Multimodal work review and propaganda control method and system

Abstract

The present application relates to the technical field of work review, and more particularly to a multi-modal work review and promotion control method and system. The method obtains multi-modal input data of a work to be reviewed, establishes a complementary verification mechanism between modes through deep semantic analysis to generate a content risk assessment vector, and then constructs a risk propagation influence model to predict scenario-based risks. Accordingly, a gradual promotion control scheme is constructed by binding the display permission and the propagation time window, and a distribution instruction is output while monitoring public opinion feedback data in real time. The present application realizes accurate risk assessment and dynamic promotion control of the content of the work.

Description

Technical Field

[0001] This invention relates to the field of works review technology, and in particular to a method and system for reviewing and controlling the promotion of multimodal works. Background Technology

[0002] In the traditional field of multimedia content moderation and dissemination control, existing technologies typically rely on single-modal analysis methods for content filtering. Keyword matching or semantic analysis models are used for textual works, feature recognition or classifiers for image works, while audio and video are processed independently. The moderation processes for each modality are isolated, lacking information exchange and collaborative verification mechanisms. This isolated approach leads to significant uncertainty in the moderation results, as single-modal data is susceptible to noise interference, ambiguous expressions, or adversarial samples. For example, text containing metaphors or irony might be misjudged as inappropriate when separated from an image background, while a seemingly normal image might convey risky information when paired with specific audio; however, traditional methods struggle to capture such cross-modal semantic connections.

[0003] In terms of risk assessment and dissemination control, existing strategies mostly employ globally uniform thresholds or rules for judgment. Once a work is identified as high-risk, a full ban or traffic restriction is implemented; if it is determined to be low-risk, dissemination rights are completely unrestricted. This approach ignores the differentiated risk characteristics of works in different dissemination scenarios. The dissemination path and audience sensitivity response patterns of the same work on short video platforms are drastically different from those on social media platforms, but existing solutions lack dynamic modeling of scenario-based risk diffusion paths. Control measures are therefore caught in a dilemma: overly strict measures inhibit the dissemination efficiency of normal content, while overly lenient measures allow potential risks to spread unchecked.

[0004] The aforementioned deficiencies manifest primarily as an imbalance between the accuracy of review and the flexibility of dissemination. The accumulation of uncertainty in a single modality renders risk assessment unreliable, while static control strategies are unable to adapt to the ever-changing dissemination environment. Therefore, a method capable of integrating multimodal information, dynamically assessing scenario-based risks, and implementing adaptive dissemination control has become an urgent technological need. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for reviewing and controlling the promotion of multimodal works, which can solve the problems in the prior art.

[0006] A first aspect of the present invention provides a method for reviewing and controlling the promotion of multimodal works, comprising:

[0007] Obtain multimodal input data for works awaiting review;

[0008] Deep semantic parsing is performed on the multimodal input data. By establishing a complementary verification mechanism between modalities, the uncertainty judgment result of a single modality is cross-validated with the deterministic features of other modalities to generate a content risk assessment vector after multimodal collaborative verification.

[0009] Based on the content risk assessment vector, a risk propagation impact model is constructed. By analyzing the risk diffusion path and audience sensitivity response pattern of the work content in different propagation scenarios, the potential impact range of the work to be reviewed in each propagation scenario is predicted, and scenario-based risk prediction results are obtained.

[0010] Based on the scenario-based risk prediction results, a progressive publicity control scheme is constructed. The progressive publicity control scheme associates and binds the display rights of the work with the dissemination time window by setting phased release conditions for the dissemination path.

[0011] The progressive publicity control scheme is implemented to output distribution instructions to the dissemination execution system and to monitor public opinion feedback data in real time during the dissemination process.

[0012] Deep semantic parsing is performed on the multimodal input data. By establishing a complementary verification mechanism between modalities, the uncertainty judgment result of a single modality is cross-validated with the deterministic features of other modalities to generate a content risk assessment vector after multimodal collaborative verification, including:

[0013] Semantic parsing is performed on each modality of the multimodal input data to extract features that represent the semantic content and construct the semantic feature space of each modality.

[0014] A semantic association graph between modalities is established in the semantic feature space. By identifying the semantic support relationships and semantic contradiction relationships of different modal data, the semantic association state between modalities is determined.

[0015] Risk identification is performed on each modality of data. Based on the identification confidence level, the identification results are divided into deterministic features and uncertain judgment results. Other modality data associated with the uncertain judgment results are located through the semantic association graph. Deterministic features that are semantically related to the uncertain judgment results are extracted from the other modality data as verification criteria.

[0016] Based on the verification criteria, the confidence level of the uncertainty judgment result is corrected. When the certainty feature is consistent with the risk direction of the uncertainty judgment result, the confidence level is increased. When the risk direction is conflicting, the confidence level is decreased to obtain the corrected risk identification result. The corrected risk identification result is then vectorized and fused to generate a content risk assessment vector.

[0017] In the semantic feature space, a semantic association graph between modalities is established. By identifying the semantic support relationships and semantic contradiction relationships of different modal data, the semantic association state between modalities is determined, including:

[0018] In the semantic feature space, semantic deconstruction is performed on each modality data to extract semantic topics and semantic tendencies. The semantic topics represent the target object to which the content points.

[0019] Based on the semantic theme, a semantic association mapping is established between different modal data. Modal data expressing the same semantic theme are associated and labeled to construct an initial semantic association graph.

[0020] Modal data pairs are obtained through the initial semantic association graph, and a semantic tendency polarity mapping relationship is established for the modal data pairs. The consistency or opposition of semantic tendencies is determined through the polarity mapping relationship. When the consistency is determined, it is marked as a semantic support relationship; when the opposition is determined, it is marked as a semantic contradiction relationship.

[0021] A relation constraint network is constructed based on the semantic support relationship and the semantic contradiction relationship. A semantic enhancement path is established through the semantic support relationship, and a semantic conflict path is established through the semantic contradiction relationship.

[0022] The semantic association state is determined based on the connectivity structure of semantic enhancement paths and semantic conflict paths in the relation constraint network. When a connectivity structure is formed through semantic enhancement paths, it is determined to be a cooperative semantic association state, and when a connectivity structure is formed through semantic conflict paths, it is determined to be an adversarial semantic association state.

[0023] Based on the content risk assessment vector, a risk propagation impact model is constructed. By analyzing the risk diffusion path and audience sensitivity response patterns of the work content in different propagation scenarios, the potential impact range of the work to be reviewed in each propagation scenario is predicted, resulting in scenario-based risk prediction results, including:

[0024] Based on the content risk assessment vector, the risk propagation potential of the work to be reviewed is calculated, the characteristics of the propagation medium and the audience distribution characteristics of different propagation scenarios are obtained, and a scenario-based propagation resistance model is constructed based on the characteristics of the propagation medium.

[0025] The risk propagation potential energy is coupled with the propagation resistance model for analysis. The risk diffusion path of the work to be reviewed is determined by the antagonistic relationship between the risk propagation potential energy and the propagation resistance model, and a risk propagation dynamic structure is constructed.

[0026] An audience sensitivity distribution map is established based on the audience distribution characteristics. The audience sensitivity distribution map represents the spatial distribution of the audience's sensitivity to the risky content of the work to be reviewed in different communication scenarios. The audience sensitivity distribution map is used to mark the sensitive areas of the risk propagation dynamics structure and determine the target sensitive propagation areas of the work to be reviewed in each communication scenario.

[0027] Based on the spatial superposition structure of the risk diffusion path of the work to be reviewed and the target sensitive propagation area in the risk propagation dynamics structure, the potential impact range of the work to be reviewed in each propagation scenario is determined, and scenario-based risk prediction results are generated.

[0028] The risk propagation potential energy is coupled with the propagation resistance model for analysis. The antagonistic relationship between the risk propagation potential energy and the propagation resistance model is used to determine the risk diffusion path of the work under review, thus constructing a risk propagation dynamics structure, including:

[0029] Calculate the energy distribution gradient of the risk propagation potential energy and the resistance distribution gradient of the propagation resistance model, extract the difference features between the energy distribution gradient and the resistance distribution gradient, and identify the dominant propagation channel based on the difference features;

[0030] The attenuation trajectory of the risk propagation potential energy and the cumulative trajectory of the propagation resistance model are traced along the dominant propagation channel. When the attenuation trajectory and the cumulative trajectory intersect, the risk propagation critical point is determined.

[0031] The path nodes from the starting region to the risk propagation critical point are connected to form the risk diffusion path of the work to be reviewed. The number of propagation levels traversed by the risk diffusion path is counted to obtain the level depth. The number of branches of the risk diffusion path at each propagation level is counted to obtain the branch density.

[0032] Based on the hierarchical depth, the longitudinal propagation range of the risk diffusion path is marked in the propagation resistance model, and based on the branch density, the lateral diffusion intensity of the risk diffusion path is marked, thereby generating a risk propagation dynamic structure.

[0033] Based on the scenario-based risk prediction results, a progressive publicity control scheme is constructed. This scheme links the display permissions of the work with the dissemination time window by setting phased release conditions for the dissemination path, including:

[0034] The potential impact range and risk level of each communication scenario are extracted from the scenario-based risk prediction results. The communication scenarios are sorted and mapped to communication paths based on the risk level. The audience reach limit is set for each communication scenario on the communication path according to the potential impact range.

[0035] For the propagation path, scene entry conditions and scene exit conditions are set. The scene entry conditions are associated with the cumulative propagation scale and risk event frequency of the preceding propagation scene. The scene exit conditions are associated with the propagation duration of the current propagation scene and the audience reach limit. Scene switching rules are constructed based on the scene entry conditions and scene exit conditions.

[0036] Time windows are allocated to each dissemination scenario and divided into a permission restriction period and a permission open period. During the permission restriction period, the display permission is limited to the initial range of the audience reach limit. During the permission open period, the display permission is extended to the full range of the audience reach limit according to the execution result of the scenario switching rule.

[0037] The propagation path, the scene switching rules, and the time window are combined to generate a progressive propaganda control scheme.

[0038] A second aspect of the present invention provides a multimodal work review and promotion control system, comprising:

[0039] The data acquisition unit is used to acquire multimodal input data of the works to be reviewed;

[0040] The risk assessment unit is used to perform deep semantic parsing on the multimodal input data. By establishing a complementary verification mechanism between modalities, the uncertainty judgment result of a single modality is cross-verified with the deterministic features of other modalities to generate a content risk assessment vector after multimodal collaborative verification.

[0041] The scenario prediction unit is used to construct a risk propagation impact model based on the content risk assessment vector. By analyzing the risk diffusion path and audience sensitivity response pattern of the work content in different propagation scenarios, it predicts the potential impact range of the work to be reviewed in each propagation scenario and obtains scenario-based risk prediction results.

[0042] The control scheme unit is used to construct a progressive publicity control scheme based on the scenario-based risk prediction results. The progressive publicity control scheme associates and binds the display permissions of the work with the dissemination time window by setting phased release conditions for the dissemination path.

[0043] The distribution monitoring unit is used to output distribution instructions to the dissemination execution system in accordance with the progressive publicity control scheme, and to monitor the public opinion feedback data of the work in real time during the dissemination process.

[0044] A third aspect of the present invention provides an electronic device, comprising:

[0045] processor;

[0046] Memory used to store processor-executable instructions;

[0047] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0048] 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.

[0049] Deep semantic parsing of multimodal input data and complementary verification mechanisms between modalities significantly improve the accuracy and robustness of content risk assessment. Ambiguous or missing information in a single modality is effectively eliminated through cross-validation with deterministic features from other modalities, thus eliminating the risk of misjudgment and omission, enabling the content risk assessment vector to truly reflect the potential harm intensity of the work.

[0050] A risk propagation impact model built upon content risk assessment vectors analyzes key nodes of risk amplification and user group emotional response patterns in different scenarios. This model quantifies the potential scope of impact and the probability of public opinion collapse, generating a risk warning map with scenario tags. The constructed progressive publicity control scheme dynamically binds work display permissions to dissemination time windows. By setting phased release conditions for the dissemination path, it achieves controllable, step-by-step promotion.

[0051] Based on scenario-based risk predictions, the system automatically adjusts the initial exposure scope, audience level, and content unlocking pace. For example, high-risk scenarios may be limited to internal or small-scale testing, with the dissemination radius gradually expanded once public opinion feedback meets the standards. This flexible control mechanism ensures both the effectiveness of the publicity and prevents sudden public opinion crises, significantly reducing the risk of cross-scenario dissemination while preserving the positive dissemination potential of the work.

[0052] By monitoring public opinion feedback data in real time during the dissemination process, the system can dynamically calibrate the publicity control strategy and form a closed-loop optimization. This adaptive mechanism continuously optimizes the collaborative efficiency of risk assessment and publicity control, upgrading the review and control process from static approval to intelligent continuous iteration, greatly reducing the cost of manual intervention, and ensuring that the dissemination behavior is always within the safe threshold. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the multimodal work review and promotion control method according to an embodiment of the present invention;

[0054] Figure 2 This is a flowchart illustrating the method for generating content risk assessment vectors after multimodal collaborative verification in an embodiment of the present invention. Detailed Implementation

[0055] 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.

[0056] 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.

[0057] Figure 1 This is a flowchart illustrating the multimodal work review and promotion control method according to an embodiment of the present invention. The multimodal work review and promotion control method provided by this embodiment includes:

[0058] Obtain multimodal input data for works awaiting review;

[0059] Deep semantic parsing is performed on the multimodal input data. By establishing a complementary verification mechanism between modalities, the uncertainty judgment result of a single modality is cross-validated with the deterministic features of other modalities to generate a content risk assessment vector after multimodal collaborative verification.

[0060] Based on the content risk assessment vector, a risk propagation impact model is constructed. By analyzing the risk diffusion path and audience sensitivity response pattern of the work content in different propagation scenarios, the potential impact range of the work to be reviewed in each propagation scenario is predicted, and scenario-based risk prediction results are obtained.

[0061] Based on the scenario-based risk prediction results, a progressive publicity control scheme is constructed. The progressive publicity control scheme associates and binds the display rights of the work with the dissemination time window by setting phased release conditions for the dissemination path.

[0062] The progressive publicity control scheme is implemented to output distribution instructions to the dissemination execution system and to monitor public opinion feedback data in real time during the dissemination process.

[0063] Figure 2 This is a flowchart illustrating the method for generating a multimodal collaboratively verified content risk assessment vector according to an embodiment of the present invention. The method involves performing deep semantic parsing on the multimodal input data, establishing a complementary verification mechanism between modalities, and cross-validating the uncertainty judgment result of a single modality with the deterministic features of other modalities to generate a multimodal collaboratively verified content risk assessment vector, including:

[0064] Semantic parsing is performed on each modality of the multimodal input data to extract features that represent the semantic content and construct the semantic feature space of each modality.

[0065] A semantic association graph between modalities is established in the semantic feature space. By identifying the semantic support relationships and semantic contradiction relationships of different modal data, the semantic association state between modalities is determined.

[0066] Risk identification is performed on each modality of data. Based on the identification confidence level, the identification results are divided into deterministic features and uncertain judgment results. Other modality data associated with the uncertain judgment results are located through the semantic association graph. Deterministic features that are semantically related to the uncertain judgment results are extracted from the other modality data as verification criteria.

[0067] Based on the verification criteria, the confidence level of the uncertainty judgment result is corrected. When the certainty feature is consistent with the risk direction of the uncertainty judgment result, the confidence level is increased. When the risk direction is conflicting, the confidence level is decreased to obtain the corrected risk identification result. The corrected risk identification result is then vectorized and fused to generate a content risk assessment vector.

[0068] When performing deep semantic parsing on multimodal input data, independent semantic parsing processes need to be executed for different modalities such as text, images, audio, and video. Taking text modality as an example, a semantic encoding method based on a pre-trained language model is used to map the text content into high-dimensional semantic vectors, capturing the contextual dependencies and implicit semantic tendencies between words. For image modality, convolutional neural networks or visual Transformers are used to extract local texture features, global scene semantics, and attribute descriptions of salient objects. For audio modality, the textual transcription information of the speech content, speaker emotional features, and background sound effect type labels are extracted. After processing by their respective feature extractors, each modality's data yields corresponding semantic feature representations, forming a structured semantic feature space within its respective feature space. The construction of the semantic feature space needs to ensure the integrity of the feature dimensions; that is, the feature vector of each modality should not only contain the surface semantics of the content but also cover multi-level semantic dimensions such as sentiment tendency, topic category, and style attributes, thus providing a sufficient information foundation for subsequent cross-modal association analysis.

[0069] After establishing the semantic feature spaces for each modality, a semantic association graph is constructed to capture the intrinsic relationships between data from different modalities. The semantic association graph uses the semantic feature nodes of each modality as vertices and the semantic relationships between modalities as edges. The edges are categorized into two types: semantic support relationships and semantic contradiction relationships. Semantic support relationships indicate that the semantic expression directions of two modalities are consistent and mutually corroborative; for example, the scene content in an image highly matches the text description. Semantic contradiction relationships, on the other hand, indicate that the semantic orientations of two modalities clearly conflict; for example, the speech content in audio significantly differs from the context presented in a video. A cross-modal attention mechanism can be used to identify semantic support relationships and semantic contradiction relationships by calculating the semantic similarity between feature vectors from different modalities. ,in and These represent the feature vector indices of two different modalities, when Exceeding the preset support threshold When, it is determined to be a semantic support relationship; when Below the preset contradiction threshold When a relationship is found to be semantically contradictory, it is classified as semantically contradictory; otherwise, it is marked as semantically neutral and is not used as a strong constraint for subsequent verification. Through systematic analysis of the semantic relationships between all modal pairs, a complete semantic association graph between modalities is finally formed, clarifying the supporting role and potential contradictions of each modality's data in the overall semantic structure.

[0070] When performing risk identification for each modality of data, the risk identifyer for each modality will output a risk category prediction and its corresponding identification confidence level based on the semantic features of that modality. When confidence level Above the deterministic threshold When the identification result is classified as a deterministic feature, it indicates that the modality has high reliability in judging specific risk types; when the confidence level is... Below the uncertainty threshold When an identification result is identified as an uncertain judgment result, it indicates that the modality's judgment of the risk type under the current content conditions has significant ambiguity or insufficient information. For risk identification items marked as uncertain judgment results, the constructed semantic association graph is used to locate related modalities. That is, along the edges connected to the uncertain node in the semantic association graph, other modal data nodes with semantic support or semantic contradiction relationships are found. After locating other related modal data, deterministic features that are highly semantically related to the uncertain judgment result are extracted from these modal data as the basis for cross-validation. During the extraction process, it is necessary to ensure that the extracted deterministic features and the risk category pointed to by the uncertain judgment result belong to the same semantic domain to avoid introducing irrelevant features that interfere with the accuracy of the validation results.

[0071] Based on the extracted verification criteria, a confidence level correction operation is performed on the uncertainty judgment result. The correction logic follows the principle of risk orientation consistency: when the risk category indicated by the deterministic feature is the same as the risk orientation of the uncertainty judgment result, it is considered that the two form semantic support, and the confidence level of the uncertainty judgment result is positively increased, with the increase being [missing information]. The confidence level is directly proportional to the strength of the certainty feature and the strength of its semantic association. When the risk category indicated by the certainty feature conflicts with the risk orientation of the uncertainty judgment result, it is considered that the two form a semantic contradiction, and the confidence level of the uncertainty judgment result is negatively adjusted, with the magnitude of the adjustment decreasing. Similarly, it is related to the degree of contradiction and the strength of association. After confidence correction, the identification results that were originally in the uncertainty range will be reclassified as certain results or rejected, thus forming a corrected set of risk identification results. This correction mechanism effectively solves the problem of misjudgment caused by the one-sidedness of information when facing complex content with a single modality, especially in scenarios where the text expression is ambiguous but the image semantics are clear, it can eliminate the ambiguity of text recognition by using the deterministic features of the image.

[0072] The corrected risk identification results from each modality are then vectorized and fused to generate the final content risk assessment vector. During this vectorization and fusion process, the corrected risk identification results from each modality must be aligned according to the risk category dimension to ensure that the judgment results of different modalities for the same risk category can be aggregated in a unified vector space. A weighted fusion strategy is used for aggregation, with the fusion weights of each modality... The risk assessment is determined by both the information richness and recognition reliability of the modality within the current content. Information richness reflects the modality's dominance in content expression, while recognition reliability is measured by the modality's historical recognition accuracy and the current content's modality quality assessment. After weighted fusion, a content risk assessment vector covering multiple risk category dimensions is obtained. The value of each dimension in the vector represents the comprehensive risk intensity of that risk category after multimodal collaborative verification. This vector will serve as the core input for constructing subsequent risk propagation impact models, supporting the generation of scenario-based risk prediction and progressive propaganda control schemes.

[0073] In the semantic feature space, a semantic association graph between modalities is established. By identifying the semantic support relationships and semantic contradiction relationships of different modal data, the semantic association state between modalities is determined, including:

[0074] In the semantic feature space, semantic deconstruction is performed on each modality data to extract semantic topics and semantic tendencies. The semantic topics represent the target object to which the content points.

[0075] Based on the semantic theme, a semantic association mapping is established between different modal data. Modal data expressing the same semantic theme are associated and labeled to construct an initial semantic association graph.

[0076] Modal data pairs are obtained through the initial semantic association graph, and a semantic tendency polarity mapping relationship is established for the modal data pairs. The consistency or opposition of semantic tendencies is determined through the polarity mapping relationship. When the consistency is determined, it is marked as a semantic support relationship; when the opposition is determined, it is marked as a semantic contradiction relationship.

[0077] A relation constraint network is constructed based on the semantic support relationship and the semantic contradiction relationship. A semantic enhancement path is established through the semantic support relationship, and a semantic conflict path is established through the semantic contradiction relationship.

[0078] The semantic association state is determined based on the connectivity structure of semantic enhancement paths and semantic conflict paths in the relation constraint network. When a connectivity structure is formed through semantic enhancement paths, it is determined to be a cooperative semantic association state, and when a connectivity structure is formed through semantic conflict paths, it is determined to be an adversarial semantic association state.

[0079] When performing semantic deconstruction on data from different modalities in the semantic feature space, it is necessary to extract the inherent semantic themes and semantic tendencies of input data in different forms such as text, images, and audio. Semantic themes represent the target object to which the content points. For example, in the text modality, the core issues corresponding to keyword clustering can be extracted using topic modeling methods; in the image modality, the main objects and scene categories in the image can be identified through visual semantic segmentation; and in the audio modality, the contextual objects to which the speech content themes or background sound effects point can be extracted through speech recognition and acoustic event detection. Semantic tendencies describe the emotional stance or attitude orientation held by each modality towards the target object, including basic polarity categories such as positive, negative, and neutral. After completing the semantic deconstruction of a single modality, each modality of data is represented as a structured description of several semantic themes and corresponding semantic tendencies, providing a unified semantic expression basis for subsequent cross-modal association analysis.

[0080] When establishing semantic association mappings between different modalities based on semantic topics, modalities expressing the same semantic topic are associated and labeled to construct an initial semantic association graph. Specifically, the semantic topics of each modality are mapped to a unified semantic topic space. By calculating the similarity between semantic topic vectors of different modalities, it is determined whether two modalities point to the same target object. When the distance between the semantic topics of two modalities in this unified space is less than a set topic matching threshold, they are considered to have a common semantic topic, and are associated and labeled with directed edges in the initial semantic association graph. The initial semantic association graph uses modal data nodes as vertices and semantic topic matching relationships as edges, forming a cross-modal semantic topology. This graph structure preserves the topic association paths between different modalities, providing a clear set of comparison objects for subsequent semantic polarity analysis.

[0081] After obtaining modal data pairs through the initial semantic association graph, a semantic polarity mapping relationship is established for each associated modal data pair. The polarity mapping relationship is established as follows: the semantic polarity of two modal data pairs under the same semantic topic is compared. If the polarity directions are the same (e.g., both are positive or both are negative), it is determined to be semantically consistent, and the modal data pair is labeled as a semantic support relationship; if the polarity directions are opposite (e.g., one is positive and the other is negative), it is determined to be semantically contradictory, and the modal data pair is labeled as a semantic contradiction relationship. In practical processing, the intensity of semantic polarity can be quantified using sentiment scores. Let the modality... The tendency score under a certain semantic topic is Modality Tendency scores under the same semantic topic ,when When it is determined to be consistent, The condition is determined to be contradictory. After completing polarity mapping for all associated modal data pairs, each edge of the initial semantic association graph is assigned a supporting or contradictory relation label, forming a semantic association graph structure with attributes.

[0082] When constructing a relation constraint network based on semantic support and semantic contradiction relationships, the attributed semantic association graph is further divided into two types of subgraph structures. Semantic enhancement paths are established through semantic support relationships, extracting all edges labeled as support relationships to form a semantic enhancement subgraph. The interconnected nodes in this subgraph represent modal data sets that mutually reinforce each other in content expression and point to a consistent stance. Semantic conflict paths are established through semantic contradiction relationships, extracting all edges labeled as contradiction relationships to form a semantic conflict subgraph. The interconnected nodes in this subgraph represent modal data sets that exhibit significant opposition in content expression and may lead to disagreements in content interpretation. The relation constraint network simultaneously includes both types of subgraphs and retains the attribution information of each node in both types of paths, providing a structured basis for subsequent determination of semantic association states. During construction, it is important to note that the same modal data node may participate in both semantic enhancement and semantic conflict paths simultaneously. This dual attribution reflects the semantic mediating role played by the modal data in the overall content and requires special handling during the state determination stage.

[0083] The final semantic association state is determined based on the connectivity structure of semantic enhancement paths and semantic conflict paths in the relation constraint network. When semantic enhancement paths form a connected structure in the relation constraint network, i.e., multiple modal data nodes are connected to each other through supporting relation edges, forming a complete strongly connected component, the overall semantic association state is determined as a cooperative semantic association state. A cooperative semantic association state means that the data from each modality are highly consistent in content orientation and sentiment stance, and the multimodal content as a whole presents a consistent semantic force. For such works, multimodal consistency can be used as a basis for increasing confidence during the risk assessment stage. When semantic conflict paths form a connected structure in the relation constraint network, i.e., multiple modal data nodes are connected to each other through contradictory relation edges, the overall semantic association state is determined as an adversarial semantic association state. An adversarial semantic association state means that there is cross-modal content opposition in the work, with different modalities conveying contradictory positions or information. Such works often have higher interpretive complexity and potential risk uncertainty during content review, requiring higher weighting in the subsequent risk propagation impact modeling stage.

[0084] In practical applications, the determination of semantic association states directly affects the generation method of content risk assessment vectors. Under collaborative semantic association states, the risk features of each modality can be weighted and fused through semantic enhancement paths, allowing features of highly consistent modalities to make a greater contribution to the fusion. Under adversarial semantic association states, independent risk assessments need to be conducted on conflicting modal pairs, and the conflict information itself should be included as an independent risk dimension in the assessment vector. Through the construction of the semantic association graph and the state determination mechanism described above, the semantic contradictions and collaborative relationships hidden in multimodal content can be effectively captured, providing a more accurate content understanding foundation for subsequent progressive propaganda control schemes.

[0085] Based on the content risk assessment vector, a risk propagation impact model is constructed. By analyzing the risk diffusion path and audience sensitivity response patterns of the work content in different propagation scenarios, the potential impact range of the work to be reviewed in each propagation scenario is predicted, resulting in scenario-based risk prediction results, including:

[0086] Based on the content risk assessment vector, the risk propagation potential of the work to be reviewed is calculated, the characteristics of the propagation medium and the audience distribution characteristics of different propagation scenarios are obtained, and a scenario-based propagation resistance model is constructed based on the characteristics of the propagation medium.

[0087] The risk propagation potential energy is coupled with the propagation resistance model for analysis. The risk diffusion path of the work to be reviewed is determined by the antagonistic relationship between the risk propagation potential energy and the propagation resistance model, and a risk propagation dynamic structure is constructed.

[0088] An audience sensitivity distribution map is established based on the audience distribution characteristics. The audience sensitivity distribution map represents the spatial distribution of the audience's sensitivity to the risky content of the work to be reviewed in different communication scenarios. The audience sensitivity distribution map is used to mark the sensitive areas of the risk propagation dynamics structure and determine the target sensitive propagation areas of the work to be reviewed in each communication scenario.

[0089] Based on the spatial superposition structure of the risk diffusion path of the work to be reviewed and the target sensitive propagation area in the risk propagation dynamics structure, the potential impact range of the work to be reviewed in each propagation scenario is determined, and scenario-based risk prediction results are generated.

[0090] After obtaining the content risk assessment vector through multimodal collaborative verification, it is necessary to further map the risk information carried by this vector onto real-world dissemination scenarios to predict the actual impact range of the work under review in different dissemination environments. To this end, the risk propagation potential of the work under review is calculated based on the content risk assessment vector. Risk propagation potential reflects the inherent driving force within the work's content itself that triggers risk diffusion during dissemination; its magnitude is determined by the combined values ​​of each risk dimension in the content risk assessment vector. Let the content risk assessment vector contain... The first risk dimension, the The values ​​of each risk dimension are The corresponding risk dimension transmission sensitivity coefficient is Then the risk propagation potential It can be represented as ,in The data, derived from the actual diffusion rate of content within this risk dimension in historical dissemination data, reflects the differences in the driving force of different types of risky content during dissemination. Higher risk dissemination potential indicates that the content is more likely to overcome dissemination obstacles and achieve widespread diffusion during the dissemination process.

[0091] After calculating the potential energy of risk propagation, it is necessary to obtain the characteristics of the propagation medium and audience distribution in different propagation scenarios, and then construct a scenario-based propagation resistance model based on the characteristics of the propagation medium. The characteristics of the propagation medium encompass multiple dimensions, including the strength of the platform's content moderation mechanism, the distribution of platform user activity, the diffusion amplification factor of the content recommendation algorithm, and the platform's content lifecycle. For example, short video platforms have a high recommendation algorithm amplification factor and a relatively short content lifecycle, while text and image forum platforms have a longer content accumulation period and a relatively low instantaneous propagation rate. For each propagation scenario… Its propagation resistance The strength coefficient of the review mechanism in this scenario Content diffusion attenuation coefficient and platform content filtering efficiency A joint decision can be expressed as ,in A larger value indicates a stronger natural diffusion and amplification effect of the content within that scenario, and relatively lower propagation resistance. By establishing propagation resistance models for multiple propagation scenarios, a scenario-based propagation resistance set covering major propagation channels is formed.

[0092] The core of coupling risk propagation potential with propagation resistance models lies in determining the risk diffusion path through their antagonistic relationship. This involves analyzing the risk propagation potential of a work awaiting review in a specific propagation scenario. The propagation resistance is significantly higher than that in this scenario. At that time, it was determined that the scenario had a high-risk diffusion channel; when and When they approach, they are identified as potential risk diffusion channels; when Significantly lower At that time, it was determined that the propagation resistance in the scenario was sufficient to suppress the spread of risk. Based on this, the assessment results of the diffusion channels in each scenario were linked together according to the cross-platform flow patterns of content across different platforms, forming a risk diffusion path network. This network uses each propagation scenario as a node and the probability of cross-platform content migration as the edge weight, characterizing the dynamic structure of risky content spreading from its initial release scenario to other scenarios—that is, the risk propagation dynamic structure. The risk propagation dynamic structure not only describes the directional path of risk diffusion but also records the changing pattern of risk intensity at each path node as propagation resistance decreases, providing structural support for subsequent sensitive area labeling.

[0093] Building an audience sensitivity distribution map based on audience distribution characteristics is a crucial step in predicting the potential scope of influence. Audience distribution characteristics include information such as user age structure, cultural background, interest tag clustering, and historical content interaction patterns across various communication scenarios. The audience sensitivity distribution map transforms these characteristics into a spatialized expression of sensitivity, where "space" does not refer to physical geographic space, but rather a semantic space constructed using the communication scenario and audience group category as coordinate axes. For communication scenarios… Audience categories Its sensitivity to specific risky content The result is calculated by combining the group's historical risk content interaction rate, emotional activation intensity, and content secondary dissemination tendency coefficient. Let the historical risk content interaction rate be... The emotional activation intensity coefficient is The secondary propagation tendency coefficient is ,but .when Exceeding the preset sensitivity threshold In this process, the corresponding audience area in the dissemination scenario is marked as a highly sensitive area, and sensitive area markers are superimposed on the corresponding nodes of the risk dissemination dynamics structure to ultimately determine the target sensitive dissemination area of ​​the work to be reviewed in each dissemination scenario. The determination of the target sensitive dissemination area means that once risky content reaches this area, it will have a high probability of triggering a strong reaction from the audience and triggering secondary dissemination, thereby significantly amplifying the scope of the risk's impact.

[0094] After determining the risk propagation dynamics and the target sensitive propagation area, spatial overlay analysis is performed to determine the potential impact range of the work under review in various propagation scenarios. The core logic of spatial overlay is to find the intersection of the set of propagation scenario nodes covered by the risk diffusion path and the set of highly sensitive nodes marked in the target sensitive propagation area. The intersection nodes represent the areas where the risky content is most likely to have a substantial negative impact during propagation. For each propagation scenario within the intersection nodes, the risk intensity attenuation value and audience sensitivity are comprehensively considered. and the size of the audience for this scenario. Calculate the potential impact level in this scenario. , can be represented as ,in This reflects the remaining diffusion energy of risky content after overcoming propagation resistance in a given scenario. The potential impact magnitudes of each propagation scenario are aggregated to form a scenario-based risk prediction result covering multiple propagation scenarios. This result is output in structured data form, including risk level labels for each propagation scenario, coordinates of target sensitive propagation areas, numerical values ​​of potential impact magnitudes, and a risk diffusion path topology map, providing accurate scenario-based input for the subsequent construction of progressive propaganda control plans. Through the complete modeling process of the above risk propagation impact model, abstract content risk assessment results can be transformed into scenario-specific propagation risk predictions, significantly improving the accuracy and foresight of propaganda control decisions.

[0095] The risk propagation potential energy is coupled with the propagation resistance model for analysis. The antagonistic relationship between the risk propagation potential energy and the propagation resistance model is used to determine the risk diffusion path of the work under review, thus constructing a risk propagation dynamics structure, including:

[0096] Calculate the energy distribution gradient of the risk propagation potential energy and the resistance distribution gradient of the propagation resistance model, extract the difference features between the energy distribution gradient and the resistance distribution gradient, and identify the dominant propagation channel based on the difference features;

[0097] The attenuation trajectory of the risk propagation potential energy and the cumulative trajectory of the propagation resistance model are traced along the dominant propagation channel. When the attenuation trajectory and the cumulative trajectory intersect, the risk propagation critical point is determined.

[0098] The path nodes from the starting region to the risk propagation critical point are connected to form the risk diffusion path of the work to be reviewed. The number of propagation levels traversed by the risk diffusion path is counted to obtain the level depth. The number of branches of the risk diffusion path at each propagation level is counted to obtain the branch density.

[0099] Based on the hierarchical depth, the longitudinal propagation range of the risk diffusion path is marked in the propagation resistance model, and based on the branch density, the lateral diffusion intensity of the risk diffusion path is marked, thereby generating a risk propagation dynamic structure.

[0100] After independently constructing the risk propagation potential energy and propagation resistance models, a coupled analysis of the two is needed. This analysis, through a quantitative comparison of their antagonistic relationship, reveals the actual diffusion path of risk content in various propagation scenarios. Coupled analysis involves placing the "driving force field" represented by the risk propagation potential energy and the "impedance field" represented by the propagation resistance model within the same spatial coordinate system. The relative strength of the two is compared node by node to determine whether the risk content possesses the driving force for continued propagation at that node. When the driving force field strength significantly exceeds the impedance field strength, the node is considered an active propagation zone; when the two tend to be in equilibrium, the node is considered a critical propagation zone; and when the impedance field strength continuously suppresses the driving force field, propagation actually terminates at that node.

[0101] To quantify this antagonistic relationship, it is necessary to calculate the energy distribution gradient of the risk propagation potential and the resistance distribution gradient of the propagation resistance model, respectively. The energy distribution gradient describes the rate of change of the risk propagation potential between adjacent propagation nodes, reflecting the increasing or decreasing trend of the driving energy carried by the risk content as it propagates to the next level. Let the nodes in the propagation graph... Its neighboring nodes The energy distribution gradient between them is ,but ,in and They are nodes and nodes The potential energy value for risk propagation at a given location. Similarly, the gradient of the resistance distribution. Describing propagation resistance from nodes To the node The cumulative change, of which and These represent the propagation resistance values ​​for the corresponding nodes.

[0102] When extracting differential features, use the differential gradient. As a basis for judgment. When It is a positive value and exceeds the preset dominant channel recognition threshold. At that time, on the side Marked as a candidate edge for the dominant propagation channel; when Negative values ​​or absolute values ​​less than In this case, the edge is considered a blocked channel or a weak propagation channel. By calculating the difference gradient for each node pair in the propagation graph, a set of continuous dominant propagation channels can be selected. These channels constitute the backbone path along which the risky content is most likely to spread in the overall propagation network. The value of should be calibrated in conjunction with the actual propagation data of the platform, and is usually taken as the average value of the difference between the energy gradient and the drag gradient in historical high propagation events as a reference benchmark.

[0103] When tracing the decay trajectory of risk propagation potential energy along the dominant propagation channel, the process of potential energy decreasing with increasing propagation level is recorded node by node, forming a monotonically decreasing decay curve. Simultaneously, the cumulative trajectory of the propagation resistance model is traced; that is, as the propagation path extends, the total accumulated propagation resistance at each node continuously increases, forming a monotonically increasing cumulative curve. The two trajectories intersect at a certain node on the propagation graph, i.e., the point where the potential energy decay value equals the resistance accumulation value; this point is defined as the risk propagation critical point. Before the critical point, the propagation momentum of the risk content remains sufficient, and propagation can continue; after the critical point, resistance exceeds the driving force, and propagation tends to naturally decay until it stops. The precise location of the critical point is crucial for the formulation of subsequent propaganda control plans, as it directly determines the boundaries of the propagation levels that a gradual propaganda control plan needs to intervene in.

[0104] After determining the risk propagation threshold, all path nodes from the starting region (i.e., the source node or set of source nodes where the work was first published) to the threshold are sequentially connected to form a complete risk diffusion path. This path is not a single linear chain, but a set of directed paths with a tree-like branching structure, because in actual propagation networks, a single node may simultaneously propagate content to multiple downstream nodes. Two dimensions of statistical analysis are performed on this path set: hierarchy depth and branch density. Hierarchy Depth Defined as the maximum number of propagation levels a risk spreads from its source node to its critical point, reflecting the depth to which risk content can penetrate vertically; branch density. Defined as the risk diffusion path in the 1st The number of path branches at each propagation level reflects the lateral diffusion capability of risky content at that level. This is determined through layer-by-layer statistical analysis. By averaging across all levels, the average branch density of the entire diffusion path can be obtained. It is used to assess the overall lateral diffusion intensity.

[0105] Based on hierarchy depth In the propagation resistance model, the longitudinal propagation range of the risk diffusion path is marked. Specifically, the vertical axis coordinate of the resistance model is extended from level 1 to level 2. The hierarchy is defined by highlighting all nodes and their resistance values ​​within a longitudinal range, allowing subsequent control schemes to intuitively identify the longitudinal propagation hierarchy intervals requiring key intervention. This is based on branch density. When labeling the lateral diffusion intensity, the diffusion width of each level is marked in the propagation graph using a visualization method based on the node radiation radius or edge weight. The larger the value, the wider the horizontal labeling range, indicating a stronger ability for the content to spread at that level, requiring more intensive propagation interception measures to be deployed horizontally.

[0106] By overlaying and integrating the longitudinal propagation range annotations and the lateral diffusion intensity annotations, a risk propagation dynamics structure is ultimately generated. This structure uses a propagation graph as its underlying framework, the differential distribution of energy gradients and resistance gradients as its driving layer, dominant propagation channels as its skeleton, critical points as boundary markers, and layer depth and branch density as quantitative descriptive indicators, comprehensively presenting the dynamic diffusion behavior pattern of the work under review in the current propagation environment. The output form of the risk propagation dynamics structure can be a directed graph data structure with attribute annotations, where each edge carries a differential gradient value. Each node carries a potential energy value and a resistance value. Each node marked as a critical point additionally carries a critical marker attribute. Each set of nodes at each level carries a corresponding branch density value. This structured output will directly serve as the input basis for the construction phase of the progressive propaganda control scheme, enabling the phased release conditions of the dissemination path to accurately correspond with the actual risk diffusion dynamics, thereby achieving refined and dynamic control over the dissemination process of the work.

[0107] Based on the scenario-based risk prediction results, a progressive publicity control scheme is constructed. This scheme links the display permissions of the work with the dissemination time window by setting phased release conditions for the dissemination path, including:

[0108] The potential impact range and risk level of each communication scenario are extracted from the scenario-based risk prediction results. The communication scenarios are sorted and mapped to communication paths based on the risk level. The audience reach limit is set for each communication scenario on the communication path according to the potential impact range.

[0109] For the propagation path, scene entry conditions and scene exit conditions are set. The scene entry conditions are associated with the cumulative propagation scale and risk event frequency of the preceding propagation scene. The scene exit conditions are associated with the propagation duration of the current propagation scene and the audience reach limit. Scene switching rules are constructed based on the scene entry conditions and scene exit conditions.

[0110] Time windows are allocated to each dissemination scenario and divided into a permission restriction period and a permission open period. During the permission restriction period, the display permission is limited to the initial range of the audience reach limit. During the permission open period, the display permission is extended to the full range of the audience reach limit according to the execution result of the scenario switching rule.

[0111] The propagation path, the scene switching rules, and the time window are combined to generate a progressive propaganda control scheme.

[0112] When extracting the potential impact scope and risk level of each communication scenario from the scenario-based risk prediction results, it is necessary to perform structured analysis on the prediction output of each communication scenario. Each communication scenario corresponds to a comprehensive risk level score, which is jointly determined by the potential impact magnitude and the audience's sensitivity response pattern. All communication scenarios are sorted in descending order according to their risk level scores, with the highest risk level corresponding to the strictest control starting point and the lowest risk level corresponding to the most lenient control endpoint. After sorting, the sequence of communication scenarios is mapped to an ordered communication path, where each node represents a communication scenario arranged by risk level. The path direction points from high-risk scenarios to low-risk scenarios, reflecting a gradual logic of expanding from controlled communication to open communication.

[0113] For each dissemination scenario along the dissemination path, an audience reach limit is set based on the potential impact of that scenario. The audience reach limit represents the maximum number of people allowed to access the work's content within that dissemination scenario, denoted as [missing information]. ,in Index for propagation scenarios. The value of is negatively correlated with the risk level of the scenario; that is, the higher the risk level, the lower the risk level. The smaller the size, the more strictly the spread of high-risk scenarios will be contained within a controllable range. For high-risk scenarios at the starting point of the transmission path, It can be configured to allow only internal auditors or a small test audience to access it; for low-risk scenarios at the end of the path, This tiered system can be expanded to the entire user base of the platform. This tiered approach ensures that works will not trigger uncontrollable public opinion risks due to large-scale exposure in the early stages of dissemination.

[0114] The design of the scene admission conditions revolves around the actual operational results of the preceding propagation scene. For the first... In each scenario, the admission criteria involve two core indicators: the cumulative spread of the preceding scenario. Frequency of risk events occurring in the preceding scenario Cumulative scale of transmission This reflects the total number of audiences actually reached by the work in the preceding scenes, and the frequency of risk events. This reflects the number of reports, violation flags, or negative public opinion events that occurred in the preceding scenarios. Scenario entry requirements. Reaching the maximum audience reach of the preceding scenarios a certain proportion That's all. At the same time, it is required Not exceeding the preset risk event frequency limit ,Right now Both conditions must be met simultaneously to trigger the process towards the first... The switching of scenes. This is the threshold for the cumulative scale of transmission. Both the upper limit for the frequency of risk events are pre-configured based on the type of work and the platform strategy.

[0115] The design of the exit conditions for a given scenario is linked to the duration of the current dissemination scenario and the upper limit of audience reach. For the first... In each propagation scenario, the exit condition consists of the disjunction of two sub-conditions: the first being the actual propagation duration of the current scenario. Exceeding the preset maximum stay time ,Right now Secondly, the actual audience reach in the current scenario. The audience reach limit has been reached. ,Right now If any sub-condition is met, the exit process for the current scenario is triggered. After the exit process is initiated, if the entry conditions for the next scenario are also met, a scenario switch is executed; if the entry conditions are not yet met, the current scenario enters a waiting state, suspending new audience outreach until the entry conditions are met or manual intervention is initiated. Based on the above entry and exit conditions, a complete scenario switching rule is constructed. This rule describes the transition logic between scenarios along the propagation path in the form of a finite state machine. Each state corresponds to a propagation scenario, and the trigger condition for state transition is the combined judgment result of the entry and exit conditions.

[0116] The allocation of time windows is performed after the scene switching rules are determined. An independent time window is allocated to each propagation scene along the propagation path. Each time window consists of two phases: a permission restriction period and a permission unlocking period. The permission restriction period is located at the beginning of the time window, and its duration is denoted as [duration missing]. During this period, display permissions are limited to the maximum audience reach. The initial range, that is, only allows The number of audiences accessing the work's content, for The initial restricted ratio is usually set to Smaller fractions, for example ,in The initial open ratio coefficient and During the restricted access period, we will continuously collect audience feedback, reports, and interaction data to assess the actual risk performance of the work in the current context.

[0117] The access period is located in the latter part of the time window, and its duration is denoted as [duration missing]. The condition for entering the open access period is: the frequency of risk events collected cumulatively during the restricted access period does not exceed [a certain number]. Furthermore, the overall sentiment index of audience feedback did not trigger the warning threshold. Once the above conditions are met, the display permission is increased from [previous setting]. Expand to audience reach limit The full range, i.e., the maximum allowed. The number of audience members accessing the content. If the frequency of risk events exceeds the limit during the access restriction period. If an emotional alert is triggered, the access period for that scenario is canceled, the dissemination of the work in the current scenario is suspended, and a review request is sent to the manual review process. This two-stage time window design enables each dissemination scenario to have adaptive risk response capabilities, rather than relying solely on cross-scenario switching rules.

[0118] By combining dissemination paths, scene switching rules, and time windows, a complete progressive publicity control plan is generated. This plan is stored in the form of structured data, including an ordered sequence of scenes along the dissemination path and the audience reach limit for each scene. Scene admission criteria parameters and Scene exit condition parameters and Time window parameters and and the initial open ratio coefficient The overall logic of the progressive publicity control scheme ensures the orderly migration of works from high-risk controlled scenarios to low-risk open scenarios. Each expansion step is based on the actual risk performance of the preceding scenarios, avoiding the uncontrollable risks caused by a one-time full release. In practical applications, for the same work, different regions, different platform types, or different audience groups can correspond to different progressive publicity control scheme instances. Each instance operates independently and does not interfere with each other, thereby achieving refined and differentiated dissemination management.

[0119] A second aspect of the present invention provides a multimodal work review and promotion control system, comprising:

[0120] The data acquisition unit is used to acquire multimodal input data of the works to be reviewed;

[0121] The risk assessment unit is used to perform deep semantic parsing on the multimodal input data. By establishing a complementary verification mechanism between modalities, the uncertainty judgment result of a single modality is cross-verified with the deterministic features of other modalities to generate a content risk assessment vector after multimodal collaborative verification.

[0122] The scenario prediction unit is used to construct a risk propagation impact model based on the content risk assessment vector. By analyzing the risk diffusion path and audience sensitivity response pattern of the work content in different propagation scenarios, it predicts the potential impact range of the work to be reviewed in each propagation scenario and obtains scenario-based risk prediction results.

[0123] The control scheme unit is used to construct a progressive publicity control scheme based on the scenario-based risk prediction results. The progressive publicity control scheme associates and binds the display permissions of the work with the dissemination time window by setting phased release conditions for the dissemination path.

[0124] The distribution monitoring unit is used to output distribution instructions to the dissemination execution system in accordance with the progressive publicity control scheme, and to monitor the public opinion feedback data of the work in real time during the dissemination process.

[0125] A third aspect of the present invention provides an electronic device, comprising:

[0126] processor;

[0127] Memory used to store processor-executable instructions;

[0128] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0129] 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.

[0130] 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.

[0131] 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 the promotion of multimodal works, characterized in that, include: Obtain multimodal input data for works awaiting review; Deep semantic parsing is performed on the multimodal input data. By establishing a complementary verification mechanism between modalities, the uncertainty judgment result of a single modality is cross-validated with the deterministic features of other modalities to generate a content risk assessment vector after multimodal collaborative verification. Based on the content risk assessment vector, a risk propagation impact model is constructed. By analyzing the risk diffusion path and audience sensitivity response pattern of the work content in different propagation scenarios, the potential impact range of the work to be reviewed in each propagation scenario is predicted, and scenario-based risk prediction results are obtained. Based on the scenario-based risk prediction results, a progressive publicity control scheme is constructed. The progressive publicity control scheme associates and binds the display rights of the work with the dissemination time window by setting phased release conditions for the dissemination path. The progressive publicity control scheme is implemented to output distribution instructions to the dissemination execution system and to monitor public opinion feedback data in real time during the dissemination process.

2. The method according to claim 1, characterized in that, Deep semantic parsing is performed on the multimodal input data. By establishing a complementary verification mechanism between modalities, the uncertainty judgment result of a single modality is cross-validated with the deterministic features of other modalities to generate a content risk assessment vector after multimodal collaborative verification, including: Semantic parsing is performed on each modality of the multimodal input data to extract features that represent the semantic content and construct the semantic feature space of each modality. A semantic association graph between modalities is established in the semantic feature space. By identifying the semantic support relationships and semantic contradiction relationships of different modal data, the semantic association state between modalities is determined. Risk identification is performed on each modality of data. Based on the identification confidence level, the identification results are divided into deterministic features and uncertain judgment results. Other modality data associated with the uncertain judgment results are located through the semantic association graph. Deterministic features that are semantically related to the uncertain judgment results are extracted from the other modality data as verification criteria. Based on the verification criteria, the confidence level of the uncertainty judgment result is corrected. When the certainty feature is consistent with the risk direction of the uncertainty judgment result, the confidence level is increased. When the risk direction is conflicting, the confidence level is decreased to obtain the corrected risk identification result. The corrected risk identification result is then vectorized and fused to generate a content risk assessment vector.

3. The method according to claim 2, characterized in that, In the semantic feature space, a semantic association graph between modalities is established. By identifying the semantic support relationships and semantic contradiction relationships of different modal data, the semantic association state between modalities is determined, including: In the semantic feature space, semantic deconstruction is performed on each modality data to extract semantic topics and semantic tendencies. The semantic topics represent the target object to which the content points. Based on the semantic theme, a semantic association mapping is established between different modal data. Modal data expressing the same semantic theme are associated and labeled to construct an initial semantic association graph. Modal data pairs are obtained through the initial semantic association graph, and a semantic tendency polarity mapping relationship is established for the modal data pairs. The consistency or opposition of semantic tendencies is determined through the polarity mapping relationship. When the consistency is determined, it is marked as a semantic support relationship; when the opposition is determined, it is marked as a semantic contradiction relationship. A relation constraint network is constructed based on the semantic support relationship and the semantic contradiction relationship. A semantic enhancement path is established through the semantic support relationship, and a semantic conflict path is established through the semantic contradiction relationship. The semantic association state is determined based on the connectivity structure of semantic enhancement paths and semantic conflict paths in the relation constraint network. When a connectivity structure is formed through semantic enhancement paths, it is determined to be a cooperative semantic association state, and when a connectivity structure is formed through semantic conflict paths, it is determined to be an adversarial semantic association state.

4. The method according to claim 1, characterized in that, Based on the content risk assessment vector, a risk propagation impact model is constructed. By analyzing the risk diffusion path and audience sensitivity response patterns of the work content in different propagation scenarios, the potential impact range of the work to be reviewed in each propagation scenario is predicted, resulting in scenario-based risk prediction results, including: Based on the content risk assessment vector, the risk propagation potential of the work to be reviewed is calculated, the characteristics of the propagation medium and the audience distribution characteristics of different propagation scenarios are obtained, and a scenario-based propagation resistance model is constructed based on the characteristics of the propagation medium. The risk propagation potential energy is coupled with the propagation resistance model for analysis. The risk diffusion path of the work to be reviewed is determined by the antagonistic relationship between the risk propagation potential energy and the propagation resistance model, and a risk propagation dynamic structure is constructed. An audience sensitivity distribution map is established based on the audience distribution characteristics. The audience sensitivity distribution map represents the spatial distribution of the audience's sensitivity to the risky content of the work to be reviewed in different communication scenarios. The audience sensitivity distribution map is used to mark the sensitive areas of the risk propagation dynamics structure and determine the target sensitive propagation areas of the work to be reviewed in each communication scenario. Based on the spatial superposition structure of the risk diffusion path of the work to be reviewed and the target sensitive propagation area in the risk propagation dynamics structure, the potential impact range of the work to be reviewed in each propagation scenario is determined, and scenario-based risk prediction results are generated.

5. The method according to claim 4, characterized in that, The risk propagation potential energy is coupled with the propagation resistance model for analysis. The antagonistic relationship between the risk propagation potential energy and the propagation resistance model is used to determine the risk diffusion path of the work under review, thus constructing a risk propagation dynamics structure, including: Calculate the energy distribution gradient of the risk propagation potential energy and the resistance distribution gradient of the propagation resistance model, extract the difference features between the energy distribution gradient and the resistance distribution gradient, and identify the dominant propagation channel based on the difference features; The attenuation trajectory of the risk propagation potential energy and the cumulative trajectory of the propagation resistance model are traced along the dominant propagation channel. When the attenuation trajectory and the cumulative trajectory intersect, the risk propagation critical point is determined. The path nodes from the starting region to the risk propagation critical point are connected to form the risk diffusion path of the work to be reviewed. The number of propagation levels traversed by the risk diffusion path is counted to obtain the level depth. The number of branches of the risk diffusion path at each propagation level is counted to obtain the branch density. Based on the hierarchical depth, the longitudinal propagation range of the risk diffusion path is marked in the propagation resistance model, and based on the branch density, the lateral diffusion intensity of the risk diffusion path is marked, thereby generating a risk propagation dynamic structure.

6. The method according to claim 1, characterized in that, Based on the scenario-based risk prediction results, a progressive publicity control scheme is constructed. This scheme links the display permissions of the work with the dissemination time window by setting phased release conditions for the dissemination path, including: The potential impact range and risk level of each communication scenario are extracted from the scenario-based risk prediction results. The communication scenarios are sorted and mapped to communication paths based on the risk level. The audience reach limit is set for each communication scenario on the communication path according to the potential impact range. For the propagation path, scene entry conditions and scene exit conditions are set. The scene entry conditions are associated with the cumulative propagation scale and risk event frequency of the preceding propagation scene. The scene exit conditions are associated with the propagation duration of the current propagation scene and the audience reach limit. Scene switching rules are constructed based on the scene entry conditions and scene exit conditions. Time windows are allocated to each dissemination scenario and divided into a permission restriction period and a permission open period. During the permission restriction period, the display permission is limited to the initial range of the audience reach limit. During the permission open period, the display permission is extended to the full range of the audience reach limit according to the execution result of the scenario switching rule. The propagation path, the scene switching rules, and the time window are combined to generate a progressive propaganda control scheme.

7. A multimodal work review and promotion control system, used to implement the method as described in any one of claims 1-6, characterized in that, include: The data acquisition unit is used to acquire multimodal input data of the works to be reviewed; The risk assessment unit is used to perform deep semantic parsing on the multimodal input data. By establishing a complementary verification mechanism between modalities, the uncertainty judgment result of a single modality is cross-verified with the deterministic features of other modalities to generate a content risk assessment vector after multimodal collaborative verification. The scenario prediction unit is used to construct a risk propagation impact model based on the content risk assessment vector. By analyzing the risk diffusion path and audience sensitivity response pattern of the work content in different propagation scenarios, it predicts the potential impact range of the work to be reviewed in each propagation scenario and obtains scenario-based risk prediction results. The control scheme unit is used to construct a progressive publicity control scheme based on the scenario-based risk prediction results. The progressive publicity control scheme associates and binds the display permissions of the work with the dissemination time window by setting phased release conditions for the dissemination path. The distribution monitoring unit is used to output distribution instructions to the dissemination execution system in accordance with the progressive publicity control scheme, and to monitor the public opinion feedback data of the work in real time during the dissemination process.

8. 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 6.

9. 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 6.

Images