Intelligent paper marking system and method for talent selection and recruitment

Abstract

The invention relates to the technical field of human resource management, and provides an intelligent paper marking system and method for talent selection and recruitment. The method comprises the following steps: cutting test paper according to a preset test paper template by using an image processing technology to obtain a plurality of plates corresponding to the test paper; converting the image content of each section of the cut test paper into a text format which can be edited and processed by adopting a character recognition technology; setting a preliminary scoring standard according to question type information and knowledge point information corresponding to the test paper, and optimizing the preliminary scoring standard by using a large language model; performing semantic understanding and logic analysis on the converted test paper text answers from a plurality of scoring dimensions including semantic accuracy, logic integrity and knowledge point coverage by using two preset scoring large models to give corresponding scoring information, and marking advantage information and defect information in the answers; and obtaining a comprehensive score corresponding to the test paper according to score information printed by each score big model for each question of each test paper.

Description

Technical Field

[0001] The present application relates to the technical field of human resource management, and in particular to an intelligent marking system and method for talent selection and recruitment. Background Art

[0002] In the talent recruitment process, grading subjective exams is a key step in assessing candidate capabilities. Traditional grading methods, which rely primarily on manual labor, have the following significant drawbacks: 1. Inefficiency: Manually reviewing thousands of test papers takes weeks, making it difficult to meet the timeliness requirements of large-scale recruitment scenarios. 2. Highly subjective: Scoring results are influenced by the examiner's professional level and subjective preferences, which can easily lead to scoring bias and result in a less objective assessment of candidate ability. 3. High cost: A large amount of manpower is required for repetitive tasks such as test paper cutting, content recognition, and score calculation, and the manual review process further increases time and financial costs. 4. Disconnected from job requirements: Traditional scoring criteria rely on manual experience and lack a systematic mapping of job competency models (such as communication skills, logical reasoning skills, depth of professional knowledge, etc.), making it difficult to accurately match the personalized needs of talent selection.

[0003] While some existing technologies have attempted automated grading based on OCR technology, these efforts are limited to objective question recognition or simple keyword matching, failing to address the core pain points of subjective question grading, including: 1. Existing solutions for generating scoring criteria often rely on manual pre-setting or fixed rules, failing to implement intelligent dynamic optimization based on job competency models. 2. The scoring process generally uses a single model, lacking a collaborative mechanism for multi-dimensional scoring models and unable to reduce subjective errors through inter-model score verification; 3. A closed-loop "scoring-feedback-optimization" mechanism has not been established, making it difficult to improve scoring accuracy through iterative use of historical data. This is especially true in talent recruitment scenarios, where personalized scoring strategies cannot be developed for different job requirements.

[0004] Therefore, a method is urgently needed to solve at least one of the above problems. Summary of the Invention

[0005] The present application provides an intelligent marking system and method for talent selection and recruitment, aiming to solve the problem that, although there are some attempts at automated marking based on OCR technology in the existing technology, they only stay at the level of objective question recognition or simple keyword matching, and cannot solve multiple core pain points in subjective question scoring.

[0006] In a first aspect, an embodiment of the present application provides an intelligent examination paper marking method for talent selection, the method comprising: Image processing technology is used to cut the test paper according to the preset test paper template to obtain multiple sections corresponding to the test paper; text recognition technology is used to convert the image content of each section of the cut test paper into an editable text format; Set preliminary scoring criteria based on the question type and knowledge point information corresponding to the test paper, and use the large language model to optimize the preliminary scoring criteria in multiple scoring dimensions based on question type, knowledge points, and answer logic; Use at least two pre-set scoring models to perform semantic understanding and logical analysis on the converted test paper text answers from multiple scoring dimensions, including semantic accuracy, logical completeness, and knowledge point coverage, to provide corresponding scoring information and mark the advantages and disadvantages of the answers; The comprehensive score corresponding to each test paper is obtained based on the scoring information of each scoring model for each question on each test paper.

[0007] In some embodiments, the test paper is cut according to a preset test paper template using image processing technology to obtain multiple sections corresponding to the test paper, including: determining the specific positions of each question area, answer area and personal information area in the test paper by identifying the positioning marks and boundary features preset in the test paper template, and segmenting the test paper image based on the positions to obtain independent sections corresponding to each area.

[0008] In some embodiments, the text recognition technology is used to convert the image content of each section of the cut test paper into an editable text format, including: performing noise reduction, tilt correction and illumination equalization pre-processing on the cut test paper section image, recognizing the text content in the image line by line and word by word through a text recognition model, generating text data containing coordinate position information, and performing semantic error correction and format standardization on the recognition results.

[0009] In some embodiments, the preliminary scoring criteria are set according to the question type information and knowledge point information corresponding to the test paper, including: parsing the test paper text structure through natural language processing technology, automatically identifying the question type category, test knowledge point labels and answer requirements corresponding to each question, combining the competency model database of talent selection positions, matching the preset question type scoring rule template, and automatically generating preliminary scoring criteria including scoring points, key scoring items and deduction rules based on knowledge point weights, answer logic complexity and job ability requirement parameters; the question type categories include subjective questions and objective questions.

[0010] In some embodiments, the use of a large language model to optimize the preliminary scoring criteria in multiple scoring dimensions from the perspective of question type, knowledge point, and answer logic includes: inputting the preliminary scoring criteria into the large language model, analyzing the relationship between question type and knowledge point, and the rationality requirements of answer logic through the model, adjusting the weight distribution of scoring points, keyword extraction rules, and semantic similarity calculation methods, and fine-tuning the model in the field based on historical high-quality answer data to generate optimized scoring criteria.

[0011] In some embodiments, the use of at least two preset scoring models to perform semantic understanding and logical analysis on the converted test paper text answers from multiple scoring dimensions including semantic accuracy, logical completeness, and knowledge point coverage to provide corresponding scoring information, including: using the first scoring model to analyze the semantic accuracy and knowledge point coverage of the answers, and using the second scoring model to analyze the logical completeness and argument rationality of the answers, the first scoring model and the second scoring model respectively calculate scores according to the preset scoring dimension weights to generate independent scoring information; wherein, the first scoring model is a natural language processing model based on the Transformer architecture, and the second scoring model is a logical reasoning model based on deep learning.

[0012] In some embodiments, marking the advantage information and disadvantage information in the answer includes: comparing the answer text with the optimized scoring criteria, identifying key sentences in the answer that meet the scoring points through a model as advantage information, identifying content that deviates from the scoring points, logical contradictions or knowledge point errors as disadvantage information, and generating annotation marks at corresponding positions in the answer text.

[0013] In some embodiments, the comprehensive score corresponding to the test paper is obtained based on the scoring information of each scoring model for each question on each test paper, including: calculating the difference in scores of the two scoring models for the same question, if the difference is less than or equal to a preset threshold determined based on historical scoring data statistics, taking the average of the two scores as the score for the question; if the difference exceeds the preset threshold, the manual review process is triggered, and the final score is determined manually by combining the marked advantage information and defect information.

[0014] In some embodiments, the method further includes: establishing a historical scoring database, associating and storing the test paper text answers, scoring information, manual review results and scoring standard versions generated for each grading, analyzing historical data through machine learning algorithms, identifying unreasonable scoring points and model parameter deviations in the scoring standards, and automatically iteratively optimizing the scoring standards and weight parameters of the large scoring model.

[0015] In a second aspect, the present application provides an intelligent examination system for talent selection and recruitment, the system comprising: The image cutting unit is used to cut the test paper according to the preset test paper template using image processing technology to obtain multiple sections corresponding to the test paper; and use text recognition technology to convert the image content of each section of the cut test paper into an editable text format; The standard setting unit is used to set preliminary scoring standards based on the question type and knowledge point information corresponding to the test paper, and use the large language model to optimize the preliminary scoring standards in multiple scoring dimensions based on question type, knowledge point, and answer logic; A scoring acquisition unit is used to use at least two preset scoring models to perform semantic understanding and logical analysis on the converted test paper text answers from multiple scoring dimensions including semantic accuracy, logical completeness, and knowledge point coverage to provide corresponding scoring information and mark the advantages and disadvantages of the answers; The scoring completion unit is used to obtain the comprehensive score corresponding to the test paper based on the scoring information of each question on each test paper given by each scoring model.

[0016] An intelligent grading system and method for talent selection provided in an embodiment of the present application shortens the large-scale test grading cycle from several weeks of traditional manual grading to a single day through automated test paper segmentation, text recognition, and model scoring, thereby meeting the efficiency requirements of talent selection. At least two large scoring models are used to independently score from multiple dimensions such as semantic accuracy and logical integrity, and a difference verification mechanism is used to reduce the subjective bias of a single model. Scoring standards are generated in combination with a job competency model, so that the scoring results are more in line with job capability requirements. Manual intervention is reduced, and the high labor cost of traditional grading is converted into low marginal costs of model training and system maintenance, significantly improving the economy of the selection process. Scoring standards and model parameters are reversely optimized through historical scoring data, so that the system gradually improves scoring accuracy as the number of uses increases, adapting to the personalized selection requirements of different positions. Scoring standards are automatically generated based on the job competency model, and the test paper scores are directly linked to the capabilities required for the position (such as logical reasoning, application of professional knowledge, etc.), thereby improving the pertinence and accuracy of talent assessment.

[0017] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0019] Figure 1This is a flowchart illustrating the steps of an intelligent examination paper marking method for talent selection provided by one embodiment of the present application; Figure 2 This is a schematic block diagram of the structure of an intelligent examination system for talent selection and recruitment provided by an embodiment of the present application; Figure 3 This is a schematic block diagram of the structure of a computer device provided in one embodiment of the present application.

[0020] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0022] The flowcharts shown in the accompanying drawings are for illustrative purposes only and do not necessarily include all contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may be decomposed, combined, or partially merged, so the actual execution order may vary depending on the actual situation.

[0023] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present invention, in the embodiments of the present invention, terms such as "first" and "second" are used to distinguish between identical or similar items having substantially the same functions and effects. Those skilled in the art will understand that terms such as "first" and "second" do not limit the quantity or order of execution, and that terms such as "first" and "second" do not necessarily define differences.

[0024] It should be understood that the terms used in this specification are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in this specification and the appended claims, the singular forms "a", "an", and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0025] It will also be understood that the term "and / or" as used in this specification and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0026] In the talent recruitment process, grading subjective exams is a critical step in assessing candidate capabilities. Traditional grading methods, which primarily rely on manual effort, suffer from significant drawbacks, including: low efficiency: manually grading thousands of exams can take weeks, making it difficult to meet the timeliness requirements of large-scale recruitment scenarios; high subjectivity: grading results are influenced by the examiner's professional level and subjective preferences, which can easily lead to biased scoring and a less objective assessment of candidate capabilities; high cost: a significant amount of manpower is required for repetitive tasks such as exam paper segmentation, content recognition, and score compilation, and manual review further increases time and financial costs; and disconnection from job requirements: traditional scoring criteria rely on manual experience and lack a systematic mapping of job competency models (such as communication skills, logical reasoning skills, and depth of professional knowledge), making it difficult to accurately match the personalized needs of talent recruitment.

[0027] Although there have been some attempts at automated grading based on OCR technology in the existing technology, they have only remained at the level of objective question recognition or simple keyword matching, and are unable to solve the core pain points of subjective question scoring, including: for the generation of scoring criteria, existing solutions mostly rely on manual presets or fixed rules, and fail to achieve intelligent dynamic optimization combined with job competency models; the scoring process generally adopts a single model, lacks a collaborative mechanism for multi-dimensional scoring models, and cannot reduce subjective errors by verifying the difference between models; a closed-loop mechanism of "scoring-feedback-optimization" has not been established, making it difficult to improve scoring accuracy through historical data iteration, especially in talent recruitment scenarios, and unable to form personalized scoring strategies for different job requirements.

[0028] In current technology, there is no complete technical solution that combines "automatic generation of scoring standards based on job competency models", "collaborative verification of multiple scoring models" and "closed-loop self-learning optimization". Therefore, those skilled in the art cannot obtain inspiration from existing technologies on combining the above technical features to solve the problem of marking subjective questions in talent selection.

[0029] To resolve the above, please refer to Figure 1 The embodiments of this application provide an intelligent examination paper grading method for talent selection and recruitment, which is applied to a computer device. The computer device can be deployed on a single server or a server cluster. It can also be deployed on a handheld terminal, a laptop computer, a wearable device, or a robot. It should be noted that all information involved in the method provided in this application is extracted with the authorization of the relevant user and in compliance with relevant regulations, and does not infringe on user privacy.

[0030] The provided intelligent examination paper marking method for talent selection includes steps S101 to S104. The details are as follows: Step S101. Use image processing technology to cut the test paper according to the preset test paper template to obtain multiple sections corresponding to the test paper; use text recognition technology to convert the image content of each section of the cut test paper into an editable text format.

[0031] Specifically, image processing technology is used to achieve structured segmentation of the test paper, and then text recognition technology is used to convert the image information into processable text data, providing standardized input for the subsequent scoring process.

[0032] Test paper template matching and area cutting use computer equipment to pre-store various test paper templates (such as templates specifically for recruitment written tests). The templates contain positioning marks (such as QR codes, corner mark markers), area division rules (question area, answer area, personal information area) and coordinate mapping relationships; use image processing algorithms (such as edge detection and template matching) to pre-process the scanned or photographed test paper images, identify the positioning marks in the templates, and determine the physical coordinate range of each functional area; according to preset area division rules (such as ROI-based interest region extraction), cut the test paper image into independent sections (such as personal information section, subjective question answer section, objective question answer section) to generate standardized image slices.

[0033] Text recognition and format processing improve image clarity by performing pre-processing on the cut images of each block, such as noise reduction (median filtering), tilt correction (Hough transform), and illumination equalization (histogram equalization). The text recognition engine (such as an OCR model based on deep learning) is called to identify the text content in the image block by block, generating a text data stream with coordinate position information. The recognition results are semantically corrected (based on job-specific terminology verification) and formatted (unified punctuation, removal of redundant spaces), and finally structured text (such as JSON format) is output, including the question content, answer content, and corresponding position index.

[0034] Exemplarily, test paper templates for different types of positions are pre-stored (e.g., the "technical R&D position subjective question template" includes a code answering area and an algorithm analysis area; the "management position template" includes a case analysis area and a strategy planning area), and position identification fields (e.g., position ID, capability dimension label) are embedded in the templates; for the scanned candidate test paper image, the position type is automatically identified through a template matching algorithm (e.g., reading the "Position: Product Manager" field in the test paper header), and the corresponding template is called for area cutting: Technical position test paper: Accurately cut the "programming question code area" and "technical solution discussion area", retaining key information such as code format indentation and formula numbering; Management position test paper: Cut the "case analysis answering area" and "team management strategy writing area", and identify paragraph logic markers (such as "first point", "in summary", and other hierarchical keywords).

[0035] During text recognition, targeted error correction is performed against the job-specific terminology database (e.g., for technical positions, the accuracy of recognition of terms such as "microservices" and "algorithm complexity" is verified; for management positions, the names of management tools such as "OKR" and "SWOT analysis" are verified). When outputting structured text, the candidate's personal information (name, job ID) is linked with the answer content to generate answer data with job tags (e.g., {Job ID: G001, Question type: Essay, Knowledge point: Team conflict management, Answer text: "..."}), providing a data foundation for subsequent job competency matching.

[0036] Step S102: Set preliminary scoring criteria based on the question type information and knowledge point information corresponding to the test paper, and use the large language model to optimize the preliminary scoring criteria in multiple scoring dimensions from the perspective of question type, knowledge point, and answer logic.

[0037] Specifically, preliminary scoring criteria are automatically generated based on the test content and job competency model, and multi-dimensional optimization is performed through a large language model to ensure that the scoring rules are accurately matched with talent selection needs.

[0038] The preliminary scoring criteria are generated by using computer equipment to analyze the test paper text structure through natural language processing (NLP) technology, identify the corresponding question type labels (such as "essay question", "case analysis question"), knowledge point labels (such as "human resource planning", "labor law compliance") and answer requirements (such as "clear logic", "integration with examples"); retrieve the competency model database of talent selection positions (pre-stored core capabilities required for the position, such as communication skills, depth of professional knowledge, problem-solving ability, etc.), match the preset scoring rule templates according to the question type and knowledge points (such as the subjective question template includes scoring dimensions such as "argument completeness" and "argument relevance"); based on the weight of knowledge points in job competency (such as core job skills corresponding to higher scores), the complexity of answer logic (such as multi-step argumentation corresponding to tiered scoring) and job capability requirement parameters (such as management positions focus on logical integrity, technical positions focus on knowledge point coverage), automatically generate preliminary scoring criteria (such as a structured rule file in JSON format) that include scoring points, key scoring items, and deduction rules.

[0039] The large language model optimizes the scoring criteria by inputting the preliminary scoring criteria into a pre-trained large language model (such as a domain-specific model based on the Transformer architecture). The model analyzes the relationship between question types and knowledge points (such as "labor law case questions" need to be associated with "compliance" and "risk identification" points), the rationality requirements of the answer logic (such as whether the causal relationship is established), and dynamically adjusts the weight distribution of scoring points (such as core knowledge points account for 60% of the weight, expression fluency accounts for 20%) and keyword extraction rules (such as extracting required keywords from the job competency dictionary). The model is fine-tuned in the field based on the historical high-quality answer database (including high-scoring answer samples for different positions), and the semantic similarity calculation logic is optimized (such as matching the semantic relevance of answers and scoring points based on cosine similarity). Finally, the optimized scoring criteria are generated and a structured file containing the weights of each scoring dimension, keyword sets, and scoring thresholds is output.

[0040] For example, the test paper content is parsed through NLP to extract the ability test points in the questions, such as "a question about designing a risk response plan for a certain project" corresponds to job competency indicators such as "risk identification ability", "resource coordination ability", and "emergency handling ability"; the competency database is matched by retrieving the "project manager position" competency model (pre-defined core competency weights for this position: risk identification 30%, logical planning 25%, communication and expression 20%, and industry knowledge 25%), and automatically generating preliminary scoring standards: scoring points: including "completeness of risk point enumeration" (corresponding to risk identification ability), "feasibility of response measures" (corresponding to resource coordination ability), and "logical hierarchy of solutions" (corresponding to logical planning ability); key scoring items: such as "correctly citing more than 3 project management methodologies (such as PDCA, agile development)" and "clarifying responsible persons and time nodes in measures"; deduction rules: such as "risk points omitting core business risks (such as technology selection risks) will be deducted 10 points" and "measures without implementation details will be deducted 5 points."

[0041] The large language model optimization (combined with historical recruitment data) inputs high-scoring answer sheets for "project manager positions" from the past three years (answer samples marked with "excellent manager characteristics"). The large language model analyzes frequently occurring capability mapping patterns (such as "excellent answers generally include the design of cross-departmental collaboration mechanisms") and dynamically adjusts the weights of scoring points: "cross-departmental collaboration" is promoted from a non-core point to a core scoring item (weight +15%); in response to special position requirements (such as a company requiring project managers to have "cloud computing project experience"), the model automatically adds a verification rule for "whether cloud architecture deployment risks are mentioned" to the scoring criteria, achieving personalized adaptation of "one company, one position, one standard".

[0042] Step S103. Use at least two preset scoring models to perform semantic understanding and logical analysis on the converted test paper text answers from multiple scoring dimensions including semantic accuracy, logical completeness, and knowledge point coverage to provide corresponding scoring information, and mark the advantages and disadvantages in the answers.

[0043] Specifically, at least two large scoring models are used to independently score from different dimensions, and the advantages and disadvantages of the answers are marked in combination with the scoring criteria to form a multi-perspective evaluation result.

[0044] The multi-model scoring dimension division is achieved by presetting a first large scoring model (such as a semantic understanding model based on BERT), which is responsible for evaluating "semantic accuracy" (whether the answer accurately expresses the knowledge points) and "knowledge point coverage" (whether all key test points are covered); presetting a second large scoring model (such as a logical reasoning model based on graph neural networks), which is responsible for evaluating "logical integrity" (whether the argumentation process is coherent) and "argument validity" (whether the case citations support the argument); the two models respectively load the optimized scoring criteria and parse the answer text based on their respective dimensions: the first model extracts the knowledge point keywords in the answer through word vector encoding and attention mechanism, compares them with the core test points in the scoring criteria, and calculates the knowledge point coverage score; the second model identifies the premise-conclusion relationship in the answer through syntactic analysis and logical chain construction, evaluates logical faults or contradictions, and calculates the logical integrity score.

[0045] Scoring information generation and content annotation: Each model calculates the final score based on the preset scoring dimension weights (such as knowledge point coverage accounting for 40% and logical integrity accounting for 30%), and generates scoring information containing the scores of each dimension (such as XML format); the answer text is compared sentence by sentence with the scoring points in the scoring criteria, and a sequence annotation model (such as BiLSTM-CRF) is used to identify key sentences that meet the scoring points (marked as "advantage information", such as "correctly citing XX regulatory provisions required for the position") and content that deviates from the key points (marked as "defect information", such as "not combined with the actual job scenario analysis"); annotation marks are generated at the corresponding positions of the answer text (such as adding comment tags), and specific scoring rules are associated (such as "reason for deduction: missing knowledge points") to form annotated answer text data.

[0046] For example, the first model (knowledge + potential model) evaluates the "knowledge point coverage" based on the basic knowledge point library of the position (such as "corporate strategy" and "financial foundation"), and calculates the fit between the answer and the "management trainee training goals" (such as whether it reflects learning ability and innovative thinking) through semantic vector similarity; the second model (logic + value model) evaluates the "logical integrity" through the logical reasoning engine to analyze the answer structure (such as whether the "argument-evidence-conclusion" is complete), and calls the corporate value dictionary (such as "customer first" and "team win-win") to identify the value matching in the answer (such as whether it emphasizes collaboration rather than individuals).

[0047] If the answer contains "propose solutions based on the management trainee's rotation experience", it will be marked as "Position adaptability advantage: reflects the ability to transform into practice"; if the "team collaboration question" only emphasizes personal contribution, it will be marked as "Position ability deficiency: lack of team awareness, deviation from the management trainee training direction"; the marked result is associated with the "ability shortcoming label" in the competency model and directly output to the candidate's ability assessment report, providing HR with a clear screening basis (such as "recommendation for elimination: mismatch of values").

[0048] Step S104: Obtain the comprehensive score corresponding to each test paper according to the score information generated by each scoring model for each question on each test paper.

[0049] Specifically, a comprehensive score is calculated based on the multi-model scoring results, the score reliability is ensured through a score difference verification mechanism, and finally a standardized scoring report is output.

[0050] The score difference check and comprehensive score calculation are carried out by extracting the scores of the two large scoring models for the same question through computer equipment, and calculating the absolute value of the score difference (for example, model A scores 8 points, model B scores 7 points, and the score difference is 1); the preset score difference threshold is determined by historical scoring data statistics (such as calculating the reasonable score difference range within the 95% confidence interval based on the normal distribution). If the score difference is ≤ the threshold, the weighted average of the two scores is taken (the weight is dynamically allocated according to the training effect of the model in the corresponding dimension) as the score for the question; if the score difference is greater than the threshold, the manual review process is triggered: the system automatically marks the question and the annotated advantages and disadvantages information, and pushes it to the manual marking terminal, and the reviewer confirms the final score based on the job requirements.

[0051] Comprehensive scoring generation and report output summarizes the scores of each question according to the test paper structure (such as question type and module), calculates the total score of the test paper and the detailed score of each dimension (such as knowledge point score and logic score); combines the marked strengths and weaknesses information to generate a structured scoring report (such as PDF format), which includes the candidate's answer highlights, ability shortcomings and improvement suggestions; the scoring results and related data are encrypted and stored in the historical scoring database to provide data support for subsequent scoring standard optimization and model training.

[0052] For example, dynamic weighted comprehensive scoring (adaptive job screening stage) includes: initial screening stage: setting a strict score difference threshold (such as 5 points), giving priority to retaining high-scoring papers with consistent scores from both models (reducing the pressure of manual review), and quickly filtering out candidates who obviously do not meet the basic requirements of the position; final interview stage: lowering the score difference threshold (such as 2 points), triggering more manual reviews (such as for "R&D position coding questions", when the score difference exceeds the threshold, it is pushed to the technical interviewer, and the value of the innovation point is judged in combination with the code implementation details); when calculating the comprehensive score, the weight configuration of the current recruitment stage of the position is automatically loaded (such as the initial screening focuses on knowledge point scores, and the final interview focuses on logic and value scores), to achieve dynamic matching of "screening stage-scoring strategy".

[0053] Recruitment decision-making data is output by generating a report containing a "Job Competency Radar Chart" that visually displays candidate scores in dimensions such as "professional knowledge," "logical thinking," and "value alignment," assisting HR in cross-candidate comparisons. Historical scoring data is categorized and stored by job type and competency dimension (e.g., "High-Frequency Flaws in Technical Positions of the Class of 2024: Insufficient Algorithm Complexity Analysis"), providing data support for optimizing next year's recruitment plan (e.g., increasing the weight of questions related to "algorithm complexity"). Sensitive data processing automatically blurs private information such as candidate names and ID numbers, retaining only job-related competency assessment data, in compliance with relevant regulatory requirements.

[0054] In some embodiments, the test paper is cut according to a preset test paper template using image processing technology to obtain multiple sections corresponding to the test paper, including: determining the specific positions of each question area, answer area and personal information area in the test paper by identifying the positioning marks and boundary features preset in the test paper template, and segmenting the test paper image based on the positions to obtain independent sections corresponding to each area.

[0055] By identifying the positioning marks (such as QR codes, corner points, watermarks) and boundary features (such as area dividing lines, fixed format borders) in the test paper template, the physical coordinate positions of the question area, answer area, and personal information area in the test paper are determined, and the image is segmented based on the coordinates to generate independent sections (such as personal information section and subjective question answer section).

[0056] In the talent recruitment scenario, template customization is achieved by pre-setting exclusive templates for different position exams (e.g., the "Technical Position Programming Question Template" includes code answer box positioning marks, and the "Management Trainee Comprehensive Question Template" includes logical layering boundary lines), with position type identifiers embedded in the templates (e.g., the position ID code in the header). Precise regional segmentation: When identifying personal information areas, sensitive information such as the candidate's name and applied position is extracted through positioning marks (subsequently desensitized) to ensure association with the answer sheet content; when segmenting the subjective question answer area, large text areas in management position case analysis questions are precisely segmented through boundary features (e.g., the words "Answer Area: Start Here") to avoid overlapping of adjacent question content; for special formats such as formulas and code indentations in technical position exams, positioning marks are used to retain the answer structure (e.g., the line number of the code block and the position of the formula number), providing a complete context for subsequent semantic analysis.

[0057] In some embodiments, the text recognition technology is used to convert the image content of each section of the cut test paper into an editable text format, including: performing noise reduction, tilt correction and illumination equalization pre-processing on the cut test paper section image, recognizing the text content in the image line by line and word by word through a text recognition model, generating text data containing coordinate position information, and performing semantic error correction and format standardization on the recognition results.

[0058] By preprocessing the cut images (noise reduction, tilt correction, and lighting equalization), the text is recognized line by line through a text recognition model (such as CRNN and Tesseract) to generate text data with coordinates. Semantic error correction is then performed based on the job-specific terminology library (such as correcting the misidentification of "agile development" as "agile development") and the format is unified (such as removing unnecessary line breaks and standardizing punctuation).

[0059] In the talent recruitment scenario, by pre-processing and adapting handwriting, a handwriting recognition optimization module is added to target the large number of handwritten answer sheets in campus recruitment (such as generating a simulated handwriting data training model through the GAN network), thereby improving the recognition accuracy of cursive and cursive characters; the terminology library targeted error correction includes: technical position test papers: verifying the recognition results of terms such as "microservice architecture" and "algorithm time complexity", and automatically correcting the misidentification of "throughput" as "throughput"; management position test papers: semantic verification of management terms such as "SWOT analysis" and "OKR goals" to ensure that "strategic planning" is not misidentified as "strategic regulation"; format standardization processing: the answer texts of different candidates are unified into plain text format, retaining paragraph level markers (such as "one," "(1)", etc.) to facilitate subsequent logical structure analysis (such as the hierarchical recognition of arguments in management position answers).

[0060] In some embodiments, the preliminary scoring criteria are set according to the question type information and knowledge point information corresponding to the test paper, including: parsing the test paper text structure through natural language processing technology, automatically identifying the question type category, test knowledge point labels and answer requirements corresponding to each question, combining the competency model database of talent selection positions, matching the preset question type scoring rule template, and automatically generating preliminary scoring criteria including scoring points, key scoring items and deduction rules based on knowledge point weights, answer logic complexity and job ability requirement parameters; the question type categories include subjective questions and objective questions.

[0061] Through NLP analysis of the test paper structure, the question type (subjective question / objective question), knowledge point tags (such as "Labor Law", "Organizational Behavior"), and answer requirements (such as "analyze with examples") are identified. The job competency model is called to match the scoring rule template. Based on the knowledge point weights (core job skills correspond to high weights), the complexity of the answer logic, and the job ability requirement parameters (such as technical positions focus on "knowledge point coverage" and management positions focus on "logical integrity"), preliminary scoring standards (including scoring points, scoring items, and deduction rules) are generated.

[0062] In the talent recruitment scenario, the job competency model mapping identifies the knowledge points of "training needs analysis" and "curriculum system construction" in the "employee training plan design" question in the "HR specialist position" exam, and matches them with the "training planning ability" (weighted 40%) and "communication and coordination ability" (weighted 30%) in the competency model. When generating preliminary standards, "whether the training needs research method is complete" is set as the core scoring item (corresponding to "training planning ability"), and "whether the plan clearly defines the collaboration process between departments" is set as the secondary scoring item (corresponding to "communication and coordination ability"). Differentiated processing of question types includes: objective questions (such as multiple-choice questions): focusing on the coverage of knowledge points, setting the "number of core test points hit" as the scoring point; subjective questions (such as essay questions): breaking down into multi-dimensional scoring points such as "argument accuracy", "evidence relevance", and "conclusion completeness" to match the "logical expression ability" and "problem-solving ability" required for the position.

[0063] In some embodiments, the use of a large language model to optimize the preliminary scoring criteria in multiple scoring dimensions from the perspective of question type, knowledge point, and answer logic includes: inputting the preliminary scoring criteria into the large language model, analyzing the relationship between question type and knowledge point, and the rationality requirements of answer logic through the model, adjusting the weight distribution of scoring points, keyword extraction rules, and semantic similarity calculation methods, and fine-tuning the model in the field based on historical high-quality answer data to generate optimized scoring criteria.

[0064] By inputting preliminary scoring criteria into a large language model (such as the GPT-4 domain fine-tuning model), analyzing the relationship between question types and knowledge points (such as "salary design questions" need to be associated with "labor law compliance" and "enterprise cost control" knowledge points), adjusting the weights of scoring points (such as increasing the weight of core knowledge points by 10%), optimizing keyword extraction rules (dynamically extracting required keywords from the job competency dictionary), and combining historical high-quality answer sheets (answers marked with high-scoring features of the job) to fine-tune the model, an optimized standard (including dimension weights and semantic matching thresholds) is generated.

[0065] In the talent recruitment scenario, the personalized needs of enterprises are integrated. For example, when an Internet company recruits for a "back-end development position", the large language model analyzes its historical high-quality answers and finds that the key points related to "code security" are mentioned frequently. It automatically adds "whether SQL injection protection is considered" to the scoring criteria, with a weight set to 20%; for the "management trainee rotation plan design question", the model upgrades "cross-departmental learning path design" from a non-core point to a core scoring item (weight +15%) by analyzing the answers of outstanding management trainees in previous years, reflecting the company's demand for the comprehensive capabilities of management trainees; dynamic weight adjustment: adjust the standards according to the progress of job recruitment, such as reducing the weight of "expression fluency" (focusing on knowledge points) in the initial screening, and increasing the weight of "logical innovation" (focusing on potential assessment) in the final interview stage.

[0066] In some embodiments, the use of at least two preset scoring models to perform semantic understanding and logical analysis on the converted test paper text answers from multiple scoring dimensions including semantic accuracy, logical completeness, and knowledge point coverage to provide corresponding scoring information, including: using the first scoring model to analyze the semantic accuracy and knowledge point coverage of the answers, and using the second scoring model to analyze the logical completeness and argument rationality of the answers, the first scoring model and the second scoring model respectively calculate scores according to the preset scoring dimension weights to generate independent scoring information; wherein, the first scoring model is a natural language processing model based on the Transformer architecture, and the second scoring model is a logical reasoning model based on deep learning.

[0067] The first large scoring model (such as the BERT-based semantic model) evaluates "semantic accuracy" (whether the answer accurately corresponds to the knowledge point) and "knowledge point coverage" (the number of core test points hit), and the second large scoring model (such as the neural network logical reasoning model in the figure) evaluates "logical integrity" (whether the argument chain is coherent) and "argument rationality" (whether the evidence supports the argument). The two models independently calculate scores (according to the preset dimension weights) to generate scoring information.

[0068] In the talent recruitment scenario, the division of labor and customization of job models include: technical positions (such as algorithm engineers): the first model: parse the answers to code questions, and calculate the scores of "algorithm correctness" and "complexity analysis accuracy" (knowledge point coverage); the second model: analyze the logical steps of the algorithm derivation process, and evaluate "whether the assumptions are reasonable" and "whether there are gaps in the derivation process" (logical integrity); management positions (such as regional managers): the first model: identify the accuracy of knowledge points such as "market trend judgment" and "competitive product analysis" in case analysis answers (semantic accuracy); the second model: construct the "problem-strategy-expected effect" logic chain in the answer, and evaluate "whether the strategy targets the root cause of the problem" and "whether the expected effect is quantifiable" (argument rationality); dimension weight differentiation: the weight of "knowledge point coverage" in the technical position test paper is set to 60%, and the weight of "logical integrity" in the management position test paper is set to 50%, matching the core ability requirements of the position.

[0069] In some embodiments, marking the advantage information and disadvantage information in the answer includes: comparing the answer text with the optimized scoring criteria, identifying key sentences in the answer that meet the scoring points through a model as advantage information, identifying content that deviates from the scoring points, logical contradictions or knowledge point errors as disadvantage information, and generating annotation marks at corresponding positions in the answer text.

[0070] By comparing the answer text with the optimized scoring criteria sentence by sentence, a sequence labeling model (such as BiLSTM-CRF) is used to identify key sentences that meet the scoring requirements (marked as "advantage information", such as "correctly citing the XX management theory required for the position"), identify content that deviates from the key points, has logical contradictions (such as conflicts between previous and subsequent arguments), or has knowledge point errors (such as incorrect citation of legal clauses) (marked as "defect information"), and generate annotations (such as adding XML tags) at the corresponding positions in the text.

[0071] In the talent recruitment scenario, visual annotation of job competencies includes: "User Growth Plan Design Question" for "Marketing Position": Advantages: If the answer contains "Designing the conversion path based on the AARRR model in combination with job requirements", it will be marked as "Advantages: Complete application of core methodology"; Defects: If the plan does not mention "target user portrait" (the core point of job competency), it will be marked as "Defects: missing user positioning, not meeting the basic competency requirements of the position"; Prioritize sensitive defects by adding red warning labels to defects involving the core competencies of the position (such as "incorrect citation of accounting standards" in the answer for finance positions), and automatically push them to HR as a key screening basis, thereby improving resume screening efficiency.

[0072] In some embodiments, the comprehensive score corresponding to the test paper is obtained based on the scoring information of each scoring model for each question on each test paper, including: calculating the difference in scores of the two scoring models for the same question, if the difference is less than or equal to a preset threshold determined based on historical scoring data statistics, taking the average of the two scores as the score for the question; if the difference exceeds the preset threshold, the manual review process is triggered, and the final score is determined manually by combining the marked advantage information and defect information.

[0073] Calculate the score difference between the two models for the same question. If the score difference is ≤ the preset threshold (determined through historical data statistics, such as mean ± 2σ), take the average score; if the score difference is greater than the threshold, trigger the manual review process, and manually determine the final score based on the marked advantages and disadvantages information to ensure the reliability of the scoring.

[0074] In the talent recruitment scenario, dynamic adjustment of thresholds in stages includes: large-scale initial screening (such as online application for campus recruitment): preset a higher threshold (such as 8 points), quickly pass the test papers with high scores and small score differences in the two models, and filter out the test papers with large or low score differences (reducing the amount of manual review by 90%); precise final interview screening: lower the threshold to 3 points, and trigger manual review for "technical position algorithm questions" and "management position strategy questions" with score differences exceeding the standard. Technical experts / senior managers will confirm the final score based on the marked code logic loopholes or strategy feasibility defects; standardization of the review process: the manual review interface automatically presents the test paper text, the difference points of the two-model scores, and the advantages and disadvantages annotations (such as "Model A believes the logic is complete, and Model B points out that the arguments are insufficient"), assisting reviewers to quickly locate controversial points and improve decision-making efficiency by more than 50%.

[0075] In some embodiments, the method further includes: establishing a historical scoring database, associating and storing the test paper text answers, scoring information, manual review results and scoring standard versions generated for each grading, analyzing historical data through machine learning algorithms, identifying unreasonable scoring points and model parameter deviations in the scoring standards, and automatically iteratively optimizing the scoring standards and weight parameters of the large scoring model.

[0076] Establish a historical scoring database to store answer texts, scoring information, manual review results, and scoring standard versions. Analyze data through machine learning (such as random forests and reinforcement learning) to identify unreasonable points in the scoring standards (such as a knowledge point with too high a weight but low actual discrimination) and model parameter deviations (such as a model that gives a low score to "innovative answers"), and automatically iterate and optimize the standards and model weights.

[0077] In the talent recruitment scenario, closed-loop optimization of job data targeted historical data for the "Customer Service Supervisor Position" and found that the score difference between the two models for the "Customer Complaint Response Process Design Question" continued to exceed the threshold. Analysis found that the "response time quantitative indicator" in the scoring criteria was not clear, resulting in inconsistent model evaluations. The system automatically supplemented this point and adjusted the weight. Analysis of historical high-quality answers to "R&D Position Code Questions" revealed that "exception handling mechanism"-related content appeared frequently, while the weight of this point in the original standard was only 5%. Through reinforcement learning, the weight was increased to 15%, enhancing the assessment of core job capabilities. Compliance data storage: The database encrypts candidate information, sets access permissions according to relevant regulations, and retains only desensitized answer data and capability assessment labels for model training and standard optimization to ensure data security.

[0078] In some embodiments, by adopting the improved YOLOv8 multimodal target detection model, after inputting the test paper image, visual features (positioning marks, borders, text blocks) and semantic features (job keywords and question type identifiers pre-extracted by OCR) are simultaneously identified to construct a "visual-semantic joint positioning model" to achieve precise cutting of irregular layout test papers (such as handwritten additional questions and cross-page answer areas).

[0079] Model training includes: collecting historical talent recruitment test papers (including complex layouts such as code sketches for technical positions and mind map answer areas for management positions), and marking more than 20 position-specific answer area categories such as "code blocks", "case analysis boxes", and "formula areas"; during training, image pixel data + OCR pre-extracted text embedding vectors are input (such as the text "Position: Architect" is converted into a semantic vector), and the model associates visual bounding boxes with semantic labels through a cross-modal attention mechanism (such as the area near the text "Algorithm Derivation" is preferentially identified as the "Logical Deduction Answer Area").

[0080] In the talent recruitment scenario, for hand-drawn algorithm flowcharts in technical position examinations, the model recognizes the "flowchart border + 'time complexity' keyword" and automatically cuts them into independent "algorithm visualization answer sections", retaining details such as arrows and node numbers; when processing open-ended examination questions for management trainees, it recognizes the "mind map answer area" (through node symbols and connecting line visual features) and divides it into multi-level sections of "core argument-sub-argument-case", providing structured data for subsequent logical analysis.

[0081] In some embodiments, by constructing a "handwriting style transfer GAN", the candidate's personalized handwritten answer sheet image is converted into a standard printed style, combined with adaptive threshold super-resolution reconstruction (ESRGAN) to solve the recognition problems of cursive and low-contrast handwriting, and output a standardized image for OCR model processing.

[0082] Data augmentation and model building include: collecting multiple samples of handwritten answers for school recruitment, classifying them into two categories according to the clarity of the handwriting: "high quality" and "fuzzy", training the GAN generator to convert the fuzzy handwriting into clear print, and the discriminator to distinguish between real print and generated images; introducing position-specific constraints: for example, the technical position answer sheet focuses on retaining formatting features such as code indentation and formula subscripts, and the management position retains the hierarchical mark integrity of paragraph numbers ("I," "(1)").

[0083] In talent recruitment scenarios, for candidate answers with "sloppy handwriting but high-quality content" (such as R&D coding questions), the handwriting is first standardized through GAN and then recognized by OCR, which increases the recognition accuracy from 65% to 92%; for candidate answers with "mixed traditional and simplified Chinese characters", conditional GAN ​​automatically converts them into a unified font format to avoid semantic misjudgments due to font differences (such as confusion in the recognition of "strategy" and "strategy").

[0084] In some embodiments, by constructing a "job competency knowledge graph" containing 2,000+ capability nodes (such as "data modeling ability" and "cross-cultural communication ability") and the relationship between nodes (such as the inference relationship of "Python development" → "data analysis ability"), combined with the test paper text parsing results, a scoring standard containing implicit capability mapping is automatically generated.

[0085] The construction of the knowledge graph includes: underlying data: integrating corporate job descriptions, industry capability models, and historical high-performance employee characteristics, defining the three-level mapping relationship of "knowledge point-capability-position" (such as "SQL statement" → "data processing capability" → "BI engineer position"); inference rules: setting up a rule engine (such as "mastering more than 3 data visualization tools → enhancing the 'data presentation capability' score"), and supporting dynamic expansion of corporate personalized capability labels (such as a company adding a "strategic understanding" capability node).

[0086] In the talent recruitment scenario, when processing the "New Energy Industry Analyst Position" test paper, if the answer mentions "Carbon Tariff Policy Analysis", the knowledge graph automatically associates "Policy Interpretation Ability" and "Industry Trend Prediction Ability", and adds corresponding scoring points (weight 15%) to the scoring criteria; for the "Management Trainee General Ability Test", through graph analysis of the implicit connection between "Team Collaboration Case Description" → "Conflict Resolution Ability" → "Leadership Potential", "Richness of Details in the Collaboration Process" is added to the scoring criteria as an indirect scoring item for leadership potential.

[0087] In some embodiments, by utilizing the Meta-Learning framework, the scoring criteria of historically mature positions (such as "Java development position") are used as meta-knowledge. For emerging positions (such as "AIGC product manager position"), exclusive scoring criteria are quickly fine-tuned and generated through small samples (only 50 historical answer sheets are required), solving the problem of lack of data for emerging positions.

[0088] Meta-learning model training includes: basic model: pre-training on the scoring standard data of 100+ mature positions, learning the general mapping pattern of "knowledge point → ability dimension → scoring rule" (such as "technical solution design questions" generally correspond to the "logical rigor" and "innovative thinking" scoring dimensions); fine-tuning mechanism: input 50 benchmark answer sheets for emerging positions (labeled with core ability labels), and quickly update the model parameters (such as the MAML algorithm) to generate scoring standards suitable for the "AIGC product manager position", and automatically identify exclusive scoring points such as "prompt design ability" and "multimodal content understanding".

[0089] In the talent recruitment scenario, for example, when a company is recruiting for a "Web3.0 operation position", based on the meta-knowledge of "Internet operation position" and combined with 30 industry benchmark answer sheets for fine-tuning, it quickly generates scoring standards that include emerging knowledge points such as "smart contract basics" and "DAO governance understanding", and the time taken is shortened from 2 weeks in the traditional solution to 48 hours; the standard automatically associates "accuracy of blockchain terminology use" and "depth of industry awareness" to avoid evaluation bias caused by omitting key points of emerging technologies due to manual standard setting.

[0090] In some embodiments, by constructing an "argumentation logic graph neural network", the answers to subjective questions are converted into a directed graph structure of "thesis-evidence-conclusion". The logical dependency strength between nodes is calculated through GNN (such as the support of evidence for the thesis, and the consistency of conclusion and argument), and multi-dimensional scoring is performed in combination with the job competency model.

[0091] The construction of the logical graph includes: text preprocessing: using dependency syntax analysis to disassemble the answer sentences, extracting entities and relationships such as "core arguments", "supporting evidence", and "inference relationships" (such as the causal relationship of "measure A → solving problem B"); graph structure definition: nodes are arguments / evidence / conclusions, edges are logical relationships such as "support", "refutation", and "supplement", and the weight is the relationship strength (determined by calculating semantic similarity through the pre-trained language model).

[0092] In the talent recruitment scenario, the scoring of "strategic decision-making questions" for management positions: GNN analyzes the integrity of the logical chain of "market positioning → competitive strategy → resource allocation" in the answer. If the "competitive strategy is not derived based on market positioning", it will be marked as a "logical fault defect" and the "strategic thinking ability" score will be deducted; the scoring of "system architecture design questions" for technical positions: the dependency relationship of "demand analysis → module division → performance optimization" is identified through the graph structure. If the "performance optimization node is not connected to the core demand node", it will be judged that "the solution is not feasible enough" and the scoring weight of "engineering practice ability" will be reduced.

[0093] In some embodiments, by designing a "scoring strategy reinforcement learning agent", the weight distribution of multiple scoring models (such as semantic models, logical models, and value models) is dynamically adjusted according to the recruitment stage (initial screening / re-screening / final interview), job scarcity, and the company's real-time employment needs, thereby achieving intelligent scoring of "thousands of people, thousands of positions, and thousands of strategies."

[0094] The state and action definitions include: state space: including 10+ parameters such as recruitment stage (3 types), job level (junior / senior), resume library size, historical pass rate, etc.; action space: dimensional weights of each scoring model (such as the "knowledge point coverage" weight is dynamically adjustable from 50% to 80%); reward function: with "screening efficiency (pass volume / time consumption) + evaluation accuracy (consistency with the final interview results)" as the optimization goals, the intelligent agent is trained through historical recruitment data.

[0095] In the talent recruitment scenario, large-scale initial screening of campus recruitment (status: resume library of 100,000+, junior positions): the intelligent agent automatically increases the weight of the "knowledge point coverage" model to 70%, quickly filters out resumes that obviously do not meet the basic requirements, and improves screening efficiency by 40%; final interview scoring of scarce positions (such as "quantum computing engineer"): reduce the weight of "knowledge point coverage" to 30%, and increase the weight of the "innovative thinking" and "potential assessment" models (a total of 60%) to avoid missing out on high-potential candidates due to insufficient existing knowledge reserves; if the pass rate of a batch of resumes is lower than 5% (triggering a status warning), the intelligent agent automatically fine-tunes the model weight, adds the "implicit ability deduction" dimension, and explores potential matching points in the resumes.

[0096] In some implementations, a "recruitment industry federated learning platform" allows companies to jointly update scoring models and criteria without sharing raw candidate data. Each company's local model uploads gradient parameters, which are aggregated and distributed to a central server. This improves the accuracy of scoring for common industry competencies (such as "communication skills") while complying with data privacy regulations.

[0097] The federated learning architecture includes: Participants: Each enterprise acts as a federated node, locally storing desensitized answer sheet features (such as "logical structure vector" and "knowledge point label"), removing identification information such as name and resume number; Training process: The central server regularly sends the initial model, the enterprise node trains on local data, encrypts and uploads the model parameter differences, and generates a global optimization model after server aggregation (such as general rules for improving the "teamwork ability" score).

[0098] In the talent recruitment scenario, after participating in federated learning, small and medium-sized enterprises can obtain a "leadership potential assessment model" with the same accuracy as that of leading companies in the industry without having to build their own large-scale data sets, thus solving the problem of insufficient data for small and medium-sized enterprises. After aggregating data across enterprises, it was found that "the frequency of the keyword 'project review' in management trainee resumes is positively correlated with performance after being converted to a regular employee." The "depth of review" scoring point has been added to the industry's overall scoring standards, with a weight of 10%, to improve the industry consistency of the potential assessment of recent graduates.

[0099] In some embodiments, the "job review generation model" fine-tuned by GPT-4 is used to input candidate answer score data, strengths and weaknesses annotation information, and job competency model to automatically generate a personalized evaluation report, including an interpretation of ability strengths (such as "demonstrating a solid algorithm foundation and engineering implementation thinking in technical solution design") and improvement suggestions (such as "cross-departmental collaboration cases need to be supplemented to reflect management potential").

[0100] Model training and customization include: training data: collecting historical manual comments, classifying them by job type (technical / management / functional), and injecting job competency dictionaries (such as technical positions include terms such as "code readability" and "system scalability"); generating strategies by combining the dimension scores output by the scoring model (such as "logical integrity 92 points"), calling the "advantage expression sentences" in the template library (such as "outstanding performance in XX dimension, specifically reflected in..."), and combining advantages and disadvantages annotations to generate concrete descriptions (such as "correctly applying the SWOT model in case analysis to clearly distinguish internal advantages from external opportunities").

[0101] In talent recruitment scenarios, such as batch screening for campus recruitment, an assessment report of approximately 300 words is automatically generated for each candidate, replacing time-consuming manual review writing and improving efficiency by 90%. The report content is deeply related to job requirements: technical position reports focus on "code implementation details" and "algorithm innovations", while management position reports focus on "strategy logic" and "team collaboration considerations". HR can directly use the report as a basis for interview questions (such as designing follow-up questions regarding the "inadequate risk assessment" point in the report).

[0102] See also Figure 2 As shown, Figure 22 is a schematic diagram of the structure of an intelligent marking system 200 for talent selection and recruitment provided in an embodiment of the present application. The intelligent marking system 200 for talent selection and recruitment is used to execute the steps of the intelligent marking method for talent selection and recruitment shown in the above embodiments. The intelligent marking system 200 for talent selection and recruitment can be a single server or a server cluster, or the intelligent marking system 200 for talent selection and recruitment can be a terminal, such as a handheld terminal, a laptop computer, a wearable device, or a robot.

[0103] like Figure 2 As shown, the intelligent marking system 200 for talent selection includes: The image cutting unit 201 is used to cut the test paper according to the preset test paper template using image processing technology to obtain multiple sections corresponding to the test paper; and convert the image content of each section of the cut test paper into an editable text format using text recognition technology; The standard setting unit 202 is used to set preliminary scoring standards based on the question type information and knowledge point information corresponding to the test paper, and optimize the preliminary scoring standards in multiple scoring dimensions based on question type, knowledge point, and answer logic using a large language model; The scoring acquisition unit 203 is configured to use at least two preset scoring models to perform semantic understanding and logical analysis on the converted test paper text answers based on multiple scoring dimensions, including semantic accuracy, logical completeness, and knowledge point coverage, to provide corresponding scoring information, and to mark the advantages and disadvantages of the answers; The scoring completion unit 204 is used to obtain the comprehensive score corresponding to the test paper based on the scoring information generated by each scoring model for each question on each test paper.

[0104] It should be noted that technical personnel in the relevant field can clearly understand that for the convenience and conciseness of description, the specific working processes of the intelligent marking system for talent selection and recruitment and each module described above can refer to the corresponding contents in the various embodiments of the intelligent marking method for talent selection and recruitment, and will not be repeated here.

[0105] The above-mentioned intelligent marking method for talent selection can be implemented in the form of a computer program. Figure 2 Run on the device shown.

[0106] See also Figure 3 , Figure 3 1 is a schematic block diagram of the structure of a computer device provided in an embodiment of the present application. The computer device includes a processor, a memory, and a network interface connected via a device bus, wherein the memory may include a storage medium and an internal memory.

[0107] The storage medium can store an operating device and a computer program. The computer program includes program instructions, which, when executed, can cause a processor to execute any one of the intelligent examination paper marking methods for talent selection.

[0108] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.

[0109] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any intelligent marking method for talent selection.

[0110] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 3 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the terminal to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0111] It should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0112] In one embodiment, the processor is configured to execute a computer program stored in the memory to implement the following steps: Image processing technology is used to cut the test paper according to the preset test paper template to obtain multiple sections corresponding to the test paper; text recognition technology is used to convert the image content of each section of the cut test paper into an editable text format; Set preliminary scoring criteria based on the question type and knowledge point information corresponding to the test paper, and use the large language model to optimize the preliminary scoring criteria in multiple scoring dimensions based on question type, knowledge points, and answer logic; Use at least two pre-set scoring models to perform semantic understanding and logical analysis on the converted test paper text answers from multiple scoring dimensions, including semantic accuracy, logical completeness, and knowledge point coverage, to provide corresponding scoring information and mark the advantages and disadvantages of the answers; The comprehensive score corresponding to each test paper is obtained based on the scoring information of each scoring model for each question on each test paper.

[0113] In some embodiments, the test paper is cut according to a preset test paper template using image processing technology to obtain multiple sections corresponding to the test paper, including: determining the specific positions of each question area, answer area and personal information area in the test paper by identifying the positioning marks and boundary features preset in the test paper template, and segmenting the test paper image based on the positions to obtain independent sections corresponding to each area.

[0114] In some embodiments, the text recognition technology is used to convert the image content of each section of the cut test paper into an editable text format, including: performing noise reduction, tilt correction and illumination equalization pre-processing on the cut test paper section image, recognizing the text content in the image line by line and word by word through a text recognition model, generating text data containing coordinate position information, and performing semantic error correction and format standardization on the recognition results.

[0115] In some embodiments, the preliminary scoring criteria are set according to the question type information and knowledge point information corresponding to the test paper, including: parsing the test paper text structure through natural language processing technology, automatically identifying the question type category, test knowledge point labels and answer requirements corresponding to each question, combining the competency model database of talent selection positions, matching the preset question type scoring rule template, and automatically generating preliminary scoring criteria including scoring points, key scoring items and deduction rules based on knowledge point weights, answer logic complexity and job ability requirement parameters; the question type categories include subjective questions and objective questions.

[0116] In some embodiments, the use of a large language model to optimize the preliminary scoring criteria in multiple scoring dimensions from the perspective of question type, knowledge point, and answer logic includes: inputting the preliminary scoring criteria into the large language model, analyzing the relationship between question type and knowledge point, and the rationality requirements of answer logic through the model, adjusting the weight distribution of scoring points, keyword extraction rules, and semantic similarity calculation methods, and fine-tuning the model in the field based on historical high-quality answer data to generate optimized scoring criteria.

[0117] In some embodiments, the use of at least two preset scoring models to perform semantic understanding and logical analysis on the converted test paper text answers from multiple scoring dimensions including semantic accuracy, logical completeness, and knowledge point coverage to provide corresponding scoring information, including: using the first scoring model to analyze the semantic accuracy and knowledge point coverage of the answers, and using the second scoring model to analyze the logical completeness and argument rationality of the answers, the first scoring model and the second scoring model respectively calculate scores according to the preset scoring dimension weights to generate independent scoring information; wherein, the first scoring model is a natural language processing model based on the Transformer architecture, and the second scoring model is a logical reasoning model based on deep learning.

[0118] In some embodiments, marking the advantage information and disadvantage information in the answer includes: comparing the answer text with the optimized scoring criteria, identifying key sentences in the answer that meet the scoring points through a model as advantage information, identifying content that deviates from the scoring points, logical contradictions or knowledge point errors as disadvantage information, and generating annotation marks at corresponding positions in the answer text.

[0119] In some embodiments, the comprehensive score corresponding to the test paper is obtained based on the scoring information of each scoring model for each question on each test paper, including: calculating the difference in scores of the two scoring models for the same question, if the difference is less than or equal to a preset threshold determined based on historical scoring data statistics, taking the average of the two scores as the score for the question; if the difference exceeds the preset threshold, the manual review process is triggered, and the final score is determined manually by combining the marked advantage information and defect information.

[0120] In some embodiments, the method further includes: establishing a historical scoring database, associating and storing the test paper text answers, scoring information, manual review results and scoring standard versions generated for each grading, analyzing historical data through machine learning algorithms, identifying unreasonable scoring points and model parameter deviations in the scoring standards, and automatically iteratively optimizing the scoring standards and weight parameters of the large scoring model.

[0121] The present application also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the processor implements the steps of the intelligent marking method for talent selection and recruitment provided in any embodiment of the present application.

[0122] The computer-readable storage medium may be an internal storage unit of the computer device described in the aforementioned embodiment, such as a hard disk or memory of the computer device. The computer-readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, a SmartMedia Card (SMC), a Secure Digital (SD) card, a flash memory card, etc., equipped on the computer device.

[0123] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present application, and such modifications or substitutions should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. An intelligent marking method for talent selection, characterized in that: include: Using image processing technology to cut the test paper according to the preset test paper template, multiple sections corresponding to the test paper are obtained; Use text recognition technology to convert the image content of each section of the cut test paper into an editable text format; Set preliminary scoring criteria based on the question type and knowledge point information corresponding to the test paper, and use the large language model to optimize the preliminary scoring criteria in multiple scoring dimensions based on question type, knowledge points, and answer logic; Use at least two pre-set scoring models to perform semantic understanding and logical analysis on the converted test paper text answers from multiple scoring dimensions, including semantic accuracy, logical completeness, and knowledge point coverage, to provide corresponding scoring information and mark the advantages and disadvantages of the answers; The comprehensive score corresponding to each test paper is obtained based on the scoring information of each scoring model for each question on each test paper.

2. The method according to claim 1, characterized in that The test paper is cut according to a preset test paper template using image processing technology to obtain multiple sections corresponding to the test paper, including: By identifying the preset positioning marks and boundary features in the test paper template, the specific positions of each question area, answer area and personal information area in the test paper are determined, and the test paper image is segmented based on the said positions to obtain independent sections corresponding to each area.

3. The method according to claim 1, characterized in that The text recognition technology is used to convert the image content of each section of the cut test paper into an editable text format, including: The cut test paper section images are pre-processed with noise reduction, tilt correction and illumination equalization. The text content in the image is recognized line by line and word by word through the text recognition model to generate text data containing coordinate position information. The recognition results are then semantically corrected and formatted.

4. The method according to claim 1, wherein The preliminary scoring criteria are set based on the question type information and knowledge point information corresponding to the test paper, including: By analyzing the test paper text structure through natural language processing technology, the corresponding question type, test knowledge point labels and answer requirements of each question are automatically identified. Combined with the competency model database of the talent recruitment position, the preset question type scoring rule template is matched. Based on the knowledge point weight, answer logic complexity and job ability requirement parameters, the preliminary scoring criteria including scoring points, key scoring items and deduction rules are automatically generated; The question type categories include subjective questions and objective questions.

5. The method according to claim 1, wherein The use of a large language model to optimize the preliminary scoring criteria across multiple scoring dimensions, including question type, knowledge points, and answer logic, includes: The preliminary scoring criteria are input into the large language model. The model analyzes the relationship between question types and knowledge points, and the rationality requirements of the answer logic. The weight distribution of scoring points, keyword extraction rules and semantic similarity calculation methods are adjusted. The model is fine-tuned in the field based on historical high-quality answer data to generate optimized scoring criteria.

6. The method according to claim 1, characterized in that The method uses at least two preset scoring models to perform semantic understanding and logical analysis on the converted test paper text answers from multiple scoring dimensions including semantic accuracy, logical completeness, and knowledge point coverage to provide corresponding scoring information, including: The first scoring model is used to analyze the semantic accuracy and knowledge point coverage of the answers, and the second scoring model is used to analyze the logical integrity and argument rationality of the answers. The first and second scoring models respectively calculate scores based on the preset scoring dimension weights to generate independent scoring information; Among them, the first scoring model is a natural language processing model based on the Transformer architecture, and the second scoring model is a logical reasoning model based on deep learning.

7. The method according to claim 1, characterized in that The advantages and disadvantages of the answers are marked, including: The answer text is compared with the optimized scoring criteria. The model is used to identify key sentences in the answer that meet the scoring points as advantage information, identify content that deviates from the scoring points, has logical contradictions, or has knowledge point errors as defect information, and generate annotation marks at the corresponding positions in the answer text.

8. The method according to claim 1, characterized in that The step of obtaining the comprehensive score corresponding to each test paper according to the score information of each question on each test paper according to each scoring model includes: Calculate the difference in scores between the two scoring models for the same question. If the difference is less than or equal to a preset threshold determined based on historical scoring data statistics, take the average of the two scores as the score for the question; If the score difference exceeds the preset threshold, the manual review process is triggered, and the final score is determined manually by combining the marked advantage information and defect information.

9. The method according to claim 1, characterized in that The method further comprises: Establish a historical scoring database, associate and store the test paper text answers, scoring information, manual review results and scoring standard versions generated for each grading, analyze historical data through machine learning algorithms, identify unreasonable scoring points and model parameter deviations in the scoring standards, and automatically iteratively optimize the scoring standards and weight parameters of the large scoring model.

10. An intelligent examination system for talent selection, characterized in that: include: An image cutting unit is used to cut the test paper according to a preset test paper template using image processing technology to obtain multiple sections corresponding to the test paper; Use text recognition technology to convert the image content of each section of the cut test paper into an editable text format; The standard setting unit is used to set preliminary scoring standards based on the question type and knowledge point information corresponding to the test paper, and use the large language model to optimize the preliminary scoring standards in multiple scoring dimensions based on question type, knowledge point, and answer logic; A scoring acquisition unit is used to use at least two preset scoring models to perform semantic understanding and logical analysis on the converted test paper text answers from multiple scoring dimensions including semantic accuracy, logical completeness, and knowledge point coverage to provide corresponding scoring information and mark the advantages and disadvantages of the answers; The scoring completion unit is used to obtain the comprehensive score corresponding to the test paper based on the scoring information of each question on each test paper given by each scoring model.

Images