Aesthetic Education Evaluation System and Method Based on Multimodal Evidence Chain and Multi-Agent Collaboration

By constructing a multimodal evidence chain and multi-agent collaborative processing, the problem of lack of continuous correlation and divergence localization of multimodal data in aesthetic education evaluation is solved, thereby achieving the stability and accuracy of the evaluation and reducing duplicate data collection.

CN123118036APending Publication Date: 2026-11-17NANJING COLLEGE OF INFORMATION TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611546672.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-08
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing methods of aesthetic education evaluation lack continuous correlation of multimodal creative data, making it difficult to pinpoint discrepancies between evaluation models and to conduct targeted supplementary evidence and stability assessments when evidence is insufficient.

Method used

By constructing a multimodal evidence chain, aesthetic intentions, creative behavior, and work status are organized into a temporally related evidence chain. These evidence chains are then processed collaboratively by multiple agents to generate evaluation data packages, locate discrepancies, and generate supplementary evidence tasks until a stable evaluation state is reached.

Benefits of technology

It achieves continuous correlation of multimodal creative data, can locate evaluation discrepancies and provide targeted supplementary evidence, ensures the stability and accuracy of evaluation results, and avoids duplicate data collection and unnecessary evaluation coercion.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118036A_ABST
    Figure CN123118036A_ABST
Patent Text Reader

Abstract

This invention discloses an aesthetic education evaluation system and method based on multimodal evidence chains and multi-agent collaboration, relating to the field of AI-assisted education. By acquiring aesthetic education task information and multimodal creative data, the system converts the multimodal creative data into evidence units and organizes them into a multimodal evidence chain containing nodes of aesthetic intent, creative behavior, work status, and expression results. Multiple evaluation agents generate evaluation data packets to locate the smallest evidence sub-chain maintaining evaluation disagreement. Based on the expected reduction in disagreement and execution cost, a target supplementary evidence task is determined, and the supplementary evidence units are integrated into the multimodal evidence chain for re-evaluation. The evaluation stability state is determined based on changes in evaluation scores and remaining evaluation disagreements, and the aesthetic education evaluation result or insufficient evidence information is output. The corresponding system includes an evidence graph library, an evidence view distributor, an agent collaboration cluster, a dispute routing controller, a branch supplementary evidence scheduler, a parallel verification container, and a version release controller.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence-assisted education technology, and in particular to an aesthetic education evaluation system and method based on multimodal evidence chains and multi-agent collaboration. Background Technology

[0002] Art education evaluation typically requires a comprehensive assessment of students' aesthetic concepts, creative process, and final artwork. With the increasing adoption of digital teaching platforms in digital painting, photography, music composition, dance choreography, and integrated arts projects, various data are generated during the teaching process, including text, images, audio, video, and creative operation records. Existing evaluation methods usually convert these data into features and then assign weighted scores, or rely solely on the final artwork and the teacher's evaluation scale to arrive at an evaluation result.

[0003] The above methods have at least the following shortcomings. First, there is a lack of continuous data correlation between theme interpretation, creative behavior, interim works, and the final expression result, making it difficult for evaluation models to determine whether the characteristics of the final work are generated by the student's prior aesthetic intentions and actual creative behavior. Second, when different evaluation models give significantly different evaluation results for the same evaluation dimension, existing methods usually determine the final result by averaging or voting, failing to pinpoint the specific evidence causing the evaluation disagreement. Finally, when evaluation data is missing, conflicting, or has an excessively high proportion of a single modality, existing methods usually directly lower the evaluation confidence level or require the resubmission of all data, failing to generate supplementary tasks for controversial points or determine whether the evaluation results have reached a stable state based on supplementary data. Summary of the Invention

[0004] The purpose of this invention is to provide an aesthetic education evaluation system and method based on multimodal evidence chains and multi-agent collaboration, in order to solve the technical problems in the existing aesthetic education evaluation process, such as the lack of continuous correlation of multimodal creative data, the difficulty in locating the source of discrepancies between evaluation models, and the inability to conduct targeted supplementary evidence and stability judgment when evidence is insufficient.

[0005] To address the aforementioned technical problems, this invention provides an aesthetic education evaluation method based on multimodal evidence chains and multi-agent collaboration, comprising:

[0006] S1. Obtain aesthetic education task information and multimodal creation data, convert the multimodal creation data into evidence units, and organize the evidence units into a multimodal evidence chain including aesthetic intention nodes, creative behavior nodes, work status nodes, and expression result nodes;

[0007] S2. According to the evaluation dimensions in the aesthetic education task information, use multiple evaluation agents to process the multimodal evidence chain to obtain an evaluation data package including evaluation dimensions, evaluation scores, cited evidence nodes, and evidence completeness.

[0008] S3. Compare multiple evaluation data packets, determine the evaluation dimensions with evaluation disagreements, locate the minimum evidence subchain that maintains the evaluation disagreement, and generate dispute information based on the minimum evidence subchain; if there are no evaluation disagreements and no dispute information generated by stability verification, then use the multimodal evidence chain as the updated multimodal evidence chain and proceed to step S5; if there is dispute information generated by stability verification, then proceed to step S4.

[0009] S4. When the disputed information exists, generate candidate supplementary evidence tasks based on the disputed information, determine the target supplementary evidence task based on the expected reduction in disagreement and execution cost of each candidate supplementary evidence task, obtain the supplementary evidence unit formed by executing the target supplementary evidence task, and connect it to the multimodal evidence chain to obtain an updated multimodal evidence chain.

[0010] S5. Using multiple evaluation agents to process the updated multimodal evidence chain, obtain updated evaluation data packets, determine the evaluation stability state based on the changes in evaluation scores before and after supplementary evidence and the remaining evaluation discrepancies between the updated evaluation data packets, and generate dispute information based on the evidence nodes that failed verification or the missing necessary evidence relationships when the state is unstable.

[0011] S6. When the evaluation stability state is reached, output the aesthetic education evaluation result; when the evaluation stability state is not reached and the number of supplementary evidence attempts does not reach the threshold, use the updated multimodal evidence chain and the updated evaluation data package as the new multimodal evidence chain and evaluation data package respectively and return to step S3; when the number of supplementary evidence attempts reaches the threshold, output the evidence insufficiency information.

[0012] Furthermore, step S1 specifically includes:

[0013] Read the evaluation dimension identifier, unified evaluation score range, and the required evidence relationship type corresponding to each evaluation dimension from the aesthetic education task information;

[0014] Obtain the theme explanation, emotional expression goals, creative instructions, operation process, works at each stage, version records and the final work;

[0015] Extract the modal feature vectors corresponding to each data point, and perform hash operations on the original encoded sequences of each data point to obtain the source summary;

[0016] The modal feature vector, the source summary, the collection time, the student identifier, and the work stage are encapsulated into the evidence unit, and the work change range determined by adjacent version records and corresponding operation trajectories is written into the corresponding evidence unit.

[0017] Furthermore, step S1 also includes: converting the evidence unit into evidence nodes of the corresponding category based on the data content and the stage of the work;

[0018] When the collection time of the creative behavior node is between adjacent work status nodes, and the semantic similarity between the creative behavior node and the corresponding work change range reaches the association threshold, a behavior change evidence edge is established.

[0019] Evidence edges for intention and behavior are established based on the semantic similarity between aesthetic intention nodes and creative behavior nodes, and evidence edges for change results are established based on the local feature similarity between work state nodes and expression result nodes.

[0020] The association strength of each evidence edge is determined based on semantic similarity and node time interval, and each evidence node is connected according to the direction of the evidence edge to form the multimodal evidence chain.

[0021] Further, step S2 specifically includes: extracting evidence subchains from the multimodal evidence chain according to the evaluation dimension and evidence edge direction, shielding evidence nodes that do not belong to the current evidence subchain, and obtaining an isolated evidence view;

[0022] The isolated evidence view is respectively input into the aesthetic intention analysis agent, the creative process analysis agent, the work expression analysis agent, and the evidence verification agent;

[0023] The evaluation scores output by each evaluation agent are normalized to the unified evaluation score range;

[0024] The ratio of the number of necessary evidence relationship types belonging to the corresponding evaluation dimension to the total number of necessary evidence relationship types is determined as the evidence completeness.

[0025] When the ratio is zero, the preset lower limit of positive completeness is determined as the completeness of the evidence.

[0026] Furthermore, step S3 specifically includes:

[0027] Extract the evaluation score of the i-th evaluation agent for the k-th evaluation dimension from the valid evaluation data packets of the same evaluation dimension. and completeness of evidence ;according to Determine weights ;

[0028] according to Obtain the weighted average score ;according to Obtaining the degree of divergence Where M is the number of evaluation agents that successfully output evaluation scores for the k-th evaluation dimension and M is not less than 2, i and r are the evaluation agent numbers and k is the evaluation dimension number;

[0029] When the degree of divergence When the divergence threshold is exceeded, the referenced evidence nodes in multiple evaluation data packets and the evidence edges connecting the referenced evidence nodes are combined to form a candidate evidence subchain.

[0030] The evidence edges in the candidate evidence subchain are sequentially masked and the divergence degree is recalculated. Evidence edges whose divergence degree still exceeds the divergence threshold after masking are deleted, and evidence edges whose divergence degree does not exceed the divergence threshold after masking are restored.

[0031] Repeat the traversal until no evidence edge is deleted in a consecutive round. The remaining evidence nodes and evidence edges are determined as the minimum evidence subchain, and the dispute information is generated based on the minimum evidence subchain.

[0032] Furthermore, step S4 specifically includes:

[0033] The following are identified as the corresponding dispute types: missing evidence edges, associated nodes with semantic similarity below the conflict threshold, inconsistent source summary verification, or the proportion of cited evidence nodes in the same data modality reaching the modality threshold.

[0034] Call the supplementary evidence task template corresponding to the dispute type, such as partial reproduction, partial description, source confirmation, or equivalent modality expression, to generate the candidate supplementary evidence task;

[0035] Based on the similarity of evidence subchain features, a weighted average is taken of the non-negative cutoff values ​​of the difference in divergence before and after supplementary evidence in historical records with the same dispute type and supplementary evidence task template, to obtain the expected reduction in divergence for the j-th candidate supplementary evidence task. ;according to Calculate the priority of supplementary evidence ;in, , and These represent the normalized and non-negative execution time, data collection volume, and creative interference volume, respectively, where α, β, γ, and ε are all positive numbers;

[0036] The candidate supplementary evidence task with the highest priority is determined as the target supplementary evidence task, and the corresponding supplementary evidence unit is incrementally connected to the multimodal evidence chain to form the updated multimodal evidence chain.

[0037] Furthermore, step S5 specifically includes:

[0038] The updated multimodal evidence chain is determined as the benchmark evidence chain, and the benchmark evidence chain is processed by multiple evaluation agents to obtain a benchmark evaluation data package;

[0039] Evidence nodes whose source digests are inconsistent, whose evidence completeness is below the verification threshold, or whose evidence nodes are only cited by one evaluation agent are identified as evidence nodes to be verified.

[0040] The nodes of evidence to be verified are sequentially blocked to form a node-blocked evidence chain;

[0041] Based on the local features of the artwork and the time of collection, the same content anchor points are determined for the evidence nodes that represent the same artistic expression object;

[0042] The evidence nodes to be verified are replaced with evidence nodes that have the same stage of work and content anchor point but different data modalities, forming a modality replacement evidence chain;

[0043] The evaluation agents are used to process the node masking evidence chain and the modality replacement evidence chain. The maximum absolute difference between the obtained evaluation score and the corresponding evaluation score in the benchmark evaluation data packet is determined as the evaluation change, and the change of the corresponding evaluation score before and after the supplementary evidence is recorded.

[0044] When at least one node-masking evidence chain or modal replacement evidence chain is formed, and the remaining divergence between the updated evaluation data packets does not exceed the divergence threshold and the evaluation change does not exceed the stability threshold, the evaluation stability state is determined to be reached. When there are no evidence nodes to be verified, the necessary evidence relationships are complete and the source digest verification is consistent, and the remaining divergence does not exceed the divergence threshold, the evaluation stability state is also determined to be reached.

[0045] Furthermore, step S6 specifically includes:

[0046] When the evaluation is stable, the key evidence path is formed by tracing back from the expression result node along the evidence edge whose association strength reaches the path threshold to the aesthetic intention node.

[0047] The aesthetic education evaluation result is output according to the updated evaluation data package, and the number of unconnected evidence nodes, the remaining degree of divergence, and the number of supplementary evidence attempts are encapsulated as evaluation uncertainty information.

[0048] When the evaluation stability state has not been reached and the number of re-certification attempts has reached the threshold, the evaluation interval is determined by the lowest and highest evaluation scores in each evaluation data package after re-certification.

[0049] The insufficient evidence identifier, the evaluation range, and the unresolved dispute location are encapsulated as the insufficient evidence information.

[0050] Another objective of this invention is to provide an aesthetic education evaluation system based on multimodal evidence chains and multi-agent collaboration, comprising:

[0051] An evidence graph library stores the version relationships of evidence nodes, evidence edges, and multimodal evidence chains, and marks active versions.

[0052] An evidence view distributor extracts evidence subchains from the activity version according to the evaluation dimension, and generates isolated evidence views for different evaluation agents to read.

[0053] The intelligent agent collaborative cluster processes the isolated evidence view in parallel and generates an evaluation data packet carrying version identifier, evaluation dimension, evaluation score, evidence completeness and referenced evidence node;

[0054] The dispute routing controller organizes the evaluation data packet into an evaluation score matrix with evaluation agents as rows, evaluation dimensions as columns, and evaluation scores as elements. When there is an evaluation disagreement, it locates the minimum evidence subchain and generates dispute routing information. When there is no evaluation disagreement, it sends the activity version into the parallel verification container.

[0055] A branch supplementary evidence scheduler generates a target supplementary evidence task based on the disputed routing information and incrementally writes the supplementary evidence units into the candidate branch versions in the evidence graph library.

[0056] The parallel verification container uses the candidate branch version or the active version before the verification is completed as the base version to form a node masking version and a modality replacement version, and calls the intelligent agent collaborative cluster to generate a base evaluation data packet, a disturbance evaluation data packet and a stable state identifier.

[0057] A version release controller, when the stability status indicator indicates stability, switches the verified candidate branch version to the active version, or keeps the active version; when the stability status indicator indicates instability and the number of supplementary certifications has not reached the threshold, the version release controller sends the candidate branch version or the active version without supplementary certification and the benchmark evaluation data packet to the dispute routing controller; when the number of supplementary certifications reaches the threshold, insufficient evidence information is generated.

[0058] Furthermore, the evidence map library includes: original evidence partitions for storing evidence units and source summaries in an uncoverable manner;

[0059] The node-edge index partition is used to store evidence nodes, evidence edges, and modal indexes;

[0060] Version genealogy table is used to record the parent-child relationship between active versions and candidate branch versions, as well as newly added evidence nodes and evidence edges;

[0061] The branch supplementary evidence scheduler forms the candidate branch version by referencing existing evidence nodes and adding evidence nodes corresponding to the supplementary evidence units.

[0062] The version release controller verifies that the source digest referenced by the candidate branch version is consistent with the corresponding source digest in the original evidence partition, and verifies that the version identifier in the evaluation data packet is consistent with the version identifier of the candidate branch version; after verification, the active version tag in the version genealogy table is changed in an atomic update manner.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] 1. This invention organizes aesthetic intentions, creative behavior, staged works and expression results into a multimodal evidence chain with temporal and content-related connections, transforming evaluation data from mutually independent multimodal features into an evidence structure with sequential relationships, thus providing a continuous data foundation for evaluation results;

[0065] 2. This invention compares evaluation data packets generated by multiple evaluation agents and locates the smallest evidence subchain that maintains the evaluation disagreement, so that the evaluation disagreement can be mapped to specific evidence nodes and evidence relationships, rather than masking the evaluation disagreement through simple averaging or voting.

[0066] 3. This invention generates candidate supplementary evidence tasks based on the minimum evidence subchain and determines the target supplementary evidence task based on the expected reduction in disagreement and the execution cost. It can obtain supplementary evidence for disputed positions and reduce the repeated collection of data that is irrelevant to the dispute.

[0067] 4. This invention reprocesses and updates the multimodal evidence chain, and determines the stable state of evaluation by combining the changes in evaluation scores before and after supplementary evidence and the remaining evaluation disagreements, so that supplementary evidence, re-evaluation and termination of judgment form a closed loop; when the number of times reaches the threshold and still is not stable, the evidence is insufficient information is output to prevent the forced formation of a single evaluation result under the condition of insufficient evidence. Attached Figure Description

[0068] Figure 1 This is the overall flowchart of the aesthetic education evaluation method based on multimodal evidence chain and multi-agent collaboration of the present invention;

[0069] Figure 2 This is a schematic diagram of the node and evidence edge structure of the multimodal evidence chain in an embodiment of the present invention;

[0070] Figure 3 This is a flowchart illustrating the multi-agent evaluation, dispute localization, and proactive supplementary evidence processing in an embodiment of the present invention.

[0071] Figure 4 This is a structural block diagram of the aesthetic education evaluation system based on multimodal evidence chains and multi-agent collaboration of the present invention;

[0072] Figure 5 This is a schematic diagram of the version hierarchy and parallel verification structure of the evidence graph library in an embodiment of the present invention. Detailed Implementation

[0073] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0074] like Figure 1 As shown, this embodiment provides an aesthetic education evaluation method based on multimodal evidence chains and multi-agent collaboration, including steps S1 to S6. The aesthetic education task can be painting, photography, calligraphy, digital art, music composition, dance choreography, theatrical performance, or a comprehensive art project. Multimodal creative data includes at least two of the following: text, images, audio, video, and creative operation records. Text includes theme explanations, creative instructions, and reflection records; images include sketches, interim works, and final works; audio includes oral explanations or musical works; video includes the painting, dance, or theatrical creation process; and creative operation records include brushstroke trajectories, layer operations, material calls, undo operations, and file version records.

[0075] Step S1 involves acquiring aesthetic education task information and multimodal creative data, converting the multimodal creative data into evidence units, and organizing these evidence units into a multimodal evidence chain that includes nodes for aesthetic intention, creative behavior, work status, and expression results. The aesthetic education task information is used to define the evaluation object, art medium, evaluation dimensions, score range, and necessary evidentiary relationships. An evidence unit is a data object obtained by encapsulating a piece of original data along with its source, time, modality, and stage of creation.

[0076] The aesthetic intention node represents the theme, emotion, narrative relationship, or formal goal that the student wishes to express; the creative behavior node represents material selection, brushstroke manipulation, layer modification, action rehearsal, or audio editing; the work status node represents the stage of the work formed during the creative process; and the expression result node represents the visual, auditory, action, or narrative effect presented in the final work. The nodes are connected in the order of aesthetic intention, creative behavior, work changes, and expression result, thus forming a directional multimodal chain of evidence.

[0077] Step S2: According to the evaluation dimensions in the aesthetic education task information, multiple evaluation agents process the multimodal evidence chain to obtain an evaluation data package. Evaluation dimensions include aesthetic conception, expression transformation, creative adjustment, and work expression. Each evaluation data package includes at least an evaluation dimension identifier, an evaluation score, a cited evidence node identifier, evidence completeness, and a multimodal evidence chain version identifier. Cited evidence nodes record the scope of evidence upon which the evaluation score is based, and evidence completeness records the degree to which this scope of evidence covers the necessary evidentiary relationships.

[0078] Step S3: Compare multiple evaluation data packets to determine the evaluation dimensions with evaluation disagreements, locate the minimum evidence subchain maintaining the evaluation disagreement, and generate dispute information based on the minimum evidence subchain. Evaluation disagreement is the state where the disagreement degree of multiple evaluation agents on the same evaluation dimension exceeds the disagreement threshold. If there is no evaluation disagreement for any evaluation dimension and no dispute information generated from the previous stability check, then proceed directly to the stability check in step S5 using the existing multimodal evidence chain. If there is dispute information generated from the previous stability check, proceed to step S4, and record the change in evaluation score before and after supplementary evidence as zero. When re-entering step S3, compare the updated evaluation data packet obtained in the previous round of step S5. The minimum evidence subchain consists of evidence nodes and evidence edges that can still maintain the evaluation disagreement. Dispute information includes dispute type, disputed evaluation dimension, disputed node position, missing evidence relationship type, and current disagreement degree.

[0079] Step S4: Generate candidate supplementary evidence tasks based on the disputed information, determine the target supplementary evidence task based on the expected reduction in disagreement and execution cost, obtain the supplementary evidence units formed by executing the target supplementary evidence task, and connect them to the multimodal evidence chain to obtain an updated multimodal evidence chain. Candidate supplementary evidence tasks are generated only for the disputed position in the smallest evidence subchain and do not require repeated collection of the complete creation process. Upon completion of the target supplementary evidence task, supplementary evidence units are formed in the form of text, images, audio, video, or operation records.

[0080] Step S5 involves using multiple evaluation agents to process and update the multimodal evidence chain, obtaining an updated evaluation data package. The evaluation stability state is determined based on the changes in evaluation scores before and after supplementary evidence and the remaining evaluation discrepancies between the updated evaluation data packages. The evaluation stability state indicates that the supplementary evidence has brought the evaluation result to a preset stability condition. Changes in evaluation scores before and after supplementary evidence are recorded to characterize the impact of the supplementary evidence; stability conditions include that the remaining discrepancy does not exceed a discrepancy threshold, and the maximum evaluation change corresponding to the available perturbation evidence chain does not exceed a stability threshold. Furthermore, node masking and modal replacement can be used to test the sensitivity of the updated evaluation result to the evidence to be verified. If the verification is unstable, dispute information is generated based on the node of the evidence to be verified that caused the evaluation change to exceed the limit or the missing necessary evidence relationship.

[0081] Step S6: When the evaluation reaches a stable state, output the aesthetic education evaluation result; if the evaluation has not reached a stable state and the number of supplementary evidence attempts has not reached the threshold, update the multimodal evidence chain and update the evaluation data package as the new multimodal evidence chain and evaluation data package, respectively, and return to step S3; when the number of supplementary evidence attempts reaches the threshold, output insufficient evidence information. This loop condition makes the output of each round of supplementary evidence the input for the next round of dispute localization, while limiting the maximum number of loops through the threshold.

[0082] In one embodiment, step S1 further includes the following processing: First, read the evaluation dimension identifier, unified evaluation score range, and the required evidence relationship type corresponding to each evaluation dimension from the aesthetic education task information. The unified evaluation score range can be set to 0-1. The required evidence relationship corresponding to the aesthetic conception dimension includes at least the aesthetic intention node and the intention behavior relationship; the required evidence relationship corresponding to the expression transformation dimension includes at least the creative behavior node, the work state node, and the behavior change relationship; the creative adjustment dimension corresponds to at least two work state nodes and the modification behavior between them; the work expression dimension corresponds to the work state node, the expression result node, and the change result relationship.

[0083] Secondly, the process involves acquiring students' thematic explanations, emotional expression goals, creative instructions, operational processes, interim works, version records, and final works. Modal feature vectors corresponding to each data point are extracted, and the original encoded sequences of each data point are hashed to obtain a source summary. This source summary can be generated using the SHA-256 hash algorithm and is used to verify whether the data source has changed; it does not directly participate in the aesthetic education ability assessment.

[0084] Text data is converted into text feature vectors by a text encoder; image data is converted into visual feature vectors by a visual encoder; audio data is first converted into a Mel spectrogram, and then audio feature vectors are obtained by an audio encoder; video data is processed through temporal sampling and a video encoder to obtain video feature vectors; operation trajectories are encoded into operation sequence features based on operation type, coordinates, duration, and target. The outputs of each encoder can be mapped to a unified 256-dimensional feature vector through a fully connected projection layer.

[0085] Modal feature vectors, source summaries, collection time, student identifiers, and work stages are encapsulated into evidence units. The content differences between adjacent version records are compared, and the work's change range is determined based on the operation trajectory within the corresponding time period. This change range is then written into the corresponding evidence unit. For example, if a new blue area is added between two adjacent work versions, and the operation trajectory indicates that a blue brush operation was performed during the corresponding time period, then this area, this time period, and this brush operation together define the work's change range.

[0086] In one embodiment, such as Figure 2As shown, based on the data content and the stage of the work, the evidence unit is converted into an aesthetic intention node, a creative behavior node, a work status node, or an expression result node. When the collection time of the creative behavior node is between adjacent work status nodes, and the semantic similarity between the creative behavior node and the corresponding work change interval reaches the association threshold, a behavior change evidence edge is established. The semantic similarity can be calculated using cosine similarity, and the association threshold can be determined through a validation set, which is set to 0.65 in one embodiment.

[0087] Evidence edges for intentional behavior are established based on the semantic similarity between aesthetic intention nodes and creative behavior nodes, and evidence edges for change results are established based on the local feature similarity between work state nodes and expression result nodes. For image works, local features may include color distribution, texture, outline, and spatial location; for musical works, local features may include pitch, timbre, rhythm, and dynamics; for dance or theatrical works, local features may include movement speed, movement amplitude, body direction, and spatial location.

[0088] The association strength of evidence edges is determined based on semantic similarity and node time intervals. Specifically, semantic similarity and temporal proximity are first normalized, and then weighted using example weights of 0.7 and 0.3, respectively. The smaller the node time interval, the higher the temporal proximity. Association strength is used to filter key evidence paths and does not replace the evaluation score output by the evaluation agent. Connecting each evidence node according to the evidence edge direction yields a multimodal evidence chain. Evidence nodes that do not meet the connection conditions are still saved and marked as unconnected evidence nodes.

[0089] In one embodiment, such as Figure 3 As shown, evidence subchains are extracted from the multimodal evidence chain according to the evaluation dimension and evidence edge direction, and evidence nodes that do not belong to the current evidence subchain are hidden to obtain an isolated evidence view. The isolated evidence view only contains the nodes and evidence edges required for the current evaluation dimension, and does not include the evaluation scores already output by other evaluation agents.

[0090] The isolated evidence views are input into the aesthetic intent analysis agent, the creative process analysis agent, the work expression analysis agent, and the evidence verification agent, respectively. For the same evaluation dimension, the four evaluation agents provide evaluation scores, cited evidence nodes, and evidence completeness based on their readable isolated evidence views, in order to compare discrepancies within the same dimension. The aesthetic intent analysis agent focuses on theme interpretation, emotional expression goals, creative explanations, and the relationship between intention and behavior; the creative process analysis agent focuses on operation trajectories, version records, the range of changes in the work, and the relationship between behavioral changes; the work expression analysis agent focuses on the final image, audio, video, or motion data; the evidence verification agent checks the source summary, node time sequence, evidence edge direction, and the type of necessary evidence relationship, and outputs evaluation scores and evidence completeness based on the verified evidence relationships. The evaluation scores of each agent are trained or calibrated using a scale for the same evaluation dimension.

[0091] In one model implementation, the text encoder employs a 12-layer Transformer encoding structure, the image encoder uses a visual Transformer or a convolutional neural network, the audio encoder uses a convolutional network with Mel spectrograms as input, the video encoder uses a spatiotemporal convolutional network or a video Transformer, the authoring process parsing agent uses a temporal Transformer, and the evidence verification agent uses a rule verification module and a three-layer graph neural network. The graph neural network distinguishes between intention / behavior evidence edges, behavior change evidence edges, and change result evidence edges during message passing.

[0092] The training data consists of historical art education projects, each including multimodal creative data, manually labeled evidence relationships, teacher evaluation scores, and evaluation dimensions. Student identity information is de-identified before training. The dataset can be divided into training, validation, and test sets at a ratio of 70%, 15%, and 15%, respectively. Mean squared error loss is used for evaluation score prediction, cross-entropy loss for evidence relationship identification, and binary cross-entropy loss for citation evidence node identification. The training loss is formed by weighting these three losses. Creative operations and stage-specific work features labeled as the same range of variation for each work are paired for training, ensuring that the cross-modal features used for association judgment are in a comparable representation space. The optimizer can be AdamW, with an initial learning rate of 0.0001, a batch size of 32, and a maximum training epoch of 100; training stops when the validation set loss does not decrease for 10 consecutive epochs.

[0093] The evaluation scores output by each evaluation agent are normalized to a unified evaluation score range. Evidence completeness is the ratio of the number of existing types belonging to the required evidence relationship type for the corresponding evaluation dimension in the isolated evidence view referenced by the agent to the number of required types; each evaluation dimension is pre-configured with at least one required type, and this ratio is between 0 and 1. When the ratio is zero, a preset lower limit for positive completeness is determined as evidence completeness, for example, 0.05, to ensure that the denominator in subsequent weight calculations remains greater than zero.

[0094] In one embodiment, the evaluation score of the i-th evaluation agent for the k-th evaluation dimension is extracted from the evaluation data packet. and completeness of evidence Weights are calculated based on the completeness of the evidence. The weighted average and degree of disagreement were calculated. The specific calculations are as follows:

[0095]

[0096]

[0097]

[0098] Where M is the number of evaluation agents that successfully output valid scores for the k-th evaluation dimension, and M is not less than 2; i and r are both evaluation agent numbers; k is the evaluation dimension number; For evaluation scores; For the completeness of evidence; For normalized weights; The weighted average of the evaluations of multiple evaluation agents on the k-th evaluation dimension; The degree of divergence is given. Since the completeness of evidence is not less than the lower limit of positive completeness, the denominator of the weights is greater than zero; the sum of all weights is 1; the degree of divergence is a weighted variance and its value is not less than zero.

[0099] When the divergence degree exceeds the divergence threshold, the referenced evidence nodes in each evaluation data packet and the evidence edges connecting them are combined into a candidate evidence subchain. Evidence edges in the candidate evidence subchain are temporarily masked sequentially according to their identification order. For each masked evidence edge, an isolated evidence view is regenerated, multiple evaluation agents are re-invoked, and the divergence degree is recalculated. If the recalculated divergence degree still exceeds the divergence threshold, the evidence edge is deleted; if the divergence degree drops below the divergence threshold, the evidence edge is restored.

[0100] The evidence edge traversal is repeated until no evidence edge is deleted in a consecutive round. The remaining evidence nodes and evidence edges are then determined as the minimum evidence subchain. Here, the minimum evidence subchain means that after the final round of traversal, the evidence subchain that cannot maintain the evaluation divergence by deleting any single remaining evidence edge. Each change to the evidence subchain triggers a re-evaluation, and the divergence degree before the evidence edge was masked is not used.

[0101] In one embodiment, the following are identified as dispute types: evidence edges that should exist but have not been established according to the required evidence relationship type in the minimum evidence subchain; associated nodes with semantic similarity below the conflict threshold; inconsistent source summary verification; or the proportion of cited evidence nodes of the same data modality reaching the modality threshold. In one embodiment, the conflict threshold is set to 0.25, and the modality threshold is set to 0.70.

[0102] Missing evidence edges correspond to local reproduction tasks, semantic conflicts correspond to local explanation tasks, inconsistent source summary verification corresponds to source confirmation tasks, and an excessively high proportion of a single data modality corresponds to equivalent modality representation tasks. The system calls the corresponding supplementary evidence task template to generate candidate supplementary evidence tasks; if the disagreement does not correspond to the above dispute types, a local explanation task is generated for the disputed position of the smallest evidence subchain.

[0103] Historical completion data includes at least the characteristics of the historical minimum evidence subchain, the dispute type, the supplementary evidence task template identifier, the degree of divergence before supplementary evidence, and the degree of divergence after supplementary evidence. The difference in historical divergence is obtained by subtracting the degree of divergence after supplementary evidence from the degree of divergence before supplementary evidence; when this difference is negative, it is truncated to zero. Based on the feature similarity between the historical evidence subchain and the current minimum evidence subchain, a weighted average is calculated for the non-negative divergence differences in historical records with the same dispute type and supplementary evidence task template to obtain the expected reduction in divergence for the j-th candidate supplementary evidence task. If there are no matching historical records or the sum of similarity weights is zero, the preset initial value corresponding to the supplementary evidence task template can be used.

[0104] The priority of supplementary certification for candidate certification tasks is calculated according to the following formula:

[0105]

[0106] in, The priority of the supplementary certification task for the j-th candidate; The expected discrepancy reduction for the j-th candidate supplementary certification task; To normalize the execution time; This represents the normalized data collection volume. α represents the normalized creative interference; α, β, and γ are the positive weights of the three cost terms; ε is a positive number to prevent the denominator from being zero, and can be set to 0.001. Since execution time, data collection volume, and creative interference are all non-negative, the denominator of the formula is greater than zero.

[0107] The candidate task with the highest priority for supplementary evidence is selected as the target task. If two candidate tasks have the same priority, they are selected sequentially based on execution time, amount of creative interference, and amount of data collected, from smallest to largest. The target task is executed to obtain supplementary evidence units, which are then incrementally integrated into the multimodal evidence chain to form an updated multimodal evidence chain.

[0108] In one embodiment, the updated multimodal evidence chain is determined as the baseline evidence chain. Multiple evaluation agents process the baseline evidence chain to obtain a baseline evaluation data package. Evidence nodes with inconsistent source digest verification, referenced by evaluation data packages with evidence completeness below the verification threshold, or referenced by only one evaluation agent are determined as evidence nodes to be verified. Evidence nodes to be verified do not include core nodes that have been consistently referenced by multiple evaluation agents and constitute a continuous evidence path.

[0109] The nodes to be verified are sequentially masked to form a node-masked evidence chain. Based on the local features of the work and the collection time, the same content anchor points are determined for the evidence nodes representing the same artistic expression. The nodes to be verified are replaced with evidence nodes that have the same work stage and content anchor points but different data modalities to form a modality replacement evidence chain. For example, audio description nodes and text description nodes for the same area of ​​the work can have the same content anchor points.

[0110] The node-masked evidence chain and modal replacement evidence chain are processed using the same evaluation agent and model parameters as the baseline evidence chain. The absolute difference between the perturbed evaluation score and the baseline evaluation score is calculated, and the maximum value among the absolute differences for all evaluation dimensions and all evaluation agents is taken as the evaluation change. A stable evaluation state is determined when at least one node-masked evidence chain or modal replacement evidence chain is formed, the remaining divergence between updated evaluation data packets does not exceed the divergence threshold, and the evaluation change does not exceed the stability threshold. If there are no evidence nodes to be verified, and the necessary evidence relationships are complete, the source summaries are consistent, and the remaining divergence does not exceed the divergence threshold, stability is directly determined. If this step is entered without supplementary verification, the same verification is performed using the active version as the baseline evidence chain. If there are evidence nodes to be verified but no usable perturbed evidence chain is formed, or the evaluation change exceeds the stability threshold, dispute information is generated for the corresponding node. If the necessary evidence relationship is missing, dispute information is generated for the missing relationship; stability is not directly determined in either case.

[0111] In one embodiment, when a stable evaluation state is reached, the path is traced backward from the expression result node along the evidence edge whose association strength reaches the path threshold to the aesthetic intention node, forming a key evidence path. The key evidence path should continuously pass through the expression result node, work state node, creative behavior node, and aesthetic intention node, without crossing missing evidence edges. In one embodiment, the path threshold is set to 0.60.

[0112] The aesthetic education evaluation results are output based on the updated evaluation data package. These results include the weighted average of each evaluation dimension, key evidence paths, and evaluation status. The number of disconnected evidence nodes, remaining disagreement, and number of supplementary evidence attempts are encapsulated as evaluation uncertainty information. The remaining disagreement is calculated from the updated evaluation data package using the aforementioned weighted variance calculation method.

[0113] When the evaluation has not reached a stable state and the number of supplementary evidence requests has reached the threshold, a single, fixed score is not output. Instead, the evaluation range is determined by the lowest and highest evaluation scores in each evaluation data package after supplementary evidence requests. The evidence insufficiency flag, the evaluation range, and the location of unresolved disputes are then encapsulated as evidence insufficiency information.

[0114] like Figure 4 As shown, this invention also provides an aesthetic education evaluation system based on multimodal evidence chains and multi-agent collaboration, including a processor 100, a memory 200, a data acquisition module 300, an evidence graph library 310, an evidence view distributor 320, an agent collaboration cluster 330, a dispute routing controller 340, a branch supplementary evidence scheduler 350, a parallel verification container 360, a version release controller 370, and a result output module 400. The processor 100 is communicatively connected to the memory 200 and each of the above modules. The memory 200 stores program instructions and model parameters executed by the processor 100.

[0115] The data acquisition module 300 acquires aesthetic education task information and multimodal creative data, and generates evidence units. The evidence graph library 310 stores the version relationships of evidence nodes, evidence edges, and multimodal evidence chains, and marks the activity version. The evidence view distributor 320 extracts evidence sub-chains from the activity version according to the evaluation dimension, and generates isolated evidence views for different evaluation agents to read.

[0116] The intelligent agent collaborative cluster 330 includes an aesthetic intent parsing intelligent agent 331, a creative process parsing intelligent agent 332, a work expression parsing intelligent agent 333, and an evidence verification intelligent agent 334. It is used to process isolated evidence views in parallel, and each intelligent agent generates an evaluation data packet containing evaluation scores for the same evaluation dimension. The dispute routing controller 340 organizes the evaluation data packets into an evaluation score matrix with evaluation intelligent agents as rows, evaluation dimensions as columns, and evaluation scores as elements. It calculates the degree of divergence, locates the smallest evidence subchain when divergence exists, and generates dispute routing information; when no divergence exists, the active version is handed over to the parallel verification container 360 for verification; if verification is unstable, dispute routing information is generated based on the returned dispute information.

[0117] The branch supplementary evidence scheduler 350 generates a target supplementary evidence task based on the disputed routing information and writes the incremental supplementary evidence unit into the candidate branch version in the evidence graph library 310. The parallel verification container 360 uses the candidate branch version or the active version without supplementary evidence as the baseline version to form a node masking version and a modality replacement version, and calls the intelligent agent collaborative cluster 330 to generate a baseline evaluation data packet, a disturbance evaluation data packet, and a stable state identifier.

[0118] When the stable state indicator indicates stability, the version release controller 370 switches the verified candidate branch version to the active version, or keeps the original active version, and sends the benchmark evaluation data packet to the result output module 400. When the stable state indicator indicates instability and the number of supplementary verifications has not reached the threshold, the candidate branch version or the active version without supplementary verification and the benchmark evaluation data packet are sent to the dispute routing controller 340. When the number of supplementary verifications reaches the threshold, insufficient evidence information is generated. The result output module 400 outputs the aesthetic education evaluation result or insufficient evidence information.

[0119] In one embodiment, such as Figure 5 As shown, the evidence graph library 310 includes an original evidence partition 311, a node-edge index partition 312, and a version genealogy table 313. The original evidence partition 311 stores evidence units and source summaries in an append-only, non-overwrite manner. The node-edge index partition 312 stores evidence nodes, evidence edges, evidence unit reference addresses, and modal indexes. The version genealogy table 313 records the parent-child relationship between active versions and candidate branch versions, newly added evidence nodes, newly added evidence edges, creation time, and activity status.

[0120] The branch supplementary evidence scheduler 350 forms candidate branch versions by referencing existing evidence nodes and adding evidence nodes corresponding to supplementary evidence units. The candidate branch version only records nodes and evidence edges added relative to the parent version. The version release controller 370 verifies that the source digest referenced by the candidate branch version matches the corresponding source digest in the original evidence partition 311, and verifies that the version identifier in the evaluation data packet matches the version identifier of the candidate branch version. If the verification matches, the active version marker in the version genealogy table 313 is changed atomically; if the verification does not match, the original active version is retained.

[0121] Specific application examples:

[0122] Example 1: This example uses digital painting to depict the tranquil atmosphere of a school campus after the rain as an aesthetic education task. Evaluation dimensions include aesthetic conception, expressive transformation, creative adjustment, and artistic expression. The unified evaluation score range is 0-1, the threshold for supplementary evidence is 3, the threshold for divergence is 0.015, and the threshold for stability is 0.05.

[0123] The data acquisition module 300 obtains the theme explanation text, oral explanation, sketches, three stages of the work, the final work, and layer and brush operation records. In the theme explanation, the use of low-saturation cool colors to express tranquility forms an aesthetic intention node; reducing background saturation, removing highlight decorations, and adding reflections form creative behavior nodes; the three stages of the work form work status nodes; and the cool tones, low contrast, and static composition in the final work form expression result nodes. The evidence graph library 310 establishes a multimodal evidence chain based on the temporal and content relationships between nodes.

[0124] The intelligent agent collaborative cluster 330 processes the corresponding isolated evidence views. Regarding the expression transformation dimension, the aesthetic intent analysis agent 331 outputs 0.82, the creation process analysis agent 332 outputs 0.76, the work expression analysis agent 333 outputs 0.48, and the evidence verification agent 334 outputs 0.70. The lower-scoring cited evidence from the work expression analysis agent 333 includes locally high-saturation areas in the final work.

[0125] The dispute routing controller 340 calculates weights, weighted average evaluations, and divergence based on the completeness of evidence. In this example, the four evaluation data packets in this dimension have the same completeness of evidence, and the divergence is 0.0165. After exceeding 0.015, the evidence edges in the candidate evidence subchains are sequentially masked according to the order of the evidence edge identifiers and re-evaluated. The evaluation divergence is determined to be concentrated in the smallest evidence subchain composed of the creative behavior node that deletes the highlighted decoration, the second-stage work status node, and the final work expression result node.

[0126] The branch-based verification scheduler 350 identified a lack of creative action nodes between the second-stage work and the final work that could explain the reappearance of high-saturation areas. Based on this, it generated a local explanation task and a local reproduction task. Historical records show that the expected divergence reduction for the local explanation task was 0.010, and the expected divergence reduction for the local reproduction task was 0.018. After substituting the normalized execution time, data collection volume, and creative interference into the verification priority formula, the local reproduction task had a higher verification priority and was therefore identified as the target verification task.

[0127] The student re-demonstrates the modification process for the corresponding area and explains that the area is used to represent the reflection of lights after rain. The data acquisition module 300 converts the reproduced video, audio description, and new operation records into supplementary evidence units. The branch supplementary evidence scheduler 350 writes the supplementary evidence units into the candidate branch version.

[0128] After supplementary verification, the intelligent agents 331 (aesthetic intent analysis), 332 (creative process analysis), 333 (work expression analysis), and 334 (evidence verification) output scores of 0.80, 0.78, 0.74, and 0.77 respectively for the expression transformation dimension, with the remaining divergence score decreasing below the divergence threshold. The parallel verification container 360 further masks voice description nodes referenced by only one evaluation agent and performs modal replacement using reproducible video nodes with the same content anchor points. The evaluation change for each perturbation version does not exceed 0.05, generating a stable state identifier representing stability.

[0129] After verifying the source summary and version identifier, the version release controller 370 switches the candidate branch version to the active version. The result output module 400 outputs the expression transformation evaluation results, the key evidence path from the aesthetic intention node through the creative behavior node and the work status node to the expression result node, as well as the remaining degree of divergence and the number of supplementary evidence attempts. In this embodiment, the inputs and outputs of each stage can be directly used by the next stage, and the supplementary evidence loop has a clear termination condition.

[0130] Example 2: This example uses a cross-media aesthetic education task involving creating an image and a musical fragment based on the same theme. The task requires students to first draw an image depicting a city in the early morning, and then create a musical fragment based on the spatial layers, light and shadow variations, and emotional progression within the image. Evaluation dimensions include aesthetic conception, cross-media expression transformation, creative adjustment, and artistic expression. A unified evaluation score range is 0-1, with a divergence threshold of 0.005 and a stability threshold of 0.05.

[0131] The data acquisition module 300 acquires the theme interpretation text, image sketches, image stage versions, final images, music project files, audio track editing records, final music fragments, and verbal explanations. The content in the theme interpretation that depicts the city's awakening from darkness to light forms the aesthetic intention node; the operations in the images that adjust brightness and spatial layers, and the operations in the music that add voices, change rhythms, and adjust volume envelopes, form the creative action node; the image stage versions and music project versions form the work status node; and the brightness and darkness levels of the final image and the dynamic range of the final music fragment form the expression result node.

[0132] The system encodes image regions and musical time segments into visual feature vectors and audio feature vectors, respectively, and establishes content anchors based on the correspondences given by students in their verbal explanations. For example, a region in the background of an image that transitions from dark to light is assigned the same content anchor as a time segment in a musical clip where the string section gradually intensifies. Content anchors only represent the cross-media representational correspondences explicitly established by students, and do not mechanically set color and pitch as fixed correspondences.

[0133] The aesthetic intent analysis agent 331 outputs a cross-media expression transformation score of 0.79 based on the theme explanation and verbal description; the creative process analysis agent 332 outputs 0.75 based on image modification records and audio track editing records; the work expression analysis agent 333 outputs 0.56 based on the final image and final music fragment; and the evidence verification agent 334 outputs 0.72. In this example, the evidence completeness of the four evaluation data packets in this dimension is the same, and the degree of divergence before supplementary evidence is 0.007625, which is greater than 0.005. The dispute routing controller 340 determines that the divergence mainly stems from the lack of direct creative process evidence between changes in image brightness and changes in music dynamics.

[0134] The branch-complement scheduler 350 generates a local description task and an equivalent modality representation task. The local description task requires students to select a region in an image and indicate its corresponding time interval and sound changes within a musical clip. The equivalent modality representation task requires students to re-express the emotional changes in that region by adjusting the dynamics and timbre of the music, without modifying the image. Based on the expected reduction in divergence and the execution cost, the system selects the local description task.

[0135] After students complete the partial description task, the system obtains the region selection coordinates, music time interval, voice description, and operation time. This data forms supplementary evidence units, establishing supplementary evidence relationships between image nodes and audio nodes with the same content anchor point. After a second evaluation, the four evaluation agents output scores of 0.78, 0.76, 0.73, and 0.75 for the cross-media expression transformation dimension, respectively.

[0136] The parallel verification container 360 masks individual voice description nodes and replaces them with audio track editing records under the same content anchor point. After supplementary verification, the remaining divergence degree is 0.000325, which is lower than the divergence threshold of 0.005; the maximum evaluation change before and after perturbation is 0.03, which is lower than the stability threshold of 0.05. Based on this, the version release controller 370 releases candidate branch versions, and the result output module 400 outputs the cross-media expression transformation evaluation results and corresponding key evidence paths.

[0137] It should be noted that the association threshold, conflict threshold, divergence threshold, modality threshold, verification threshold, stability threshold, path threshold, and frequency threshold can be stored as configuration parameters for the art education task in memory 200. During initial deployment, candidate thresholds can be iterated over on the validation set, and parameters can be selected based on the error between the evaluation score and the manual review result, the evidence missing identification rate, and the average number of supplementary verifications. Different art media can use different parameter sets, but the same parameter set should be used for all supplementary verification rounds within the same art education task.

[0138] When data for a particular modality cannot be read, the data acquisition module 300 records a reading failure flag and does not generate false modality feature vectors; the evidence graph library 310 marks the corresponding necessary evidence relationships as missing. When an evaluation agent times out, the agent collaboration cluster 330 retains the evaluation data packets of other evaluation agents, excludes the timed-out agent from the divergence calculation of the current dimension, and generates model execution anomaly information; if there are fewer than two valid evaluation data packets, insufficient evidence information is output. Model execution anomaly information is not used as a basis for student ability evaluation.

[0139] When historical completion data is insufficient to calculate the expected reduction in divergence, the branch supplementary evidence scheduler 350 uses a task template to preset initial values. After completing the actual supplementary evidence, the divergence degree before and after supplementary evidence, the minimum evidence subchain characteristics, and the actual execution cost are written into the historical completion data. Thus, the historical completion data is gradually updated along with the completed aesthetic education tasks, but the published activity versions and their evaluation results are not automatically changed due to subsequent changes in historical data.

[0140] The aesthetic education evaluation system can be deployed on a teaching platform server, or the data acquisition module 300 and part of the modal encoder can be deployed on terminal devices, while the evidence graph library 310, the intelligent agent collaborative cluster 330, and the dispute routing controller 340 can be deployed on the server. The terminal devices upload modal feature vectors, source summaries, and necessary evidence unit identifiers, and the server completes evidence chain construction, intelligent agent evaluation, and supplementary evidence scheduling. For implementations that require retaining the original video or audio, the original data can be stored in access-controlled object storage, and the evidence graph library 310 stores the access address and source summary.

[0141] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for evaluating aesthetic education based on multimodal evidence chains and multi-agent collaboration, characterized in that, include: S1. Obtain aesthetic education task information and multimodal creation data, convert the multimodal creation data into evidence units, and organize the evidence units into a multimodal evidence chain including aesthetic intention nodes, creative behavior nodes, work status nodes, and expression result nodes; S2. According to the evaluation dimensions in the aesthetic education task information, use multiple evaluation agents to process the multimodal evidence chain to obtain an evaluation data package including evaluation dimensions, evaluation scores, cited evidence nodes, and evidence completeness. S3. Compare multiple evaluation data packets, determine the evaluation dimensions with evaluation disagreements, locate the minimum evidence subchain that maintains the evaluation disagreement, and generate dispute information based on the minimum evidence subchain; if there are no evaluation disagreements and no dispute information generated by stability verification, then use the multimodal evidence chain as the updated multimodal evidence chain and proceed to step S5; if there is dispute information generated by stability verification, then proceed to step S4. S4. When the disputed information exists, generate candidate supplementary evidence tasks based on the disputed information, determine the target supplementary evidence task based on the expected reduction in disagreement and execution cost of each candidate supplementary evidence task, obtain the supplementary evidence unit formed by executing the target supplementary evidence task, and connect it to the multimodal evidence chain to obtain an updated multimodal evidence chain. S5. Using multiple evaluation agents to process the updated multimodal evidence chain, obtain updated evaluation data packets, determine the evaluation stability state based on the changes in evaluation scores before and after supplementary evidence and the remaining evaluation discrepancies between the updated evaluation data packets, and generate dispute information based on the evidence nodes that failed verification or the missing necessary evidence relationships when the state is unstable. S6. When the evaluation stability state is reached, output the aesthetic education evaluation result; when the evaluation stability state is not reached and the number of supplementary evidence attempts does not reach the threshold, use the updated multimodal evidence chain and the updated evaluation data package as the new multimodal evidence chain and evaluation data package respectively and return to step S3; when the number of supplementary evidence attempts reaches the threshold, output the evidence insufficiency information.

2. The aesthetic education evaluation method based on multimodal evidence chains and multi-agent collaboration as described in claim 1, characterized in that, Step S1 specifically includes: Read the evaluation dimension identifier, unified evaluation score range, and the required evidence relationship type corresponding to each evaluation dimension from the aesthetic education task information; Obtain the theme explanation, emotional expression goals, creative instructions, operation process, works at each stage, version records and the final work; Extract the modal feature vectors corresponding to each data point, and perform hash operations on the original encoded sequences of each data point to obtain the source summary; The modal feature vector, the source summary, the collection time, the student identifier, and the work stage are encapsulated into the evidence unit, and the work change range determined by adjacent version records and corresponding operation trajectories is written into the corresponding evidence unit.

3. The aesthetic education evaluation method based on multimodal evidence chains and multi-agent collaboration according to claim 1, characterized in that, Step S1 further includes: converting the evidence unit into evidence nodes of the corresponding category based on the data content and the stage of the work; When the collection time of the creative behavior node is between adjacent work status nodes, and the semantic similarity between the creative behavior node and the corresponding work change range reaches the association threshold, a behavior change evidence edge is established. Evidence edges for intention and behavior are established based on the semantic similarity between aesthetic intention nodes and creative behavior nodes, and evidence edges for change results are established based on the local feature similarity between work state nodes and expression result nodes. The association strength of each evidence edge is determined based on semantic similarity and node time interval, and each evidence node is connected according to the direction of the evidence edge to form the multimodal evidence chain.

4. The aesthetic education evaluation method based on multimodal evidence chains and multi-agent collaboration according to claim 1, characterized in that, Step S2 specifically includes: extracting evidence sub-chains from the multimodal evidence chain according to the evaluation dimension and evidence edge direction, blocking evidence nodes that do not belong to the current evidence sub-chain, and obtaining an isolated evidence view; The isolated evidence view is respectively input into the aesthetic intention analysis agent, the creative process analysis agent, the work expression analysis agent, and the evidence verification agent; The evaluation scores output by each evaluation agent are normalized to the unified evaluation score range; The ratio of the number of necessary evidence relationship types belonging to the corresponding evaluation dimension to the total number of necessary evidence relationship types is determined as the evidence completeness. When the ratio is zero, the preset lower limit of positive completeness is determined as the completeness of the evidence.

5. The aesthetic education evaluation method based on multimodal evidence chains and multi-agent collaboration according to claim 4, characterized in that, Step S3 specifically includes: Extract the evaluation score of the i-th evaluation agent for the k-th evaluation dimension from the valid evaluation data packets of the same evaluation dimension. and completeness of evidence ;according to Determine weights ; according to Obtain the weighted average score ;according to Obtaining the degree of divergence Where M is the number of evaluation agents that successfully output evaluation scores for the k-th evaluation dimension and M is not less than 2, i and r are the evaluation agent numbers and k is the evaluation dimension number; When the degree of divergence When the divergence threshold is exceeded, the referenced evidence nodes in multiple evaluation data packets and the evidence edges connecting the referenced evidence nodes are combined to form a candidate evidence subchain. The evidence edges in the candidate evidence subchain are sequentially masked and the divergence degree is recalculated. Evidence edges whose divergence degree still exceeds the divergence threshold after masking are deleted, and evidence edges whose divergence degree does not exceed the divergence threshold after masking are restored. Repeat the traversal until no evidence edge is deleted in a consecutive round. The remaining evidence nodes and evidence edges are determined as the minimum evidence subchain, and the dispute information is generated based on the minimum evidence subchain.

6. The aesthetic education evaluation method based on multimodal evidence chains and multi-agent collaboration according to claim 5, characterized in that, Step S4 specifically includes: The following are identified as the corresponding dispute types: missing evidence edges, associated nodes with semantic similarity below the conflict threshold, inconsistent source summary verification, or the proportion of cited evidence nodes in the same data modality reaching the modality threshold. Call the supplementary evidence task template corresponding to the dispute type, such as partial reproduction, partial description, source confirmation, or equivalent modality expression, to generate the candidate supplementary evidence task; Based on the similarity of evidence subchain features, a weighted average is taken of the non-negative cutoff values ​​of the difference in divergence before and after supplementary evidence in historical records with the same dispute type and supplementary evidence task template, to obtain the expected reduction in divergence for the j-th candidate supplementary evidence task. ;according to Calculate the priority of supplementary evidence ;in, , and These represent the normalized and non-negative execution time, data collection volume, and creative interference volume, respectively, where α, β, γ, and ε are all positive numbers; The candidate supplementary evidence task with the highest priority is determined as the target supplementary evidence task, and the corresponding supplementary evidence unit is incrementally connected to the multimodal evidence chain to form the updated multimodal evidence chain.

7. The aesthetic education evaluation method based on multimodal evidence chains and multi-agent collaboration according to claim 1, characterized in that, Step S5 specifically includes: The updated multimodal evidence chain is determined as the benchmark evidence chain, and the benchmark evidence chain is processed by multiple evaluation agents to obtain a benchmark evaluation data package; Evidence nodes whose source digests are inconsistent, whose evidence completeness is below the verification threshold, or whose evidence nodes are only cited by one evaluation agent are identified as evidence nodes to be verified. The nodes of evidence to be verified are sequentially blocked to form a node-blocked evidence chain; Based on the local features of the artwork and the time of collection, the same content anchor points are determined for the evidence nodes that represent the same artistic expression object; The evidence nodes to be verified are replaced with evidence nodes that have the same stage of work and content anchor point but different data modalities, forming a modality replacement evidence chain; The evaluation agents are used to process the node masking evidence chain and the modality replacement evidence chain. The maximum absolute difference between the obtained evaluation score and the corresponding evaluation score in the benchmark evaluation data packet is determined as the evaluation change, and the change of the corresponding evaluation score before and after the supplementary evidence is recorded. When at least one node-masking evidence chain or modal replacement evidence chain is formed, and the remaining divergence between the updated evaluation data packets does not exceed the divergence threshold and the evaluation change does not exceed the stability threshold, the evaluation stability state is determined to be reached. When there are no evidence nodes to be verified, the necessary evidence relationships are complete and the source digest verification is consistent, and the remaining divergence does not exceed the divergence threshold, the evaluation stability state is also determined to be reached.

8. The aesthetic education evaluation method based on multimodal evidence chains and multi-agent collaboration according to claim 1, characterized in that, Step S6 specifically includes: When the evaluation is stable, the key evidence path is formed by tracing back from the expression result node along the evidence edge whose association strength reaches the path threshold to the aesthetic intention node. The aesthetic education evaluation result is output according to the updated evaluation data package, and the number of unconnected evidence nodes, the remaining degree of divergence, and the number of supplementary evidence attempts are encapsulated as evaluation uncertainty information. When the evaluation stability state has not been reached and the number of re-certification attempts has reached the threshold, the evaluation interval is determined by the lowest and highest evaluation scores in each evaluation data package after re-certification. The insufficient evidence identifier, the evaluation range, and the unresolved dispute location are encapsulated as the insufficient evidence information.

9. An aesthetic education evaluation system based on multimodal evidence chains and multi-agent collaboration, used to implement the aesthetic education evaluation method based on multimodal evidence chains and multi-agent collaboration as described in any one of claims 1 to 8, characterized in that the system... include: An evidence graph library stores the version relationships of evidence nodes, evidence edges, and multimodal evidence chains, and marks active versions. An evidence view distributor extracts evidence subchains from the activity version according to the evaluation dimension, and generates isolated evidence views for different evaluation agents to read. The intelligent agent collaborative cluster processes the isolated evidence view in parallel and generates an evaluation data packet carrying version identifier, evaluation dimension, evaluation score, evidence completeness and referenced evidence node; The dispute routing controller organizes the evaluation data packet into an evaluation score matrix with evaluation agents as rows, evaluation dimensions as columns, and evaluation scores as elements. When there is an evaluation disagreement, it locates the minimum evidence subchain and generates dispute routing information. When there is no evaluation disagreement, it sends the activity version into the parallel verification container. A branch supplementary evidence scheduler generates a target supplementary evidence task based on the disputed routing information and incrementally writes the supplementary evidence units into the candidate branch versions in the evidence graph library. The parallel verification container uses the candidate branch version or the active version before the verification is completed as the base version to form a node masking version and a modality replacement version, and calls the intelligent agent collaborative cluster to generate a base evaluation data packet, a disturbance evaluation data packet and a stable state identifier. A version release controller, when the stability status indicator indicates stability, switches the verified candidate branch version to the active version, or keeps the active version; when the stability status indicator indicates instability and the number of supplementary certifications has not reached the threshold, the version release controller sends the candidate branch version or the active version without supplementary certification and the benchmark evaluation data packet to the dispute routing controller; when the number of supplementary certifications reaches the threshold, insufficient evidence information is generated.

10. The aesthetic education evaluation system based on multimodal evidence chains and multi-agent collaboration according to claim 9, characterized in that, The evidence map library includes: original evidence partitions, used to store evidence units and source summaries in an uncoverable manner; The node-edge index partition is used to store evidence nodes, evidence edges, and modal indexes; Version genealogy table is used to record the parent-child relationship between active versions and candidate branch versions, as well as newly added evidence nodes and evidence edges; The branch supplementary evidence scheduler forms the candidate branch version by referencing existing evidence nodes and adding evidence nodes corresponding to the supplementary evidence units. The version release controller verifies that the source digest referenced by the candidate branch version is consistent with the corresponding source digest in the original evidence partition, and verifies that the version identifier in the evaluation data packet is consistent with the version identifier of the candidate branch version; after verification, the active version tag in the version genealogy table is changed in an atomic update manner.