Human resource information processing method, device, equipment and medium

Abstract

This invention discloses a method, apparatus, equipment, and medium for processing human resource information, belonging to the field of information processing technology. It includes: collecting raw human resource data and digitizing unstructured documents using optical character recognition technology based on deep convolutional neural networks; preprocessing and extracting features from the data to construct a multi-dimensional feature vector space; building a talent-job matching model based on deep learning networks and multi-head attention mechanisms to calculate the correlation score between candidate characteristics and job requirements; introducing internal enterprise evaluation standards and job weight factors to weight and correct the matching score; combining long short-term memory neural networks to predict the candidate's career development potential and stability, and outputting a recommended ranking list; and continuously iterating and optimizing the model by tracking post-employment performance data and feeding it back to the matching model. This significantly improves the efficiency and accuracy of human resource information processing and enhances the system's self-learning ability and data security.

Description

Technical Field

[0001] This invention belongs to the field of information processing technology, and specifically relates to a method, apparatus, equipment and medium for processing human resources information. Background Technology

[0002] In modern corporate governance systems, human resource management, as a key link in optimizing organizational structure and enhancing core competitiveness, encompasses multiple dimensions such as talent recruitment, job allocation, performance evaluation, and career development planning. With the accelerating pace of global economic digital transformation, leveraging advanced information technology to efficiently manage and deeply mine massive amounts of heterogeneous human resource data has become an essential means to ensure the development of corporate talent pipelines and the achievement of strategic goals.

[0003] Human resource information processing technology focuses on the structured analysis of multi-dimensional information such as candidate qualifications, professional skills, behavioral characteristics, and job requirements, aiming to build a precise mapping relationship between the talent supply side and the job demand side. Through scientific evaluation models and intelligent matching algorithms, this technology strives to achieve the optimal solution for person-job matching in complex organizational environments, thereby improving the operational efficiency of the entire talent management system.

[0004] Existing human resource information processing technologies still face numerous serious challenges in practical applications: First, the information processing model heavily relies on human experience and basic data statistics, resulting in overly limited evaluation dimensions and a high susceptibility to subjective bias, making it difficult to objectively and comprehensively measure the overall quality of talent; second, the lack of deep semantic understanding of unstructured resume data and dynamic business needs leads to a significant lag and rigidity in job matching logic, making it unable to cope with the precise staffing needs brought about by rapid business iterations; third, insufficient real-time data processing and system collaboration capabilities make it difficult to quickly capture key features from massive amounts of information, greatly increasing screening costs and reducing decision-making efficiency; finally, due to the lack of an effective feedback loop mechanism, the system struggles to continuously optimize the matching model based on actual performance, easily leading to serious mismatches or hidden losses of human resources.

[0005] In view of the shortcomings of traditional technologies, such as single evaluation dimensions, large subjective bias, insufficient semantic understanding and lack of feedback mechanism, the human resource information processing method and device proposed in this invention are particularly important. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a human resources information processing method, apparatus, equipment and medium that addresses the shortcomings of the prior art and can effectively solve the problems in the prior art.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A human resources information processing method includes the following steps: Step 1: Collect raw human resources data. Obtain candidates' resumes, job descriptions, and historical recruitment data from companies through multi-source data interfaces. Use optical character recognition technology based on deep convolutional neural networks to extract and convert the digital features of unstructured paper documents to obtain a structured raw human resources dataset. Step 2, Preprocessing and Feature Extraction: The original human resources dataset is subjected to data deduplication, missing value imputation and semantic standardization. Natural language processing technology is used to extract candidates' professional skills, work experience, education level and behavioral feature labels. At the same time, hard indicators and soft quality requirements in job requirements are identified to construct a multi-dimensional feature vector space. Step 3: Construct a talent-job matching model. Based on the multi-dimensional feature vector space, a semantic association model between talent and job is constructed using a deep learning network. A multi-head attention mechanism is used to perform deep semantic fusion of candidate features and job requirements, calculating the correlation score between the candidate feature vector and the job requirement vector to generate a preliminary matching score. Specifically, the deep semantic fusion of candidate features and job requirements using the multi-head attention mechanism includes: the deep learning network employs a dual-tower structure based on a Transformer encoder to independently encode the candidate features and job requirements, obtaining a candidate feature vector sequence and a job feature vector sequence. The calculation formula for multi-head attention mechanism is as follows: ; in, , , They are generated by linear transformations of candidate characteristics and job characteristics, respectively. For the attention head dimension, eight parallel attention heads are set up, each capturing the correlation information of eight dimensions: education matching degree, skill matching degree, experience matching degree, project complexity matching degree, technology stack overlap degree, soft quality matching degree, cultural adaptability matching degree, and career development matching degree. Then, the outputs of the eight attention heads are dynamically weighted and fused using a learnable gated fusion network to obtain the final fused feature vector: ; Among them, gating weights , For the first The output vector of each attention head. Represents the candidate feature vector Job Feature Vector splicing, Here is the weight matrix of the gated network. For bias vectors, For the first The learnable temperature coefficient of an individual attention head; Calculate the cosine similarity between the candidate feature vector and the job requirement vector. ; in Representative candidate characteristic components, Represents the weight of job requirements. These are the dimension weights determined through feature importance assessment; Step four: Multi-dimensional evaluation and ranking. Introduce the company's internal talent evaluation standards and job weight factors to weight and correct the preliminary matching scores. Combine long short-term memory neural networks to analyze the candidates' career trajectory sequences, predict the candidates' career development potential and stability, and output a ranked list of recommended candidates. The job weighting factor is dynamically adjusted based on the urgency and strategic importance of the business department, and the adjustment formula is as follows: ; For the position The basic weighting factor, This is an adjustment factor based on business urgency. For the position Business urgency rating Rate the highest urgency level among all open positions. For the position Strategic importance score Rate the highest strategic importance among all open positions. This is a coefficient for adjusting for talent scarcity. For the position The market scarcity index of the required core skills. The maximum scarcity index is preset; The long short-term memory neural network employs a loss function with added stability regularization during training: ; in , The total number of training samples, For the first The candidate's true credentials The probability of leaving the company is predicted by LSTM. The regularization coefficient is . Let L2 be the norm of the final hidden state vector of the LSTM. For indicator functions; Step 5: Feedback loop and model iteration. Real-time tracking of candidate performance data after onboarding. The performance data is normalized and then fed back to the talent-job matching model as a monitoring signal. The model parameters are adjusted through the backpropagation algorithm to achieve continuous optimization of the talent-job matching model.

[0008] As a further preferred embodiment of the human resource information processing method of the present invention, in step one, This paper utilizes optical character recognition technology based on deep convolutional recurrent neural networks (CRNNs) to extract and transform digital features from unstructured paper documents. The CRNN consists of a cascaded convolutional neural network (CNN) and a recurrent neural network (RNN). The CNN part uses a ResNet-34 residual network structure to extract feature maps from the document image, while the RNN part uses a bidirectional long short-term memory (BiLSTM) network to perform temporal modeling of the feature sequence. Character sequence alignment and decoding are achieved through a connection-based temporal classification (CTC) loss function. During recognition, the residual network structure analyzes the physical structure of the document, identifying paragraph boundaries, table row and column relationships, and headshot regions in the image. Table row and column relationships are identified using a TableNet table parsing model based on graph neural networks, and headshot regions are located and cropped using a lightweight object detection network, YOLOv8. The process of obtaining information through multi-source data interfaces includes: Call the application programming interface of a third-party recruitment platform to synchronously update the candidate application status at a preset frequency; Using web crawler technology to retrieve publicly available information from industry talent pools within authorized scope; Obtain existing talent data through the database connection protocol of the enterprise's internal office system; The optical character recognition technology uses a residual network structure with a predetermined number of layers to extract features from resume images, recognize and convert various document formats, including portable document formats, image formats and documents generated by word processing software; During the recognition process, the residual network structure parses the physical structure of the document and identifies paragraph boundaries, table row and column relationships, and headshot regions in images.

[0009] As a further preferred embodiment of the human resources information processing method of the present invention, in step two, the professional skills, work experience, education level, and behavioral feature tags of the candidates are extracted using natural language processing technology. Specifically, this includes: segmenting the original text using a word segmentation algorithm based on a Hidden Markov Model or Conditional Random Field (CRF) combined with a custom human resources professional dictionary; mapping discrete text words into 768-dimensional continuous real-valued vectors using a BERT pre-trained language model, wherein the BERT model has undergone domain-adaptive fine-tuning on a corpus of millions of human resources words; extracting behavioral feature tags using a sentiment analysis algorithm, which is based on the TextCNN convolutional neural network structure and extracts sentiment features in the text through multi-scale convolutional kernels, outputting personality tendency scores of the candidates in five dimensions: openness, conscientiousness, extraversion, agreeableness, and neuroticism; and extracting quantitative scores in teamwork, leadership, stress resistance, and innovative thinking dimensions by text mining of the candidates' past project experiences using a hierarchical attention network (HAN), wherein the hierarchical attention network performs attention weighting at the sentence level and document level respectively. The process of semantic standardization and feature extraction includes: Establish a human resources professional dictionary containing a predetermined number of entries; use word segmentation algorithms to segment the original text and remove stop words and meaningless characters; By using word embedding technology, discrete text words are mapped to continuous real vectors of a preset dimension, ensuring that the distance between semantically similar words in the vector space is within a preset range. Extracting behavioral characteristic tags from candidates includes: analyzing candidates' public activities on social media or professional forums, and using sentiment analysis algorithms to identify their career values ​​and personality tendencies; By analyzing the text of candidates’ past project experiences, we extracted quantitative scores for their teamwork, leadership, stress resistance, and innovative thinking. The hard requirements extracted from job descriptions include: automatically identifying educational requirements, major requirements, years of work experience requirements, and specific qualification certificate requirements; Extracting soft skills requirements includes identifying job descriptions for communication skills, logical thinking skills, and organizational and coordination skills.

[0010] As a further preferred embodiment of the human resources information processing method of the present invention, the process of constructing a talent-job matching model in step three includes: The deep learning network employs a multi-head attention mechanism, capturing matching features of different dimensions through a predetermined number of parallel attention heads; The calculation process of the correlation score is as follows: when calculating the similarity between the candidate feature vector and the job requirement vector, different initial weights are set for different feature dimensions. The feature dimensions include professional skills dimension, work experience dimension, education level dimension and other dimensions, and each dimension is set with a corresponding preset weight ratio. The intelligent matching process uses a high-performance graphics processor to perform large-scale matrix operations, supports model files generated by various deep learning frameworks, and generates a correlation score matrix between candidates and positions. The different attention heads in the multi-head attention mechanism focus on capturing the correlation information of educational background matching degree, project complexity, and technology stack overlap.

[0011] As a further preferred embodiment of the human resources information processing method of the present invention, in step four, the candidate's career trajectory sequence is analyzed in conjunction with a long short-term memory neural network. Specifically, the input layer of the long short-term memory neural network contains the candidate's career trajectory sequence sorted by time over the past 10 years. Each trajectory record includes features in six dimensions: tenure, job level, salary increase, company size, industry category, and number of promotions. The sequence length is uniformly padded or truncated to 20 time steps. The LSTM network adopts a three-layer stacked structure with 128 hidden units in each layer. Temporal features are extracted through the synergistic effect of the forget gate, input gate, and output gate. An attention mechanism is added to the top layer to weight key career turning points. The turnover risk prediction adopts a binary classification output layer with cross-entropy loss as the loss function. The comprehensive evaluation process also includes the introduction of an employee turnover early warning mechanism. By mining the desensitized data of departing employees, the Isolation Forest algorithm is used to identify abnormal features such as fluctuations in leave frequency, increased social network activity, and decreased emotional value of communication language. The detected abnormal features are converted into negative feature vectors and input into the long short-term memory neural network.

[0012] The process of multi-dimensional evaluation and ranking includes: The company's internal talent evaluation criteria include the alignment of the company's core values, organizational culture adaptability score, and talent structure balance factor within departments. The job weighting factors are dynamically adjusted based on the urgency and strategic importance of the business departments, with the adjustment range within a preset proportion. The input layer of the long short-term memory neural network contains the candidate's career trajectory sequence within a predetermined period. Through the synergistic effect of the forget gate, input gate, and output gate, it predicts the candidate's risk of leaving the company within a predetermined period. The comprehensive evaluation process also includes the introduction of an employee turnover early warning mechanism. By mining the de-identified data of departing employees, the mechanism identifies characteristics such as fluctuations in leave frequency, increased social network activity, and decreased emotional value in communication language. These characteristics are then transformed into negative feature vectors and input into the long short-term memory neural network.

[0013] As a further preferred embodiment of the human resources information processing method of the present invention, in step five, the model parameters are adjusted through the backpropagation algorithm. Specifically, the backpropagation algorithm adopts the adaptive moment estimation optimizer Adam, with an initial learning rate set to 0.001. By training on at least 10,000 real onboarding samples, the average absolute error of the talent-job matching model is reduced to below 5% within 6 months. The feedback optimization process also adopts a federated learning framework, specifically the federated averaging algorithm FedAvg. When each branch office trains the model locally, it uses local performance data and only uploads the encrypted parameter gradients to the headquarters server for weighted average aggregation through homomorphic encryption or secure aggregation protocol. This achieves the co-evolution of the global talent matching model without aggregating the original performance data. To prevent catastrophic forgetting, the elastic weight consolidation EWC algorithm is introduced during the model update process, applying regularization constraints to important parameters when training new samples. The feedback loop and model iteration process includes: Regularly and automatically acquire candidates' key performance indicator achievement rates, comprehensive environmental impact assessment scores, and qualitative evaluation data from their superiors; The backpropagation algorithm uses stochastic gradient descent and a preset learning rate. By training on a predetermined scale of real job samples, the average absolute error of the talent-job matching model is reduced within a predetermined time. The feedback optimization process also employs a federated learning framework. When processing data distributed across different data centers, each branch trains its model locally and only uploads the encrypted parameter gradients to the headquarters server for aggregation. This enables the co-evolution of the global talent matching model without aggregating the original performance data.

[0014] As a further preferred embodiment of the human resources information processing method of the present invention, it also includes a data security protection step, a visualization interaction step, and a distributed computing scheduling step: The data security protection steps include: encrypting and storing the collected personal privacy data using an encryption algorithm; The data security protection steps include: encrypting and storing the collected personal privacy data using the AES-256-GCM encryption algorithm; the encryption key is independently generated by the hardware security module HSM and stored in protected non-volatile memory; employing differential privacy technology during data processing, setting a privacy budget ε=1.0, and adding random noise conforming to a Laplace distribution to the original data to ensure individual-level privacy protection; using a blockchain ledger to record the resume collection time, source channel, original hash value, and key node information during processing; using the Practical Byzantine Fault Tolerance (PBFT) consensus mechanism; and using smart contracts to achieve automatic on-chain and tamper-proof storage of operation logs. The distributed computing scheduling steps include: using a master-slave scheduling model to monitor the resource usage of the computing server; using the FairScheduler scheduling algorithm; splitting the job matching task into multiple sub-task packages according to the candidate ID hash; and distributing them to various computing nodes for parallel processing via a Kafka message queue. Differential privacy technology is employed in the data processing to add random noise conforming to a Laplace distribution to the original data; a blockchain ledger is used to record the collection time, source channel, original hash value, and key node information of the resumes during processing; the visualization interaction steps include: displaying the matching profile of candidates and positions through a graphical interface, supporting human resource managers to adjust the weights of each evaluation dimension and update the ranking results in real time; the distributed computing scheduling steps include: using a master-slave scheduling model to monitor the resource usage of the computing server, splitting the job matching task into multiple sub-task packages, and distributing them to various computing nodes for parallel processing through a message queue.

[0015] An apparatus based on a human resources information processing method includes: The hardware support platform includes a central processing unit (CPU) cluster, high-speed random access memory (RAM), AI acceleration chips, and a large-capacity solid-state storage array. The platform ensures continuous operation through redundant power supplies and a dual-machine hot standby mechanism. Regarding algorithm deployment, the AI ​​acceleration chips (GPU / NPU) handle the inference tasks of deep learning models, including CRNN-OCR recognition, BERT feature extraction, Transformer matching models, and computationally intensive LSTM potential prediction modules. Model quantization technology is used to compress the FP32 floating-point model into an INT8 fixed-point model, increasing inference speed by 4 times. The CPU cluster handles data preprocessing, distributed scheduling, and encryption / decryption tasks. The large-capacity solid-state storage array employs tiered storage for hot and cold data; hot data is stored on NVMe SSDs, while cold data is automatically migrated to large-capacity HDDs, reducing storage costs by over 40%. The data acquisition module is used to acquire candidates' resume information, job requirement description information, and historical recruitment data of enterprises through multi-source interface units, and to digitize unstructured documents using image scanning units and structured processing units. The feature extraction module, including a natural language processing engine, a feature mapping unit, and a label generation unit, is used to preprocess the original dataset and extract candidates' professional skills, work experience, education level, and behavioral feature labels to construct a multi-dimensional feature vector space. The intelligent matching module includes a vector calculation unit, a deep learning inference unit, and a matching scoring unit, which are used to construct a semantic association model between talents and positions based on the multi-dimensional feature vector space and generate preliminary matching scores. The comprehensive evaluation module, including a weight management unit, a potential prediction unit, and a ranking output unit, is used to incorporate the company's internal talent evaluation standards and job weight factors, to weight and correct the matching scores, and to use a recurrent neural network to predict the candidate's career development potential and stability. The feedback optimization module includes a performance data acquisition unit, a parameter update unit, and a model version control unit. It is used to track the candidate's performance data and feed it back to the intelligent matching module to adjust the model parameters through a backpropagation algorithm.

[0016] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the human resources information processing method.

[0017] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects: 1. This invention fundamentally changes the traditional screening model that relies on subjective human experience by constructing a multi-dimensional feature vector space and a deep semantic association model. Utilizing semantic vectors of preset dimensions and a multi-head attention mechanism, the system can deeply understand the inherent logical connection between resumes and job requirements, significantly improving the accuracy of talent-job matching. Simultaneously, by introducing internal corporate evaluation standards and behavioral characteristic analysis, it effectively eliminates subjective biases caused by interviewer preferences, ensuring the fairness and scientific rigor of talent selection.

[0018] 2. The high-performance optical character recognition technology and distributed computing architecture integrated in this invention enable rapid processing of massive amounts of human resource data. The time required for job matching tasks of a specific scale is reduced to a predetermined timeframe, significantly improving efficiency compared to traditional manual screening methods. The automated data collection and preprocessing process greatly reduces the transactional workload of human resource managers, allowing them to focus on higher-value strategic decision-making tasks.

[0019] 3. The feedback loop mechanism and model iteration process established in this invention enable the system to learn and evolve on its own. By collecting performance data after onboarding in real time and updating it in reverse, the model can automatically capture the changing trends of corporate talent needs and achieve dynamic optimization of accuracy. The prediction of talent potential and stability by the long short-term memory neural network provides reliable data support for enterprises to build a stable talent pipeline, effectively reducing talent turnover and recruitment costs;

[0020] 4. The advanced encryption standards, differential privacy technology, and blockchain evidence storage mechanism employed in this invention provide comprehensive security protection for massive amounts of sensitive human resource information. The visually appealing interactive interface and multi-dimensional profile display provide management with intuitive and detailed decision-making support, supporting dynamic weight adjustments and real-time simulation predictions, significantly improving the organization's human resource governance level and responsiveness to market changes. Attached Figure Description

[0021] Figure 1 This is a flowchart of the human resources information processing method proposed in this invention; Figure 2 This is a schematic diagram of the talent-job matching model based on deep learning in this invention; Figure 3 This is a logical flowchart of the human resources information processing and model feedback closed-loop iteration in this invention. Detailed Implementation

[0022] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings: The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention. The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. The purpose and effects of the present invention will become clearer. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0023] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0024] A human resources information processing method, such as Figure 1 As shown, it includes the following steps: Step 1: Collect raw human resources data. Obtain candidates' resumes, job descriptions, and historical recruitment data from companies through multi-source data interfaces. Use optical character recognition technology based on deep convolutional neural networks to extract and convert the digital features of unstructured paper documents to obtain a structured raw human resources dataset. This paper utilizes optical character recognition technology based on deep convolutional recurrent neural networks (CRNNs) to extract and transform digital features from unstructured paper documents. The CRNN consists of a cascaded convolutional neural network (CNN) and a recurrent neural network (RNN). The CNN part uses a ResNet-34 residual network structure to extract feature maps from the document image, while the RNN part uses a bidirectional long short-term memory network (BiLSTM) to perform temporal modeling of the feature sequence. Character sequence alignment and decoding are achieved through a connection-time classification (CTC) loss function. During recognition, the residual network structure analyzes the physical structure of the document, identifying paragraph boundaries, table row and column relationships, and headshot regions in the image. Table row and column relationships are identified using a table parsing model based on graph neural networks (TableNet), and headshot regions are located and cropped using a lightweight object detection network (YOLOv8).

[0025] Step 2, Preprocessing and Feature Extraction: The original human resources dataset is subjected to data deduplication, missing value imputation and semantic standardization. Natural language processing technology is used to extract candidates' professional skills, work experience, education level and behavioral feature labels. At the same time, hard indicators and soft quality requirements in job requirements are identified to construct a multi-dimensional feature vector space. Natural language processing (NLP) techniques are used to extract candidates' professional skills, work experience, education level, and behavioral characteristic tags. Specifically, this includes: segmenting the original text using word segmentation algorithms based on Hidden Markov Models (HMMs) or Conditional Random Fields (CRFs) combined with a custom human resources professional dictionary; mapping discrete text words to 768-dimensional continuous real-valued vectors using a BERT pre-trained language model, where the BERT model has undergone domain-adaptive fine-tuning on a corpus of millions of human resources words; extracting behavioral characteristic tags using sentiment analysis algorithms, which are based on the TextCNN convolutional neural network structure and extract sentiment features from the text through multi-scale convolutional kernels, outputting personality tendency scores for candidates in five dimensions: openness, conscientiousness, extraversion, agreeableness, and neuroticism; and extracting quantitative scores for candidates in teamwork, leadership, stress resistance, and innovative thinking dimensions through text mining of candidates' past project experiences using a hierarchical attention network (HAN), where the hierarchical attention network performs attention weighting at both the sentence and document levels.

[0026] Step 3: Construct a talent-job matching model. Based on the multi-dimensional feature vector space, a semantic association model between talent and job is constructed using a deep learning network. A multi-head attention mechanism is used to perform deep semantic fusion of candidate features and job requirements, calculating the correlation score between the candidate feature vector and the job requirement vector to generate a preliminary matching score. Specifically, the deep semantic fusion of candidate features and job requirements using the multi-head attention mechanism includes: the deep learning network employs a dual-tower structure based on a Transformer encoder to independently encode the candidate features and job requirements, obtaining a candidate feature vector sequence and a job feature vector sequence. The calculation formula for multi-head attention mechanism is as follows: ; in, , , They are generated by linear transformations of candidate characteristics and job characteristics, respectively. For the attention head dimension, eight parallel attention heads are set up, each capturing the correlation information of eight dimensions: education matching degree, skill matching degree, experience matching degree, project complexity matching degree, technology stack overlap degree, soft quality matching degree, cultural adaptability matching degree, and career development matching degree. Then, the outputs of the eight attention heads are dynamically weighted and fused using a learnable gated fusion network to obtain the final fused feature vector: ; Among them, gating weights , For the first The output vector of each attention head. Represents the candidate feature vector Job Feature Vector splicing, Here is the weight matrix of the gated network. For bias vectors, For the first The learnable temperature coefficient of an individual attention head; Calculate the cosine similarity between the candidate feature vector and the job requirement vector. ; in Representative candidate characteristic components, Represents the weight of job requirements. These are the dimension weights determined through feature importance assessment; Step four: Multi-dimensional evaluation and ranking. Introduce the company's internal talent evaluation standards and job weight factors to weight and correct the preliminary matching scores. Combine long short-term memory neural networks to analyze the candidates' career trajectory sequences, predict the candidates' career development potential and stability, and output a ranked list of recommended candidates. The job weighting factor is dynamically adjusted based on the urgency and strategic importance of the business department, and the adjustment formula is as follows: ; For the position The basic weighting factor, This is an adjustment factor based on business urgency. For the position Business urgency rating Rate the highest urgency level among all open positions. For the position Strategic importance score Rate the highest strategic importance among all open positions. This is a coefficient for adjusting for talent scarcity. For the position The market scarcity index of the required core skills. The maximum scarcity index is preset; The long short-term memory neural network employs a loss function with added stability regularization during training: ; in , The total number of training samples, For the first The candidate's true credentials The probability of leaving the company is predicted by LSTM. The regularization coefficient is . Let L2 be the norm of the final hidden state vector of the LSTM. For indicator functions; The analysis of candidates' career trajectory sequences is combined with a Long Short-Term Memory (LSTM) neural network. Specifically, the input layer of the LSTM neural network contains the candidate's career trajectory sequence sorted chronologically over the past 10 years. Each trajectory record includes features in six dimensions: tenure, job level, salary increase, company size, industry category, and number of promotions. The sequence length is uniformly padded or truncated to 20 time steps. The LSTM network adopts a three-layer stacked structure with 128 hidden units in each layer. Temporal features are extracted through the synergistic effect of the forget gate, input gate, and output gate. An attention mechanism is added to the top layer to weight key career turning points. The turnover risk prediction adopts a binary classification output layer with cross-entropy loss as the loss function. The comprehensive evaluation process also includes the introduction of an employee turnover early warning mechanism. By mining the desensitized data of departing employees, the Isolation Forest algorithm is used to identify abnormal features such as fluctuations in leave frequency, increased social network activity, and decreased emotional value of communication language. The detected abnormal features are transformed into negative feature vectors and input into the LSTM neural network.

[0027] Step 5: Feedback loop and model iteration. Real-time tracking of candidate performance data after onboarding. The performance data is normalized and then fed back to the talent-job matching model as a monitoring signal. The model parameters are adjusted through the backpropagation algorithm to achieve continuous optimization of the talent-job matching model.

[0028] The model parameters are adjusted through a backpropagation algorithm. Specifically, the backpropagation algorithm uses an adaptive moment estimator (Adam) with an initial learning rate of 0.001. By training on at least 10,000 real onboarding samples, the mean absolute error of the talent-job matching model is reduced to below 5% within 6 months. The feedback optimization process also adopts a federated learning framework, specifically the federated averaging algorithm (FedAvg). When training the model locally, each branch uses local performance data and only uploads the encrypted parameter gradients to the headquarters server for weighted average aggregation through homomorphic encryption or a secure aggregation protocol. This achieves the co-evolution of the global talent matching model without aggregating the original performance data. To prevent catastrophic forgetting, the elastic weight consolidation (EWC) algorithm is introduced during the model update process to apply regularization constraints to important parameters when training new samples.

[0029] It also includes data security protection steps, visualization and interaction steps, and distributed computing scheduling steps: The data security protection steps include: encrypting and storing the collected personal privacy data using an encryption algorithm; The data security protection steps include: encrypting and storing the collected personal privacy data using the AES-256-GCM encryption algorithm; the encryption key is independently generated by the hardware security module HSM and stored in protected non-volatile memory; employing differential privacy technology during data processing, setting a privacy budget ε=1.0, and adding random noise conforming to a Laplace distribution to the original data to ensure individual-level privacy protection; using a blockchain ledger to record the resume collection time, source channel, original hash value, and key node information during processing; using the Practical Byzantine Fault Tolerance (PBFT) consensus mechanism; and using smart contracts to achieve automatic on-chain and tamper-proof storage of operation logs. The distributed computing scheduling steps include: using a master-slave scheduling model to monitor the resource usage of the computing server; using the FairScheduler scheduling algorithm; splitting the job matching task into multiple sub-task packages according to the candidate ID hash; and distributing them to various computing nodes for parallel processing via a Kafka message queue. Differential privacy technology is employed in the data processing to add random noise conforming to a Laplace distribution to the original data; a blockchain ledger is used to record the collection time, source channel, original hash value, and key node information of the resumes during processing; the visualization interaction steps include: displaying the matching profile of candidates and positions through a graphical interface, supporting human resource managers to adjust the weights of each evaluation dimension and update the ranking results in real time; the distributed computing scheduling steps include: using a master-slave scheduling model to monitor the resource usage of the computing server, splitting the job matching task into multiple sub-task packages, and distributing them to various computing nodes for parallel processing through a message queue.

[0030] Example 1: This embodiment provides a human resources information processing method and apparatus, which is mainly applied to the digital governance scenarios of large multinational enterprises or super-large human resources service organizations. In these scenarios, due to the daily number of resumes received reaching tens of thousands or more, and the job requirements involving laws, regulations, language habits, and professional skills requirements in multiple regions around the world, traditional manual processing methods can no longer meet the real-time and accuracy requirements of the business.

[0031] This embodiment constructs a highly integrated hardware support platform as the physical foundation of the entire information processing system. The platform adopts a distributed computing architecture, deployed in a cluster of 12 high-performance industrial-grade servers. Each server is equipped with two 64-core central processing units (CPUs), totaling 1536 cores, with a stable clock speed above 3.2 GHz, ensuring the throughput of large-scale parallel computing. To address the high computing power requirements of deep learning model inference, the system integrates four dedicated AI acceleration chips in each server, with a single chip computing power of no less than 200 trillion floating-point operations per second (TFLOPS). In terms of memory, the system is equipped with 512 gigabytes of high-speed random access memory (DDR4 3200 MHz), achieving ultra-fast data exchange through 8-channel technology. The storage subsystem consists of an array of 24 7.68 terabyte embedded multimedia storage devices or enterprise-grade solid-state drives, with a measured sequential read / write speed greater than or equal to 5000 megabytes per second and random read performance (…). The platform surpasses the 1 million level in throughput. It connects to external networks via dual redundant 10 Gigabit Ethernet switches and is equipped with dual UPS uninterruptible power supplies and an intelligent temperature-controlled fire suppression system, ensuring over 99.99% system availability even in complex physical environments. Regarding algorithm deployment, the AI ​​acceleration chip (GPU / NPU) handles the inference tasks of deep learning models, including computationally intensive modules such as CRNN-OCR recognition, BERT feature extraction, Transformer matching models, and LSTM potential prediction. Model quantization technology compresses the FP32 floating-point model into an INT8 fixed-point model, increasing inference speed by 4 times. The central processing unit cluster handles data preprocessing, distributed scheduling, encryption, and decryption. The large-capacity solid-state storage array employs tiered storage for hot and cold data; hot data is stored on NVMe SSDs, while cold data is automatically migrated to large-capacity HDDs, reducing storage costs by over 40%.

[0032] Based on the above hardware architecture, the human resource information processing method of this embodiment is officially running. First, step (1) is executed to collect the original human resource data. The data collection module calls 25 predefined application programming interfaces (APIs) through the multi-source interface unit. These interfaces are deeply connected to mainstream global recruitment platforms, professional social networks, and the enterprise's internal office automation system. The system is set to automatically poll and synchronize the latest candidate delivery status and job change information at a frequency of once every 10 minutes. For a large number of paper archives or unstructured image documents积压 in the enterprise, the system performs digital acquisition through a high-resolution (not less than 600 DPI) document scanner connected by the image scanning unit. During this process, the system calls the built-in 50-layer Residual Neural Network (ResNet-50) model for optical character recognition. This model used more than 10 million resume images containing various complex layouts, handwritten signatures, and seal interferences during the training phase, and its recognition accuracy for specific terms in the human resource field reached 99.5%. The recognition process is not limited to text extraction but also includes parsing the physical structure of the document, such as automatically identifying paragraph boundaries, table row and column relationships, and avatar areas in images.

[0033] The collected original data then enters step (2) for preprocessing and feature extraction. The structured processing unit in the feature extraction module first performs duplicate checking on the data and uses the Bloom filter algorithm to determine whether the newly collected data is a duplicate entry within milliseconds. Subsequently, the natural language processing engine intervenes. This engine integrates a human resource-specific dictionary with more than 500,000 professional entries, covering thousands of job names and skill keywords from traditional manufacturing to the cutting-edge Internet industry. In the word segmentation stage, the system uses an improved Hidden Markov Model to segment the text and automatically removes more than 1,200 Chinese stop words such as "的", "在", "以及", and meaningless punctuation marks.

[0034] To achieve deep semantic understanding, the feature mapping unit maps discrete text words into 512-dimensional continuous real-valued vectors using word embedding technology. In this embodiment, these vectors not only contain the meaning of the words themselves but also extract professional value tags through sentiment analysis of candidates' public social media activity. For example, by analyzing more than 50 technical blog posts published by candidates on professional forums, a Long Short-Term Memory (LSTM) network is used to identify the rigor of their logical thinking, sensitivity to new technologies, and tendency towards teamwork. The system ultimately generates a quantitative scoring matrix for each candidate, containing eight core dimensions (such as leadership, resilience, and innovative thinking), with the scoring range normalized to between 0 and 1. Simultaneously, based on the job requirements description, the system uses the same logic to extract hard indicators (such as: full-time bachelor's degree or above, more than 5 years of Java development experience, and holding a senior system architect certificate) and soft skills requirements, constructing a multi-dimensional feature vector space.

[0035] Next, the system executes step (3) to construct a talent-job matching model. For example... Figure 2 As shown, the vector computation unit in the intelligent matching module utilizes a high-performance graphics processing unit (GPU) for large-scale matrix operations. This embodiment employs a deep semantic association model based on a multi-head attention mechanism. This model is configured with eight parallel attention heads, each focusing on capturing the correlation of different feature dimensions. For example, the first head might focus on matching educational background with professional requirements, while the fifth head focuses on the overlap between the candidate's past project complexity and the current job's technical stack. The system calculates the cosine similarity between the candidate's feature vector and the job requirement vector, using the following formula:

[0036]

[0037] in, Representative candidate characteristic components, Represents the weight of job requirements. These are preset dimension weights. In the initial configuration of this embodiment, the weight of the professional skills dimension is... Set to 0.4, with work experience as the weight. The weight of education level is set to 0.3. The initial score is set to 0.2, while other dimensions such as geographical location preference and salary expectation are weighted at 0.1. Through this calculation, the system generates a preliminary matching score for each person-job pairing.

[0038] Subsequently, in step (4) multi-dimensional evaluation and ranking, the comprehensive evaluation module intervenes. This module not only considers static matching scores but also introduces dynamic corporate evaluation standards. The weight management unit allows different business units to adjust the evaluation weights according to their current strategic priorities, with an adjustment range between 0.8 and 1.5 times. The potential prediction unit uses a recurrent neural network (RNN) with 3 hidden layers and 128 neurons in each layer to analyze the candidate's career trajectory sequence. By comprehensively calculating the candidate's average job-hopping frequency, promotion speed, and industry prosperity over the past 10 years, the system predicts their risk of leaving within the next 2 years. If the predicted probability of leaving is higher than 25%, the system will impose a corresponding penalty on the final score. The ranking output unit finally generates a detailed analysis report of the top 50 candidates based on the corrected total score and displays it on the visual interactive interface in the form of radar charts and bar charts.

[0039] Finally, the system achieves feedback loop and model iteration through step (5). The feedback optimization module is deeply integrated with the enterprise's internal performance management system (KPI System) and office automation system (OA). Whenever a candidate has been employed for 3, 6, and 12 months, the performance data acquisition unit automatically captures their key performance indicator achievement rate, 360-degree environmental assessment score, and supervisor evaluation. The parameter update unit uses these real performance results as labels for supervised learning and uses the stochastic gradient descent algorithm (SGD) to backpropagate and update the 512-dimensional vector weights and attention head parameters in the matching model. In the test of this embodiment, through continuous training on 12,000 real onboarding samples, the model's prediction error within 6 months decreased from the initial 15% to 4.2%, significantly improving the return on investment in recruitment.

[0040] Example 2: Based on Example 1, this example further details the specific implementation details of the system in terms of data security, privacy protection, and distributed task scheduling, in order to meet the industry standards of finance, defense, and other sectors with extremely high data sensitivity requirements.

[0041] The human resources information processing device in this embodiment adds a dedicated security encryption module at the hardware level. Based on Hardware Security Module (HSM) technology, this module ensures that all collected personal privacy data (such as ID numbers, contact numbers, and home addresses) undergoes hardware-level encryption using the 256-bit Advanced Encryption Standard (AES-256) before entering the database. The encryption key is independently generated by the hardware module and stored in protected non-volatile memory, preventing any external process from directly reading the plaintext data. During data processing, the system incorporates differential privacy technology. Specifically, when the feature extraction module calculates group statistical characteristics (such as the average salary level of a certain industry in a certain region), it automatically adds random noise conforming to a Laplace distribution to the original data. This processing method ensures that, without affecting macroscopic statistical characteristics, even if an attacker obtains part of the processing results, they cannot deduce the privacy information of a specific individual through reverse engineering.

[0042] At the methodological level, this embodiment extends the data acquisition logic in step (1). To ensure the authenticity and immutability of the data, the system records the collection time, source channel, original hash value, and key node information of each resume in real time in a blockchain ledger built based on private chain technology. Whenever a resume is updated or manually accessed, the system generates a new block and synchronizes it to the 12 nodes in the cluster. This traceability mechanism not only prevents internal personnel from maliciously modifying candidate evaluations but also provides legal-level electronic evidence for potential subsequent labor arbitration or qualification reviews.

[0043] To address the demands of large-scale concurrent processing, this embodiment employs a master-slave scheduling model in its distributed computing architecture. The scheduling center monitors the CPU utilization, memory availability, and network bandwidth of more than 10 computing servers in real time. When executing job matching tasks on a scale of tens of thousands, the scheduling center breaks the task down into more than 128 sub-task packages and distributes them to each computing node via a high-speed message queue. After processing, each node returns the intermediate results (i.e., a 512-dimensional vector) to the aggregation unit for final sorting. Real-world testing shows that when processing full matching of 50,000 resumes with 100 job positions, the system's total processing time is only 4.8 seconds, significantly improving the response speed of the human resources department under urgent recruitment tasks.

[0044] In the semantic standardization process of step (2), this embodiment refines the multilingual processing logic for a globalized context. The system incorporates a deep learning-based neural machine translation engine, capable of automatically recognizing and converting resume content in more than 12 mainstream languages. When constructing the multidimensional feature vector space, the system considers not only the literal meaning of words but also cultural context factors. For example, in the North American recruitment context, "Team Player" may emphasize communication and collaboration, while in the East Asian context, it may imply an adaptation to overtime culture. Through pre-training on different language corpora, the system achieves cross-language semantic alignment, ensuring fair evaluation of the global talent pool under unified standards.

[0045] Furthermore, the visual interaction steps in this embodiment have been significantly enhanced. The graphical interface not only displays the candidate's current status but also supports a "simulation prediction" function. Human resource managers can dynamically adjust the weight of a specific dimension using a slider (for example, temporarily increasing the weight of "innovative thinking" from 0.1 to 0.3), and the system will recalculate and update the ranking list within 500 milliseconds. This real-time interactive capability benefits from the front-end's use of a high-performance WebGL-based rendering engine, which can smoothly display knowledge graph relationships containing thousands of nodes, helping decision-makers clearly see the skill complementarity between candidates and existing team members.

[0046] Example 3: This embodiment focuses on describing the adaptive capabilities of the present invention in complex organizational structures and its in-depth application in long-term career management. In some large enterprises with dozens of subsidiaries, the talent needs of different business units (BUs) vary greatly. This embodiment solves this problem through a dynamic weight adjustment mechanism and an incremental learning algorithm.

[0047] At the system architecture level, the weight management unit in the comprehensive evaluation module is designed in a multi-tenant mode. Each subsidiary or business unit can define its own exclusive "evaluation template." For example, in the R&D department's template, algorithm capabilities and the number of patents account for as much as 60% of the weight, while the sales department's template focuses on past performance achievement rate and the breadth of network resources. The system supports dynamic switching and hot reloading of these templates, enabling seamless updates of the evaluation logic without restarting the service.

[0048] In terms of methodology and process, such as Figure 3As shown, this embodiment deeply explores the feedback loop of step (5). In addition to conventional performance appraisal data, the system also introduces an "employee turnover early warning" mechanism. By mining anonymized data of employees who have left the company in the past 20 years, the system discovered some implicit signs of impending departure (such as abnormal fluctuations in leave frequency, sudden increases in activity on professional social networks, and a decrease in the emotional value of communication language in internal collaboration tools). These features are transformed into a 128-dimensional negative feature vector and input into the long short-term memory neural network in real time. When the system detects that the probability of a core talent leaving the company exceeds 70%, it will automatically push an early warning report to the human resources partner (HRBP) to which the talent belongs, and recommend internal transfer opportunities that are more suitable for the talent's career development from the internal job pool based on the talent's current skill characteristics, thereby achieving precise talent retention.

[0049] The hardware support platform in this embodiment has also been specifically optimized in terms of storage architecture. Considering the significant time-sensitive nature of human resources data, the system adopts a tiered storage technology for hot and cold data. Active candidate data from the past two years is stored in an NVMe solid-state drive array to ensure extremely fast retrieval and matching speeds; while existing data that has not been updated for more than five years is automatically migrated to a large-capacity hard disk drive array or tape library. In this way, the system reduces overall storage costs by more than 40% while ensuring core business performance.

[0050] In the intelligent matching module, this embodiment introduces "Knowledge Graph" technology. The system constructs an industry talent knowledge graph containing 10 million entities and 50 million relationships through automated mining of millions of resumes and industry research reports. Entities include skill points, company names, school names, project types, etc., while relationships include "prerequisite skills," "competitor relationships," "alumni relationships," etc. When performing semantic matching in step (3), the system not only calculates vector similarity but also reasones along the paths of the knowledge graph. For example, if a job requires "deep learning" skills, but the candidate's resume only lists "PyTorch" and "convolutional neural networks," the system can accurately determine that the candidate possesses the core skills required for the job through the association logic of the knowledge graph, thereby avoiding the omission problem caused by traditional keyword matching.

[0051] Finally, this embodiment also provides a remote interview assistance function based on augmented reality (AR). When HR managers conduct remote interviews, the system overlays a multi-dimensional matching profile of the candidate onto the interviewer's terminal screen in real time via an encrypted network channel. Based on the real-time speech recognition results during the interview, the system automatically prompts the interviewer to ask follow-up questions targeting the candidate's weaknesses. For example, if the system detects a low score for a candidate in the "team conflict handling" dimension, it will automatically display a set of targeted behavioral interview questions. This human-machine collaborative working mode increases the standardization of interviews by more than 50%, significantly reducing recruitment risks caused by varying interviewer skill levels.

[0052] Example 4: This embodiment aims to illustrate the robustness of the present invention under large-scale concurrency and high-load environments, and to describe in detail the system's defense mechanism against non-standardized data and malicious data injection.

[0053] In this embodiment, a dual-machine hot standby and heartbeat detection mechanism is introduced into the hardware support platform. A cluster of 10 or more servers is divided into two independent logical partitions, located on different physical racks and powered by different power distribution units. The master scheduling node sends a heartbeat signal to the slave nodes every 100 milliseconds. Once a master node failure is detected, the standby node will take over all system services within 2 seconds, ensuring the continuity of matching tasks on a scale of tens of thousands. Furthermore, to address the heat generated by large-scale matrix operations, the hardware platform is equipped with a closed-loop liquid cooling system. Through coolant circulation, the operating temperature of the computing chips is kept below 65 degrees Celsius. Compared to traditional air-cooling solutions, its energy efficiency ratio is significantly higher. The value was reduced to 1.15.

[0054] In step (1) of the method execution, the system adds a rigorous data cleaning and integrity verification process. For unauthorized or incorrectly formatted interference data obtained by web crawlers, the structured processing unit employs an anomaly detection model based on Generative Adversarial Networks (GANs). This model can identify "fake resumes" generated through artificial intelligence or "resume bombs" maliciously filled with keywords. Once an anomaly is detected, the system automatically isolates it in a sandbox environment for further manual review to prevent malicious data from contaminating the parameters of the talent-job matching model.

[0055] In the feature extraction process of step (2), this embodiment strengthens the quantitative modeling of "soft skills". The system analyzes the candidate's role positioning and contribution description in past projects, and uses a graph neural network (GNN) to calculate the centrality of the candidate in the professional social network. This centrality is regarded as a quantitative indicator to measure the candidate's industry influence and resource integration ability. At the same time, the system also introduces a reputation assessment algorithm, which automatically generates a credit score by cross-checking the consistency between the project experience filled in by the candidate and the publicly available bidding information and news reports of the company. If the credit score is lower than 60 points, the system will place the candidate at the bottom in the final ranking and mark the potential risk of resume fraud.

[0056] In the deep learning inference stage of step (3), the system employs model quantization and pruning techniques to further improve computational efficiency. Without sacrificing matching accuracy, the 32-bit floating-point model is compressed into an 8-bit fixed-point model, resulting in a 4-fold increase in processing speed on the dedicated AI acceleration chip. This means that under the same hardware conditions, the system can simultaneously support more concurrent human resource management personnel performing real-time queries, and support at least 1000 concurrent users receiving feedback results within 2 seconds.

[0057] In the comprehensive evaluation of step (4), this embodiment adds a "organizational culture fit" evaluation submodule. This module constructs a "cultural gene map" of the enterprise by capturing text features from internal forums, email correspondence (after anonymization), and team-building activities. The system matches the candidate's behavioral characteristic tags with the enterprise's cultural genes. For example, if the enterprise culture emphasizes "results-oriented" and "rapid iteration," while the candidate's characteristics show "perfectionism" and "long-term research tendency," the system will provide corresponding suitability suggestions. This deep-level matching logic not only focuses on whether the candidate "can do" the current job, but also on whether they "can stay" in the current enterprise environment.

[0058] Finally, step (5) of the feedback optimization module in this embodiment introduces a federated learning framework. When processing data from multinational corporations distributed across data centers in different countries, to comply with local data export security regulations, the system no longer directly aggregates raw performance data. Instead, each branch trains its model locally, only uploading the encrypted parameter gradients to the headquarters server for aggregation. This approach protects the privacy of employees in different countries while enabling the co-evolution of the global talent matching model.

[0059] In summary, this invention, by constructing a high-performance hardware support platform and combining advanced technologies such as deep learning, natural language processing, knowledge graphs, and blockchain, realizes an efficient, objective, and continuously evolving method and apparatus for processing human resource information. It not only solves the pain points of low efficiency and significant subjective bias in traditional manual screening but also provides comprehensive decision support for corporate talent strategies through feedback loops and potential prediction.

[0060] This application provides a schematic diagram of the structure of an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned human resources information processing method. The electronic device can be a server, a personal computer, or a cloud server.

[0061] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps in the above-described method embodiments. The computer-readable storage medium can be any medium capable of storing data, such as ROM, RAM, magnetic disk, optical disk, etc.

[0062] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

[0063] It will be understood by those skilled in the art that the above descriptions are merely preferred examples of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention. All technical features in this embodiment can be freely combined according to actual needs.

[0064] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for processing human resource information, characterized in that, Includes the following steps: Step 1: Collect raw human resources data. Obtain candidates' resumes, job descriptions, and historical recruitment data from companies through multi-source data interfaces. Use optical character recognition technology based on deep convolutional neural networks to extract and convert the digital features of unstructured paper documents to obtain a structured raw human resources dataset. Step 2, Preprocessing and Feature Extraction: The original human resources dataset is subjected to data deduplication, missing value imputation and semantic standardization. Natural language processing technology is used to extract candidates' professional skills, work experience, education level and behavioral feature labels. At the same time, hard indicators and soft quality requirements in job requirements are identified to construct a multi-dimensional feature vector space. Step 3: Construct a talent-job matching model. Based on the multi-dimensional feature vector space, a semantic association model between talent and job is constructed using a deep learning network. A multi-head attention mechanism is used to perform deep semantic fusion of candidate features and job requirements, calculating the correlation score between the candidate feature vector and the job requirement vector to generate a preliminary matching score. Specifically, the deep semantic fusion of candidate features and job requirements using the multi-head attention mechanism includes: the deep learning network employs a dual-tower structure based on a Transformer encoder to independently encode the candidate features and job requirements, obtaining a candidate feature vector sequence and a job feature vector sequence. The calculation formula for multi-head attention mechanism is as follows: ; in, , , They are generated by linear transformations of candidate characteristics and job characteristics, respectively. For the attention head dimension, eight parallel attention heads are set up, each capturing the correlation information of eight dimensions: education matching degree, skill matching degree, experience matching degree, project complexity matching degree, technology stack overlap degree, soft quality matching degree, cultural adaptability matching degree, and career development matching degree. Then, the outputs of the eight attention heads are dynamically weighted and fused using a learnable gated fusion network to obtain the final fused feature vector: ； Among them, gating weights , For the first The output vector of each attention head. Represents the candidate feature vector Job Feature Vector splicing, Here is the weight matrix of the gated network. For bias vectors, For the first The learnable temperature coefficient of an individual attention head; Calculate the cosine similarity between the candidate feature vector and the job requirement vector. ; in Representative candidate characteristic components, Represents the weight of job requirements. These are the dimension weights determined through feature importance assessment; Step four: Multi-dimensional evaluation and ranking. Introduce the company's internal talent evaluation standards and job weight factors to weight and correct the preliminary matching scores. Combine long short-term memory neural networks to analyze the candidates' career trajectory sequences, predict the candidates' career development potential and stability, and output a ranked list of recommended candidates. The job weighting factor is dynamically adjusted based on the urgency and strategic importance of the business department, and the adjustment formula is as follows: ; For the position The basic weighting factor, This is an adjustment factor based on business urgency. For the position Business urgency rating Rate the highest urgency level among all open positions. For the position Strategic importance score Rate the highest strategic importance among all open positions. This is a coefficient for adjusting for talent scarcity. For the position The market scarcity index of the required core skills. The maximum scarcity index is preset; The long short-term memory neural network employs a loss function with added stability regularization during training: ； in , The total number of training samples, For the first The candidate's true credentials The probability of leaving the company is predicted by LSTM. The regularization coefficient is . Let L2 be the norm of the final hidden state vector of the LSTM. For indicator functions; Step 5: Feedback loop and model iteration. Real-time tracking of candidate performance data after onboarding. The performance data is normalized and then fed back to the talent-job matching model as a monitoring signal. The model parameters are adjusted through the backpropagation algorithm to achieve continuous optimization of the talent-job matching model.

2. The human resources information processing method according to claim 1, characterized in that, In step one, This paper utilizes optical character recognition technology based on deep convolutional recurrent neural networks (CRNNs) to extract and transform digital features from unstructured paper documents. The CRNN consists of a cascaded convolutional neural network (CNN) and a recurrent neural network (RNN). The CNN part uses a ResNet-34 residual network structure to extract feature maps from the document image, while the RNN part uses a bidirectional long short-term memory (BiLSTM) network to perform temporal modeling of the feature sequence. Character sequence alignment and decoding are achieved through a connection-based temporal classification (CTC) loss function. During recognition, the residual network structure analyzes the physical structure of the document, identifying paragraph boundaries, table row and column relationships, and headshot regions in the image. Table row and column relationships are identified using a TableNet table parsing model based on graph neural networks, and headshot regions are located and cropped using a lightweight object detection network, YOLOv8. The process of obtaining information through multi-source data interfaces includes: Call the application programming interface of a third-party recruitment platform to synchronously update the candidate application status at a preset frequency; Using web crawler technology to retrieve publicly available information from industry talent pools within authorized scope; Obtain existing talent data through the database connection protocol of the enterprise's internal office system; The optical character recognition technology uses a residual network structure with a predetermined number of layers to extract features from resume images, recognize and convert various document formats, including portable document formats, image formats and documents generated by word processing software; During the recognition process, the residual network structure parses the physical structure of the document and identifies paragraph boundaries, table row and column relationships, and headshot regions in images.

3. The human resources information processing method according to claim 1, characterized in that, In step two, natural language processing (NLP) techniques are used to extract candidates' professional skills, work experience, education level, and behavioral feature tags. Specifically, this includes: segmenting the original text using a word segmentation algorithm based on a Hidden Markov Model (HMM) or Conditional Random Field (CRF) combined with a custom human resources professional dictionary; mapping discrete text words to 768-dimensional continuous real-valued vectors using a BERT pre-trained language model, where the BERT model has undergone domain-adaptive fine-tuning on a corpus of millions of human resources words; extracting behavioral feature tags using a sentiment analysis algorithm based on the TextCNN convolutional neural network structure, which extracts sentiment features from the text through multi-scale convolutional kernels, outputting personality tendency scores for candidates in five dimensions: openness, conscientiousness, extraversion, agreeableness, and neuroticism; and extracting quantitative scores in teamwork, leadership, resilience, and innovative thinking dimensions using a hierarchical attention network (HAN) through text mining of candidates' past project experiences, where the hierarchical attention network performs attention weighting at both the sentence and document levels. The process of semantic standardization and feature extraction includes: Establish a human resources professional dictionary containing a predetermined number of entries; use word segmentation algorithms to segment the original text and remove stop words and meaningless characters; By using word embedding technology, discrete text words are mapped to continuous real vectors of a preset dimension, ensuring that the distance between semantically similar words in the vector space is within a preset range. Extracting behavioral characteristic tags from candidates includes: analyzing candidates' public activities on social media or professional forums, and using sentiment analysis algorithms to identify their career values ​​and personality tendencies; By analyzing the text of candidates’ past project experiences, we extracted quantitative scores for their teamwork, leadership, stress resistance, and innovative thinking. The hard requirements extracted from job descriptions include: automatically identifying educational requirements, major requirements, years of work experience requirements, and specific qualification certificate requirements; Extracting soft skills requirements includes identifying job descriptions for communication skills, logical thinking skills, and organizational and coordination skills.

4. The human resources information processing method according to claim 1, characterized in that, In step three, the process of constructing the talent-job matching model includes: The deep learning network employs a multi-head attention mechanism, capturing matching features of different dimensions through a predetermined number of parallel attention heads; The calculation process of the correlation score is as follows: when calculating the similarity between the candidate feature vector and the job requirement vector, different initial weights are set for different feature dimensions. The feature dimensions include professional skills dimension, work experience dimension, education level dimension and other dimensions, and each dimension is set with a corresponding preset weight ratio. The intelligent matching process uses a high-performance graphics processor to perform large-scale matrix operations, supports model files generated by various deep learning frameworks, and generates a correlation score matrix between candidates and positions. The different attention heads in the multi-head attention mechanism focus on capturing the correlation information of educational background matching degree, project complexity, and technology stack overlap.

5. The human resources information processing method according to claim 1, characterized in that, In step four, the candidate's career trajectory sequence is analyzed using a Long Short-Term Memory (LSTM) neural network. Specifically, the input layer of the LSTM neural network contains the candidate's career trajectory sequence sorted chronologically over the past 10 years. Each trajectory record includes features in six dimensions: tenure, job level, salary increase, company size, industry category, and number of promotions. The sequence length is uniformly padded or truncated to 20 time steps. The LSTM network adopts a three-layer stacked structure with 128 hidden units per layer. Temporal features are extracted through the synergistic effect of the forget gate, input gate, and output gate. An attention mechanism is added to the top layer to weight key career turning points. The turnover risk prediction uses a binary classification output layer with cross-entropy loss as the loss function. The comprehensive evaluation process also includes the introduction of an employee turnover early warning mechanism. By mining the desensitized data of departing employees, the Isolation Forest algorithm is used to identify abnormal features such as fluctuations in leave frequency, increased social network activity, and decreased emotional value of communication language. The detected abnormal features are then converted into negative feature vectors and input into the long short-term memory neural network. The process of multi-dimensional evaluation and ranking includes: The company's internal talent evaluation criteria include the alignment of the company's core values, organizational culture adaptability score, and talent structure balance factor within departments. The job weighting factors are dynamically adjusted based on the urgency and strategic importance of the business departments, with the adjustment range within a preset proportion. The input layer of the long short-term memory neural network contains the candidate's career trajectory sequence within a predetermined period. Through the synergistic effect of the forget gate, input gate, and output gate, it predicts the candidate's risk of leaving the company within a predetermined period. The comprehensive evaluation process also includes the introduction of an employee turnover early warning mechanism. By mining the de-identified data of departing employees, the mechanism identifies characteristics such as fluctuations in leave frequency, increased social network activity, and decreased emotional value in communication language. These characteristics are then transformed into negative feature vectors and input into the long short-term memory neural network.

6. The human resources information processing method according to claim 1, characterized in that, In step five, the model parameters are adjusted using the backpropagation algorithm. Specifically, the backpropagation algorithm employs the Adam adaptive moment estimator optimizer with an initial learning rate of 0.

001. By training on at least 10,000 real onboarding samples, the mean absolute error of the talent-job matching model is reduced to below 5% within six months. The feedback optimization process also employs a federated learning framework, specifically the FedAvg federated averaging algorithm. Each branch uses local performance data for model training, and only uploads the encrypted parameter gradients to the headquarters server for weighted average aggregation via homomorphic encryption or a secure aggregation protocol. This achieves the co-evolution of the global talent matching model without aggregating the original performance data. To prevent catastrophic forgetting, the Elastic Weight Consolidation (EWC) algorithm is introduced during model updates, applying regularization constraints to important parameters when training new samples. The feedback loop and model iteration process includes: Regularly and automatically acquire candidates' key performance indicator achievement rates, comprehensive environmental impact assessment scores, and qualitative evaluation data from their superiors; The backpropagation algorithm uses stochastic gradient descent and a preset learning rate. By training on a predetermined scale of real job samples, the average absolute error of the talent-job matching model is reduced within a predetermined time. The feedback optimization process also employs a federated learning framework. When processing data distributed across different data centers, each branch trains its model locally and only uploads the encrypted parameter gradients to the headquarters server for aggregation. This enables the co-evolution of the global talent matching model without aggregating the original performance data.

7. The human resources information processing method according to claim 1, characterized in that, It also includes data security protection steps, visualization and interaction steps, and distributed computing scheduling steps: The data security protection steps include: encrypting and storing the collected personal privacy data using an encryption algorithm; The data security protection steps include: encrypting and storing the collected personal privacy data using the AES-256-GCM encryption algorithm; the encryption key is independently generated by the hardware security module HSM and stored in protected non-volatile memory; employing differential privacy technology during data processing, setting a privacy budget ε=1.0, and adding random noise conforming to a Laplace distribution to the original data to ensure individual-level privacy protection; using a blockchain ledger to record the resume collection time, source channel, original hash value, and key node information during processing; using the Practical Byzantine Fault Tolerance (PBFT) consensus mechanism; and using smart contracts to achieve automatic on-chain and tamper-proof storage of operation logs. The distributed computing scheduling steps include: using a master-slave scheduling model to monitor the resource usage of the computing server; using the FairScheduler scheduling algorithm; splitting the job matching task into multiple sub-task packages according to the candidate ID hash; and distributing them to various computing nodes for parallel processing via a Kafka message queue. Differential privacy technology is employed in the data processing to add random noise conforming to a Laplace distribution to the original data; a blockchain ledger is used to record the collection time, source channel, original hash value, and key node information of the resumes during processing; the visualization interaction steps include: displaying the matching profile of candidates and positions through a graphical interface, supporting human resource managers to adjust the weights of each evaluation dimension and update the ranking results in real time; the distributed computing scheduling steps include: using a master-slave scheduling model to monitor the resource usage of the computing server, splitting the job matching task into multiple sub-task packages, and distributing them to various computing nodes for parallel processing through a message queue.

8. An apparatus based on the human resources information processing method according to any one of claims 1 to 7, characterized in that, include: The hardware support platform includes a central processing unit (CPU) cluster, high-speed random access memory (RAM), AI acceleration chips, and a large-capacity solid-state storage array. The platform ensures continuous operation through redundant power supplies and a dual-machine hot standby mechanism. Regarding algorithm deployment, the AI ​​acceleration chips (GPU / NPU) handle the inference tasks of deep learning models, including CRNN-OCR recognition, BERT feature extraction, Transformer matching models, and computationally intensive LSTM potential prediction modules. Model quantization technology is used to compress the FP32 floating-point model into an INT8 fixed-point model, increasing inference speed by 4 times. The CPU cluster handles data preprocessing, distributed scheduling, and encryption / decryption tasks. The large-capacity solid-state storage array employs tiered storage for hot and cold data; hot data is stored on NVMe SSDs, while cold data is automatically migrated to large-capacity HDDs, reducing storage costs by over 40%. The data acquisition module is used to acquire candidates' resume information, job requirement description information, and historical recruitment data of enterprises through multi-source interface units, and to digitize unstructured documents using image scanning units and structured processing units. The feature extraction module, including a natural language processing engine, a feature mapping unit, and a label generation unit, is used to preprocess the original dataset and extract candidates' professional skills, work experience, education level, and behavioral feature labels to construct a multi-dimensional feature vector space. The intelligent matching module includes a vector calculation unit, a deep learning inference unit, and a matching scoring unit, which are used to construct a semantic association model between talents and positions based on the multi-dimensional feature vector space and generate preliminary matching scores. The comprehensive evaluation module, including a weight management unit, a potential prediction unit, and a ranking output unit, is used to incorporate the company's internal talent evaluation standards and job weight factors, to weight and correct the matching scores, and to use a recurrent neural network to predict the candidate's career development potential and stability. The feedback optimization module includes a performance data acquisition unit, a parameter update unit, and a model version control unit. It is used to track the candidate's performance data and feed it back to the intelligent matching module to adjust the model parameters through a backpropagation algorithm.

9. An electronic device, characterized in that, include: A processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the human resources information processing method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the human resources information processing method as described in any one of claims 1 to 7.

Images