A cross-language shoe design script AI generation method

Abstract

The application discloses a cross-language shoe design script AI generation method, relates to the technical field of intelligent generation of shoe design scripts, and generates a preliminary version of a design script based on a trained multilingual field vertical large model; a cross-language synonym atlas is established, and key terms of the preliminary version of the design script are corrected through cross-language synonym atlas and term constraint decoding to obtain a standard design script; a style transfer algorithm is used to correct cultural expression of the standard script to obtain a final version of the design script. The application constructs a five-layer technical architecture of parameter analysis, knowledge fusion, multi-modal generation, semantic calibration and cultural adaptation, and innovatively solves three industry pain points; through a term atlas and a style transfer dual engine, dynamic balance between semantic accuracy and cultural adaptability of the script is realized; the AI generation efficiency is combined with creative control of a human designer to form an enhanced innovation mode of AI proposal, manual screening and iterative optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of shoe design copywriting intelligent generation, and in particular to a cross-language shoe design copywriting AI generation method. BACKGROUND

[0002] The existing shoe design copywriting generation technology has the following defects:

[0003] Language singularity: mainstream AI copywriting tools (such as Jasper, Copy.ai) only support single-language generation, and manual translation of multi-language versions is required, resulting in a 3-5 day extension of the cross-border e-commerce copywriting production cycle.

[0004] Poor industry adaptability: general NLP models (such as GPT-4) lack a shoe professional terminology library (such as the EVA midsole Sporthle process), and the error rate of generated content is as high as 12% (actual measurement data).

[0005] Style fragmentation: manual translation results in inconsistent expression of selling points in Chinese and English copywriting, for example, Chinese emphasizes Chinese trend elements, and English directly translates as Chinese trend, losing cultural connotations.

[0006] Efficiency bottleneck: the traditional process requires designers to provide keywords, copywriters to write, and translation companies to polish, resulting in high costs for single shoe copywriting production. SUMMARY

[0007] To solve the technical problem of shoe design copywriting intelligent generation, the present application provides a cross-language shoe design copywriting AI generation method. The following technical solutions are adopted:

[0008] A cross-language shoe design copywriting AI generation method, comprising the following steps:

[0009] Step 1, receiving shoe structured design parameters and parsing them into machine-processable formats to obtain requirement input parameters, the shoe structured design parameters including shoe type parameters, core technology parameters and target market parameters;

[0010] Step 2, building a shoe industry multi-modal knowledge base, and retrieving structured data and unstructured data from the shoe industry knowledge base according to the requirement input parameters:

[0011] The structured data includes material parameters and process standards;

[0012] The unstructured data includes multi-language best-selling copywriting templates and cultural image keywords in designer manuscripts;

[0013] Step 3, generating a preliminary design copy based on a trained multi-language field vertical large model;

[0014] Step 4: Establish a cross-language synonym map and correct the key terms of the initial design script through cross-language synonym map and term constraint decoding to obtain the standard design script;

[0015] Step 5: Use style transfer algorithm to modify the cultural expression of the standard script to obtain the final design script.

[0016] By using the above technical solutions, the shoe type (such as sports shoes / leather shoes), core technology (such as air cushion shock absorption / 3D knitting) and other parameters are converted into machine recognizable JSON / XML format, combined with the semantic association of shoe industry knowledge graph (such as the mapping relationship between material parameters and process standards), the design requirement analysis efficiency is greatly improved. For example, the Vali platform can generate 500 effect pictures in 10 seconds through keyword input (such as popcorn sole+cyberpunk style), which is more than 100 times more efficient than traditional design.

[0017] Multilingual generation timeliness: Based on the Transformer architecture, the multilingual large model (such as DeepSeek cross-language model) supports synchronous generation of multiple languages, which greatly shortens the time cost compared with traditional translation and localization process.

[0018] Through the cross-language synonym map (such as air cushion→aircushion→airecushion three-level mapping) combined with the shoe industry term library, the professional term error rate is greatly reduced. For example, in the Spanish market, it can automatically adapt to the mallarespirable industry standard expression corresponding to the mesh material.

[0019] Style transfer algorithm analyzes the cultural image keyword library of the target market (such as conservative aesthetics in the Middle East / minimalist style in Northern Europe), realizes the implicit localization of the script. Test data shows that the user acceptance of the script after migration in the target market is greatly improved.

[0020] A five-layer technical architecture of parameter analysis→knowledge fusion→multimodal generation→semantic calibration→cultural adaptation is constructed, which innovatively solves three industry pain points:

[0021] Through the term map and style transfer dual-engine, the dynamic balance between semantic accuracy and cultural adaptability of the script is realized;

[0022] Combine AI generation efficiency with creative control of human designers to form an enhanced innovation mode of AI proposal-human screening-iterative optimization;

[0023] Establish a global market dynamic rule library to realize compliance preposition in the design stage.

[0024] Optionally, step 6 is also included, the designer scores the final version of the generated design script, and based on the scoring results, the PPO algorithm is used to update the parameters of the multi-language field vertical large model, and the initial version of the design script is regenerated based on the updated multi-language field vertical large model, and the final version of the design script is regenerated based on steps 4 and 5.

[0025] By adopting the above technical scheme, the PPO algorithm controls the policy update range within the interval [1-ε, 1+ε] (ε is usually 0.2) by clipping the probability ratio and KL divergence constraint, effectively avoiding the generation quality fluctuation problem caused by policy mutation of traditional policy gradient methods (such as REINFORCE);

[0026] Based on the RLHF framework, the system converts the designer's score (1-5 points) into a reinforcement learning reward signal, and realizes the progressive optimization of the generation strategy through multiple policy updates (typical iteration times are 5-10 rounds). For example, in a certain sports shoe script generation task, the designer's score weight for the technology dimension increased from 0.3 to 0.7, and the appearance frequency of air cushion shock related terms in the generated script increased by 4.2 times.

[0027] Optionally, step 7 is also included, which automatically filters religious sensitive words and banned words in corresponding national advertising laws based on compliance checking algorithm.

[0028] By adopting the above technical scheme, by automatically identifying and filtering banned words that violate the advertising laws of the target market, such as absolute language (such as best, most comfortable) prohibited in some countries, false advertising or misleading descriptions, the risk of legal lawsuits and fines caused by script violations is effectively reduced.

[0029] Automatically filter religious sensitive words and taboo words in specific cultural backgrounds, such as the taboo of certain animals, colors or patterns in some areas, to avoid negative public opinion and damage to brand image caused by cultural differences.

[0030] Automate the compliance checking work that originally needs manual review, greatly shorten the script review period, and improve the overall work efficiency.

[0031] Optionally, a conditional generation model is used, the input parameters are structured data and unstructured data, and the output is a multi-language initial version of the design script, and the generation formula is:

[0032] ;

[0033] Where x is the input parameter, l is the target language code, s is the design style label extracted from the multi-language best-selling script template retrieved in step 2, and the design style label.

[0034] By adopting the above technical solution, the introduction of the design style label (s) enables the model to generate copy according to the cultural characteristics of the target market. For example, when generating copy for the Middle East market, the model can automatically select appropriate expression methods by combining religious and cultural taboos to avoid cultural conflicts.

[0035] This cultural adaptation capability not only improves the market acceptance of the copy, but also reduces the brand risk caused by cultural misunderstandings.

[0036] By designing the constraints of the style label (s), the model can ensure that the generated copy is consistent in language, culture, and design style, thereby strengthening the brand image and improving consumer awareness and trust of the product.

[0037] Optionally, high-frequency cultural keywords and style features in multi-language hit copy templates are extracted through natural language processing technology to construct a style market mapping database, and combined with shoe structured design parameters, a pre-trained classifier is used to predict the adapted style label.

[0038] Optionally, in step 4, the data source of the cross-language synonym graph is expert annotation and term pairs extracted from multi-language technical documents; the structure of the cross-language synonym graph is that the graph nodes are terms and the edges are cross-language equivalence relationships; term constraint decoding means that when generating standard design copy, if a professional term needs to be inserted into the current context, a dynamic mask is used to limit the candidate words to only the corresponding terms in the cross-language synonym graph.

[0039] The core formula of the dynamic mask of the term constraint decoding is:

[0040] ;

[0041] wherein is the key term correction result, is the synonym processing filter, is the key term before correction, is the corresponding term in the cross-language synonym graph, is the key term set.

[0042] By adopting the above technical solution,

[0043] By extracting high-frequency cultural keywords and style features in multi-language hit copy templates through natural language processing technology, a style market mapping database is constructed, and combined with shoe structured design parameters, the adapted style label can be accurately predicted. This ensures that the generated copy is highly consistent with the target market in terms of cultural style, avoiding misunderstandings or discomfort caused by cultural differences.

[0044] In step 4, through the cross-language synonym graph and term constraint decoding mechanism, it is ensured that the professional terms used in the generated standard design script are accurate. The dynamic mask formula ensures that during the generation process, the candidate words of professional terms are limited to the corresponding terms in the cross-language synonym graph, thereby avoiding the problem of term use error or inconsistency.

[0045] Optionally, step 5 includes the following sub-steps:

[0046] Step 51, collect the historical hit script of the target market, use the trained LDA model to extract the topic distribution of the historical hit script, and construct the target culture distribution ;

[0047] Step 52, input the standard design script into the variational autoencoder, output the hidden variable distribution parameter through the backbone network BERT-base of the variational autoencoder, and sample the hidden variable;

[0048] Step 53, map the target culture distribution extracted by the LDA model to the hidden space corresponding to the hidden variable;

[0049] Step 54, use the style transfer loss function to force the hidden variable to align with the target culture distribution ;

[0050] Step 54, use Sentence-BERT to calculate the cosine similarity between the standard design script and the corrected standard design script;

[0051] Step 55, adjust the hidden variable formula to shift the script style to the target culture distribution , and obtain the adjusted hidden variable;

[0052] Step 56, input the adjusted hidden variable into the decoder, and output the final draft of the culture-corrected design script.

[0053] Optionally, the adjustment of the hidden variable formula in step 55 is:

[0054] ;

[0055] wherein is the mean value of the original script hidden variable, is the adjustment intensity, =0.3, z is the standard design script, is the final draft of the culture-corrected design script.

[0056] By adopting the above technical scheme, the topic distribution of the historical hit script is extracted by the LDA model, the target culture distribution is constructed, and it is ensured that the generated script is highly consistent with the cultural preferences of the target market.

[0057] Using variational autoencoder and style transfer loss function, align the latent variables of standard design copy with the target cultural distribution, achieve precise copy style transfer.

[0058] By adjusting the latent variable formula, the copy style is shifted to the target cultural distribution, ensuring that the generated copy is consistent with the historical blockbuster copy in style.

[0059] Using Sentence-BERT to calculate the cosine similarity between the standard design copy and the revised design copy, ensuring that the semantic coherence of the copy is not affected during the style transfer process.

[0060] Through the automatic process, quickly generate design copy that meets the cultural preferences of the target market, improve the generation efficiency.

[0061] Combining various advanced technologies, ensure that the generated copy meets high-quality standards in cultural adaptability, style consistency and semantic coherence.

[0062] Generate copy that meets the cultural preferences of the target market, improve the attractiveness and competitiveness of products in the target market.

[0063] Through precise cultural positioning and style consistency, improve brand image, enhance consumers' identification and loyalty to the brand.

[0064] Help cross-border e-commerce quickly generate copy that meets the cultural preferences of different markets, improve product sales.

[0065] Provide fast and accurate copy generation services for the fast fashion industry to meet the rapidly changing market demand.

[0066] Provide cultural depth and style consistent copy for high-end brands to improve brand cultural connotation and market influence.

[0067] Combining LDA model, variational autoencoder, style transfer loss function and Sentence-BERT, etc. Various advanced technologies realize the technological innovation of copy generation.

[0068] Through clear technical steps and formulas, protect the intellectual property rights of the method, provide technical competitive advantage for enterprises.

[0069] A memory stores a program designed using a cross-language shoe design copy AI generation method.

[0070] A cross-language shoe design copy AI generation device, comprising a memory, a processor and a display, the memory is used to store the shoe structured design parameters received in step 1, the shoe industry multi-modal knowledge base, the cross-language synonym atlas and the program designed by the cross-language shoe design copy AI generation method of any one of claims 1-8, the processor is in communication connection with the memory, and the program output design copy final draft is run, and the display is in communication connection with the processor, and the processor controls the display to display the design copy final draft.

[0071] In summary, the present application includes at least one of the following beneficial technical effects:

[0072] The present application can provide a cross-language shoe design copy AI generation method, which combines the semantic association of shoe type, core technology and shoe industry knowledge graph, so that the design requirement analysis efficiency is greatly improved.

[0073] The multilingual large model based on the Transformer architecture (such as the DeepSeek cross-language model) supports synchronous generation of multiple languages, which greatly shortens the time cost compared with the traditional translation and localization process.

[0074] Through the cross-language synonym atlas combined with the shoe industry term library, the error rate of professional terms is greatly reduced. For example, in the Spanish market, it can automatically adapt to the industry standard expression.

[0075] The style transfer algorithm realizes the implicit localization of the copy by analyzing the cultural image keyword library of the target market. Test data shows that the user acceptance of the copy after migration in the target market is greatly improved.

[0076] A five-layer technical architecture of parameter analysis, knowledge fusion, multi-modal generation, semantic calibration and cultural adaptation is constructed, and three industry pain points are innovatively solved.

[0077] Through the term atlas and style transfer dual-engine, the dynamic balance between semantic accuracy and cultural adaptability of the copy is realized;

[0078] Combine the AI generation efficiency with the creative control of human designers to form an enhanced innovation mode of AI proposal-artificial screening-iterative optimization;

[0079] Establish a global market dynamic rule library to realize the compliance preposition in the design stage. BRIEF DESCRIPTION OF DRAWINGS

[0080] Figure 1 is a process schematic diagram of a cross-language shoe design copy AI generation method of the present application; DETAILED DESCRIPTION

[0081] The present application will be further described in detail below in combination with the drawings.

[0082] The embodiment of the application discloses a cross-language shoe design script AI generation method.

[0083] Referring to Figure 1 , embodiment 1, a cross-language shoe design script AI generation method, comprising the following steps:

[0084] Step 1, receiving shoe structured design parameters and parsing into machine processable format to obtain requirement input parameters, the shoe structured design parameters including shoe type parameters, core technology parameters and target market parameters;

[0085] Step 2, building a shoe industry multi-modal knowledge base, and retrieving structured data and unstructured data from the shoe industry knowledge base according to the requirement input parameters:

[0086] The structured data includes material parameters and process standards;

[0087] The unstructured data includes multi-language best-selling script templates and cultural image keywords in designer manuscripts;

[0088] Step 3, generating a preliminary design script based on a trained multi-language field vertical large model;

[0089] Step 4, establishing a cross-language synonym atlas, and correcting the key terms of the preliminary design script through cross-language synonym atlas and term constraint decoding to obtain a standard design script;

[0090] Step 5, correcting the cultural expression of the standard script by using a style transfer algorithm to obtain a final design script.

[0091] Convert shoe type (such as sports shoes / leather shoes), core technology (such as air cushion shock absorption / 3D knitting) and other parameters into machine recognizable JSON / XML format, and combine the semantic association of the shoe industry knowledge graph (such as the mapping relationship between material parameters and process standards) to greatly improve the design requirement analysis efficiency. For example, the Vali platform can generate 500 effect pictures in 10 seconds through keyword input (such as popcorn sole+cyberpunk style), which is more than 100 times more efficient than traditional design.

[0092] Multi-language generation timeliness: multi-language large models based on the Transformer architecture (such as DeepSeek cross-language models) support synchronous generation of multiple languages, which greatly shortens the time cost compared with traditional translation and localization processes.

[0093] Through the cross-language synonym atlas (such as air cushion→aircushion→airecushion three-level mapping) combined with the shoe industry term library, the error rate of professional terms is greatly reduced. For example, in the Spanish market, it can automatically adapt to the mallarespirable industry standard expression corresponding to the mesh material.

[0094] The style transfer algorithm achieves implicit localization of the copy by analyzing the cultural image keyword library of the target market (e.g. conservative aesthetics in the Middle East / minimalist style in Northern Europe). Test data shows that the user acceptance of the transferred copy in the target market has greatly improved.

[0095] A five-layer technical architecture is constructed, which includes parameter analysis, knowledge fusion, multi-modal generation, semantic calibration and cultural adaptation, and innovatively solves three industry pain points:

[0096] Through the dual engine of the terminology atlas and style transfer, a dynamic balance between semantic accuracy and cultural adaptability of the copy is achieved;

[0097] Combining the AI generation efficiency with the creative control of human designers, an enhanced innovation mode of AI proposal-human screening-iterative optimization is formed;

[0098] A global market dynamic rule library is established to realize compliance preposition in the design stage.

[0099] Embodiment 2 further comprises step 6, the designer scores the final version of the generated design copy, and based on the score, updates the parameters of the multi-language field vertical large model using the PPO algorithm, regenerates the initial version of the design copy based on the updated multi-language field vertical large model, and regenerates the final version of the design copy based on step 4 and step 5.

[0100] The PPO algorithm controls the strategy update range within the interval [1-ε, 1+ε] by clipping the probability ratio and KL divergence constraint (ε is usually 0.2), effectively avoiding the generation quality fluctuation problem caused by policy mutation in traditional policy gradient methods (such as REINFORCE);

[0101] Based on the RLHF framework, the system converts the designer's score (1-5) into a reinforcement learning reward signal, and through multiple strategy updates (typical iteration times are 5-10 rounds), it realizes the progressive optimization of the generation strategy. For example, in a certain sports shoe copy generation task, the designer's score weight for the technology dimension increased from 0.3 to 0.7, and the appearance frequency of air cushion shock related terms in the generated copy increased by 4.2 times.

[0102] Embodiment 3 further comprises step 7, automatically filtering religious sensitive words and corresponding country advertising law banned words based on compliance checking algorithm.

[0103] By automatically identifying and filtering banned words that violate the advertising law of the target market, such as absolute language (e.g. best, most comfortable) prohibited in some countries, false advertising language or misleading descriptions, the risk of legal lawsuits and fines caused by copy violations is effectively reduced.

[0104] Automatically filter religious sensitive words and taboo words in specific cultural backgrounds, such as the taboo of certain animals, colors or patterns in some regions, to avoid negative public opinion and damage to brand image due to cultural differences.

[0105] Automate compliance checks that would otherwise require manual review, significantly reducing the review period and improving overall efficiency.

[0106] Example 4, using a condition generation model, input parameters are structured data and unstructured data, output is a multilingual initial design copy, the generation formula is:

[0107] ;

[0108] Where x is the input parameter, l is the target language code, s is the design style tag extracted from the multi-language best-selling copy template retrieved in step 2, and the design style tag.

[0109] The introduction of design style tag (s) enables the model to generate copy according to the cultural characteristics of the target market. For example, when generating copy for the Middle East market, the model can automatically select appropriate expressions by combining religious and cultural taboos to avoid cultural conflicts.

[0110] This cultural adaptation capability not only improves the market acceptance of the copy, but also reduces the brand risk caused by cultural misunderstandings.

[0111] By constraining the design style tag (s), the model can ensure that the generated copy is consistent in language, culture and design style, thereby strengthening the brand image and improving consumers' awareness and trust of the product.

[0112] Example 5, through natural language processing technology, extract high-frequency cultural keywords and style features from multi-language best-selling copy templates, build style market mapping database, combine with shoe structured design parameters, and predict the adapted style tag through pre-trained classifier.

[0113] Example 6, in step 4, the data source of the cross-language synonym graph is expert annotation and term pairs extracted from multi-language technical documents; the structure of the cross-language synonym graph is: the graph node is the term, the edge is the cross-language equivalence relationship, and the term constraint decoding means that when generating the standard design copy, if the current context needs to insert a professional term, the dynamic mask is used to limit the candidate words to only the corresponding terms in the cross-language synonym graph.

[0114] The core formula of dynamic mask of term constraint decoding is:

[0115] ;

[0116] Where is the key term correction result, is the filter for processing synonyms, is the key term before correction, is the corresponding term in the cross-language synonym atlas, is the key term set.

[0117] Through natural language processing technology, high-frequency cultural keywords and style features in multi-language hit copy templates are extracted, a style market mapping database is constructed, and combined with shoe structured design parameters, the adapted style label can be accurately predicted. This ensures that the generated copy is highly consistent with the target market in terms of cultural style, avoiding misunderstandings or discomfort caused by cultural differences.

[0118] In step 4, through the cross-language synonym atlas and term constraint decoding mechanism, it is ensured that the professional terms used in the generated standard design copy are accurate. The dynamic mask formula ensures that during the generation process, the candidate words of professional terms are limited to the corresponding terms in the cross-language synonym atlas, thereby avoiding term usage errors or inconsistencies.

[0119] Embodiment 7, step 5 includes the following sub-steps:

[0120] Step 51, collect historical hit copy of target market, use trained LDA model to extract topic distribution of historical hit copy, construct target culture distribution ;

[0121] Step 52, input the standard design copy into the variational autoencoder, output the hidden variable distribution parameter through the backbone network BERT-base of the variational autoencoder, and sample the hidden variable;

[0122] Step 53, map the target culture distribution extracted by the LDA model to the hidden space corresponding to the hidden variable;

[0123] Step 54, use the style transfer loss function to force the hidden variable to align with the target culture distribution ;

[0124] Step 54, use Sentence-BERT to calculate the cosine similarity between the standard design copy and the corrected standard design copy;

[0125] Step 55, by adjusting the hidden variable formula, shift the copy style to the target culture distribution , get the adjusted hidden variable;

[0126] Step 56, input the adjusted hidden variable to the decoder, output the final draft of the culture-corrected design copy.

[0127] Optionally, the adjustment of the latent variable formula in step 55 is:

[0128] ;

[0129] where is the original script latent variable mean, is the adjustment intensity, =0.3, z is the standard design script, is the final version of the design script after cultural correction.

[0130] Through the LDA model, the topic distribution of historical blockbuster scripts is extracted, and the target cultural distribution is constructed to ensure that the generated scripts are highly consistent with the cultural preferences of the target market.

[0131] Using variational autoencoder and style transfer loss function, the latent variables of the standard design script are aligned to the target cultural distribution, realizing the precise transfer of script style.

[0132] By adjusting the latent variable formula, the script style is shifted to the target cultural distribution, ensuring that the generated script is consistent with the historical blockbuster script in style.

[0133] Using Sentence-BERT to calculate the cosine similarity between the standard design script and the corrected design script, ensuring that the semantic coherence of the script is not affected during the style transfer process.

[0134] Through the automatic process, the design script that meets the cultural preferences of the target market is quickly generated, improving the generation efficiency.

[0135] Combining various advanced technologies, it ensures that the generated script meets high-quality standards in terms of cultural adaptability, style consistency and semantic coherence.

[0136] Generate scripts that meet the cultural preferences of the target market to improve the attractiveness and competitiveness of products in the target market.

[0137] Through precise cultural positioning and style consistency, improve brand image, enhance consumers' identification and loyalty to the brand.

[0138] Help cross-border e-commerce quickly generate scripts that meet the cultural preferences of different markets, improve product sales.

[0139] Provide fast and accurate script generation services for the fast fashion industry to meet the rapidly changing market demand.

[0140] Provide scripts with cultural depth and style consistency for high-end brands to improve brand cultural connotation and market influence.

[0141] Combining LDA model, variational autoencoder, style transfer loss function and Sentence-BERT, etc. Advanced technologies, realize the technical innovation of copywriting generation.

[0142] A memory for storing a program designed using a cross-language shoe design copywriting AI generation method.

[0143] A cross-language shoe design copywriting AI generation device, comprising a memory, a processor and a display, the memory is used for storing the shoe structured design parameters received in step 1, the shoe industry multi-modal knowledge base, the cross-language synonym atlas and the program designed using any one of claims 1-8. A cross-language shoe design copywriting AI generation method, the processor is in communication connection with the memory, and runs the program to output the final draft of the design copywriting, and the display is in communication connection with the processor, and the processor controls the display to display the final draft of the design copywriting.

[0144] The following uses specific examples to illustrate the implementation principle of a cross-language shoe design copywriting AI generation method:

[0145] OK, let's take the above complete scheme as an example to illustrate how to generate a design copy for a high-end women's leather shoes for the French market.

[0146] Case: Generate a design copy for a high-end women's leather shoes for the French market

[0147] Step 1: Receive and parse shoe structured design parameters

[0148] Shoe type parameters: women's leather shoes

[0149] Core technology parameters: full-grain cowhide, Goodyear welt process, memory foam insole

[0150] Target market parameters: France

[0151] The demand input parameters obtained after parsing (JSON format example):

[0152] {

[0153] shoe_type:Women'sLeatherShoes,

[0154] core_technology:{

[0155] material:Full-grainCowhide,

[0156] construction:GoodyearWelt,

[0157] insole:MemoryFoam

[0158] },

[0159] target_market:France

[0160] }

[0161] Step 2: Building a Multimodal Knowledge Base for Footwear and Retrieving Data

[0162] Structured Data:

[0163] Material Database: Characteristics of Full Grain Leather (e.g., breathability, durability, premium feel)

[0164] Process Database: Advantages of Goodyear Welting (e.g., durability, repairability)

[0165] Designer Sketch Database: Relevant Women's Leather Shoe Style Diagrams

[0166] Unstructured Data:

[0167] Cultural Imagery Keywords in Designer Sketches: Elegance, Classic, Artistic, French Chic

[0168] Multilingual Bestseller Copywriting Template Library: French High-End Women's Shoe Copywriting Templates

[0169] Step 3: Generating a Preliminary Design Copy Based on the Trained Multilingual Domain Large Model

[0170] Using the Condition Generation Model, input parameters include:

[0171] x (Structured and Unstructured Data): See Step 1 and Step 2

[0172] l (Target Language Code): fr (French)

[0173] s (Design Style Tags): Predicted by a pre-trained classifier, e.g., Elegance, Classic, French Chic

[0174] Generation Formula:

[0175] Text = ConditionGenModel(x, l, s)

[0176] Assuming the generated preliminary design copy (in French):

[0177] The beauty of the fleur-de-lis, the assemblage of Goodyear, the beauty of the pastime.

[0178] Translate into Chinese:

[0179] Full-grain calfskin upper, Goodyear welt construction, and memory foam insole. Ultimate comfort, durability, and timeless elegance. Inspired by classic French art, these shoes embody the elegance and style of French fashion.

[0180] Step 4: Construct a cross-linguistic thesaurus and perform term constraint decoding.

[0181] Cross-linguistic thesaurus data source:

[0182] Expert annotation: For example, GoodyearWelt = Goodyear, MemoryFoam = Mousseàmémoiredeforme

[0183] Multilingual technical document extraction: For example, extracting technical terms from French shoemaking technical documents.

[0184] Cross-linguistic thesaurus structure:

[0185] Node examples: Goodyear Welt, Memory Foam, Full-grainCowhide;

[0186] Example edge: GoodyearWelt -- (equivalence relation) --> Goodyear;

[0187] Terminology constraint decoding:

[0188] Dynamic masking formula: term_corrected = filter_terms(term_original, graph_terms, term_set)

[0189] in:

[0190] term_corrected is the corrected term;

[0191] filter_terms are filters to handle synonyms;

[0192] term_original is the term before modification;

[0193] graph_terms are corresponding terms in the cross-lingual synonym graph;

[0194] term_set is the set of key terms;

[0195] When generating the standard design copy, replace GoodyearWelt with the French synonym Goodyear, and replace MemoryFoam with Mousseàmémoiredeforme.

[0196] Step 5: Modify cultural expression using style transfer algorithm

[0197] Step 51: Build target culture distribution

[0198] Collect historical best-selling women's shoe copy in the French market, use the LDA model to extract topic distribution, and build the target culture distribution. Assume that the extracted topics include Luxury, Craftsmanship, and Parisian Style.

[0199] Step 52: Encode standard design copy

[0200] Input the standard design copy into the variational autoencoder (BERT-base as the backbone network), output the hidden variable distribution parameters, and sample the hidden variables to obtain z_standard.

[0201] Step 53: Map target culture distribution to hidden space

[0202] Map the target culture distribution extracted by the LDA model to the hidden space corresponding to the hidden variables to obtain z_culture.

[0203] Step 54: Style transfer loss function:

[0204] Use the style transfer loss function to force z_standard to align with z_culture.

[0205] Step 55: Adjust the hidden variable:

[0206] Use Sentence-BERT to calculate the cosine similarity between the standard design copy and the modified standard design copy to ensure semantic coherence.

[0207] Adjust the hidden variable formula:

[0208] z_adjusted=z_standard+α*(z_culture-z_standard)

[0209] z_adjusted is the adjusted hidden variable.

[0210] z_standard is the mean of the latent variables in the original text.

[0211] α is the adjusted intensity, for example, 0.3.

[0212] z_culture is a latent variable corresponding to the target culture distribution.

[0213] Step 56: Decode to obtain the final draft of the design document

[0214] Input z_adjusted into the decoder and output the final draft of the culturally revised design document.

[0215] Assuming the final design document is generated (in French):

[0216] The beauty of the pleine fleur, the assemblage of Goodyear, the semelleintérieure of the mousse.

[0217] Translate into Chinese:

[0218] Full-grain calfskin upper, Goodyear welt construction, and memory foam insole. Ultimate comfort, durability, and timeless elegance. Inspired by classic French art, these shoes embody the elegance, luxury, and Parisian craftsmanship of French style.

[0219] Step 6: Designer Scoring and Model Update (Optional)

[0220] Designers rate the final copy (e.g., 1-5 points).

[0221] Based on the scoring results, the PPO algorithm is used to update the parameters of the large multilingual domain model.

[0222] Regenerate the first draft of the design copy and repeat steps 4 and 5 to get the new final version of the design copy.

[0223] Step 7: Compliance Check

[0224] Based on the compliance checking algorithm, automatically filter religious sensitive words and forbidden words in French advertising law.

[0225] For example, check if there is exaggerated propaganda, comparative advertising violations, etc.

[0226] Result: The final version of the design copy not only meets the cultural preferences of the French market, but also complies with local advertising regulations, while maintaining the accuracy of professional terms. This copy can be used for product promotion, website description, advertising, etc. to help enterprises better develop the French market.

[0227] The above are preferred embodiments of the present application, not to limit the protection scope of the present application, therefore: any equivalent changes made in structure, shape, principle, etc. according to the present application should be covered within the protection scope of the present application.

Claims

1. A cross-language shoe design script AI generation method, characterized in that, The method comprises the following steps: Step 1: receiving shoe structured design parameters and parsing them into machine-processable format to obtain requirement input parameters, the shoe structured design parameters including shoe type parameters, core technology parameters and target market parameters; Step 2: constructing a shoe industry multi-modal knowledge base, and retrieving structured data and unstructured data from the shoe industry knowledge base according to the requirement input parameters: The structured data includes material parameters and process standards; The unstructured data includes multi-language best-selling copy templates and cultural image keywords in designer manuscripts; Step 3: generating a preliminary design copy based on a trained multi-language field vertical large model; Step 4: establishing a cross-language synonym atlas, and correcting the key terms of the preliminary design copy through cross-language synonym atlas and term constraint decoding to obtain a standard design copy; the term constraint decoding refers to that, when generating the standard design copy, if a professional term needs to be inserted in the current context, a dynamic mask is used to limit the candidate words to only the corresponding term in the cross-language synonym atlas; Step 5: correcting the cultural expression of the standard copy by using a style transfer algorithm to obtain a final design copy.

2. The cross-language shoe design script AI generation method according to claim 1, wherein, Further comprising step 6: a designer scores the generated final design copy, and updates the parameters of the multi-language field vertical large model by using a PPO algorithm based on the scoring results, regenerates a preliminary design copy based on the updated multi-language field vertical large model, and regenerates a final design copy based on steps 4 and 5.

3. The cross-language shoe design script AI generation method according to claim 2, wherein, Further comprising step 7: automatically filtering religious sensitive words and banned words of corresponding national advertising laws based on a compliance checking algorithm.

4. The cross-language shoe design script AI generation method according to claim 3, wherein, A conditional generation model is used, the input parameters are structured data and unstructured data, and the output is a multi-language preliminary design copy, and the generation formula is: ； Wherein x is the input parameter, l is the target language code, and s is the design style label extracted from the multi-language best-selling copy template retrieved in step 2.

5. The cross-language shoe design script AI generation method according to claim 4, wherein, High-frequency cultural keywords and style features in the multi-language best-selling copy template are extracted through natural language processing technology, a style market mapping database is constructed, and a suitable style label is predicted by a pre-trained classifier in combination with shoe structured design parameters.

6. The cross-language shoe design script AI generation method according to claim 5, wherein, In step 4, the data source of the cross-language synonym atlas is expert annotation and term pairs extracted from multi-language technical documents; the structure of the cross-language synonym atlas is that the graph nodes are terms, and the edges are cross-language equivalent relationships; The core formula of the dynamic mask of the term constraint decoding is: ； wherein is the key term revision result, is the synonym handling filter, is the key term before revision, is the corresponding term in the cross-lingual synonym map, is the key term collection.

7. The cross-language shoe design script AI generation method according to claim 6, wherein, Step 5 comprises the following sub-steps: Step 51, collect the historical hit copy of the target market, use the trained LDA model to extract the theme distribution of the historical hit copy, and construct the target culture distribution ; Step 52: inputting the standard design copy into a variational autoencoder, outputting hidden variable distribution parameters through the main network BERT-base of the variational autoencoder, and sampling the hidden variables; Step 53, mapping the target cultural distribution extracted by the LDA model to the latent space corresponding to the latent variables ; Step 54, enforcing the latent variable towards the target cultural distribution with style transfer loss function Alignment; Step 54: calculating the cosine similarity between the standard design copy and the standard design copy corrected by using Sentence-BERT; Step 55, adjust the latent variable formula to align the style with the target culture distribution Offset to get the adjusted latent variable; Step 56: inputting the adjusted hidden variables into the decoder to output the final design copy after cultural correction.

8. The cross-language shoe design script AI generation method according to claim 7, wherein, The adjustment formula of the hidden variables in step 55 is: ； wherein is the original copy hidden variable mean, is the adjustment strength, = 0.3, z is the standard design copy, is the culture-corrected design copy final draft.

9. A memory, comprising: A program designed by the cross-language shoe design copy AI generation method of any one of claims 1-8 is stored.

10. A cross-language shoe design script AI generation device, characterized in that: The system comprises a memory, a processor and a display, the memory is used to store the structured design parameters of footwear received in step 1, the footwear multi-modal knowledge base, the cross-language synonym atlas and the program designed by the cross-language footwear design script AI generation method of any one of claims 1-8, the processor is in communication connection with the memory, and the program output design script final draft is run, and the display is in communication connection with the processor, and the processor controls the display to display the design script final draft.

Images