An evaluation data processing and stratified report generation method and system

Abstract

The application provides a kind of evaluation data processing and layered report generation method and system, adopt " known layer + core layer + auxiliary layer " three-layer architecture.Publicly known layer includes professional interest evaluation and cognitive function evaluation, respectively extract first and second feature vectors.Core layer is self-adaptive weighted fusion calibration algorithm, dynamically adjusts the weight of two feature vectors, deviation correction and threshold judgment, and outputs the integrated features of interest, cognition, assets and reality constraints.The auxiliary layer converts the integrated features into interpretable and executable layered reports through mechanisms such as deep judgment, layered classification, parallel generation, alignment verification, cropping and translation.The application overcomes the problems of dimension fragmentation, recommendation disconnection from reality constraints and model illusion by using the core fusion algorithm to connect two evaluations, combined with large language model parallel calling and alignment verification, improving the accuracy, executability and credibility of the recommendations.

Description

Technical Field

[0001] This invention relates to a method and system for processing assessment data and generating stratified reports, relating to the interdisciplinary fields of artificial intelligence and online assessment. More specifically, this invention relates to a method and system for fusing and mapping the outputs of a first assessment algorithm (a career interest assessment based on a six-dimensional interest model) and a second assessment algorithm (a cognitive function assessment based on cognitive function preferences) through an adaptive weighted fusion calibration algorithm described in this invention, and combining user-real-world constraints with the parallel scheduling of a large language model (LLM), to output an interpretable, comparable, and executable stratified report. Background Technology

[0002] There are already numerous publicly available technologies in the field of online assessment and career recommendation. During their long-term research and development process, the inventors of this application have identified the following typical existing technical solutions and discovered their respective shortcomings: (Category 1 Existing Technology: Single-Dimensional Interest Assessment) is represented by "Holland's Theory of Vocational Interests" (Holland, JL, Making Vocational Choices, 1973, and the subsequent O*NET Interest Inventory series). This approach categorizes users' interests into six dimensions and matches them with pre-defined career paths based on their scores. Its shortcomings include: using only "interest" as the matching criterion, ignoring users' cognitive function preferences, personal assets, and real-world constraints (education, income, location); recommendations often exceed the user's current education or income level (e.g., recommending "researcher" to a frontline operator), lacking practicality; and two users with the same interests may have vastly different career options, which single-dimensional matching cannot differentiate.

[0003] The second type of existing technology (single-dimensional personality / cognitive function assessment) is represented by the MBTI type index (based on Jung's 1923 published theory of psychological types) or assessments based on Jung's cognitive functions (Ni / Ne / Ti / Te / Fi / Fe / Si / Se), which categorize users into a certain type and recommend careers accordingly. Its shortcomings are: relying solely on "cognitive function preferences" is disconnected from users' actual interests, skills, and real-world constraints; users of the same type have vastly different actual career preferences, leading to a high rate of misjudgment based solely on type recommendations; and it lacks quantifiable matching scores, with outputs mostly being qualitative "type-career list" mappings, making horizontal comparisons difficult.

[0004] Furthermore, with the development of Large Language Models (LLMs), conversational solutions have emerged that directly ask the LLM "What career is suitable for me?" The drawbacks of such solutions are: LLMs suffer from a serious "illusion" problem when generating multi-career content, outputting fictitious careers not in the candidate pool; LLM outputs lack interpretable evidence of the user's actual profile; and LLM calls are time-consuming per instance, with multiple sequential calls severely slowing down the response.

[0005] In summary, the existing technologies generally suffer from the following shortcomings: First, fragmented dimensions—interests, cognition, assets, and constraints are either treated separately or simply pieced together, lacking organic integration and bias correction of multi-source heterogeneous data; Second, "one-size-fits-all" data collection—existing systems typically require users to complete full-scale assessments to obtain high-quality reports, failing to dynamically adjust report depth based on data completeness; Third, flattened output and simplistic action plans—outputs are mostly flat lists sorted by similarity, without dividing into subsets with different decision-making implications, and usually only generate a general plan for a single recommended profession, failing to meet users' needs for comparing and deciding among multiple alternative professions. Existing technologies lack a complete technical solution that can simultaneously achieve "organic integration of multi-source data, dynamic adjustment of depth, hierarchical differentiated recommendations, parallel generation of multiple objects, and output alignment verification." Summary of the Invention

[0006] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for assessment data processing and hierarchical report generation. The core innovation of this invention lies not in any single assessment algorithm (both the first and second assessment algorithms are well-known technologies in the field, and this invention does not claim any rights to them), but rather in using the adaptive weighted fusion calibration algorithm described in this invention as the core correction / fusion module. This module fuses and maps the outputs of the first and second assessment algorithms, and combines user-defined constraints with the parallel scheduling of a large language model to form a complete assessment data processing pipeline with differentiated technical effects. The method of this invention adopts a three-layer architecture: a "well-known layer + a core layer + an auxiliary layer." The well-known layer includes the first and second assessment algorithms; the core layer is the adaptive weighted fusion calibration algorithm; and the auxiliary layer includes mechanisms such as depth determination, hierarchical classification, parallel generation, alignment verification, and depth-based pruning (preferred implementation methods for each step are detailed in the specific implementation methods).

[0007] To achieve the above objectives, this invention provides a method for processing assessment data and generating tiered reports, comprising the following steps: receiving raw assessment data from users, the raw data including a first type of assessment data, a second type of assessment data, and real-world constraints from interactive questionnaires; using a first assessment algorithm to extract dimensions from the first type of assessment data to obtain a first feature vector; using a second assessment algorithm to extract dimensions from the second type of assessment data to obtain a second feature vector; using an adaptive weighted fusion calibration algorithm to fuse and map the first feature vector and the second feature vector, outputting fused features, the adaptive weighted fusion calibration algorithm including at least dynamic weight adjustment, bias correction, and segmented threshold judgment logic; determining the report generation depth based on the data completeness of this assessment; using the fused features and the real-world constraints as input, calling a large language model through parallel scheduling to generate multiple results for at least two candidate objects, the multiple results including at least tiered classification results and object-by-object planning results; applying alignment verification to the multiple results output by the large language model, pruning the integrated results according to the report generation depth, and outputting a tiered report.

[0008] The present invention also provides a method and system for processing assessment data and generating tiered reports, as well as a computer-readable storage medium.

[0009] Compared with the prior art, the present invention has the following beneficial effects: (1) The core fusion calibration algorithm is linked with the evaluation algorithm to solve the problem of dimensional fragmentation. Existing technologies either use interest evaluation alone, or cognitive function evaluation alone, or simply splice the outputs of the two; this invention uses an adaptive weighted fusion calibration algorithm to dynamically adjust the weights, correct deviations and determine segmented thresholds of the outputs of the two evaluation algorithms. The output reflects the unified fusion characteristics of interest, cognition, assets and constraints, which fundamentally solves the problem of organic fusion of multi-source heterogeneous data.

[0010] (2) Deviation correction and segmented threshold improve the operability of recommendations. This invention introduces a deviation correction term and a segmented attenuation coefficient based on real-world constraints in the fusion calibration, so that the recommendation results automatically decay with the user's real-world conditions (education level, job level, income), avoiding recommendations that are out of touch with reality.

[0011] (3) Three-layer hierarchical recommendation and independent generation of multiple objects to meet the needs of comparative decision-making. This invention divides the recommended objects into three categories according to real-world constraints: primary recommendation, advanced career choice, and track switching. For the top K objects, a large language model is independently invoked to generate corresponding development paths and action plans, enabling users to horizontally compare the execution paths of multiple alternative objects.

[0012] (4) Output alignment verification to suppress the illusion of large language models. This invention applies identifier alignment verification to each development path and action plan output by the large language model, and forcibly overwrites or discards fictitious object names in the model to ensure that the objects finally displayed to the user all come from the real candidate pool.

[0013] (5) Robust design of dynamic adjustment of report depth and parallel scheduling, balancing quality and conversion. This invention automatically determines multiple levels of depth based on the completeness of data collection and differentiates the report chapters accordingly, so that users can obtain valuable results without completing the full evaluation; at the same time, it adopts a thread pool to concurrently call large language models and configures a rollback mechanism to ensure that the overall process is not interrupted. Attached Figure Description

[0014] Figure 1 is a schematic diagram of the overall process of the method described in an embodiment of the present invention.

[0015] Figure 2 is a schematic diagram of the three-layer architecture of "public knowledge layer + core layer + auxiliary layer" as described in the embodiment of the present invention.

[0016] Figure 3 is a schematic diagram of the internal flow of the adaptive weighted fusion calibration algorithm in step S3 of the embodiment of the present invention.

[0017] Figure 4 is a schematic diagram of the scheduling of parallel generation of multiple types of results in step S5 of the embodiment of the present invention.

[0018] Figure 5 is a logical diagram of hierarchical classification and parallel generation of multiple objects in step S5 of the embodiment of the present invention.

[0019] Figure 6 is a schematic diagram of the report hierarchical structure according to an embodiment of the present invention.

[0020] Figure 7 is a schematic diagram of the ideal object matching degree analysis sub-process according to an embodiment of the present invention.

[0021] Figure 8 is a schematic diagram of the module structure of the system described in an embodiment of the present invention. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments described are only for explaining the present invention and do not limit the scope of protection of the present invention. The core innovation of the present invention does not lie in a single evaluation algorithm itself, but in using the adaptive weighted fusion calibration algorithm described in the present invention as the core module to organically fuse the outputs of the evaluation algorithms, supplemented by a variety of auxiliary mechanisms.

[0023] As shown in Figures 1 and 2, the method described in this embodiment adopts a three-layer architecture of "public knowledge layer + core layer + auxiliary layer".

[0024] (In a preferred embodiment of the present invention) the first assessment algorithm is a career interest assessment algorithm based on a six-dimensional interest model (e.g., a known version of the Holland RIASEC test), which categorizes the user's response data into six dimensions and outputs the six-dimensional scores as the first feature vector; the second assessment algorithm is a cognitive function assessment algorithm based on cognitive function preferences (e.g., a known version based on Jungian cognitive function Ni / Ne / Ti / Te / Fi / Fe / Si / Se or MBTI types), which categorizes the user's response data into a certain type or a certain dominant function and outputs the type identifier and dominant function identifier as the second feature vector. Both the first and second assessment algorithms are known technologies in the art. The present invention only uses their response data and dimension extraction results, without modifying their internal calculation logic, and does not claim any rights protection for them.

[0025] (Core Layer) The adaptive weighted fusion calibration algorithm described in this invention performs a fusion mapping between the first feature vector and the second feature vector (and optional asset dimension features and real-world constraint feasibility features). As shown in Figure 3, the algorithm includes: Sub-step S31: Dynamic weight adjustment—Dynamically determine the initial weights α, β, γ, and δ based on the stated real-world constraints, and use these initial weights to weight and fuse the matching degrees of each dimension to obtain the initial fused feature M = α·R_match + β·J_match + γ·A_match + δ·F_match; where R_match is the first feature matching degree, J_match is the second feature matching degree, A_match is the asset dimension matching degree, and F_match is the feasibility score of the real-world constraints; Sub-step S32: Deviation correction—Calculate the deviation correction term Δ = f(education gap, job level gap, income gap, regional gap) based on the actual constraints, and apply the deviation correction to the initial fusion feature to obtain M' = M - Δ; when the gap exceeds the preset threshold, Δ increases significantly. Sub-step S33: Segmentation threshold judgment - The bias-corrected feature value M' is divided into segments according to the segmentation threshold (e.g., M' ≥ 70 is high matching segment, 40 ≤ M'<70 is medium matching segment, M'<40 is exploration segment). Different segments correspond to different attenuation coefficients k_i, and the final fused feature M'' = M' × k_i is obtained. Sub-step S34: Output the fused feature M''.

[0026] The specific values ​​of the deviation correction term Δ and the attenuation coefficient k_i, as well as the specific values ​​of the segmentation threshold, are core proprietary parameters that determine the evaluation accuracy. They are preferably protected as trade secrets and are not included in the specific values ​​in this specification.

[0027] This embodiment provides a method for processing assessment data and generating tiered reports. The execution entity is a consulting generation service deployed on the server side (e.g., an API service implemented using Python and FastAPI), and it interacts with the user through a mini-program deployed on the terminal side. The method includes: (Step S1) Data Collection. The terminal-side mini-program collects data in multiple rounds during the user's conversation: the first round collects the first type of assessment data (e.g., response data based on a six-dimensional interest model); the second round collects the second type of assessment data (e.g., response data based on cognitive function preferences); the third round collects personal asset inventory data (organized into two quadrants: "growth assets" and "consumable assets"); the fourth round collects real-world constraints through an interactive questionnaire. These real-world constraints include age, education level, current city, years of work experience, current occupational level, current industry, current annual income, expected income changes, desired city preference, desired occupational level, desired industry direction, preferred work style, preferred work content, work environment to avoid, and the name and reasons for the user's ideal partner.

[0028] (Step S2) Evaluation Algorithm Dimension Extraction. The first evaluation algorithm is used to extract dimensions from the first type of evaluation data to obtain the first feature vector; the second evaluation algorithm is used to extract dimensions from the second type of evaluation data to obtain the second feature vector.

[0029] (Step S3) Core Fusion Calibration. Using the aforementioned adaptive weighted fusion calibration algorithm, the first feature vector and the second feature vector (as well as asset dimension features and real-world constraint feasibility features) are fused and mapped to output fused features.

[0030] (Step S4) Depth Determination. The report generation depth is determined based on the completeness of the data collected in this assessment: if only the quick questionnaire is completed and the complete assessment is not, the depth is determined to be lightweight; if the first and second types of assessments are completed but asset inventory data collection is not completed, the depth is determined to be standard; if the first type of assessment, the second type of assessment, and asset inventory data collection are completed, the depth is determined to be complete. The corresponding rendering markers and chapter sets are then selected according to the stated depth. When the depth is lightweight or standard, the answers to the quick questionnaire and the formal questionnaire are merged at the field level before proceeding to step S2.

[0031] (Step S5) Parallel Generation of Multiple Classes. A thread pool (max_workers preferably 6) is used to schedule large language models in parallel, generating multiple results for at least two candidate objects. As shown in Figure 4, the parallel generation tasks include, but are not limited to: generating narrative analysis, generating a single action plan, generating an ideal job profile, generating development paths for multiple objects (inputting the top K recommended objects and outputting K corresponding paths), generating action plans independently for the top K objects, and generating industry analysis when the depth is complete. Each call is standardized using pre-agreed preset data format constraints, and a local fallback template (rollback mechanism) is configured. When there is a network error or the return does not meet the constraints, the fallback template is used as a substitute.

[0032] Furthermore, as shown in Figure 5, this step also includes hierarchical classification: the candidate object list sorted by fusion features is input into the large language model, and it is instructed to divide the candidate objects into three categories according to "same level of reality constraints": the first category (primary recommendation), which corresponds to the candidate object set with the highest matching degree with the user's current reality constraints and the easiest to enter; the second category (advanced career selection), which corresponds to the candidate object set that requires additional preset resources to enter; and the third category (track switching), which corresponds to the candidate object set that matches the user's interests but crosses the reality constraint level.

[0033] For each of the first type of objects, a preset number of K objects (K is a positive integer, preferably 3) are independently generated using a large language model. Each path includes why_fit, key_skills_gap, stages, branches, and milestones. Alignment validation is performed on each path: the path's identifier field is matched with the identifier field at the corresponding position in the input list. If the match fails, the input identifier is forcibly overwritten with the output identifier, or the path is discarded.

[0034] Similarly, for each of the first category's K objects (based on a preset number), a large language model is independently invoked to generate K corresponding action plans. Each action plan includes short-term phase tasks (days 1-90), medium-term phase tasks (3-12 months, including recommendation certificates and networking strategies), and long-term phase tasks (1-3 years, including North Star, annual milestones, path branch selection, and risk response plans). The content of the first action plan is synchronously written into the top-level field of the report to maintain backward compatibility.

[0035] Furthermore, as shown in Figure 7, this step also includes an ideal object matching degree analysis sub-process: extracting the ideal object text and reason text input by the user; matching the ideal object text with the candidate object set to obtain the closest object and its matching degree score; calling a large language model, combining the fusion features with the closest object, to generate a fit analysis text, advantage matching items, gap analysis items, and at least one alternative object suggestion; and mounting the ideal object matching degree analysis results as an independent column in the hierarchical classification column group.

[0036] Step S6) Report Integration and Trimming. The multiple results are integrated into a hierarchical report according to a pre-defined report chapter structure. The hierarchical report is then trimmed according to the report generation depth: the lightweight version retains only the first category and ideal object columns, as well as the development path and short-term action plan; the standard version retains the first, second, third, and ideal object columns, as well as the development path, short-term, and medium-term action plans; the complete version retains all chapters and adds industry analysis and appendices. The display order of each column in the report is consistent with the order of the objects within that column, and the order of the objects within the column is based on the hierarchical classification results.

[0037] Furthermore, a three-level fallback strategy is adopted for localization translation of the identifier field: the first level searches for the localized result corresponding to the original identifier in the preset mapping table; the second level searches for the localized title in the object description database indexed by the unique code when the first level fails; the third level directly outputs the original identifier when the second level fails or the title is a preset placeholder; and the separator in the hit result is uniformly replaced with a comma.

[0038] As shown, this embodiment provides a method and system for assessment data processing and stratified report generation, including: a data acquisition module 910, used to receive a user's first type of assessment data, second type of assessment data, personal asset inventory data, and real-world constraints from interactive questionnaires; a first assessment module 920, used to extract dimensions from the first type of assessment data using a first assessment algorithm to obtain a first feature vector; a second assessment module 930, used to extract dimensions from the second type of assessment data using a second assessment algorithm to obtain a second feature vector; and a fusion calibration module 940, used for adaptive weighted fusion calibration. The algorithm performs a fusion mapping between the first feature vector and the second feature vector, and outputs a fused feature; a depth determination module 950 is used to determine the report generation depth based on data completeness; a parallel generation module 960 is used to call a large language model in parallel to generate hierarchical classification results and object-by-object planning results; an alignment verification module 970 is used to perform alignment verification on the identifier field output of the large language model; a report integration module 980 is used to integrate the multi-class results into a hierarchical report and prune it according to the report generation depth; and a localization translation module 990 is used to perform localization translation on the identifier field according to a three-level fallback strategy.

[0039] The fusion calibration module 940 also includes a weight dynamic adjustment submodule 941, a deviation correction submodule 942, and a segmented threshold judgment submodule 943.

[0040] This embodiment provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the steps of the method described in Embodiment 1, or enables the operation of the system described in Embodiment 2. The computer-readable storage medium includes, but is not limited to, ROM, RAM, magnetic disk, optical disk, solid-state drive, cloud storage, etc.

[0041] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention. Scope of Protection and Equivalent Description

[0042] General description of equivalent range The various technical features / technical steps / technical modules described in this invention should be understood to include any equivalent substitutions that "achieve substantially the same function and substantially the same effect by substantially the same means"; unless otherwise expressly stated in the specification, equivalent solutions that can be obviously conceived by those skilled in the art based on the disclosure of this invention should be included in the protection scope of this invention. The "scheduling method for calling a large language model" described in step S5 of claim 1 can be equivalent to, but is not limited to:

[0043] Parallel scheduling: thread pool concurrency, coroutine concurrency, asynchronous task queues, etc. Quasi-parallel scheduling: pipelined segmented execution, batch execution Hybrid scheduling: partially parallel + partially serial scheduling, based on user response priority. An extreme serial approach that simply "executes all tasks sequentially" without employing any parallel / asynchronous / pipeline mechanisms does not constitute an equivalent replacement for this invention. The "alignment check" described in step S6 of claim 1 may be equivalent to, but is not limited to:

[0044] Exact matching of identifier fields (preferred in the instruction manual) Fuzzy matching of identifier fields (regular expressions, substrings) Candidate pool whitelist comparison (output object must exist in the pre-defined object pool) Semantic similarity comparison (embedding similarity) Multi-strategy combination verification (whitelist first, then fuzzy matching) The algorithm described in step S3 of claim 1, and its equivalents include, but are not limited to:

[0045] The implementation of dynamically adjusting weights based on real-world constraints can include: rule-based adjustment, machine learning-based adjustment, and online feedback-based adjustment. The input dimensions for deviation correction items may include: education level, job level, income, region, and other equivalent constraints (age, family, health, etc.). The specific value of the segment threshold may vary depending on the product form. **Specific numerical parameters** — including but not limited to:

[0046] The specific values ​​of the initial weights α, β, γ, and δ for each dimension. Private coefficient in the formula for calculating the deviation correction term Δ The specific values ​​of the high / medium / low thresholds for the segmentation (e.g., 70 / 40) and the corresponding attenuation coefficients k_1 / k_2 / k_3. The specific numerical values ​​are not included in this specification. Instead, their **structure and position** are described only in sub-steps S31–S33 of Example 1 using functional placeholders such as "M = α·R_match + β·J_match + γ·A_match + δ·F_match", "Δ = f(…)", and "M' ≥ 70, etc." These parameters are protected separately by the applicant as trade secrets through internal confidentiality systems, employee confidentiality agreements, and code permission levels. They do not constitute a legal defect of "insufficient disclosure" or "lack of support from the specification" in this specification. Those skilled in the art can implement various implementations with the same structure but equivalent parameter substitutions based on the algorithm logic disclosed in the specification, achieving the same algorithm structure and business process as this invention. This includes user responses and system-derived data.

[0047] "Candidates": refers to candidate professions, research directions, and job positions that can be recommended and compared. "Tiered Report": This refers to a report format that presents information in chapters based on report depth (lightweight / standard / complete) or recommendation category (primary / advanced / track switching / ideal target). The goal of "alignment verification" is to ensure that the output object name is traceable in the real candidate pool.

Claims

1. A method and system for processing assessment data and generating stratified reports, characterized in that, Includes the following steps: S1: Receive the user's raw assessment data, which includes a first type of assessment data, a second type of assessment data, and the real-world constraints of the interactive questionnaire collection. S2: The first evaluation algorithm is used to extract dimensions from the first type of evaluation data to obtain a first feature vector; the second evaluation algorithm is used to extract dimensions from the second type of evaluation data to obtain a second feature vector; S3: The first feature vector and the second feature vector are fused and mapped by an adaptive weighted fusion calibration algorithm to output fused features. The adaptive weighted fusion calibration algorithm includes at least dynamic weight adjustment, deviation correction and segmented threshold judgment logic. The dynamic weight adjustment includes dynamically determining the weights according to the real constraints. S4: Determine the report generation depth based on the completeness of the data in this assessment. The report generation depth includes at least two levels. S5: Using the fusion features and the real-world constraints as input, a large language model is invoked through a scheduling method to generate multiple results for at least two candidate objects. The multiple results include at least hierarchical classification results and object-by-object planning results. The scheduling method is preferably parallel scheduling. S6: Apply alignment verification to the multiple results output by the large language model, trim the integrated results according to the report generation depth, and output a hierarchical report.

2. The method according to claim 1, characterized in that, In step S3, the adaptive weighted fusion calibration algorithm is implemented as follows: The initial weights are dynamically determined based on the real-world constraints, and the first feature vector and the second feature vector are weighted and fused using the initial weights to obtain the initial fused features. The deviation correction term is calculated based on the stated real-world constraints, and the deviation correction is applied to the initial fusion features. The bias-corrected eigenvalues ​​are divided into segments based on threshold values, with different attenuation coefficients corresponding to different segments. Output the fusion features.

3. The method according to claim 1, characterized in that, In step S5, the hierarchical classification result is obtained in the following way: The candidate object list sorted according to the fusion features is input into the large language model; The large language model is instructed to divide the candidate objects into at least three categories based on the real-world constraints. The at least three categories include: a first category, which corresponds to the set of candidate objects that best matches the user's current real-world constraints; and a second category, which corresponds to the set of candidate objects that require additional preset resources to enter. The third category corresponds to a set of candidate objects that match the user's interests but cross real-world constraints.

4. The method according to claim 3, characterized in that, For each of the first category of K candidate objects (where K is a positive integer), the large language model is independently invoked to generate K corresponding development paths. The alignment check is then performed on each development path, and the alignment check includes: Match the identifier field of each development path with the identifier field at the corresponding position in the first type of input list; When a match fails, the input identifier is forcibly overwritten with the output identifier, or the development path is discarded.

5. The method according to claim 3, characterized in that, For each of the first category of K candidate objects (where K is a positive integer), the large language model is independently invoked to generate K corresponding action plans. Each action plan includes short-term phase tasks, medium-term phase tasks, long-term phase tasks, recommended resources, path branch selection, and risk response plans. The content of the first action plan will be synchronously written into the top-level field of the report.

6. The method according to claim 3, characterized in that, Step S5 also includes an ideal object matching degree analysis sub-step: Extract the text of the user's ideal partner and reasons entered in the questionnaire; The ideal object text is matched with the candidate object set to obtain the closest object and its matching score; The large language model is invoked, and the fusion features are combined with the closest object to generate a fit analysis, advantage matching items, gap analysis items, and alternative object suggestions. The results of the ideal object matching degree analysis are displayed as an independent column in the hierarchical category column group.

7. The method according to claim 1, characterized in that, In step S4, the report generation depth is determined according to the following rules: If only the quick questionnaire is completed in this assessment, the report generation depth is set to the first level. When the first type of assessment is completed but the asset inventory data collection is not completed, the report generation depth is determined to be the second level. When the first type of assessment, the second type of assessment, and the asset inventory data collection are completed in this assessment, the report generation depth is determined to be the third level. Furthermore, in both the first and second tiers, the answers to the quick questionnaire and the formal questionnaire are merged at the field level before proceeding to step S2.

8. The method according to claim 1, characterized in that, In step S5, the scheduling method is implemented using a thread pool, and the maximum concurrency of the thread pool is greater than or equal to the total number of tasks for the multi-class results; each call to the large language model is standardized by a pre-agreed preset data format constraint; and each call is configured with a fallback mechanism, so that when the call fails or the returned content does not conform to the preset data format constraint, a preset local fallback template is used instead.

9. A system for processing assessment data and generating stratified reports, characterized in that, include: The data acquisition module is used to receive the user's first type of assessment data, second type of assessment data, and the real-world constraints of the interactive questionnaire collection. The first evaluation module is used to extract dimensions from the first type of evaluation data using a first evaluation algorithm to obtain a first feature vector. The second evaluation module is used to extract dimensions from the second type of evaluation data using the second evaluation algorithm to obtain the second feature vector. The fusion calibration module is used to perform fusion mapping on the first feature vector and the second feature vector through an adaptive weighted fusion calibration algorithm and output fusion features. The algorithm includes at least dynamic weight adjustment, deviation correction and segmented threshold judgment logic. The depth determination module is used to determine the report generation depth based on data completeness. The parallel generation module is used to call a large language model through a scheduling method to generate hierarchical classification results and object-by-object planning results, wherein the scheduling method is preferably parallel scheduling. The alignment verification module is used to perform alignment verification on the identifier fields of the output of large language models; The report assembly module is used to integrate the multiple types of results into a hierarchical report and trim it according to the report generation depth.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-8, or implements the operation of the system as described in claim 9.

Images