An AI large model-based talent matching and intelligent recruitment method and system

Abstract

The application relates to a talent matching and intelligent recruitment method and system based on an AI large model, which comprises the following steps: acquiring multi-modal data of candidates and post demand data of a recruitment party; performing semantic extraction and feature fusion on the multi-modal data through a multi-modal fusion encoder to generate a comprehensive feature vector containing semantic correlation features and behavior features; performing field adaptation processing on the comprehensive feature vector based on a dynamically updated industry knowledge graph to generate a skill label set of the candidates; adopting a reinforcement learning model to calculate the dynamic matching degree of the skill label set and the post demand data according to historical recruitment feedback data to generate a recommendation list; and outputting the recommendation list to a terminal of the recruitment party and online optimizing the reinforcement learning model based on operation behavior data of the recruitment party. The application realizes intelligent matching of job seekers and recruitment requirements, and improves the efficiency and accuracy of recruitment.

Description

Technical Field

[0001] This application relates to the technical field of the intersection of artificial intelligence and human resources, and in particular to a talent matching and intelligent recruitment method and system based on a large AI model. Background Technology

[0002] With the expanding job market and intensifying competition, both businesses and job seekers face challenges in recruitment and job seeking. Traditional recruitment methods typically involve manual resume review and initial screening, which is not only time-consuming but also prone to human error and bias. Large Language Models (LLMs) are AI systems designed to understand and generate human language. They can learn from large amounts of text data and perform various language tasks such as text generation, translation, and question answering. In recent years, the superior capabilities of artificial intelligence (AI) technologies, especially large language models (LLMs), in processing natural language have made them ideal tools for automating the recruitment process.

[0003] Existing intelligent recruitment systems generally adopt technical approaches based on keyword matching or simple semantic analysis. For example, they may only perform word segmentation and statistics on resume text or extract text features through pre-trained language models. Such technical solutions are difficult to capture the implicit connections between resumes and job descriptions, such as unstated skill requirements in project experience. This may result in the omission of candidates with experience in mathematical modeling competitions for the "algorithm engineer" position. Summary of the Invention

[0004] To achieve intelligent matching of job seekers and recruitment needs and improve the efficiency and accuracy of recruitment, this application provides a talent matching and intelligent recruitment method and system based on an AI large model.

[0005] The above-mentioned objective of this application is achieved through the following technical solution:

[0006] A talent matching and intelligent recruitment method based on AI large-scale models includes the following steps:

[0007] The multimodal data of candidates and the job requirements data of recruiters are obtained. The multimodal data includes at least two of the following: text resumes, voice interview records and video interview recordings.

[0008] The multimodal data is semantically extracted and feature fused using a multimodal fusion encoder to generate a comprehensive feature vector containing semantic association features and behavioral features.

[0009] Based on a dynamically updated industry knowledge graph, the comprehensive feature vector is subjected to domain adaptation processing to generate a set of candidate skill tags.

[0010] A reinforcement learning model is used to dynamically calculate the matching degree between the skill tag set and the job requirement data based on historical recruitment feedback data, and a recommendation list is generated.

[0011] The recommendation list is output to the recruiter's terminal, and the reinforcement learning model is optimized online based on the recruiter's operational behavior data.

[0012] By adopting the above technical solutions and integrating multimodal data from text, voice, and video, the limitations of traditional single resume screening are overcome. This allows for the comprehensive capture of candidates' skills, emotional expressions, and behavioral characteristics, breaking through the limitations of traditional text matching. Based on reinforcement learning models, the matching strategy is optimized in real time, adapting to the rapid iteration of industry skills and the dynamic changes in job requirements. Real-time data updates solve the timeliness problem of traditional static graphs, improving the industry adaptability of skill tags and enhancing the efficiency and accuracy of talent resource matching. Recruiter feedback is directly injected into model iteration, forming a complete closed loop of "data collection-matching-feedback-optimization," continuously improving recommendation accuracy.

[0013] In a preferred embodiment, the steps of obtaining multimodal data of candidates and job requirement data of recruiters specifically include the following steps:

[0014] The process involves converting resumes of different formats into text files of a unified format, using a natural language processing model to identify the core modules in the text, extracting standardized entity information from each module, and generating text resume features.

[0015] The interview video is segmented into consecutive frames along the timeline, and facial recognition technology is used to track the candidate's facial expression changes to generate video facial expression features.

[0016] The emotional tendency is quantified by the spectral energy distribution of the voice in the interview audio, and the voice content is combined with the emotional features to generate voice emotional features.

[0017] By adopting the above technical solutions, through standardized resume formatting, keyframe extraction from videos, and speech denoising, the interference of raw data quality fluctuations on analysis is reduced. Implicit ability indicators (such as stress resistance and communication affinity) are extracted from non-textual data such as video micro-expressions and emotional fluctuations in speech, making up for the information gaps in traditional resumes and generating structured text, video, and speech feature vectors, providing consistent input for subsequent multimodal fusion.

[0018] In a preferred embodiment of this application, the step of performing semantic extraction and feature fusion on the multimodal data using a multimodal fusion encoder to generate a comprehensive feature vector containing semantic association features and behavioral features specifically includes the following steps:

[0019] Text resume features, video expression features, and voice emotion features are mapped to a unified semantic space. The weight ratio of each modality is automatically adjusted according to the job requirements being processed. A preliminary fused feature vector is generated through weighted feature fusion.

[0020] Construct cross-modal positive and negative sample pairs, and through comparative learning, shorten the cross-modal feature distance of positive sample pairs while widening the feature distance of negative sample pairs;

[0021] The initial fused feature vectors are mapped to a low-dimensional space, and redundant information is removed by a clustering algorithm to generate a compact comprehensive feature vector.

[0022] By adopting the above technical solutions, modal weights are dynamically adjusted according to job type (e.g., technical positions emphasize textual skills, sales positions strengthen voice emotion analysis), avoiding matching bias caused by fixed weights. Through comparative learning, the differences in expression between modalities are eliminated (e.g., the contradiction between "proactive" in a resume and passive body language in a video), improving the credibility of feature fusion, compressing redundant information (e.g., repeatedly described skills), generating a comprehensive feature vector with high information density, and reducing the complexity of subsequent calculations.

[0023] In a preferred embodiment of this application, the step of performing domain adaptation processing on the comprehensive feature vector based on a dynamically updated industry knowledge graph to generate a set of candidate skill tags specifically includes the following steps:

[0024] Real-time acquisition of publicly available data from industry recruitment platforms, technology forums, and companies, combined with graph neural networks to uncover the relationships between skills;

[0025] The node and edge weights in the knowledge graph are updated through incremental learning, and the accuracy of the updated graph prediction is verified by backtesting with historical data.

[0026] Align the candidate's comprehensive feature vector with the skill nodes in the knowledge graph, extract explicit labels by directly matching the candidate's skills with the graph nodes, and infer potential abilities and generate implicit labels based on the association relationship;

[0027] Adjust tag priorities based on real-time industry needs to generate a set of skill tags with confidence levels.

[0028] By adopting the above technical solutions, the incremental update mechanism enables the knowledge graph to keep up with industry trends, avoids the lag of static graphs, infers the undisclosed cross-disciplinary potential of candidates through skill association networks, broadens the dimensions of talent evaluation, and automatically adjusts the label weights according to market demand fluctuations, thereby improving the matching efficiency of in-demand positions.

[0029] In a preferred embodiment of this application, the step of using a reinforcement learning model to dynamically calculate the matching degree between the skill tag set and the job requirement data based on historical recruitment feedback data, and generating a recommendation list, specifically includes the following steps:

[0030] The candidate's skill tag set and job requirement data are converted into a unified coding format, executable matching strategy adjustment actions are defined, and reward functions are dynamically generated based on the recruitment success rate and the candidate's performance data after joining the company.

[0031] Extract historical recruitment feedback data and randomly try various matching strategies. Through repeated simulations of historical trajectories, update the strategy network parameters so that the reinforcement learning model tends to choose actions that maximize cumulative rewards.

[0032] Input the candidate's skill tag set and job requirement data to generate an initial feature combination state. Select the optimal action according to the model strategy and calculate the comprehensive matching score between the candidate and the job based on the adjusted strategy.

[0033] A recommendation list is generated based on the matching score in descending order.

[0034] By adopting the above technical solutions, the reward function integrates the instant recruitment success rate with long-term employee performance, avoiding the trap of "high matching but low retention after onboarding". Through the mapping of feature combination states and action spaces, it provides understandable matching logic, balances historical experience and exploration of new strategies in model training, and prevents innovation stagnation caused by over-reliance on past successful paths.

[0035] In a preferred embodiment of this application, the step of outputting the recommendation list to the recruiter's terminal and optimizing the reinforcement learning model online based on the recruiter's operational behavior data specifically includes the following steps:

[0036] The recommended list is output to the recruiter's terminal, and the recruiter's operations are tracked throughout the entire process, with key behavioral information recorded.

[0037] By binding key behavioral information with candidate characteristics and job requirements in the recommendation list, data triples are formed to construct a negative sample dataset.

[0038] Based on the deviation between negative sample features and model prediction results, the degree of decision error of the current strategy is measured, and it is analyzed which strategy adjustments can reduce similar errors in the future, and the network weights of the strategy are adjusted accordingly.

[0039] By adopting the above technical solution, the employer's rejection behavior is transformed into model parameter adjustment, which can quickly correct the matching strategy bias. A high-quality negative sample library is built through implicit feedback (such as quick skipping) and explicit labeling to enhance the model's noise resistance.

[0040] In a preferred embodiment, this application further includes the following step before the step of performing semantic extraction and feature fusion on the multimodal data using a multimodal fusion encoder to generate a comprehensive feature vector containing semantic association features and behavioral features:

[0041] Multimodal data is split into independent modules according to its source, and sensitive information is scanned for in each module. The data is then classified according to the degree of sensitivity of the information.

[0042] Sensitive information is processed using differential privacy techniques, and homomorphic encryption is used in the feature fusion stage to ensure that the data is computed in an encrypted state.

[0043] By adopting the above technical solutions, differential privacy destroys the traceability of the original data, homomorphic encryption ensures that the entire computation process is encrypted, supports cross-institutional data fusion without leaking plaintext information, breaks down data silos, and the layered encryption strategy preserves the availability of key features while protecting privacy, avoiding excessive desensitization that could lead to analysis failure.

[0044] The second objective of this invention is achieved through the following technical solution:

[0045] A talent matching and intelligent recruitment system based on an AI big data model includes:

[0046] The data collection module is used to acquire multimodal data of candidates and job requirements data of recruiters. The multimodal data includes at least two of the following: text resumes, voice interview records, and video interview recordings.

[0047] The data processing module is used to perform semantic extraction and feature fusion on the multimodal data through a multimodal fusion encoder to generate a comprehensive feature vector containing semantic association features and behavioral features;

[0048] The skill tag extraction module is used to perform domain adaptation processing on the comprehensive feature vector based on a dynamically updated industry knowledge graph to generate a set of skill tags for candidates.

[0049] The intelligent recommendation module is used to use a reinforcement learning model to dynamically calculate the matching degree between the skill tag set and the job requirement data based on historical recruitment feedback data, and generate a recommendation list.

[0050] The feedback module is used to output the recommendation list to the recruiter's terminal and optimize the reinforcement learning model online based on the recruiter's operational behavior data.

[0051] The above-mentioned objective three of this application is achieved through the following technical solution:

[0052] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the aforementioned AI-based large-scale model-based talent matching and intelligent recruitment method.

[0053] The fourth objective of this application is achieved through the following technical solution:

[0054] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned AI-based large-scale model-based talent matching and intelligent recruitment method.

[0055] In summary, this application includes at least one of the following beneficial technical effects:

[0056] 1. By integrating multimodal data from text, voice, and video, it breaks through the limitations of traditional single resume screening, comprehensively capturing candidates' skills, emotional expressions, and behavioral characteristics. It also overcomes the limitations of traditional text matching, optimizing matching strategies in real time based on reinforcement learning models. This adapts to the rapid iteration of industry skills and the dynamic changes in job requirements. Real-time data updates solve the timeliness problem of traditional static graphs, improve the industry adaptability of skill tags, and enhance the efficiency and accuracy of talent resource matching. Recruiter feedback is directly injected into model iteration, forming a complete closed loop of "data collection-matching-feedback-optimization" to continuously improve recommendation accuracy.

[0057] 2. The incremental update mechanism enables the knowledge graph to keep up with industry trends, avoiding the lag of static graphs. It infers the undisclosed cross-disciplinary potential of candidates through skill association networks, broadens the dimensions of talent evaluation, and automatically adjusts the tag weights according to market demand fluctuations, thereby improving the matching efficiency of in-demand positions.

[0058] 3. The reward function integrates the immediate recruitment success rate with long-term employee performance to avoid the trap of "high matching but low retention after onboarding". By mapping the feature combination state to the action space, it provides an understandable matching logic and balances historical experience and new strategy exploration in model training to prevent innovation stagnation caused by over-reliance on past successful paths.

[0059] 4. Transform the employer's rejection behavior into model parameter adjustments to quickly correct matching strategy biases. Enhance the model's robustness against noise by building a high-quality negative sample library through implicit feedback (such as quick skipping) and explicit labeling. Attached Figure Description

[0060] Figure 1 This is a flowchart of an embodiment of a talent matching and intelligent recruitment method and system based on an AI large model, as described in this application.

[0061] Figure 2This is a flowchart of step S30 in an embodiment of a talent matching and intelligent recruitment method and system based on an AI large model according to this application;

[0062] Figure 3 This is a flowchart of step S40 in an embodiment of a talent matching and intelligent recruitment method and system based on an AI large model according to this application;

[0063] Figure 4 This is a schematic diagram of a talent matching and intelligent recruitment system based on an AI large model, as described in this application.

[0064] Figure 5 This is a schematic block diagram of a computer device according to this application. Detailed Implementation

[0065] The present application will be further described in detail below with reference to the accompanying drawings.

[0066] In the example, such as Figure 1-3 As shown, this application discloses a talent matching and intelligent recruitment method based on an AI large-scale model, which specifically includes the following steps:

[0067] S10: Obtain multimodal data of candidates and job requirements data of recruiters, wherein the multimodal data includes at least two of the following: text resumes, voice interview records, and video interview recordings;

[0068] S20: The multimodal data is semantically extracted and feature fused using a multimodal fusion encoder to generate a comprehensive feature vector containing semantic association features and behavioral features;

[0069] S30: Based on the dynamically updated industry knowledge graph, perform domain adaptation processing on the comprehensive feature vector to generate a set of candidate skill tags;

[0070] S40: Using a reinforcement learning model, dynamically calculate the matching degree between the skill tag set and the job requirement data based on historical recruitment feedback data, and generate a recommendation list;

[0071] S50: Output the recommendation list to the recruiter's terminal, and optimize the reinforcement learning model online based on the recruiter's operational behavior data.

[0072] In this embodiment, natural language processing technology is used to perform structured parsing of text resumes, and computer vision and speech recognition technologies are used to extract candidates' facial expression features, body language features, and emotional features of speech from video interview recordings and voice interview records, respectively. An attention mechanism is used to dynamically allocate feature weights for data from different modalities, and a cross-modal contrastive learning algorithm is used to eliminate semantic biases between modalities, generating a unified comprehensive feature vector. Data from industry recruitment platforms, technical forums, and publicly available enterprise data are crawled in real time, and graph neural networks are used to mine the correlations between skills. The node and edge weights in the knowledge graph are updated through incremental learning. The operation of the reinforcement learning model further includes: defining the matching degree calculation as a Markov decision process, where the state space is the feature combination of the candidate and the position, the action space is the adjustment direction of the matching strategy, and the reward function is dynamically generated based on the recruitment success rate and the candidate's performance data after joining the company; a negative sample dataset is constructed based on the recruiter's click, filter, and feedback behavior on the recommendation list, and the policy network parameters of the reinforcement learning model are adjusted in real time using an online gradient descent algorithm.

[0073] In one embodiment, step S10 specifically includes the following steps:

[0074] S11: Convert resumes of different formats into text files of a unified format, use a natural language processing model to identify the core modules in the text, extract standardized entity information from each module, and generate text resume features;

[0075] S12: Segment the interview video into consecutive frame images along the timeline, track the candidate's facial expression changes using facial recognition technology, and generate video facial expression features;

[0076] S13: Quantify the emotional tendency based on the spectral energy distribution of the sound in the interview audio, combine the speech content with the emotional features, and generate speech emotional features.

[0077] In this embodiment, the core modules in the text include “educational background”, “work experience”, and “skills”; standardized entity information includes educational background: extracting school name, major, degree, and time range; work experience: parsing company name, position, tenure, and job description; and skill tags: mapping free text (such as “familiar with machine learning algorithms”) to an industry standard skill library (such as “machine learning: advanced”).

[0078] Specifically, resumes in different formats (such as PDF and Word documents) are converted into text files of a unified format to eliminate the interference of layout differences on parsing. For example, style information of images and tables is removed, and plain text content is retained. Natural language processing models are used to identify the core modules in the text, and semantic analysis is used to distinguish descriptive content (such as "led a team to complete a project") from factual data (such as "Python programming: 5 years of experience"), extracting standardized entity information from each module to generate text resume features. Interview videos are segmented into continuous frame images along the timeline, and key segments (such as segments of candidates answering questions and interaction scenes with interviewers) are identified. Facial recognition technology is used to track changes in candidates' facial expressions, including basic emotion recognition: detecting common emotion labels (such as confidence, nervousness, focus); micro-expression capture: analyzing brief facial muscle movements (such as momentary frowning, twitching of the corners of the mouth) to assess stress coping ability. A spatiotemporal model of candidate posture is constructed, including motion capture: recognizing gestures (such as clasped hands, frequent nodding), sitting posture (leaning forward / backward); dynamic trajectory analysis: statistically analyzing the frequency and amplitude of limb movements (such as excessive gesture amplitude may reflect emotional fluctuations). The interview audio is noise-reduced to separate the candidate's voice from environmental noise, ensuring clarity. Pitch variations (e.g., a lowering of the final syllable of a statement) and speech rate fluctuations (e.g., increased speech rate when nervous) are analyzed. Emotional tendency is quantified based on the spectral energy distribution of the voice (e.g., increased high-frequency energy may reflect excitement). The analysis combines voice content with emotional features, including inconsistency detection: identifying inconsistencies between the emotional tone of the voice and the content of the answer (e.g., using a flat tone to describe "enthusiastic work experience"). Stress response assessment: analyzing the degree of voice tremor and recovery speed when answering unexpected questions.

[0079] In one embodiment, the step S20 is further included before step S20:

[0080] S14: Divide multimodal data into independent modules according to their source, scan each module for sensitive information, and classify them according to the degree of sensitivity of the information;

[0081] S15: Perform differential privacy processing on sensitive information and use homomorphic encryption technology in the feature fusion stage to ensure that the data is computed in an encrypted state.

[0082] In this embodiment, sensitive information scanning includes: text data: identifying plaintext sensitive fields such as name, ID number, contact information, and address; video data: detecting faces and identifiable backgrounds (such as company logos and home environments); and voice data: marking voiceprint features and privacy-related dialogue content (such as salary discussions). Differential privacy processing includes: data perturbation noise injection: partially masking highly sensitive fields (such as mobile phone numbers) and adding random noise to ensure that the original data cannot be restored; dynamically blurring or replacing facial areas with virtual images; preserving body language but eliminating identity features; adjusting tone and speech rate to destroy the uniqueness of voiceprints while preserving semantic content; statistical protection: adding Laplace noise to query results during aggregation analysis (such as statistically analyzing the proportion of candidates for a certain skill) to ensure that individual contributions cannot be reverse-inferred; and privacy budget management: setting a global privacy budget (such as ε = 0.1) and allocating it to each modal data processing step to prevent cumulative privacy leakage caused by multiple queries.

[0083] Specifically, multimodal data (such as text resumes, video interviews, and voice recordings) is split into independent modules according to its source, and each module is scanned for sensitive information. The data is then categorized based on its sensitivity level (e.g., high sensitivity: ID number; medium sensitivity: work experience; low sensitivity: skill tags) to provide a basis for subsequent differentiated processing. Differential privacy processing is applied to sensitive information, and encryption public keys and decryption private keys are generated for the recruiter, candidate, and system respectively. This ensures that the data provider (candidate) cannot access the fusion results, and the recruiter can only decrypt the final matching information. The processed multimodal features (such as text skill tags, video emotion scores, and voice fluency indicators) are encrypted into ciphertext to ensure that single-modal data cannot be independently decrypted. During the feature fusion stage, it supports weighted summation of encrypted feature values ​​(e.g., skill matching degree = encrypted "programming ability" × weight 1 + encrypted "communication ability" × weight 2), and allows conditional judgment (e.g., encrypted "years of experience" > 3) and logical operations for dynamic candidate screening. If the data is distributed across different institutions (e.g., recruitment platforms and enterprise databases), cross-domain encrypted fusion is achieved through security protocols to avoid plaintext transmission, and feature splicing, normalization and other operations are performed directly on the encrypted data to output a comprehensive feature vector in the encrypted state.

[0084] In one embodiment, step S20 specifically includes the following steps:

[0085] S21: Map text resume features, video expression features, and voice emotion features to a unified semantic space, automatically adjust the weight ratio of each modality according to the job requirement type being processed, and generate a preliminary fused feature vector through weighted feature fusion;

[0086] S22: Construct cross-modal positive and negative sample pairs, and through comparative learning, shorten the cross-modal feature distance of positive sample pairs while widening the feature distance of negative sample pairs;

[0087] S23: Map the initial fused feature vector to a low-dimensional space, remove redundant information through a clustering algorithm, and generate a compact comprehensive feature vector.

[0088] In this embodiment, positive sample pairs are: extracting semantically consistent features from the multimodal data of the same candidate (such as "strong communication skills" in the resume and natural gestures and fluent speech in the video); negative sample pairs are: randomly combining modal features of different candidates (such as pairing the text description of "introverted personality" of candidate A with a video clip of candidate B's heated debate).

[0089] Specifically, different modalities of data are mapped to a unified semantic space to ensure alignment of feature dimensions across modalities. Initial weights are assigned to each modality (e.g., text has higher weights, video / audio has lower weights). Based on the job requirements being processed (e.g., technical positions emphasize textual skills, while management positions emphasize leadership performance in video), the weight ratios of each modality are automatically adjusted. Low-quality data (e.g., blurry videos, noisy audio) has its weights automatically reduced to avoid interfering with the overall analysis. The adjusted weights are then multiplied element-wise with each modal feature to generate a preliminary fusion feature vector. The core information of high-weight modalities (e.g., programming ability tags for technical positions) is retained, while secondary information of low-weight modalities (e.g., excessive body language for sales positions) is weakened. Constructing cross-modal positive and negative sample pairs, the system uses contrastive learning to narrow the cross-modal feature distance of positive sample pairs (e.g., the text "team collaboration" versus multi-person interaction scenes in videos), while widening the feature distance of negative sample pairs (e.g., the text "working independently" versus compliant body language with frequent nodding). It detects and corrects contradictory expressions between modalities; for example, if a resume states "proficient in Python" but the candidate cannot answer basic syntax questions in a video interview, the credibility of the text skill label is reduced; if the voice sentiment analysis shows "confidence" but the candidate frequently lowers their head and avoids eye contact in the video, the sentiment feature weights are recalibrated. The fused multimodal features are mapped to a low-dimensional space, and redundant information (e.g., the repetitive expression of "leadership") is removed using clustering algorithms, generating a compact comprehensive feature vector. A visual interface is provided to display the weight allocation logic (e.g., "In technical job matching, text skills account for 70% of the weight, and video coding ability accounts for 20% of the weight"), allowing recruiters to manually adjust the weights and use the attention mechanism and contrastive learning strategy to correct the results based on actual recruitment outcomes (e.g., post-employment performance data).

[0090] In one embodiment, step S30 specifically includes the following steps:

[0091] S31: Real-time acquisition of publicly available data from industry recruitment platforms, technology forums, and companies, combined with graph neural networks to uncover the relationships between skills;

[0092] S32: Update the node and edge weights in the knowledge graph through incremental learning, and verify the accuracy of the updated graph prediction through backtesting with historical data;

[0093] S33: Align the candidate's comprehensive feature vector with the skill nodes in the knowledge graph, extract explicit labels by directly matching the candidate's skills with the graph nodes, deduce potential abilities based on the association relationship, and generate implicit labels;

[0094] S34: Adjust tag priority based on real-time industry needs and generate a set of skill tags with confidence levels.

[0095] In this embodiment, targeted data crawling includes: recruitment platforms: real-time crawling of structured data such as job descriptions, skill requirements, and salary ranges; technical forums: collection of unstructured text such as technical discussion hotspots and feedback on the use of emerging tools in developer communities; and publicly available corporate data: extraction of job updates, internal skills training plans, and industry white papers published on corporate websites.

[0096] Specifically, it acquires real-time data from industry recruitment platforms, technology forums, and publicly available corporate data, removes advertisements, duplicate content, and irrelevant noise data (such as fake job postings on recruitment platforms), extracts key entities from unstructured text (such as forum discussions) (such as "AIGC" and "low-code platform"), and prioritizes data based on the credibility of the data source (such as authoritative corporate websites vs. anonymous forums) and the freshness of the data (such as information updated within 24 hours). Skill terms (such as "Python" and "project management") are abstracted into graph nodes, and initial edge connections are established based on co-occurrence relationships (such as "Python" often appearing alongside "data analysis"). Graph neural networks are used to mine the relationships between skills, including local association expansion: analyzing the context of skill nodes (such as the competitive substitution relationship between "TensorFlow" and "PyTorch" in the "machine learning engineer" job title); global semantic reasoning: identifying cross-domain skill associations (such as the cross-domain application dependence between "blockchain development" and "financial risk control"); and adjusting the connection strength between nodes based on changes in skill popularity (such as the surge in searches for "ChatGPT") and trend predictions in industry reports (such as strengthening the association between "AI drawing" and "design tools"). When new skill terms (such as "quantum computing engineer") or correlation patterns deviating from historical patterns (such as the emerging combination of "Java" and "embedded development") are detected, a knowledge graph update is initiated: newly added skill nodes only update local subgraphs to avoid the resource consumption of full graph reconstruction; historical correlation weights are preserved (such as the long-term stable relationship of "Python-data analysis"), and only edges that have changed significantly (such as the increased demand for "Rust-system development") are recalibrated; the accuracy of the updated graph predictions is verified through backtesting with historical data (such as verifying the correlation between "AIGC skills" and the growth in job demand). Align the candidate's comprehensive feature vector (such as skill mastery and project experience) with skill nodes in the knowledge graph, filter relevant subgraphs (such as focusing on the "Internet industry" skill tree), extract explicit labels by directly matching candidate skills with graph nodes (such as "proficient in Python"), infer potential abilities based on the relationship (such as "familiar with TensorFlow" can infer "possessing a foundation in deep learning") and generate implicit labels, adjust the label priority according to real-time industry demand (such as a 30% month-on-month increase in recruitment volume for a certain skill), and generate a set of skill labels with confidence (such as "front-end development (core) | UI design (auxiliary)").

[0097] In one embodiment, step S40 specifically includes the following steps:

[0098] S41: Transform the candidate's skill tag set and job requirement data into a unified coding format, define executable matching strategy adjustment actions, and dynamically generate reward functions based on recruitment success rate and candidate performance data after onboarding.

[0099] S42: Extract historical recruitment feedback data and randomly try multiple matching strategies. Through repeated simulations of historical trajectories, update the strategy network parameters so that the reinforcement learning model tends to choose actions that maximize cumulative rewards.

[0100] S43: Input the candidate's skill tag set and job requirement data, generate an initial feature combination state, select the optimal action according to the model strategy, and calculate the comprehensive matching score between the candidate and the job based on the adjusted strategy.

[0101] S44: Generate a recommendation list in descending order of matching score.

[0102] In this embodiment, the matching strategy adjustment actions include increasing the weight of specific skills (such as increasing the proportion of "deep learning framework" in the matching of technical positions), relaxing secondary conditions (such as lowering the requirement for "internship experience" to expand the candidate pool), and introducing cross-domain skill associations (such as linking "data analysis" with "business insights" for evaluation).

[0103] Specifically, the candidate's skill tag set and job requirement data are transformed into a unified encoding format to form a multi-dimensional feature vector. For example, candidate features include: skill mastery, project experience duration, and certificate level; job features include: a list of required skills, experience thresholds, and team culture preferences. Executable matching strategy adjustment actions are defined, and immediate rewards are calculated based on recruitment success rates, such as the proportion of candidates accepting offers and the conversion rate from interview to hiring. Delayed rewards are generated based on post-employment performance data (such as quarterly assessment results and project contributions), dynamically balancing short-term efficiency and long-term hiring quality. Complete matching path data is extracted from past recruitment records, including: initial state (candidate-job feature combination); execution action (matching strategy adjustment); result state (recruitment success); and final reward (performance data converted into numerical scores). In the early stages of the model, various matching strategies (such as different skill combination weights) are randomly tried to accumulate diverse decision samples. As data accumulates, the model gradually favors historical strategies with high rewards (such as successful cases where "algorithm ability" weight is emphasized in a certain type of job). Through repeated simulations of historical trajectories, the strategy network parameters are updated, causing the reinforcement learning model to tend to select actions that maximize cumulative rewards (such as prioritizing candidates with high performance potential). Input the current candidate's skill tags and job requirements to generate an initial feature combination state. Based on the model strategy, select the optimal action (e.g., adjust the weight of "communication skills" in management positions to 40%). Combined with the adjusted strategy, calculate the candidate's overall match score with the job, including: core skill match (hard requirement), cultural fit (soft indicator), and potential prediction (based on the growth trajectory of similar historical candidates). Generate a Top-N recommendation list in descending order of match score, with interpretable explanations (e.g., "Recommendation reason: Project experience and team technology stack are highly compatible"), and support multi-dimensional filtering (e.g., generating differentiated lists based on "stability priority" and "innovation potential priority").

[0104] In one embodiment, step S50 specifically includes the following steps:

[0105] S51: Output the recommended list to the recruiter's terminal, track the recruiter's operations throughout the entire process, and record key behavioral information;

[0106] S52: Bind key behavioral information with candidate characteristics and job requirements in the recommendation list to form data triples and construct a negative sample dataset;

[0107] S53: Based on the deviation between negative sample features and model prediction results, measure the degree of decision error of the current strategy, analyze which strategy adjustments can reduce similar errors in the future, and adjust the network weights of the strategy.

[0108] In this embodiment, the recruiter's terminal supports the following interactive functions: Details Click: Recruiters can click on a candidate's card to view a complete resume, interview video, and other detailed information; Conditional Filtering: Filter recommended results by skill tags, years of experience, etc.; Proactive Feedback: Mark candidates as "suitable," "pending," or "unsuitable," and fill in the reasons (e.g., "skills mismatch," "cultural incompatibility"). Key behavioral information includes: Implicit negative feedback: Candidates are quickly skipped (e.g., dwell time is below a threshold), not clicked, or excluded by filtering conditions; Explicit negative feedback: Candidates are directly marked as "unsuitable" or a negative evaluation is provided; Positive signals: Initiating an interview invitation or adding a candidate to the talent pool. Data triplets include: Input features: Candidate skill tags, job requirements, and matching score; Operational behavior: Clicking, filtering, and marking results; Contextual information: Operation time and recruiter account attributes (e.g., company size, industry). The negative sample dataset includes: strong negative samples: recommendations explicitly marked as "inappropriate"; weak negative samples: candidates who were not clicked and whose matching degree was higher than the threshold (presumably implicit mismatch); conflicting samples: the recruiter's screening conditions contradict the model's recommendation logic (e.g., requiring "bachelor's degree or above" but recommending candidates with associate degrees).

[0109] Specifically, the recommendation list generated by the reinforcement learning model is sorted by matching degree and displayed in pages on the recruiter's terminal interface. The recruiter's operations are tracked throughout the process, and key behavioral information is recorded. The recruiter's behavior is bound to the candidate characteristics and job requirements in the recommendation list to form data triples, constructing a negative sample dataset. The negative sample data is injected into the model training pipeline in batches according to time windows (e.g., 5 minutes) to avoid system load fluctuations caused by high-frequency updates. Based on the deviation between the negative sample features and the model prediction results, the degree of decision error of the current strategy is measured (e.g., high-match candidates are frequently rejected). It is analyzed which strategy adjustments (e.g., skill weight allocation, cultural adaptation threshold) can reduce similar errors in the future. The network weights are slightly adjusted in the opposite direction of the gradient (e.g., reducing the weight of a skill tag that is rejected a lot) to limit the magnitude of a single update and prevent model oscillation caused by noisy data (e.g., erroneous negative labels). The updated model is run in parallel with the old version, and the following indicators are compared: recommendation acceptance rate: the proportion of the recruiter's adoption of the recommendation results; negative feedback reduction rate: the reduction of "unsuitable" labels for similar positions; long-term retention association: the performance change of newly recommended candidates after joining the company. If key metrics (such as matching accuracy) drop significantly after the update, the system will automatically switch to a stable historical version and trigger manual checks (such as checking the quality of negative sample data).

[0110] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0111] In one embodiment, a talent matching and intelligent recruitment system based on an AI large-scale model is provided. This device corresponds one-to-one with the talent matching and intelligent recruitment method based on an AI large-scale model described in the above embodiments. Figure 4 As shown, the system includes:

[0112] The data collection module is used to acquire multimodal data of candidates and job requirements data of recruiters. The multimodal data includes at least two of the following: text resumes, voice interview records, and video interview recordings.

[0113] The data processing module is used to perform semantic extraction and feature fusion on the multimodal data through a multimodal fusion encoder to generate a comprehensive feature vector containing semantic association features and behavioral features;

[0114] The skill tag extraction module is used to perform domain adaptation processing on the comprehensive feature vector based on a dynamically updated industry knowledge graph to generate a set of skill tags for candidates.

[0115] The intelligent recommendation module is used to use a reinforcement learning model to dynamically calculate the matching degree between the skill tag set and the job requirement data based on historical recruitment feedback data, and generate a recommendation list.

[0116] The feedback module is used to output the recommendation list to the recruiter's terminal and optimize the reinforcement learning model online based on the recruiter's operational behavior data.

[0117] For specific limitations regarding the AI-based large-scale model talent matching and intelligent recruitment system, please refer to the limitations of the AI-based large-scale model talent matching and intelligent recruitment method mentioned above, which will not be repeated here. Each module in the aforementioned AI-based large-scale model talent matching and intelligent recruitment system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0118] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a talent matching and intelligent recruitment method based on a large AI model.

[0119] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a talent matching and intelligent recruitment method based on an AI big model.

[0120] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements a talent matching and intelligent recruitment method based on an AI large model.

[0121] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0122] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to several functional units or modules as needed, that is, the internal structure of the device can be divided into several functional units or modules to complete all or part of the functions described above.

[0123] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A talent matching and intelligent recruitment method based on an AI large-scale model, characterized in that, Including the following steps: The multimodal data of candidates and the job requirements data of recruiters are obtained. The multimodal data includes at least two of the following: text resumes, voice interview records and video interview recordings. The process involves converting resumes of different formats into text files of a unified format, using a natural language processing model to identify the core modules in the text, extracting standardized entity information from each module, and generating text resume features. The interview video is segmented into consecutive frames along the timeline, and facial recognition technology is used to track the candidate's facial expression changes to generate video facial expression features. The emotional tendency is quantified based on the spectral energy distribution of the sound in the interview audio, and the speech content is combined with the emotional features to generate speech emotional features. Multimodal data is split into independent modules according to its source, and sensitive information is scanned for in each module. The data is then classified according to the degree of sensitivity of the information. Differential privacy processing is applied to sensitive information, and homomorphic encryption technology is used in the feature fusion stage to ensure that the data is computed in an encrypted state. The multimodal data is semantically extracted and feature fused using a multimodal fusion encoder to generate a comprehensive feature vector containing semantic association features and behavioral features. Text resume features, video expression features, and voice emotion features are mapped to a unified semantic space. The weight ratio of each modality is automatically adjusted according to the job requirements being processed. A preliminary fused feature vector is generated through weighted feature fusion. Construct cross-modal positive and negative sample pairs, and through comparative learning, shorten the cross-modal feature distance of positive sample pairs while widening the feature distance of negative sample pairs; The initial fused feature vectors are mapped to a low-dimensional space, and redundant information is removed by clustering algorithms to generate compact comprehensive feature vectors. Based on a dynamically updated industry knowledge graph, the comprehensive feature vector is subjected to domain adaptation processing to generate a set of candidate skill tags. Real-time acquisition of publicly available data from industry recruitment platforms, technology forums, and companies, combined with graph neural networks to uncover the relationships between skills; The node and edge weights in the knowledge graph are updated through incremental learning, and the accuracy of the updated graph prediction is verified by backtesting with historical data. Align the candidate's comprehensive feature vector with the skill nodes in the knowledge graph, extract explicit labels by directly matching the candidate's skills with the graph nodes, and infer potential abilities and generate implicit labels based on the association relationship; Adjust tag priorities based on real-time industry needs to generate a set of skill tags with confidence levels; A reinforcement learning model is used to dynamically calculate the matching degree between the skill tag set and the job requirement data based on historical recruitment feedback data, and a recommendation list is generated. The candidate's skill tag set and job requirement data are converted into a unified coding format, executable matching strategy adjustment actions are defined, and reward functions are dynamically generated based on the recruitment success rate and the candidate's performance data after joining the company. Extract historical recruitment feedback data and randomly try various matching strategies. Through repeated simulations of historical trajectories, update the strategy network parameters so that the reinforcement learning model tends to choose actions that maximize cumulative rewards. Input the candidate's skill tag set and job requirement data to generate an initial feature combination state. Select the optimal action according to the model strategy and calculate the comprehensive matching score between the candidate and the job based on the adjusted strategy. A recommendation list is generated in descending order of matching score; The recommendation list is output to the recruiter's terminal, and the reinforcement learning model is optimized online based on the recruiter's operational behavior data.

2. The talent matching and intelligent recruitment method based on an AI large-scale model according to claim 1, characterized in that, The step of outputting the recommendation list to the recruiter's terminal and optimizing the reinforcement learning model online based on the recruiter's operational behavior data specifically includes the following steps: The recommended list is output to the recruiter's terminal, and the recruiter's operations are tracked throughout the entire process, with key behavioral information recorded. By binding key behavioral information with candidate characteristics and job requirements in the recommendation list, data triples are formed to construct a negative sample dataset. Based on the deviation between negative sample features and model prediction results, the degree of decision error of the current strategy is measured, and it is analyzed which strategy adjustments can reduce similar errors in the future, and the network weights of the strategy are adjusted accordingly.

3. A talent matching and intelligent recruitment system based on an AI large-scale model, characterized in that, include: The data collection module is used to acquire multimodal data of candidates and job requirements data of recruiters. The multimodal data includes at least two of the following: text resumes, voice interview records, and video interview recordings. The process involves converting resumes of different formats into text files of a unified format, using a natural language processing model to identify the core modules in the text, extracting standardized entity information from each module, and generating text resume features. The interview video is segmented into consecutive frames along the timeline, and facial recognition technology is used to track the candidate's facial expression changes to generate video facial expression features. The emotional tendency is quantified based on the spectral energy distribution of the sound in the interview audio, and the speech content is combined with the emotional features to generate speech emotional features. Multimodal data is split into independent modules according to its source, and sensitive information is scanned for in each module. The data is then classified according to the degree of sensitivity of the information. Differential privacy processing is applied to sensitive information, and homomorphic encryption technology is used in the feature fusion stage to ensure that the data is computed in an encrypted state. The data processing module is used to perform semantic extraction and feature fusion on the multimodal data through a multimodal fusion encoder to generate a comprehensive feature vector containing semantic association features and behavioral features; Text resume features, video expression features, and voice emotion features are mapped to a unified semantic space. The weight ratio of each modality is automatically adjusted according to the job requirements being processed. A preliminary fused feature vector is generated through weighted feature fusion. Construct cross-modal positive and negative sample pairs, and through comparative learning, shorten the cross-modal feature distance of positive sample pairs while widening the feature distance of negative sample pairs; The initial fused feature vectors are mapped to a low-dimensional space, and redundant information is removed by clustering algorithms to generate compact comprehensive feature vectors. The skill tag extraction module is used to perform domain adaptation processing on the comprehensive feature vector based on a dynamically updated industry knowledge graph to generate a set of skill tags for candidates. Real-time acquisition of publicly available data from industry recruitment platforms, technology forums, and companies, combined with graph neural networks to uncover the relationships between skills; The node and edge weights in the knowledge graph are updated through incremental learning, and the accuracy of the updated graph prediction is verified by backtesting with historical data. Align the candidate's comprehensive feature vector with the skill nodes in the knowledge graph, extract explicit labels by directly matching the candidate's skills with the graph nodes, and infer potential abilities and generate implicit labels based on the association relationship; Adjust tag priorities based on real-time industry needs to generate a set of skill tags with confidence levels; The intelligent recommendation module is used to use a reinforcement learning model to dynamically calculate the matching degree between the skill tag set and the job requirement data based on historical recruitment feedback data, and generate a recommendation list. The candidate's skill tag set and job requirement data are converted into a unified coding format, executable matching strategy adjustment actions are defined, and reward functions are dynamically generated based on the recruitment success rate and the candidate's performance data after joining the company. Extract historical recruitment feedback data and randomly try various matching strategies. Through repeated simulations of historical trajectories, update the strategy network parameters so that the reinforcement learning model tends to choose actions that maximize cumulative rewards. Input the candidate's skill tag set and job requirement data to generate an initial feature combination state. Select the optimal action according to the model strategy and calculate the comprehensive matching score between the candidate and the job based on the adjusted strategy. A recommendation list is generated in descending order of matching score; The feedback module is used to output the recommendation list to the recruiter's terminal and optimize the reinforcement learning model online based on the recruiter's operational behavior data.

4. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the AI-based large model-based talent matching and intelligent recruitment method as described in any one of claims 1 to 2.

5. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the AI-based large model-based talent matching and intelligent recruitment method as described in any one of claims 1 to 2.

Images