improved BERT-Graph knowledge graph fusion theory is adopted to construct a technology stack association network in the recruitment field. The resume text embedding vector output by the large model is fused with the knowledge graph node embedding vector to obtain the initial feature vector hr. The initial feature vector hr is then optimized by a triple attention mechanism at the word, sentence, and paragraph levels. The attention weight is calculated using the Softmax function, with the formula αi = exp (wT・hr,i) / Σexp (wT・hr,k), where αi is the attention weight of the i-th feature dimension, w is the attention weight parameter vector, and hr,i is the initial feature vector of the i-th dimension of the resume r. The BiLSTM-CRF model is then used to accurately extract time-series features (work experience, education experience) to obtain a structured feature vector, which is then fused with the weighted initial feature vector to generate the final multi-dimensional feature vector vr of the resume. Step 3: Using the same large model and feature extraction method, extract the multi-dimensional feature vectors vj of each job description; Step 4: Based on the cosine similarity algorithm combined with the Hungarian algorithm, construct a multi-dimensional matching model to calculate the matching degree between the resume feature vector and the job feature vector, thereby realizing resume screening and job matching; use the cosine similarity algorithm to calculate the single-dimensional feature matching degree, the formula is sim (vr,k,vj,k) = (vr,k・vj,k) / (||vr,k||・||vj,k||); use the analytic hierarchy process (AHP) to determine the weight ωk of each feature dimension, satisfying Σωk=1, and calculate the comprehensive matching degree between the resume and the job, the formula is match (r,j) = Σωk・sim (vr,k,vj,k); use the Hungarian algorithm to perform global optimal matching between the resume and the job, construct the matching matrix S, where Si,j=match (ri,jj), and solve for the one-to-one matching scheme that maximizes the total matching degree; Step 5: Set a matching threshold τ, filter out resume-job matching pairs with a comprehensive matching score match(r,j)≥τ, complete intelligent filtering and matching, and output the feature matching details and comprehensive evaluation score of each matching pair.
2. The method according to claim 1, characterized in that, The large model in step 2 is a large language model fine-tuned by SFT in the recruitment domain. The fine-tuning dataset adopts the triple (instruction, input, output) format, containing more than 15,000 resumes and corresponding structured labels. The fine-tuning process uses the AdamW optimizer, with a learning rate set to 1e-5 to 5e-5, and 30 to 50 iterations. The fine-tuning loss function adopts the cross-entropy loss function.
3. The method according to claim 1, characterized in that, The matching threshold τ in step 5 is dynamically determined using an adaptive threshold algorithm. It is iteratively optimized based on the statistical characteristics of historical matching data and the principle of maximizing the F1 score. The F1 score is calculated as F1 = 2PR / (P + R), where P is the matching precision and R is the matching recall.
4. The method according to claim 1, characterized in that, The implicit ability features in the multi-dimensional features are obtained through large model semantic reasoning. Based on natural language reasoning (NLI) theory, the communication skills, learning skills and problem-solving skills implied in the resume are inferred in combination with job requirements. The reasoning confidence threshold is set to 0.
75. Implicit ability features below this threshold are not included in the final feature vector.