A content generation quality evaluation and optimization method based on a multi-modal large model

Abstract

The application discloses a content generation quality evaluation and optimization method based on a multi-modal large model, and belongs to the technical field of artificial intelligence, comprising the following steps: performing semantic analysis and cross-modal feature extraction on an input multi-modal content generation request instruction to generate a structured feature vector; obtaining initial generated content output by a multi-modal large model, performing full-dimensional quality scoring to obtain an initial quality score vector and a quality short board dimension; based on the structured feature vector and the initial quality score vector, matching a reference case from a historical optimization case library, extracting an optimization parameter sequence of the reference case, and calculating dynamic regulation weights of each optimization dimension according to a score distribution; constructing a multi-objective comprehensive optimization objective function, combining the optimization parameter sequence of the reference case to generate a composite optimization instruction to execute multi-round iterative content optimization, and outputting optimized content meeting preset quality threshold requirements; and effectively improving the quality stability and scene adaptability of the generated content.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for evaluating and optimizing the quality of content generation based on a multimodal large model. Background Technology

[0002] With the rapid development of multimodal large-scale modeling technology, content generation technologies that integrate multiple modalities such as text, images, and audio have been widely applied in content creation, intelligent interaction, digital media, online education, and other fields, becoming an important support for intelligent content production in various industries. At the same time, the quality requirements for generated content at the business implementation level continue to increase. The accuracy, consistency, compliance, and scenario adaptability of content are gradually becoming core indicators for measuring the effectiveness of multimodal generation. How to effectively control and iteratively optimize the quality of generated content has become one of the core research directions in this field.

[0003] Currently, quality control of multimodal generated content in the industry mainly focuses on single-point verification and manual correction after generation. In the instruction parsing stage, existing solutions generally rely on surface semantic recognition, with limited depth in mining the scene rules and constraints implicit in user needs. The semantic consistency verification mechanism between cross-modal inputs is relatively weak, making it difficult to provide complete and accurate feature basis for subsequent quality control from the input source.

[0004] In the quality assessment stage, most existing assessment systems focus on conducting quality testing independently for single-modal content, and do not adequately consider the impact of cross-modal semantic consistency and multi-dimensional quality coupling. The weight configuration of assessment dimensions often adopts fixed prior rules, which are difficult to dynamically adapt to the quality focus of different business scenarios. This results in insufficient scenario fit and quantitative accuracy of the assessment results, and cannot accurately and comprehensively reflect the true quality level of the generated content.

[0005] In the content optimization stage, mainstream optimization methods mostly rely on manual adjustment of prompts or repeated generation, resulting in a low degree of automation. Effective experience from historical optimization processes lacks a systematic accumulation and reuse mechanism, making it difficult to accumulate reusable strategies for optimizing similar scenarios and quality issues. At the same time, the optimization process often focuses on correcting single problems, with less systematic consideration of the ripple effects of optimization actions on the overall quality of the content. The convergence efficiency of optimization iterations and the stability of the final results are significantly limited, making it difficult to form a closed loop of quality improvement with full-chain automation.

[0006] Therefore, there is an urgent need for a content generation quality assessment and optimization method based on a multimodal large model to solve the above problems. Summary of the Invention

[0007] This invention provides a method for quality assessment and optimization of content generation based on a multimodal large model. It generates structured feature vectors by performing explicit and implicit semantic mining, cross-modal conflict resolution, and scenario-based feature reconstruction on multimodal generation instructions. This allows for comprehensive quality quantification of the initial generated content through multi-dimensional coupling calibration, accurately identifying quality weaknesses. Historical high-value optimization cases are matched, and dynamically adjusted weights are calculated. A multi-objective optimization function is constructed that balances the benefits of improving weaknesses with the costs of parameter adjustment. It combines cross-modal collaborative constraint manifold space iterative optimization to generate composite optimization instructions that drive model iterative optimization. This method solves the problems of incomplete instruction feature representation, insufficient objectivity and scenario adaptability in quality assessment, low reuse rate of historical optimization experience, and easy quality decay in non-weakness dimensions during the optimization process in existing technologies. It achieves fully automated quality assessment and targeted optimization across the entire multimodal content generation chain, effectively improving the quality stability and scenario adaptability of generated content and reducing the cost of manual intervention.

[0008] This invention provides a method for content generation quality assessment and optimization based on a multimodal large model, including: The system performs semantic parsing and cross-modal feature extraction on the input multimodal content to generate a structured feature vector; the structured feature vector includes business scenario tags, modal attributes, and compliance constraints. The initial generated content output by the multimodal large model is obtained, and the initial generated content is subjected to a full-dimensional quality score to obtain an initial quality score vector including the scores of each dimension, and the quality weakness dimensions below the qualified threshold are identified. Based on structured feature vectors and initial quality score vectors, reference cases with similar scene features and similar quality shortcomings are matched from the historical optimization case library. The optimization parameter sequence of the reference cases is extracted, and the dynamic adjustment weight of each optimization dimension is calculated according to the score distribution of each quality shortcoming dimension of the current content. A multi-objective comprehensive optimization objective function is constructed using dynamically adjusted weights as coefficients. Combined with the optimization parameter sequence of the reference case, a composite optimization instruction is generated to drive the multimodal large model to perform multiple rounds of iterative content optimization and output optimized content that meets the preset quality threshold requirements.

[0009] The beneficial effects of this invention are: This invention provides a method for content generation quality assessment and optimization based on a multimodal large model. It obtains structured feature vectors containing business scenarios, modal attributes, and compliance constraints by performing semantic parsing and cross-modal feature extraction on multimodal generation requests. Combined with the full-dimensional quality score of the initial generated content, it identifies the dimensions of quality shortcomings. Then, it matches similar historical optimization cases and calculates dynamically adjusted weights based on the score distribution to construct a multi-objective comprehensive optimization objective function. This generates composite optimization instructions that drive the model through multiple rounds of iterative optimization. This method can adapt to the differentiated quality requirements of multimodal content under different business scenarios, enabling targeted control of quality shortcomings. It improves the comprehensiveness of multimodal content generation quality assessment and the automation of the optimization process, reduces manual intervention, and efficiently outputs generated content that meets quality and compliance requirements.

[0010] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0011] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0012] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a content generation quality assessment and optimization method based on a multimodal large model, as described in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the generation of structured feature vectors in an embodiment of the present invention; Figure 3 This is a flowchart illustrating how to obtain an initial quality score vector and identify quality bottleneck dimensions below the acceptable threshold in an embodiment of the present invention. Figure 4 This is a flowchart illustrating the effective optimization actions for each case in this embodiment of the invention. Figure 5 This is a flowchart illustrating the process of obtaining a standard case library in an embodiment of the present invention. Detailed Implementation

[0013] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention. Furthermore, the business scenarios described in the embodiments of this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the emergence of new business scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0014] For ease of understanding, the technical terms used in the embodiments of this application will be introduced below.

[0015] Multimodal large models: refer to large-scale artificial intelligence models that can simultaneously receive, process and generate multiple modal data (such as text, images, audio, video, etc.), and have comprehensive capabilities for cross-modal understanding and generation.

[0016] Cross-modal feature extraction: Extracting features from data of different modalities (such as text and images) and mapping them to a unified semantic space so that computers can compute and understand the associations and similarities between data of different modalities.

[0017] Semantic anchors: In a multimodal feature sequence, the key feature points with the highest global semantic importance are calculated through an attention mechanism. They represent the most essential semantic information in that modality of data and are used for cross-modal alignment and conflict resolution.

[0018] Modal conflict resolution mechanism: When the similarity of the core semantics (semantic anchors) extracted from different modalities (such as text descriptions and reference images) is lower than a threshold, the system automatically triggers an error correction mechanism. By introducing contextual confidence, the features of low-confidence modalities are directionally corrected based on high-confidence modalities, thus eliminating semantic contradictions.

[0019] Implicit Constraint Reasoning Graph: A pre-constructed knowledge graph structure, including nodes such as scenarios, modal attributes, and compliance constraints, as well as weighted edges representing the strength of associations. It is used to uncover potential generation requirements not explicitly expressed by users (e.g., "e-commerce marketing" implicitly implies "compliance with advertising law's extreme wording").

[0020] Graph Neural Networks: A deep learning model specifically designed for processing graph-structured data. In this approach, information from neighboring nodes is aggregated through graph convolution operations to achieve constraint diffusion and reasoning.

[0021] Node diffusion inference: In graph neural networks, the initial node features (explicit user intent) are propagated and aggregated along the edges of the graph to neighboring nodes, ultimately activating and inferring potential associated constraint nodes.

[0022] Structured feature vectors: After parsing, fusing and reducing the dimensionality of unstructured multimodal user requests, they are transformed into fixed-length, machine-searchable and computationally comprehensible standardized numerical vectors that contain explicit identifiers such as business scenarios, modal attributes and compliance constraints.

[0023] Primitive feature extraction: Performing basic feature encoding on the raw multimodal data. For example, text is encoded using BERT after word segmentation, images are encoded using ViT after patching, and audio is encoded using Wav2Vec after frame segmentation.

[0024] Principal Component Analysis (PCA): A classic linear dimensionality reduction algorithm. It projects the original high-dimensional features onto a low-dimensional orthogonal basis with the largest variance through orthogonal transformation, preserving key information while removing redundant features and reducing computational complexity.

[0025] Dynamic scaling and dimensionality reduction reconstruction: Based on the weight coefficients of the business scenario, the different dimensions of the feature vector are non-uniformly scaled (highlighting the core identifier and compressing redundant information), and then PCA dimensionality reduction is performed to achieve adaptive optimization of feature representation.

[0026] Multi-granularity quality feature deconstruction: The generated content is decomposed from different levels (single-modal fine-grained, cross-modal association, compliance special) into multiple quantifiable indicators (such as word error rate, image-text semantic consistency, sensitive word hit count) to form a comprehensive set of quality features.

[0027] Dual-track timeline: Two parallel timelines are constructed in the case log analysis: one records the fluctuation of quality scores (short-board fluctuation timeline), and the other records the parameter adjustment actions (optimization operation timeline), which are used for causal correlation analysis.

[0028] Time-series steady-state characteristics: After the optimization operation is executed, the stability indicators of the quality score within a specific time window (such as the number of subsequent operations, the standard deviation of score fluctuation, and the mean score) are used to distinguish between "one-time achievement" and "multiple rounds of convergence".

[0029] Comprehensive Value Index: A composite indicator that evaluates the "value" of a historical optimization case. It is calculated by weighting three factors: scenario adaptability value, optimization effectiveness, and universal value of shortcomings, and is adjusted by a dynamic verification coefficient (considering the proportion of invalid operations).

[0030] Standard Case Ontology: A data entity that is a standardized encapsulation of high-value cases, including structured fields such as unique identifiers, scene feature vector groups, quality profile vector groups, optimization parameter sequences, effect verification tags, and value attribute tags.

[0031] Hierarchical retrieval index: a multi-level tree-like or dictionary-like index structure. This solution is divided into three levels: Level 1 is routing based on business scenarios; Level 2 is grouping by state compression encoding (7-bit binary to decimal integer) based on the combination of bottleneck dimensions; and Level 3 is tiered based on quality improvement rate to achieve millisecond-level accurate retrieval.

[0032] Multi-objective integrated optimization objective function: A mathematical function that simultaneously contains multiple conflicting objectives. In this scheme, it is transformed into a single-objective fitness function: a weighted combination of maximizing the benefit of the weakest link (first objective) and minimizing the cost and side effects of parameter adjustment (second objective).

[0033] Deviation penalty term: During the optimization process, a penalty mechanism is set to prevent the generated parameters from deviating excessively from the baseline values ​​of historical successful reference cases. It usually takes the form of squared error to ensure the stability and interpretability of the optimization.

[0034] Standard genetic algorithm: a global heuristic search algorithm that simulates the biological evolution process (selection, crossover, mutation) in nature. In this scheme, it is used to iteratively search within the initial solution space to find the optimal combination of parameters that maximizes the fitness function.

[0035] Cross-modal collaborative constraint manifold space: There are inherent physical or logical constraints among multimodal parameters (such as resolution upper limits, style parameter summation of 1). These constraints form a low-dimensional smooth surface (manifold) in a high-dimensional space. Restricting the optimization process within this manifold space ensures that the generated parameters are always legal and valid.

[0036] Fitness: In evolutionary algorithms, fitness is a numerical value used to evaluate the quality of each individual (i.e., a set of parameters) in the population. In this scheme, it is calculated by normalizing the gain, cost, and modality conflict penalty term.

[0037] Lévy flight: a random walk model whose step size follows a heavy-tailed probability distribution (mostly small step sizes, occasionally extremely large step sizes). In optimization algorithms, it is used for "global exploration," helping the algorithm escape local optima.

[0038] Exponential and logarithmic mappings: concepts in differential geometry. Exponential mappings are used to map vectors (such as update step sizes) in the tangent space (flat Euclidean space) of a manifold back onto the surface of the manifold, ensuring that the updated points still satisfy the constraints; logarithmic mappings are the inverse process, mapping points on the manifold back to the tangent space.

[0039] Geodesic: The "shortest path" between two points in manifold space (equivalent to a straight line in a plane). When updating parameters in manifold space, numerical integration along geodesics (such as the Euler method) can ensure that the parameter update trajectory is always smooth and strictly satisfies all coupling constraints.

[0040] like Figure 1 As shown, this invention provides a method for content generation quality assessment and optimization based on a multimodal large model, including: The system performs semantic parsing and cross-modal feature extraction on the input multimodal content to generate a structured feature vector; the structured feature vector includes business scenario tags, modal attributes, and compliance constraints. The initial generated content output by the multimodal large model is obtained, and the initial generated content is subjected to a full-dimensional quality score to obtain an initial quality score vector including the scores of each dimension, and the quality weakness dimensions below the qualified threshold are identified. Based on structured feature vectors and initial quality score vectors, reference cases with similar scene features and similar quality shortcomings are matched from the historical optimization case library. The optimization parameter sequence of the reference cases is extracted, and the dynamic adjustment weight of each optimization dimension is calculated according to the score distribution of each quality shortcoming dimension of the current content. A multi-objective comprehensive optimization objective function is constructed using dynamically adjusted weights as coefficients. Combined with the optimization parameter sequence of the reference case, a composite optimization instruction is generated to drive the multimodal large model to perform multiple rounds of iterative content optimization and output optimized content that meets the preset quality threshold requirements.

[0041] Specifically, the system performs semantic parsing and cross-modal feature extraction on the input multimodal content generation request command to generate a structured feature vector. This structured feature vector includes business scenario tags, modal attributes, and compliance constraints. The input is the user-submitted multimodal content generation request command, supporting mixed multimodal inputs such as text descriptions, reference images, and reference audio. The output is a fixed-length structured feature vector. Its function is to transform unstructured user needs into standardized features that are searchable and computable by machines, providing a unified feature benchmark for subsequent quality assessment and case matching.

[0042] The system acquires initial generated content output from a multimodal large model, performs a full-dimensional quality score on this initial generated content, obtaining an initial quality score vector including scores for each dimension, and identifies quality deficiency dimensions below the acceptable threshold. The input is the initial generated content output from the multimodal large model, and the output is the initial quality score vector and the set of quality deficiency dimensions. This method replaces manual sampling inspection, performing multi-dimensional automated quantitative evaluation of the generated content to accurately locate quality defects.

[0043] Based on structured feature vectors and initial quality score vectors, reference cases with similar scene features and similar quality shortcomings are matched from the historical optimization case library. The optimization parameter sequences of the reference cases are extracted, and the dynamic adjustment weights for each optimization dimension are calculated based on the score distribution of each quality shortcoming dimension of the current content. Here, the acceptable threshold is set as... (e.g., 60 points) Current shortcomings The score is The lower the score, the larger the gap, and the higher the urgency of optimization. Calculate the weakness dimension. The scoring gap: ;like ,but ; Calculate the basic urgency weights: ;in, For dimension The scoring gap; Basic urgency weight; It is the set of all the dimensions of the weakness; For dimension The score gap. Search for the weakness dimension from the "Dual-Dimensional Feature Value Benchmark Library". Dimensional value coefficient The initial dynamic weights are obtained by linearly fusing the basic urgency weights with the dimensional value coefficients. ;in, For dimension The initial dynamic weights; This is an urgency adjustment factor with a preset value of [value missing]. This means it focuses more on the current score gap; for all dimensions of weaknesses... After normalization, the final dynamic adjustment weight vector is obtained. : ; For dimension The initial dynamic weights; A multi-objective comprehensive optimization objective function is constructed using dynamically adjusted weights as coefficients. Combined with the optimization parameter sequence from a reference case, composite optimization instructions are generated to drive a multimodal large-scale model to perform multiple rounds of iterative content optimization, outputting optimized content that meets a preset quality threshold. The input consists of the optimization parameter sequence and dynamically adjusted weights, while the output is the optimized content that meets the quality threshold. Quality optimization is transformed into a mathematical optimization problem, driving multiple rounds of iteration in the large-scale model to achieve the optimal balance between improving weaknesses and preserving strengths.

[0044] The beneficial effects of the above technical solution are as follows: By performing semantic parsing and cross-modal feature extraction on multimodal generation requests, a structured feature vector containing business scenarios, modal attributes, and compliance constraints is obtained. Combined with the full-dimensional quality score of the initial generated content, the dimension of quality shortcomings is located. Then, similar historical optimization cases are matched, and the weights are dynamically adjusted based on the score distribution to construct a multi-objective comprehensive optimization objective function. A composite optimization instruction-driven model is generated for multiple rounds of iterative optimization. This can adapt to the differentiated quality requirements of multimodal content under different business scenarios, realize targeted adjustment of quality shortcomings, improve the comprehensiveness of multimodal content generation quality assessment and the degree of automation of the optimization process, reduce manual intervention, and efficiently output generated content that meets quality and compliance requirements.

[0045] Furthermore, such as Figure 2 As shown, the step of performing semantic parsing and cross-modal feature extraction on the input multimodal content generation request command to generate a structured feature vector includes: The multimodal content generation request instruction is subjected to explicit semantic parsing to extract surface intent features; at the same time, the surface intent features are input into a pre-constructed implicit constraint reasoning graph, and node diffusion reasoning is performed through graph neural network to determine the potential generation constraints not explicitly expressed by the user. Primitive features are extracted from the text modality, image modality, and audio modality of the instruction respectively to obtain the initial feature sequence of each modality; in the initial feature sequence, multiple semantic anchors representing the core semantics are identified based on the cross-modal attention mechanism; Calculate the semantic similarity matrix of semantic anchors between different modalities. When the similarity between cross-modal semantic anchors is detected to be lower than the preset alignment threshold, trigger the modal conflict resolution mechanism to generate conflict-free cross-modal fusion features. The surface intent features, potential generation constraints, and conflict-free cross-modal fusion features are concatenated to obtain an initial structured feature vector; Based on the parsed business scenario labels, the corresponding dimension adjustment coefficients are queried from the pre-set scenario-dimensional weight mapping table; the redundant feature subspace associated with business scenario labels, modal attributes and compliance constraints in the initial structured feature vector is dynamically scaled and dimensionality reduced and reconstructed using the dimension adjustment coefficients to generate a structured feature vector.

[0046] Specifically, the multimodal content generation request instruction is subjected to explicit semantic parsing to extract surface intent features. Simultaneously, these surface intent features are input into a pre-constructed implicit constraint reasoning graph, and node diffusion reasoning is performed using a graph neural network to determine potential generation constraints not explicitly expressed by the user. The explicit semantic parsing employs a fine-tuned BERT-based entity recognition model to perform sequence labeling on the text portion of the request instruction, extracting core entities and intent tags, and outputting surface intent features. These surface intent features include three categories: business scenario entities, modal requirement entities, and compliance requirement entities, which are converted into a 32-dimensional one-hot + semantic vector concatenation feature vector. The pre-constructed implicit constraint reasoning graph includes nodes and edges; the nodes are divided into scenario nodes, modal attribute nodes, and compliance constraint nodes; scenario nodes include e-commerce marketing, education and training, etc.; modal attribute nodes include high-definition images, short videos, etc.; compliance constraint nodes include advertising law compliance, content security, etc.; the edges are directed weighted edges, with weight values ​​ranging from [0,1], representing the correlation strength from scenario to constraint, obtained by domain expert annotation combined with historical data statistics, for example, the edge weight from the e-commerce marketing node to the advertising law extreme word compliance node is 0.92; the complete set of nodes. .in, This includes all predefined business scenarios (such as e-commerce, education, etc.). Includes all modal attributes, Includes all compliance terms; side weight The calculation is performed using a combination of statistical probability and expert knowledge. in, Scenes in historical data With constraints The number of times they appear together For the scene The total number of occurrences (i.e., calculating the conditional probability); The domain experts score the strength of this association (normalized to 0-1). The statistical weighting coefficient is set to 0.7. The system recalculates the statistics based on newly added historical case logs every preset period (e.g., 30 days). And using the moving average formula The edge weights are updated smoothly to adapt to changes in business trends.

[0047] Node diffusion inference is performed using a graph neural network to determine potential generation constraints not explicitly expressed by the user. Specifically, a two-layer graph convolutional network (GCN) is used for node diffusion inference. The extracted surface intent features are used as the initial node feature vector X, with the dimension being the total number of nodes. The initial value of the hit nodes is 1, and the rest are 0. The first layer of graph convolution is calculated as follows: ;in, This is the activation value vector calculated after convolution of the first layer graph; It is the ReLU activation function, used to introduce nonlinear transformations; The adjacency matrix for incorporating self-loops; ; This is the original adjacency matrix; It is the identity matrix; For the corresponding The degree matrix is ​​used to perform symmetric normalization on the adjacency matrix; The initial node feature vector matrix has a dimension equal to the total number of nodes. The initial value of the hit nodes is 1, and the rest are 0. The trainable weight matrix for the first layer of graph convolution is given. The second layer of graph convolution is calculated as follows: ;in, This is the activation value vector for all constraint nodes; The trainable weight matrix is ​​the second-layer graph convolution; the activation threshold is set to 0.6, and the constraint terms corresponding to the constraint nodes with activation values ​​≥ 0.6 are the latent generative constraints obtained through inference. The latent generative constraints include latent modal attribute expectations and latent compliance boundaries.

[0048] Primitive feature extraction is performed on the text, image, and audio modalities of the instruction to obtain initial feature sequences for each modality. In these initial feature sequences, multiple semantic anchors representing the core semantics are identified based on a cross-modal attention mechanism. Specifically, primitive feature extraction for each modality yields initial feature sequences for each modality, including initial text, initial image, and initial audio feature sequences. The text modality uses Jieba word segmentation to obtain a word sequence, and each word is encoded using a BERT model to obtain a 768-dimensional feature vector. The initial feature sequence for the text is generated; for the image modality, the ViT-B / 16 model is used to divide the image into 16×16 patches, and each patch is encoded into a 768-dimensional feature vector, forming the initial feature sequence for the image; for the audio modality, the Wav2Vec2.0 model is used to encode the audio in frames, obtaining a 768-dimensional feature vector for each frame, forming the initial feature sequence for the audio; a cross-modal attention mechanism is used to calculate the global attention weights of each modality feature, with the text feature sequence as the query Q and the image / audio feature sequence as the key K and value V, to calculate the cross-modal attention weight matrix. ;in, The calculated cross-modal attention weight matrix; This is the query vector, which is a sequence of text features. This is the key vector, which is a sequence of image or audio features. Let be the transpose of the key vectors; The dimension of the key vector (768-dimensional in this case) is used to scale the dot product result and prevent gradient vanishing. The average of the attention weights at each feature location is taken to obtain the global semantic importance score of that feature. An importance threshold of 0.7 is set. Feature points with a score ≥ 0.7 are semantic anchors, and each semantic anchor corresponds to a 768-dimensional feature vector.

[0049] The semantic similarity matrix of semantic anchors between different modalities is calculated. When the similarity between cross-modal semantic anchors is detected to be lower than a preset alignment threshold, a modal conflict resolution mechanism is triggered to generate conflict-free cross-modal fusion features. Specifically, for the text and image modalities, an N×M dimensional semantic similarity matrix S is constructed, where N is the number of text anchors and M is the number of image anchors. ;in, For the first The text anchor point and the first The cosine similarity between image anchor points, after non-negative truncation, has a value range of [0,1]. , They are vectors and The modulus length; For the first Each text anchor feature vector For the first Each image anchor point feature vector; a preset alignment threshold of 0.7; if the maximum similarity value in the matrix is ​​<0.7, a cross-modal semantic anchor point conflict is determined, triggering a conflict resolution mechanism; the modal conflict resolution mechanism is as follows: extract the contextual features of the current instruction, combine them with a preset scenario prior probability distribution, calculate the contextual confidence of each modal semantic anchor point, and use the semantic anchor point feature with the highest contextual confidence as a benchmark to perform directional correction on the modal features with lower confidence, generating conflict-free cross-modal fusion features; the contextual confidence calculation is performed by extracting the contextual features of the current instruction, querying a preset scenario prior probability distribution table, and obtaining the prior confidence of the text modality in that scenario. Prior confidence of image modality Combining the anchor's own feature confidence (normalized feature vector magnitude), the final contextual confidence is calculated: ;in, For the first The final contextual confidence of each semantic anchor point; This represents the prior confidence level of the modality (text or image) in the current scene. For vectors The modulus length; The maximum value of the feature vector magnitude of all anchor points in the current modality is used for normalization; the anchor point with the highest context confidence is selected as the baseline anchor point. For the remaining anchor features with confidence levels below the threshold, corrections are made: ;in, For the first time before the correction Anchor point feature vectors; For the revised first Anchor point feature vectors; The feature vector of the benchmark anchor point with the highest confidence level in the context; To preserve the coefficients, the values ​​are taken as the confidence level of the anchor point itself, ranging from [0,1]. After the correction, all anchor point features are concatenated and then uniformly mapped to 768 dimensions through a pre-set linear dimensionality reduction mapping layer to obtain conflict-free cross-modal fusion features.

[0050] The surface intent features, latent generation constraints, and conflict-free cross-modal fusion features are concatenated to obtain an initial structured feature vector; specifically, the surface intent features (32 dimensions), latent generation constraints (15 dimensions), and conflict-free cross-modal fusion features (768 dimensions) are concatenated end to end to obtain an 815-dimensional initial structured feature vector. Based on the parsed business scenario tags, the corresponding dimension adjustment coefficients are retrieved from a pre-set scenario-dimensional weight mapping table. Using these dimension adjustment coefficients, the redundant feature subspace associated with the business scenario tags, modal attributes, and compliance constraints in the initial structured feature vector is dynamically scaled and dimensionality reduced for reconstruction, while retaining the explicit identification information of the business scenario tags, modal attributes, and compliance constraints, thus generating a structured feature vector. Specifically, based on the business scenario tags, the pre-set scenario-dimensional weight mapping table is retrieved to obtain the corresponding 815-dimensional dimension adjustment coefficient vector. Each element takes a value in the range [0.1, 1.0], and the core identifier coefficient is fixed at 1.0; where the core identifier represents the dimensions corresponding to the business scenario, modal attribute, and compliance constraint; the initial feature vector is element-wise scaled. ;in, The scaled feature vector; The initial structured feature vector is 815 dimensions; The vector consists of 815-dimensional feature adjustment coefficients, with the core identifier coefficient fixed at 1.0 and the remaining values ​​ranging from 0.1 to 1.0. The PCA algorithm is used to reduce the dimensionality of the scaled features to 128 dimensions, while retaining three explicit identifier fields: business scenario labels, modal attributes, and compliance constraints. The final output is a structured feature vector. A pre-defined scenario-dimensional weight mapping table construction method is included, among others, for explicit identifier dimensions directly corresponding to business scenario labels, modal attributes, and compliance constraints, their adjustment coefficients... The coefficients are fixed at 1.0 to ensure that critical information is not lost. For the remaining redundant feature subspaces, their coefficients... The calculation is based on the information gain or variance contribution rate of this dimension in a specific scenario from historical data. ;in, Information gain represents the contribution of the feature to distinguishing the current scenario; the lower limit of the coefficient is set to 0.1 to prevent the feature from being completely erased, and the upper limit is set to 1.0 (after normalization); for compliance-related scenarios (such as finance and healthcare), the coefficients of all compliance-related feature dimensions are manually increased to the range of 0.8-1.0.

[0051] The beneficial effects of the above technical solution are as follows: It extracts surface intent through explicit semantic parsing, mines potential generation constraints not directly expressed by the user through graph neural network reasoning using implicit constraint reasoning graphs, extracts multimodal primitive features and locates semantic anchors using cross-modal attention mechanisms, detects and resolves modal conflicts based on semantic similarity matrices to obtain cross-modal fusion features, and finally combines dimensionality adjustment coefficients corresponding to the business scenario to complete dynamic scaling and dimensionality reduction reconstruction of features to generate structured feature vectors. This approach not only fully mines the explicit and implicit information of instructions, effectively solving the semantic misalignment and modal conflict problems existing in multimodal instructions, but also eliminates redundant features and compresses invalid feature subspaces in conjunction with business scenarios, improving the semantic integrity, modal consistency, and feature expression accuracy of structured feature vectors. This provides a highly reliable feature input foundation for subsequent quality assessment, case matching, and content optimization.

[0052] Furthermore, such as Figure 3 As shown, the initial generated content output by the multimodal large model is obtained, and a full-dimensional quality score is performed on the initial generated content to obtain an initial quality score vector including the scores of each dimension. Furthermore, the quality bottleneck dimensions below the acceptable threshold are identified, including: Obtain the initial generated content output by the multimodal large model, perform multi-granularity quality feature deconstruction on the initial generated content, and extract single-modal fine-grained features, cross-modal correlation features and compliance-specific features respectively to obtain a full-dimensional quality feature set; Using seven quality dimensions—content relevance, factual accuracy, logical coherence, expression standardization, cross-modal consistency, information timeliness, and compliance effectiveness—as orthogonal coordinate axes, a seven-dimensional quality weight space is constructed. The set of all-dimensional quality features is then matched to the corresponding quality dimensions of the seven-dimensional quality weight space through a dimension-feature mapping relationship. For each quality dimension, principal component analysis is used to determine the optimal projection direction of the feature set under that dimension. The features are mapped along the optimal projection direction to generate a feature point cloud distribution. The Mahalanobis distance from each feature point to the ideal reference point of the corresponding dimension is directly calculated using the inverse matrix of the feature covariance matrix. After normalization, the feature fit of each feature is obtained; The pre-trained random forest feature importance model is called to output the intra-dimensional importance weights of each feature. The feature fit is weighted and summed to obtain the original fit scores for each dimension. The coupling influence coefficient between dimensions is calculated based on the pre-constructed dimensional correlation strength matrix, and the initial baseline weight vector for the scenario is coupled and corrected to obtain the coupled corrected weight vector. Based on the full-dimensional quality feature set, the information entropy of each dimension is calculated. An objective weight vector is generated by the entropy weight method. The modified weight vector and the objective weight vector are fused and coupled according to a preset ratio to obtain the final dimension weight vector. The original fit scores of each dimension are mapped to percentage scores and normalized to generate preliminary dimension scores. Then, the preliminary dimension scores are weighted and calculated in combination with the final dimension weight vector to generate a comprehensive quality score for all dimensions. The initial dimensional score vectors are compared one by one with the corresponding pass thresholds for each dimension, and the dimensions with scores below the pass thresholds are marked as quality weakness dimensions.

[0053] Specifically, the initial generated content output by the multimodal large model is obtained, and the initial generated content is deconstructed into multi-granularity quality features. Single-modal fine-grained features, cross-modal association features, and compliance-specific features are extracted to obtain a full-dimensional quality feature set. Specifically, features are extracted from the initial generated content at three granularities, resulting in a total of 87 quality features: the single-modal fine-grained features are 52-dimensional, including text, image, and audio dimensions; the text dimension includes 24 dimensions such as word error rate, syntactic complexity, and logical coherence; the image dimension includes 20 dimensions such as clarity, color fidelity, and compositional balance; the audio dimension includes 8 dimensions such as signal-to-noise ratio, naturalness, and speech rate fluctuation rate; the cross-modal association features are 18-dimensional, including text-image semantic consistency, audio-visual synchronization deviation, and multimodal information complementarity; and the compliance-specific features are 17-dimensional, including the number of sensitive word hits, the number of extreme word hits, and intellectual property risk scores.

[0054] A seven-dimensional quality weight space is constructed; the set of full-dimensional quality features is matched to the corresponding quality dimensions of the seven-dimensional quality weight space through a dimension-feature mapping relationship; wherein, a seven-dimensional Euclidean space is constructed with seven dimensions as orthogonal coordinate axes: content relevance, factual accuracy, logical coherence, expression standardization, cross-modal consistency, information timeliness, and compliance effectiveness; a dimension-feature mapping table is pre-established to assign the 87-dimensional quality features to the corresponding quality dimensions, for example, word error rate is assigned to expression standardization, and semantic consistency between text and images is assigned to cross-modal consistency.

[0055] For each quality dimension, principal component analysis (PCA) is used to determine the optimal projection direction of the feature set under that dimension. The features are then mapped along the optimal projection direction to generate a feature point cloud distribution. Using the inverse of the feature covariance matrix, the Mahalanobis distance from each feature point to the ideal reference point of the corresponding dimension is directly calculated. After normalization, the feature fit of each feature is obtained. Specifically, for each quality dimension, PCA is used to solve for the optimal projection direction, and all features under that dimension are used to construct a sample matrix. , where n is the number of historical samples and d is the number of features in this dimension; Calculate the covariance matrix ,in, Let be the covariance matrix of the feature set; This represents the number of historical samples. This is a sample matrix composed of all features under this dimension; Sample matrix Transpose the vector; perform eigenvalue decomposition on it, and take the top k eigenvectors with a cumulative contribution rate ≥ 90% to form the projection matrix P; project the eigenvectors of the current sample onto the principal component space to obtain the projection points. ,in, The projection point of the current sample feature vector onto the principal component space; This is the original feature vector of the current sample; The projection matrix is ​​composed of the top contributors with a cumulative contribution rate ≥ 90%. The feature points are composed of 10 feature vectors; the projection of all historical samples forms a feature point cloud distribution. The mean projection of the full-scoring samples in this dimension is taken as the ideal reference point. The perfect score sample is the sample with a human score of 95 or higher; calculate the Mahalanobis distance from the current sample point to the reference point: ;in, The Mahalanobis distance from the current sample point to the ideal reference point; This is the projection point of the current sample; This is the projected mean of the perfect score samples in this dimension; This is the transpose of the matrix; This is the inverse of the covariance matrix of the projected point cloud; to avoid the covariance matrix... When the feature dimension is greater than the number of samples, singularities occur, leading to irreversibility. This solution uses a regularization method to address this: the corrected covariance matrix calculation formula is as follows: ;in, The covariance matrix after regularization is guaranteed to be non-singular and invertible. This is the regularization coefficient, with a default value of [value missing]. ; It is the identity matrix; The covariance matrix of the original projected point cloud is given; the corrected Mahalanobis distance calculation formula is as follows: ; Let be the inverse of the regularized covariance matrix; take the historical maximum Mahalanobis distance in this dimension as . Calculate the feature fit: By introducing a truncation function, the range of values ​​is strictly limited to the interval [0,1]. For feature fit, the value range is strictly limited to [0,1]; The Mahalanobis distance for the current sample; This is the largest Mahalanobis distance in history for this dimension.

[0056] The pre-trained random forest feature importance model is used to output the intra-dimensional importance weights of each feature. The feature fit scores are then weighted and summed to obtain the raw fit scores for each dimension. The pre-trained random forest feature importance model uses historical manually labeled quality scores as labels and each quality feature as input to train a random forest regression model (100 trees, maximum depth 10). After training, it outputs the Gini importance of each feature, which is then normalized to obtain the intra-dimensional importance weights of each feature. The original fit score for this dimension. ;in, This is the original fit score for this dimension; For the first Importance weights within the dimensions of each quality feature; For the first The feature fit of each quality characteristic.

[0057] The inter-dimensional coupling influence coefficient is calculated based on a pre-constructed dimensional correlation strength matrix. The initial baseline weight vector for the scenario is then coupled and corrected to obtain a coupled-corrected weight vector. Information entropy for each dimension is calculated based on the full-dimensional quality feature set. An objective weight vector is generated using the entropy weighting method. The coupled-corrected weight vector and the objective weight vector are then fused according to a preset ratio to obtain the final dimensional weight vector. The pre-constructed dimensional correlation strength matrix R (7×7) uses Pearson correlation coefficients calculated for pairwise dimensional scores based on historical scoring data as matrix elements. The value ranges from [-1, 1]. Row normalization of R yields the correlation influence matrix. ; Set the initial baseline weight vector for the specific scenario ( (7-dimensional), calculate the coupling correction weights: ;in, This is the weight vector after coupling correction; This is the initial baseline weight vector for the specific scenario. The matrix represents the dimensional correlation influence matrix after row normalization. Based on the full-dimensional quality feature set, the information entropy of each dimension is calculated. An objective weight vector is generated using the entropy weighting method. The modified weight vector and the objective weight vector are then fused and coupled according to a preset ratio to obtain the final dimensional weight vector. Specifically, the 7-dimensional scores of the m samples in the current batch are standardized to eliminate the influence of dimensions. ;in, For the first The first dimension Standardized score of each sample; This is the original score; , The first Minimum and maximum scores for each dimension; calculate the score for the 1st dimension. Dimension 1 Percentage of each sample: ;in, For the first Dimension 1 The proportion of each sample; This represents the total number of samples in the current batch; to avoid the undefined logarithm, the percentage is smoothed using a minimum value. ;in, This represents the percentage after smoothing, used to avoid the undefined logarithm. For the first Dimension 1 The proportion of each sample; To prevent the denominator from being zero; calculate the first... Information entropy of a dimension: ;in, This represents the total number of samples in the current batch. For the first Information entropy of dimensions; The percentage after smoothing; calculate the difference coefficient: ; For the first The difference coefficient of a dimension; the smaller the information entropy, the larger the difference coefficient, and the higher the weight. Normalization yields the objective weight vector. ;in, For the first Objective weights for dimensions; subjective and objective weights are fused based on a 6:4 ratio: ; This is the final fused dimension weight vector; , These are the coupling correction weight vector and the objective weight vector, respectively.

[0058] The original fit scores of each dimension are mapped to percentage scores and normalized to generate preliminary dimension scores. Then, the preliminary dimension scores are weighted and calculated in conjunction with the final dimension weight vector to generate a comprehensive quality score for all dimensions. The preliminary dimension score vector is compared with the corresponding qualified threshold for each dimension, and dimensions with scores below the qualified threshold are marked as quality weakness dimensions. Among them, the original fit scores of each dimension are multiplied by 100 to obtain preliminary dimension scores of 0-100. ;in, The overall quality score is calculated across all dimensions. For the first Preliminary dimension scores for each quality dimension; For the first The final weights of each quality dimension; The initial scores of each dimension are compared with the preset passing threshold (60 points) one by one. Dimensions with scores <60 points are marked as quality weakness dimensions.

[0059] The beneficial effects of the above technical solution are as follows: It deconstructs the initial generated content using multi-granularity quality features, extracts single-modal, cross-modal, and compliance-specific features to construct a full-dimensional quality feature set, completes dimensional mapping of features based on a seven-dimensional quality weight space, calculates feature fit using principal component analysis and Mahalanobis distance, obtains feature importance weights using a random forest model to get the original fit scores for each dimension, and then combines the dimensional correlation strength matrix to achieve dimensional coupling correction. It integrates the objective weight vector obtained by the entropy weight method to obtain the final dimensional weight vector, and obtains the quality score vector through score mapping calibration and weighting, thus screening out dimensions with quality weaknesses. This approach can take into account differences in feature distribution, inter-dimensional coupling correlation, and objective data information, reducing subjective bias caused by a single evaluation method. It achieves multi-dimensional, fine-grained, and objective quantitative quality assessment of multimodal content, accurately identifies defective quality weaknesses, and provides a reliable quality basis for subsequent case matching and targeted optimization.

[0060] Furthermore, before matching reference cases from the historical optimization case library, the method further includes the step of constructing the historical optimization case library, including: Obtain all historical multimodal generation optimization sample source data, and perform deduplication, outlier removal and format normalization preprocessing on the sample source data to obtain a standardized case source dataset; the standardized case source dataset includes request-side structured feature vector, initial generated content, initial full-dimensional quality score vector, full-link optimization operation log, optimized generated content and optimized full-dimensional quality score vector. The structured feature vector and the initial full-dimensional quality score vector are dimensionally split to obtain the scene feature dimension set and the shortcoming dimension set, respectively. Based on historical full-scale optimization big data, the influence of matching of each dimension in the set of scene feature dimensions and the set of shortcoming dimensions on the optimization effect is statistically analyzed, and a dual-dimensional feature value benchmark library is constructed. For each case in the standardized case source dataset, the optimization actions are filtered to obtain the effective optimization actions corresponding to each case; Calculate the overall optimization effectiveness of the effective optimization actions corresponding to each case, and generate a bottleneck-optimization parameter association mapping table; Based on the comprehensive optimization effectiveness, the comprehensive value index of each case is calculated, and each case is value-screened based on the comprehensive value index to obtain high-value cases; combined with the short-board-optimization parameter association mapping table, the high-value cases are standardized and structured to generate a standard case ontology, which is then stored to obtain a standard case library; A hierarchical search index and a global index metadata table are constructed based on all standard case ontology in the standard case library; By integrating the dual-dimensional feature value benchmark library, standard case library, hierarchical retrieval index, and global index metadata table, a historical optimization case library is obtained.

[0061] Specifically, the system acquires all historical multimodal generation optimization sample source data, performs deduplication, outlier removal, and format normalization preprocessing on the sample source data to obtain a standardized case source dataset. The standardized case source dataset includes a request-side structured feature vector, initial generated content, initial full-dimensional quality score vector, end-to-end optimization operation log, optimized generated content, and optimized full-dimensional quality score vector. Specifically, it pulls all completed optimization tasks within a specified time period from the business log system, model inference platform, and quality evaluation engine. Each original sample must include six core fields: request-side structured feature vector, initial generated content, initial full-dimensional quality score vector, end-to-end optimization operation log, optimized generated content, and optimized full-dimensional quality score vector. The sample source data undergoes deduplication, outlier removal, and format normalization preprocessing to obtain the standardized case source dataset. The structured fields for each sample are fixed as: {Sample ID, Structured Feature Vector, Initial Generated Content Resource Address, Initial Quality Score Vector, Optimization Operation Log Array, Optimized Content Resource Address, Optimized Quality Score Vector}.

[0062] The structured feature vector and the initial full-dimensional quality score vector are dimensionally split to obtain a set of scenario feature dimensions and a set of shortcomings dimensions. The set of scenario feature dimensions includes three primary dimensions: business scenario, modal attribute, and compliance constraint. Each primary dimension has corresponding secondary sub-dimensions. The set of shortcomings dimensions includes dimensions of content relevance, factual accuracy, logical coherence, expression standardization, cross-modal consistency, information timeliness, and compliance effectiveness. Specifically, a two-level hierarchical architecture is adopted, divided into three primary dimensions, each with corresponding secondary sub-dimensions, and each dimension is assigned a globally unique code. Primary dimension S01 is the business scenario category, and the secondary sub-dimensions include: e-commerce marketing, education and training, industrial design, cultural and entertainment content, government publicity, medical science popularization, and financial services, totaling seven categories; corresponding codes... The first-level dimension S02 is the modal attribute class, and the second-level sub-dimensions include four categories: generated modality type, resolution level, duration level, and style type; the corresponding codes are... The modal types include text and images, short videos, 3D models, digital humans, and voice; the first-level dimension S03 is the compliance constraint category, and the second-level sub-dimensions include content security compliance, advertising law compliance, intellectual property compliance, and industry-specific compliance, for a total of 4 categories; corresponding codes All second-level sub-dimensions are the smallest granularity dimensional units, and subsequent value calculations and feature matching are all based on these second-level sub-dimensions. The weakness dimension set is split into 7 independent dimensions, each assigned a unique code. Among them, D01 is content relevance; D02 is factual accuracy; D03 is logical coherence; D04 is expression standardization; D05 is cross-modal consistency; D06 is information timeliness; and D07 is compliance validity. Two dimension mapping tables are output: a scenario dimension coding table and a weakness dimension coding table, including fields such as dimension ID, dimension name, level, and parent dimension ID, which serve as a unified dimension dictionary for the entire database.

[0063] Based on historical full-scale optimization data, the impact of matching each dimension in the scene feature dimension set and the shortcoming dimension set on the optimization effect is statistically analyzed, and a dual-dimensional feature value benchmark library is constructed. Specifically, based on historical full-scale optimization data, the impact of matching each dimension in the scene feature dimension set and the shortcoming dimension set on the optimization effect is statistically analyzed, and the objective contribution value of each dimension is calculated. The importance evaluation values ​​of each dimension output by the expert review node pool are obtained, and a preset first weight is assigned to the objective contribution value, and a preset second weight is assigned to the importance evaluation value. The dimension value coefficient of each dimension is calculated by weighting the values; the first weight is greater than the second weight. The unique identifier of each dimension, the dimension name, and the corresponding dimension value coefficient are mapped and stored one-to-one, generating a scene feature value table including each dimension in the scene feature dimension set and a table including the shortcoming dimension set. The value tables of each dimension in the set of shortcomings are integrated to obtain a dual-dimensional feature value benchmark library. Specifically, the objective contribution value is calculated based on historical full-scale optimization effect big data. The objective contribution value of each dimension is calculated using the controlled variable statistical method. For a single target dimension, two sets of control samples are split from the full set of historical cases. The two sets of control samples are a matching group and a non-matching group. The matching group is the set of cases where the feature of the dimension is completely matched. The non-matching group is the set of cases where the feature of the dimension is not matched. The average improvement rate of the two sets of samples is calculated. The difference between the average improvement rates of the two groups is calculated. This difference is the original objective contribution value of the dimension. The larger the difference, the greater the impact of whether the dimension is matched or not on the optimization effect, and the higher the objective value of the dimension itself. The original objective contribution values ​​of all dimensions are normalized by min-max. ;in, The normalized objective contribution value is mapped to the [0,1] interval; This is the original objective contribution value; , The minimum and maximum values ​​among the original objective contribution values ​​for all dimensions are mapped to the [0,1] interval to accommodate negative values, resulting in the final objective contribution value. For example, in the business scenario dimension, the average improvement rate of the matched group is 28.3%, and the average improvement rate of the unmatched group is 12.7%, with a difference of 15.6%. The maximum difference among all dimensions is 20.1%, and the minimum difference is 5.1%. Therefore, the normalized objective contribution value for this dimension is... An expert review pool was constructed, consisting of three categories of personnel, with no fewer than 5 people in each category, for a total of no less than 15 people: multimodal algorithm engineers, content operation experts, and compliance review experts, covering the three perspectives of technology, business, and compliance. The Delphi method was adopted, with experts independently scoring the importance of each secondary scenario dimension and each weakness dimension on a scale of 0 to 10. After collecting all scores, the highest and lowest scores were removed, and the arithmetic mean of the remaining scores was taken to obtain the original expert evaluation value for that dimension. The original expert evaluation value was divided by 10 and normalized to the [0,1] interval to obtain the final expert importance evaluation value. The first weight (objective contribution value weight) is preset to 0.7, and the second weight (expert importance evaluation value weight) is preset to 0.3, satisfying that the first weight is greater than the second weight; the value coefficients of all scenario sub-dimensions and all weakness dimensions are calculated one by one; the unique identifier of the dimension, the dimension name and the corresponding dimension value coefficient are mapped and stored one by one, and a scenario feature value table including each dimension in the scenario feature dimension set and a weakness dimension value table including each dimension in the weakness dimension set are generated respectively. The two value tables are logically integrated and a unified query interface is provided to the outside world, which is the dual-dimensional feature value benchmark library. For each case in the standardized case source dataset, optimization actions are filtered to obtain the effective optimization actions corresponding to each case. The comprehensive optimization effectiveness of the effective optimization actions corresponding to each case is calculated, and a bottleneck-optimization parameter association mapping table is generated. Based on the comprehensive optimization effectiveness, the comprehensive value index of each case is calculated, and each case is value-filtered based on the comprehensive value index to obtain high-value cases. Combining the bottleneck-optimization parameter association mapping table, the high-value cases are standardized and structured to generate a standard case ontology, which is then stored to obtain a standard case library. Based on all the standard case ontologs in the standard case library, a hierarchical retrieval index and a global index metadata table are constructed. The dual-dimensional feature value benchmark library, the standard case library, the hierarchical retrieval index, and the global index metadata table are integrated to obtain a historical optimization case library.

[0064] The beneficial effects of the above technical solution are as follows: By deduplicating, removing anomalies, and normalizing the format of historical multimodal generated optimization samples, a standardized case source dataset is obtained. This dataset is then split into scene feature and shortcoming dimension sets. Based on historical optimization effect big data, a two-dimensional feature value benchmark library is constructed. Effective optimization actions are selected, and the comprehensive optimization effectiveness is calculated to establish a shortcoming-optimization parameter association mapping table. Based on the comprehensive value index, high-value cases are selected and standardized to obtain a standard case ontology. Then, a hierarchical retrieval index and a global index metadata table are constructed and integrated to form a historical optimization case library. This can filter invalid noise samples, quantify the relationship between scenes, shortcomings, and optimization actions, retain high-value effective cases, establish a multi-level indexing mechanism, improve the retrieval efficiency and matching accuracy of subsequent case matching, ensure the effectiveness and reliability of reference cases, and provide high-quality data support for subsequent retrieval of suitable optimization parameter sequences.

[0065] Furthermore, such as Figure 4 As shown, optimization actions are filtered for each case in the standardized case source dataset to obtain the effective optimization actions for each case, including: For each case in the standardized case source dataset, a dual-track timeline is constructed; the dual-track timeline includes a short-board fluctuation timeline and an optimization operation timeline. The quality score results after each parameter adjustment in the entire optimization process are mapped to the short board fluctuation time series axis by timestamp, generating multiple short board record items carrying dimension scores and time nodes; Map each parameter adjustment action in the end-to-end optimization operation log to the optimization operation timeline by timestamp, generating multiple operation record items carrying the optimization parameter category and adjustment range; Centered on each operation record item, the short-board record items within the preset effective time window before and after are extracted as the associated observation set. The consistency between the parameter adjustment direction in the associated observation set and the expected change direction of the short-board dimension score in the corresponding short-board dimension set is verified, and the effective optimization actions corresponding to each case are screened out.

[0066] Specifically, a dual-track timeline is constructed for each case in the standardized case source dataset; the dual-track timeline includes a short-board fluctuation timeline and an optimization operation timeline; wherein, a millisecond-level Unix timestamp is used as a unified time base to construct two parallel timelines; the short-board fluctuation timeline stores short-board records in ascending order of time, and each record corresponds to a full-dimensional score of a quality assessment; the optimization operation timeline stores operation records in ascending order of time. The quality score results after each parameter adjustment in the entire optimization process are mapped to the short-board fluctuation timeline by timestamp, generating multiple short-board record items carrying dimension scores and time nodes; each parameter adjustment action in the full-link optimization operation log is mapped to the optimization operation timeline by timestamp, generating multiple operation record items carrying optimization parameter categories and adjustment magnitudes; wherein, the entire optimization process is the complete adjustment process recorded in the full-link optimization operation log from the initial generated content output to the final optimized content generation; the optimization parameters are controllable parameters used to regulate the multimodal generation effect, including rendering parameters, style weight parameters, content constraint parameters, and generation accuracy parameters; each parameter adjustment in the full-link optimization operation log is inserted sequentially into the optimization operation timeline by timestamp; the quality evaluation results after each parameter adjustment in the entire optimization process are inserted sequentially into the short-board fluctuation timeline by timestamp.

[0067] Centered on each operation record item, the short-board record items within the preset effective time window before and after are extracted as the associated observation set. The consistency between the parameter adjustment direction in the associated observation set and the expected change direction of the short-board dimension score in the corresponding short-board dimension set is verified, and the effective optimization actions corresponding to each case are screened out. The preset effective time window is set to 5 seconds before the operation time point to 60 seconds after the operation time point. For each operation record, extract all the weakest record items within the preset effective time window interval before and after the operation as the associated observation set. If the associated observation set is empty, the operation is directly determined to be invalid and removed. The preset parameter-dimensional expected direction mapping table specifies the expected change direction of each optimization parameter for a specific quality dimension. For example, increasing the "rendering resolution" parameter is expected to increase the score of the audiovisual quality dimension; increasing the "compliance constraint strength" parameter is expected to increase the score of compliance effectiveness. Take the dimension score of the first record in the observation set as the initial score and the dimension score of the last record as the final score. The actual change is equal to the final score minus the initial score. If the sign of the actual change is consistent with the expected direction and the absolute change is greater than or equal to the preset minimum effective threshold, it is determined to be a valid optimization action. If the direction is opposite or the change is less than the threshold, it is determined to be an invalid / erroneous operation and removed.

[0068] The beneficial effects of the above technical solution are as follows: It constructs a dual-track timeline for each case, including a short-board fluctuation timeline and an optimization operation timeline. The quality score results and parameter adjustment actions are mapped to short-board record items and operation record items according to timestamps. A set of related observations within an effective time window is extracted, centered on the operation record item. This verifies whether the direction of parameter adjustment is consistent with the expected change direction of the short-board dimension score, thus filtering out effective optimization actions. This approach can identify optimization actions that truly improve quality from a timeline correlation perspective, eliminating invalid and reverse-interference parameter adjustments, preventing non-causal pseudo-correlation optimization actions from being included in subsequent mapping relationships, improving the authenticity of optimization action selection, and ensuring the data credibility of the subsequent short-board-optimization parameter correlation mapping table.

[0069] Furthermore, the comprehensive optimization effectiveness of the effective optimization actions corresponding to each case is calculated, and a bottleneck-optimization parameter association mapping table is generated, including: Extract the temporal steady-state features corresponding to effective optimization actions; Based on the steady-state characteristics of the time series, effective optimization actions are divided into two categories: steady-state optimal adjustment and iterative convergence adjustment, which are respectively assigned a preset first basic weight and a second basic weight; the first basic weight is greater than the second basic weight. The score improvement rate of each short board in the set of short board dimensions corresponding to the effective optimization action is calculated based on the short board record item. For each case, the sum of the products of the base weights of all effective optimization actions for the same shortcoming dimension in the set of shortcoming dimensions and the corresponding shortcoming score improvement rate is calculated to obtain the dimension optimization effectiveness of each shortcoming dimension. Then, based on the shortcoming dimension value coefficients in the dual-dimensional feature value benchmark library, the dimension optimization effectiveness of each shortcoming dimension is weighted and summed to obtain the comprehensive optimization effectiveness of each case. At the same time, based on the parameter-shortcoming correspondence relationship of effective optimization actions, the co-occurrence frequency of each optimization parameter and the corresponding shortcoming dimension is counted to generate a shortcoming-optimization parameter association mapping table.

[0070] Specifically, the temporal steady-state features corresponding to effective optimization actions are extracted. Based on these features, effective optimization actions are categorized into two types: steady-state optimal adjustment and iterative convergence adjustment, which are assigned preset first and second basic weights, respectively. The temporal steady-state features refer to the fluctuation amplitude and sustained stable state characteristics of the corresponding weakness dimension score over time after the optimization operation. The steady-state optimal adjustment is defined as the absence of subsequent adjustments in the same dimension within a preset verification time window after the operation, with the weakness dimension score consistently maintained above the acceptable threshold. The iterative convergence adjustment is defined as the presence of subsequent fine-tuning actions in the same dimension within a preset verification time window after the operation, with the score gradually converging to the acceptable threshold. The above describes the process for extracting steady-state features based on a preset verification time window of 300 seconds. Three steady-state indicators are extracted within this window after the operation. These indicators include the number of subsequent operations in the same dimension, the standard deviation of score fluctuation, and the mean score. The number of subsequent operations in the same dimension is the number of parameter adjustment operations in the same dimension within the window. The standard deviation of score fluctuation is the standard deviation of the score in that dimension within the window. The mean score is the average score in that dimension within the window. The optimal steady-state adjustment satisfies three conditions: the number of subsequent operations in the same dimension = 0, the mean score ≥ 60 points, and the standard deviation of the score ≤ 2 points. This represents achieving the target and stability with a single adjustment, and is assigned a first basic weight of 1.2. The iterative convergence adjustment satisfies two conditions: the number of subsequent operations in the same dimension ≥ 1, and the score monotonically increases after each operation, ultimately reaching a score ≥ 60 points. This represents gradually achieving the target through multiple iterations, and is assigned a second basic weight of 0.8.

[0071] The score improvement rate of each shortcoming in the set of shortcoming dimensions corresponding to the effective optimization action is calculated based on the shortcoming record item; wherein, the shortcoming score improvement rate is the difference between the final score of the shortcoming dimension after parameter adjustment and the score before adjustment, divided by the score before adjustment; for a single effective optimization action, the score improvement rate of the corresponding shortcoming dimension is calculated as follows: ;in, as a shortcoming dimension The rate of improvement in scores; This is the initial score for this dimension before the operation; To determine the final score for this dimension after the operation converges, a minimum value is introduced. To avoid the risk of the denominator being 0.

[0072] For each case, the sum of the products of the base weights of all effective optimization actions for the same shortcoming dimension in the set of shortcoming dimensions and the corresponding shortcoming score improvement rate is calculated to obtain the dimensional optimization effectiveness of each shortcoming dimension. Then, based on the shortcoming dimension value coefficients in the dual-dimensional feature value benchmark library, the dimensional optimization effectiveness of each shortcoming dimension is weighted and summed to obtain the comprehensive optimization effectiveness of each case. Simultaneously, based on the parameter-shortcoming correspondence relationship of effective optimization actions, the co-occurrence frequency of each optimization parameter and the corresponding shortcoming dimension is counted to generate a shortcoming-optimization parameter association mapping table. The single-dimensional optimization effectiveness is calculated by weighting the improvement sum for all effective optimization actions under the same shortcoming dimension. ;in, as a shortcoming dimension Dimensional optimization effectiveness; For the first Each action in dimension Improvement rate of scores on weaknesses; The base weights for the i-th effective optimization action. To optimize the number of actions effectively within this dimension. The overall optimization effectiveness is determined by the value coefficient of the weakest dimension based on the dual-dimensional feature value benchmark library. Weighted summation of validity across all dimensions: ; To optimize the effectiveness of the case study; For the two-dimensional feature value benchmark library, the dimensions The value coefficient; Statistical table structure: ; Traverse all valid optimization actions, and for each occurrence of the correspondence "a certain parameter is used to improve a certain weakness", the corresponding... Add 1, and normalize all parameters under the same weakness dimension according to the number of co-occurrences to obtain the adaptation weight of that parameter for that weakness: ;in, The weights for adapting a certain optimization parameter to a specific weakness dimension; This represents the frequency of co-occurrence between this parameter and this weakness dimension.

[0073] The beneficial effects of the above technical solution are as follows: It extracts the temporal steady-state features of effective optimization actions, divides optimization actions into steady-state optimal adjustments and iterative convergence adjustments, assigns differentiated basic weights, obtains the improvement rate of the short-board score by combining the short-board record items, and obtains the dimensional optimization effectiveness by summing the product of the basic weights and the score improvement rate. Then, it calculates the comprehensive optimization effectiveness of the case based on the short-board dimension value coefficients of the dual-dimensional feature value benchmark library. Simultaneously, it generates a short-board-optimization parameter association mapping table by statistically analyzing the co-occurrence frequency of optimization parameters and short-board dimensions. This allows for the differentiation of the contribution differences of different types of optimization actions, quantitatively assesses the actual improvement degree of each effective optimization action on the short-board dimension, takes into account the value weight of different short-board dimensions themselves, objectively outputs the overall optimization effectiveness of the case, and establishes a statistical correlation between parameters and short-boards, providing a quantifiable basis for subsequent matching of reference cases and retrieval of suitable optimization parameter sequences.

[0074] Furthermore, such as Figure 5 As shown, based on the comprehensive optimization effectiveness, a comprehensive value index is calculated for each case, and each case is value-screened based on the comprehensive value index to obtain high-value cases; combined with the shortcoming-optimization parameter association mapping table, the high-value cases are standardized and structuredly encapsulated to generate a standard case ontology, which is then stored to obtain a standard case library, including: Three core value factors are extracted from each case and normalized in parallel; the three core value factors include scenario adaptation value factors, optimization effectiveness factors, and universally applicable weakness value factors. The comprehensive value index of each case is calculated based on the three normalized core value factors; the comprehensive value index is compared with the preset case entry value threshold, and cases with a comprehensive value index greater than or equal to the entry value threshold are judged as high-value cases. High-value cases are standardized and structured to generate standard case ontology; All standard case ontologies are written in batches to the basic storage layer of the case library to obtain the standard case library.

[0075] Specifically, three types of core value factors are extracted from each case and normalized in parallel. The three types of core value factors include scenario adaptation value factors, optimization effectiveness factors, and universal weakness value factors. Among them, the scenario adaptation value factor is calculated by weighting the scenario feature dimensions in the scenario feature dimension set covered by the case and the corresponding dimension value coefficients in the dual-dimensional feature value benchmark library. ;in, Adapt value factors to different scenarios; The value coefficient of the s-th scenario dimension in the dual-dimensional value benchmark library; The variable is 0-1; it is 1 if the case covers this dimension, and 0 otherwise. Adapt value factors to different scenarios; The optimization efficiency factor is obtained by normalizing the overall optimization efficiency; specifically, the overall optimization efficiency for all cases is normalized using a min-max method: ;in, To optimize the effectiveness factor; To optimize the effectiveness of the current case; , These represent the minimum and maximum values ​​of the overall optimization effectiveness across all cases. To optimize the effectiveness of the current case; The universal value factor for shortcomings is calculated by weighting the shortcomings dimension in the set of shortcomings included in the case and the corresponding dimension value coefficient in the dual-dimensional feature value benchmark library. ;in, As a universally applicable value factor for addressing shortcomings; The variable is 0-1; if the case includes the dimension of the weakness, the value is 1, otherwise it is 0. as a shortcoming dimension The value coefficient.

[0076] The comprehensive value index for each case is calculated based on three normalized core value factors. This comprehensive value index is then compared to a preset case inclusion value threshold; cases with a comprehensive value index greater than or equal to this threshold are classified as high-value cases. The comprehensive value index for each case is as follows: ;in, The comprehensive value index of the case; Adapt value factors to different scenarios; To optimize the effectiveness factor; As a universally applicable value factor for addressing shortcomings; The index weights of value factors are adapted to different scenarios; To optimize the exponential weights of the effectiveness factor; The index weights of universally applicable value factors for addressing shortcomings; pre-set , , ,satisfy >1> > >0, highlighting the core importance of optimization effectiveness. Dynamic verification coefficient. Using a three-stage calculation, if the percentage of invalid operations... And the percentage of dimensions that did not meet the standards : , The number of effective optimization parameters included in the case study; The number of valid optimization parameters included in the case; if the percentage of invalid operations... And the percentage of dimensions that did not meet the standards : In other cases: the dynamic verification coefficients are calculated using a linear interpolation formula, so that... from Smooth to , ;when , When approaching 0, ;when , When it approaches 1, .

[0077] High-value cases are standardized and structurally encapsulated to generate a standard case ontology. This standard case ontology includes a unique case identifier, a scene feature vector group, a quality profile vector group, an optimization parameter sequence, an effect verification label, and a value attribute label. The scene feature vector group consists of second-level sub-dimensional features from the scene feature dimension set. The quality profile vector group consists of the weakness dimension set and the scores for each dimension. The effect verification label includes the overall quality improvement rate after optimization. The value attribute label includes the comprehensive value index. The optimization parameter sequence carries the adaptation weights of each parameter determined based on a weakness-optimization parameter association mapping table.

[0078] All standard case ontologies are written in batches to the basic storage layer of the case library to obtain the standard case library. Specifically, all standard case ontologies are written in batches to MySQL structured storage + Milvus vector storage to form the standard case library.

[0079] The beneficial effects of the above technical solution are as follows: It extracts three core value factors—scene adaptation value factors, optimization effectiveness factors, and universal shortcoming value factors—and performs parallel normalization processing to calculate a comprehensive case value index. By comparing this index with the entry value threshold, high-value cases are selected. These high-value cases are then standardized and structurally encapsulated to generate standard case ontologies, which are then batch-stored into the basic storage layer of the case library to obtain a standard case library. This allows for a comprehensive evaluation of case value from multiple dimensions, including scene adaptation capability, actual optimization effect, and universality of shortcoming reuse. It filters out low-quality, poorly reusable, and invalid cases, achieving standardized and unified encapsulation and storage of cases. This ensures the overall quality and reusability of cases within the standard case library, facilitating subsequent retrieval and improving the reliability of case matching.

[0080] Furthermore, a hierarchical retrieval index and a global index metadata table are constructed based on all standard case ontology in the standard case library, including: Extract the business scenario tags from the scenario feature vector group of each standard case ontology, divide them into first-level categories according to business scenario classes, and assign each business scenario to an independent case sub-library and routing node; the case sub-library is associated with the corresponding standard case ontology through the unique case identifier, and the first-level index layer of the hierarchical retrieval index is determined. Within each case sub-library, the combination of weak dimensions in the quality profile vector group of each standard case ontology is extracted. The state compression code of the combination of weak dimensions is used as the index key to construct a secondary index node. The unique identifiers of standard case ontologs with the same combination of weak dimensions are grouped into the same secondary index node. The unique identifiers of cases in each secondary index node are arranged in descending order according to the comprehensive value index in the value attribute label of the corresponding standard case ontology, thus determining the secondary index layer of the hierarchical retrieval index. Under each secondary index node, the overall quality improvement rate after optimization is extracted from the effect verification tag of the corresponding standard case ontology. The improvement rate is divided into three graded subsets: significant improvement, effective improvement, and slight improvement. During retrieval, the graded subsets are matched in descending order of priority to determine the three-level index layer of the hierarchical retrieval index. A complete hierarchical retrieval index is composed of a first-level index layer, a second-level index layer, and a third-level index layer. Based on the nodes of each level of the hierarchical retrieval index, a global index metadata table is generated, which records the number of standard case entities, average comprehensive value index, and average quality improvement rate of each case sub-database and each second-level index node.

[0081] Specifically, business scenario tags are extracted from the scenario feature vector group of each standard case ontology. These are then divided into primary categories based on business scenario classes, with each business scenario corresponding to an independent case sub-library and routing node. The case sub-library is associated with the corresponding standard case ontology through a unique case identifier, determining the primary index layer of the hierarchical retrieval index. This primary index layer employs a precise routing mechanism, parsing the business scenario tag of the current request and directly routing to the corresponding scenario sub-library for retrieval, avoiding the erroneous logic of traversing all sub-libraries in descending order of global popularity. The primary index layer provides scenario-based routing, dividing the data into 12 case sub-libraries based on major business scenario categories. Each sub-library corresponds to a unique sub-library ID and stores all cases within the corresponding scenario. A primary routing table is constructed with the following structure: Routing priority calculation: Higher call frequency means higher priority, used for load balancing scheduling of multiple sub-libraries in the same scenario; sub-libraries are accessed through... The standard case ontology corresponding to the association mapping.

[0082] Within each case sub-database, the weakness dimension combinations are extracted from the quality profile vector groups of each standard case ontology. Secondary index nodes are constructed using the state compression encoding of these weakness dimension combinations as index keys. Case unique identifiers of standard case ontologs with the same weakness dimension combinations are grouped into the same secondary index node. The case unique identifiers within each secondary index node are sorted in descending order according to the comprehensive value index in the value attribute tags of the corresponding standard case ontology, thus determining the secondary index layer of the hierarchical retrieval index. Specifically, the secondary index layer is a weakness combination state compression index. The weakness dimension combination encoding involves converting the seven dimensions (whether they are weaknesses) into 7-bit binary identifiers (1 represents a weakness, 0 represents no weakness), and then converting these into decimal integers as state compression keys. For example, if the weakness is the 1st and 3rd dimensions, the binary value is 1010000, corresponding to the key value 80. Within each scenario sub-database, an ordered dictionary index is constructed using the key values ​​(the underlying red-black tree structure ensures insertion and query efficiency). Each tree node carries the corresponding weakness combination. List. Within a node The list is sorted in descending order of comprehensive value index, and the search will directly return the top high-value cases.

[0083] Under each secondary index node, the overall quality improvement rate after optimization is extracted from the effect verification tags of the corresponding standard case ontology. This improvement rate is then divided into three tiered subsets: significant improvement, effective improvement, and slight improvement. During retrieval, these subsets are matched sequentially from highest to lowest priority to determine the three-tiered index layer of the hierarchical retrieval index. Within each secondary index node, the overall quality improvement rate is further divided into three partition lists: a significant improvement partition, an effective improvement partition, and a slight improvement partition. A significant improvement partition is defined as an improvement rate ≥ 30%; an effective improvement partition is defined as 10% ≤ improvement rate < 30%; and a slight improvement partition is defined as an improvement rate < 10%. During retrieval, the significant improvement partition is matched first, and if no match is found, the search proceeds downwards sequentially.

[0084] A complete hierarchical retrieval index is composed of a first-level index layer, a second-level index layer, and a third-level index layer. A global index metadata table is generated based on the statistical data of each level of the hierarchical retrieval index. This table records statistical information such as the number of standard case entities, average comprehensive value index, and average quality improvement rate corresponding to each case sub-database and each second-level index node. This information is used for routing decisions and result pre-screening during the retrieval process. The structure of the global index metadata table is as follows:

[0085] Traverse all first-level sub-databases and second-level nodes, count the number of cases, average comprehensive value index, and average improvement rate within each node, and write them to the metadata table; query the metadata table before retrieval to quickly determine the number and quality of cases in a node, skip nodes with no cases or low quality, and realize routing decisions and result pre-screening.

[0086] The beneficial effects of the above technical solution are as follows: First-level index layer division is completed based on business scenario tags, corresponding case sub-libraries and routing nodes are constructed, and then second-level index nodes are built within each sub-library using state compression encoding based on the weakness dimension combination as keywords. Case identifiers are arranged in descending order according to the comprehensive value index. Furthermore, based on the overall quality improvement rate after optimization, three hierarchical subsets are divided to form a third-level index layer, which is then combined to obtain a hierarchical retrieval index. Simultaneously, the information of each level of nodes is statistically analyzed to generate a global index metadata table. This enables multi-level progressive retrieval based on business scenarios, weakness combinations, and optimization improvement effects. The retrieval process prioritizes high-value cases with good optimization effects, narrowing the retrieval scope and reducing matching computation overhead. The global index metadata table facilitates quick understanding of the case distribution and quality statistics of each index node, significantly improving the retrieval speed of subsequent reference case matching and the overall quality of recalled cases.

[0087] Furthermore, the step of constructing a multi-objective comprehensive optimization objective function using dynamically adjusted weights as coefficients, and generating composite optimization instructions by combining the optimization parameter sequence of the reference case, includes: A multi-objective comprehensive optimization objective function is constructed, which includes a first objective function that maximizes the expected benefit of the weakest dimension and a second objective function that minimizes the parameter adjustment cost. The first objective function is the sum of the products of the expected score improvement for each quality weakness dimension and the corresponding dynamic adjustment weight; the expected score improvement is calculated based on the historical average improvement rate of each optimization parameter for the corresponding weakness dimension in the optimization parameter sequence of the reference case; The second objective function is the weighted sum of the deviation penalty term between the current parameter to be optimized and the corresponding parameter in the optimized parameter sequence of the reference case, and the penalty term for the score decay of non-weakness dimension caused by parameter adjustment; The calculation results of the first objective function and the second objective function are normalized to unify the dimensions. Then, based on the dynamically adjusted weights, the standard genetic algorithm is used to solve the normalized multi-objective comprehensive optimization objective function. The optimal parameter combination is obtained by iterative optimization using the optimization parameter sequence of the reference case as the initial solution space. The optimal parameter combination is converted into a format of prompts or control parameters that can be parsed by a multimodal large model, and a composite optimization instruction is generated.

[0088] Specifically, a multi-objective comprehensive optimization objective function is constructed, comprising a first objective function that maximizes the expected gains of the weakest dimension and a second objective function that minimizes the cost of parameter adjustment. The first objective function is the sum of the products of the expected score improvement of each quality weakest dimension and the corresponding dynamic adjustment weight. The expected score improvement is calculated based on the historical average improvement rate of each optimized parameter for the corresponding weakest dimension in the optimization parameter sequence of the reference case. The second objective function is the weighted sum of the deviation penalty term between the current parameter to be optimized and the corresponding parameter in the optimization parameter sequence of the reference case, and the penalty term for score decay in non-weakest dimension caused by parameter adjustment. The first objective function is to maximize the expected gains of the weakest dimension. ;in, Expected returns based on the shortcomings; A set of dimensions representing quality shortcomings; For dimension Dynamic adjustment weights; For dimension The expected increase in score; ; For parameters The adjustment amount, For parameters Dimension Historical average sensitivity; The second objective function is to minimize the cost of parameter adjustment: ;in, The cost of parameter adjustment; , The penalty coefficient weight; This is a penalty item for deviation. , To reference the baseline values ​​of the corresponding parameters in the case study, a minimum value smoothing term is introduced. To avoid a denominator of 0; to measure the degree to which the parameter deviates from the reference case. ; This is a penalty item for side effects; ;in, This represents the expected score decay for non-weakness dimensions. The penalty coefficient for non-weakest link dimensions; , The penalty coefficient is used to avoid the inconsistency in dimensions and the difference in order of magnitude caused by directly adding the deviation penalty term and the score decay penalty term. The sub-terms are first normalized within the second objective function and then weighted and summed.

[0089] The calculation results of the first and second objective functions are normalized to unify the dimensions. Then, based on the dynamically adjusted weights, a standard genetic algorithm is used to solve the normalized multi-objective comprehensive optimization objective function. Using the optimization parameter sequence of the reference case as the initial solution space, the optimal parameter combination is obtained through iterative optimization. The first and second objective functions are then fused into a single-objective fitness function through weights. ;in, The fitness value of the individual parameter; The weight for earnings is set to 0.6 by default. , The expected return and parameter adjustment cost are normalized to the [0,1] interval; respectively for and Perform min-max normalization to the [0,1] interval, use the standard genetic algorithm (SGA), population size 50, maximum number of iterations 100, crossover probability 0.8, mutation probability 0.1.

[0090] An adaptive dynamic initial solution space is constructed centered on the optimization parameter sequence of the reference case. As a benchmark, set the basic fluctuation range. ; Calculate the score for the current content weakness dimension. With the qualified threshold (e.g., 60 points) gap ratio .like (If the score is below 42 points, indicating a serious defect), then the search range coefficient is expanded. ;like ,but ;like ,but (Fine-tuning); The final initial solution space range is: ; This serves as the baseline value for the corresponding parameter in the reference case. The coefficients are adaptive expansion coefficients. To ensure global search capability, the initial population (50 individuals) is generated using a hybrid strategy: 70% of the individuals (35) are randomly generated within the adaptive range calculated above. 30% of the individuals (15) are randomly generated within the globally valid range allowed by the physical parameters. After iteration, the individual with the highest fitness is selected as the optimal parameter combination.

[0091] The optimal parameter combination is transformed into a format of parseable prompts or control parameters for a multimodal large model, generating a composite optimization instruction. Specifically, the optimal parameter combination is mapped to a control instruction that can be parsed by the multimodal large model, output in a JSON structure, including a prompt optimization segment, a model parameter segment, and a multimodal control segment. The prompt optimization segment converts the optimization direction into natural language prompts and appends them to the original request. The model parameter segment adjusts generated parameters such as temperature, top_p, and CFG_scale. The multimodal control segment sets multimodal-specific parameters such as resolution, sampling steps, and style intensity.

[0092] The beneficial effects of the above technical solution are as follows: A multi-objective comprehensive optimization objective function is constructed, comprising a first objective function that maximizes the expected return of the weakest dimension and a second objective function that minimizes the parameter adjustment cost. The expected score improvement of the weakest dimension is calculated based on the historical average improvement rate of reference cases. Simultaneously, a parameter deviation penalty term and a score decay penalty term for non-weakest dimensions are introduced. After normalizing and unifying the dimensions of the dual-objective calculation results, the optimal parameter combination is obtained using the optimized parameter sequence of reference cases as the initial solution space. A standard genetic algorithm is then used to find the optimal solution and convert it into model-recognizable prompts or control parameters to generate composite optimization instructions. This approach prioritizes maximizing the quality improvement gains of the weakest dimension while constraining the parameter adjustment range, avoiding the problem of quality degradation in non-weakest dimensions due to over-optimization. Optimization within the existing case solution space using a genetic algorithm improves the rationality of the parameter combination and ensures that the generated composite optimization instructions can be directly parsed and executed by a multimodal large model, achieving targeted optimization that balances optimization gains and adjustment costs.

[0093] Furthermore, using the optimized parameter sequence of the reference case as the initial solution space, the optimal parameter combination is obtained through iterative optimization, including: An initial parameter population is constructed based on the optimized parameter sequence of the reference case, and the cross-modal parameter sensitivity matrix of each optimized parameter to each quality bottleneck dimension is calculated. During the iterative optimization process, a cross-modal collaborative constraint manifold space is constructed; the initial parameter population is mapped to the cross-modal collaborative constraint manifold space, and the fitness of each parameter individual is calculated within the manifold space; An adaptive step-size update strategy is used to update the positions of individual parameters in the manifold space; When the preset iteration termination condition is met, the parameter individual with the highest fitness in the manifold space is inversely mapped back to the original parameter space through the parameterized inverse function to obtain the optimal parameter combination.

[0094] Specifically, an initial parameter population is constructed based on the optimized parameter sequence of the reference case, and the cross-modal parameter sensitivity matrix of each optimized parameter to each quality bottleneck dimension is calculated. The cross-modal parameter sensitivity matrix represents the rate of change of the partial derivative of the score of each quality bottleneck dimension caused by a unit adjustment of a single optimized parameter. Specifically, based on the optimized parameter sequence of the reference case, Gaussian noise is added to generate 50 individual parameters, forming the initial population, each individual being an M-dimensional parameter vector. An M×N-dimensional sensitivity matrix S is constructed (M is the number of parameters, N is the 7 quality dimensions), with elements... Representing the The first parameter unit adjustment caused by the first The rate of change of scores in each dimension was obtained by averaging historical case data using the finite difference method: ;in, For the first The change in the score of each quality dimension is calculated by subtracting the score of the dimension before adjustment from the dimension score after parameter adjustment. For the first The adjustment amount of each optimization parameter is calculated by subtracting the original value from the adjusted value.

[0095] During the iterative optimization process, a cross-modal collaborative constraint manifold space is constructed. The initial parameter population is mapped to the cross-modal collaborative constraint manifold space, and the fitness of each individual parameter is calculated within the manifold space. The fitness is calculated by the multi-objective comprehensive optimization objective function, with an inter-modal conflict penalty term superimposed. The inter-modal conflict penalty term is calculated based on the product of the cross-modal consistency dimension score decay value caused by parameter adjustment and a preset inter-modal conflict penalty coefficient. Specifically, inherent coupling constraints between multi-modal parameters are embedded, including equality constraints and inequality constraints, for example: equality constraint: style intensity parameter + realism intensity parameter = 1; inequality constraint: rendering resolution ≤ maximum resolution corresponding to the upper limit of video memory. All constraints together constitute a low-dimensional smooth manifold in a high-dimensional parameter space. The manifold parameterization uses intrinsic coordinate parameterization, using... Low-dimensional vectors Representing high-dimensional parameters ,Right now , For smooth mapping functions, for The low-dimensional intrinsic coordinate vector transforms the optimization space from a high-dimensional constrained space into an unconstrained low-dimensional manifold tangent space. The parameterized mapping function... The specific construction methods include using a combination of constraint elimination and projection methods to parameterize constraints in the manifold space, and for equality constraints (such as style intensity)... + Realism ), using variable substitution for elimination. Let the low-dimensional coordinates be... ,but At this point, the mapping function is... In this way, The dimensional parameter space is reduced to Dimension; for inequality constraints (such as rendering resolution) Instead of directly performing parameterized elimination, it uses the projection operator after optimization and update. Force the parameters back into the feasible region. That is, if Then let The final mapping function It consists of the equality constraint variables after elimination and the inequality constraint variables before elimination.

[0096] The fitness formula for each individual parameter is as follows: ;in, , The benefits and costs of normalization; The weight for revenue is set at a preset value of 0.6; inter-modal conflict penalty term: ; This is a penalty term for inter-modal conflict. This represents the expected score decay across the modal consistency dimension. This is the penalty coefficient, with a default value of 1.5.

[0097] An adaptive step-size update strategy is used to update the positions of individual parameters within the manifold space. This strategy involves: calculating the difference in average fitness between the current iteration and the previous iteration; if the difference is greater than a preset exploration threshold, calculating the Levy flight step size within the manifold tangent space and mapping the updated tangent vector back to the manifold space via exponential mapping for global exploration updates; if the difference is less than or equal to a preset development threshold, calculating the Gaussian variable step size within the manifold tangent space and mapping the updated tangent vector back to the manifold space via exponential mapping for local development updates. Specifically, this involves calculating the difference in average fitness between the current generation and the previous generation. The search mode is adaptively switched; the search modes include global exploration mode, balanced exploration mode and local development mode. The global exploration mode is as follows: The step size is generated using Lévy flight, and the Lévy distribution follows... ; Step scaling factor 0.1: Generates a Lévy random step size vector v in the tangent space of the current point; maps the tangent vector back to the manifold space through an exponential mapping. ; For the updated individual parameter positions; For exponential mapping functions, the tangent space vector is mapped back to the manifold space; Let be the perturbation step size vector within the tangent space; it is implemented using the Euler method with numerical integration along the geodesic to ensure that the updated parameters still satisfy all constraints. The exponential mapping is achieved using the Euler method with numerical integration along the geodesic; specifically, the update vector is calculated within the tangent space of the manifold. Then, the first-order Euler method is used to approximate the exponential mapping, with the specific formula as follows: ;in, These are parameter points on the current manifold; For the update vector in the tangent space (such as the Levy flight step size or Gaussian variable step size). The integration step size (learning rate) is preset to 0.05 (range 0.01~0.1), which is used to control the distance traveled along the geodesic line to prevent the step size from being too large and causing overflow of the manifold boundary; This is a projection operator used to project temporary points in Euclidean space. Reprojection back to manifold space (That is, perform the above equality constraint substitution and inequality constraint truncation) to ensure the updated parameters All cross-modal cooperative constraints are strictly satisfied. The equilibrium exploration mode is as follows: The step size is generated by a linear combination of Lévy flight and Gaussian mutation: ;in, For the mixed perturbation step size in the balanced exploration mode; Global exploration step size generated for Levi's flight; The local development step size for Gaussian mutation generation; These are the mixing coefficients used for linear interpolation between exploration and development; A mixed perturbation step size v is generated within the tangent space; similarly, the manifold space is mapped back through an exponential mapping to complete the equilibrium update; the local development mode is as follows: The step size is generated using Gaussian mutation and follows a normal distribution. Generate a Gaussian perturbation step size v in the tangent space; similarly, map back to the manifold space through an exponential mapping to complete a local fine update.

[0098] When the preset iteration termination condition is met, the parameter individual with the highest fitness in the manifold space is inversely mapped back to the original parameter space through a parameterized inverse function to obtain the optimal parameter combination. The iteration terminates when either of the following conditions is met: the maximum number of iterations (200) is reached; or the change in the optimal fitness of the population over 20 consecutive generations is less than 1e-6. The inverse mapping back to the original parameter space is performed by selecting the parameter individual with the highest fitness and directly mapping it back through the parameterized inverse function. This transformation yields the optimal parameter combination in the original parameter space; where, The optimal combination of parameters is used for inverse mapping back to the original parameter space; It is a parameterized inverse function; The parameter individual with the highest fitness within the manifold space.

[0099] The beneficial effects of the above technical solution are as follows: An initial parameter population is constructed based on the optimized parameter sequence of the reference case, and the cross-modal parameter sensitivity matrix of the optimized parameters for each quality bottleneck dimension is calculated. During the iterative optimization phase, a cross-modal collaborative constraint manifold space is constructed, and the initial parameter population is mapped to this manifold space to complete the individual fitness calculation. An adaptive step-size update strategy is used to adjust the position of individual parameters. After the iteration terminates, the parameter individual with the highest fitness in the manifold space is inversely mapped back to the original parameter space through a parameterized inverse function to obtain the optimal parameter combination. This fully considers the sensitivity differences of different parameters to each bottleneck dimension, realizes cross-modal parameter collaborative constraints in the manifold space, avoids the problem of mutual conflict between modes during parameter iteration, and balances the convergence speed and optimization accuracy with the adaptive step-size strategy, effectively improving the cross-modal adaptability of the optimal parameter combination, preventing local optima, and ensuring the actual optimization effect of the output parameters.

[0100] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for content generation quality assessment and optimization based on a multimodal large model, characterized in that, include: The system performs semantic parsing and cross-modal feature extraction on the input multimodal content generation request command to generate a structured feature vector; The Structured feature vectors include business scenario labels, modal attributes, and compliance constraints; The initial generated content output by the multimodal large model is obtained, and the initial generated content is subjected to a full-dimensional quality score to obtain an initial quality score vector including the scores of each dimension, and the quality weakness dimensions below the qualified threshold are identified. Based on structured feature vectors and initial quality score vectors, reference cases with similar scene features and similar quality shortcomings are matched from the historical optimization case library. The optimization parameter sequence of the reference cases is extracted, and the dynamic adjustment weight of each optimization dimension is calculated according to the score distribution of each quality shortcoming dimension of the current content. A multi-objective comprehensive optimization objective function is constructed using dynamically adjusted weights as coefficients. Combined with the optimization parameter sequence of the reference case, a composite optimization instruction is generated to drive the multimodal large model to perform multiple rounds of iterative content optimization and output optimized content that meets the preset quality threshold requirements.

2. The content generation quality assessment and optimization method based on a multimodal large model as described in claim 1, characterized in that, The step of semantically parsing and extracting cross-modal features from the input multimodal content generation request command to generate a structured feature vector includes: The multimodal content generation request instruction is subjected to explicit semantic parsing to extract surface intent features; at the same time, the surface intent features are input into a pre-constructed implicit constraint reasoning graph, and node diffusion reasoning is performed through graph neural network to determine the potential generation constraints not explicitly expressed by the user. Primitive features are extracted from the text modality, image modality, and audio modality of the instruction respectively to obtain the initial feature sequence of each modality; in the initial feature sequence, multiple semantic anchors representing the core semantics are identified based on the cross-modal attention mechanism; Calculate the semantic similarity matrix of semantic anchors between different modalities. When the similarity between cross-modal semantic anchors is detected to be lower than the preset alignment threshold, trigger the modal conflict resolution mechanism to generate conflict-free cross-modal fusion features. The surface intent features, potential generation constraints, and conflict-free cross-modal fusion features are concatenated to obtain an initial structured feature vector; Based on the parsed business scenario labels, the corresponding dimension adjustment coefficients are queried from the pre-set scenario-dimensional weight mapping table; the redundant feature subspace associated with business scenario labels, modal attributes and compliance constraints in the initial structured feature vector is dynamically scaled and dimensionality reduced and reconstructed using the dimension adjustment coefficients to generate a structured feature vector.

3. The content generation quality assessment and optimization method based on a multimodal large model as described in claim 1, characterized in that, Obtain the initial generated content output by the multimodal large model, perform a full-dimensional quality score on the initial generated content to obtain an initial quality score vector including scores for each dimension, and identify the quality weakness dimensions that are below the acceptable threshold, including: Obtain the initial generated content output by the multimodal large model, perform multi-granularity quality feature deconstruction on the initial generated content, and extract single-modal fine-grained features, cross-modal correlation features and compliance-specific features respectively to obtain a full-dimensional quality feature set; Using seven quality dimensions—content relevance, factual accuracy, logical coherence, expression standardization, cross-modal consistency, information timeliness, and compliance effectiveness—as orthogonal coordinate axes, a seven-dimensional quality weight space is constructed. The set of all-dimensional quality features is then matched to the corresponding quality dimensions of the seven-dimensional quality weight space through a dimension-feature mapping relationship. For each quality dimension, principal component analysis is used to determine the optimal projection direction of the feature set under that dimension. The features are mapped along the optimal projection direction to generate a feature point cloud distribution. The Mahalanobis distance from each feature point to the ideal reference point of the corresponding dimension is directly calculated using the inverse matrix of the feature covariance matrix. After normalization, the feature fit of each feature is obtained. The pre-trained random forest feature importance model is called to output the intra-dimensional importance weights of each feature. The feature fit is weighted and summed to obtain the original fit scores for each dimension. The coupling influence coefficient between dimensions is calculated based on the pre-constructed dimensional correlation strength matrix, and the initial baseline weight vector for the scenario is coupled and corrected to obtain the coupled corrected weight vector. Based on the full-dimensional quality feature set, the information entropy of each dimension is calculated. An objective weight vector is generated by the entropy weight method. The modified weight vector and the objective weight vector are fused and coupled according to a preset ratio to obtain the final dimension weight vector. The original fit scores of each dimension are mapped to percentage scores and normalized to generate preliminary dimension scores. Then, the preliminary dimension scores are weighted and calculated in combination with the final dimension weight vector to generate a comprehensive quality score for all dimensions. The initial dimensional score vectors are compared one by one with the corresponding pass thresholds for each dimension, and the dimensions with scores below the pass thresholds are marked as quality weakness dimensions.

4. The content generation quality assessment and optimization method based on a multimodal large model as described in claim 1, characterized in that, Before matching reference cases from the historical optimization case library, the method further includes the step of constructing the historical optimization case library, including: Obtain all historical multimodal generation optimization sample source data, and perform deduplication, outlier removal and format normalization preprocessing on the sample source data to obtain a standardized case source dataset; the standardized case source dataset includes request-side structured feature vector, initial generated content, initial full-dimensional quality score vector, full-link optimization operation log, optimized generated content and optimized full-dimensional quality score vector. The structured feature vector and the initial full-dimensional quality score vector are dimensionally split to obtain the scene feature dimension set and the shortcoming dimension set, respectively. Based on historical full-scale optimization big data, the influence of matching of each dimension in the set of scene feature dimensions and the set of shortcoming dimensions on the optimization effect is statistically analyzed, and a dual-dimensional feature value benchmark library is constructed. For each case in the standardized case source dataset, the optimization actions are filtered to obtain the effective optimization actions corresponding to each case; Calculate the overall optimization effectiveness of the effective optimization actions corresponding to each case, and generate a bottleneck-optimization parameter association mapping table; Based on the comprehensive optimization effectiveness, the comprehensive value index of each case is calculated, and each case is value-screened based on the comprehensive value index to obtain high-value cases; combined with the short-board-optimization parameter association mapping table, the high-value cases are standardized and structured to generate a standard case ontology, which is then stored to obtain a standard case library; A hierarchical search index and a global index metadata table are constructed based on all standard case ontology in the standard case library. By integrating the dual-dimensional feature value benchmark library, standard case library, hierarchical retrieval index, and global index metadata table, a historical optimization case library is obtained.

5. The content generation quality assessment and optimization method based on a multimodal large model as described in claim 4, characterized in that, For each case in the standardized case source dataset, optimization actions are filtered to obtain the effective optimization actions corresponding to each case, including: For each case in the standardized case source dataset, a dual-track timeline is constructed; the dual-track timeline includes a short-board fluctuation timeline and an optimization operation timeline. The quality score results after each parameter adjustment in the entire optimization process are mapped to the short board fluctuation time series axis by timestamp, generating multiple short board record items carrying dimension scores and time nodes; Map each parameter adjustment action in the end-to-end optimization operation log to the optimization operation timeline by timestamp, generating multiple operation record items carrying the optimization parameter category and adjustment range; Centered on each operation record item, the short-board record items within the preset effective time window before and after are extracted as the associated observation set. The consistency between the parameter adjustment direction in the associated observation set and the expected change direction of the short-board dimension score in the corresponding short-board dimension set is verified, and the effective optimization actions corresponding to each case are screened out.

6. The content generation quality assessment and optimization method based on a multimodal large model as described in claim 4, characterized in that, Calculate the overall optimization effectiveness of the effective optimization actions corresponding to each case, and generate a bottleneck-optimization parameter association mapping table, including: Extract the temporal steady-state features corresponding to effective optimization actions; Based on the steady-state characteristics of the time series, effective optimization actions are divided into two categories: steady-state optimal adjustment and iterative convergence adjustment, which are respectively assigned a preset first basic weight and a second basic weight; the first basic weight is greater than the second basic weight. The score improvement rate of each short board in the set of short board dimensions corresponding to the effective optimization action is calculated based on the short board record item. For each case, the sum of the products of the base weights of all effective optimization actions for the same shortcoming dimension in the set of shortcoming dimensions and the corresponding shortcoming score improvement rate is calculated to obtain the dimension optimization effectiveness of each shortcoming dimension. Then, based on the shortcoming dimension value coefficients in the dual-dimensional feature value benchmark library, the dimension optimization effectiveness of each shortcoming dimension is weighted and summed to obtain the comprehensive optimization effectiveness of each case. At the same time, based on the parameter-shortcoming correspondence relationship of effective optimization actions, the co-occurrence frequency of each optimization parameter and the corresponding shortcoming dimension is counted to generate a shortcoming-optimization parameter association mapping table.

7. The content generation quality assessment and optimization method based on a multimodal large model as described in claim 4, characterized in that, Based on the comprehensive optimization effectiveness, a comprehensive value index is calculated for each case, and each case is then value-screened based on the comprehensive value index to obtain high-value cases. Combining this with the shortcoming-optimization parameter association mapping table, the high-value cases are standardized and structurally encapsulated to generate a standard case ontology, which is then stored to obtain a standard case library, including: Three core value factors are extracted from each case and normalized in parallel; the three core value factors include scenario adaptation value factors, optimization effectiveness factors, and universally applicable weakness value factors. The comprehensive value index of each case is calculated based on the three normalized core value factors; the comprehensive value index is compared with the preset case entry value threshold, and cases with a comprehensive value index greater than or equal to the entry value threshold are judged as high-value cases. High-value cases are standardized and structured to generate standard case ontology; All standard case ontologies are written in batches to the basic storage layer of the case library to obtain the standard case library.

8. The content generation quality assessment and optimization method based on a multimodal large model as described in claim 4, characterized in that, A hierarchical search index and a global index metadata table are constructed based on all standard case ontology in the standard case library, including: Extract the business scenario tags from the scenario feature vector group of each standard case ontology, divide them into first-level categories according to business scenario classes, and assign each business scenario to an independent case sub-library and routing node; the case sub-library is associated with the corresponding standard case ontology through the unique case identifier, and the first-level index layer of the hierarchical retrieval index is determined. Within each case sub-library, the combination of weak dimensions in the quality profile vector group of each standard case ontology is extracted. The state compression code of the combination of weak dimensions is used as the index key to construct a secondary index node. The unique identifiers of standard case ontologs with the same combination of weak dimensions are grouped into the same secondary index node. The unique identifiers of cases in each secondary index node are arranged in descending order according to the comprehensive value index in the value attribute label of the corresponding standard case ontology, thus determining the secondary index layer of the hierarchical retrieval index. Under each secondary index node, the overall quality improvement rate after optimization is extracted from the effect verification tag of the corresponding standard case ontology. The improvement rate is divided into three graded subsets: significant improvement, effective improvement, and slight improvement. During retrieval, the graded subsets are matched in descending order of priority to determine the three-level index layer of the hierarchical retrieval index. A complete hierarchical retrieval index is composed of a first-level index layer, a second-level index layer, and a third-level index layer. Based on the nodes of each level of the hierarchical retrieval index, a global index metadata table is generated, which records the number of standard case entities, average comprehensive value index, and average quality improvement rate of each case sub-database and each second-level index node.

9. The content generation quality assessment and optimization method based on a multimodal large model as described in claim 1, characterized in that, The process of constructing a multi-objective comprehensive optimization objective function using dynamically adjusted weights as coefficients, and generating composite optimization instructions by combining the optimization parameter sequence of the reference case, includes: A multi-objective comprehensive optimization objective function is constructed, which includes a first objective function that maximizes the expected benefit of the weakest dimension and a second objective function that minimizes the parameter adjustment cost. The first objective function is the sum of the products of the expected score improvement for each quality weakness dimension and the corresponding dynamic adjustment weight; the expected score improvement is calculated based on the historical average improvement rate of each optimization parameter for the corresponding weakness dimension in the optimization parameter sequence of the reference case; The second objective function is the weighted sum of the deviation penalty term between the current parameter to be optimized and the corresponding parameter in the optimized parameter sequence of the reference case, and the penalty term for the score decay of non-weakness dimension caused by parameter adjustment; The calculation results of the first objective function and the second objective function are normalized to unify the dimensions. Then, based on the dynamically adjusted weights, the standard genetic algorithm is used to solve the normalized multi-objective comprehensive optimization objective function. The optimal parameter combination is obtained by iterative optimization using the optimization parameter sequence of the reference case as the initial solution space. The optimal parameter combination is converted into a format of prompts or control parameters that can be parsed by a multimodal large model, and a composite optimization instruction is generated.

10. The content generation quality assessment and optimization method based on a multimodal large model as described in claim 9, characterized in that, Using the optimized parameter sequence of the reference case as the initial solution space, the optimal parameter combination is obtained through iterative optimization, including: An initial parameter population is constructed based on the optimized parameter sequence of the reference case, and the cross-modal parameter sensitivity matrix of each optimized parameter to each quality bottleneck dimension is calculated. During the iterative optimization process, a cross-modal collaborative constraint manifold space is constructed; the initial parameter population is mapped to the cross-modal collaborative constraint manifold space, and the fitness of each parameter individual is calculated within the manifold space; An adaptive step-size update strategy is used to update the positions of individual parameters in the manifold space; When the preset iteration termination condition is met, the parameter individual with the highest fitness in the manifold space is inversely mapped back to the original parameter space through the parameterized inverse function to obtain the optimal parameter combination.

Images