Multimodal employment service data intelligent processing platform based on large language model

Abstract

The application discloses a multi-modal employment service data intelligent processing platform based on a large language model and particularly relates to the technical field of artificial intelligence and education information, and comprises the following: a semantic deconstruction and representation module, which utilizes a large language model to perform multi-level semantic deconstruction on post recruitment texts and talent training plans, extracts nodes such as tasks, capabilities, knowledge and skills, constructs a semantic graph and generates normalized semantic vectors; an asymmetric gap quantization module, which calculates bidirectional coverage based on the semantic vectors, identifies three types of gaps, namely, insufficient training, excessive training and misaligned training, and performs gap propagation by utilizing an attention mechanism and a hierarchical relationship of the semantic graph to generate a structured analysis report; and a meta-capability path planning and revision module.The application can accurately quantify semantic gaps between professional training and post requirements, drive automatic revision of training plans and improve the matching degree of vocational education and industrial demand.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and educational informatization, and more specifically, to a multimodal employment service data intelligent processing platform based on a large language model. Background Technology

[0002] Currently, vocational education generally faces the problem of a disconnect between curriculum design and industry needs. When adjusting their curricula, vocational schools often rely on manual experience, lacking scientific decision-making support based on big data analysis of industry talent demands. Existing technologies primarily use keyword matching or simple statistical analysis to compare job postings with talent development plans to determine the degree of match between majors and industries. However, this approach can only identify literal identical skill terms and cannot understand the deeper semantic connections behind the text. For example, teamwork skills and cross-departmental collaborative task completion are semantically highly related, but traditional methods treat them as different strings, leading to distorted matching results.

[0003] To address the aforementioned issues, some technologies have attempted to incorporate natural language processing models for text analysis, but significant shortcomings remain. First, they fail to construct structured semantic graphs of capabilities, hindering the hierarchical deconstruction and relational modeling of job tasks, abilities, and knowledge, resulting in a singular analytical dimension. Second, they lack quantitative methods for assessing the asymmetric relationship between professional training capabilities and job requirements, making it difficult to accurately identify which capabilities are over-developed, under-developed, or misaligned. Third, the analysis results fail to effectively drive subsequent revisions to talent development programs; the entire decision-making process still requires substantial manual intervention, leading to inefficiency and a lack of practicality.

[0004] Therefore, how to leverage the semantic understanding capabilities of large language models to deeply deconstruct and accurately align industry job requirements with professional training programs, automatically identify capability gaps, and generate actionable revision suggestions has become a pressing technical problem in the field of dynamic adjustment of vocational education majors. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a multimodal employment service data intelligent processing platform based on a large language model.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The multimodal employment service data intelligent processing platform based on a large language model includes the following modules:

[0008] Semantic deconstruction and representation module: used to perform multi-level semantic deconstruction of job recruitment texts and talent training programs, extract task nodes, ability nodes, and knowledge, skill, and quality nodes, construct job ability semantic graphs and professional ability semantic graphs, and generate normalized semantic vectors for each node;

[0009] The asymmetric gap quantification module is used to calculate the bidirectional coverage of job competency and professional competency based on the semantic vector, identify three types of gaps: insufficient training, excessive training, and misaligned training, and use attention mechanism and hierarchical relationship of semantic graph to propagate the gap and generate a structured gap analysis report.

[0010] The Meta-Competency Path Planning and Revision Module is used to plan a meta-competency learning path from existing competencies to target competencies with the goal of cultivating insufficient competencies. Based on the target position and the module to be revised selected by the user, it automatically generates a revised talent development plan text.

[0011] Specifically, the semantic deconstruction and representation module performs the following:

[0012] Construct first-level prompt words and extract typical work task nodes from the job description text;

[0013] For each work task node, construct a second-level prompt word and extract the required professional skills, general skills and new productivity elements nodes;

[0014] For each capability node, construct third-level prompts and extract the knowledge nodes, skill nodes, and quality nodes that support that capability.

[0015] All extracted nodes and their relationships are stored in a graph database using an attribute graph model, thus completing the semantic graph construction.

[0016] Specifically, the semantic deconstruction and representation module employs a multiple independent sampling and voting mechanism to quantify node reliability:

[0017] For the same capability node, the temperature parameters of the large language model are set to 0.3, 0.6, and 0.9 respectively for three independent calls;

[0018] Count the number of times each knowledge node, skill node, or quality node appears in the three results, and use the number of occurrences divided by three as the confidence level.

[0019] If the confidence level is greater than or equal to 0.67, the node is stored in the graph database and a confidence level attribute is attached.

[0020] If the confidence level of all nodes is less than 0.67, then the vote is abandoned, and the result of the call with a temperature parameter of 0.6 is directly adopted as the final output, and the confidence level of all nodes in that result is marked as 0.5.

[0021] Specifically, the asymmetric gap quantification module identifies under-cultivation and over-cultivation in the following way:

[0022] Calculate the cosine similarity between each job competency node vector and all professional competency node vectors, and take the maximum value as the coverage score;

[0023] A preset semantic similarity threshold is set. If the coverage score is less than the threshold, the job competency node is marked as underdeveloped, and the professional competency node with the highest similarity is recorded as the closest existing competency.

[0024] Calculate the cosine similarity between each professional competency node vector and all job competency node vectors, and take the maximum value as the demand relevance score;

[0025] If the demand relevance score is less than the threshold, the professional competency node is marked as overtrained.

[0026] Specifically, the asymmetric gap quantification module identifies misaligned training through a cross-attention mechanism:

[0027] Stack the vectors of all ability nodes of the job into a first matrix, stack the vectors of all ability nodes of the profession into a second matrix, and calculate the attention weight matrix.

[0028] For each professional competency node, identify the two job competency nodes with the highest attention weight;

[0029] If the professional competency node is not marked as under-trained or over-trained, but the job competency node with the highest attention is not the same node as the closest existing competency node or the semantic distance between the two is greater than a preset second threshold, then it is determined that the professional competency node has been mistrained.

[0030] Specifically, the asymmetric gap quantization module performs gap propagation after completing gap marking:

[0031] Downward propagation: Mark the knowledge nodes, skill nodes, and quality nodes supported by the ability nodes marked as underdeveloped or misplaced as the affected weak elements;

[0032] Upward propagation: Mark typical work task nodes associated with nodes that are marked as having excess capacity as tasks that may be over-supported or obsolete;

[0033] Cross-propagation: Utilizing the hierarchical structure of the professional competence semantic graph, the gap is propagated from the competence node to the graduation requirement node and then to the training objective node. Each propagation layer is multiplied by a preset attenuation factor, and the cumulative influence weight of each training objective node is recorded.

[0034] Specifically, the method by which the meta-capability path planning and revision module plans the meta-capability learning path is as follows:

[0035] Extract all job competency nodes marked as insufficiently developed from the gap analysis report as the target competency set;

[0036] Identify all capability nodes that are not marked as over-cultivated from the professional capability semantic graph and use them as the starting capability set;

[0037] Call the large language model to construct prompt words that include starting ability, target ability, intermediate ability category constraints, path length constraints and reference to national professional teaching standards. The output should be the meta-ability path and difficulty coefficient in JSON format.

[0038] The generated meta-capability paths are stored in the professional capability completion knowledge graph.

[0039] Specifically, the automated revision method performed by the meta-capability path planning and revision module is as follows:

[0040] It provides an interactive interface that displays a scatter plot of professional matching job analysis and a structured catalog of talent training programs, and receives one or more target jobs selected by the user and one or more modules to be revised.

[0041] Extract all meta-competency paths required to be competent for the selected target position from the professional competency completion knowledge graph, and sort them by difficulty level;

[0042] Construct revision prompts, including role settings, relevant parts of the gap analysis report related to the target position, extracted meta-competency path sets, and the original text of the selected modules in the current talent development program;

[0043] Input the revision prompts into the large language model to obtain the revised text blocks;

[0044] Use document processing tools to replace the content of the corresponding modules in the original talent training program document with the revised text blocks, and add an AI revision record containing the revision time and revision description at the end of the document to generate a downloadable revised document.

[0045] The technical effects and advantages of this invention are as follows:

[0046] This invention constructs a multimodal capability semantic deconstruction and representation module, utilizing a large language model to perform multi-level semantic deconstruction of job recruitment texts and talent training programs. A voting mechanism is employed to quantify node reliability, significantly improving the accuracy of extracting elements such as tasks, capabilities, and knowledge from unstructured texts and enhancing the precision of semantic representation. Building upon this foundation, an asymmetric gap quantification module introduces bidirectional coverage calculation, cross-attention mechanisms, and hierarchical propagation of semantic graphs. This module accurately identifies three types of implicit gaps: under-training, over-training, and misaligned training. It overcomes the limitations of traditional keyword matching methods in capturing semantic relationships, providing a scientific and fine-grained quantitative basis for dynamic professional adjustments.

[0047] Furthermore, this invention, through a meta-competency path planning and revision module, aims to cultivate insufficient capabilities. It automatically plans a meta-competency learning path from existing capabilities to target capabilities using a large language model. Based on the user-selected target position and revision module, it automatically generates a revised talent development plan text with accompanying revision instructions. The entire process achieves a closed-loop drive from industry demand analysis to professional plan revision, significantly reducing manual decision-making costs, improving the responsiveness and operability of professional adjustments, and effectively promoting the precise matching of talent development with job requirements. Attached Figure Description

[0048] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] like Figure 1 As shown, the modules of the multimodal employment service data intelligent processing platform based on a large language model are as follows:

[0051] The semantic deconstruction and representation module utilizes a large language model to perform multi-level semantic deconstruction of job recruitment texts and talent development plans, extracting nodes such as tasks, abilities, knowledge, and skills, and constructing a semantic graph. A multi-sample voting mechanism is used to improve node reliability, and an embedding model is invoked to map all nodes into normalized semantic vectors, achieving a structured representation of unstructured data. The specific process is as follows:

[0052] The system initializes by loading a finely tuned base large language model (such as LLaMA-2-13B or its derivatives) with 13 billion (13B) parameters, and a dedicated model for semantic embedding (such as BAAI / bge-large-zh-v1.5) with the output vector dimension set to 1024.

[0053] The system acquires multimodal input data through data interfaces or user uploads. For each professional position to be analyzed (e.g., "Artificial Intelligence Algorithm Engineer"), the system uses a web crawler module to collect no fewer than 200 valid job posting texts for that position from publicly available recruitment websites such as 51job and Zhaopin, constructing a job requirement corpus called Job_Corpus. Simultaneously, the system receives PDF or Word documents of current talent training programs for the target major (e.g., "Artificial Intelligence Technology Application") uploaded by users, converts them to plain text format, and constructs a professional training corpus called Major_Corpus.

[0054] The system processes each text in Job_Corpus Perform multi-level capability semantic deconstruction. Specific sub-steps are as follows:

[0055] The system constructs a first-level prompt, Prompt_Task, with the following content: "You are a vocational education expert. Please extract 3-5 of the most core and typical work tasks from the following job description text. Describe each task in a sentence containing both the action and the object, such as 'Design and optimize machine learning models'. Please only output the task list, one task per line, without any other explanations. Text: [Text to be analyzed]". Enter Prompt_Task and input it into the large language model, obtain the output, and after cleaning, obtain the set of task nodes for this position. .

[0056] For each task The system constructs a second-level prompt, Prompt_Capability, which reads: "Based on the task [task description], please list the requirements for completing this task:"

[0057] 1. Professional skills (2-3, such as "Python programming");

[0058] 2. General skills (1-2, such as "teamwork");

[0059] 3. New quality factors of production (0-2, such as "data thinking").

[0060] Output format: Professional competence: Competence 1, Competence 2, General competence: ..., New quality productivity: ...”. Enter the description, obtain the output and parse it to get the capability nodes associated with the task.

[0061] For each capability node The system constructs a third-level prompt, Prompt_KSQ, with the following content: In order to cultivate the ability [ability description], please list its supporting elements:

[0062] 1. Knowledge (such as "probability theory and mathematical statistics");

[0063] 2. Skills (e.g., "using the PyTorch framework");

[0064] 3. Qualities (such as "continuous learning").

[0065] Output format: 'Knowledge: Knowledge 1, Knowledge 2', 'Skills: ...', 'Qualities: ...'.

[0066] Obtain the output and parse it to get the knowledge, skills, and qualities nodes associated with the ability nodes.

[0067] Furthermore, to quantify the reliability of each extraction node and reduce the randomness of a single output of the large language model, a multi-stage independent sampling and voting mechanism is employed during the above deconstruction process. Specifically: for the same capability node... Construct the same third-level prompt words, but set the generation temperature parameters of the large language model to 0.3, 0.6, and 0.9 respectively, and call them three times independently;

[0068] Each call returns a list of knowledge, skill, and quality nodes. The three returns are then combined and statistically analyzed: for each extracted knowledge node, the number of times it appears in the three results (k) is recorded, and the confidence level of that knowledge node is set to conf = k / 3. If the confidence level is ≥ 0.67 (i.e., it appears at least twice), the node is officially stored in the graph database with a confidence attribute attached; if the confidence level of all nodes is less than 0.67 (i.e., the three results do not contain duplicate nodes), voting is abandoned, and the result of the call with a temperature parameter of 0.6 is directly adopted as the final output, and the confidence level of all nodes in that result is marked as 0.5 (indicating that the result comes from a single inference). This voting mechanism is applied to both the task layer and the capability layer deconstruction.

[0069] All nodes (tasks, abilities, knowledge, skills, qualities) extracted in the above steps, along with their "inclusion" and "support" relationships, are stored in a graph database (such as Neo4j) in the form of an attribute graph model, thereby constructing a complete job competency semantic graph. .

[0070] Repeat the above deconstruction process for Major_Corpus, but replace the first-level prompts with the major's specific talent development plan: "Please extract the following from the talent development plan for this major: 1. Core training objectives (1-2); 2. Graduation requirements (3-5, such as 'possessing...ability'); 3. Core professional courses (5-8). Text: [Text to be analyzed]". Based on the subsequent prompt chain, deconstruct the nodes of abilities, knowledge, and skills covered by the professional training, and construct a semantic graph of professional abilities. .

[0071] Call the semantic embedding model bge-large-zh-v1.5 to analyze the graph. and All node text (such as "machine learning model design", "Python programming skills", etc.) is processed in batches. The processing flow includes:

[0072] The node text is concatenated with the instruction "Generate a vector that represents the semantics of the following text:"; this is input into the embedding model, which outputs a 1024-dimensional floating-point vector; L2 normalization is then performed on this vector. Finally, the system generates a normalized semantic vector for each node. The vector is then stored as an attribute on the corresponding node of the graph, thus completing the structured representation of the multimodal data.

[0073] The asymmetric gap quantification module calculates the bidirectional coverage of job skills and professional abilities based on semantic vectors, identifying three types of gaps: insufficient training, excessive training, and misaligned training. It introduces an attention mechanism and semantic graph hierarchical relationships to propagate the gap downwards to knowledge points and upwards to typical work tasks and training objectives, generating a structured gap analysis report. The specific process is as follows:

[0074] Load from graph database and and the normalized semantic vectors of all its nodes. .

[0075] for Each capability node in and its vector The system is All capability nodes set In the middle, calculation With each Cosine similarity: Take the maximum value among them. As a score indicating the degree of coverage; where For cosine similarity, For job competency semantic graph A certain capability node (such as The normalized semantic vector of ) has a dimension of 1024. For vector dot product operation, For professional competence semantic graph A certain capability node (such as The normalized semantic vector of ) has a dimension of 1024. To achieve the maximum similarity, This is the set of similarities calculated between the job competency node and each of the various professional competency nodes. The subscripts 1, 2, ... are used only to distinguish different professional competency nodes. The system has a preset semantic similarity threshold. (This value can be adjusted experimentally according to the type of major; the optimal range is...) ).like Then the node Mark as insufficient culture and record how to achieve it. of node The closest existing capability.

[0076] for Each capability node in and its vector The system is All capability nodes set In the middle, calculation With each The cosine similarity is used. The maximum value is taken as the demand relevance score. If this maximum value... Then the node Marked as over-cultivation.

[0077] Introducing a cross-attention mechanism: First, ... The vectors of all capable nodes are stacked into a matrix. (dimension is) ),Will The vectors of all capable nodes are stacked into a matrix. (dimension is) Then, the system calculates the attention weight matrix: ;in Here is the attention weight matrix. For normalized exponential functions, It is a matrix composed of the vectors of all capability nodes in the professional capability semantic graph, with dimension 1. ×1024 ( (Total number of professional capability nodes) This refers to matrix multiplication. A stacked matrix of vectors representing all capability nodes in the job capability semantic graph. (dimension) ×1024, The transpose of the total number of job competency nodes, the transpose dimension is 1024× ;

[0078] The dimension of this matrix is , of which Line 1 Column elements Indicates the professional level The first ability for the position The level of attention given to each ability. Each node in The system found it in The largest Top-K ( The weights corresponding to ) Node set .

[0079] Finally, the system determines: if Unmarked (i.e., it is different from all others) Maximum similarity of nodes (), but its highest attention node The closest existing capability node obtained Not the same node (or the semantic distance is greater than the second threshold) If so, then determine Misaligned training has occurred. This means that although the ability has been trained to the level required by the market (high similarity), the core job competency it is associated with is incorrect (high attention weight points to the wrong ability). For example, a large number of "Java programming" skills have been trained, but the market requires "Python programming" skills, and the semantic vectors of the two are far apart.

[0080] Furthermore, after completing the gap labeling of capability nodes, gap propagation and impact analysis are performed using the hierarchical relationships in the constructed semantic graph. Specifically, this includes:

[0081] Downward propagation—marking the knowledge, skills, and qualities nodes supported by the capability nodes marked as underdeveloped or misplaced as the affected weak elements;

[0082] Upward propagation—marking typical work task nodes associated with nodes that are marked as having excess capacity as tasks that may be over-supported or obsolete;

[0083] Cross-propagation—Utilizing the hierarchical structure of the professional competence semantic graph, the gap is propagated to graduation requirements and training objectives nodes. The initial influence weight of each job-side gap node on its corresponding professional competence node is set to 1.0, and then propagates upwards along the graph: from the competence node to the graduation requirement node is the first layer, and from the graduation requirement node to the training objective node is the second layer. For each upward propagation layer, the current weight is multiplied by a decay factor λ = 0.8. Finally, each training objective node records a cumulative influence weight (i.e., the sum of the weights of all incoming paths); a higher weight indicates a greater impact of the current competence gap on that training objective.

[0084] The system summarizes all marked nodes indicating insufficient, excessive, and misaligned training, and generates a structured JSON file containing a professional-job competency gap analysis report. Each gap record includes: node ID, original text source of the node, path of the node in the whole graph, gap type, and related similarity score or attention weight.

[0085] The meta-competency path planning and revision module aims to cultivate deficient competencies by using a large language model to plan a meta-competency learning path from existing competencies to target competencies. The system provides an interactive interface for users to select the target job and the revision plan module. Based on the gap analysis and path planning results, it automatically generates a revised training plan text and outputs a Word document with revision explanations. The specific process is as follows:

[0086] Load the generated professional-job competency gap analysis report and map , Simultaneously, the system loads and stores the gap propagation results (including affected weak elements, potentially over-supported tasks, and the cumulative impact weight of each training objective) as a reference for determining priorities and supplementary constraints during subsequent path planning. All tasks marked as under-trained are extracted from the report. Nodes form a set of target capabilities. .

[0087] for Each target capability node in The system executes the following sub-steps:

[0088] from In the process, identify all ability nodes that are not marked as over-cultivated, and use them as the starting ability set. .

[0089] The large language model (the same 13B model as in Module 1) is used as the inference engine. A detailed path planning prompt word, `Prompt_Path`, is constructed in the following format:

[0090] You are a vocational education curriculum expert. Task: Design a learning path for a student who already possesses the [starting-point abilities], enabling them to acquire the [target abilities].

[0091] [Starting Capabilities]: {A list of capability texts in Start_Set};

[0092] [Target Capability]: {Textual description of N_target};

[0093] [Available intermediate competency library]: Please select from the following competency categories: basic theoretical competency, tool application competency, specific task competency, comprehensive project competency, metacognitive competency.

[0094] [Constraints]: The path length shall not exceed 4 steps, and each step shall be a teachable and assessable meta-capability.

[0095] [National Professional Teaching Standards Reference]: {Retrieve course standard text fragments related to N_target from the database};

[0096] Output a path in the following JSON format:

[0097] {"path": ["Meta-ability 1", "Meta-ability 2", "Meta-ability 3"]", "difficulty": "Intermediate"}; Note: "difficulty" is divided into beginner, intermediate, and advanced levels. Output only JSON, do not output any other interpretations.

[0098] Input Prompt_Path into the large language model, parse its output JSON, and obtain a meta-capability path. and its difficulty level Treat this path as an edge. ,connect The optimal starting point (chosen by internal model inference) and Store a new professional skill to complete the knowledge graph. middle.

[0099] Complete the task After path planning for all nodes, integrate all to form a complete .

[0100] The front-end page displays an interactive interface. The first area shows a scatter plot of professional matching job postings, with the horizontal axis representing job demand popularity and the vertical axis representing the current matching degree. Users can click or select from various options... Select one or more target positions in the first area. The second area displays the structured directory of the current talent training program (such as "I. Training Objectives", "II. Training Specifications", "III. Core Courses", "IV. Practical Teaching Components"). Users can select one or more modules that need to be revised (for example, only "III. Core Courses").

[0101] Automated revisions are performed based on the user-selected target roles and modules to be revised:

[0102] from Extract all the meta-capability paths that need to be supplemented in order to be competent for the selected target position, and sort them by difficulty level.

[0103] Construct the revision prompt word Prompt_Revise. This prompt word consists of three parts:

[0104] Character description: "You are a professional manager with 20 years of experience";

[0105] Upstream analysis results: the portion of the generated gap analysis report relevant to the selected target position;

[0106] Planned path: i.e., the set of extracted meta-capability paths;

[0107] Original text: The complete text of the revision module selected by the user in the current talent development program; Instruction: "Please revise the original text based on the capability gap in the [Upstream Analysis Results] and the meta-capabilities in the [Planning Path]."

[0108] Revisions should specifically add or replace course / practice projects / competency descriptions. At the end of the revision section, begin with "[Revision Notes]", explaining point by point which specific competency gap your changes fill or which meta-competency path they follow.

[0109] Input Prompt_Revise into the large language model. The model outputs the revised text block.

[0110] After receiving the revised text block, the backend uses the python-docx library to open the original talent development program Word document in memory, locates the user-selected revision module, replaces the original content with the new text block output by the model, and adds an AI revision record table at the end, including the revision time, the gap ID used as the basis, and the revision description. Finally, the modified document stream is pushed to the frontend for the user to download. The downloaded document is an editable Microsoft Word document with complete revision traces and descriptions.

[0111] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, ATA hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state ATA hard disk.

[0112] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0113] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0114] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0115] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0116] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0117] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable ATA hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0118] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A multimodal employment service data intelligent processing platform based on a large language model, characterized in that: Includes the following modules: Semantic deconstruction and representation module: used to perform multi-level semantic deconstruction of job recruitment texts and talent training programs, extract task nodes, ability nodes, and knowledge, skill, and quality nodes, construct job ability semantic graphs and professional ability semantic graphs, and generate normalized semantic vectors for each node; The asymmetric gap quantification module is used to calculate the bidirectional coverage of job competency and professional competency based on the semantic vector, identify three types of gaps: insufficient training, excessive training, and misaligned training, and use attention mechanism and hierarchical relationship of semantic graph to propagate the gap and generate a structured gap analysis report. The Meta-Competency Path Planning and Revision Module is used to plan a meta-competency learning path from existing competencies to target competencies with the goal of cultivating insufficient competencies. Based on the target position and the module to be revised selected by the user, it automatically generates a revised talent development plan text.

2. The multimodal employment service data intelligent processing platform based on a large language model according to claim 1, characterized in that, The semantic deconstruction and representation module is specifically executed as follows: Construct first-level prompt words and extract typical work task nodes from the job description text; For each work task node, construct a second-level prompt word and extract the required professional skills, general skills and new productivity elements nodes; For each capability node, construct third-level prompts and extract the knowledge nodes, skill nodes, and quality nodes that support that capability. All extracted nodes and their relationships are stored in a graph database using an attribute graph model, thus completing the semantic graph construction.

3. The intelligent processing platform for multimodal employment service data based on a large language model according to claim 2, characterized in that, The semantic deconstruction and representation module uses a multiple independent sampling and voting mechanism to quantify node reliability. For the same capability node, the temperature parameters of the large language model are set to 0.3, 0.6, and 0.9 respectively for three independent calls; Count the number of times each knowledge node, skill node, or quality node appears in the three results, and use the number of occurrences divided by three as the confidence level. If the confidence level is greater than or equal to 0.67, the node is stored in the graph database and a confidence level attribute is attached. If the confidence level of all nodes is less than 0.67, then the vote is abandoned, and the result of the call with a temperature parameter of 0.6 is directly adopted as the final output, and the confidence level of all nodes in that result is marked as 0.

5.

4. The intelligent processing platform for multimodal employment service data based on a large language model according to claim 1, characterized in that, The specific method by which the asymmetric gap quantification module identifies under-cultivation and over-cultivation is as follows: Calculate the cosine similarity between each job competency node vector and all professional competency node vectors, and take the maximum value as the coverage score; A preset semantic similarity threshold is set. If the coverage score is less than the threshold, the job competency node is marked as underdeveloped, and the professional competency node with the highest similarity is recorded as the closest existing competency. Calculate the cosine similarity between each professional competency node vector and all job competency node vectors, and take the maximum value as the demand relevance score; If the demand relevance score is less than the threshold, the professional competency node is marked as overtrained.

5. The intelligent processing platform for multimodal employment service data based on a large language model according to claim 4, characterized in that, The asymmetric gap quantification module identifies misaligned training through a cross-attention mechanism. Stack the vectors of all ability nodes of the job into a first matrix, stack the vectors of all ability nodes of the profession into a second matrix, and calculate the attention weight matrix. For each professional competency node, identify the two job competency nodes with the highest attention weight; If the professional competency node is not marked as under-trained or over-trained, but the job competency node with the highest attention is not the same node as the closest existing competency node or the semantic distance between the two is greater than a preset second threshold, then it is determined that the professional competency node has been mistrained.

6. The intelligent processing platform for multimodal employment service data based on a large language model according to claim 5, characterized in that, The asymmetric gap quantization module performs gap propagation after completing gap marking: Downward propagation: Mark the knowledge nodes, skill nodes, and quality nodes supported by the ability nodes marked as underdeveloped or misplaced as the affected weak elements; Upward propagation: Mark typical work task nodes associated with nodes that are marked as having excess capacity as tasks that may be over-supported or obsolete; Cross-propagation: Utilizing the hierarchical structure of the professional competence semantic graph, the gap is propagated from the competence node to the graduation requirement node and then to the training objective node. Each propagation layer is multiplied by a preset attenuation factor, and the cumulative influence weight of each training objective node is recorded.

7. The intelligent processing platform for multimodal employment service data based on a large language model according to claim 1, characterized in that, The specific method by which the meta-capability path planning and revision module plans the meta-capability learning path is as follows: Extract all job competency nodes marked as insufficiently developed from the gap analysis report as the target competency set; Identify all capability nodes that are not marked as over-cultivated from the professional capability semantic graph and use them as the starting capability set; Call the large language model to construct prompt words that include starting ability, target ability, intermediate ability category constraints, path length constraints and reference to national professional teaching standards. The output should be the meta-ability path and difficulty coefficient in JSON format. The generated meta-capability paths are stored in the professional capability completion knowledge graph.

8. The multimodal employment service data intelligent processing platform based on a large language model according to claim 7, characterized in that, The specific method by which the meta-capability path planning and revision module performs automated revision is as follows: It provides an interactive interface that displays a scatter plot of professional matching job analysis and a structured catalog of talent training programs, and receives one or more target jobs selected by the user and one or more modules to be revised. Extract all meta-competency paths required to be competent for the selected target position from the professional competency completion knowledge graph, and sort them by difficulty level; Construct revision prompts, including role settings, relevant parts of the gap analysis report related to the target position, extracted meta-competency path sets, and the original text of the selected modules in the current talent development program; Input the revision prompts into the large language model to obtain the revised text blocks; Use document processing tools to replace the content of the corresponding modules in the original talent training program document with the revised text blocks, and add an AI revision record containing the revision time and revision description at the end of the document to generate a downloadable revised document.

Images