A large model and knowledge base-based intelligent design method for graphic publicity materials

Abstract

The application belongs to the technical field of data processing, and particularly relates to a graphic propaganda material intelligent design method based on a large model and a knowledge base. A front end receives user text description, which is preprocessed into structured requirements through grammar correction and intention recognition. An initial design parameter set is generated by inputting the propaganda material design field into a large model (optimized through a genetic algorithm), and a preference correction is performed in combination with a user historical record statistical parameter acceptance frequency. Finally, a graphic propaganda material type size template is automatically adapted, a picture is generated, and output to the front end, thereby realizing the unification of efficient, personalized and professional design.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data processing, and particularly relates to a graphic propaganda material intelligent design method based on a large model and a knowledge base. BACKGROUND

[0002] Current graphic propaganda material design mainly relies on manual design or simple template tools: manual design has problems of low efficiency, high cost and high professional threshold, and is difficult to meet the rapid design needs of small and medium-sized enterprises or individual users; although the template tool can improve efficiency, it has the defect of template solidification and cannot flexibly generate a design scheme according to the personalized text needs of users, resulting in uneven quality of the output propaganda materials and difficulty in balancing personalization and professionalism. SUMMARY

[0003] The application proposes a graphic propaganda material intelligent design method based on a large model and a knowledge base to solve the technical problems in the above background art.

[0004] In order to achieve the above purpose, the technical scheme adopted by the application comprises the following steps:

[0005] The front end receives the text description of the graphic propaganda material input by the user, pre-processes the text description, and outputs a structured demand text, wherein the pre-processing includes grammar correction and intent recognition;

[0006] The grammar correction is used to correct ambiguities and wrong characters and determine the real requirements of the user;

[0007] The intent recognition is used to extract core demand types and clarify the intent information of the graphic propaganda material;

[0008] The structured demand text is input into a large model adapted to the field of propaganda material design, and an initial design parameter set is output; the initial design parameter set includes a color system, a layout structure, a font type and a list of visual elements;

[0009] The initial design parameter set is corrected according to the user's preferences, and a corrected design parameter set is output;

[0010] Based on the corrected design parameter set, a design scheme is generated and converted into a picture file, a size template is automatically adapted based on the type of the propaganda material, and the final picture is output to the front end.

[0011] As a preferred embodiment, the implementation of the grammar correction for correcting ambiguities and wrong characters and determining the real requirements of the user specifically comprises:

[0012] First, high-frequency error cases in the description of propaganda material requirements are collected to construct a field error expression set, a unique corresponding propaganda material design field standard expression is matched for each error expression, and a corpus is obtained by combining the field error expression set and the corresponding standard expression.

[0013] Based on the edit distance algorithm combined with the field corpus, the accurate recognition and replacement of the surface error of the demand text are defined as the text to be corrected , wherein is the number of text characters, and a candidate correct text segment set is extracted from the field corpus , wherein is the number of candidate segments, , wherein is the number of characters of the kth candidate segment;

[0014] For the text to be corrected , the minimum edit distance of each candidate segment , wherein k represents the kth candidate segment, and the minimum edit distance is calculated as follows:

[0015] , wherein represents the minimum number of edits of the first i characters of S and the first j characters of , when the text to be corrected has i characters, but the candidate correct text has no characters to compare, i deletions of the first i characters of the text to be corrected are required; when the text to be corrected has no characters, but the candidate correct text has j characters, j insertions of the first j characters of the candidate correct text are required; when both texts have characters, the minimum number of operations is selected from the three possibilities of deletion, insertion, and replacement / no operation, the deletion operation is to calculate the distance between the first i-1 characters of the text to be corrected and the first j characters of the candidate , and add one deletion operation; the insertion operation is that the text to be corrected lacks the jth character of the candidate text, which needs to be inserted and then matched, that is, the distance between the first i characters of the text to be corrected and the first j-1 characters of the candidate is calculated , and add one insertion operation; the replacement / no operation is to determine whether the ith character of the text to be corrected is consistent with the jth character of the candidate text, if consistent, no operation is required, if not consistent, replacement is required, the distance between the first i-1 characters of the text to be corrected and the first j-1 characters of the candidate is calculated , and 0 or 1 is added according to whether the characters are consistent or not is an indicator function;

[0016] An edit distance threshold is set, and the smallest candidate segment with an edit distance less than the threshold is selected , the error part in is replaced to obtain the corrected text of the syntax correction .

[0017] As an optimization, the intent recognition is used to extract the core demand type, and the specific implementation of the intent information of the graphic propaganda product includes:

[0018] First, build the core library, including promotional material type label and type identification keywords;

[0019] Corrected text for grammar correction Perform extraction operation, extract type identification keywords and type identification keywords in core library for matching calculation, calculate matching degree and select candidate types with matching degree greater than set threshold;

[0020] If the candidate type is unique, the type is the core demand type, if the candidate type is multiple, select the core demand type according to the size of the matching degree, if there is no candidate type, output demand supplement prompt.

[0021] As preferred, the operation of matching calculation of the extracted type identification keywords in the corrected text with the type identification keywords in the core library is:

[0022] Generate keyword vector of type identification keyword in core library and keyword vector in user corrected text by pre-trained BERT model;

[0023] Combine the keywords of each type in the core library to get the keyword vector set of the type , wherein is the number of types, is the number of keywords of each type, and the type center vector of each type is calculated , wherein is the keyword Importance weight of type q;

[0024] The keyword vector set extracted from the corrected text Calculate the demand aggregation vector of user keywords , wherein is the frequency weight of user mention of this keyword;

[0025] For each type q, first calculate and the cosine similarity of type center vector and the Euclidean distance , normalize the Euclidean distance to get the distance matching score ;

[0026] Finally, the cosine similarity and the distance matching score are weighted and fused to get the final matching degree.

[0027] As preferred, the specific operation of inputting the structured demand text into the promotional product design field adaptive large model and outputting the initial design parameter set is: ​

[0028] The pre-processed structured demand text is input into the text embedding layer of the promotional product design field adaptive large model to generate a demand feature vector;

[0029] The adaptive large model generates a candidate set based on the demand feature vector, and calculates the adaptation probability of each candidate for each parameter of each type: , wherein is the yth candidate of a certain type of parameter, is the feature vector, is the model output layer function;

[0030] For each type of parameter, the top 3 candidates are selected to form a parameter candidate set, and the candidate set is subjected to multi-objective optimization, and the obtained promotional product design score is used as the target for optimization by a genetic algorithm to screen the optimal parameter, which is denoted as the initial design parameter set.

[0031] As an optimization, the initial design parameter set is subjected to user preference correction, and the implementation of the corrected design parameter set is:

[0032] The historical generation records of the currently logged-in user are identified, and the acceptance frequency of the user for the design parameters in all records is counted, the acceptance frequency being the ratio of the number of times the user selects the parameter in the historical generation record process to the total number of selections, if the acceptance frequency is greater than a predetermined threshold, the user preference correction is performed, and the initial parameter set is replaced with the design parameter after the user preference correction.

[0033] Compared with the prior art, the advantages and positive effects of the present application are that the present application optimizes the design process through multi-link innovation. In the front-end preprocessing stage, the promotional product field error corpus is combined with the edit distance algorithm to accurately correct errors, and the pre-trained BERT model is matched with the core library keywords to calculate the intent recognition, ensuring accurate demand input; then the structured demand is input into the promotional product design field adaptive large model to generate parameter candidates, which are subjected to multi-objective optimization by a genetic algorithm with the design score as the target, to screen the initial parameter set containing color, layout, etc., taking into account professionalism and diversity; the parameter acceptance frequency in the user's historical records is also counted to correct the preference of the initial parameter set, improving personalization; finally, the promotional product type size template is automatically adapted to generate a picture, realizing the unification of efficient, personalized and professional design, and meeting the rapid design needs of small and medium-sized enterprises and individuals. BRIEF DESCRIPTION OF DRAWINGS

[0034] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0035] Figure 1 A structural flowchart of a graphic propaganda intelligent design method based on a large model and a knowledge base is shown. DETAILED DESCRIPTION

[0036] In order to enable a more clear understanding of the above-mentioned objects, features and advantages of the present application, the present application will be further described below in conjunction with the accompanying drawings and embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0037] In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the present application, however, the present application can also be implemented in other ways different from those described herein, and therefore, the present application is not limited to the specific embodiments disclosed in the following description.

[0038] In the current digital dissemination and new media operation environment, graphic propaganda has become an important carrier for enterprises, government agencies and individual brands to display and transmit information to the outside. However, the existing graphic propaganda design method mainly relies on the experience of manual designers or simple generation tools based on limited template library. Although manual design can provide a certain professionalism, it has problems such as high cost, low efficiency, high professional threshold of design personnel, and is not suitable for the rapid promotion needs of small and medium-sized enterprises or individual users. While the template design tool improves efficiency to a certain extent, it is difficult to accurately respond to the personalized needs of users, which can easily lead to serious homogenization of propaganda, lack of innovation and pertinence. Therefore, the present application proposes a graphic propaganda intelligent design method based on a large model and a knowledge base, and the specific implementation process is as shown in Figure 1 .

[0039] In order to overcome the ambiguity, misspelling and semantic ambiguity in user demand description, and ensure the input quality of the subsequent large model, grammar correction and intent recognition technology is used. First, the front end receives the graphic propaganda text description input by the user, pre-processes the text description, and outputs a structured demand text, the pre-processing includes grammar correction and intent recognition.

[0040] The grammar correction is used to correct ambiguity and misspelling, and to determine the user's real requirements. Specifically, first, collect high-frequency error cases in propaganda demand description to construct a domain error expression set, match a unique corresponding propaganda design domain specification expression for each error expression, and obtain a corpus combining the domain error expression set and the corresponding specification expression; based on the edit distance algorithm combined with the domain corpus, accurately identify and replace the surface errors of the demand text, and define the text to be corrected , wherein is the number of text characters, and a candidate correct text segment set is extracted from the domain corpus wherein is the number of candidate segments, wherein is the number of characters of the kth candidate segment; for the text to be corrected and each candidate segment has a minimum edit distance, wherein k represents the kth candidate segment, and the minimum edit distance is calculated as follows:

[0041] wherein, represents the minimum number of edits between the first i characters of S and the first j characters of , when the text to be corrected has i characters, but the candidate correct text has no characters to compare, i deletions of the first i characters of the text to be corrected are needed; when the text to be corrected has no characters, but the candidate correct text has j characters, j insertions of the first j characters of the candidate correct text are needed; when both the text to be corrected and the candidate correct text have characters, the minimum number of operations is selected from the three possibilities of deletion, insertion, and replacement / no operation, the deletion operation is to first calculate the distance between the first i-1 characters of the text to be corrected and the first j characters of the candidate , and then add one deletion operation; the insertion operation is that the text to be corrected lacks the jth character of the candidate text, which needs to be inserted and then matched, that is, the distance between the first i characters of the text to be corrected and the first j-1 characters of the candidate is first calculated , and then one insertion operation is added; the replacement / no operation is to determine whether the ith character of the text to be corrected is consistent with the jth character of the candidate text, if consistent, no operation is needed, if not consistent, replacement is needed, the distance between the first i-1 characters of the text to be corrected and the first j-1 characters of the candidate is calculated , and then 0 or 1 is added according to whether the characters are consistent or not, is an indicator function; a threshold value of the edit distance is set, and the edit distance less than the threshold value is selected to obtain the minimum candidate segment , the error part in is replaced to obtain the corrected text of the syntax correction Finally, the candidate segment with the minimum edit distance less than the preset threshold value is taken as the optimal replacement object, and the error part in the text to be corrected is corrected, so as to obtain the corrected text with complete semantics and no ambiguity. Through the process, the accuracy of the demand input can be significantly improved, and the interference of wrong characters or ambiguous expressions on subsequent intent recognition and design generation can be avoided.

[0042] The intent recognition is used to extract the core demand type and clearly indicate the intent information of the graphic propaganda product. First, a core library is constructed, including a propaganda product type label and a type recognition keyword; the corrected text of the syntax correction The extraction operation is performed, the type identification keyword is matched with the type identification keyword in the core library, the matching degree is calculated, and the candidate types with a matching degree greater than a set threshold are screened; if the candidate type is unique, the type is the core demand type, if the candidate type is multiple, the core demand type is selected according to the size of the matching degree, and if there is no candidate type, a demand supplement prompt is output. Specifically, first, the core library is constructed, which contains promotional material type labels (such as posters, brochures, social media covers, etc.) and their corresponding type identification keyword sets. After receiving the text corrected by grammar correction, the system performs word segmentation and feature extraction on the text, extracts the type identification keywords that may be involved, and matches them one by one with the keywords in the core library. The matching method uses a text similarity measurement method based on cosine similarity or Jaccard coefficient to calculate the matching degree of the extracted keywords and the keywords in the library. If the matching degree is greater than the set threshold, the corresponding promotional material type is taken as the candidate type. When the candidate type is unique, the type is directly determined as the core demand type of the user; when there are multiple candidate types, the type with the highest matching degree is selected as the final demand type according to the matching degree; if no candidate type can be matched, the system outputs a prompt information to ask the user to supplement the design requirement. In this way, the system can maintain robustness while effectively avoiding demand classification errors, ensuring that the subsequent design generation link takes accurate type parameters as input and improving the overall design effect and user satisfaction. The matching calculation operation is as follows: the pre-trained BERT model is used to generate the keyword vector of the type identification keyword in the core library and the keyword vector of the identification keyword in the user corrected text; the keywords of each type in the core library are combined to obtain the keyword vector set of the type wherein is the number of types, is the number of keywords for each type, and the type center vector of each type is calculated wherein is the keyword is the importance weight of the type q; and the keyword vector set extracted from the corrected text The demand aggregation vector of the user keyword is calculated wherein is the mention frequency weight of the user for this keyword; for each type q, first calculate and the cosine similarity of the type center vector and the Euclidean distance , normalize the Euclidean distance to obtain the distance matching score ; finally, the cosine similarity and the distance matching score ​The final matching degree is obtained by weighted fusion. The method takes into account the semantic relevance and spatial distance difference at the same time, can effectively improve the accuracy of demand type recognition, avoid the deviation caused by single similarity measurement, and thus improve the robustness and accuracy of user real demand type determination.

[0043] In order to convert the natural language demand into a form that can be processed by the model, the structured demand text is input into the propaganda product design field adaptive large model to output an initial design parameter set. The initial design parameter set includes a color system, a layout structure, a font type, and a list of visual elements. The specific operation of inputting the structured demand text into the propaganda product design field adaptive large model to output the initial design parameter set is to input the preprocessed structured demand text into the text embedding layer of the propaganda product design field adaptive large model to generate a demand feature vector. The adaptive large model generates a candidate set based on the demand feature vector, and calculates the adaptation probability of each candidate of each parameter to the demand: wherein is the yth candidate of a certain parameter, is the feature vector, is a model output layer function; for each type of parameter, select the top 3 candidates in terms of probability to form a parameter candidate set, and perform multi-objective optimization on the candidate set, and through the genetic algorithm, the score of the obtained promotional product design is optimized as the target to screen out the optimal parameters, which are recorded as the initial design parameter set. Specifically, first, the structured demand text obtained after the grammar correction, type identification, and demand extraction, etc. is input into the large model adapted to the promotional product design field. In the input stage, the structured demand text first enters the text embedding layer of the model, which is based on word embedding and context modeling mechanism, and converts the demand description into a unified high-dimensional demand feature vector, ensuring that the demand under different expression methods can be accurately represented in the semantic space. Subsequently, the large model generates a candidate set for each of the four types of parameters, namely color system, layout structure, font type, and visual elements, based on the demand feature vector. For each candidate in each type of parameter, the large model calculates its adaptation probability with the demand feature vector through the output layer function. Through the above calculation, the matching degree of each type of parameter candidate and the user demand can be obtained. In the candidate screening stage, the system sorts each type of parameter candidate according to the adaptation probability, and selects the top 3 candidates to form a parameter candidate set. This process ensures the retention of diversified parameter selection closest to the demand, avoids design rigidity caused by a single candidate, and provides a richer search space for subsequent combination optimization. After obtaining the candidate set of each type of parameter, the system further introduces a multi-objective optimization mechanism. The comprehensive promotional product design score is taken as the objective function. The specific optimization adopts a genetic algorithm, the core idea of which is to simulate the natural selection and genetic mechanism of biological evolution to search for the optimal solution in a vast candidate parameter space. The implementation process is as follows: first, encode each group of parameter candidates into chromosome individuals to construct an initial population, and the population size is pre-set according to the computing resources and accuracy requirements. Subsequently, calculate the objective function for each individual as the individual fitness. In the evolution process, the algorithm preferentially retains individuals with higher fitness through selection operations, exchanges part of the gene fragments between two individuals through crossover operations to generate new solution combinations, and randomly changes individual genes through mutation operations to increase population diversity and avoid falling into local optima. After several iterations, the individuals in the population gradually converge to the optimal solution. Finally, the individual with the highest fitness is the initial design parameter set obtained by optimization. The advantage of this method is that it can realize global search under complex multi-objective conditions, ensuring the coordination and innovation of parameter combination. Finally, the genetic algorithm outputs the optimal solution after several iterations, which is the initial design parameter set. This parameter set not only ensures high matching degree with user demand, but also realizes global coordination among design parameters under the constraint of multi-objective optimization, thereby providing stable and reliable input for subsequent specific design generation.

[0044] Then the initial design parameter set is corrected according to user preferences, and a corrected design parameter set is output. The historical generation records of the currently logged-in user are identified, and the acceptance frequency of the user for the design parameters in all records is counted. The acceptance frequency is the ratio of the number of times the user selects the parameter in the historical generation record process to the total number of selections. If the acceptance frequency is greater than a preset threshold, the user preference correction is performed, and the initial parameter set is replaced by the design parameter corrected according to the user preference. Specifically, the system automatically identifies the historical generation records of the currently logged-in user, and extracts the design parameters involved therein. Subsequently, the acceptance frequency of the user for each type of parameter is counted. The acceptance frequency is defined as the ratio of the number of times the user selects a certain parameter in the historical records to the total number of selections of this type of parameter. For example, if the user selects a certain font type in ten generations seven times, the acceptance frequency of the font is 0.7. The system presets an acceptance frequency threshold. When the acceptance frequency of a certain parameter is greater than the threshold, it means that the parameter can fully represent the user's stable preference. At this time, the system replaces the dimension of this type of parameter in the initial design parameter set with the user preference parameter, thereby generating a corrected design parameter set. This process ensures that the generation result not only retains the global rationality of the large model and the optimization algorithm, but also incorporates the user's long-term personalized preferences, realizes the organic combination of model output and user aesthetic habits, and improves the practicality and satisfaction of the final design scheme.

[0045] Finally, based on the corrected design parameter set, a design scheme is generated and converted into a picture file, a size template is automatically adapted based on the type of promotional materials, and the final picture is output to the front end. Specifically, after obtaining the corrected design parameter set, the system takes it as the core input to drive the automatic generation of promotional material design schemes. First, the system calls the design generation engine to convert abstract parameters into specific design schemes according to the design parameters. The engine automatically renders color matching, layout distribution, and element arrangement through the combination of rule templates and graphic rendering models, forming a complete promotional material visual composition. Subsequently, the system automatically adapts the corresponding size template according to the promotional material type (such as posters, flyers, social media pictures, etc.) identified in the user requirements, ensuring that the generated design scheme meets the usage specifications of this type in terms of resolution, proportion, and margin. Finally, the system renders the generated design scheme into a standardized picture file (such as PNG or JPEG format) and outputs it to the front-end interface for the user to directly preview and download for use. This process realizes an automatic closed loop from parameters to pictures, ensuring the efficiency of design generation and improving the practicality and usability of promotional material products.

[0046] The above merely describes preferred embodiments of the present application, but is not intended to limit the present application to other forms, any person skilled in the art can make changes or modifications to the above disclosed technical contents into equivalent embodiments with equivalent changes, and apply to other fields, but any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application without departing from the technical solution content of the present application still belongs to the protection scope of the present application technical solution.

Claims

1. A large model and knowledge base based intelligent design method for graphic publicity materials, characterized in that, The method comprises the following steps: The front end receives a text description of a graphic publicity product input by a user, pre-processes the text description, and outputs a structured demand text, wherein the pre-processing comprises syntax error correction and intent recognition; The syntax error correction is used to correct ambiguities and wrong characters and determine the real requirements of the user; The intent recognition is used to extract core demand types and determine the intent information of the graphic publicity product; The structured demand text is input into a publicity product design field adaptive large model, and an initial design parameter set is output; The initial design parameter set comprises a color system, a layout structure, a font type, and a list of visual elements; The initial design parameter set is corrected according to user preferences, and a corrected design parameter set is output; Based on the corrected design parameter set, a design scheme is generated and converted into a picture file, a size template is automatically adapted based on the type of the publicity product, and a final picture is output to the front end; The implementation of the syntax error correction for correcting ambiguities and wrong characters and determining the real requirements of the user specifically comprises: First, collect high-frequency error cases in publicity product demand descriptions to construct a domain error expression set, match a unique corresponding publicity product design domain specification expression for each error expression, and obtain a corpus by combining the domain error expression set and the corresponding specification expression; Based on the edit distance algorithm combined with the field corpus, the accurate recognition and replacement of the surface error of the demand text are defined as the text to be corrected wherein is the number of text characters, and a candidate correct text segment set is extracted from the field corpus wherein is the number of candidate segments, wherein is the number of characters of the kth candidate segment; For the text to be corrected the minimum edit distance of each candidate segment where k represents the kth candidate segment, and the minimum edit distance is calculated in a manner that wherein, denotes the minimum edit distance between the first i characters of S and the first j characters of T, when there are i characters in S to be corrected, but T has no characters to compare with, i deletions are needed to delete the first i characters of S; when S has no characters to be corrected, but T has j characters, j insertions are needed to insert the first j characters of T; when both S and T have characters, the minimum edit distance is selected from the three possibilities of deletion, insertion and substitution / no operation, the deletion is calculated by first calculating the distance between the first i-1 characters of S and the first j characters of T , then adding one deletion; the insertion is calculated by first calculating the distance between the first i characters of S and the first j-1 characters of T , then adding one insertion; the substitution / no operation is calculated by judging whether the i-th character of S is identical to the j-th character of T, if identical, no operation is needed, if not identical, substitution is needed, the distance between the first i-1 characters of S and the first j-1 characters of T is calculated , then adding 0 or 1 according to whether the characters are identical or not, is an indicator function. Setting an edit distance threshold, selecting an edit distance less than the threshold, and obtaining the smallest candidate segment , the error part in the is replaced to obtain the corrected text of the syntax error correction .

2. The method according to claim 1, characterized in that, The specific implementation of the intent recognition for extracting core demand types and determining the intent information of the graphic publicity product comprises: First, construct a core library comprising publicity product type labels and type recognition keywords; Corrected text for grammatical correction The extraction operation is performed, the type identification keyword in the extraction type is matched with the type identification keyword in the core library to calculate a matching degree, and a candidate type with a matching degree greater than a set threshold is screened. If the candidate type is unique, the type is the core demand type; if the candidate type is multiple, the core demand type is selected according to the size of the matching degree; if there is no candidate type, a demand supplement prompt is output.

3. The method according to claim 2, characterized in that, The operation of matching and calculating the type recognition keywords in the corrected text with the type recognition keywords in the core library is: Generate keyword vectors of the type recognition keywords in the core library and recognition keyword vectors in the user corrected text through a pre-trained BERT model; combining the keywords of each type in the core library to obtain a keyword vector set of the type wherein is the number of types, is the number of keywords of each type, and a type center vector of each type is calculated wherein is a keyword importance weight of the type q; a set of keyword vectors extracted from the revised text computing a demand aggregation vector for the user keyword wherein is the frequency of the user's mentions of this keyword For each type q, first calculate and the cosine similarity of the type center vector and the Euclidean distance , normalize the Euclidean distance to obtain the distance matching score ;​ Finally, the cosine similarity is matched with the distance score to obtain the final matching degree.

4. The method according to claim 1, characterized in that, The specific operation of inputting the structured demand text into the publicity product design field adaptive large model to output the initial design parameter set is: Input the pre-processed structured demand text into the text embedding layer of the publicity product design field adaptive large model to generate a demand feature vector; The adaptation large model generates a candidate set based on the demand feature vector, and calculates the adaptation probability of each candidate of each type of parameter with the demand: wherein is the yth candidate of a type of parameter, is the feature vector, is a model output layer function; For each type of parameter, select the top 3 candidates in terms of probability to form a parameter candidate set, and perform multi-objective optimization on the candidate set to obtain a publicity product design score as the target for optimization, and select the optimal parameter as the initial design parameter set.

5. The method according to claim 1, characterized in that, The implementation of correcting the initial design parameter set according to user preferences and outputting the corrected design parameter set is: Recognize the historical generation records of the currently logged-in user, and count the acceptance frequency of the user for the design parameters in all records, wherein the acceptance frequency is the ratio of the number of times that the user selects the parameter in the historical generation record process to the total number of selections; if the acceptance frequency is greater than a preset threshold, the user preference correction is performed, and the initial parameter set is replaced with the user preference corrected design parameter.

Images