Social media copywriting intelligent generation method and system based on large language model

Abstract

The invention discloses a social media copywriting intelligent generation method and system based on a big language model, and relates to the technical field of electronic commerce, and the method comprises the steps: generating a first copywriting of a preset commodity based on the correlation information and image of the preset commodity through a preset big language model, the first copywriting meets basic copywriting requirements of the target social media platform; adding a preset element of the target social media platform in the first copywriting to obtain a second copywriting; performing semantic analysis on the second copywriting to generate a title of the second copywriting; and combining the second copywriting with the title of the second copywriting to generate a target copywriting of the preset commodity. According to the method and the device, diversified results can be generated according to different input contents through the preset large language model so as to meet different requirements and scenes, and the first copywriting generally has relatively high semantic coherence and naturalness and meets the feature and style requirements of the target social media platform.

Description

Technical Field

[0001] The present invention relates to the field of e-commerce technology, and in particular to a method and system for intelligently generating social media copy based on a large language model. Background Art

[0002] In the e-commerce sector, creating social media content is a crucial step in product promotion. With the continuous expansion and iteration of product categories, the volume of content required is rapidly increasing. Traditional content creation relies primarily on the independent work of marketers, but this manual writing model is facing efficiency bottlenecks. Especially in scenarios such as advertising marketing and hot topic operations, the market demand for the rapid generation of high-quality, diverse copy is becoming increasingly urgent.

[0003] Currently, rule-based automation tools are often used to improve the efficiency of writing product copy. Rule-based automation tools rely on predefined rules and templates to generate copy. This method usually generates different content by setting a fixed template, combined with specific grammatical structures and vocabulary replacements. The significant advantages include: 1. Efficiency: It can quickly generate a large amount of copy, especially suitable for scenarios that require a large amount of repetitive content, such as product descriptions and specifications. 2. Consistency: The generated content has a unified style and structure, which helps maintain the consistency of the brand image. Although rule-based automation tools perform well in terms of efficiency and consistency, they are insufficient in flexibility and personalization. Due to the reliance on fixed templates, the generated content usually lacks innovation and diversity, and is difficult to customize according to specific scenarios or user needs. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to address the deficiencies of the existing technology and specifically provide a method and system for intelligently generating social media copy based on a large language model, as follows:

[0005] 1) In a first aspect, the present invention provides a method for intelligently generating social media copy based on a large language model. The specific technical solution is as follows:

[0006] Based on the relevant information and images of the preset product and using a preset large language model, a first copy of the preset product is generated, wherein the first copy meets the basic copy requirements of the target social media platform;

[0007] Add the preset elements of the target social media platform to the first copy to get the second copy;

[0008] Performing semantic analysis on the second copy to generate a title for the second copy;

[0009] Combine the second copy and the title of the second copy to generate the target copy for the preset product.

[0010] The application provides a social media script intelligent generation method based on a large language model.

[0011] The preset large language model can generate diversified results according to different input contents to meet different needs and scenes, and the generated first script usually has high semantic coherence and naturalness. In addition, the preset elements of the target social media platform are added in the first script to ensure that the generated target script of the preset commodity meets the characteristics and style requirements of the target social media platform, thereby further improving user satisfaction.

[0012] Based on the above scheme, the social media script intelligent generation method based on the large language model can be further improved as follows.

[0013] Further, based on the association information and the image of the preset commodity, and by using the preset large language model, a first script of the preset commodity is generated, which includes:

[0014] When the reference script of the preset commodity is received, the writing detail information of the reference script is identified by using the preset large language model, and based on the writing detail information of the reference script and the association information and the image of the preset commodity, a third script of the preset commodity is generated, and the third script is used as the first script.

[0015] When the reference script of the preset commodity is not received, based on the association information and the image of the preset commodity, a fourth script of the preset commodity is generated by using the preset large language model, and the fourth script is used as the first script.

[0016] The beneficial effects of the above further scheme are as follows: when the user provides the reference script of the preset commodity, the reference script can provide rich materials and accurate information, so that the third script has more key points of the preset commodity, highlights the advantages of the commodity, and can ensure the authenticity and feasibility of the third script, thereby avoiding exaggerated or false descriptions and effectively helping sales growth. When the user does not provide the reference script of the preset commodity, the fourth script of the preset commodity is automatically generated based on the association information and the image of the preset commodity, which can significantly improve the writing efficiency and save time and effort.

[0017] Further, before the third script is used as the first script, the correctness, tone and diction of the third script are evaluated to obtain a first evaluation result, and the sentence structure of the third script and the reference script is compared and analyzed to obtain a comparison and analysis result. According to the first evaluation result and the comparison and analysis result, it is determined whether to optimize the third script. If yes, the third script is optimized according to the first evaluation result and / or the comparison and analysis result, and the optimized third script is used as the first script.

[0018] Before the fourth script is taken as the first script, further comprising: evaluating the correctness, tone and wording of the fourth script to obtain a second evaluation result; determining whether to optimize the fourth script according to the second evaluation result, if yes, optimizing the fourth script according to the second evaluation result, and then taking the optimized fourth script as the first script.

[0019] The beneficial effects of the further scheme are: according to the first evaluation result and the comparative analysis result, the third script is optimized, which can ensure that the optimized third script is consistent with the reference script in style, and can ensure that the optimized third script has appropriate tone and wording, and can ensure the correctness of the optimized third script. According to the second evaluation result, the fourth script is optimized, which can ensure that the optimized fourth script has appropriate tone and wording, and can ensure the correctness of the optimized fourth script.

[0020] Further, before the semantic analysis of the second script, further comprising:

[0021] The second script is subjected to forbidden word identification, and when the forbidden word is identified, the identified forbidden word is processed to obtain a fifth script, and the fifth script is taken as the second script for semantic analysis.

[0022] The beneficial effects of the further scheme are: the compliance and safety of the second script can be ensured.

[0023] Further, further comprising:

[0024] Receiving feedback information of the user on the target script;

[0025] According to the feedback information, the target script is optimized.

[0026] The beneficial effects of the further scheme are:

[0027] According to the feedback information of the user on the target script, the target script can be rewritten and optimized in a targeted manner, so as to ensure that the final script (optimized target script) is more in line with the expectations and needs of the user, and improve the user experience.

[0028] 2) In a second aspect, the present application also provides a social media script intelligent generation system based on a large language model, and the specific technical scheme is as follows:

[0029] Comprising a first script generation module, a second script generation module, a title generation module and a target script generation module;

[0030] The first text generation module is configured to generate a first text of the preset commodity based on the associated information and the image of the preset commodity and by using a preset large language model, wherein the first text meets the basic requirements of the text of the target social media platform.

[0031] The second text generation module is configured to add preset elements of the target social media platform to the first text to obtain a second text.

[0032] The title generation module is configured to perform semantic analysis on the second text to generate a title of the second text.

[0033] The target text generation module is configured to combine the second text and the title of the second text to generate a target text of the preset commodity.

[0034] Based on the above scheme, the social media text intelligent generation system based on the large language model can be further improved as follows.

[0035] Further, the first text generation module is specifically configured to:

[0036] When the reference text of the preset commodity is received, the preset large language model is used to identify the writing detail information of the reference text, and the third text of the preset commodity is generated based on the writing detail information of the reference text and the associated information and the image of the preset commodity, and the third text is used as the first text.

[0037] When the reference text of the preset commodity is not received, the fourth text of the preset commodity is generated based on the associated information and the image of the preset commodity by using the preset large language model, and the fourth text is used as the first text.

[0038] Further, the system further comprises a first evaluation module and a second evaluation module.

[0039] The first evaluation module is configured to evaluate the correctness, tone and diction of the third text before the third text is used as the first text to obtain a first evaluation result, and to compare and analyze the sentence structure of the third text with the reference text to obtain a comparison and analysis result, and to determine whether to optimize the third text based on the first evaluation result and the comparison and analysis result, and if so, to optimize the third text based on the first evaluation result and / or the comparison and analysis result, and the first text generation module is further configured to use the optimized third text as the first text.

[0040] The second evaluation module is configured to evaluate the correctness, tone and diction of the fourth text before the fourth text is used as the first text to obtain a second evaluation result, and to determine whether to optimize the fourth text based on the second evaluation result, and if so, to optimize the fourth text based on the second evaluation result, and the first text generation module is further configured to use the optimized fourth text as the first text.

[0041] Further, the application further comprises a forbidden word identification processing module, which is configured to: before performing semantic analysis on the second script, perform forbidden word identification on the second script, when a forbidden word is identified, perform processing on the identified forbidden word, obtain a fifth script, and use the fifth script as the second script for semantic analysis.

[0042] Further, the application further comprises a target script optimization module, which is configured to: receive feedback information of the target script from a user, and optimize the target script according to the feedback information.

[0043] 3) In a third aspect, the application further provides an electronic device, which comprises a processor and a memory coupled to the processor, and the memory stores at least one computer program, and the at least one computer program is loaded and executed by the processor to enable the electronic device to implement any of the above-mentioned social media script intelligent generation methods based on a large language model.

[0044] 4) In a fourth aspect, the application further provides a computer readable storage medium, and the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement any of the above-mentioned social media script intelligent generation methods based on a large language model.

[0045] It should be noted that the technical solutions of the second to fourth aspects of the application and the corresponding possible implementation manners have the beneficial effects as described above for the first aspect and the corresponding possible implementation manners, and will not be described here. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the application, the following will briefly introduce the drawings needed to be used in the description of the embodiments of the application:

[0047] Figure 1 FIG. 1 is a flowchart of a social media script intelligent generation method based on a large language model according to an embodiment of the application;

[0048] Figure 2 FIG. 2 is a structural diagram of a social media script intelligent generation system based on a large language model according to an embodiment of the application;

[0049] Figure 3 FIG. 3 is a structural diagram of an electronic device according to an embodiment of the application. DETAILED DESCRIPTION

[0050] The principles and features of the application are described below, and the examples are only used to explain the application and are not used to limit the scope of the application.

[0051] The technical solutions of the present application and how the technical solutions of the present application solve the above technical problems will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present application will be described below with reference to the accompanying drawings.

[0052] As shown in Figure 1 The method for generating a social media script based on a large language model according to an embodiment of the present application comprises the following steps:

[0053] S1, based on the associated information and image of the preset commodity, and using a preset large language model, a first script of the preset commodity is generated, wherein the first script meets the basic requirements of the script of the target social media platform;

[0054] The associated information of the preset commodity includes the name of the preset commodity, the target audience, the use scenario, the pain point, the selling point, the endorsement, and the script style (also known as the note style), etc. The preset commodity can be clothes, electronic products, or kitchen utensils, etc. The associated information of the preset commodity is provided by the user.

[0055] The preset large language model can be GPT-3, GPT-4, Deepseek V3, or Doubao, etc. It can be selected according to actual conditions.

[0056] S2, adding preset elements of the target social media platform to the first script to obtain a second script;

[0057] The preset elements of the target social media platform include topic labels (#), user markers (@user), and interactive dialogues, etc. They can be set according to actual conditions.

[0058] S3, performing semantic analysis on the second script to generate a title of the second script;

[0059] S4, combining the second script and the title of the second script to generate a target script of the preset commodity.

[0060] In this embodiment, the preset large language model can generate diversified results according to different input contents to meet different needs and scenarios. The first script generated usually has high semantic coherence and naturalness. Moreover, by adding preset elements of the target social media platform to the first script, it is ensured that the target script of the preset commodity generated meets the characteristics and style requirements of the target social media platform, further improving user satisfaction.

[0061] Optionally, in S1, based on the associated information and image of the preset commodity, and using a preset large language model, a first script of the preset commodity is generated, comprising:

[0062] 1) When the reference text of the preset commodity is received, the writing detail information of the reference text is identified by using the preset large language model, and the third text of the preset commodity is generated according to the writing detail information of the reference text and the associated information and image of the preset commodity, and the third text is taken as the first text;

[0063] 2) When the reference text of the preset commodity is not received, the fourth text of the preset commodity is generated by using the preset large language model based on the associated information and image of the preset commodity, and the fourth text is taken as the first text.

[0064] When the user provides the reference text of the preset commodity, the reference text can provide rich materials and accurate information, so that the third text has more key points of the preset commodity, highlights the advantages of the commodity, and can ensure the authenticity and feasibility of the third text, thereby avoiding exaggerated or false description and effectively helping sales growth. When the user does not provide the reference text of the preset commodity, the fourth text of the preset commodity is automatically generated based on the associated information and image of the preset commodity, which can significantly improve the writing efficiency and save time and effort.

[0065] In another embodiment, when the reference text of the preset commodity is not received, the process of generating the third text of the preset commodity includes:

[0066] ①Obtain the structured text data and unstructured text data of the preset commodity, the structured text data includes the commodity name, target group, use scene and endorsement, etc., and the unstructured text data includes user evaluation text and marketing keywords;

[0067] ②The feature encoding of the structured text data is performed by a natural language processing model to generate a first text feature vector, and a second text feature vector of the unstructured text data is extracted by a semantic analysis model;

[0068] Wherein, when processing the structured text data, the natural language processing model will first perform word segmentation on the text, and split the sentence into word or sub-word units. Then, by using pre-trained word embedding technology (such as Word2Vec, GloVe, etc.) or the embedding layer of the model itself, each word is converted into a fixed-dimensional vector to preliminarily capture the semantic information of the word. Then, the natural language processing model will encode the text sequence through a recurrent neural network (RNN), a long short-term memory network (LSTM), a Transformer, etc. For example, the self-attention mechanism in the Transformer analyzes the mutual relationship between the words, combines the information according to the weight of the words in different positions, and updates the vector representation of each word, thereby generating a first text feature vector containing rich semantic and context association.

[0069] First, the text is preprocessed, including tokenization, stopword removal, etc. Then, a pre-trained semantic analysis model (such as BERT, Transformer, etc.) is used to encode the text. The model calculates the semantic representation of each word based on its context position in the text through multiple layers of neural networks. For example, the self-attention mechanism of the Transformer model analyzes the relationship between each word and other words to generate a vector reflecting the semantic and contextual association of the word. Finally, the model integrates these word vectors to generate a second text feature vector that can represent the semantic information of the entire text.

[0070] ③Convolutional neural network is used to extract multi-scale features from the product image to generate an image feature vector with spatial semantic association. Specifically:

[0071] The product image is input into the convolutional neural network, which contains multiple convolutional layers and pooling layers. The convolutional layer uses different size convolutional kernels to perform convolution operations on the image to extract local features such as edges and textures. Different size convolutional kernels can capture features at different scales, achieving multi-scale feature extraction. The pooling layer then down-samples the feature map after convolution to reduce the dimension of the feature map, preserving important feature information and enhancing the model's robustness to image size changes. As the network depth increases, the convolutional layer continuously extracts higher-level semantic features such as shape and category. Finally, the features extracted from each layer are integrated to generate an image feature vector with spatial semantic association through full connection layer operations, etc. This vector can comprehensively and compactly represent the features of the product image and can be used for image classification, retrieval, etc.

[0072] ④Construct a multi-modal feature fusion matrix to cross-modal attention weight the first text feature vector, the second text feature vector and the image feature vector to generate a dimension-aligned multi-modal joint feature vector. Specifically:

[0073] First, the first text feature vector, the second text feature vector and the image feature vector are uniformly adjusted to the same dimension, which is usually achieved through full connection layer or normalization operation. Then, a multi-modal feature fusion matrix is constructed to combine the three feature vectors into a matrix form, with each feature vector as a row or column of the matrix. Then, the cross-modal attention mechanism is applied. A self-attention model is constructed to calculate the correlation of each feature vector in different modalities to obtain an attention weight matrix. These weights reflect the importance of each feature vector in the fusion process. Finally, the original feature vector is multiplied by the attention weight and weighted sum or splicing operation is performed to generate a dimension-aligned multi-modal joint feature vector. This vector integrates the semantic information of text and image and can be used for multi-modal tasks such as product classification and recommendation.

[0074] The multi-modal joint feature vector is input into a preset large language model, and a candidate script set is generated through a dynamic temperature sampling algorithm, specifically:

[0075] When the multi-modal joint feature vector is input into the preset large language model, the feature vector is first adapted to the input layer of the model, which may be adjusted in dimension through linear transformation. The large language model uses its internal self-attention mechanism and multi-layer Transformer architecture to perform deep semantic understanding and modeling on the input feature vector. In the text generation phase, a dynamic temperature sampling algorithm is used, and the temperature parameter controls the degree of dispersion of the sampling probability distribution. A higher value generates more random and diverse candidate scripts, and a lower value generates more certain and stable scripts. Through multiple sampling, a candidate script set containing various styles and contents is generated based on the probability distribution of the model output, to meet the needs of script diversification in different application scenarios.

[0076] The candidate script is iteratively optimized using a keyword coverage-based evaluation model to output a fourth script that meets the preset marketing indicators, specifically:

[0077] First, key marketing keywords are extracted from marketing goals and product characteristics to build a keyword library. Then, based on the keyword coverage evaluation model, the number and proportion of keywords contained in the candidate script are calculated to evaluate the coverage of the script on marketing points. According to the evaluation results, the candidate script is iteratively optimized using the text generation model, such as adjusting the script content to increase keyword coverage. Iterative optimization is repeated until the script meets the preset marketing indicators, such as keyword coverage meeting the standard and script quality meeting the requirements, and finally a fourth script that highlights marketing points and has high quality is output to achieve precise marketing.

[0078] Optionally, in the above technical solution, before the third script is used as the first script, the correctness, tone, and diction of the third script are evaluated to obtain a first evaluation result, and the sentence structure of the third script is compared and analyzed with the reference script to obtain a comparative analysis result. According to the first evaluation result and the comparative analysis result, it is determined whether to optimize the third script. If so, the third script is optimized according to the first evaluation result and / or the comparative analysis result, and the third script is used as the first script, including: using the optimized third script as the first script. According to the first evaluation result and the comparative analysis result, the third script is optimized, which can ensure that the optimized third script is consistent with the reference script in style, has appropriate tone and diction, and is correct.

[0079] Optionally, in the above technical solution, before the fourth script is taken as the first script, further comprising: evaluating the correctness, tone and diction of the fourth script to obtain a second evaluation result; determining whether to optimize the fourth script according to the second evaluation result, if yes, optimizing the fourth script according to the second evaluation result, and then taking the optimized fourth script as the first script.

[0080] Optionally, in the above technical solution, before the semantic analysis of the second script, further comprising:

[0081] The second script is subjected to forbidden word identification, and when a forbidden word is identified, the identified forbidden word is processed to obtain a fifth script, and the fifth script is taken as the second script for semantic analysis. The compliance and safety of the second script can be ensured.

[0082] Optionally, in the above technical solution, further comprising:

[0083] S5, receiving feedback information of the user on the target script, and optimizing the target script according to the feedback information. According to the feedback information of the user on the target script, the target script can be rewritten and optimized, so as to ensure that the final script (optimized target script) is more in line with the expectations and needs of the user, and improve the user experience.

[0084] The method is specifically described through the following embodiments, comprising the following steps:

[0085] S10, receiving and analyzing the associated information and image of the preset commodity input by the user, which is specifically realized by calling the application program interface (MLLM API) of the multi-modal large language model, and combining the name of the preset commodity to identify the real features of the preset commodity, such as color, material, etc. The extracted real features will be used as context data in the subsequent script generation process.

[0086] S11, judging whether the user inputs a reference script (the reference script can also be called a reference note) of the preset commodity, obtaining a judgment result, and then:

[0087] ① When the judgment result is yes, the writing detail information of the reference script is identified by using the preset large language model, and a third script (the third script can also be called an imitated note) of the preset commodity is generated according to the writing detail information of the reference script and the associated information and image of the preset commodity, specifically:

[0088] A pre-set copywriting prompt template is called, which aims to help the pre-set large language model understand and imitate the style of the reference note. The present application is based on the BROKE framework to build the basic structure of the copywriting prompt template, and integrates the associated information and image description of the pre-set commodity (i.e. the real characteristics of the pre-set commodity) into the copywriting prompt template. By explicitly describing the task goal, the pre-set large language model is guided to focus on the writing detail information of the reference script, including sentence style, sentence structure, word selection and tone, etc. According to the writing detail information of the reference script and the associated information and image of the pre-set commodity, the third script of the pre-set commodity is generated to ensure that the third script not only contains the information of the pre-set commodity, and ensures that the third script is consistent with the reference script in tone and style.

[0089] Among them, the associated information (such as name, characteristics, purpose, etc.) and image description of the pre-set commodity are integrated to construct a comprehensive description containing text information and visual information of the commodity. The copywriting prompt template is designed based on the BROKE framework, which emphasizes innovation under resource constraints, so the template will guide the model to creatively expand on the basis of existing information. The template may contain fixed prompt structures, such as "Commodity name: [name], characteristics: [characteristics list], image description: [image content], please generate an attractive script". The integrated information is filled into the copywriting prompt template to form a complete prompt, which is then input into the large language model. The purpose of this step is to use the text generation capability of the model to generate relevant third scripts based on the given prompt.

[0090] In another embodiment, the reference text is input into the preset large language model, which encodes the reference text and converts the text into vector form. Then, by analyzing the attention mechanism inside the model, the characteristics of the style of the text in the text are identified, such as whether it is formal, humorous, etc. Then, using the syntactic analysis ability of the model, the sentence structure is parsed, such as determining whether it is a simple sentence or a complex sentence. In terms of word selection, the model will determine the selection of words in the text according to the semantic and frequency information of the words. For tone recognition, the preset large language model will determine the tone of the text according to the emotional tendency and context of the words, whether it is positive, negative, or neutral. Finally, the identification results are integrated to obtain the writing detail information of the reference text. Then, the reference text, the associated information and the image of the preset commodity are input into the large language model. The model first encodes the text information and extracts semantic features; the image information is extracted to obtain visual features, and the writing detail information of the reference text (sentence style, sentence structure, word selection, tone), text features and image features of the commodity associated information are fused to construct a multi-modal feature vector, providing comprehensive semantic and visual basis for text generation. Based on the fused feature vector, the text generation capability of the large language model is used to follow the style characteristics of the reference text, combine the commodity information and image content, and generate commodity text word by word. During the generation process, appropriate words and sentence patterns are explored through sampling or beam search strategies to ensure that the text meets the target style and contains key information about the commodity. After generating candidate texts, an evaluation model is used to filter and optimize the quality of the text, such as checking the coverage of key words, the fluency of sentences, and the relevance to marketing, etc., and finally outputting the third text that meets the preset marketing goals and style requirements.

[0091] ②When the judgment result is no, based on the associated information and the image of the preset commodity, a fourth text of the preset commodity is generated using the preset large language model, specifically:

[0092] First, a pre-set note generation prompt template is called, which uses the BROKE framework to build the basic structure of the note generation prompt template by describing background information, determining responsibilities, clarifying goals, and setting key results. The associated information and image of the preset commodity (i.e. the real characteristics of the preset commodity) are used as basic data, which are integrated into the note generation prompt template to generate a corresponding prompt instance, and the prompt instance is input into the preset large language model. Finally, a fourth text of the preset commodity (the fourth text can also be referred to as a basic note) is generated, which meets the basic requirements of the text of the target social media platform.

[0093] S12, text verification, specifically:

[0094]

[0095] The correctness, tone and wording of the third script are evaluated using the preset large language model to obtain a first evaluation result, and the sentence structure of the third script and the reference script are compared and analyzed to obtain a comparison and analysis result. According to the first evaluation result and the comparison and analysis result, it is determined whether to optimize the third script. If yes, the third script is optimized according to the first evaluation result and / or the comparison and analysis result, and the third script is taken as the first script, specifically:

[0096] The third script and the associated information of the preset commodity are input into the preset imitation note verification template, and the template integrates these information into a prompt instance. The role of this prompt instance is to provide clear input and task instructions for the large language model, so that the model can analyze and optimize the script for specific targets. The prompt instance is input into the preset large language model. The model will first encode the prompt instance and extract the key information and semantic features in it, including the content of the third script and the associated information of the commodity. The preset large language model compares and analyzes the sentence structure of the third script and the reference script according to the instructions in the prompt instance. The model will identify the similarities and differences between the two scripts in terms of sentence complexity, sentence type usage (such as declarative sentences, interrogative sentences, etc.), sentence component structure (such as the arrangement of subject-predicate-object), etc., and generate a comparison and analysis result. According to the comparison and analysis result, if there is a significant difference between the sentence structure of the third script and the reference script, it means that the third script may not meet the expected style or expression effect. At this time, the large language model is used to optimize the third script and adjust its sentence structure to make it closer to the style of the reference script, while retaining the key information and marketing points of the commodity.

[0097] The quantitative indicators of whether there is a significant difference between the sentence structure of the third script and the reference script include sentence complexity difference, sentence type distribution difference, sentence component difference, sentence length difference and sentence similarity, specifically:​

[0098] 1) Sentence complexity difference: The number of complex sentences, including clauses, parallel structures, and modifiers, was calculated between the third and reference texts. If the proportion of complex sentences in the third text exceeded 30% compared to the reference text, a significant difference was considered.

[0099] 2) Sentence pattern distribution differences: The proportions of declarative sentences, interrogative sentences, imperative sentences, and exclamatory sentences in the third copy and the reference copy were counted. If the difference in the distribution of major sentence patterns between the third copy and the reference copy exceeded 20%, it was considered to be a significant difference.

[0100] 3) Sentence Component Differences: Analyze the frequency of sentence components (such as subject, predicate, object, attributive, adverbial, etc.) between the third and reference texts. If the difference in the frequency of major sentence components between the third and reference texts exceeds 25%, it is considered a significant difference.

[0101] 4) Sentence length difference: The average length of the sentences (in words or characters) was calculated. If the average sentence length of the third copy differed from that of the reference copy by more than 30%, it was considered to be significantly different.

[0102] 5) Sentence similarity: The third copy and the reference copy are analyzed using syntax tree similarity or dependency-based similarity metrics. If the sentence structure similarity is less than 0.6 (assuming the similarity range is 0 to 1), it is considered to be significantly different.

[0103] When the comparative analysis results show that there are no significant differences in the sentence structure between the third copy and the reference copy, and when the first evaluation result shows that the third copy has no problems with its correctness, tone, and wording, the third copy can be directly used as the first copy. When the comparative analysis results show that there are significant differences in the sentence structure between the third copy and the reference copy, and when the first evaluation result shows that there are no problems with its correctness, tone, and wording, the third copy will be optimized based on the comparative analysis results. When the comparative analysis results show that there are no significant differences in the sentence structure between the third copy and the reference copy, and when the first evaluation result shows that there are problems with its correctness, tone, and wording, the third copy will be optimized based on the first evaluation result.

[0104] ② When generating the fourth copy, the correctness, tone, and wording of the fourth copy are evaluated to obtain a second evaluation result; based on the second evaluation result, determining whether to optimize the fourth copy; if so, optimizing the fourth copy based on the second evaluation result, and using the fourth copy as the first copy, including: using the optimized fourth copy as the first copy, specifically:

[0105] The preset generation note verification template is adopted. The associated information of the preset commodity and the fourth script are input into the preset generation note verification template together to form a prompt instance, and the prompt instance is input into the preset large language model. The fourth script is comprehensively checked by using the preset large language model to ensure the accuracy and consistency of the information in the optimized fourth script, so as to avoid any misleading or inaccurate content. Moreover, the tone and wording of the fourth script are also evaluated to ensure that the commodity advantages are conveyed while maintaining a true and credible style.

[0106] When the second evaluation result is that the correctness, tone and wording of the fourth script have problems, the fourth script is optimized according to the second evaluation result. When the second evaluation result is that the correctness, tone and wording of the fourth script have no problems, the fourth script is directly used as the second script.

[0107] S13, adding preset elements of the target social media platform in the first script to obtain a second script, specifically:

[0108] The preliminary optimized note, that is, the first script, is input into the preset specific element adding template to form a prompt instance, and the prompt instance is input into the preset large language model to make the preset large language model modify and supplement the details of the first script. That is, preset elements of the target social media platform, such as topic labels (#), user markers (@ user), and interactive dialogues, are added to the first script to ensure that the second script meets the characteristic and style requirements of the target social media platform. Moreover, the emojis used in the first script are also reviewed to ensure that the emojis used in the first script are used appropriately and meet the specifications of the target social media platform. According to the review results, the preset large language model is also used to make necessary fine-tuning and optimization of the first script to ensure that the second script is both attractive and meets the platform's usage standards.

[0109] S14, performing forbidden word identification on the second script, and when a forbidden word is identified, processing the identified forbidden word to obtain a fifth script, specifically:

[0110] The second document is combined with a pre-configured banned word template to form a prompt instance. This instance is input into a preset large language model. The preset large language model is used to comprehensively scan the second document to identify whether the second document includes banned words listed in the banned word template and words similar to those listed in the banned word template. The banned words listed in the banned word template and words similar to those listed in the banned word template that are identified are all considered as identified banned words. Once banned words are detected, a replacement operation is performed to replace these banned words with appropriate vocabulary to ensure the compliance and security of the fifth document. The fifth document is then used as the second document for semantic analysis, and then S15 is executed.

[0111] In another embodiment, the specific implementation process of identifying banned words in the second copy is as follows:

[0112] ① Construct a multi-dimensional banned word library, which includes a basic banned word list, a variant word list, and a context-sensitive word list. Among them, the variant word list is generated by performing homophonic replacement, similar character variation, and semantic equivalent expansion on basic banned words through a pre-trained language model.

[0113] Building a multi-dimensional banned word library requires the comprehensive application of natural language processing technology. First, basic banned words are collected to form a basic banned word list. These words are mostly derived from laws, regulations, industry standards, etc. Then, with the help of a preset large language model (such as BERT), variants of the basic banned words are generated. Specifically:

[0114] 1) Homophone replacement: Leveraging the language model's dictionary and pronunciation knowledge, some characters in banned words are replaced with homophones. For example, "sell medicine" might be replaced with "sell yao." The model must be able to recognize the pronunciation of characters, based on its learning of pronunciation patterns in a large amount of text during training.

[0115] 2) Similar Character Mutation: Based on the similarity of character shapes and leveraging the model's understanding of font features (which relies on learning text patterns related to character shapes), banned words can be mutated. For example, "药" (medicine) can be changed to "荮" (荮). However, note that not all similar character substitutions are idiomatic, and the model must consider the context to determine if they are appropriate.

[0116] 3) Semantic Equivalence Extension: Based on semantic understanding, the model can expand banned words by mining synonyms and antonyms. For example, "selling banned items" can be expanded to "selling banned items." This requires the model to accurately grasp the semantic core of words and phrases.

[0117] The construction of a context-sensitive vocabulary list requires combining context analysis and using the context understanding capabilities of the language model to identify words that only have prohibited meanings in specific scenarios. For example, "special services" may imply illegal behavior in certain contexts.

[0118] Finally, the basic forbidden word list, variant word list, and context-sensitive word list are integrated into a multi-dimensional forbidden word library, which is continuously updated and optimized to cope with language evolution and the emergence of new forbidden words. This process requires a combination of technology and manual review to ensure the accuracy and effectiveness of the forbidden word library.

[0119] ② Use a bidirectional attention mechanism to perform hierarchical forbidden word recognition on the second text. First, match the basic forbidden word list using regular expressions, then detect the variant word list using semantic similarity calculation, and finally identify context-sensitive words based on dependency syntax analysis.

[0120] Use regular expressions to perform a preliminary scan of the second text to quickly locate words that exactly match the basic forbidden word list. This step uses the pattern matching capabilities of regular expressions to efficiently identify the basic forbidden words present in the second text. With the help of a pre-set large language model, convert the words and phrases in the second text into semantic vectors. By calculating the semantic similarity of these vectors with the words in the variant word list, detect semantically equivalent or similar variant words. Set a similarity threshold (such as 0.8), and when the similarity exceeds the threshold, determine it as a variant word. Perform dependency syntax analysis on the second text to construct a dependency relationship tree for the sentence, and analyze the grammatical and semantic relationships between words. Based on the dependency relationship and context, identify words that only have forbidden meaning in a specific context. This step requires contextual understanding to determine the actual meaning and purpose of the words.

[0121] ③ Implement hierarchical processing strategies for the identified forbidden words: high-frequency core forbidden words are directly deleted, low-frequency variant words are replaced based on word vector space similarity matching pre-set replacement words, and context-sensitive words are replaced by generating semantically equivalent replacement phrases using a generative adversarial network. Specifically:

[0122] First, according to the pre-set forbidden word library, identify high-frequency core forbidden words. These words have a high frequency of occurrence in the forbidden word library. Once these words are identified, they are directly deleted from the text. This step is simple and direct, ensuring that these high-risk words are not spread. For low-frequency variant words, use word vector technology (such as Word2Vec or GloVe) to map words to a high-dimensional space. Calculate the word vector similarity between the variant word and the pre-set replacement word, and select the replacement word with the highest similarity for replacement. Ensure that the replacement word is semantically close to the original word while avoiding forbidden content. For context-sensitive words, use the generator in the generative adversarial network (GAN) to generate semantically equivalent replacement phrases.

[0123] ④ Perform consistency checking on the processed text using a contrastive learning model to verify the semantic similarity before and after replacement. If it is lower than the pre-set threshold, trigger the manual review interface.

[0124] The processed text is compared with the second text to identify the replaced prohibited words and their replacement words or phrases. A contrast learning model is used to encode the text segments before and after replacement, converting them into fixed-dimensional semantic vectors. The contrast learning model, trained on large-scale data, can learn the deep semantic representation of the processed text. The similarity between the semantic vectors of the text segments before and after replacement is calculated. Common methods include cosine similarity, Euclidean distance, etc. Set a similarity threshold (such as 0.9), if the semantic similarity before and after replacement is higher than the threshold, it is considered that the replaced text maintains consistency with the original text in semantics; otherwise, it is considered that the replacement may affect the original meaning of the second script, and the replacement strategy needs to be further adjusted.

[0125] ⑤Repairing word boundaries and optimizing syntactic structure for the verified script, outputting the fifth script that meets regulatory requirements, specifically:

[0126] a. Use regular expressions or specialized word segmentation tools to identify abnormal boundaries in the text caused by prohibited word replacement (such as missing spaces, punctuation errors, etc.). Repair strategy: use the word boundary patterns learned by the pre-set large language model on large-scale text to automatically repair or manually correct, ensuring that English, Chinese, and numerical elements meet standard language habits.

[0127] b. Use dependency syntax analysis and other techniques to identify sentence component misplacement (such as subject-verb inconsistency) and improper collocation caused by replacement. Optimization adjustment: based on syntactic analysis results, language model-generated optimization suggestions, or manual intervention, adjust sentence structure to make it more natural and fluent while preserving the semantic integrity of the original script.

[0128] It should be noted that when no prohibited words are identified from the second script, S15 is executed.

[0129] S15, generating a title for the second script, specifically:

[0130] Combine the second script with the pre-set title generation template to form a prompt instance. Input this embodiment into the pre-set large language model, use the pre-set large language model to perform semantic analysis on the content of the second script, and generate multiple candidate titles through deep understanding of the second script. These candidate titles not only accurately reflect the core content of the article, but also have enough appeal for users to choose from. The user determines the title of the second script from these candidate titles, or uses the pre-set large language model to perform semantic analysis on the content of the second script to generate a candidate title, which is directly used as the title of the second script.

[0131] S16, combine the second script and the title of the second script to generate a target script of the preset commodity, provide the target script to the user, so that the user can evaluate the content of the target script in detail and determine whether to be satisfied, if yes, take the target script as the final script, so that the user uses it in the target social media platform, if not, explain the feedback information of the user, the feedback information of the user includes: the content marked in the target script that needs to be modified, the reason description of dissatisfaction and the expected improvement direction. Combine the feedback information of the user with the preset note modification template to form a prompt instance. Input the prompt instance into the preset large language model, so that the preset large language model can deeply understand the feedback information of the user and make targeted rewriting and optimization to obtain the final script, thereby ensuring that the final script is more in line with the expectations and needs of the user.

[0132] The present application has the following beneficial effects:

[0133] 1) Construct an efficient, controllable and intelligent script generation workflow, decompose the script generation task into multiple ordered and refined stages, and combine advanced large language models (including multi-modal capabilities) and structured prompt engineering to automatically generate high-quality, platform-specific, compliant and flexible user-optimized social media script content. That is, the present application establishes a multi-stage, refined and structured content generation, verification and optimization process to ensure the factual accuracy, logical consistency, style adaptability and effective communication of core advantages of the generated script.

[0134] Deep fusion analysis of user input information and multi-modal image data of the commodity is realized to provide rich and accurate context information for script generation. A dynamic, context-based compliance review and replacement mechanism is introduced to reduce the risk of generated content and ensure compliance with platform and regulatory requirements.

[0135] An efficient and flexible human-computer collaborative content iteration optimization mode is constructed, which enables users to control and accurately modify the automatically generated content with fine granularity, improving the personalization and user satisfaction of the final content, significantly improving the efficiency and quality of social media content production, reducing labor costs, and enhancing the compliance and dissemination effect of the content.

[0136] Users only need to provide basic information and images to automatically generate notes that conform to the style of social media platforms, greatly reducing the user's editing workload. The automatic content generation and optimization process reduces the time and cost of manual editing, improves the efficiency of marketing, and through the user feedback mechanism, the system can optimize the content according to the specific needs and preferences of the user, improving user satisfaction.

[0137] 2) The invention develops a set of automated copywriting workflow for specific social media platforms (such as Xiaohongshu), which can be realized through Python scripts, Dify development platform and preset large language models. The invention realizes high customization and standardization by decomposing the copywriting task into multiple ordered and interdependent stages (such as information analysis, content generation, verification, adaptation, compliance review and title generation). Each stage uses a carefully designed structured prompt template, combined with domain knowledge and task objectives, to guide the preset large language model to generate, analyze and optimize. This cascading processing mode combined with specific prompt engineering improves the efficiency, controllability and quality of copywriting.

[0138] The invention realizes the deep fusion analysis of user input text and product image by using multi-modal large language model. The system intelligently analyzes the product picture, identifies and extracts key visual features (such as color, style, material, details, etc.), and integrates them with user text description and product attributes to generate multi-dimensional product context representation. This integrated data is used for copywriting and optimization, improving the accuracy and vividness of content in describing product features.

[0139] The invention introduces a unique multi-stage, fine-grained, structured content optimization and verification process, which includes: ① Basic content and demand compliance verification to ensure the factual accuracy, information consistency, core selling point transmission effect and basic style compliance of the copy; ② Platform-specific element adaptation and enhancement, through intelligent identification and adjustment of platform-specific elements such as topic tags, user labels, guided interactive language and emoji usage specifications, to make the copy more consistent with the platform ecology and communication habits; ③ Dynamic compliance review and replacement based on context, using the semantic understanding ability of large language model to deeply scan the content to identify potential prohibited words or risky expressions, and perform context-aware intelligent replacement. In addition, the system also checks the reference style and structure accuracy of the imitation process. This multiple filtering and enhancement mechanism together constitutes a highly robust copy quality control closed loop.

[0140] The invention has a highly realistic reference note imitation capability. In the imitation generation stage, the LLM is guided by specific prompt templates to deeply deconstruct and learn the internal style parameters of the reference note, including but not limited to: sentence structure characteristics, word selection tendency, tone emotion, narrative rhythm, etc. More innovatively, in the imitation verification stage, the system performs detailed structured comparison analysis between the generated imitation copy and the original reference note (for example, based on sentence length distribution, key rhetorical structure, paragraph organization method, etc.). If significant deviations from the reference style or structure are detected, the system will use the LLM to adjust and reshape the imitation content, ensuring that while integrating new product information, the style and structure of the reference note are maximally reproduced or even surpassed.

[0141] The present application provides an efficient and flexible human-computer collaborative content iteration optimization mode. After the LLM generates the preliminary or optimized copy and title, the user can enter the interactive interface to mark the unsatisfactory places at the "atomic level" granularity (i.e., for specific words, sentences, paragraphs, or even punctuation marks), and provide modification reasons, expected effects, or specific instructions in natural language form. The system combines these structured user feedback with preset modification templates, uses the LLM to deeply understand the user's modification intentions, and performs accurate and targeted rewriting and optimization based on the original copy context. This mechanism greatly improves the user's control over the final content, achieving personalized customization and iterative improvement of the copy. In the above embodiments, although the steps are numbered S1, S2, etc., it is only a specific embodiment given by the present application, and those skilled in the art can adjust the execution order of S1, S2, etc. according to the actual situation, which is within the protection scope of the present application. It can be understood that in some embodiments, some or all of the above embodiments can be included.

[0142] As shown in Figure 2 The social media copy intelligent generation system 200 based on a large language model according to an embodiment of the present application includes a first copy generation module 201, a second copy generation module 202, a title generation module 203, and a target copy generation module 204.

[0143] The first copy generation module 201 is configured to generate a first copy of a preset product based on associated information and images of the preset product and using a preset large language model, wherein the first copy meets the basic requirements of the copy of a target social media platform.

[0144] The second copy generation module 202 is configured to add preset elements of the target social media platform to the first copy to obtain a second copy.

[0145] The title generation module 203 is configured to perform semantic analysis on the second copy to generate a title of the second copy.

[0146] The target copy generation module 204 is configured to combine the second copy and the title of the second copy to generate a target copy of the preset product.

[0147] Optionally, in the above technical solution, the first copy generation module 201 is specifically configured to:

[0148] When a reference copy of the preset product is received, the preset large language model is used to identify the writing detail information of the reference copy, and a third copy of the preset product is generated based on the writing detail information of the reference copy and the associated information and images of the preset product, and the third copy is used as the first copy.

[0149] When the reference copy of the preset product is not received, a fourth copy of the preset product is generated based on the associated information and image of the preset product using the preset large language model, and the fourth copy is used as the first copy.

[0150] Furthermore, it also includes a first evaluation module and a second evaluation module;

[0151] The first evaluation module is configured to: before using the third copy as the first copy, evaluate the correctness, tone, and wording of the third copy to obtain a first evaluation result, and compare and analyze the sentence structure of the third copy with the reference copy to obtain a comparative analysis result; determine whether to optimize the third copy based on the first evaluation result and the comparative analysis result; if so, optimize the third copy based on the first evaluation result and / or the comparative analysis result. The first copy generation module 201 is further configured to: use the optimized third copy as the first copy;

[0152] The second evaluation module is used to: before using the fourth copy as the first copy, evaluate the correctness, tone and wording of the fourth copy to obtain a second evaluation result; based on the second evaluation result, determine whether to optimize the fourth copy; if so, optimize the fourth copy based on the second evaluation result, and the first copy generation module 201 is further specifically used to: use the optimized fourth copy as the first copy.

[0153] Furthermore, it also includes a banned word identification and processing module, which is used to: identify banned words in the second copy before performing semantic analysis on the second copy; when banned words are identified, process the identified banned words to obtain a fifth copy, and use the fifth copy as the second copy for semantic analysis.

[0154] Furthermore, it also includes a target copy optimization module, which is used to: receive user feedback information on the target copy; and optimize the target copy according to the feedback information.

[0155] It should be noted that the beneficial effects of the social media copywriting intelligent generation system based on a large language model provided by the above embodiment are the same as the beneficial effects of the social media copywriting intelligent generation method based on a large language model, which will not be repeated here. In addition, when the system provided by the above embodiment realizes its functions, it only uses the division of the above functional modules as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the system can be divided into different functional modules according to actual conditions to complete all or part of the functions described above. In addition, the system and method embodiments provided by the above embodiment belong to the same concept. The specific implementation process is detailed in the method embodiment, which will not be repeated here.

[0156] In the present application, the social media script intelligent generation system based on large language model can be a computer program (including program code) running in a computer device. For example, the social media script intelligent generation system based on large language model is an application software, which can be used to execute the corresponding steps in the social media script intelligent generation method based on large language model.

[0157] In some embodiments, the social media script intelligent generation system based on large language model can be implemented in a combination of software and hardware. For example, the social media script intelligent generation system based on large language model can be a hardware decoding processor programmed to execute the social media script intelligent generation method based on large language model. For example, the hardware decoding processor can be one or more application specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field programmable gate arrays (FPGAs), or other electronic elements.

[0158] In the present application, the modules described in the embodiments can be implemented by software or hardware. In some cases, the name of the module does not limit the module itself.

[0159] The electronic device of the present application can include a processor and a memory. The memory is used to store a computer program. The processor is used to execute the social media script intelligent generation method based on large language model by calling the computer program.

[0160] In an optional embodiment, an electronic device is provided, as shown in Figure 3 Figure 3 ​The electronic device 4000 shown includes a processor 4001 and a memory 4003. The processor 4001 and the memory 4003 are connected, for example, via a bus 4002. Optionally, the electronic device 4000 can also include a transceiver 4004, which can be used for data interaction, such as data transmission and / or data reception, between the electronic device and other electronic devices. It should be noted that the transceiver 4004 is not limited to one in actual applications, and the structure of the electronic device 4000 does not constitute a limitation on the embodiments of the present application.

[0161] The processor 4001 can be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute various exemplary logical blocks, modules and circuits described in connection with the present disclosure. The processor 4001 can also be a combination of computing functions, such as one or more microprocessor combinations, combinations of DSP and microprocessor, etc.

[0162] The bus 4002 can include a path for transmitting information between the above-mentioned components. The bus 4002 can be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The bus 4002 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 3 Only one bus 4002 is represented by a thick line, but it does not mean that there is only one bus or only one type of bus.

[0163] The memory 4003 can be a ROM (Read Only Memory) or other type of static storage device that can store static information and instructions, a RAM (Random Access Memory) or other type of dynamic storage device that can store information and instructions, an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory) or other optical disk storage, a magnetic disk storage or other magnetic storage devices, or any other medium capable of storing desired program code in the form of instructions or data structures and that can be accessed by a computer, but is not limited thereto.

[0164] The memory 4003 is used to store application program codes (computer programs) for implementing the solutions of the present application, and is controlled by the processor 4001 to perform. The processor 4001 is used to execute the application program codes stored in the memory 4003 to realize the contents shown in the foregoing method embodiments.

[0165] The electronic device can also be a terminal device, and the terminal device can be any device that can install an application, including at least one of a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, a smart television, and a smart vehicle device.

[0166] It should be noted that, Figure 3 The electronic device shown is only an example and should not limit the functions and use range of the embodiments of the present application.

[0167] The computer readable storage medium of the embodiments of the present application, the computer readable storage medium stores a computer program, and the computer program is executed by the processor to realize any one of the above-mentioned social media script intelligent generation methods based on a large language model.

[0168] Optionally, the computer readable storage medium can be a read-only memory (Read-Only Memory, ROM), a random access memory (Random Access Memory, RAM), a compact disc read-only memory (Compact Disc Read-Only Memory, CD-ROM), a magnetic tape, a floppy disk, and an optical data storage device, etc.

[0169] In an example embodiment, a computer program product or computer program is also provided, the computer program product or computer program comprising computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to cause the electronic device to perform any of the above-described methods for intelligent generation of social media copy based on a large language model.

[0170] Computer program code for carrying out operations of the present application can be written in any combination of one or more programming languages, including an object oriented programming language such as Java, Smalltalk, C++ or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider).

[0171] It should be understood that the flowchart and block diagrams in the drawings illustrate the architecture, functionality, and operation of possible implementations of various embodiments of the present application. In this regard, each block in the flowchart and block diagrams can represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions noted in the flowchart or block diagrams can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks can sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flowchart illustrations, and combinations thereof, can be implemented by a dedicated hardware-based system that performs the specified functions or operations, or combinations of hardware and software.

[0172] The computer readable storage medium provided by the embodiments of the present application can be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or apparatus, or any combination of the above. More specific examples of the computer readable storage medium can include, but are not limited to, an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present application, the computer readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus or device.

[0173] The computer readable storage medium described above carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the methods shown in the embodiments described above.

[0174] The above description is merely the preferred embodiments of the present application and the explanation of the technical principles used. It should be understood by those skilled in the art that the disclosed scope of the present application is not limited to the technical solutions formed by the specific combinations of the technical features described above, and should also cover other technical solutions formed by any combinations of the technical features described above or their equivalent features without departing from the disclosed concept. For example, the technical solutions formed by replacing the above features with the technical features disclosed in the present application (but not limited to) having similar functions.

[0175] It should be noted that the terms "first", "second" and the like in the specification and claims of the present application are used to distinguish similar objects, and represent no specific order or sequence. The order of use of similar objects can be interchanged in appropriate cases, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described.

[0176] Those skilled in the art know that the present application can be implemented as a system, a method or a computer program product, so the present application can be specifically implemented as follows: it can be a complete hardware, a complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, which is generally referred to as "circuit", "module" or "system" in this paper. In addition, in some embodiments, the present application can also be implemented as a computer program product in one or more computer readable media, which contains computer readable program code.

[0177] Although the embodiments of the present application have been shown and described above, it is understood that the above-described embodiments are exemplary and are not to be construed as limiting the present application, and that variations, modifications, substitutions and changes can be made by those skilled in the art without departing from the scope of the present application.

Claims

1. A method for intelligently generating social media copy based on a large language model, characterized in that: include: Based on the associated information and image of the preset product and using a preset large language model, generating a first copy for the preset product, wherein the first copy meets the basic copy requirements of the target social media platform; Adding preset elements of the target social media platform to the first copy to obtain a second copy; Performing semantic analysis on the second copy to generate a title for the second copy; The second copy and the title of the second copy are combined to generate a target copy for the preset product.

2. The method for intelligently generating social media copy based on a large language model according to claim 1, characterized in that: Based on the associated information and image of the preset product and using the preset large language model, a first copy for the preset product is generated, including: When receiving the reference copy of the preset product, using the preset large language model to identify the writing details of the reference copy, and generating a third copy of the preset product based on the writing details of the reference copy and the associated information and image of the preset product, and using the third copy as the first copy; When the reference copy of the preset product is not received, a fourth copy of the preset product is generated based on the associated information and image of the preset product using the preset large language model, and the fourth copy is used as the first copy.

3. The method for intelligently generating social media copy based on a large language model according to claim 2, characterized in that: Before using the third copy as the first copy, the method further includes: evaluating the correctness, intonation, and wording of the third copy to obtain a first evaluation result, and performing a comparative analysis on the sentence structure of the third copy and the reference copy to obtain a comparative analysis result; determining whether to optimize the third copy based on the first evaluation result and the comparative analysis result; and if so, optimizing the third copy based on the first evaluation result and / or the comparative analysis result, and using the third copy as the first copy, including: using the optimized third copy as the first copy; Before using the fourth copy as the first copy, the method further includes: evaluating the correctness, tone and wording of the fourth copy to obtain a second evaluation result; determining whether to optimize the fourth copy based on the second evaluation result, and if so, optimizing the fourth copy based on the second evaluation result, and using the fourth copy as the first copy, including: using the optimized fourth copy as the first copy.

4. The method for intelligently generating social media copy based on a large language model according to any one of claims 1 to 3, characterized in that: Before performing semantic analysis on the second copy, the method further includes: The second text is subjected to banned word recognition. When banned words are recognized, the recognized banned words are processed to obtain a fifth text, and the fifth text is used as the second text for semantic analysis.

5. The method for intelligently generating social media copy based on a large language model according to any one of claims 1 to 3, characterized in that: Also includes: receiving user feedback on the target copy; The target copy is optimized according to the feedback information.

6. A social media copywriting intelligent generation system based on a large language model, characterized by: It includes a first copy generation module, a second copy generation module, a title generation module and a target copy generation module; The first copywriting generation module is configured to generate a first copywriting for the preset product based on the associated information and image of the preset product and using a preset large language model, wherein the first copywriting meets the basic copywriting requirements of the target social media platform; The second copy generation module is configured to add preset elements of the target social media platform to the first copy to obtain a second copy; The title generation module is used to: perform semantic analysis on the second copy to generate a title for the second copy; The target copy generation module is used to combine the second copy and the title of the second copy to generate the target copy of the preset product.

7. The social media copywriting intelligent generation system based on a large language model according to claim 6 is characterized in that: The first copywriting generation module is specifically used to: When receiving the reference copy of the preset product, using the preset large language model to identify the writing details of the reference copy, and generating a third copy of the preset product based on the writing details of the reference copy and the associated information and image of the preset product, and using the third copy as the first copy; When the reference copy of the preset product is not received, a fourth copy of the preset product is generated based on the associated information and image of the preset product using the preset large language model, and the fourth copy is used as the first copy.

8. The social media copywriting intelligent generation system based on a large language model according to claim 7 is characterized in that: Also includes a first evaluation module and a second evaluation module; The first evaluation module is configured to: before using the third copy as the first copy, evaluate the correctness, intonation, and wording of the third copy to obtain a first evaluation result, and compare and analyze the sentence structure of the third copy with that of the reference copy to obtain a comparative analysis result; determine whether to optimize the third copy based on the first evaluation result and the comparative analysis result; if so, optimize the third copy based on the first evaluation result and / or the comparative analysis result, and then the first copy generation module is further configured to: use the optimized third copy as the first copy; The second evaluation module is used to: evaluate the correctness, tone and wording of the fourth copy before using it as the first copy to obtain a second evaluation result; determine whether to optimize the fourth copy based on the second evaluation result; if so, optimize the fourth copy based on the second evaluation result, and the first copy generation module is further specifically used to: use the optimized fourth copy as the first copy.

9. An electronic device, characterized in that: The invention comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for intelligently generating social media copy based on a large language model as described in any one of claims 1 to 5 is implemented.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, it implements the method for intelligently generating social media copy based on a large language model as described in any one of claims 1 to 5.

Images