Multimodal ai vertical class model review control method and system

Abstract

The present application relates to the technical field of multi-modal content review, and more particularly to a multi-modal AI vertical class model review control method and system. Multi-modal content data of a work to be reviewed, target special topic planning intention and historical special topic propagation data are obtained; feature extraction and semantic analysis are performed to obtain work content features, special topic demand features and audience features; the three are mapped to a unified semantic space, and a special topic adaptability representation vector is generated through ternary semantic association calculation; an integrated model containing a multi-path performance prediction layer and a hierarchical review decision layer is constructed, the representation vector is input to predict performance distribution and multi-objective optimization is solved to generate hierarchical review decisions; according to the decision execution control action, the integrated model is updated end-to-end and the mapping parameters are adjusted synchronously by using the deviation between the predicted performance and the actual performance, thereby improving the accuracy and self-adaptive optimization capability of multi-modal content review.

Description

Technical Field

[0001] This invention relates to the field of multimodal content moderation technology, and in particular to a method and system for moderation control of multimodal AI vertical models. Background Technology

[0002] In the field of digital content moderation, existing practices mainly focus on rule-based judgment of single-modal or simple multimodal data. Moderation systems typically rely on pre-defined sensitive word libraries, image classification models, or audio fingerprint databases to independently detect and filter text, images, and audio. These methods depend on manually defined features and thresholds, resulting in relatively fixed moderation standards that struggle to handle complex and ever-changing content scenarios across different categories. When it comes to thematic planning, moderators often manually translate planning intentions into moderation rules, such as setting specific keywords or image tags for a particular topic. There is a lack of systematic correlation modeling between the moderation process and the thematic dissemination goals. Audience characteristics are usually considered indirectly during the operational phase through statistical indicators such as historical click-through rates and dwell time, rather than being actively incorporated during the moderation decision-making stage.

[0003] Existing practices suffer from two significant drawbacks. First, they exhibit fragmentation of the feature space and a semantic gap. Multimodal content data, thematic planning intentions, and audience preferences each reside within different feature representation systems. Current methods merely perform simple splicing or weighted fusion, failing to establish deep semantic connections between these three elements. For instance, visual elements and textual descriptions in a video may imply a potential resonance with a specific thematic audience, but current technologies struggle to capture this cross-modal, cross-dimensional matching information. This leads to review decisions being based solely on surface features, making binary "yes or no" judgments while ignoring the adaptive relationship between content, the thematic focus, and the audience. Second, there is a complete separation between review decisions and dissemination effectiveness prediction. In practice, content review and dissemination effectiveness evaluation are handled independently by different systems or processes. Content that passes review may experience inefficient recommendations or unforeseen risks during its release and dissemination due to its incompatibility with audience interests or thematic needs. Existing methods cannot proactively assess content performance across different dissemination paths during the review phase. Consequently, they cannot dynamically adjust the rigor or stratification of review based on expected effectiveness, often passively accepting the effects or risks of subsequent dissemination. This disconnect means that the review system lacks a closed-loop mechanism for continuous optimization based on actual feedback, resulting in a long-term misalignment between review strategies and communication goals. Summary of the Invention

[0004] This invention provides a method and system for reviewing and controlling multimodal AI vertical models, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a method for reviewing and controlling multimodal AI vertical models, comprising:

[0006] Obtain multimodal content data of works to be reviewed, the planning intent of the target topic, and the dissemination data of historical topics;

[0007] Feature extraction is performed on the multimodal content data to obtain the content features of the work; semantic parsing is performed on the planning intent to obtain the theme demand features; and audience characteristics are obtained by establishing audience preference representations based on the dissemination data of the historical themes.

[0008] The content features of the works, the theme requirements, and the audience features are mapped to a unified semantic space. A deep matching relationship between works, themes, and audiences is established through ternary semantic association calculation, and a theme adaptability representation vector is generated.

[0009] An integrated model for predicting dissemination effectiveness and making review decisions is constructed. The integrated model includes a multi-path effectiveness prediction layer and a hierarchical review decision layer. The topic adaptability representation vector is input into the multi-path effectiveness prediction layer to predict the effectiveness distribution of the work to be reviewed under different dissemination paths. The topic adaptability representation vector and the effectiveness distribution are input into the hierarchical review decision layer for multi-objective optimization solution to generate hierarchical review decisions.

[0010] Based on the hierarchical review decision, corresponding control actions are performed on the works to be reviewed, the deviation between the predicted effectiveness distribution and the actual dissemination performance is calculated, the deviation is used to update the integrated model end-to-end, and the mapping parameters of the ternary semantic association calculation are adjusted synchronously.

[0011] Feature extraction is performed on the multimodal content data to obtain content features of the works; semantic parsing is performed on the planning intent to obtain thematic demand features; and audience characteristics are obtained by establishing audience preference representations based on the dissemination data of the historical themes.

[0012] Cross-modal semantic parsing is performed on multimodal content data. By establishing deep semantic associations between modalities, multidimensional content features of works containing thematic semantics, expression style and emotional tendency are extracted, and a structured representation vector of the multidimensional content features of works is constructed.

[0013] The planning intent is hierarchically semantically decomposed to identify the core theme, target audience, and communication goals within the planning intent, and the core theme, target audience, and communication goals are mapped into constraint expressions of thematic demand characteristics.

[0014] Based on historical thematic dissemination data, an audience behavior analysis framework is constructed. By mining the interactive response patterns and preference evolution of the audience to different works, audience characteristics that characterize the distribution of audience group preferences are generated.

[0015] Semantically align the topic semantics in the structured representation vector with the topic core in the constraint expression to establish the association between the work and the topic in the topic dimension;

[0016] The structured representation vector and the constraint expression are weighted and adjusted using the preference distribution in the audience features to generate audience features.

[0017] Based on historical thematic dissemination data, an audience behavior analysis framework is constructed. By mining the interactive response patterns and preference evolution of audiences to different works, audience characteristics that characterize the distribution of audience preferences are generated, including:

[0018] By extracting audience interaction behavior sequences across different work feature dimensions from historical dissemination data, and analyzing the relationship between behavior intensity changes and feature correlations in these interaction behavior sequences, an audience behavior analysis framework containing a binary mapping between audience and work features is constructed.

[0019] Based on the audience behavior analysis framework, the interaction behavior sequences of the audience are clustered and grouped to identify the differences in response of different audience groups to the features of the work, and the group response feature vector in the interaction response pattern is extracted.

[0020] The group response feature vector is time-series tracked to capture the dynamic change trajectory of different audience groups’ preferences for work features within a continuous thematic period, and a time-series evolution representation that depicts the evolution of preferences is established.

[0021] The group response feature vector and the temporal evolution representation are fused in multiple dimensions. During the fusion process, the group response feature vector is used to locate the current preference focus, and the temporal evolution representation is used to predict the preference trend direction.

[0022] Based on the results of the multi-dimensional fusion, the preference distribution density of different audience groups in the feature space of the work is calculated to generate audience features.

[0023] The content features of the works, the thematic requirements, and the audience features are mapped to a unified semantic space. A deep matching relationship between works, themes, and audiences is established through ternary semantic association calculation, generating a thematic suitability representation vector including:

[0024] Cross-domain semantic alignment is performed on the content features of the work, the theme requirements features, and the audience features. By establishing semantic transformation relationships between feature domains, the three are projected into a unified semantic space, and semantic vectors of the work, theme, and audience are generated in the unified semantic space.

[0025] The matching degree between the work and the topic in the content theme dimension is calculated based on the semantic vector of the work and the semantic vector of the audience. The fit between the work and the audience in the preference response dimension is calculated based on the semantic vector of the work and the semantic vector of the audience. The consistency between the topic and the audience in the target positioning dimension is calculated based on the semantic vector of the topic and the semantic vector of the audience.

[0026] The matching degree, the fit degree, and the consistency degree are constructed into a ternary association structure. By performing multi-level semantic interaction analysis on the ternary association structure, the synergistic enhancement effect and constraint check effect between the work, the topic, and the audience are captured, and a deep matching relationship that describes the deep coupling relationship among the three is established.

[0027] The core semantic factors reflecting the work's suitability to the topic are extracted from the deep matching relationship, and the core semantic factors are vectorized and encoded to generate a topic suitability representation vector.

[0028] The matching degree, the fit degree, and the consistency degree are constructed into a ternary association structure. Through multi-level semantic interaction analysis of the ternary association structure, the synergistic enhancement effect and constraint check effect between the work, the topic, and the audience are captured, and a deep matching relationship that describes the deep coupling relationship among the three is established, including:

[0029] A ternary association structure is constructed by using matching degree, fit degree, and consistency degree as ternary nodes. A bidirectional interaction path is established between nodes in the ternary association structure. Through the bidirectional interaction path, a matching degree-fit degree interaction substructure, a fit degree-consistency interaction substructure, and a matching degree-consistency interaction substructure are constructed to form an association network containing three sets of interaction substructures.

[0030] The first layer of semantic interaction analysis is performed on the aforementioned network. In the matching degree-fit degree interaction substructure, the positive promotion strength of the content of the work meeting the theme requirements on the audience's preference response is calculated. In the fit degree-consistency degree interaction substructure, the gain effect of the synergistic consistency between the audience's preferences and the theme positioning on the dissemination of the work is calculated. In the matching degree-consistency degree interaction substructure, the direct driving force of the matching between the work and the theme on the audience's recognition is calculated, and the synergistic enhancement effect is extracted.

[0031] A second-layer semantic interaction analysis is performed on the aforementioned network. In the three sets of interaction substructures, the inhibitory effect of excessive matching degree optimization on fit degree, the restrictive effect of fit degree bias on consistency degree, and the constraint relationship of consistency degree rigidity on matching degree flexibility are identified, and the constraint and balance effect is extracted.

[0032] The synergistic enhancement effect and the constraint and balancing effect are coupled and modeled to construct a ternary coupling function that simultaneously integrates the enhancement and constraint effects;

[0033] The deep coupling degree between the work, the topic, and the audience is calculated using the ternary coupling function, and the deep coupling degree is used as the core representation of the deep matching relationship.

[0034] The topic-adaptability representation vector is input into the multi-path effectiveness prediction layer to predict the effectiveness distribution of the work to be reviewed under different dissemination paths. The topic-adaptability representation vector and the effectiveness distribution are then input into the hierarchical review decision layer for multi-objective optimization to generate hierarchical review decisions, including:

[0035] The topic-adaptability representation vector is input into the multi-path effectiveness prediction layer. By establishing a nonlinear mapping relationship between the topic-adaptability representation vector and the propagation path features, the propagation effectiveness index of the work to be reviewed under different propagation paths is predicted, and the effectiveness distribution is generated.

[0036] A path-to-path efficiency comparison analysis was performed on the efficiency distribution to extract the efficiency advantages and efficiency constraints of different propagation paths and to construct a path efficiency difference characterization.

[0037] The topic adaptability representation vector, the effectiveness distribution, and the path effectiveness difference representation are input into the hierarchical review decision layer. A multi-objective optimization model with topic adaptability and dissemination effectiveness as dual constraints is established in the hierarchical review decision layer. The topic adaptability representation vector is used as the adaptability constraint term, and the effectiveness distribution is used as the effectiveness constraint term. The multi-objective optimization model is solved to obtain the Pareto optimal solution set.

[0038] Decision schemes that meet different review level standards are extracted from the Pareto optimal solution set. Works are classified based on the adaptability score and the effectiveness score in the decision scheme. Works whose adaptability score and effectiveness score both meet the high-quality entry threshold are included in the high-quality selection decision. Works whose adaptability score or effectiveness score meets the candidate range are included in the candidate retention decision.

[0039] For works that do not meet the candidate retention criteria, optimization suggestions are generated based on the performance constraints in the path performance difference representation, and a hierarchical review decision is generated that includes the selection decision for the best works, the candidate retention decision, and the optimization suggestions.

[0040] Based on the hierarchical review decision, corresponding control actions are executed on the works to be reviewed. The deviation between the predicted effectiveness distribution and the actual dissemination performance is calculated. The deviation is used to update the integrated model end-to-end, and the mapping parameters of the ternary semantic association calculation are adjusted synchronously, including:

[0041] Based on the decisions regarding the selection of high-quality works, the decision to retain candidates, and the optimization suggestions in the tiered review process, release and push control actions, delayed review control actions, and content optimization control actions are executed on the works to be reviewed.

[0042] Collect path-level dissemination data of the work to be reviewed during the actual dissemination process after the execution of corresponding control actions, and construct the actual dissemination performance based on the path-level dissemination data;

[0043] The predicted performance distribution is aligned and compared with the actual propagation performance at the propagation path level. The path performance deviation between the predicted performance index and the actual performance index under different propagation paths is calculated. The path performance deviation is constructed as a deviation that includes path identifier and deviation quantification value.

[0044] The deviation is analyzed for error source tracing to identify the generation level and propagation link of the deviation in the integrated model, and an error backpropagation path is generated from the multi-path performance prediction layer to the topic-adaptive coding layer.

[0045] Based on the error backpropagation path, the deviation is backpropagated to the integrated model, and the prediction parameters of the multi-path effectiveness prediction layer, the decision parameters of the hierarchical review decision layer, and the coding parameters of the topic adaptation coding layer are updated by gradient, and the mapping deviation component related to the ternary semantic association calculation is extracted from the deviation.

[0046] The mapping parameter correction value in the ternary semantic association calculation is calculated using the mapping deviation component, and the mapping parameter correction value is applied to the synchronous adjustment of the matching degree calculation mapping parameter, the fit degree calculation mapping parameter, and the consistency degree calculation mapping parameter.

[0047] A second aspect of this invention provides a multimodal AI vertical model review and control system, comprising:

[0048] The data acquisition unit is used to acquire multimodal content data of works to be reviewed, the planning intent of the target topic, and the dissemination data of historical topics;

[0049] The feature extraction unit is used to extract features from the multimodal content data to obtain content features of the work, perform semantic analysis on the planning intent to obtain thematic demand features, and establish audience preference representation based on the dissemination data of the historical themes to obtain audience features.

[0050] The semantic association unit is used to map the content features of the work, the topic requirements, and the audience features to a unified semantic space, and to establish a deep matching relationship between the work, topic, and audience through ternary semantic association calculation, thereby generating a topic suitability representation vector.

[0051] The prediction and decision-making unit is used to construct an integrated model for predicting dissemination effectiveness and making review decisions. The integrated model includes a multi-path effectiveness prediction layer and a hierarchical review decision-making layer. The topic adaptability representation vector is input into the multi-path effectiveness prediction layer to predict the effectiveness distribution of the work to be reviewed under different dissemination paths. The topic adaptability representation vector and the effectiveness distribution are input into the hierarchical review decision-making layer for multi-objective optimization to generate hierarchical review decisions.

[0052] The feedback update unit is used to perform corresponding control actions on the work to be reviewed based on the hierarchical review decision, calculate the deviation between the predicted effectiveness distribution and the actual dissemination performance, use the deviation to update the integrated model end-to-end, and synchronously adjust the mapping parameters of the ternary semantic association calculation.

[0053] A third aspect of the present invention provides an electronic device, comprising:

[0054] processor;

[0055] Memory used to store processor-executable instructions;

[0056] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0057] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0058] The ternary semantic association calculation of multimodal content features, topic demand features, and audience preferences effectively improves the matching accuracy between work review and topic planning intentions and target audiences. The deep matching relationship established after mapping works, topics, and audiences to a unified semantic space means that the review results no longer rely on a single dimension for judgment, but comprehensively integrate content quality, topic orientation, and dissemination adaptability. The collaborative mechanism of the layered review decision layer and the multi-path effectiveness prediction layer ensures that the review process can not only predict the performance differences under different dissemination paths, but also optimize the balance of multiple objectives such as review pass rate and dissemination effect, thereby significantly reducing the cost of manual intervention and the frequency of re-review.

[0059] The performance distribution output by the multi-path performance prediction layer in the integrated model provides a dynamic basis for tiered review decisions. Review decisions are based on multi-objective optimization solutions using the work-topic-audience suitability representation vector and performance distribution, automatically generating the optimal review strategy. This mechanism avoids the rigidity caused by fixed rules in traditional review processes, allowing review results to adapt to real-time fluctuations in the needs of different topics. The same work can obtain differentiated review results under different topics, significantly improving the flexibility and targeting of review strategies while reducing the risk of incorrect or missed reviews.

[0060] A feedback mechanism that tracks the discrepancy between predicted effectiveness distribution and actual dissemination performance drives continuous end-to-end updates to the integrated model. Synchronous adjustment of mapping parameters in ternary semantic association computation enables the system to adaptively learn new content features, evolving topic trends, and shifts in audience preferences. This closed-loop self-optimizing architecture maintains long-term accuracy in review and timeliness in dissemination prediction without requiring manual reconfiguration of review rules. The topic-specific adaptability representation vector evolves continuously with model updates, further enhancing the foresight and robustness of review decisions. Attached Figure Description

[0061] Figure 1 This is a flowchart illustrating the multimodal AI vertical model review and control method according to an embodiment of the present invention.

[0062] Figure 2 This is a flowchart illustrating the review process for adapting works to specific topics based on ternary semantic association, as described in this embodiment of the invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0065] Figure 1 This is a flowchart illustrating the multimodal AI vertical model review and control method according to an embodiment of the present invention.

[0066] Multimodal AI vertical model review and control methods include:

[0067] Obtain multimodal content data of works to be reviewed, the planning intent of the target topic, and the dissemination data of historical topics;

[0068] Feature extraction is performed on the multimodal content data to obtain the content features of the work; semantic parsing is performed on the planning intent to obtain the theme demand features; and audience characteristics are obtained by establishing audience preference representations based on the dissemination data of the historical themes.

[0069] The content features of the works, the theme requirements, and the audience features are mapped to a unified semantic space. A deep matching relationship between works, themes, and audiences is established through ternary semantic association calculation, and a theme adaptability representation vector is generated.

[0070] An integrated model for predicting dissemination effectiveness and making review decisions is constructed. The integrated model includes a multi-path effectiveness prediction layer and a hierarchical review decision layer. The topic adaptability representation vector is input into the multi-path effectiveness prediction layer to predict the effectiveness distribution of the work to be reviewed under different dissemination paths. The topic adaptability representation vector and the effectiveness distribution are input into the hierarchical review decision layer for multi-objective optimization solution to generate hierarchical review decisions.

[0071] Based on the hierarchical review decision, corresponding control actions are performed on the works to be reviewed, the deviation between the predicted effectiveness distribution and the actual dissemination performance is calculated, the deviation is used to update the integrated model end-to-end, and the mapping parameters of the ternary semantic association calculation are adjusted synchronously.

[0072] In one optional implementation, feature extraction is performed on the multimodal content data to obtain content features of the work, semantic parsing is performed on the planning intent to obtain topic demand features, and audience features are obtained by establishing audience preference representations based on the dissemination data of the historical topics, including:

[0073] Cross-modal semantic parsing is performed on multimodal content data. By establishing deep semantic associations between modalities, multidimensional content features of works containing thematic semantics, expression style and emotional tendency are extracted, and a structured representation vector of the multidimensional content features of works is constructed.

[0074] The planning intent is hierarchically semantically decomposed to identify the core theme, target audience, and communication goals within the planning intent, and the core theme, target audience, and communication goals are mapped into constraint expressions of thematic demand characteristics.

[0075] Based on historical thematic dissemination data, an audience behavior analysis framework is constructed. By mining the interactive response patterns and preference evolution of the audience to different works, audience characteristics that characterize the distribution of audience group preferences are generated.

[0076] Semantically align the topic semantics in the structured representation vector with the topic core in the constraint expression to establish the association between the work and the topic in the topic dimension;

[0077] The structured representation vector and the constraint expression are weighted and adjusted using the preference distribution in the audience features to generate audience features.

[0078] Feature extraction for multimodal content data requires cross-modal semantic parsing of raw data from different modalities, such as images, videos, audio, and text. Each modality's data first undergoes a modality-specific encoder to extract primary feature representations. For example, image content uses a visual encoder to extract visual semantic embeddings, text content uses a language model to extract word semantic context representations, and audio content uses an acoustic feature extractor to obtain acoustic semantic descriptions such as pitch and rhythm. After primary feature extraction, a cross-modal attention mechanism is used to establish deep semantic associations between features from different modalities. This allows visual elements in images to form corresponding semantic bindings with text descriptions, and audio emotional expressions to complement and enhance the emotional tendencies of visual images.

[0079] After establishing cross-modal semantic associations, three dimensions of work content features are extracted from the fused multimodal representation: thematic semantic features, expressive style features, and emotional tendency features. Thematic semantic features describe the core content orientation of the work, such as thematic categories like science and technology, culture, and people's livelihood, and their fine-grained sub-themes; expressive style features depict the expressive characteristics of the work in terms of narrative style, visual composition, and language style, including documentary style, emotional narrative style, and data visualization style; emotional tendency features quantify the work's distribution position in the emotional space, covering emotional polarities and their intensity, such as positive encouragement, neutral objectivity, and critical reflection. These three types of features are concatenated and processed through a structured encoding layer to obtain a structured representation vector of the multidimensional work content features. This vector fully expresses the content attributes of the work within a unified semantic embedding space.

[0080] When performing hierarchical semantic decomposition of planning intent, the planning intent text is viewed as a structured expression containing multiple layers of semantic constraints. Planning intent is typically described in natural language by media editors or content planners, implicitly containing three key types of information: core theme, target audience, and communication objectives. Through semantic role labeling and dependency parsing, the core theme vocabulary cluster is identified from the planning intent text to determine the content focus of this special feature; audience targeting information, including age group, areas of interest, and geographic distribution, is extracted using an audience intent identification model; and communication objectives, such as increasing topic popularity, strengthening brand awareness, and guiding public discussion, are identified through an objective expression extraction model.

[0081] The identified core themes, target audience, and communication objectives are mapped to constraint expressions representing thematic demand characteristics. The core themes are encoded as thematic constraint vectors. This expresses the expected scope of the topic in terms of content theme; audience positioning is encoded as an audience constraint vector. The target audience's attributes and expected characteristics are described; the communication objective is encoded as an effectiveness constraint vector. This describes the quantitative requirements of the topic in terms of its dissemination effect. The three types of constraint vectors together constitute the constraint expression of the topic's demand characteristics. This provides a structured requirement description basis for subsequent matching calculations between works and topics. The advantage of hierarchical decomposition lies in transforming vague natural language planning intentions into computable constraint representations, giving the semantic alignment process a clear optimization objective.

[0082] When constructing an audience behavior analysis framework based on historical thematic dissemination data, multi-dimensional interactive response data such as click-through rate, completion rate, number of reposts, and sentiment tendency of comments for various works within the historical themes are collected. This data is then associated with the content characteristics of the corresponding works and stored to form a historical sample library of "work characteristics - audience behavior." Through a combination of collaborative filtering and sequence modeling, the preference response patterns of different audience groups to different combinations of work characteristics are mined. For example, a certain audience group shows a significantly higher completion rate for works with a strong emotional narrative style and a theme focusing on people's livelihood; this pattern constitutes the preference characteristic of this group. Simultaneously, time-series analysis is used to mine the evolution of audience preferences, identifying the dynamic changes in preferences over time, topic popularity, social events, and other factors. The above analytical results are encoded into an audience characteristic matrix. Each row corresponds to an audience segment, and each column corresponds to a preference weight on a feature dimension of a work, thereby generating an audience feature representation that depicts the distribution of audience group characteristics and preferences.

[0083] In obtaining structured representation vectors With constraint expressions Then, the topic semantic sub-vector in the structured representation vector is combined with the topic core constraint vector in the constraint expression. Semantic alignment is performed. Semantic alignment is achieved by calculating the similarity distribution between the two elements in a shared semantic space. Specifically, a soft alignment mechanism is used instead of hard matching, allowing for semantically similar but differently expressed topics between the work and the topic. A topical semantic graph is introduced as auxiliary knowledge during the alignment process, enabling the semantic distance measurement to consider the hierarchical and co-occurrence relationships between topical concepts, thereby more accurately establishing the association between the work and the topic in the topical dimension. The aligned topic association score serves as one of the inputs for subsequent weighted adjustments, ensuring that works with highly matched topics receive higher relevance weights in the final feature representation.

[0084] Using audience feature matrix The preference distribution in the structured representation vector With constraint expressions When making weighted adjustments, first consider the target audience of the topic. From the audience feature matrix The search retrieves the preference weight distribution that best matches the target audience. Distribute preference weights Acting on structured representation vectors Each dimension of the work is assigned a different audience preference sensitivity, resulting in higher representation weights for feature dimensions that are more popular among the target audience. Similarly, the preference weights are distributed... For constraint expressions Adjustments were made to calibrate the strength of the constraints of thematic requirements with the strength of audience preferences. After audience-driven weighted adjustments, an audience feature representation that integrates the audience preference perspective was obtained. This representation includes both the content attributes of the work itself and the expected preference response of the target audience to the corresponding content attributes, providing a complete and consistent feature input for the deep matching of work-theme-audience in subsequent ternary semantic association calculations.

[0085] The entire feature extraction and demand analysis process forms a complete mapping link from raw data to structured feature representation, ensuring that the three types of heterogeneous information—multimodal work content, special planning intentions, and audience behavioral preferences—can be used for subsequent association calculations and decision inferences within a unified semantic framework.

[0086] In one optional implementation, an audience behavior analysis framework is constructed based on historical thematic dissemination data. By mining the audience's interactive response patterns and preference evolution laws to different works, audience characteristics that characterize the distribution of audience preferences are generated, including:

[0087] By extracting audience interaction behavior sequences across different work feature dimensions from historical dissemination data, and analyzing the relationship between behavior intensity changes and feature correlations in these interaction behavior sequences, an audience behavior analysis framework containing a binary mapping between audience and work features is constructed.

[0088] Based on the audience behavior analysis framework, the interaction behavior sequences of the audience are clustered and grouped to identify the differences in response of different audience groups to the features of the work, and the group response feature vector in the interaction response pattern is extracted.

[0089] The group response feature vector is time-series tracked to capture the dynamic change trajectory of different audience groups’ preferences for work features within a continuous thematic period, and a time-series evolution representation that depicts the evolution of preferences is established.

[0090] The group response feature vector and the temporal evolution representation are fused in multiple dimensions. During the fusion process, the group response feature vector is used to locate the current preference focus, and the temporal evolution representation is used to predict the preference trend direction.

[0091] Based on the results of the multi-dimensional fusion, the preference distribution density of different audience groups in the feature space of the work is calculated to generate audience features.

[0092] Extracting audience interaction sequences from historical dissemination data is the starting point for establishing an audience behavior analysis framework. Dissemination data typically encompasses multi-dimensional interaction records of audiences across various works, including clicks, dwell times, reposts, comments, and favorites. Each record is associated with a feature dimension of the work, including visual style, narrative structure, emotional tone, and content theme. For each audience member, their interaction records across different feature dimensions of different works are organized into an ordered sequence of interaction behaviors. Each element in the sequence not only includes the interaction type but also carries a corresponding work feature tag and interaction intensity value. Changes in behavior intensity are characterized by statistical analysis of the frequency and depth of interactions within a continuous time window. For example, an increase in the duration of an audience's dwell time or a rise in the reposting rate on a certain feature dimension are both identified as positive signals of change in behavior intensity on that dimension. Feature correlations are established by calculating the co-occurrence frequency and conditional probability of different work feature dimensions in the same audience's interaction sequence, thus forming a binary mapping structure between the audience and work features, constituting the core framework of the audience behavior analysis framework.

[0093] After establishing the audience behavior analysis framework, the audience is clustered to identify response differences among different groups. The interaction sequence of each audience member is transformed into a fixed-dimensional feature representation. Specifically, the interaction intensity of each work feature dimension in the sequence is weighted statistically to generate a behavior intensity distribution vector for that audience member across each feature dimension, denoted as [vector]. ,in Number the audience. As clustering input, density-based or distance-based clustering algorithms are used to divide the audience into several groups, with high consistency in the audience's preferences for work features within each group. For each cluster, the weighted mean of the distribution vectors of all audience behavior intensity within the group is calculated to obtain the group response feature vector of that group. ,in Use group number. Each component corresponds to the average group response intensity on a feature dimension of a work, directly reflecting the group's overall preference for different content features. By comparing the responses of different groups... This can clearly identify which characteristic dimensions show significant differences in response among different groups, providing a grouping basis for subsequent preference evolution analysis.

[0094] The dynamic evolution of preferences is an indispensable dimension in audience behavior analysis. Audience preferences for work features are not static but shift with continuous updates to thematic content, changes in social hotspots, and alterations in media consumption habits. For each audience group... Along the continuous thematic cycle Track the trajectory of the change in its group response eigenvector, and denote the first... Group within a cycle The group response feature vector is By analyzing the sequence Temporal modeling is performed to capture the rate and direction of change of preference intensity across various feature dimensions. Temporal modeling can employ recurrent neural networks or temporal convolutional networks to encode the preference evolution trajectory, outputting a temporal evolution representation that characterizes the evolutionary patterns of preferences. . It not only encodes historical trends in preferences but also implicitly predicts future preferences, meaning that based on the known preference focus in the current period, it can estimate the magnitude and direction of preference drift across various feature dimensions in the next period. The introduction of temporal evolution representation makes audience characteristics forward-looking, rather than merely reflecting historical averages.

[0095] group response feature vector With temporal evolution representation Multi-dimensional fusion is a crucial step in generating the final audience characteristics. During the fusion process, It serves to locate the current focus of preference, and the absolute value of each component indicates the current intensity of attention of the group on each feature dimension of the work; This component is responsible for predicting the direction of preference trends. The sign and magnitude of each component represent the strengthening or weakening trend of preference on the corresponding feature dimension and its expected magnitude, respectively. The fusion method employs a gating mechanism, introducing a learnable fusion gate. ,right and The contribution ratios across each feature dimension are dynamically adjusted. The fused representation is... ,in This represents element-wise product. The value of is determined by both the sufficiency of data and the stability of preferences within the current thematic period: when historical data is sufficient and preferences are stable, The component of approaches 1, and the fusion result depends more on the current preference focus; when preferences are in a rapidly changing phase, The component of the fusion factor approaches zero, and the fusion result relies more on the prediction of evolutionary trends. This adaptive fusion mechanism ensures that audience characteristics maintain a reasonable balance between stability and dynamism.

[0096] Based on the results of multi-dimensional fusion Furthermore, the preference distribution density of different audience groups in the feature space of the work is calculated to generate the final audience characteristics. The calculation of preference distribution density uses a fusion representation. Using the mean center as the basis, combined with the distribution vector of individual audience behavior intensity within the group. The dispersion degree is estimated to determine the covariance structure, and a probability density distribution centered on each group is established in the feature space of the work. For any feature point in the feature space of the work, its belonging probability relative to the preference distribution of each audience group can be calculated, thus obtaining the audience coverage potential of that feature point. The preference distribution densities of all groups are integrated into an audience feature matrix. Each row of the matrix corresponds to an audience group, and each column corresponds to a feature dimension of a work. The matrix elements reflect the preference density weight of the group on the corresponding feature dimension. The construction method ensures that it can directly participate in subsequent ternary semantic association calculations, mapping and matching with the content features of the work and the theme requirements in a unified semantic space, providing a quantitative basis for the preference distribution for the deep adaptation assessment between the audience and the work. Through the above complete audience behavior analysis framework, starting from historical communication data, and through a complete link of behavior sequence extraction, clustering, temporal evolution modeling, multi-dimensional fusion to preference distribution density calculation, audience features that can characterize the distribution of audience group preferences are generated, laying the data foundation for the construction of theme adaptation representation vectors.

[0097] In one optional implementation, the content features of the work, the topic-requirement features, and the audience features are mapped to a unified semantic space. A deep matching relationship between the work, topic, and audience is established through ternary semantic association calculation, generating a topic-adaptability representation vector including:

[0098] Cross-domain semantic alignment is performed on the content features of the work, the theme requirements features, and the audience features. By establishing semantic transformation relationships between feature domains, the three are projected into a unified semantic space, and semantic vectors of the work, theme, and audience are generated in the unified semantic space.

[0099] The matching degree between the work and the topic in the content theme dimension is calculated based on the semantic vector of the work and the semantic vector of the audience. The fit between the work and the audience in the preference response dimension is calculated based on the semantic vector of the work and the semantic vector of the audience. The consistency between the topic and the audience in the target positioning dimension is calculated based on the semantic vector of the topic and the semantic vector of the audience.

[0100] The matching degree, the fit degree, and the consistency degree are constructed into a ternary association structure. By performing multi-level semantic interaction analysis on the ternary association structure, the synergistic enhancement effect and constraint check effect between the work, the topic, and the audience are captured, and a deep matching relationship that describes the deep coupling relationship among the three is established.

[0101] The core semantic factors reflecting the work's suitability to the topic are extracted from the deep matching relationship, and the core semantic factors are vectorized and encoded to generate a topic suitability representation vector.

[0102] like Figure 2 As shown, the method includes:

[0103] The content features, thematic demand features, and audience features originate from three heterogeneous feature domains: visual, linguistic, and behavioral. Significant semantic gaps exist among these three domains in the original feature space, making direct correlation computation impossible. To eliminate cross-domain semantic misalignment, an independent semantic projection network is constructed for each feature domain, mapping the content features, thematic demand features, and audience features to a unified semantic space. Specifically, the content features undergo multi-layer nonlinear transformation for semantic enhancement, integrating multimodal components such as visual, textual, and audio features into a unified expression. For thematic demand features, a semantic alignment layer transforms the concatenated expression of theme constraints, audience constraints, and effectiveness constraints into a vector form compatible with the work's semantic space. The group fusion representation encoded in the audience feature matrix is ​​mapped to the same semantic space via the audience domain projection network. During training, the three projection networks collaboratively optimize through shared semantic space metric constraints, ensuring that the work's semantic vectors... Thematic semantic vectors With audience semantic vector All of them are in the same metric space, thus laying the foundation for subsequent ternary semantic association calculations.

[0104] Within a unified semantic space, semantic association metrics are established between pairs of works from three perspectives: content theme dimension, preference response dimension, and target positioning dimension. The degree of matching between works and topics in the content theme dimension is also considered. This reflects whether the content and theme of the work align with the planning direction of the special topic, through calculation. and The semantic similarity between them is obtained, and a topic weight matrix is ​​introduced. Different weights are assigned to different thematic dimensions, allowing the matching degree calculation to focus on the thematic subspaces most relevant to the topic. This assesses the alignment between the work and the audience in terms of preference response. This reflects whether the content style, narrative structure, and emotional tone of a work align with the historical preference distribution of the target audience, through calculation. and The similarity is obtained by projecting it onto the audience preference subspace, which is spanned by the main preference directions extracted from the audience feature matrix. This also determines the consistency between the topic and the audience in the target positioning dimension. This reflects the degree of match between the communication goals of the special topic and the characteristics of the target audience, through calculation. and The degree of semantic alignment between them is obtained, and the focus is on measuring the degree of consistency between the audience profile pre-set by the special topic planning intention and the actual audience group preference representation.

[0105] Will , and Constructed as a ternary association structure This structure not only stores three scalar metrics but also retains intermediate semantic attention weights generated during the calculation of each metric, for subsequent multi-level semantic interaction analysis. Multi-level semantic interaction is performed on the ternary association structure, specifically divided into two stages: a local interaction layer and a global collaboration layer. The local interaction layer performs refined semantic analysis on each of the three pairwise associations, identifying the semantic dimension that contributes most to the association degree in each pair of semantic vectors through an attention mechanism, and extracting local semantic interaction features. , and These correspond to the local deep semantic expressions of three pairs of relationships: content-topic, content-audience, and topic-audience. The global collaboration layer fuses the three local semantic interaction features and captures the high-order synergistic effects among the work, topic, and audience through a multi-head attention mechanism, identifying the triadic synergistic enhancement effect and constraint check effect that cannot be reflected individually in local pairwise associations.

[0106] The synergistic enhancement effect refers to the positive cumulative effect when the content of a work highly matches the topic and also aligns well with the preferences of the target audience. This results in a significantly higher overall suitability than a simple weighted sum of any single binary association. The constraint and balancing effect, on the other hand, refers to the constraint and balancing effect when there are local deviations between the topic's target positioning and the characteristics of the target audience. Even if the work has a high degree of individual match with the topic or the audience, the overall suitability will be constrained, preventing situations where the content is good but the direction is off. The global synergy layer works by analyzing three local interactive features. , , The joint modeling outputs a deep matching representation that includes information on synergistic enhancement and constraint checks and balances. This representation fully depicts the deep coupling relationship between the work, the topic, and the audience, and is a high-order compression and refinement of the ternary semantic association calculation results.

[0107] From deep matching representation The process of extracting core semantic factors employs a semantic factor decomposition mechanism. The input factor extraction network uses a learnable semantic factor selection matrix. Projecting the deep matching representation identifies the core semantic components that contribute most to the topic suitability judgment. The extraction process of core semantic factors also introduces sparsity constraints to ensure high semantic interpretability of the extracted factors and avoid redundancy and overlap between factors. The resulting set of core semantic factors... It covers multiple dimensions of adaptability information, including topic fit factor, audience response expectation factor, communication goal alignment factor, and three-element synergistic gain factor.

[0108] For the core semantic factor set During vectorization encoding, a factor fusion encoder is used to integrate the semantic representations of each factor into a fixed-dimensional topic-adaptive representation vector. During the coding process, the contribution weights of different factors are dynamically adjusted by an adaptive weight allocation mechanism according to the planning intent of the current topic, so that... The system automatically prioritizes the most critical adaptability dimensions across different thematic scenarios. For example, for topics primarily aimed at expanding audience reach, the weights of audience response expectation factors and communication goal alignment factors are increased accordingly; for topics centered on content depth, the weight of thematic relevance factors takes precedence. The resulting thematic adaptability representation vector... As a unified semantic interface, it delivers complete deep matching information of works, topics, and audiences to the downstream multi-path effectiveness prediction layer and hierarchical review decision layer, ensuring that subsequent review decisions are based on sufficient ternary semantic association analysis.

[0109] In one optional implementation, the matching degree, the fit degree, and the consistency degree are constructed into a ternary association structure. By performing multi-level semantic interaction analysis on the ternary association structure, the synergistic enhancement effect and constraint check effect between the work, the topic, and the audience are captured, and a deep matching relationship that describes the deep coupling relationship among the three is established, including:

[0110] A ternary association structure is constructed by using matching degree, fit degree, and consistency degree as ternary nodes. A bidirectional interaction path is established between nodes in the ternary association structure. Through the bidirectional interaction path, a matching degree-fit degree interaction substructure, a fit degree-consistency interaction substructure, and a matching degree-consistency interaction substructure are constructed to form an association network containing three sets of interaction substructures.

[0111] The first layer of semantic interaction analysis is performed on the aforementioned network. In the matching degree-fit degree interaction substructure, the positive promotion strength of the content of the work meeting the theme requirements on the audience's preference response is calculated. In the fit degree-consistency degree interaction substructure, the gain effect of the synergistic consistency between the audience's preferences and the theme positioning on the dissemination of the work is calculated. In the matching degree-consistency degree interaction substructure, the direct driving force of the matching between the work and the theme on the audience's recognition is calculated, and the synergistic enhancement effect is extracted.

[0112] A second-layer semantic interaction analysis is performed on the aforementioned network. In the three sets of interaction substructures, the inhibitory effect of excessive matching degree optimization on fit degree, the restrictive effect of fit degree bias on consistency degree, and the constraint relationship of consistency degree rigidity on matching degree flexibility are identified, and the constraint and balance effect is extracted.

[0113] The synergistic enhancement effect and the constraint and balancing effect are coupled and modeled to construct a ternary coupling function that simultaneously integrates the enhancement and constraint effects;

[0114] The deep coupling degree between the work, the topic, and the audience is calculated using the ternary coupling function, and the deep coupling degree is used as the core representation of the deep matching relationship.

[0115] Matching degree Compatibility With consistency As a ternary node, construct a ternary association structure. The ternary association structure is not simply a concatenation of three scalar values; rather, it treats each node as a structural unit carrying semantic information, with nodes interconnected through bidirectional interaction pathways. Bidirectional interaction pathways mean that nodes... For nodes There is an impact path, and the nodes... For nodes There are also reverse influence paths, with each of the three groups of nodes establishing such bidirectional pathways. This forms three interactive substructures: matching degree-fit degree, fit degree-consistency degree, and matching degree-consistency degree. These three interactive substructures together constitute a complete association network. The topology of this association network retains the inherent logic of interdependence and mutual constraint among the three dimensions, providing a structural foundation for subsequent multi-level semantic interaction analysis.

[0116] In the first-level semantic interaction analysis, the synergistic enhancement effects of the three sets of interaction substructures in the association network are extracted. In the matching-fitness interaction substructure, the focus is on analyzing how the degree to which the content of the work meets the thematic requirements positively promotes the audience's preference response, i.e., when... At a higher level, the work closely aligns with the theme and content of the topic. This alignment further activates the audience's preference and resonance for this type of content, thereby... Achieving positive gain. This positive facilitation strength can be understood as the semantic synergy coefficient between two nodes, achieved by... and Corresponding local semantic interaction features and This is obtained by performing joint attention calculations; specifically, the calculations... exist The intensity of the projection response in the direction quantifies the cooperative direction of the two in the semantic space.

[0117] Within the relevance-consistency interaction substructure, this study analyzes how the synergistic consistency between audience preferences and topic positioning generates a beneficial effect on the dissemination of the work. and Simultaneously, a high level indicates that the work not only reaches the target audience's preferred focus but also aligns its thematic positioning closely with the audience's values. The combined effect of these two factors significantly amplifies the work's dissemination potential. This gain effect is achieved through... and The bilinear interactive operation is quantified. Within the matching-consistency interactive substructure, the direct driving force of the match between the work and the topic on audience acceptance is analyzed. and The combined effect characterizes the degree of alignment between the thematic content framework and the audience's cognitive framework, and the corresponding driving force is through and The weighted inner product operation is used for extraction. The positive cooperative signals extracted from the above three sets of interactive substructures are spliced ​​and fused to form a comprehensive expression of the cooperative enhancement effect, denoted as... This vector characterizes the overall cooperative strength of the ternary nodes in the mutually reinforcing direction.

[0118] The second layer of semantic interaction analysis focuses on the identification and extraction of constraint and balance effects, with a key focus on capturing the negative coupling relationships of mutual inhibition and restriction among ternary nodes. In the matching degree-fit degree interaction substructure, it identifies the inhibitory effect of over-optimization of matching degree on fit degree: when... When overemphasized, the content of the work excessively caters to the formal requirements of the theme, resulting in a highly templated characteristic. This weakens the precise response to the preferences of a specific audience, making... It is suppressed. This suppression is calculated... The portion exceeding the threshold The negative perturbation intensity is used for quantification, and a suppression coefficient is introduced. This indicates the constraint weight in that direction.

[0119] In the fit-consistency interaction substructure, we identify the constraining effect of fit bias on consistency: when a work excessively adapts to the preferences of a specific audience segment, A high degree of specialization can cause a work to deviate from its overall positioning of targeting a broad audience, thus affecting... A constraint is generated, and the effect of this constraint is calculated. right The directional offset is used to characterize the value, and a constraint coefficient is introduced. In the matching-consistency interaction substructure, the constraint relationship between consistency rigidity and matching flexibility is identified: when the positioning consistency between the topic and the audience... When a theme exhibits high rigidity, the constraints imposed by the thematic scope on the content of the work tend to solidify, limiting the space for flexible exploration within the thematic dimension, thus impacting... The adjustable range forms a constraint, and a constraint coefficient is introduced. Integrating the inhibition signals from the three sets of constraint directions forms a comprehensive expression of the constraint and balance effect, denoted as... This vector characterizes the overall constraint strength of the ternary nodes in the mutual constraint directions.

[0120] Synergistic enhancement effect With the effect of constraints and checks and balances Coupling modeling is performed, constructing a ternary coupling function. The design principle of the ternary coupling function is that the enhancing effect and the balancing effect are not simply superimposed, but rather jointly determine the final deep coupling degree through a dynamic trade-off. Specifically, learnable coupling balance parameters are introduced. The ternary coupling function is used to adjust the relative contribution ratio of the enhancement effect and the balancing effect in the coupling calculation. ,in To collaboratively enhance the projection matrix, To constrain the balancing projection matrix, both the enhancement effect and the balancing effect are mapped to a unified coupled representation space. During model training, adaptive learning is performed through end-to-end backpropagation, enabling the model to dynamically adjust the balance between enhancement and restraint based on different thematic scenarios.

[0121] The output vector calculated using the ternary coupling function, after undergoing nonlinear activation processing, serves as a vector representation of the deep coupling degree, denoted as . . Simultaneously encoding the synergistic and mutually restrictive information between the three nodes, it provides a complete depiction of the deep coupling relationship between the work, the topic, and the audience. As the core representation of deep matching relationships, it is input into subsequent topic-specific adaptation representation vectors. The generation process provides deep matching signals with rich semantic layers for predicting dissemination effectiveness and making review decisions. Compared to shallow splicing schemes that only use three scalar values—matching degree, fit degree, and consistency degree—the process built based on a ternary coupling function... It can capture the non-linear interaction patterns among the three, significantly improving the accuracy and robustness of topic suitability judgment.

[0122] In one optional implementation, the topic-adaptability representation vector is input into the multi-path effectiveness prediction layer to predict the effectiveness distribution of the work to be reviewed under different dissemination paths, and the topic-adaptability representation vector and the effectiveness distribution are input into the hierarchical review decision layer for multi-objective optimization to generate hierarchical review decisions, including:

[0123] The topic-adaptability representation vector is input into the multi-path effectiveness prediction layer. By establishing a nonlinear mapping relationship between the topic-adaptability representation vector and the propagation path features, the propagation effectiveness index of the work to be reviewed under different propagation paths is predicted, and the effectiveness distribution is generated.

[0124] A path-to-path efficiency comparison analysis was performed on the efficiency distribution to extract the efficiency advantages and efficiency constraints of different propagation paths and to construct a path efficiency difference characterization.

[0125] The topic adaptability representation vector, the effectiveness distribution, and the path effectiveness difference representation are input into the hierarchical review decision layer. A multi-objective optimization model with topic adaptability and dissemination effectiveness as dual constraints is established in the hierarchical review decision layer. The topic adaptability representation vector is used as the adaptability constraint term, and the effectiveness distribution is used as the effectiveness constraint term. The multi-objective optimization model is solved to obtain the Pareto optimal solution set.

[0126] Decision schemes that meet different review level standards are extracted from the Pareto optimal solution set. Works are classified based on the adaptability score and the effectiveness score in the decision scheme. Works whose adaptability score and effectiveness score both meet the high-quality entry threshold are included in the high-quality selection decision. Works whose adaptability score or effectiveness score meets the candidate range are included in the candidate retention decision.

[0127] For works that do not meet the candidate retention criteria, optimization suggestions are generated based on the performance constraints in the path performance difference representation, and a hierarchical review decision is generated that includes the selection decision for the best works, the candidate retention decision, and the optimization suggestions.

[0128] Topic-adaptability representation vector When inputting the multi-path performance prediction layer, a propagation path feature library needs to be pre-constructed. This feature library covers structured descriptions of various propagation channels, including social media platform paths, video streaming paths, news aggregation platform paths, and vertical community paths. Each propagation path is constructed as a path feature vector based on dimensions such as user activity time distribution, content consumption habits, interaction behavior patterns, and algorithm recommendation mechanisms. ,in Number the propagation paths. The multi-path effectiveness prediction layer establishes... and The nonlinear mapping relationship between them deeply integrates the adaptability semantic information of the work with the propagation mechanism information of the path, predicting the work to be reviewed in the [number]th [stage]. Dissemination effectiveness indicators under different dissemination paths Communication effectiveness indicators It is a multi-dimensional metric, including sub-metrics such as expected exposure, expected interaction rate, expected sharing depth, and audience retention rate, which together constitute the performance vector under this path. The efficiency vectors of all paths are converged to obtain the complete efficiency distribution matrix. , of which Column corresponding to the first The efficiency vector of each propagation path .

[0129] Obtain the performance distribution matrix Subsequently, a comparative analysis of the effectiveness of different propagation paths was conducted. Specifically, the relative strength of different propagation paths across various effectiveness sub-indicators was calculated, and paths that significantly outperformed the average level in one or more sub-indicators were identified. The corresponding feature information was then extracted as an effectiveness advantage feature vector. Simultaneously, paths with significant weaknesses in key performance sub-indicators are identified, and their limiting factors are extracted into a performance limiting factor vector. By concatenating the feature vector of efficiency advantages with the vector of efficiency constraints, a representation of path efficiency differences is constructed. This characterization fully depicts the effectiveness differentiation pattern among different propagation paths, providing a differentiated basis at the path level for subsequent review decisions. The effectiveness comparison between paths not only focuses on absolute numerical differences, but also measures the similarity of paths in effectiveness structure by calculating the cosine distance between path effectiveness vectors, thereby identifying effective complementary path combinations and effective redundant path combinations, providing support for the formulation of multi-path joint propagation strategies.

[0130] Topic-adaptability representation vector Performance distribution matrix and the representation of differences in path effectiveness After the shared input is processed at the hierarchical review and decision-making level, a multi-objective optimization model is established at this level, with both topic suitability and dissemination effectiveness as constraints. This model aims to maximize the topic suitability score. And maximizing the overall communication effectiveness score It is a biobjective function, where Depend on The results, obtained through a compatibility assessment network mapping, reflect the overall compatibility level of the work across three dimensions: content theme, audience fit, and thematic positioning. From the efficiency distribution matrix The efficiency aggregation function is used to calculate the peak efficiency of the work under the optimal propagation path and the efficiency balance under multiple path combinations. Because... and The correlation between the two is not entirely positive—some content has extremely high adaptability but its dissemination effectiveness is low due to path limitations, while some content has outstanding dissemination effectiveness but deviates somewhat from the topic's positioning. Therefore, a Pareto multi-objective optimization framework is used to solve the model, and the Pareto optimal solution set is obtained through iterative search. Each solution in this solution set corresponds to a decision scheme that achieves a specific trade-off between adaptability and effectiveness. No solution can further improve the other objective without sacrificing one objective.

[0131] From the Pareto optimal solution set When extracting decision-making schemes that meet different review level standards, a high-quality product admission threshold is pre-set. With candidate interval For each decision scheme in the solution set, extract its corresponding fitness score. Performance Score Works are classified according to the following rules: When and At the same time, the work was included in the selection decision for outstanding works, meaning that the work met the standards for outstanding works in both the dimensions of thematic suitability and dissemination effectiveness, and could directly enter the thematic release process; when or At least one condition is met, and the score in the other dimension is not lower than At that time, the work is included in the candidate retention decision, meaning that the work has potential in at least one dimension and can enter the candidate pool to await further evaluation or publication after certain conditions are triggered. (High-quality entry threshold) With candidate interval boundary , All parameters are configurable and can be dynamically adjusted according to the quality requirements and dissemination goals of different topics, thereby achieving differentiated review standard management.

[0132] For the fit score Performance Score None of them met the lower bound of the candidate retention criteria. The work is based on the representation of path performance differences. The efficiency constraint vector extracted from Generate targeted optimization suggestions. The logic for generating optimization suggestions is as follows: [The following text appears to be a separate, unrelated section:] ... The constraints of the logo and the characteristics of the work's content were cross-analyzed to identify the specific content attributes that led to insufficient effectiveness. These included insufficient adaptability of the visual presentation to the target dissemination path, low coverage of title semantics and topic keywords, and a mismatch between content rhythm and audience consumption habits. Simultaneously, [the analysis was combined with...] The detailed scores for each sub-dimension identify the specific semantic layers where adaptability shortcomings lie, and provide suggestions for content adjustment, such as supplementing content modules more closely related to the core topic and optimizing the expression density of emotional resonance points with the audience. The optimization suggestions are output in structured text format, including a description of the problem, an explanation of the improvement direction, and an estimated value of the expected performance improvement after the improvements. .

[0133] Finally, the decision sets for selected high-quality works, the decision sets for retaining candidates, and the optimization suggestions for each substandard work are summarized to generate a complete tiered review decision output. This output not only includes clear instructions for each work awaiting review but also an appropriate score to support the decision. Performance score The system provides auxiliary information such as optimal dissemination path recommendations and optimization suggestions, offering actionable review references for content editing teams and enabling a leap from simple pass / rejection judgments to multi-level, refined review management capabilities.

[0134] In one optional implementation, corresponding control actions are performed on the work to be reviewed based on the hierarchical review decision, the deviation between the predicted effectiveness distribution and the actual dissemination performance is calculated, the deviation is used to update the integrated model end-to-end, and the mapping parameters of the ternary semantic association calculation are adjusted synchronously, including:

[0135] Based on the decisions regarding the selection of high-quality works, the decision to retain candidates, and the optimization suggestions in the tiered review process, release and push control actions, delayed review control actions, and content optimization control actions are executed on the works to be reviewed.

[0136] Collect path-level dissemination data of the work to be reviewed during the actual dissemination process after the execution of corresponding control actions, and construct the actual dissemination performance based on the path-level dissemination data;

[0137] The predicted performance distribution is aligned and compared with the actual propagation performance at the propagation path level. The path performance deviation between the predicted performance index and the actual performance index under different propagation paths is calculated. The path performance deviation is constructed as a deviation that includes path identifier and deviation quantification value.

[0138] The deviation is analyzed for error source tracing to identify the generation level and propagation link of the deviation in the integrated model, and an error backpropagation path is generated from the multi-path performance prediction layer to the topic-adaptive coding layer.

[0139] Based on the error backpropagation path, the deviation is backpropagated to the integrated model, and the prediction parameters of the multi-path effectiveness prediction layer, the decision parameters of the hierarchical review decision layer, and the coding parameters of the topic adaptation coding layer are updated by gradient, and the mapping deviation component related to the ternary semantic association calculation is extracted from the deviation.

[0140] The mapping parameter correction value in the ternary semantic association calculation is calculated using the mapping deviation component, and the mapping parameter correction value is applied to the synchronous adjustment of the matching degree calculation mapping parameter, the fit degree calculation mapping parameter, and the consistency degree calculation mapping parameter.

[0141] Based on the different decision types in the tiered review process, differentiated control actions are implemented for works awaiting review. For a "Selected for Excellence" decision, an immediate release and push control action is triggered, pushing the work to the corresponding dissemination path and recording the push timestamp and path identifier. For a "Candidate Retention" decision, a delayed review control action is triggered, temporarily storing the work in a queue awaiting further review, pending supplementary information or a second evaluation before further action. For an "Optimization Suggestion" decision, a content optimization control action is triggered, sending optimization suggestions back to the content creator in a structured feedback format, specifying the content dimensions that need improvement, and setting up a process node for resubmitting for review after optimization. All three types of control actions are executed through a unified control command interface, ensuring that the control process is traceable and reversible.

[0142] Once the work enters the actual dissemination process, path-level dissemination data is continuously collected, including multi-dimensional indicators such as exposure, click-through rate, completion rate, repost rate, comment sentiment distribution, and audience profile coverage for each dissemination path. Path-level dissemination data is organized using the dissemination path identifier as the primary key, with each path maintaining an independent record of actual performance indicators. Based on this path-level dissemination data, actual dissemination performance is constructed according to the same indicator system as the predicted performance distribution, ensuring strict alignment between the prediction and actual sides in terms of indicator dimensions, path granularity, and time windows. This provides a consistent basis for subsequent deviation calculations.

[0143] When aligning and comparing the predicted effectiveness distribution with the actual propagation performance at the propagation path level, the path effectiveness deviation between the predicted effectiveness index and the actual effectiveness index is calculated for each propagation path. Let the first... The prediction efficiency vector of each propagation path is The corresponding actual performance vector is Then the path efficiency deviation vector of this path Defined as ,in Each component corresponds to the prediction bias in different performance dimensions. The bias vectors of all paths are packaged together with the path identifiers to construct a structured bias representation containing both the path identifiers and the quantified bias values. The path identifier is used to locate the source of the deviation propagation in the subsequent error tracing stage, and the deviation quantization value is used to drive the gradient update of the model parameters.

[0144] Structured deviation When conducting error source analysis, the generation level and propagation link of each path deviation within the integrated model are traced. Specifically, it is analyzed whether the deviation mainly originates from the prediction error of the multi-path effectiveness prediction layer, the offset of the decision boundary delineation by the hierarchical review decision layer, or the semantic deviation introduced by the topic adaptation coding layer in the adaptation representation vector generation stage. By calculating the contribution ratio of each layer's output to the final deviation, the dominant level of deviation is identified, and an error backpropagation path is constructed backward along the model's forward computation graph. This path starts from the multi-path effectiveness prediction layer, passes through the hierarchical review decision layer, and traces back layer by layer to the topic adaptation coding layer, forming a complete end-to-end error backpropagation link.

[0145] Based on the established error backpropagation path, the structured deviations are... The parameters are propagated layer by layer to the parameter layers of the integrated model via backpropagation. Gradient updates are performed on the prediction parameters of the multi-path performance prediction layer to more accurately estimate the performance distribution under each path in the next prediction; gradient updates are also performed on the decision parameters of the hierarchical review decision layer to adjust the premium admission threshold. Lower bound of candidate interval upper bound of candidate interval Dynamically setting the decision boundary makes the decision results more consistent with the actual dissemination performance; gradient updates are performed on the coding parameters of the topic-adaptive coding layer to improve the topic-adaptive representation vector. The precision of expressing the matching relationship between work, topic, and audience. During the aforementioned end-to-end update process, simultaneously from... The mapping bias component related to the ternary semantic association calculation is extracted. This component specifically characterizes the bias contribution caused by the inaccuracy of the semantic space mapping parameters. It is decoupled from the performance prediction error and the decision parameter error and is used independently to drive the synchronous adjustment of the mapping parameters.

[0146] After extracting the mapping bias components, the correction values ​​for each mapping parameter in the ternary semantic association calculation are calculated. The correction values ​​for the mapping parameters used to calculate the matching degree are determined based on the content-topic semantic alignment bias, adjusting the content-topic matching degree accordingly. The topic weight matrix during the calculation process This makes the semantic vector of the work With topic semantic vectors Inner product operations in a unified semantic space more accurately reflect the degree of matching between actual content and themes; the correction value of the mapping parameter for the fit calculation is determined based on the content-audience semantic alignment deviation, adjusting the fit between the work and the audience in the preference response dimension. The calculation relies on projection parameters to make the semantic vector of the work... With audience semantic vector The similarity calculation more closely reflects the actual audience response patterns; the correction value of the mapping parameter in the consistency calculation is determined based on the semantic alignment deviation between the topic and the audience, adjusting the consistency between the topic and the audience in the target positioning dimension. The computation depends on the mapping parameters, which make the thematic semantic vectors... With audience semantic vector The correlation metric more accurately captures the actual correspondence between topic positioning and audience expectations. The correction values ​​for the three types of mapping parameters are applied simultaneously to avoid triggering ternary semantic association structures when adjusting a single mapping parameter. Internal semantic inconsistencies are addressed to ensure that the entire ternary matching framework maintains internal consistency during continuous updates.

[0147] The aforementioned end-to-end update mechanism and mapping parameter synchronization adjustment mechanism work together to form a closed-loop online learning process. Whenever new actual propagation data accumulates to a certain batch, a complete process of deviation calculation, error tracing, backpropagation, and parameter update is triggered, enabling the integrated model to continuously adapt to the dynamic changes in the propagation environment and gradually improve prediction accuracy and the quality of review decisions.

[0148] A second aspect of this invention provides a multimodal AI vertical model review and control system, comprising:

[0149] The data acquisition unit is used to acquire multimodal content data of works to be reviewed, the planning intent of the target topic, and the dissemination data of historical topics;

[0150] The feature extraction unit is used to extract features from the multimodal content data to obtain content features of the work, perform semantic analysis on the planning intent to obtain thematic demand features, and establish audience preference representation based on the dissemination data of the historical themes to obtain audience features.

[0151] The semantic association unit is used to map the content features of the work, the topic requirements, and the audience features to a unified semantic space, and to establish a deep matching relationship between the work, topic, and audience through ternary semantic association calculation, thereby generating a topic suitability representation vector.

[0152] The prediction and decision-making unit is used to construct an integrated model for predicting dissemination effectiveness and making review decisions. The integrated model includes a multi-path effectiveness prediction layer and a hierarchical review decision-making layer. The topic adaptability representation vector is input into the multi-path effectiveness prediction layer to predict the effectiveness distribution of the work to be reviewed under different dissemination paths. The topic adaptability representation vector and the effectiveness distribution are input into the hierarchical review decision-making layer for multi-objective optimization to generate hierarchical review decisions.

[0153] The feedback update unit is used to perform corresponding control actions on the work to be reviewed based on the hierarchical review decision, calculate the deviation between the predicted effectiveness distribution and the actual dissemination performance, use the deviation to update the integrated model end-to-end, and synchronously adjust the mapping parameters of the ternary semantic association calculation.

[0154] A third aspect of the present invention provides an electronic device, comprising:

[0155] processor;

[0156] Memory used to store processor-executable instructions;

[0157] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0158] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0159] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0160] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for reviewing and controlling multimodal AI vertical models, characterized in that, include: Obtain multimodal content data of works to be reviewed, the planning intent of the target topic, and the dissemination data of historical topics; Feature extraction is performed on the multimodal content data to obtain the content features of the work; semantic parsing is performed on the planning intent to obtain the theme demand features; and audience preference representation is established based on the dissemination data of the historical themes to obtain audience features. The content features of the works, the theme requirements, and the audience features are mapped to a unified semantic space. A deep matching relationship between works, themes, and audiences is established through ternary semantic association calculation, and a theme adaptability representation vector is generated. An integrated model for predicting dissemination effectiveness and making review decisions is constructed. The integrated model includes a multi-path effectiveness prediction layer and a hierarchical review decision layer. The topic adaptability representation vector is input into the multi-path effectiveness prediction layer to predict the effectiveness distribution of the work to be reviewed under different dissemination paths. The topic adaptability representation vector and the effectiveness distribution are input into the hierarchical review decision layer for multi-objective optimization solution to generate hierarchical review decisions. Based on the hierarchical review decision, corresponding control actions are performed on the works to be reviewed, the deviation between the predicted effectiveness distribution and the actual dissemination performance is calculated, the deviation is used to update the integrated model end-to-end, and the mapping parameters of the ternary semantic association calculation are adjusted synchronously.

2. The method according to claim 1, characterized in that, Feature extraction is performed on the multimodal content data to obtain content features of the works; semantic parsing is performed on the planning intent to obtain thematic demand features; and audience characteristics are obtained by establishing audience preference representations based on the dissemination data of the historical themes. Cross-modal semantic parsing is performed on multimodal content data. By establishing deep semantic associations between modalities, multidimensional content features of works containing thematic semantics, expression style and emotional tendency are extracted, and a structured representation vector of the multidimensional content features of works is constructed. The planning intent is hierarchically semantically decomposed to identify the core theme, target audience, and communication goals within the planning intent, and the core theme, target audience, and communication goals are mapped into constraint expressions of thematic demand characteristics. Based on historical thematic dissemination data, an audience behavior analysis framework is constructed. By mining the interactive response patterns and preference evolution of the audience to different works, audience characteristics that characterize the distribution of audience group preferences are generated. Semantically align the topic semantics in the structured representation vector with the topic core in the constraint expression to establish the association between the work and the topic in the topic dimension; The structured representation vector and the constraint expression are weighted and adjusted using the preference distribution in the audience features to generate audience features.

3. The method according to claim 2, characterized in that, A framework for audience behavior analysis is constructed based on historical thematic dissemination data. By mining the interactive response patterns and preference evolution of audiences to different works, audience characteristics that characterize the distribution of audience preferences are generated, including: By extracting audience interaction behavior sequences across different work feature dimensions from historical dissemination data, and analyzing the relationship between behavior intensity changes and feature correlations in these interaction behavior sequences, an audience behavior analysis framework containing a binary mapping between audience and work features is constructed. Based on the audience behavior analysis framework, the interaction behavior sequences of the audience are clustered and grouped to identify the differences in response of different audience groups to the features of the work, and the group response feature vector in the interaction response pattern is extracted. The group response feature vector is time-series tracked to capture the dynamic change trajectory of different audience groups’ preferences for work features within a continuous thematic period, and a time-series evolution representation that depicts the evolution of preferences is established. The group response feature vector and the temporal evolution representation are fused in multiple dimensions. During the fusion process, the group response feature vector is used to locate the current preference focus, and the temporal evolution representation is used to predict the preference trend direction. Based on the results of the multi-dimensional fusion, the preference distribution density of different audience groups in the feature space of the work is calculated to generate audience features.

4. The method according to claim 1, characterized in that, The content features of the works, the thematic requirements, and the audience features are mapped to a unified semantic space. A deep matching relationship between works, themes, and audiences is established through ternary semantic association calculation, generating a thematic suitability representation vector including: Cross-domain semantic alignment is performed on the content features of the work, the theme requirements features, and the audience features. By establishing semantic transformation relationships between feature domains, the three are projected into a unified semantic space, and semantic vectors of the work, theme, and audience are generated in the unified semantic space. The matching degree between the work and the topic in the content theme dimension is calculated based on the semantic vector of the work and the semantic vector of the audience. The fit between the work and the audience in the preference response dimension is calculated based on the semantic vector of the work and the semantic vector of the audience. The consistency between the topic and the audience in the target positioning dimension is calculated based on the semantic vector of the topic and the semantic vector of the audience. The matching degree, the fit degree, and the consistency degree are constructed into a ternary association structure. By performing multi-level semantic interaction analysis on the ternary association structure, the synergistic enhancement effect and constraint check effect between the work, the topic, and the audience are captured, and a deep matching relationship that describes the deep coupling relationship among the three is established. The core semantic factors reflecting the work's suitability to the topic are extracted from the deep matching relationship, and the core semantic factors are vectorized and encoded to generate a topic suitability representation vector.

5. The method according to claim 4, characterized in that, The matching degree, the fit degree, and the consistency degree are constructed into a ternary association structure. Through multi-level semantic interaction analysis of the ternary association structure, the synergistic enhancement effect and constraint check effect between the work, the topic, and the audience are captured, and a deep matching relationship that describes the deep coupling relationship among the three is established, including: A ternary association structure is constructed by using matching degree, fit degree, and consistency degree as ternary nodes. A bidirectional interaction path is established between nodes in the ternary association structure. Through the bidirectional interaction path, a matching degree-fit degree interaction substructure, a fit degree-consistency interaction substructure, and a matching degree-consistency interaction substructure are constructed to form an association network containing three sets of interaction substructures. The first layer of semantic interaction analysis is performed on the aforementioned network. In the matching degree-fit degree interaction substructure, the positive promotion strength of the content of the work meeting the theme requirements on the audience's preference response is calculated. In the fit degree-consistency degree interaction substructure, the gain effect of the synergistic consistency between the audience's preferences and the theme positioning on the dissemination of the work is calculated. In the matching degree-consistency degree interaction substructure, the direct driving force of the matching between the work and the theme on the audience's recognition is calculated, and the synergistic enhancement effect is extracted. A second-layer semantic interaction analysis is performed on the aforementioned network. In the three sets of interaction substructures, the inhibitory effect of excessive matching degree optimization on fit degree, the restrictive effect of fit degree bias on consistency degree, and the constraint relationship of consistency degree rigidity on matching degree flexibility are identified, and the constraint and balance effect is extracted. The synergistic enhancement effect and the constraint and balancing effect are coupled and modeled to construct a ternary coupling function that simultaneously integrates the enhancement and constraint effects; The deep coupling degree between the work, the topic, and the audience is calculated using the ternary coupling function, and the deep coupling degree is used as the core representation of the deep matching relationship.

6. The method according to claim 1, characterized in that, The topic-adaptability representation vector is input into the multi-path effectiveness prediction layer to predict the effectiveness distribution of the work to be reviewed under different dissemination paths. The topic-adaptability representation vector and the effectiveness distribution are then input into the hierarchical review decision layer for multi-objective optimization to generate hierarchical review decisions, including: The topic-adaptability representation vector is input into the multi-path effectiveness prediction layer. By establishing a nonlinear mapping relationship between the topic-adaptability representation vector and the propagation path features, the propagation effectiveness index of the work to be reviewed under different propagation paths is predicted, and the effectiveness distribution is generated. A path-to-path efficiency comparison analysis was performed on the efficiency distribution to extract the efficiency advantages and efficiency constraints of different propagation paths and to construct a path efficiency difference characterization. The topic adaptability representation vector, the effectiveness distribution, and the path effectiveness difference representation are input into the hierarchical review decision layer. A multi-objective optimization model with topic adaptability and dissemination effectiveness as dual constraints is established in the hierarchical review decision layer. The topic adaptability representation vector is used as the adaptability constraint term, and the effectiveness distribution is used as the effectiveness constraint term. The multi-objective optimization model is solved to obtain the Pareto optimal solution set. Decision schemes that meet different review level standards are extracted from the Pareto optimal solution set. Works are classified based on the adaptability score and the effectiveness score in the decision scheme. Works whose adaptability score and effectiveness score both meet the high-quality entry threshold are included in the high-quality selection decision. Works whose adaptability score or effectiveness score meets the candidate range are included in the candidate retention decision. For works that do not meet the candidate retention criteria, optimization suggestions are generated based on the performance constraints in the path performance difference representation, and a hierarchical review decision is generated that includes the selection decision for the best works, the candidate retention decision, and the optimization suggestions.

7. The method according to claim 1, characterized in that, Based on the hierarchical review decision, corresponding control actions are executed on the works to be reviewed. The deviation between the predicted effectiveness distribution and the actual dissemination performance is calculated. The deviation is used to update the integrated model end-to-end, and the mapping parameters of the ternary semantic association calculation are adjusted synchronously, including: Based on the decisions regarding the selection of high-quality works, the decision to retain candidates, and the optimization suggestions in the tiered review process, release and push control actions, delayed review control actions, and content optimization control actions are executed on the works to be reviewed. Collect path-level dissemination data of the work to be reviewed during the actual dissemination process after the execution of corresponding control actions, and construct the actual dissemination performance based on the path-level dissemination data; The predicted performance distribution is aligned and compared with the actual propagation performance at the propagation path level. The path performance deviation between the predicted performance index and the actual performance index under different propagation paths is calculated. The path performance deviation is constructed as a deviation that includes path identifier and deviation quantification value. The deviation is analyzed for error source tracing to identify the generation level and propagation link of the deviation in the integrated model, and an error backpropagation path is generated from the multi-path performance prediction layer to the topic-adaptive coding layer. Based on the error backpropagation path, the deviation is backpropagated to the integrated model, and the prediction parameters of the multi-path effectiveness prediction layer, the decision parameters of the hierarchical review decision layer, and the coding parameters of the topic adaptation coding layer are updated by gradient, and the mapping deviation component related to the ternary semantic association calculation is extracted from the deviation. The mapping parameter correction value in the ternary semantic association calculation is calculated using the mapping deviation component, and the mapping parameter correction value is applied to the synchronous adjustment of the matching degree calculation mapping parameter, the fit degree calculation mapping parameter, and the consistency degree calculation mapping parameter.

8. A multimodal AI vertical model review and control system, used to implement the method as described in any one of claims 1-7, characterized in that, include: The data acquisition unit is used to acquire multimodal content data of works to be reviewed, the planning intent of the target topic, and the dissemination data of historical topics; The feature extraction unit is used to extract features from the multimodal content data to obtain content features of the work, perform semantic analysis on the planning intent to obtain thematic demand features, and establish audience preference representation based on the dissemination data of the historical themes to obtain audience features. The semantic association unit is used to map the content features of the work, the topic requirements, and the audience features to a unified semantic space, and to establish a deep matching relationship between the work, topic, and audience through ternary semantic association calculation, thereby generating a topic suitability representation vector. The prediction and decision-making unit is used to construct an integrated model for predicting dissemination effectiveness and making review decisions. The integrated model includes a multi-path effectiveness prediction layer and a hierarchical review decision-making layer. The topic adaptability representation vector is input into the multi-path effectiveness prediction layer to predict the effectiveness distribution of the work to be reviewed under different dissemination paths. The topic adaptability representation vector and the effectiveness distribution are input into the hierarchical review decision-making layer for multi-objective optimization to generate hierarchical review decisions. The feedback update unit is used to perform corresponding control actions on the work to be reviewed based on the hierarchical review decision, calculate the deviation between the predicted effectiveness distribution and the actual dissemination performance, use the deviation to update the integrated model end-to-end, and synchronously adjust the mapping parameters of the ternary semantic association calculation.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images