Knowledge graph and RAG-based cross-border e-commerce marketing content generation method and system

Abstract

This invention discloses a method and system for generating cross-border e-commerce marketing content based on knowledge graphs and RAG (Relational Architecture Framework), belonging to the field of cross-border e-commerce marketing technology. The method includes: acquiring target e-commerce demand information and target market localized data to construct a joint constraint knowledge graph; performing a hybrid retrieval based on the RAG architecture and the joint constraint knowledge graph to obtain candidate knowledge graph nodes and relationship paths; performing scene calibration, behavior calibration, and language calibration based on the candidate nodes and relationship paths to generate localized calibration information; constructing a constraint prompt sequence from the localized calibration information and original product data, and inputting it into a content generation model to generate e-commerce marketing content adapted to the target market; and simultaneously updating the knowledge graph based on actual delivery feedback. This invention can reduce scene mismatch, behavior mismatch, and language mismatch in cross-border e-commerce marketing content generation, and improve the consistency between the generated content and the constraints of the target market.

Description

Technical Field

[0001] This invention relates to the field of cross-border e-commerce marketing technology, and more specifically, to a method and system for generating cross-border e-commerce marketing content based on knowledge graphs and RAG. Background Technology

[0002] With the rapid development of global e-commerce, cross-border merchants often face the challenge of selling the same products to multiple countries and regions. However, the global market is becoming increasingly diverse, with consumers in different regions having different aesthetics, consumption habits, and cultural backgrounds. Traditional systems mostly process data from different countries within the same model, leading to biases in recommendation strategies in markets with different cultural backgrounds. To improve marketing relevance, marketing content (product titles, descriptions, advertising copy, etc.) must be deeply localized, not just translated at the language level.

[0003] Existing technologies mainly include machine translation (MT) technology, direct generation of general large language models (LLM), and ordinary RAG systems. However, these existing technologies still have significant shortcomings: Machine translation can only perform word-for-word translation and cannot establish a structured relationship between product attributes, target market life scenarios, user behavior, and local expressions; General-purpose large language models directly generate language based on general corpora, which can easily lead to internationally neutral expressions and lack constraints on local life scenarios and user behavior in the target market. Ordinary RAG systems mainly retrieve internal enterprise documents or general knowledge, lacking a local culture and user behavior knowledge graph specifically for cross-border e-commerce marketing content generation, and also lacking a three-layer calibration mechanism of scenario, behavior, and language.

[0004] Therefore, in order to address the above problems, there is an urgent need to propose a method and system for generating cross-border e-commerce marketing content based on knowledge graphs and RAG. Summary of the Invention

[0005] To overcome or at least partially solve the above problems, this invention provides a method and system for generating cross-border e-commerce marketing content based on knowledge graphs and RAGs, which can reduce scenario mismatch, behavior mismatch, and language mismatch in cross-border e-commerce marketing content generation, and improve the consistency between the generated content and the constraints of the target market.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: Firstly, a method for generating cross-border e-commerce marketing content based on knowledge graphs and RAGs includes the following steps: S1. Obtain target e-commerce demand information, which includes at least target market identifier, product category, product structured attribute data, product text data, content output type, and target platform identifier; S2. Obtain the target market localization data corresponding to the target market identifier; S3. Based on the target market localization data and product structured attribute data, construct a joint constraint knowledge graph of target market local culture and user behavior; S4. Construct a cross-border e-commerce retrieval model based on the RAG architecture and the aforementioned joint constraint knowledge graph; S5. Generate a query tuple based on the target e-commerce demand information; S6. Based on the query tuple, perform a hybrid search in the cross-border e-commerce retrieval model to obtain candidate knowledge graph nodes and candidate relationship paths; S7. Calculate the comprehensive fit score based on the candidate knowledge graph nodes and candidate relationship paths, and determine the target calibration node and target calibration path based on the comprehensive fit score; S8. Perform localized mapping calibration based on the target calibration node and the target calibration path to generate localized calibration information; the localized mapping calibration includes scene calibration, behavior calibration and language calibration; S9. Construct a constraint prompt sequence based on the original product data in the localized calibration information and the target e-commerce demand information, and input the constraint prompt sequence into a pre-built content generation model to generate e-commerce marketing content adapted to the target market; S10. Collect user behavior feedback data after the e-commerce marketing content is launched in the target market, and update the node weights and relationship weights in the joint constraint knowledge graph based on the user behavior feedback data.

[0007] This invention uses a combined constrained knowledge graph and RAG retrieval mechanism to perform structured calibration of product attributes, target market scenarios, user behavior, and language expression before content generation, thereby reducing scenario mismatch, behavior mismatch, and language mismatch in cross-border e-commerce marketing content generation.

[0008] Based on the first aspect, further, the aforementioned target market localization data includes at least local life scenario data, user behavior habit data, local language expression data, cultural taboo data, and platform content style data.

[0009] Based on the first aspect, the aforementioned joint constraint knowledge graph further includes product attribute nodes, life scenario nodes, user behavior nodes, language expression nodes, cultural taboo nodes, and platform constraint nodes; the nodes are connected through adaptation relationships, conflict relationships, replacement relationships, triggering relationships, and disabling relationships.

[0010] Based on the first aspect, furthermore, each of the above nodes and various relationships has a weight, confidence level, source identifier, and update time identifier.

[0011] Based on the first aspect, the above-mentioned hybrid retrieval further includes keyword recall, vector similarity recall, and graph path recall.

[0012] Based on the first aspect, the above-mentioned comprehensive adaptation score further includes semantic similarity, graph path weight, scene adaptation, behavior matching, language adaptation, and risk constraint value.

[0013] Based on the first aspect, further, the aforementioned constraint prompt sequence includes the target market, original product data, scene calibration information, behavior calibration information, language calibration information, prohibited expressions, and output formats.

[0014] Based on the first aspect, the aforementioned e-commerce marketing content further includes product titles, details page copy, advertising copy, short video scripts, live streaming scripts, and cover copy.

[0015] Secondly, this invention provides a cross-border e-commerce marketing content generation system based on knowledge graphs and RAG, including a demand data acquisition module, a localized data acquisition module, a knowledge graph construction module, a retrieval model construction module, a query tuple generation module, a hybrid retrieval module, a rating and ranking module, a localized calibration module, a marketing content generation module, and a feedback and update module, wherein: The demand data acquisition module is used to acquire target e-commerce demand information, which includes at least the target market identifier, product category, product structured attribute data, product text data, content output type, and target platform identifier. The localization data acquisition module is used to acquire target market localization data corresponding to the target market identifier; The graph construction module is used to construct a knowledge graph jointly constrained by the local culture and user behavior of the target market based on the target market localized data and product structured attribute data. The retrieval model construction module is used to construct a cross-border e-commerce retrieval model based on the RAG architecture and the joint constraint knowledge graph. The query tuple generation module is used to generate query tuples based on the target e-commerce demand information; The hybrid retrieval module is used to perform a hybrid retrieval in the cross-border e-commerce retrieval model based on the query tuple to obtain candidate knowledge graph nodes and candidate relationship paths. The scoring and ranking module is used to calculate the comprehensive adaptation score based on the candidate knowledge graph nodes and candidate relationship paths, and to determine the target calibration node and target calibration path based on the comprehensive adaptation score. The localization calibration module is used to perform localized mapping calibration based on the target calibration node and the target calibration path, and generate localized calibration information; the localized mapping calibration includes scene calibration, behavior calibration and language calibration; The marketing content generation module is used to construct a constraint prompt sequence based on the original product data in the localized calibration information and the target e-commerce demand information, and input the constraint prompt sequence into the pre-built content generation model to generate e-commerce marketing content adapted to the target market. The feedback update module is used to collect user behavior feedback data after the e-commerce marketing content is launched in the target market, and update the node weights and relationship weights in the joint constraint knowledge graph based on the user behavior feedback data.

[0016] This system, through the collaboration of multiple modules including a demand data acquisition module, a localized data acquisition module, a knowledge graph construction module, a retrieval model construction module, a query tuple generation module, a hybrid retrieval module, a scoring and ranking module, a localization calibration module, a marketing content generation module, and a feedback and update module, utilizes a joint constrained knowledge graph and RAG retrieval mechanism to perform structured calibration of product attributes, target market scenarios, user behavior, and language expression before content generation. This reduces scenario mismatch, behavior mismatch, and language mismatch in cross-border e-commerce marketing content generation.

[0017] This invention has at least the following advantages or beneficial effects: 1. This invention constructs a joint constraint knowledge graph that includes product attributes, life scenarios, user behavior, language expression, cultural taboos, and platform constraints, enabling the generative model to acquire structured local constraints before generation, thereby reducing the risk of product attributes not matching the target market scenario; 2. This invention improves the consistency between recall nodes and target market demand by using a hybrid retrieval and comprehensive adaptation scoring mechanism, which makes the retrieval results not only rely on text similarity, but also combine graph path, scene adaptation, behavior matching, language adaptation and risk constraints. 3. This invention transforms the original product selling points into marketing content that is understandable and acceptable to the target market and conforms to the platform's expression habits through scene calibration, behavior calibration, and language calibration, thereby reducing literal translation, culturally incompatible expressions, and platform-incompatible expressions. 4. This invention uses a feedback update mechanism to adjust the weights of graph nodes and relationships based on user behavior feedback after actual deployment, so that subsequent content generation can be continuously optimized as the target market changes. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1This is a flowchart illustrating a method for generating cross-border e-commerce marketing content based on knowledge graphs and RAG, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of a cross-border e-commerce marketing content generation system based on knowledge graphs and RAG, according to an embodiment of the present invention.

[0020] Figure labeling: 100, Demand data acquisition module; 200, Localized data acquisition module; 300, Graph construction module; 400, Retrieval model construction module; 500, Query tuple generation module; 600, Hybrid retrieval module; 700, Rating and sorting module; 800, Localized calibration module; 900, Marketing content generation module; 1000, Feedback update module. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0022] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0023] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0024] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the term "comprising" or any other variations thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0025] In the description of the embodiments of the present invention, "multiple" means at least two. Example

[0026] like Figure 1 As shown, in the first aspect, a method for generating cross-border e-commerce marketing content based on knowledge graphs and RAG includes the following steps: S1. Obtain target e-commerce demand information, which includes at least target market identifier, product category, product structured attribute data, product text data, content output type, and target platform identifier; S2. Obtain the target market localization data corresponding to the target market identifier; the target market localization data includes at least local life scenario data, user behavior habit data, local language expression data, cultural taboo data, and platform content style data.

[0027] S3. Based on the target market localized data and product structured attribute data, construct a joint constraint knowledge graph of target market local culture and user behavior; the joint constraint knowledge graph includes product attribute nodes, life scenario nodes, user behavior nodes, language expression nodes, cultural taboo nodes, and platform constraint nodes; the nodes are connected through adaptation relationships, conflict relationships, replacement relationships, trigger relationships, and disabling relationships. Each node and relationship has a weight, confidence level, source identifier, and update time identifier.

[0028] In some embodiments of the present invention, a knowledge graph jointly constrained by the local culture and user behavior of the target market is constructed by acquiring local information of the target market. For several preset target markets (e.g., but not limited to: the United States, Canada, Japan, Germany), a structured knowledge graph G=(V,E) is constructed. By constructing a targeted knowledge graph, effective support is provided for subsequent retrieval and analysis. The node set V includes: product attribute nodes (size, material, function), life scenario nodes (apartment, suburban detached house, commuting subway), user behavior nodes, language expression nodes, cultural taboo nodes, and platform constraint nodes. Product attribute nodes include: product category, size, weight, material, function, installation method, applicable users, and usage restrictions; lifestyle scenario nodes include: apartment, detached house, living room, bedroom, kitchen, backyard, garage, commuting scenario, office scenario, etc.; user behavior nodes include: storage, cleaning, pet interaction, home organization, outdoor use, holiday gifting, etc.; language expression nodes include: commonly used expressions in the target market, platform short phrases, category keywords, purchase trigger words, and local habitual expressions; cultural taboo nodes include: taboo expressions in the target market, exaggerated advertising expressions, absolute expressions, and inappropriate metaphors; platform constraint nodes include: different content format requirements such as TikTok short videos, Amazon product detail pages, independent website ads, and live streaming scripts. Adaptation relationship: Product attributes are suitable for a certain target market scenario; Conflict relationship: Product attributes are not suitable for a certain target market scenario; Trigger relationship: Product functions can trigger a certain user behavior or purchase motivation; Substitution relationship: Common expressions can be replaced with local expressions; Prohibition relationship: Certain expressions should be prohibited in the target market or platform.

[0029] S4. Construct a cross-border e-commerce retrieval model based on the RAG architecture and the aforementioned joint constraint knowledge graph; In some embodiments of this invention, a knowledge graph of local culture and behavior in the target market is introduced based on the existing RAG architecture to obtain a new cross-border e-commerce retrieval model. This enables the large model to function as a "cultural consultant" rather than simply a "translation assistant." The aforementioned cross-border e-commerce retrieval model refers to an architecture that combines externally verifiable data from a joint constraint knowledge graph with a large-scale language model, resulting in a mathematical model that enables more accurate retrieval analysis.

[0030] S5. Generate a query tuple based on the target e-commerce demand information; the query tuple shall include at least the target market, product category, product attributes, content output type, target platform, and target audience.

[0031] S6. Based on the query tuple, perform a hybrid retrieval in the cross-border e-commerce retrieval model to obtain candidate knowledge graph nodes and candidate relationship paths; the hybrid retrieval includes keyword recall, vector similarity recall, and graph path recall.

[0032] S7. Calculate the comprehensive adaptation score based on the candidate knowledge graph nodes and candidate relationship paths, and determine the target calibration node and target calibration path based on the comprehensive adaptation score; the comprehensive adaptation score includes semantic similarity, graph path weight, scene adaptation, behavior matching, language adaptation and risk constraint value.

[0033] In some embodiments of the present invention, query tuples are first generated based on the target e-commerce needs; then keyword retrieval, vector retrieval, and graph path retrieval are performed; then, the nodes and paths are sorted according to the comprehensive fit score; finally, the top-ranked nodes and paths are taken as calibration criteria. The comprehensive fit score formula is: Score = a × semantic similarity + b × path weight + c × scenario fit + d × behavior matching + e × language fit - f × risk constraint value, where a, b, c, d, e, and f are preset weights that can be adjusted according to product category, target market, and platform type, and Score is the comprehensive fit score.

[0034] S8. Perform localized mapping calibration based on the target calibration node and the target calibration path to generate localized calibration information; the localized mapping calibration includes scene calibration, behavior calibration and language calibration; In some embodiments of the present invention, scene calibration includes: converting product size, weight, installation method, and usage radius into product spatial constraints; converting space type, usage location, and user habits in the target market's living scenarios into scene constraints; determining whether the product spatial constraints and scene constraints match; if they match, allowing the generation of the corresponding scene expression; if they do not match, prohibiting the generation of the scene expression and reselecting other scene nodes. For example, if the product is large, an expression like "suitable for a corner of a small apartment" cannot be generated; if the product is suitable for a backyard or living room, an expression that conforms to a scenario familiar to American users should be generated. Behavior calibration includes: matching product functions with target market user behavior nodes; if the product function is "cat scratching, claw sharpening, furniture protection," do not simply translate the function, but map it to behavior expressions familiar to American users, such as "protect furniture," "indoor cat enrichment," "reduce boredom," etc.; if the product function and target market behavior do not match, reduce the weight of the behavior node and do not proceed to generate prompts. Language calibration includes: establishing a localized vocabulary database for the target market, a database of high-frequency phrases for the platform, a database of prohibited words, and a database of culturally risky words; converting generic expressions into localized expressions; filtering out exaggerated, absolute, unnatural, and culturally incompatible expressions; and outputting different formats such as short sentences, titles, detail pages, and live-streaming scripts according to the target platform. For example, "saving space" should not be simply translated as "save space," but should be changed to "space-saving design," "small-apartment friendly," etc., depending on the platform and context; expressions such as "hot seller," "number one on the entire internet," and "miracle product" should be filtered for risk and avoided from being generated directly.

[0035] S9. Construct a constraint prompt sequence based on the original product data in the localized calibration information and target e-commerce demand information, and input the constraint prompt sequence into a pre-built content generation model to generate e-commerce marketing content adapted to the target market. The constraint prompt sequence includes the target market, original product data, scene calibration information, behavior calibration information, language calibration information, prohibited expressions, and output formats. The e-commerce marketing content includes product titles, detail page copy, advertising copy, short video scripts, live streaming scripts, and cover copy.

[0036] S10. Collect user behavior feedback data after the e-commerce marketing content is launched in the target market, and update the node weights and relationship weights in the joint constraint knowledge graph based on the user behavior feedback data.

[0037] In some embodiments of the present invention, data such as clicks, dwell time, interaction, comment keywords, favorites, add-to-cart, conversion, negative feedback, and violation warnings are collected after the campaign is launched. If a certain expression or scenario receives good feedback in the target market, the weight of the corresponding node and relationship is increased. If a certain expression causes negative feedback or violation warnings, the weight of the corresponding node is reduced or a disabling relationship is established. In subsequent retrieval, the recall and ranking are affected by the updated weights.

[0038] To further improve the real-time adaptability and effectiveness of marketing content, this invention collects user feedback after marketing content delivery and dynamically updates the knowledge graph for the corresponding target market based on user behavior feedback data. This knowledge graph is then effectively optimized to improve subsequent retrieval accuracy and output more precise marketing content. Through a feedback update mechanism, this invention adjusts the weights of graph nodes and relationships based on actual user behavior feedback after delivery, enabling continuous optimization of subsequent content generation as the target market changes.

[0039] First, this invention constructs a joint constraint knowledge graph encompassing product attributes, life scenarios, user behavior, language expression, cultural taboos, and platform constraints. This allows the generative model to acquire structured localized constraints before generation, reducing the risk of product attributes not matching the target market scenario. Second, through a hybrid retrieval and comprehensive adaptation scoring mechanism, this invention ensures that retrieval results not only rely on text similarity but also incorporate graph paths, scenario adaptation, behavior matching, language adaptation, and risk constraints, improving the consistency between recall nodes and target market needs. Third, through scenario calibration, behavior calibration, and language calibration, this invention transforms the original product selling points into marketing content that is understandable, acceptable, and conforms to platform expression habits in the target market, reducing literal translations, culturally incompatible expressions, and platform-incompatible expressions. Finally, through a feedback update mechanism, this invention adjusts the graph node weights and relationship weights based on user behavior feedback after actual deployment, enabling subsequent content generation to continuously optimize as the target market changes.

[0040] This invention uses a combined constrained knowledge graph and RAG retrieval mechanism to perform structured calibration of product attributes, target market scenarios, user behavior, and language expression before content generation, thereby reducing scenario mismatch, behavior mismatch, and language mismatch in cross-border e-commerce marketing content generation.

[0041] like Figure 2 As shown, in a second aspect, embodiments of the present invention provide a cross-border e-commerce marketing content generation system based on knowledge graphs and RAG, including a demand data acquisition module 100, a localized data acquisition module 200, a knowledge graph construction module 300, a retrieval model construction module 400, a query tuple generation module 500, a hybrid retrieval module 600, a rating and ranking module 700, a localized calibration module 800, a marketing content generation module 900, and a feedback and update module 1000, wherein: The demand data acquisition module 100 is used to acquire target e-commerce demand information, which includes at least the target market identifier, product category, product structured attribute data, product text data, content output type, and target platform identifier. The localization data acquisition module 200 is used to acquire target market localization data corresponding to the target market identifier; The graph construction module 300 is used to construct a joint constraint knowledge graph of local culture and user behavior in the target market based on the target market localization data and product structured attribute data. The retrieval model construction module 400 is used to construct a cross-border e-commerce retrieval model based on the RAG architecture and the joint constraint knowledge graph. The query tuple generation module 500 is used to generate query tuples based on the target e-commerce demand information. The hybrid retrieval module 600 is used to perform a hybrid retrieval in the cross-border e-commerce retrieval model based on the query tuple to obtain candidate knowledge graph nodes and candidate relationship paths. The scoring and ranking module 700 is used to calculate the comprehensive adaptation score based on the candidate knowledge graph nodes and candidate relationship paths, and to determine the target calibration node and target calibration path based on the comprehensive adaptation score. The localization calibration module 800 is used to perform localized mapping calibration based on the target calibration node and the target calibration path, and generate localized calibration information; the localized mapping calibration includes scene calibration, behavior calibration and language calibration; The marketing content generation module 900 is used to construct a constraint prompt sequence based on the original product data in the localized calibration information and the target e-commerce demand information, and input the constraint prompt sequence into the pre-built content generation model to generate e-commerce marketing content adapted to the target market. The feedback update module 1000 is used to collect user behavior feedback data after the e-commerce marketing content is launched in the target market, and update the node weights and relationship weights in the joint constraint knowledge graph based on the user behavior feedback data.

[0042] This system, through the collaboration of multiple modules including the demand data acquisition module 100, the localized data acquisition module 200, the graph construction module 300, the retrieval model construction module 400, the query tuple generation module 500, the hybrid retrieval module 600, the rating and ranking module 700, the localization calibration module 800, the marketing content generation module 900, and the feedback and update module 1000, uses a joint constraint knowledge graph and RAG retrieval mechanism to perform structured calibration of product attributes, target market scenarios, user behavior, and language expression before content generation. This reduces scenario mismatch, behavior mismatch, and language mismatch in cross-border e-commerce marketing content generation.

[0043] To better illustrate the technical solution of this invention, the generation of a TikTok short video script for pet supplies in the United States is used as an example: Input information: Target market: USA; Target platform: TikTok; Product: L-shaped cat scratching post or automatic litter box; Product attributes: size, material, function, suitable pets, installation / placement method.

[0044] Output content: short video script, title, and cover text.

[0045] Example of a graph node: Product attribute node: cat scratch, furniture protection, indoor cat, corner placement; Life scenario nodes: apartment living room, small-space pet corner, rental home; User behavior nodes: protect furniture, indoor cat enrichment, reduce boredom; Language expression nodes: small-space friendly, furniture saver, cat enrichment; Cultural taboo nodes: overly exaggerated expression, absolute expression, unrealistic efficacy expression; Platform constraint nodes: TikTok first three seconds hook, short sentences, strong scenario expression.

[0046] Calibration process: Scene calibration: Determine if the product is suitable for scenarios such as American apartments, living rooms, and corners; Behavior calibration: Map the "cat scratching" function to "protect furniture," "provide activity for indoor cats," and "reduce boredom"; Language calibration: Change the Chinese selling points to local expressions and filter out exaggerated and literal translations.

[0047] Constraint prompt generation: Combine target market, platform, product hard attributes, scenario expression, behavior expression, language expression, and prohibited expression into a prompt sequence.

[0048] Content generation: Output TikTok short video scripts, titles, and cover text.

[0049] Feedback Update: Update the corresponding node and relationship weights based on video clicks, dwell time, comment keywords, negative feedback, etc.

[0050] In the embodiments provided in this application, it should be understood that the disclosed methods and systems can also be implemented in other ways. The method and system embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0051] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0052] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0053] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of this application is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within this application. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for generating cross-border e-commerce marketing content based on knowledge graphs and RAG, characterized in that, Includes the following steps: S1. Obtain target e-commerce demand information, which includes at least target market identifier, product category, product structured attribute data, product text data, content output type, and target platform identifier; S2. Obtain the target market localization data corresponding to the target market identifier; S3. Based on the target market localization data and product structured attribute data, construct a joint constraint knowledge graph of target market local culture and user behavior; S4. Construct a cross-border e-commerce retrieval model based on the RAG architecture and the aforementioned joint constraint knowledge graph; S5. Generate a query tuple based on the target e-commerce demand information; S6. Based on the query tuple, perform a hybrid search in the cross-border e-commerce retrieval model to obtain candidate knowledge graph nodes and candidate relationship paths; S7. Calculate the comprehensive fit score based on the candidate knowledge graph nodes and candidate relationship paths, and determine the target calibration node and target calibration path based on the comprehensive fit score; S8. Perform localized mapping calibration based on the target calibration node and the target calibration path to generate localized calibration information; the localized mapping calibration includes scene calibration, behavior calibration and language calibration; S9. Construct a constraint prompt sequence based on the original product data in the localized calibration information and the target e-commerce demand information, and input the constraint prompt sequence into a pre-built content generation model to generate e-commerce marketing content adapted to the target market; S10. Collect user behavior feedback data after the e-commerce marketing content is launched in the target market, and update the node weights and relationship weights in the joint constraint knowledge graph based on the user behavior feedback data.

2. The method for generating cross-border e-commerce marketing content based on knowledge graphs and RAG as described in claim 1, characterized in that, The target market localization data includes at least local life scenario data, user behavior data, local language expression data, cultural taboo data, and platform content style data.

3. The method for generating cross-border e-commerce marketing content based on knowledge graphs and RAG as described in claim 1, characterized in that, The joint constraint knowledge graph includes product attribute nodes, life scenario nodes, user behavior nodes, language expression nodes, cultural taboo nodes, and platform constraint nodes; the nodes are connected through adaptation relationships, conflict relationships, replacement relationships, triggering relationships, and disabling relationships.

4. The method for generating cross-border e-commerce marketing content based on knowledge graphs and RAG as described in claim 3, characterized in that, Each node and each type of relationship has a weight, confidence level, source identifier, and update time identifier.

5. The method for generating cross-border e-commerce marketing content based on knowledge graphs and RAG as described in claim 1, characterized in that, The hybrid retrieval includes keyword recall, vector similarity recall, and graph path recall.

6. The method for generating cross-border e-commerce marketing content based on knowledge graphs and RAG as described in claim 1, characterized in that, The comprehensive adaptation score includes semantic similarity, graph path weight, scene adaptation, behavior matching, language adaptation, and risk constraint value.

7. The method for generating cross-border e-commerce marketing content based on knowledge graphs and RAG as described in claim 1, characterized in that, The constraint prompt sequence includes target market, original product data, scene calibration information, behavior calibration information, language calibration information, prohibited expressions, and output formats.

8. The method for generating cross-border e-commerce marketing content based on knowledge graphs and RAG according to claim 1, characterized in that, The e-commerce marketing content includes product titles, product detail page copy, advertising copy, short video scripts, live streaming scripts, and cover copy.

9. A cross-border e-commerce marketing content generation system based on knowledge graphs and RAG, characterized in that, It includes modules for demand data acquisition, localized data acquisition, graph construction, retrieval model construction, query tuple generation, hybrid retrieval, scoring and ranking, localized calibration, marketing content generation, and feedback update. The demand data acquisition module is used to acquire target e-commerce demand information, which includes at least the target market identifier, product category, product structured attribute data, product text data, content output type, and target platform identifier. The localization data acquisition module is used to acquire target market localization data corresponding to the target market identifier; The graph construction module is used to construct a knowledge graph jointly constrained by the local culture and user behavior of the target market based on the target market localized data and product structured attribute data. The retrieval model construction module is used to construct a cross-border e-commerce retrieval model based on the RAG architecture and the joint constraint knowledge graph. The query tuple generation module is used to generate query tuples based on the target e-commerce demand information; The hybrid retrieval module is used to perform a hybrid retrieval in the cross-border e-commerce retrieval model based on the query tuple to obtain candidate knowledge graph nodes and candidate relationship paths. The scoring and ranking module is used to calculate the comprehensive adaptation score based on the candidate knowledge graph nodes and candidate relationship paths, and to determine the target calibration node and target calibration path based on the comprehensive adaptation score. The localization calibration module is used to perform localized mapping calibration based on the target calibration node and the target calibration path, and generate localized calibration information; the localized mapping calibration includes scene calibration, behavior calibration and language calibration; The marketing content generation module is used to construct a constraint prompt sequence based on the original product data in the localized calibration information and the target e-commerce demand information, and input the constraint prompt sequence into the pre-built content generation model to generate e-commerce marketing content adapted to the target market. The feedback update module is used to collect user behavior feedback data after the e-commerce marketing content is launched in the target market, and update the node weights and relationship weights in the joint constraint knowledge graph based on the user behavior feedback data.

Images