A method and system for brand retail multi-modal marketing content automated generation

Abstract

The application discloses a brand retail multi-modal marketing content automatic generation method and system, relates to the field of AIGC marketing content generation, and comprises the following steps: acquiring and aligning a plurality of selected product reasoning paths, extracting and cutting a stable path core, and replaying a selected product model; generating an evidence load unit, and transferring the evidence load unit to a theme unit through an intermediate semantic island; when the load is continuous, combining the theme unit to obtain a creative theme; constructing a shadow user focus according to a target user focus to perform a competitive route, and distributing the evidence load unit based on route differences to obtain a differentiated selling point; then generating multi-modal marketing content and an evidence load account book, and generating a constraint debt account book through compliance detection; performing a correction transaction, when a load gap occurs, compensating the corresponding evidence load unit to a replacement content carrier and then releasing the original content carrier, and outputting the multi-modal marketing content. The application has the effects of realizing the alignment of selected product evidence and creativity, the generation of a differentiated selling point, and a compliance closed loop.

Description

Technical Field

[0001] This application relates to the field of AIGC marketing content generation, and in particular to a method and system for automating the generation of multimodal marketing content for brand retail. Background Technology

[0002] In the brand retail sector, the inventory of goods is huge and marketing activities are frequent. The traditional method of relying on manual planning and production of marketing content is time-consuming and labor-intensive, and it is difficult to keep up with the rapidly changing marketing pace.

[0003] While general AI-generated content tools have significantly increased content production speed, several technical shortcomings have been exposed in practical applications. The generation process lacks unified constraints on brand guidelines such as visual style and tone, resulting in significant style fluctuations between different batches or by different operators. This often manifests as deviations in color schemes, layouts, and copywriting tone from brand guidelines, leading to blurred brand recognition. There is a severe disconnect between product selection decisions and marketing creativity. Product selection models rely on multi-dimensional reasoning based on product attributes and market trends to reach selection conclusions, but these reasoning bases and the role of product selection cannot be effectively extracted and conveyed to the creative generation stage. This results in marketing content that is completely unrelated to the core reasons for product selection, weakening its persuasiveness. Existing tools lack granular selling point generation mechanisms for differentiated user groups. When facing different channels and user profiles, they cannot build differentiated content strategies based on user focus differences, resulting in similar marketing materials that fail to achieve precise targeting and personalized communication, leading to low marketing conversion rates. Furthermore, automatically generated content also poses compliance risks, lacking a closed-loop processing capability from generation and compliance testing to correction and compensation, resulting in high review and modification costs. Summary of the Invention

[0004] To address the issues of inconsistent content styles, disconnected product selection and creativity, homogenization, and difficulty in compliance corrections arising from existing AI-generated content, this application provides a method and system for automating the generation of multimodal marketing content for brand retail.

[0005] In the first aspect, the method for automatically generating multimodal marketing content for brand retail provided in this application adopts the following technical solution: S1, obtain multiple product selection reasoning paths of the product selection model, align the product selection reasoning paths based on product evidence and product selection function, and extract stable path kernels; S2. Cut off the stable path kernel and replay the product selection model, and generate an evidence load unit containing load parameters based on the obtained product selection response; The load parameters include root evidence identifiers, coupling relationships, and load states; S3. The evidence payload unit is passed to the topic unit through the intermediate semantic island. When the payloads are continuous, the topic units are combined to obtain the creative topic. S4. Construct shadow user concerns based on the concerns of the target users, perform contention routing with the evidence payload unit as the sending end and the above two types of user concerns as the receiving end, and allocate the evidence payload unit based on the routing differences to obtain differentiated selling points. S5. Based on the brand DNA, the creative theme, and the differentiated selling points, the evidence payload unit is projected onto the multimodal content carrier to generate multimodal marketing content and evidence payload ledger, and a constraint debt ledger is generated after compliance testing; S6. Based on the evidence payload ledger and the constraint debt ledger, perform a correction transaction. When the correction generates a payload gap, first compensate the evidence payload unit corresponding to the payload gap to the alternative content carrier, then release the original content carrier, and output multimodal marketing content that passes the compliance test.

[0006] By adopting the above technical solution, a stable path kernel is extracted from the reasoning path of the product selection model, and product selection evidence, coupling relationships, and load states are encapsulated in the form of evidence payload units. This allows the core reasons for product selection decisions to be explicitly expressed and transmitted across stages, solving the problem of the separation between product selection logic and marketing creativity. Simultaneously, creative themes are constructed through intermediate semantic islands, differentiated selling points are generated using competitive routing, and corrective transactions are driven by evidence payload ledgers and constraint debt ledgers. This achieves full-process automation from evidence extraction, theme generation, selling point allocation to compliance correction, significantly improving the generation efficiency and content consistency of multimodal marketing content for brand retail.

[0007] Optionally, S1 specifically includes performing perturbations on the product attributes, candidate product context, and reasoning state while keeping the evidence of the target product unchanged, and obtaining the product selection reasoning path corresponding to each perturbation state to form a set of product selection path graphs; Based on the location of product evidence and the role of product selection, the nodes and reasoning relationships in the product selection path graph set are aligned to form an aligned path structure; The stable path kernel is obtained by extracting the common reasoning substructure that maintains the root commodity evidence and selection effect before and after the perturbation from the alignment path structure and changes the selection result after cutting.

[0008] By employing the aforementioned technical solution, perturbations are applied to product attributes, candidate product contexts, and reasoning states while maintaining the target product evidence unchanged. This allows for the acquisition of product selection reasoning paths across multiple states, forming a set of product selection path graphs. These paths are then aligned based on the position of product evidence and its role in product selection. Finally, a stable path kernel is extracted, representing a common reasoning substructure that maintains the root product evidence and its role in product selection before and after perturbation, but alters the product selection result after severance. This effectively removes redundant variables and noisy paths from the product selection reasoning process, ensuring the robustness and representativeness of the product selection evidence and making the product selection rationale upon which subsequent creative ideas rely more reliable and stable.

[0009] Optionally, the evidence payload unit specifically includes the root evidence identifier, product evidence location, product selection function type, degree of irreplaceability, payload type, coupling relationship, allowed semantic transformation type, allowed carrying modality, current content carrier, payload status, and payload version; The allowed semantic transformation type is used to construct the intermediate semantic island; the allowed bearer modality is used to determine the multimodal content carrier; the payload type, the payload state, and the payload version are used to control the correction transaction.

[0010] By adopting the above technical solution, the various parameters contained in the evidence payload unit provide unified structured metadata for semantic transformation, modality selection, and compliance correction of the evidence payload throughout the entire chain. This enables the same evidence payload to be accurately identified, traced, and controlled at different stages, avoiding ambiguity and loss during semantic transmission and improving the system's controllability under complex creative combinations and compliance constraints.

[0011] Optionally, S3 specifically includes S31, converting the evidence payload unit into an initial intermediate semantic island according to the allowed semantic transformation type, and binding the initial intermediate semantic island to the root evidence identifier; S32. Generate a next intermediate semantic island based on the previous intermediate semantic island and its load status, and record the newly added semantics, lost semantics and commodity evidence sources between its adjacent intermediate semantic islands to form a semantic island chain. S33. Detect the load continuity of the semantic island chain based on the newly added semantics, the lost semantics, and the source of the product evidence; repair the intermediate semantic islands with discontinuous loads; and convert the semantic island chain with continuous loads into the topic unit to obtain the creative topic.

[0012] By adopting the above technical solution, based on the allowed semantic transformation types of the evidence payload unit, it is converted into intermediate semantic islands and bound to the root evidence identifier. Sequential semantic island chains are generated, and newly added semantics, lost semantics, and the source of product evidence between adjacent semantic islands are recorded. This process detects and repairs discontinuous intermediate semantic islands, and finally combines continuous semantic island chains into thematic units. This achieves semantic assembly from fragmented evidence to coherent creative themes, ensuring a strict semantic correspondence between the creative theme and the product selection evidence, and avoiding the phenomenon of creative ideas deviating from the reasons for product selection.

[0013] Optionally, S33 specifically includes recovering the evidence payload unit from adjacent intermediate semantic islands in reverse, verifying the newly added semantics and the lost semantics according to the root evidence identifier and the product evidence source, and obtaining the payload continuity result; Based on the load continuity results, identify the fracture intermediate semantic islands, retain the valid intermediate semantic islands before the fracture intermediate semantic islands, and regenerate alternative intermediate semantic islands by selecting different allowed semantic transformation types. The alternative intermediate semantic islands are verified using the load continuity results. The verified alternative intermediate semantic islands are then connected to the semantic island chain, and the connected semantic island chain is converted into the topic unit.

[0014] By employing the above technical solution, evidence payload units are recovered from adjacent intermediate semantic islands. Based on root evidence identifiers and product evidence sources, newly added and missing semantics are verified to obtain payload continuity results. After identifying broken intermediate semantic islands, the preceding valid parts are retained, and different permissible semantic transformation types are selected to regenerate alternative intermediate semantic islands. Once verified, these are integrated into the semantic island chain and converted into topic units. This effectively repairs evidence breaks caused by improper semantic transformation during creative idea generation, improves the completeness and coherence of creative topic generation, and ensures that key product selection reasons are not omitted or distorted.

[0015] Optionally, S4 specifically includes keeping other user features in the target user profile unchanged and replacing the target user's focus to obtain the shadow user's focus, and establishing candidate routes from the evidence payload unit to the two types of user focus based on the bridging relationship between the user profile body and the product body. The same evidence payload unit is transmitted to the two types of user concerns along the candidate routes respectively, and the route differences are obtained based on the route reachability state, route conflict state, and coupling relationship maintenance state. Based on the routing differences, the evidence payload units with routing advantages are exclusively allocated to the target user's focus, and the allocated evidence payload units are set to an occupied state to suppress candidate selling points that repeatedly carry the evidence payload units, thereby obtaining the differentiated selling points.

[0016] By employing the above technical solution, while keeping other features in the target user profile unchanged, the focus is replaced to construct shadow user focus points. Based on the bridging relationship between the user profile and the product itself, candidate routes are established from the evidence payload unit to the two types of user focus points. Route differences are obtained through competitive routing, and evidence payload units with routing advantages are exclusively assigned to the target user focus points. Simultaneously, an occupancy state is set to suppress duplicate payloads, enabling differentiated expression of the same product selection evidence across different user groups. This fundamentally solves the problem of monotonous marketing content and enhances the uniqueness and precise targeting capabilities of selling points.

[0017] Optionally, S6 specifically includes S61, reading the evidence payload ledger and the constraint debt ledger, and generating candidate correction transactions based on the association between constraint debts and multimodal content carriers; S62. Execute the candidate correction transaction in the multimodal marketing content copy and ledger copy, detect the constraint debt, load state and coupling relationship before and after execution, and obtain the transaction verification result; S63. If the transaction verification results indicate that the constraint debt has decreased, the mandatory evidence payload unit has been retained, the coupling relationship has not been broken, and no new higher priority constraint debt has been added, submit the candidate correction transaction; otherwise, roll back the multimodal marketing content copy and the ledger copy.

[0018] By adopting the above technical solution, the evidentiary payload ledger and the constraint liability ledger are read, and candidate correction transactions are generated. These transactions are executed in the copies, and changes in constraint liabilities, payload status, and coupling relationships are detected to obtain transaction verification results. Correction transactions are submitted only when constraint liabilities are reduced, mandatory evidentiary payload units are retained, coupling relationships are not broken, and no new higher-priority constraint liabilities are added. This achieves evidence-based and controllable rollback for compliant corrections. It avoids introducing new compliance risks or disrupting existing creative structures during correction operations, significantly reducing the manual intervention and correction costs for multimodal marketing content compliance review.

[0019] Optionally, S61 specifically includes S611, constructing a content dependency graph based on the multimodal content carrier, the evidence payload unit, the constraint debt, and the evidence carrying relationship, semantic connection relationship, cross-modal correspondence relationship, and constraint dependency relationship between them; S612. Starting from the multimodal content carrier associated with the constraint debt to be processed, propagate the load missing state and constraint debt change in the content dependency graph, and truncate the propagation when it reaches the content carrier that does not carry the corresponding evidence load unit to obtain the modified influence cone; S613. Based on the modified influence cone, determine the load risk and constraint liability changes caused by each modified operation, select the modified operation that reduces the constraint liability and maintains the evidence load unit, and generate the candidate modified transaction.

[0020] By employing the aforementioned technical solution, a content dependency graph is constructed. Starting from the content carrier associated with the constraint debt to be processed, the graph propagates the state of missing payloads and changes in constraint debts, truncating at content carriers that do not carry corresponding evidence payload units. This yields a correction impact cone, which is used to filter and reduce constraint debts while maintaining correction operations for evidence payload units, generating candidate correction transactions. This precisely defines the impact boundary of each compliant correction, avoiding content collapse and cascading debts caused by excessive correction, thus ensuring the accuracy and security of the correction.

[0021] Optionally, S62 specifically includes S621, when the correction influence cone indicates that the candidate correction transaction generates a load gap, selecting the alternative content carrier from the content carrier equivalent cluster carrying the same evidence load unit, and pre-occupying the corresponding evidence load unit to the alternative content carrier; S622. Detect the payload acceptance status, compliance status, and coupling relationship of the alternative content carrier. When the detection passes, switch the alternative content carrier to an effective state, release the original content carrier, and update the evidence payload ledger and the constraint debt ledger. S623. If the detection fails, cancel the pre-occupancy and retain the original content carrier, and generate the transaction verification result based on the updated evidence payload ledger and the constraint debt ledger.

[0022] By adopting the above technical solution, when the correction impact cone indicates that a candidate correction transaction has a load gap, an alternative content carrier is selected from the equivalent cluster of content carriers carrying the same evidence load unit, and the evidence load unit is pre-occupied. After the load-bearing status, compliance status, and coupling relationship of the alternative carrier are checked and approved, it is switched to an effective state, the original content carrier is released, and the ledger is updated. If the check fails, the pre-occupancy is canceled, and the original carrier is retained. This achieves lossless migration of evidence loads during the compliance correction process. It ensures that the correction operation does not cause gaps or breaks in key evidence within the marketing content, maintaining a dynamic balance between content compliance and evidence integrity.

[0023] Secondly, the present invention provides an automated generation system for multimodal marketing content in brand retail, including a product selection reasoning module, an evidence payload generation module, a theme generation module, a selling point generation module, a content generation module, a compliance detection module, and a correction output module; The product selection reasoning module is used to obtain multiple product selection reasoning paths of the product selection model, align the product selection reasoning paths based on product evidence and product selection function, and extract stable path kernels; The evidence payload generation module is connected to the product selection reasoning module and is used to cut off the stable path kernel and replay the product selection model. Based on the obtained product selection response, an evidence payload unit containing payload parameters is generated. The payload parameters include root evidence identifier, coupling relationship and payload state. The topic generation module is connected to the evidence payload generation module and is used to transmit the evidence payload unit to the topic unit through the intermediate semantic island. When the payloads are continuous, the topic units are combined to obtain a creative topic. The selling point generation module is connected to the evidence payload generation module. It is used to construct shadow user concerns based on the target user concerns, perform contention routing with the evidence payload unit as the sending end and the target user concerns and shadow user concerns as the receiving end, and allocate the evidence payload unit based on the routing differences to obtain differentiated selling points. The content generation module is connected to the evidence payload generation module, the theme generation module, and the selling point generation module, respectively, and is used to project the evidence payload unit onto the multimodal content carrier based on the brand DNA, the creative theme, and the differentiated selling point to generate multimodal marketing content and evidence payload ledger. The compliance detection module is connected to the content generation module and is used to perform compliance detection on the multimodal marketing content and generate a constraint debt ledger. The correction output module is connected to the evidence payload generation module, the content generation module, and the compliance detection module. It is used to perform correction transactions based on the evidence payload ledger and the constraint debt ledger. When a correction generates a payload gap, it first obtains the evidence payload unit corresponding to the payload gap from the evidence payload generation module to compensate for the alternative content carrier, then releases the original content carrier, and outputs multimodal marketing content that passes the compliance detection.

[0024] The technical effects of the automated generation system for multimodal marketing content in brand retail provided in the second aspect of this invention are the same as those in the first aspect, and will not be repeated here.

[0025] In summary, this application includes at least one of the following beneficial technical effects: 1. By extracting stable path kernels from the product selection model and transforming them into structured evidence payload units, which are then transmitted to the creative theme via intermediate semantic islands, a strict alignment between product selection decision-making criteria and marketing creativity is achieved. This effectively solves the problem of product selection logic being disconnected from creativity in general generation tools, enabling the generated marketing content to faithfully reflect the reasons for product selection and enhancing the credibility of the content and brand consistency. 2. By constructing shadow user attention points and executing competitive routing, evidence payload units are exclusively assigned to target user attention points based on routing differences to generate differentiated selling points, overcoming the shortcomings of existing tools in terms of content homogenization; thus enabling the same product to automatically generate targeted selling point statements when facing different user profiles and channels, achieving fine-grained personalized marketing content production. 3. By establishing an evidentiary payload ledger and a constraint liability ledger, and executing verifiable corrective actions accordingly, when a payload gap occurs, compensation is first made to an alternative content carrier before releasing the original carrier, forming a closed-loop mechanism for compliance detection and corrective compensation. Therefore, the problems of high compliance risk and high correction cost of automatically generated content are solved, ensuring the compliance of marketing content while maintaining the integrity of the evidentiary payload and the stability of the creative structure. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a flowchart of the method for automatically generating multimodal marketing content for brand retail provided in the embodiments of this application; Figure 2 This is a flowchart of step S1 of the method for automatically generating multimodal marketing content for brand retail provided in the embodiments of this application; Figure 3 This is a flowchart of step S3 of the method for automatically generating multimodal marketing content for brand retail provided in the embodiments of this application; Figure 4 This is a flowchart of step S33 of the method for automatically generating multimodal marketing content for brand retail provided in the embodiments of this application; Figure 5 This is the flow of step S4 of the method for automatically generating multimodal marketing content for brand retail provided in the embodiments of this application; Figure 6 This is a flowchart of step S61 of the method for automatically generating multimodal marketing content for brand retail provided in this application embodiment; Figure 7 This is a flowchart of step S62 of the method for automatically generating multimodal marketing content for brand retail provided in the embodiments of this application; Figure 8 This is an overall architecture diagram of the automated generation system for multimodal marketing content in brand retail provided in this application embodiment; Figure 9 This is a schematic diagram of the internal structure of the computer device provided in the embodiments of this application.

[0028] Attached diagram labels: 1. Product selection reasoning module; 2. Evidence load generation module; 3. Theme generation module; 4. Selling point generation module; 5. Content generation module; 6. Compliance testing module; 7. Correction output module. Detailed Implementation

[0029] The following is in conjunction with the appendix Figure 1 - Appendix Figure 9 This application will be described in further detail.

[0030] This application discloses a method and system for automatically generating multimodal marketing content for brand retail. This application is applicable to the automated generation of e-commerce product detail pages, social media content, advertising images, short video scripts, video subtitles, and video footage. The method can be executed by a server, cloud computing platform, or content generation device equipped with a processor and memory, or it can be executed collaboratively by a product selection model service, a content generation model service, and a compliance detection service.

[0031] Among them, the product selection model can be a machine learning model, a large language model, or a rule model that outputs product selection results based on product attributes, market trends, sales data, user needs, and inventory data; the model used to generate copy can be a large language model; and the model used to generate image or video materials can be a diffusion model, a generative adversarial network, or other generative models.

[0032] This application does not treat product selection logic extraction, creative theme generation, differentiated selling point generation, and brand compliance correction as independent processing steps. Instead, it generates evidence payload units that can be transferred between processing steps from the product selection reasoning path. The evidence payload units enter the creative theme through an intermediate semantic island, enter the differentiated selling point through competitive routing, and are then projected onto multimodal content carriers such as text fragments, image areas, video shots, subtitles, or narration. When compliance corrections are made to multimodal marketing content, the evidence payload ledger is used to determine the current content carrier of each evidence payload unit, and the constraint debt ledger is used to determine the constraint debts to be eliminated, thereby preventing the core product evidence that determines the selection of a product from being deleted, diluted, or incorrectly replaced during multiple rounds of corrections.

[0033] See attached document Figure 1 A method for automatically generating multimodal marketing content for brand retail includes S1, obtaining multiple product selection reasoning paths of the product selection model, aligning the product selection reasoning paths based on product evidence and product selection function, and extracting stable path kernels.

[0034] The product selection reasoning path refers to the intermediate processing path by which the product selection model derives product selection results from product input information. This path can include directed reasoning relationships between product attributes, attribute combinations, intermediate judgment results, market trends, user needs, and the final product selection result. For example, the product selection reasoning path could state that "the retro washed material and high-waisted cut both conform to the retro fashion trend and meet the target user's need for body proportion enhancement, thus enabling the product to meet the selection criteria."

[0035] Product evidence refers to original or verified data that can prove a product possesses corresponding product attributes. This includes product parameter fields, target areas in product images, testing items in product inspection reports, verified product descriptions, attribute nodes in the product itself, and the data sources associated with those attribute nodes. Product evidence location refers to the addressable location of the product evidence within the original data, such as field identifiers in a product database, region coordinates in a product image, item identifiers in an inspection report, or node identifiers in the product itself.

[0036] The product selection function refers to the role that product evidence plays in the product selection reasoning process. This includes serving as a necessary condition for a product to be selected, contributing to the selection reasons along with other product evidence, enhancing the product's selection score, or excluding product attributes that do not match the selection objectives. Different product selection reasoning paths may use different textual descriptions or different intermediate nodes, but as long as they correspond to the same product evidence location and have the same product selection function, they can be considered as reasoning components from the same source.

[0037] A stable path kernel is a common inference substructure extracted from multiple product selection inference paths. Even when non-core input information, candidate product context, or model inference state changes, the stable path kernel maintains the root product evidence and product selection function. Furthermore, after the stable path kernel is cut off, the product selection result or score of the product selection model changes according to preset conditions. The stable path kernel is used to eliminate random representations, intermediate inference drift, and candidate product context bias in a single model interpretation, providing stable input for subsequent evidence payload unit generation.

[0038] See attached document Figure 2 In this embodiment, S1 specifically includes S11, performing perturbations on the product attributes, candidate product context and reasoning state to keep the evidence of the target product unchanged, and obtaining the product selection reasoning path corresponding to each perturbation state to form a product selection path graph set.

[0039] In this context, perturbation refers to altering the information view or inference environment read by the product selection model while maintaining the identity of the target product and the original product evidence records. Perturbation can include masking attributes of non-products to be tested, changing the order of candidate products, adding or removing candidate products adjacent to the target product, changing the sampling state of the product selection model, changing the organization order of product selection prompts, and replacing contextual descriptions that do not affect the authenticity of the product.

[0040] In practice, the product ontology of the target product is first obtained. The product ontology includes at least product nodes, product attribute nodes, product function nodes, usage scenario nodes, and product evidence nodes, and includes the attribute attribution relationship, function support relationship, scenario applicability relationship, attribute mutual exclusion relationship, and attribute dependency relationship between the nodes.

[0041] The system reads the candidate product attributes corresponding to the target product from the product body and generates multiple input views. In one input view, all candidate product attributes are retained; in another input view, one non-tested product attribute is hidden; in yet another input view, the order of the target product in the candidate product set is changed; alternatively, the product attributes of the target product can remain unchanged, and only other products in the candidate product set with similar attributes to the target product are replaced.

[0042] It should be noted that perturbing product attributes does not modify the original product evidence in the product database, but rather masks, replaces, or rearranges the corresponding attributes in the input copy of the product selection model. The original product evidence and its position remain unchanged during the perturbation process, so that different product selection reasoning paths can be aligned through the same product evidence position.

[0043] Each input view is input into the product selection model to obtain the product selection results, product selection scores, and product selection inference paths corresponding to each perturbation state. If the product selection model can directly output the model inference log, the product selection inference path is obtained by parsing the model inference log; if the product selection model only outputs the product selection results and product selection scores, the product selection inference path can be recovered from the input-output relationship of the product selection model using the feature attribution module, the path explanation module, or a constrained large language model.

[0044] For each product selection reasoning path, a product selection path graph is constructed. Nodes in the product selection path graph include product attribute nodes, intermediate judgment nodes, external condition nodes, and product selection result nodes; edges in the product selection path graph are used to represent the reasoning direction between nodes. When multiple product attributes jointly produce an intermediate judgment result, hyperedges connecting multiple product attribute nodes and intermediate judgment nodes can be established to preserve the synergistic effect between product attributes.

[0045] The product selection path diagrams corresponding to each disturbance state are combined into a product selection path diagram set, and the disturbance object, disturbance method, product selection result and product selection score are recorded for each product selection path diagram, so that S12 can perform path alignment.

[0046] S12. Based on the location of product evidence and the role of product selection, align the nodes and reasoning relationships in the product selection path graph set to form an aligned path structure.

[0047] The aligned path structure refers to a unified path structure formed by merging nodes and reasoning relationships from different product selection path diagrams that share the same origin but differ in their descriptions. Alignment does not require nodes to have completely identical text; rather, it is based on whether nodes point to the same product evidence location, whether they have the same product selection function, and whether they connect to the same type of downstream judgment result.

[0048] In practice, for each product attribute node in the product selection path map set, the location of the product evidence referenced by that product attribute node is queried. For example, although "high-waisted design", "high waistline position" and "high waist cut" are different in wording, they can all point to the "waist type equals high-waisted" attribute node in the product body or the waist type field in the product specification table, and therefore can be merged into the same root product evidence node.

[0049] For intermediate judgment nodes that cannot be directly linked to the location of product evidence, trace back along their upstream reasoning relationships to the product attribute nodes, and determine the product selection role based on their impact on the selection results. For example, "in line with the retro trend of the 1990s" and "has a vintage feel" are both derived from retro washed materials and high-waisted cuts, and both enhance the product selection results. Therefore, they can be aligned as intermediate judgment nodes of the same type.

[0050] The system generates alignment features for any two nodes to be aligned. Alignment features can include whether the product evidence locations are consistent, whether the node types are consistent, whether the product selection functions are consistent, whether the upstream node sets overlap, and whether the downstream node types are consistent. The system determines the node correspondence based on the alignment features and merges multiple nodes that meet the corresponding conditions into aligned nodes.

[0051] After node alignment is completed, the directed edges and hyperedges in each product selection path graph are aligned. If two inference relations connect the same aligned nodes and their inference direction and product selection function are consistent, the two inference relations are merged into an aligned inference relation. The aligned inference relation also records its existence state in each product selection path graph, its corresponding perturbation state, and the product selection model response after severance.

[0052] The alignment path structure is formed by the alignment nodes and alignment inference relationships. The alignment path structure retains both the common parts between different product inference paths and the variable parts that only appear under partial perturbation states, thus enabling S13 to identify stable path kernels.

[0053] S13. Extract the common reasoning substructure from the aligned path structure that keeps the root commodity evidence and product selection function unchanged before and after the perturbation, and changes the product selection result after cutting off, to obtain the stable path kernel.

[0054] Root product evidence refers to product evidence that can be directly traced back to the original product data and is located upstream in the product selection reasoning path. The common reasoning substructure may include a root product evidence node, multiple root product evidence nodes with coupling relationships, reasoning relationships connecting root product evidence nodes and intermediate judgment nodes, and reasoning relationships connecting intermediate judgment nodes and product selection result nodes.

[0055] In practice, the system first selects candidate inference substructures from the alignment path structure that can connect to the product selection result node under multiple perturbation states. For each candidate inference substructure, it checks whether its root product evidence position remains consistent under multiple perturbation states and whether its product selection function remains consistent. If a candidate inference substructure only changes the node text due to different natural language expressions generated by the model, but its root product evidence position and product selection function remain unchanged, it is still determined as a stable candidate structure.

[0056] Subsequently, a severing operation is performed on the stable candidate structure. This severing operation can include masking root product evidence nodes, deleting a reasoning relation, deleting hyperedges connecting multiple root product evidence nodes, or blocking the transmission from intermediate decision nodes to the product selection result node. The severed input copy is then re-inputted into the product selection model to obtain the severed product selection response.

[0057] The product selection response includes the product selection result, product selection score, product ranking position, and the product selection reasoning path regenerated by the product selection model. The product selection response before and after the severance is compared. If, after severing the candidate reasoning substructure, the target product changes from being selected to not being selected, the product selection score decreases to meet the discrimination criteria, the product ranking position shifts to meet the discrimination criteria, or the original product selection reason is replaced by other evidence, then the candidate reasoning substructure is considered to have a valid impact on the product selection result.

[0058] For candidate reasoning substructures containing multiple root product evidence nodes, single-node severing and combination relationship severing are also performed separately. If retaining each root product evidence node individually cannot produce the original product selection effect, but retaining their combination relationship can produce the original product selection effect, then the coupled reasoning relationship connecting these root product evidence nodes is retained in the stable path kernel.

[0059] Finally, candidate inference substructures that simultaneously satisfy the cross-disturbance stability condition and the cutoff response condition are merged to obtain the stable path kernel. The root product evidence nodes, coupled inference relationships, and product selection roles within the stable path kernel are provided to S2.

[0060] For example, considering a pair of high-waisted retro jeans, after changing the order of candidate products, disabling price and promotional information, and altering the model's sampling state, the product selection model might output "Retro wash and high-waisted cut conform to retro trends," "Washed denim combined with a raised waistline creates a 90s silhouette," and "Retro material and high-waisted cut meet the proportion enhancement needs of the target audience." Although the wording differs, they all point to the two product evidence locations of "retro wash material" and "high-waisted cut," and both contribute to the product's selection. After severing the combination relationship between the two product evidences, the product still possesses the washed material and high-waisted cut, but the product selection model no longer outputs "90s silhouette" as a selection reason. Therefore, the two root product evidence nodes and their combination relationship can be extracted as a stable path kernel.

[0061] It should be noted that the stable path core is not simply the intersection of textual content in multiple product selection reasoning paths. If only keyword intersection is used, different expressions may lead to the accidental deletion of truly stable root product evidence. This application's embodiment aligns different expressions by considering the location of product evidence and the role of product selection, and verifies the actual impact of the common reasoning substructure on the product selection result by cutting off playback, thereby reducing the impact of single-interpretation fluctuations and semantic expression differences on the extraction of core product selection logic.

[0062] S2. Disconnect the stable path kernel and replay the product selection model. Based on the obtained product selection response, generate evidence load units containing load parameters. Load parameters include root evidence identifier, coupling relationship, and load state.

[0063] In this embodiment, the evidence payload unit specifically includes root evidence identifier, product evidence location, product selection function type, degree of irreplaceability, payload type, coupling relationship, allowed semantic transformation type, allowed carrying modality, current content carrier, payload status, and payload version.

[0064] Semantic transformation types are allowed to be used to construct intermediate semantic islands; payload modalities are allowed to be used to determine multimodal content carriers; payload type, payload state, and payload version are used to control correction transactions.

[0065] Among them, the evidence payload unit refers to the stateful data unit generated by the root product evidence in the stable path kernel and its selection function. The evidence payload unit is not just a static label used to mark the data source, but a control object that is continuously transmitted and updated in state during the processes of creative theme generation, differentiated selling point reasoning, multimodal content generation, and compliance correction.

[0066] The root evidence identifier is a unique identifier generated for the root product evidence in the stable path kernel. It is used to associate product attributes, subject units, differentiating selling points, and multimodal content carriers with the same product evidence. The product evidence location is used to access the original record of the root product evidence.

[0067] The product selection role type describes the function of root product evidence in the product selection reasoning process. Product selection role types can include necessary role, coupling role, enhancing role, and exclusionary role. A necessary role indicates that deleting the corresponding root product evidence means the target product no longer meets the selection criteria; a coupling role indicates that the corresponding root product evidence needs to work with other root product evidence to generate a selection reason; an enhancing role indicates that the corresponding root product evidence improves the product selection score, but deletion does not necessarily change its selection status; an exclusionary role indicates that the corresponding product attribute has been excluded by the product selection logic, and subsequent content must not reintroduce it as a selling point.

[0068] The degree of nonsubstitutability is used to indicate the impact on the selected product response when the root product evidence is removed, replaced, or loses its combination relationship. The degree of nonsubstitutability can be expressed as a continuous value, a discrete level, or a ranking number.

[0069] Payload types are used to determine the retention rules for evidence payload units in subsequent processing. Payload types can include mandatory payloads, coupled payloads, transferable payloads, alternative payloads, and prohibited payloads. Mandatory payloads must be carried by at least one multimodal content carrier in the final multimodal marketing content; coupled payloads must be expressed together with the designated evidence payload unit; transferable payloads allow migration from one multimodal content carrier to another; alternative payloads allow being carried by other semantic expressions pointing to the same product evidence; prohibited payloads must not be included in creative themes, differentiating selling points, or the final multimodal marketing content.

[0070] Coupling relationships are used to record the relationships between multiple evidence payload units that must be carried together, migrated together, carried mutually exclusively, or carried in a preset order. Allowed semantic transformation types are used to define the semantic transformation directions that root commodity evidence can undergo, including transformations from physical attributes to perceived results, from functional attributes to usage results, from stylistic attributes to cultural context, from scene attributes to situational narrative, and from multiple coupled attributes to combined imagery.

[0071] The allowed carrier modality is used to define that the evidentiary payload unit can be projected onto one or more of text, images, video footage, subtitles, or narration. The current content carrier is used to point to the text fragment, image area, video shot, subtitle fragment, or narration fragment that is currently actually carrying the evidentiary payload unit.

[0072] Payload states can include unassigned, transferred, occupied, pre-occupied, valid, gapped, released, and prohibited states. Payload versions are used to record version changes of the evidentiary payload unit during semantic transformation, content projection, and compliance correction processes.

[0073] In practice, for each root product evidence node in the stable path kernel, the following steps are performed sequentially: node severing, substitution of similar evidence, severing of coupling relationships, and severing of intermediate inference relationships. Node severing is used to determine whether the root product evidence is necessary evidence; substitution of similar evidence is used to determine whether the root product evidence can be replaced by an attribute of the same category; severing of coupling relationships is used to determine whether multiple root product evidences must work together; and severing of intermediate inference relationships is used to determine how the root product evidence influences the product selection result.

[0074] The system replays each cutoff result back to the product selection model and compares the product selection responses before and after the cutoff. It determines the degree of irreplaceability based on the direction and extent of change in the product selection responses; it determines the coupling relationship based on the product selection responses corresponding to individual root product evidence and combinations of root product evidence; and it determines the load type and permissible bearing modes based on whether root product evidence must be retained and whether it can be expressed by other modes.

[0075] For example, if the target product changes from being selected to not being selected after the "high-waisted cut" is removed, then its corresponding evidentiary load unit can be set as a mandatory load. If neither "retro washed material" nor "high-waisted cut" is sufficient to generate a selection reason for "1990s silhouette" when they exist separately, but the combination of the two can generate this selection reason, then the two are set as coupled loads and a common bearing relationship is established.

[0076] If the product image clearly demonstrates the "high-waisted cut" and the copywriting effectively conveys evidence of the product by stating "improving the visual waistline," then its allowed load modality can be set to text, image, video, and subtitle. If a product's waterproof performance lacks a test report or product parameters to support it, then a valid evidence load unit will not be generated for that performance. If the product selection path explicitly excludes the "waterproof" attribute, then a prohibited load can be generated to prevent subsequent model generation from adding unsubstantiated waterproof selling points.

[0077] After the evidence payload unit is generated, its initial payload state is set to unassigned, its current content carrier is empty, and its payload version is set to the initial version. The evidence payload unit is then passed to S3.

[0078] It should be noted that the existence of "product evidence" and "product evidence being carried by the current content" can be distinguished by payload status and payload version. For example, the presence of high-waisted cropping in the product body does not equate to the final text or image effectively expressing high-waisted cropping. Only after the corresponding multimodal content carrier passes the payload acceptance detection will the evidence payload unit switch from the transmission state or pre-occupancy state to the valid state.

[0079] S3. The evidence payload unit is passed to the topic unit through the intermediate semantic island. The topic units are combined when the payloads are continuous to obtain the creative topic.

[0080] The intermediate semantic island refers to an intermediate data structure located between the root commodity evidence and the subject unit, used to receive the evidence payload unit and complete a single restricted semantic transformation. The intermediate semantic island includes at least the semantic expression, root evidence identifier, input payload state, output payload state, permitted semantic transformation type, added semantics, missing semantics, and commodity evidence source.

[0081] A semantic island chain is a semantic transmission structure composed of multiple interconnected intermediate semantic islands. The output payload state of the preceding intermediate semantic island serves as the input payload state of the following intermediate semantic island, thereby transforming the root commodity evidence step by step into an engaging creative expression.

[0082] A thematic unit is the smallest unit of creative expression formed by a chain of semantic islands with continuous payloads. A thematic unit can carry one evidential payload unit or multiple evidential payload units with coupling relationships. A creative theme is obtained by combining one or more thematic units according to the product selection function and the order of the marketing narrative.

[0083] Load continuity refers to the fact that when the evidence load unit passes through adjacent intermediate semantic islands, the root evidence identifier remains traceable, the loss of semantics does not destroy the necessary meaning of the corresponding root commodity evidence, the newly added semantics can be supported by the commodity evidence source or the permitted semantic transformation type, and the coupling relationship is not erroneously split.

[0084] See attached document Figure 3 In this embodiment, S3 specifically includes S31, converting the evidence payload unit into an initial intermediate semantic island according to the allowed semantic transformation type, and binding the initial intermediate semantic island to the root evidence identifier.

[0085] In practice, a semantic transformation operator library is pre-built. Each semantic transformation operator in the library includes the applicable product attribute type, input semantic structure, allowed output semantic type, prohibited semantic type, and corresponding validation rules.

[0086] For example, for the pattern attribute "high-waisted cut", you can select the semantic transformation operator "physical attribute to visual result" to transform it into "visual waistline enhancement"; for the material and style attribute "vintage washed material", you can select the semantic transformation operator "physical attribute to perceived result" to transform it into "naturally aged denim texture".

[0087] The system reads the root product evidence based on the product evidence location in the evidence payload unit, selects a semantic transformation operator based on the product attribute type and allowed semantic transformation types, and uses the semantic transformation operator to generate initial intermediate semantic islands. The initial intermediate semantic islands record the root evidence identifier they receive, the semantic transformation operator used, and the semantics before and after the transformation.

[0088] After the initial intermediate semantic islands are generated, the load state of the evidence payload units is switched from the unassigned state to the transitive state. If the evidence payload units are coupled, initial intermediate semantic islands can be generated for each evidence payload unit separately, or initial intermediate semantic islands that jointly carry multiple evidence payload units can be generated using the combined image semantic transformation operator.

[0089] S32. Generate a next intermediate semantic island based on the previous intermediate semantic island and its load state, and record the newly added semantics, lost semantics and commodity evidence sources between its adjacent intermediate semantic islands to form a semantic island chain.

[0090] Here, "added semantics" refers to concepts, emotions, scenarios, or effects added to the subsequent intermediate semantic island compared to the previous intermediate semantic island. "Lost semantics" refers to concepts that existed in the previous intermediate semantic island but were not carried over to the subsequent intermediate semantic island. The source of product evidence is used to explain whether the added semantics can be supported by root product evidence, other product evidence coupled with the root product evidence, or verified product functionality.

[0091] In practice, the initial intermediate semantic island is input into the next semantic transformation operator to generate the next intermediate semantic island. For example, "high-waisted cut" is transformed into "visual waistline raised" through the first intermediate semantic island, then into "body proportions extended upwards" through the second intermediate semantic island, and finally into "sleek and modern visual silhouette" through the third intermediate semantic island. After each transformation, the semantic difference module compares adjacent intermediate semantic islands. The semantic difference module can use a combination of semantic role labeling, entity relation extraction, textual implication model, and product ontology query to identify added and lost semantics.

[0092] For example, when changing from "visual waistline enhancement" to "body proportion extension", the new semantic is "proportion extension", which can be supported by the product function relationship between high-waisted cut and visual proportion. When changing from "body proportion extension" to "windproof and warm modern silhouette", the "windproof and warm" in the new semantic cannot be supported by high-waisted cut or other product evidence, and is therefore marked as a new semantic without product evidence.

[0093] The system assigns a sequence identifier to each intermediate semantic island and records the input and output load states between adjacent intermediate semantic islands, forming a semantic island chain. The semantic island chain is then provided to S33 for load continuity detection.

[0094] S33. Detect the load continuity of the semantic island chain based on the newly added semantics, missing semantics, and the source of product evidence, repair the intermediate semantic islands with discontinuous loads, and convert the semantic island chain with continuous loads into topic units to obtain creative topics.

[0095] See attached document Figure 4Specifically, S33 includes S331, recovering evidence payload units from adjacent intermediate semantic islands in reverse, and obtaining payload continuity results based on the root evidence identifier and the verification of the source of commodity evidence by adding and losing semantics.

[0096] In practice, adjacent intermediate semantic islands are input into the reverse semantic parser. The reverse semantic parser extracts product attributes, product functions, usage results, stylistic context, and attribute combinations from the semantic expressions of the intermediate semantic islands, and queries the product ontology based on the extraction results to obtain recoverable root product evidence.

[0097] The root evidence identifiers recovered from the previous and subsequent intermediate semantic islands are compared. If the subsequent intermediate semantic island can still recover the mandatory payload carried by the previous intermediate semantic island, and the newly added semantics have a commodity evidence source, then the corresponding evidence payload unit is determined to be continuous.

[0098] If a subsequent intermediate semantic island cannot directly recover the root product evidence, but can return to the previous intermediate semantic island through pre-recorded allowed semantic transformation types, and then return to the root product evidence from the previous intermediate semantic island, it can still be considered continuous. This method allows for significant textual differences between the final creative theme and the original product parameters, but still requires traceability of semantic transformations at all levels.

[0099] For missing semantics, determine whether they are essential to the root commodity evidence. If the missing semantics are merely non-essential modifiers and do not alter the root commodity evidence, they are allowed to be missing; if the missing semantics cause the mandatory load to be unrecoverable or cause the coupled load to be split, they are considered load discontinuities.

[0100] The load continuity results include continuous state, fracture state, fracture location, affected root evidence identifiers, source-free new semantics, and missing necessary semantics.

[0101] S332. Identify intermediate semantic islands in the fracture based on the load continuity results, retain the valid intermediate semantic islands before the fracture intermediate semantic islands, and regenerate alternative intermediate semantic islands by selecting different allowed semantic transformation types.

[0102] Among them, the broken intermediate semantic island refers to the intermediate semantic island that prevents the evidence payload unit from continuing to trace back to the root commodity evidence, introduces the semantics of no commodity evidence, or destroys the coupling relationship.

[0103] In practice, the intermediate semantic island where the load discontinuity first occurs is located based on the load continuity results. This intermediate semantic island and all subsequent intermediate semantic islands are temporarily removed from the semantic island chain, while retaining the valid intermediate semantic islands that have passed the detection before the break point.

[0104] The system reads the semantic transformation type used in the semantic islands in the fracture and selects an alternative semantic transformation type from other semantic transformation types allowed by the evidence payload unit. For example, when the "functional attribute to usage result" transformation produces an unfounded effect, the "functional attribute to scene perception" transformation can be used instead; when the single attribute transformation causes the loss of coupling relationship, the "coupled attribute to combined image" transformation can be used instead.

[0105] By utilizing the effective intermediate semantic islands before the alternative semantic transformation type and the fracture location, alternative intermediate semantic islands are regenerated, and the original fracture cause is used as a negative constraint input to generate the model to avoid reintroducing the same sourceless semantics.

[0106] S333. Use load continuity result verification to replace intermediate semantic islands, connect the verified intermediate semantic islands to the semantic island chain, and convert the connected semantic island chain into topic units.

[0107] In practice, reverse recovery, new semantic verification, and lost semantic verification are performed again on the replacement intermediate semantic islands. If the replacement intermediate semantic island can accommodate the corresponding evidence payload unit, it is connected to the retained valid intermediate semantic islands, and subsequent intermediate semantic islands are generated. If the verification still fails, the allowed semantic transformation type is changed, or it is reverted to an earlier valid intermediate semantic island.

[0108] When all mandatory and coupled payloads in the semantic island chain remain continuous, the semantic representation at the end of the semantic island chain is converted into a topic unit. The topic unit records the root evidence identifier it receives, the intermediate semantic island path, and the payload status.

[0109] For multiple thematic units, creative themes are obtained by combining them according to the reasoning order in the stable path kernel, the irreplaceability of the evidence payload units, and the coupling relationship. For example, the two semantic island chains carrying "retro washed material" and "high-waisted cut" can be combined to form the main theme of "returning to the modern moments of the 1990s", and form two auxiliary thematic units of "naturally aged denim texture" and "sleek, upward-shifted visual waistline" respectively.

[0110] It should be noted that this application does not require a high degree of superficial semantic similarity between the creative theme and the product attributes. If only superficial similarity is pursued, the generated results are likely to remain at the level of restating product parameters, making it difficult to form marketing creativity. This application allows product attributes to generate far-reaching creative expressions through multiple intermediate semantic islands, but requires that the evidence payload unit remain continuous in each semantic leap, thereby balancing creative distance and product authenticity.

[0111] It should also be noted that an evidentiary payload unit can be split into multiple subject units. For example, "high-waisted cropping" can be divided into "raising the visual waistline" for text and "highlighting the waistline" for images. Multiple evidentiary payload units with coupling relationships can also be merged into the same subject unit. When splitting or merging occurs, the system records the root evidence identifier and coupling relationship of each subject unit for S5 to establish an evidentiary payload ledger.

[0112] S4. Construct shadow user concerns based on the target user concerns, execute competitive routing with evidence payload units as the sending end and the above two types of user concerns as the receiving end, and allocate evidence payload units based on routing differences to obtain differentiated selling points.

[0113] Among them, "target user concerns" refer to the factors that influence users' product choices, extracted from the target user profile. "Shadow user concerns" refer to the comparative concerns obtained by replacing the target user concerns while keeping other user characteristics in the target user profile unchanged. Shadow user concerns are used to determine whether a candidate selling point is truly triggered by target user concerns, rather than a generic selling point applicable to all users.

[0114] Competition routing refers to the transmission of the same evidence payload unit to the target user's point of interest and the shadow user's point of interest respectively along the bridging relationship between the user profile ontology and the product ontology, and determining which user's point of interest is more suitable to receive the evidence payload unit based on the reachability, conflict, and coupling relationship of the two routes.

[0115] The user profile ontology includes user characteristic nodes, user focus nodes, usage scenario nodes, user pain point nodes, and content preference nodes. The product ontology includes product attribute nodes, product function nodes, perceived results nodes, usage scenario nodes, and product evidence nodes. A bridging relationship is established between the user profile ontology and the product ontology to connect user focus, user needs, product functions, product attributes, and product evidence.

[0116] See attached document Figure 5 In this embodiment, S4 specifically includes S41, keeping other user features in the target user profile unchanged and replacing the target user's focus to obtain shadow user focus, and establishing candidate routes from the evidence payload unit to the two types of user focus based on the bridging relationship between the user profile ontology and the product ontology.

[0117] In practice, the target user profile is read, and one of the target user's concerns to be tested is set as the target receiver. While keeping the age range, spending power, region, occupation, channel preferences, and other non-tested user characteristics in the target user profile unchanged, concerns different from the target user's concerns are selected from a preset set of concern replacements to generate shadow user concerns.

[0118] For example, the target user profile might be working professionals who are concerned about commuting comfort and body proportions. When testing "commuting comfort," the focus can remain on workplace, age, spending power, and body proportions, while replacing "commuting comfort" with "weekend leisure comfort" to obtain the focus of shadow users.

[0119] Subsequently, starting from the product evidence node corresponding to each evidence payload unit, a path pointing to the target user's focus is searched within the product ontology and user profile ontology. Candidate routes can sequentially pass through product evidence, product attributes, product functions, usage scenarios, user needs, and user focus. For example, "straight-cut product evidence" reaches "commuting suitability target user focus" via "neat design product function" and "suitability for formal scenarios."

[0120] The system uses the same method to search for candidate routes pointing to the shadow user's points of interest, and records the ontology nodes, bridging relationships, conflict constraints, and root evidence identifiers that each candidate route passes through.

[0121] S42. The same evidence payload unit is transmitted to two types of user concerns along the candidate routes respectively. The route differences are obtained based on the route reachability state, route conflict state, and coupling relationship maintenance state.

[0122] The route reachability state indicates whether the evidence payload unit can reach the corresponding user's point of interest via a complete path supported by product evidence. The route conflict state indicates whether the candidate route conflicts with mutual exclusion relationships, brand language specifications, product selection functions, or other evidence payload units in the product ontology. The coupling relationship preservation state indicates whether coupled evidence payload units can reach the same user's point of interest along the candidate route or can be jointly expressed in mutually related selling points.

[0123] In practice, path integrity checks are performed on candidate routes pointing to the target user's point of interest and the shadow user's point of interest, respectively. If a candidate route can continuously reach the user's point of interest from the product evidence node, and each product function or perception result is supported by upstream product evidence, it is set to a reachable state; if a candidate route needs to introduce a function that does not exist in the product itself, it is set to an unreachable state.

[0124] Then check whether the candidate route passes through nodes associated with prohibited payloads, whether it re-excludes product attributes as positive selling points, or whether it violates prohibited brand expressions. If any of the above situations exist, set the candidate route to a conflict state.

[0125] For coupled evidence load units, check whether candidate routes maintain their combined effect. For example, when "retro wash material" and "high-waisted cut" jointly produce "90s silhouette", if a candidate route only uses high-waisted cut but outputs "90s retro texture", the coupling relationship is not maintained; if a candidate route includes both retro wash material and high-waisted cut, the coupling relationship is maintained.

[0126] The system comprehensively compares the route reachability, route conflict, route length, bridging strength, and coupling maintenance status of the same evidence payload unit to two types of user concerns to obtain the route difference. The route difference is used to indicate whether the evidence payload unit is more suitable for the target user's concerns than the shadow user concerns.

[0127] S43. Based on routing differences, exclusively allocate evidence payload units with routing advantages to the target user's focus, and set the occupied status of the allocated evidence payload units to suppress candidate selling points that repeatedly carry evidence payload units, thereby obtaining differentiated selling points.

[0128] Exclusive allocation means that once an evidence payload unit is fully carried by a differentiating selling point, it enters an occupied state. Other candidate selling points cannot repeatedly consume the same evidence payload unit with similar expressions, unless other candidate selling points also carry different new evidence payload units or are used for different necessary content modalities.

[0129] In practice, candidate routes are ranked according to the irreplaceability of the evidence payload unit and the routing differences. Priority is given to candidate routes that are reachable to the target user's point of interest, have no routing conflicts, maintain coupling, and have a routing advantage relative to the shadow user's point of interest.

[0130] Candidate selling points are generated based on the selected candidate routes. Each candidate selling point simultaneously records the target user's focus, the product functions traversed, the corresponding product attributes, and the root evidence identifier. The candidate selling points are then input into a selling point reverse parser. After confirming that they can recover the target user's focus and evidence payload unit, they are set as differentiated selling points.

[0131] When a differentiating selling point has already fully carried its corresponding evidentiary payload unit, that evidentiary payload unit is set to occupied. Subsequent candidate selling points that merely carry the same evidentiary payload unit without adding new user-differentiating information are then suppressed.

[0132] For example, after generating "High-waisted cut enhances the visual waistline, suitable for a sleek commuting look" based on the high-waisted cut, the system will no longer generate two similar selling points, "High-waisted design optimizes proportions" and "Raising the waistline makes it look more streamlined," separately. If another candidate selling point also carries evidence of a straight-leg silhouette and points to the commuting scenario, then "Straight-leg silhouette maintains the sleekness of a commuting look" can be generated.

[0133] When a candidate selling point has the same or similar routing status for the concerns of both the target user and the shadow user, it should be marked as a general selling point and not prioritized as a differentiating selling point. For example, "denim material is durable" applies to both commuting users and leisure users, so it should not be considered a differentiating selling point for working users simply because the tone has been changed.

[0134] It should be noted that if the competing route indicates that the target user's focus requires a certain mandatory payload, but the creative topic generated by S3 does not yet have a suitable topic unit to carry that mandatory payload, the system can reverse-locate the corresponding intermediate semantic island along the candidate route and generate a new topic branch without changing other evidential payload units. The new topic branch is then integrated into the creative topic after passing a payload continuity check. Therefore, the differentiated selling point is not mechanically added after the creative topic, but rather can use evidential payload units to reverse-compensate for payload gaps in the creative topic.

[0135] S5. Based on brand DNA, creative themes, and differentiated selling points, the evidence payload unit is projected onto a multimodal content carrier to generate multimodal marketing content and evidence payload ledger, and a constraint debt ledger is generated after compliance testing.

[0136] Brand DNA refers to the structured brand guidelines used to define the expression of brand content, including visual guidelines and language guidelines. Visual guidelines may include brand colors, secondary colors, color range, main product position, composition, model poses, shooting angle, lighting style, background elements, logo position, and prohibited visual elements; language guidelines may include commonly used brand words, taboo words, brand slogans, tone, sentence structure, copywriting structure, intensity of expression, and prohibited promotional methods.

[0137] Multimodal content carriers refer to content units capable of carrying evidentiary payloads and recognizable by corresponding detectors, including text fragments, image regions, video footage, subtitle fragments, and narration fragments. Multimodal content carriers do not refer solely to complete text or images, but rather to content regions that can be individually located, generated, detected, and corrected.

[0138] Evidence payload unit projection refers to converting evidence payload units into modal projection codes suitable for the generation and detection of corresponding content modalities, and generating multimodal content carriers based on the modal projection codes. Modal projection codes are used to describe the identifiable representation of the corresponding evidence payload unit in a specific content modality, including text projection codes, image projection codes, video projection codes, and voice projection codes.

[0139] For example, for the evidentiary payload unit corresponding to "high-waisted cropping," its text projection code may include "raising the visual waistline" and "high-waisted silhouette"; its image projection code may include the waist cropping area being fully visible, the waist area not being obscured by the upper garment, and the waist area being in the area of ​​visual attention; its video projection code may include a display shot moving from the waist to the legs; and its audio projection code may include "high-waisted straight silhouette" for narration. Each modal projection code is associated with the same root evidence identifier.

[0140] The evidence payload ledger is a collection of state data used to record the carrying status and changes of evidence payload units in multimodal marketing content. The evidence payload ledger records at least the root evidence identifier, payload type, current content carrier, current carrying modality, payload status, coupling relationship, payload version, and historical migration records.

[0141] The constraint debt ledger is a data set used to record unmet constraints of multimodal marketing content and their processing status. The constraint debt ledger records at least the constraint debt identifier, constraint source, constraint object, constraint priority, associated multimodal content carrier, associated evidentiary payload unit, constraint conflict relationship, and constraint debt status. Constraint sources include brand DNA, channel rules, product authenticity rules, product selection logic rules, and cross-modal consistency rules.

[0142] In practice, the system first converts the various brand specifications in the brand DNA into generation control items, compliance check items, and violation remediation items. Generation control items are used to control the input or generation parameters of the generation model before generating multimodal marketing content; compliance check items are used to determine whether the corresponding brand specifications are met after generation; and violation remediation items are used to determine the appropriate local remediation methods when the checks fail.

[0143] For example, brand primary color specifications can be converted into color condition vectors used by image generation models, compliance detection items for detecting the primary color and its distribution in an image, and local color restoration items for adjusting background colors. Model pose specifications can be converted into pose control maps, human keypoint detection items, and pose region redrawing items. Taboo word specifications can be converted into forbidden word control items, text matching detection items, and sentence-level replacement items for generation models.

[0144] Subsequently, the system determines the target content carrier for each evidence payload unit based on the creative theme, differentiated selling points, brand DNA, and target channels. For product material, layout, and composition features suitable for direct visual expression, they can be prioritized for allocation to image areas or video shots; for product functions, user benefits, and brand tone, they can be prioritized for allocation to text snippets, subtitle snippets, or narration snippets.

[0145] For each evidence payload unit, a corresponding modal projection code is generated, and the modal projection code is checked to ensure it conforms to the allowed carrying modality. If the evidence payload unit is a coupled payload, it is checked whether the selected multiple target content carriers can maintain a common coupling relationship. For coupled payloads that must be expressed by the same content unit, multiple evidence payload units are merged and projected onto the same multimodal content carrier; for coupled payloads that allow cross-modal co-expression, they can be projected onto interrelated text fragments and image regions respectively, and a cross-modal correspondence is established.

[0146] When generating e-commerce detail pages, the system can project "retro washed material" onto the material close-up image area and "high-waisted cut" onto the model's front-facing image area and selling point copywriting fragments; when generating short video content, the above evidence load units can be projected onto material detail shots, waistline shots, subtitle fragments, and narration fragments, respectively.

[0147] The system constructs corresponding multimodal generation tasks based on the projection codes of each modality. Text generation tasks include at least a creative theme, differentiating selling points, target user concerns, root evidence identifiers, permissible expression range, brand language specifications, and channel structure requirements. Image generation tasks include at least the product subject, scene description, visual theme, visible area requirements corresponding to the evidence payload unit, brand visual specifications, and prohibited visual elements. Video generation tasks include at least the shot sequence, evidence payload units carried by each shot, image description, subtitles, narration, and brand visual specifications.

[0148] The multimodal generation task is input into the corresponding generation model to obtain initial multimodal marketing content. The system uses a text parser, image recognition model, and video content analysis model to detect whether each multimodal content carrier receives the assigned evidence payload unit. Only when the content in the multimodal content carrier can be used to reverse-engineer the root evidence identifier and no product function without supporting product evidence is introduced, is the corresponding payload status set to a valid state.

[0149] After confirming the carrying status, an evidence payload ledger is established. For each evidence payload unit, its current content carrier and current carrying mode are recorded; for evidence payload units jointly carried by multiple content carriers, the carrying range of each content carrier is recorded separately; for mandatory payloads that have not yet been effectively carried by any content carrier, their load status is set to a gap state.

[0150] Furthermore, compliance checks are performed on the initial multimodal marketing content. These checks include brand visual checks, brand language checks, product authenticity checks, product selection logic consistency checks, channel compatibility checks, and cross-modal consistency checks. Brand visual checks determine whether colors, composition, posture, and prohibited visual elements align with the brand DNA; brand language checks determine whether prohibited words, tone, and sentence structure are compliant; product authenticity checks determine whether the content provides evidence of the product; product selection logic consistency checks determine whether the mandatory load corresponding to the stable path core is still being carried; channel compatibility checks determine whether the content format meets the requirements of the target channel; and cross-modal consistency checks determine whether the copy, images, videos, subtitles, and narration express consistent product attributes.

[0151] Constraints that fail compliance testing are converted into constrained liabilities and recorded in the constrained liability ledger. For example, when a taboo word appears in the copy, a brand language constrained liability is established; when the waist area of ​​the product in the image is obscured, making the "high-waisted cut" unrecognizable, both a visual composition constrained liability and an evidentiary load gap are established; when the copy expresses "waterproof performance" but there is no corresponding evidence in the product itself, a product authenticity constrained liability is established.

[0152] It should be noted that the same detection problem can be associated with multiple constraint liabilities simultaneously, and the same constraint liability can also be associated with multiple multimodal content carriers. The system establishes a "product evidence - content carrier - constraint liability" relationship through the evidence payload ledger and the constraint liability ledger, thereby providing a data foundation for S6 to determine whether a correction operation will lead to the loss of core product attributes.

[0153] S6. Based on the evidence payload ledger and the constraint debt ledger, perform the correction transaction. When the correction generates a payload gap, first compensate the evidence payload unit corresponding to the payload gap to the alternative content carrier, then release the original content carrier, and output multimodal marketing content that passes the compliance test.

[0154] In this context, a corrective transaction refers to a set of content corrective operations generated for one or more constraint liabilities and executed in an isolated copy. Corrective transactions use constraint liability reduction and evidentiary payload preservation as common commit conditions, rather than solely relying on the removal of currently violating content.

[0155] A load gap refers to a situation where a mandatory load or a coupled load that needs to be maintained is not carried by any valid multimodal content carrier after the execution of a candidate correction transaction, or where, although a relevant expression exists, the state of the root evidence identifier cannot be recovered in reverse.

[0156] An alternative content carrier refers to another multimodal content carrier capable of receiving the same evidentiary payload unit as the original content carrier. Multiple multimodal content carriers that carry the same evidentiary payload unit and are mutually substitutable constitute a content carrier equivalence cluster. A content carrier equivalence cluster can include content carriers of different modalities, or different content positions within the same modality.

[0157] In this embodiment, S6 specifically includes S61, reading the evidence payload ledger and the constraint debt ledger, and generating candidate correction transactions based on the association between constraint debts and multimodal content carriers.

[0158] See attached document Figure 6 Specifically, S61 includes S611, which constructs a content dependency graph based on multimodal content carriers, evidence payload units, constraint obligations, and the evidence carrying relationships, semantic connection relationships, cross-modal correspondence relationships, and constraint dependency relationships between them.

[0159] The nodes in the content dependency graph include multimodal content carrier nodes, evidence payload unit nodes, and constraint debt nodes. Evidence payload relationships connect evidence payload units to the multimodal content carriers that carry them; semantic connection relationships connect content carriers with referential, sequential, or explanatory relationships; cross-modal correspondence relationships connect different modal content carriers expressing the same thematic unit or differentiating selling points; and constraint dependency relationships connect constraint debts to the content carriers that cause those constraint debts or generation control items.

[0160] In practice, the evidence carrying relationship between the evidence payload unit node and the current content carrier node is established based on the evidence payload ledger, the cross-modal correspondence is established based on the theme unit and selling point identifier recorded in the multimodal generation task, and the constraint dependency relationship between the constraint debt node and the illegal content carrier is established based on the constraint debt ledger, thus obtaining the content dependency graph.

[0161] S612. Starting from the multimodal content carrier associated with the constraint debt to be processed, propagate the load missing state and constraint debt changes in the content dependency graph, and truncate the propagation when it reaches the content carrier that does not carry the corresponding evidence load unit to obtain the modified influence cone.

[0162] Among them, the correction impact cone refers to the set of evidence payload units, multimodal content carriers, and constraint obligations that may be directly or indirectly affected by the candidate correction operation.

[0163] For example, when a text fragment to be deleted carries a mandatory load corresponding to "high-waisted cropping", the load missing state propagates from the text fragment to the evidence load unit, and then to the image area and subtitle fragment that have a cross-modal correspondence with it, in order to determine whether other multimodal content carriers can continue to carry the mandatory load.

[0164] If the content carrier being propagated does not carry the corresponding evidence payload unit and has no dependency relationship with the current binding debt, then propagation along that branch will stop to avoid local modifications from expanding into the overall regeneration of irrelevant content.

[0165] S613. Based on the modified impact cone, determine the load risk and constraint liability changes caused by each modified operation, select the modified operation that reduces the constraint liability and maintains the evidentiary load unit, and generate candidate modified transactions.

[0166] Load risk can be determined based on whether the correction operation involves mandatory loads, whether it disrupts coupling relationships, whether cross-modal migration is required, and whether it introduces prohibited loads. Changes in constraint liabilities include eliminated constraint liabilities, potentially new constraint liabilities, and their priorities.

[0167] For example, if the copy exceeds the length limit, corrective actions can be taken, such as deleting modifiers, merging duplicate selling points, or migrating some selling points to the subtitles. If deleting modifiers does not affect the evidentiary load, then the corresponding candidate corrective action should be generated first; if deleting core selling points will create a mandatory load gap, then the candidate corrective action should also include load compensation operations.

[0168] S62. Execute candidate correction transactions in the multimodal marketing content copy and ledger copy, detect the constraint liabilities, load states and coupling relationships before and after execution, and obtain transaction verification results.

[0169] In practice, the current multimodal marketing content, evidentiary payload ledger, and constraint liability ledger are copied to create copies of the multimodal marketing content and the ledgers. Candidate correction transactions are first executed in the copies, without directly modifying the currently valid content.

[0170] See attached document Figure 7 In this embodiment, S62 specifically includes S621, when the correction influence cone indicates that a candidate correction transaction generates a load gap, selecting an alternative content carrier from the equivalent cluster of content carriers carrying the same evidence load unit, and pre-occupying the corresponding evidence load unit to the alternative content carrier.

[0171] In practice, the equivalent cluster of content carriers is queried based on the allowed modalities of the evidence payload unit. If deleting the core selling points in the text would create a payload gap, existing image areas, video shots, or subtitle fragments can be selected as alternative content carriers; if none of the existing content carriers can effectively accommodate the payload, a new alternative content carrier is generated based on the corresponding modal projection code.

[0172] The system first generates content corresponding to the root evidence identifier in the alternative content carrier and sets the evidence payload unit to a pre-occupied state. At this time, the original content carrier remains valid to avoid temporary loss of the evidence payload unit if the generation of the alternative content carrier fails.

[0173] S622. Detect the payload acceptance status, compliance status, and coupling relationship of the alternative content carrier. When the detection is passed, switch the alternative content carrier to the valid state, release the original content carrier, and update the evidence payload ledger and the constraint debt ledger.

[0174] In practice, a detector corresponding to the alternative content carrier modality is used to determine whether the root evidence identifier can be recovered in reverse, whether the brand DNA and channel rules are met, and whether the coupling relationship of the evidence payload unit is maintained.

[0175] After the detection is passed, the replacement content carrier is switched from the pre-occupied state to the active state, and then the original content carrier is set to the released state and deleted or modified. The above "compensate first, then release" execution order can avoid the situation where the original content carrier has been deleted before the replacement content carrier has effectively carried the evidence payload unit.

[0176] Subsequently, the current content carrier in the evidence payload ledger is updated to the alternative content carrier, the payload version is added, and the migration history is recorded; at the same time, the eliminated constraint liabilities are set to the processed status, and any new detection results that may be generated by the alternative content carrier are written into the constraint liability ledger.

[0177] S623. If the test fails, cancel the pre-occupancy and retain the original content carrier, and generate the transaction verification result based on the updated evidence payload ledger and constraint debt ledger.

[0178] If the alternative content carrier fails to restore the root evidence identifier, violates the brand DNA, or disrupts the coupling relationship, the pre-positioning status is cancelled, the unsubmitted alternative content carriers are deleted, and the original content carrier is retained. The system can also select another alternative content carrier from the content carrier equivalence cluster, or re-execute compensation after changing the modal projection code.

[0179] The transaction verification results record the constraint debt status, mandatory load status, coupling relationship status, prohibited load status, and content version status before and after the execution of the candidate modified transaction, and are provided to S63.

[0180] S63. If the transaction verification results show that the constraint debt has decreased, the mandatory evidence payload unit has been retained, the coupling relationship has not been broken, and no new higher priority constraint debt has been added, submit the candidate correction transaction; otherwise, roll back the multimodal marketing content copy and the ledger copy.

[0181] In practice, the system compares the dual ledger states before and after the candidate correction transaction is executed. If the currently pending constraint debts are eliminated or reduced, all mandatory loads remain valid, coupled loads still satisfy the common bearing rules, prohibited loads are not introduced, and no new constraint debts with higher priority are generated, then the multimodal marketing content copy and the ledger copy are submitted as new valid versions.

[0182] If any of the above submission conditions are not met, the multimodal marketing content copy and ledger copy will be discarded, the multimodal marketing content, evidentiary payload ledger, and constraint liability ledger before the execution of the corrective transaction will be restored, and other candidate corrective transactions will be selected.

[0183] For example, the phrase "absolutely slimming" in the copy violates brand taboo word guidelines. While directly deleting this sentence would eliminate the brand's linguistic constraint debt, it would create a load gap if it were the sole content carrier of the mandatory load for "high-waisted cut." The system first enhances the visibility of the waist area in the product display image and generates an equivalent expression, "high-waisted straight silhouette," in the captions. Only after both the image area and the caption segment pass the load-bearing detection are the original violating phrase deleted. Thus, the taboo word is eliminated while the mandatory load corresponding to the high-waisted cut is preserved.

[0184] It should be noted that for multiple rounds of revisions, the system can construct an evidence version chain based on the payload version and compare the current content carrier with both the previous valid content carrier and the root product evidence. If the changes between the current content carrier and the previous valid content carrier are minor, but the root product evidence cannot be recovered after multiple rounds of rewriting, cumulative semantic decay is determined to have occurred. In this case, it can revert to the most recent content carrier in the evidence version chain that can recover the root product evidence, reselect a semantic transformation path, or replace the content carrier, thereby preventing the core product attributes from being gradually abstracted and eventually lost during multiple rounds of local rewriting.

[0185] Ultimately, when there are no pending high-priority constraint debts in the constraint debt ledger, all mandatory loads in the evidentiary load ledger are in a valid state, coupling relationships are maintained, and prohibitive loads are not introduced, the current valid version of multimodal marketing content is output. This achieves a closed-loop process from stable product selection logic, creative themes, differentiated selling points to brand compliance correction.

[0186] This application also provides an automated generation system for multimodal marketing content in brand retail, which can be applied to the methods of any of the above embodiments.

[0187] See attached document Figure 8 A brand retail multimodal marketing content automated generation system includes a product selection reasoning module 1, an evidence payload generation module 2, a theme generation module 3, a selling point generation module 4, a content generation module 5, a compliance detection module 6, and a correction output module 7. Product selection reasoning module 1 is used to obtain multiple product selection reasoning paths of the product selection model, align the product selection reasoning paths based on product evidence and product selection function, and extract stable path kernels; The evidence payload generation module 2 is connected to the product selection reasoning module 1. It is used to cut off the stable path kernel and replay the product selection model. Based on the obtained product selection response, it generates an evidence payload unit containing payload parameters, including root evidence identifier, coupling relationship and payload state. The topic generation module 3, connected to the evidence payload generation module 2, is used to transfer the evidence payload unit to the topic unit through the intermediate semantic island, and combine the topic units when the payloads are continuous to obtain the creative topic. The selling point generation module 4 is connected to the evidence payload generation module 2. It is used to construct shadow user concerns based on the concerns of target users, execute competitive routing with evidence payload units as the sending end and target user concerns and shadow user concerns as the receiving end, and allocate evidence payload units based on routing differences to obtain differentiated selling points. Content generation module 5 is connected to evidence payload generation module 2, theme generation module 3 and selling point generation module 4 respectively. It is used to project evidence payload units onto multimodal content carriers based on brand DNA, creative themes and differentiated selling points to generate multimodal marketing content and evidence payload ledger. The compliance detection module 6, connected to the content generation module 5, is used to perform compliance detection on multimodal marketing content and generate a constraint debt ledger. The correction output module 7 is connected to the evidence payload generation module 2, the content generation module 5, and the compliance detection module 6. It is used to perform correction transactions based on the evidence payload ledger and the constraint debt ledger. When the correction generates a payload gap, it first obtains the evidence payload unit corresponding to the payload gap from the evidence payload generation module 2 to compensate for the replacement content carrier, and then releases the original content carrier to output multimodal marketing content that passes the compliance detection.

[0188] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database contains data related to a method for automatically generating multimodal marketing content for brand retail. The network interface is used for communication with external terminals via a network connection. The computer program, when executed by the processor, implements a method for automatically generating multimodal marketing content for brand retail.

[0189] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," "third," and similar terms used in this application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. The terms "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms "comprising" or "including" and similar terms mean that the elements or objects preceding "comprising" or "including" encompass the elements or objects listed following "comprising" or "including" and their equivalents, and do not exclude other elements or objects. "Above," "below," "left," "right," etc., are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0190] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for automatically generating multimodal marketing content for brand retail, characterized in that: This includes S1, obtaining multiple product selection reasoning paths from the product selection model, aligning the product selection reasoning paths based on product evidence and product selection function, and extracting stable path kernels; S2. Cut off the stable path kernel and replay the product selection model, and generate an evidence load unit containing load parameters based on the obtained product selection response; The load parameters include root evidence identifiers, coupling relationships, and load states; S3. The evidence payload unit is passed to the topic unit through the intermediate semantic island. When the payloads are continuous, the topic units are combined to obtain the creative topic. S4. Construct shadow user concerns based on the concerns of target users, perform contention routing with the evidence payload unit as the sending end and the above two types of user concerns as the receiving end, and allocate the evidence payload unit based on the routing differences to obtain differentiated selling points. S5. Based on the brand DNA, the creative theme, and the differentiated selling points, the evidence payload unit is projected onto the multimodal content carrier to generate multimodal marketing content and evidence payload ledger, and a constraint debt ledger is generated after compliance testing; S6. Based on the evidence payload ledger and the constraint debt ledger, perform a correction transaction. When the correction generates a payload gap, first compensate the evidence payload unit corresponding to the payload gap to the alternative content carrier, then release the original content carrier, and output multimodal marketing content that passes the compliance test.

2. The method for automatically generating multimodal marketing content for brand retail according to claim 1, characterized in that: S1 specifically includes performing perturbations on the product attributes, candidate product context, and reasoning state to maintain the evidence of the target product unchanged, and obtaining the product selection reasoning path corresponding to each perturbation state to form a set of product selection path graphs; Based on the location of product evidence and the role of product selection, the nodes and reasoning relationships in the product selection path graph set are aligned to form an aligned path structure; The stable path kernel is obtained by extracting the common inference substructure that maintains the root commodity evidence and selection effect before and after the perturbation from the alignment path structure and changes the selection result after severance.

3. The method for automatically generating multimodal marketing content for brand retail according to claim 2, characterized in that: The evidence payload unit specifically includes the root evidence identifier, product evidence location, product selection function type, degree of irreplaceability, payload type, coupling relationship, allowed semantic transformation type, allowed carrying modality, current content carrier, payload status, and payload version; The allowed semantic transformation type is used to construct the intermediate semantic island; the allowed bearer modality is used to determine the multimodal content carrier; the payload type, the payload state, and the payload version are used to control the correction transaction.

4. The method for automatically generating multimodal marketing content for brand retail according to claim 3, characterized in that: S3 specifically includes S31, converting the evidence payload unit into an initial intermediate semantic island according to the allowed semantic transformation type, and binding the initial intermediate semantic island to the root evidence identifier; S32. Generate a next intermediate semantic island based on the previous intermediate semantic island and its load status, and record the newly added semantics, lost semantics and commodity evidence sources between its adjacent intermediate semantic islands to form a semantic island chain. S33. Detect the load continuity of the semantic island chain based on the newly added semantics, the lost semantics, and the source of the product evidence; repair the intermediate semantic islands with discontinuous loads; and convert the semantic island chain with continuous loads into the topic unit to obtain the creative topic.

5. The method for automatically generating multimodal marketing content for brand retail according to claim 4, characterized in that: S33 specifically includes recovering the evidence payload unit from adjacent intermediate semantic islands in reverse, verifying the newly added semantics and the lost semantics according to the root evidence identifier and the source of the commodity evidence, and obtaining the payload continuity result; Based on the load continuity results, identify the fracture intermediate semantic islands, retain the valid intermediate semantic islands before the fracture intermediate semantic islands, and regenerate alternative intermediate semantic islands by selecting different allowed semantic transformation types. The alternative intermediate semantic islands are verified using the load continuity results. The verified alternative intermediate semantic islands are then connected to the semantic island chain, and the connected semantic island chain is converted into the topic unit.

6. The method for automatically generating multimodal marketing content for brand retail according to claim 1, characterized in that: S4 specifically includes keeping other user features in the target user profile unchanged and replacing the target user's focus to obtain the shadow user's focus, and establishing candidate routes from the evidence payload unit to the two types of user focus based on the bridging relationship between the user profile body and the product body. The same evidence payload unit is transmitted to the two types of user concerns along the candidate routes respectively, and the route differences are obtained based on the route reachability state, route conflict state, and coupling relationship maintenance state. Based on the routing differences, the evidence payload units with routing advantages are exclusively allocated to the target user's focus, and the allocated evidence payload units are set to an occupied state to suppress candidate selling points that repeatedly carry the evidence payload units, thereby obtaining the differentiated selling points.

7. The method for automatically generating multimodal marketing content for brand retail according to claim 1, characterized in that: S6 specifically includes S61, reading the evidence payload ledger and the constraint debt ledger, and generating candidate correction transactions based on the association between constraint debts and multimodal content carriers; S62. Execute the candidate correction transaction in the multimodal marketing content copy and ledger copy, detect the constraint liabilities, load states and coupling relationships before and after execution, and obtain the transaction verification results; S63. If the transaction verification results indicate that the constraint debt has decreased, the mandatory evidence payload unit has been retained, the coupling relationship has not been broken, and no new higher priority constraint debt has been added, submit the candidate correction transaction; otherwise, roll back the multimodal marketing content copy and the ledger copy.

8. The method for automatically generating multimodal marketing content for brand retail according to claim 7, characterized in that: S61 specifically includes S611, constructing a content dependency graph based on the multimodal content carrier, the evidence payload unit, the constraint debt, and the evidence carrying relationship, semantic connection relationship, cross-modal correspondence relationship and constraint dependency relationship between them; S612. Starting from the multimodal content carrier associated with the constraint debt to be processed, propagate the load missing state and constraint debt change in the content dependency graph, and truncate the propagation when it reaches the content carrier that does not carry the corresponding evidence load unit to obtain the modified influence cone; S613. Based on the modified influence cone, determine the load risk and constraint liability changes caused by each modified operation, select the modified operation that reduces the constraint liability and maintains the evidence load unit, and generate the candidate modified transaction.

9. The method for automatically generating multimodal marketing content for brand retail according to claim 8, characterized in that: S62 specifically includes S621, when the correction influence cone indicates that the candidate correction transaction generates a load gap, selecting the alternative content carrier from the content carrier equivalent cluster carrying the same evidence load unit, and pre-occupying the corresponding evidence load unit to the alternative content carrier. S622. Detect the payload acceptance status, compliance status, and coupling relationship of the alternative content carrier. When the detection passes, switch the alternative content carrier to an effective state, release the original content carrier, and update the evidence payload ledger and the constraint debt ledger. S623. If the detection fails, cancel the pre-occupancy and retain the original content carrier, and generate the transaction verification result based on the updated evidence payload ledger and the constraint debt ledger.

10. A system for automatically generating multimodal marketing content for brand retail, characterized in that: It includes a product selection reasoning module (1), an evidence load generation module (2), a theme generation module (3), a selling point generation module (4), a content generation module (5), a compliance detection module (6), and a correction output module (7); The product selection reasoning module (1) is used to obtain multiple product selection reasoning paths of the product selection model, align the product selection reasoning paths based on product evidence and product selection function, and extract stable path kernels; The evidence payload generation module (2) is connected to the product selection reasoning module (1) and is used to cut off the stable path kernel and replay the product selection model. Based on the obtained product selection response, an evidence payload unit containing payload parameters is generated. The payload parameters include root evidence identifier, coupling relationship and payload state. The topic generation module (3) is connected to the evidence payload generation module (2) and is used to transmit the evidence payload unit to the topic unit through the intermediate semantic island. When the payload is continuous, the topic unit is combined to obtain the creative topic. The selling point generation module (4) is connected to the evidence payload generation module (2) and is used to construct shadow user attention points based on the target user attention points, execute competitive routing with the evidence payload unit as the sending end and the target user attention points and shadow user attention points as the receiving end, and allocate the evidence payload unit based on the routing differences to obtain differentiated selling points. The content generation module (5) is connected to the evidence payload generation module (2), the theme generation module (3) and the selling point generation module (4) respectively, and is used to project the evidence payload unit onto the multimodal content carrier based on the brand DNA, the creative theme and the differentiated selling point to generate multimodal marketing content and evidence payload ledger; The compliance detection module (6) is connected to the content generation module (5) and is used to perform compliance detection on the multimodal marketing content and generate a constraint debt ledger; The correction output module (7) is connected to the evidence payload generation module (2), the content generation module (5) and the compliance detection module (6), and is used to perform correction transactions based on the evidence payload ledger and the constraint debt ledger. When the correction generates a payload gap, the evidence payload unit corresponding to the payload gap is first obtained from the evidence payload generation module (2) to compensate the alternative content carrier, and then the original content carrier is released to output multimodal marketing content that passes the compliance detection.

Images