Enterprise brand multi-modal marketing content intelligent generation method based on large language model

Abstract

The application discloses an enterprise brand multi-modal marketing content intelligent generation method based on a large language model, relates to the technical field of multi-modal processing, and is characterized in that: a large language model is obtained based on historical marketing data, multi-modal marketing data semantics related to an enterprise brand are analyzed, brand tone vectors of various modal types are determined, a plurality of marketing theme baselines are determined, brand tone vector semantics is embedded into different word distance positions of the marketing theme baselines, a plurality of marketing events are generated, a public recommendation window for storing all marketing events is established, style anchor points are associated and set, the style anchor points are connected to a marketing demand pool, corresponding style anchor points are adapted to form an anchor point set according to customer demand, and marketing content meeting customer demand is generated based on the anchor point set. Through the mapping of the brand tone vector and the marketing theme baseline, and the adaptive retrieval of the style anchor point and the customer demand, the consistency of the marketing content, the brand positioning and the customer demand is improved, and accurate generation of the marketing content is realized.

Description

Technical Field

[0001] This invention relates to the field of multimodal processing technology, specifically to a method for intelligent generation of multimodal marketing content for enterprise brands based on a large language model. Background Technology

[0002] The generation of corporate marketing content mainly relies on manual writing, which has problems such as long cycle time, high cost, and difficulty in ensuring style consistency. Although existing AIGC tools can assist in generating text or image content, they have the following shortcomings: The lack of a precise understanding and quantitative representation of the corporate brand tone leads to content style drift, making it difficult to maintain consistency with brand positioning. There is a lack of a unified brand style constraint mechanism in the multimodal content generation process, resulting in style fragmentation between different modal content. The marketing needs of different customer groups vary greatly, and existing solutions lack a flexible mechanism to effectively match brand characteristics with customer needs, resulting in insufficient targeting of generated content.

[0003] In addition, existing tools typically only support single-modal content generation, making it difficult to meet the integration needs of multimodal marketing scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent method for generating multimodal marketing content for enterprise brands based on a large language model, in order to address the shortcomings in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent generation of multimodal marketing content for enterprise brands based on a large language model, comprising the following steps: Step S1: Perform modeling based on historical marketing data to obtain a large language model that represents all marketing methods. Perform semantic analysis on all marketing data related to the enterprise brand through the large language model to determine the brand tone vectors of multiple modal types. Step S2: Determine multiple marketing theme baselines, embed all brand tone vectors into different word distance positions of the marketing theme baselines to obtain multiple marketing events, establish a public recommendation window to store all marketing events, and set the style anchor points of the public recommendation window accordingly. Step S3: Establish a marketing demand pool, connect all style anchors to the marketing demand pool, allow customers to put forward their own demands, and adapt the corresponding style anchors to form an anchor set, and retrieve and generate corresponding marketing content that meets their needs based on the anchor set.

[0006] In a preferred embodiment, the step of determining the brand tone vectors for multiple modal types includes: Identify multiple marketing themes in marketing activities based on historical marketing data, analyze the marketing data corresponding to the marketing themes in marketing activities, use them as marketing tools, perform convolution and pooling processing on the marketing tools, determine the corresponding model layers for each, and integrate the model layers to obtain a large language model; By adapting multiple modal types using a large language model, the model analysis domain corresponding to each modal type is determined, and marketing data related to the enterprise brand from each model analysis domain is integrated to determine the corresponding model analysis layer for each model analysis domain. The model analysis layer is enabled to perform multidimensional semantic analysis on marketing data. The analysis results are then vectorized to obtain brand tone vectors that match the marketing theme of the enterprise brand. The brand tone vectors are grouped according to modality type.

[0007] In a preferred embodiment, the step of determining the corresponding model analysis layer for each model analysis domain includes: Obtain all modalities involved in the enterprise's brand marketing, including text modalities, image modalities, and audio modalities; Using the underlying shared parameters of the large language model as a common base, a corresponding modality encoder is instantiated for each modality type. Each modality encoder receives the original marketing data of the corresponding modality type and extracts modality-specific features in its own feature space. The extracted modality-specific features are mapped to the unified semantic space of the large language model. Combined with the enterprise brand identity information, the boundary of the model analysis domain corresponding to each modality type is defined in the unified semantic space, and an independent model analysis layer is assigned to each model analysis domain.

[0008] In a preferred embodiment, the step of grouping the brand tone vector based on modal type includes: Enable the model analysis layer corresponding to each modality type to perform multidimensional semantic analysis on the input marketing data of each modality, including sentiment polarity analysis, style feature analysis and rhetorical device analysis; The semantic analysis results of each dimension are mapped to numerical vector components, and the numerical vector components are combined to obtain the original semantic vector of the corresponding modality type. For each modality type, the original semantic vector is oriented and normalized to generate a brand tone vector that matches the marketing theme of the enterprise brand. Brand tone vectors belonging to the same modality type are grouped into the same vector group, and each vector group is labeled with a modality label.

[0009] In a preferred embodiment, the step of setting the style anchor point of the associated public recommendation window includes: Identify multiple marketing theme baselines involved in corporate brand marketing. These marketing theme baselines are semantic baseline vector sequences covering different marketing narrative frameworks. The brand tone vectors under each modality group are semantically embedded into the preset word distance positions of each marketing theme baseline according to the modality correspondence. During the embedding process, the brand tone vectors in adjacent positions are smoothed through an attention mechanism to generate multiple marketing events carrying modality fusion information. All marketing events are stored in a public recommendation window according to their generation sequence or theme similarity. For marketing events located at the cluster center in the public recommendation window, the corresponding key semantic features are extracted and associated as style anchors for the corresponding public recommendation window.

[0010] In a preferred embodiment, the step of generating multiple marketing events carrying modal fusion information includes: Extract multiple pre-defined word spacing slots from the baseline of each marketing theme, with each word spacing slot associated with a location code and a modality type label; Based on the modal grouping tags to which each brand tone vector belongs, the brand tone vector is matched to word distance slots with tags of the same modal type, and the initial embedding is completed by vector summation. For the adjacent word distance position vectors after the initial embedding is completed, a multi-head self-attention mechanism is used to calculate the semantic association weight between each position. Based on the association weight, a weighted smoothing process is performed on the adjacent position vectors to obtain a marketing event representation that integrates contextual semantics.

[0011] In a preferred embodiment, the step of extracting the corresponding key semantic features and associating them as style anchors for the corresponding public recommendation window includes: Using the first entry position of the public recommendation window as the starting point of the sliding window, subsequent marketing events are sequentially merged into the current sliding window according to the preset step size until the theme similarity of the marketing events in the sliding window meets the clustering condition to form an event cluster. Determine the marketing event corresponding to the cluster center of the event cluster, extract the principal components of the semantic vector of the corresponding marketing event, compress the principal components of the semantic vector into a low-dimensional feature vector, and use the low-dimensional feature vector as the style anchor point corresponding to the current event cluster. Set corresponding style anchors for all event clusters within the public recommendation window, and establish a two-way index association between each style anchor and the marketing events within its event cluster.

[0012] In a preferred embodiment, the step of generating corresponding marketing content that meets the requirements based on the anchor set retrieval includes: Establish a marketing demand pool, which is a database that stores the mapping relationship between customer marketing demand data and style anchors to be matched; All style anchors are connected to the marketing demand pool according to their respective marketing theme baseline and event cluster affiliation, and a demand matching interface is configured for each style anchor. Obtain the customer's input requirement description text, perform semantic parsing on the requirement description text, extract the requirement intent vector, calculate the similarity between the requirement intent vector and the feature vector of each style anchor point, sort according to similarity, and adapt style anchor points with similarity exceeding a preset threshold to form an anchor point set. Using marketing events corresponding to each style anchor in the anchor set as search criteria, the system retrieves corresponding marketing content fragments from the public recommendation window, splices each marketing content fragment according to the narrative logic in the requirement description text, automatically adjusts the content structure based on preset scene tags, and makes fine-tuning of the content in the annotation and review interface to generate marketing content that meets the client's needs.

[0013] In a preferred embodiment, the step of configuring a requirement matching interface for each style anchor point includes: Iterate through all event clusters in the public recommendation window, and extract the corresponding style anchor and the baseline label of the marketing theme baseline where the corresponding style anchor is located for each event cluster; Organize each style anchor point and its corresponding baseline label, event cluster identifier, and modal grouping label into a demand matching index tuple, and write all demand matching index tuples into the demand matching index table of the marketing demand pool. Configure a corresponding demand matching interface for each style anchor point to receive the demand intent vector from external input and output the matching score between the current style anchor point and the current customer demand.

[0014] In a preferred embodiment, the step of generating marketing content that meets customer needs includes: Using marketing events of each style anchor in the anchor set as search criteria, the corresponding marketing content fragments are retrieved from the public recommendation window, the narrative logic sequence is extracted, the splicing order of each marketing content fragment is determined according to the narrative logic sequence, and preliminary spliced ​​content is generated. Obtain scene tags associated with customer needs, match the corresponding content structure template, and adaptively adjust the arrangement and fusion weight of each modal fragment in the initial spliced ​​content according to the content structure template to obtain structure-adapted content; Input the structural adaptation content into the annotation and review interface. In response to the reviewer's fine-tuning instructions for any modal segment, recall semantically similar alternative segments from the public recommendation window and replace them. Once all marketing content fragments are confirmed, multimodal marketing content is rendered and output, and the generated content, along with its corresponding style anchors, narrative logic identifiers, and scene tags, is stored in the marketing demand pool.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention uses a large language model to perform semantic analysis on multimodal historical marketing data, obtains brand tone vectors for each modality, transforms the abstract style characteristics of corporate brands into quantifiable vector representations, and provides accurate brand semantic input for subsequent content generation.

[0016] 2. This invention generates marketing events by embedding the brand tone vector semantics into the word distance position of the marketing theme baseline, and stores all marketing events in a unified public recommendation window and sets style anchor points accordingly. This establishes a structured relationship between brand tone and marketing events, ensuring that the brand style maintains a stable mapping during the content generation process.

[0017] 3. This invention connects style anchors to a marketing demand pool and adapts them to customer needs to form an anchor set. The anchor set is then used to retrieve corresponding marketing events to generate marketing content. This ensures that the generated results are simultaneously constrained by brand tone and guided by customer needs, thereby improving the consistency between marketing content and brand positioning and customer needs. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0019] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1, please refer to Figure 1 As shown in this embodiment, the intelligent generation method for enterprise brand multimodal marketing content based on a large language model includes the following steps: Step S1: Perform modeling based on historical marketing data to obtain a large language model that represents all marketing methods. Perform semantic analysis on all marketing data related to the enterprise brand through the large language model to determine the brand tone vectors of multiple modal types. Step S2: Determine multiple marketing theme baselines, embed all brand tone vectors into different word distance positions of the marketing theme baselines to obtain multiple marketing events, establish a public recommendation window to store all marketing events, and set the style anchor points of the public recommendation window accordingly. Step S3: Establish a marketing demand pool, connect all style anchors to the marketing demand pool, allow customers to put forward their own demands, and adapt the corresponding style anchors to form an anchor set, and retrieve and generate corresponding marketing content that meets their needs based on the anchor set.

[0022] As described in steps S1 to S3 above, semantic analysis of multimodal historical marketing data is performed using a large language model to obtain brand tone vectors for each modality. This transforms the abstract style characteristics of the brand into quantifiable vector representations, providing precise brand semantic input for subsequent content generation. The brand tone vectors are embedded into the word distance positions of the marketing theme baseline to generate marketing events. All marketing events are uniformly stored in a public recommendation window and associated with style anchors, establishing a structured relationship between brand tone and marketing events. This ensures that the brand style maintains a stable mapping during content generation. By connecting style anchors to the marketing demand pool and adapting them to customer needs to form an anchor set, the corresponding marketing events are retrieved using the anchor set, generating marketing content that meets customer needs. This ensures that the generated results are simultaneously constrained by brand tone and guided by customer needs, improving the consistency between the generated marketing content and the needs.

[0023] It should be further explained that, in the specific implementation process, the step S1, which involves performing modeling based on historical marketing data to obtain a large language model representing all marketing methods, and then performing semantic analysis on all marketing data related to the enterprise brand using the large language model to determine the brand tone vectors of multiple modalities, includes: Step S11: Identify multiple marketing themes in marketing activities based on historical marketing data, analyze the marketing data corresponding to the marketing themes in marketing activities, use them as marketing tools, perform convolution and pooling processing on the marketing tools, determine the corresponding model layers for each, and integrate the model layers to obtain a large language model; Step S12: Adapt multiple modal types through a large language model, determine the model analysis domain corresponding to each modal type, integrate some marketing data related to the enterprise brand from each model analysis domain, and determine the corresponding model analysis layer for each model analysis domain; Step S13: Enable the model analysis layer to perform multidimensional semantic analysis on the marketing data, and vectorize the analysis results to obtain the brand tone vector that matches the marketing theme of the enterprise brand. The brand tone vector is grouped according to the modality type.

[0024] It should be noted that the marketing theme refers to the target audience and marketing channel type corresponding to the marketing behavior. Different marketing themes correspond to different marketing data distributions. The convolution processing uses a one-dimensional convolution kernel to perform sliding window calculation on the time sequence of marketing data to extract the local patterns of marketing events in the time dimension. The pooling processing uses max pooling to reduce the dimensionality of the convolution output and retain the salient features in the local patterns. The model layers corresponding to each modality are the Transformer encoding layer for text modality, the ResNet convolutional layer for image modality, and the temporal feature extraction layer for audio modality. Each model layer is trained independently according to its modality type and then integrated into a unified large language model through a parameter sharing mechanism. Modal types include text modality, image modality, and audio modality; The model analysis domain corresponding to each modality type is the parameter subspace within the large language model used to process specific modality data. Integrating the marketing data related to the enterprise brand from each model analysis domain means selecting marketing samples containing the corresponding enterprise brand logo from all marketing data, and inputting the selected marketing samples into their respective model analysis domains (i.e., the parameter subspaces of their respective specific modality data) according to modality type. The model analysis layer is an intermediate representation layer located between the shallow layer (infrastructure layer) and the deep layer (output representation layer) of the large language model. Each modality type corresponds to an independent model analysis layer, which is used to receive the intermediate feature representation of the input data of that modality type after shallow encoding. For multidimensional semantic analysis, sentiment polarity analysis is used to determine the sentiment tendency value of the text content in marketing data. The sentiment tendency value is a continuous value between negative one and positive one. Style feature analysis is used to extract the sentence length distribution, lexical richness and frequency of rhetorical devices used in marketing data, and obtain style feature values ​​after processing. Rhetorical device analysis is used to identify metaphor, parallelism and rhetorical question rhetorical markers in marketing data and generate corresponding rhetorical feature values. The specific acquisition of sentiment tendency value, style feature value, and rhetorical feature value is as follows: For sentiment inclination values, a multimodal sentiment dictionary for the marketing domain is constructed. This multimodal sentiment dictionary contains a set of positive sentiment words and a set of negative sentiment words. Each sentiment word is pre-labeled with a corresponding sentiment polarity intensity weight in the dictionary. The input marketing data is segmented, and each word in the segmentation results is traversed. Matching positive or negative words are searched in the sentiment dictionary. The sentiment polarity intensity weights corresponding to the matched words are accumulated according to their positive and negative signs to obtain positive and negative scores. The sum of the positive and negative scores is divided by the sum of their absolute values ​​to obtain the sentiment inclination value, which ranges from -1 to +1. A positive value indicates a positive inclination, a negative value indicates a negative inclination, and zero indicates neutrality. The larger the absolute value, the higher the sentiment intensity.

[0025] For style feature values, word segmentation and syntactic analysis are performed on the input marketing data, the total number of words and sentences in the text are counted, the average sentence length is calculated, the proportion of the number of unique words in the text to the total number of words is counted to obtain the lexical richness, and the average sentence length and lexical richness are weighted and summed according to the preset weight ratio to obtain the text complexity score, which is used as the style feature value. For the image modality, the style feature is the weighted sum of the standard deviation of the image's hue saturation distribution and the composition density value. The composition density value is obtained by detecting the proportion of the main area in the image to the total area of ​​the image. For the audio modality, the style feature is the weighted sum of the audio signal's speech rate and volume dynamic range. The speech rate is obtained by detecting the number of syllables per unit time, and the volume dynamic range is obtained by detecting the difference between the peak and trough values ​​of the audio signal.

[0026] For rhetorical feature values, a rhetorical tag library is pre-constructed, which includes a set of metaphor trigger words, a set of parallel sentence templates, and a set of rhetorical question templates. Syntactic analysis and sentence structure matching are performed on the input marketing data to detect whether the text contains metaphor trigger words and to match parallel sentence templates and rhetorical question templates. The proportion of the number of detected rhetorical tags to the total number of sentences is counted to obtain the numerical value corresponding to the rhetorical richness, which is used as the rhetorical feature value.

[0027] After converting the analysis results of each dimension into corresponding vector components, they are concatenated into the original semantic vector in the order of sentiment value, style feature value, and rhetorical feature value. Direction correction is performed by rotating the original semantic vector to be consistent with the core semantic direction of the corporate brand. Modulus normalization is performed by scaling the modulus of the corrected vector to a unit length to generate a brand tone vector. Brand tone vectors belonging to the same modality type are grouped into the same vector group. Each vector group is labeled with a text modality label, an image modality label, and an audio modality label based on its respective modality type.

[0028] It should be further explained that, in the specific implementation process, the step S12, which involves adapting multiple modalities through a large language model, determining the model analysis domain corresponding to each modality, integrating marketing data related to the enterprise brand from each model analysis domain, and determining the corresponding model analysis layer for each model analysis domain, includes: Step S121: Obtain all modal types involved in the enterprise's brand marketing, including text modal, image modal and audio modal; Step S122: Using the underlying shared parameters of the large language model as a common base, instantiate a corresponding modality encoder for each modality type. Each modality encoder receives the original marketing data of the corresponding modality type and extracts modality-specific features in its own feature space. Step S123: Map the extracted modality-specific features to the unified semantic space of the large language model, combine the enterprise brand identity information, delineate the model analysis domain boundary corresponding to each modality type in the unified semantic space, and assign an independent model analysis layer to each model analysis domain.

[0029] It should be noted that all modalities involved in corporate brand marketing are determined through a pre-set marketing data collection list. The marketing data collection list marks the modal type to which each marketing data belongs. The underlying shared parameters are the parameters of the general feature extraction layer at the bottom of the large language model. This general feature extraction layer performs the same initial encoding operation on the input data of all modal types. Each modal encoder performs deep feature extraction in its own independent feature space based on the initial encoding results output by the underlying shared parameters. For different modal encoders, the data extracted is different. Specifically, the text modal encoder extracts inter-word dependency features in the text sequence, the image modal encoder extracts spatial layout features in the image, and the audio modal encoder extracts temporal rhythm features and timbre features in the audio. In this context, the inter-word dependency features in the text sequence are the context semantic association weight matrix calculated through a self-attention mechanism; the spatial layout features in the image are the edge and texture distribution maps extracted by the convolution kernel; and the temporal rhythm features in the audio are the zero-crossing rate and short-time energy change curves of the audio signal. The unified semantic space is the vector space shared by each modal encoder after linear projection. The corporate brand identification information is the word vector corresponding to the corporate brand name. The geometric center is the position of the corporate brand identification information in the unified semantic space, and the radius is the distance threshold between the projected features of each modality and the geometric center. The model analysis domain boundary corresponding to each modality type is defined. Each model analysis layer shares the semantic representation within the boundary, and each analysis layer achieves differentiated analysis of the semantics within the boundary through independent trainable parameters.

[0030] It should be further explained that, in the specific implementation process, the step S13, which involves enabling the model analysis layer to perform multidimensional semantic analysis on marketing data, vectorizing the analysis results to obtain a brand tone vector matching the marketing theme of the enterprise brand, and grouping the brand tone vector based on modal type, includes: Step S131: Enable the model analysis layer corresponding to each modality type, and perform multidimensional semantic analysis on the input marketing data of each modality. Multidimensional semantic analysis includes sentiment polarity analysis, style feature analysis and rhetorical device analysis. Step S132: Map the semantic analysis results of each dimension into numerical vector components, and combine the numerical vector components to obtain the original semantic vector of the corresponding modality type; Step S133: Perform direction correction and modulus normalization on the original semantic vector of each modality type to generate a brand tone vector that matches the marketing theme of the enterprise brand. Group the brand tone vectors belonging to the same modality type into the same vector group and label each vector group with a modality label.

[0031] It should be noted that in step S133, the direction correction is to calculate the cosine value of the angle between the original semantic vector and the core semantic reference vector of the enterprise brand, construct a rotation matrix based on the cosine value of the angle, multiply the original semantic vector by the rotation matrix, so that the direction of the corrected vector is consistent with the direction of the core semantic reference vector of the enterprise brand, and the modulus normalization is to divide each component of the corrected vector by the Euclidean norm of the vector to obtain the brand tone vector of unit length.

[0032] It should be further explained that, in the specific implementation process, the step S2, which involves determining multiple marketing theme baselines, embedding the semantics corresponding to all brand tone vectors into different word distance positions of the marketing theme baselines to obtain multiple marketing events, establishing a public recommendation window to store all marketing events, and associating and setting the style anchor points of the public recommendation window, includes: Step S21: Determine multiple marketing theme baselines involved in the enterprise's brand marketing. The marketing theme baselines are semantic baseline vector sequences covering different marketing narrative frameworks. Step S22: The brand tone vectors under each modality group are semantically embedded into the preset word distance positions of each marketing theme baseline according to the modality correspondence. During the embedding process, the brand tone vectors in adjacent positions are smoothed through an attention mechanism to generate multiple marketing events carrying modality fusion information. Step S23: Store all marketing events in the public recommendation window according to their generation sequence or theme similarity. For marketing events located at the cluster center in the public recommendation window, extract the corresponding key semantic features and associate them as style anchors for the corresponding public recommendation window.

[0033] It should be noted that the marketing theme baseline is a sequence of semantic baseline vectors pre-set by the enterprise brand under different marketing narrative frameworks. All marketing theme baselines correspond to four narrative frameworks, including brand story narration, product function demonstration, user emotional resonance, and promotional campaign appeal. Each marketing theme baseline (corresponding to a semantic baseline vector sequence) consists of multiple semantic baseline vectors arranged at fixed intervals, and the dimensions of each semantic baseline vector are consistent with the dimensions of the brand tone vector. The interval between adjacent semantic reference vectors is the word distance position. The length of the word distance position is determined by the Euclidean distance between adjacent semantic reference vectors in the semantic space. Different word distance positions have a unique index number on the marketing theme baseline, which is used to identify the brand tone vector and its embedding position on the corresponding marketing theme baseline.

[0034] It should be further explained that, in the specific implementation process, the brand tone vectors under each modality group are semantically embedded into the preset word distance positions of each marketing theme baseline according to the modality correspondence. During the embedding process, the brand tone vectors at adjacent positions are smoothed through an attention mechanism to generate multiple marketing events carrying modality fusion information. Step S22 includes: Step S221: Extract multiple preset word spacing slots from each marketing theme baseline. Each word spacing slot is associated with a position code and a modality type label. Step S222: Based on the modal grouping tags to which each brand tone vector belongs, match the brand tone vector to the word distance slots with tags of the same modal type, and complete the initial embedding by vector summation; Step S223: For the adjacent word distance position vectors after the initial embedding is completed, a multi-head self-attention mechanism is used to calculate the semantic association weight between each position. Based on the association weight, a weighted smoothing process is performed on the adjacent position vectors to obtain a marketing event representation that integrates contextual semantics.

[0035] Among them, the word spacing slot is a vector space reserved at each word spacing position on the marketing theme baseline. The position code associated with each word spacing slot is a fixed position code vector generated based on the sine and cosine functions, which is used to record the relative order of the slot on the marketing theme baseline. Modality type tags are stored in the metadata field of word spacing slots. The tag value can be a text modality tag, an image modality tag, or an audio modality tag, which is used to limit the modality type that can be embedded in the brand tone vector of the corresponding word spacing slot. When the modal grouping tag of the brand tone vector is consistent with the modal type tag of the word distance slot, the brand tone vector is superimposed on the vector already stored in the word distance slot in the form of vector summation. The superimposed vector in the same word distance slot is the embedding vector of the corresponding word distance slot. The multi-head self-attention mechanism uses the embedding vector of each word distance slot as the input of the query vector, key vector, and value vector. The embedding vector of each word distance slot is compared with the embedding vector of all word distance slots on the same marketing theme baseline to calculate the attention score. The embedding vectors of all word distance slots are weighted and summed according to the attention score to obtain the context enhancement vector of each word distance slot. The context smoothing is to weight the embedding vector of each word distance slot and its context enhancement vector according to a preset smoothing coefficient. The weighted average vector is used as the smooth embedding vector of the word distance slot. Marketing event representations are formed by concatenating the smooth embedding vectors of all word spacing slots on the same marketing theme baseline in indexed order. Each marketing event representation contains fused information of brand tone vectors in three modalities: text, image, and audio.

[0036] It should be further explained that, in the specific implementation process, the step S23, which involves storing all marketing events in a public recommendation window according to their generation sequence or theme similarity, and extracting the corresponding key semantic features from the marketing events located at the cluster center within the public recommendation window and setting them as style anchors for the corresponding public recommendation window, includes: Step S231: Using the first storage position of the public recommendation window as the starting point of the sliding window, and according to the preset step size, the subsequent marketing events are successively merged into the current sliding window until the theme similarity of the marketing events in the sliding window meets the clustering condition to form an event cluster. Step S232: Determine the marketing event corresponding to the cluster center of the event cluster, extract the principal components of the semantic vector of the corresponding marketing event, compress the principal components of the semantic vector into a low-dimensional feature vector, and use the low-dimensional feature vector as the style anchor point corresponding to the current event cluster. Step S233: Set corresponding style anchors for all event clusters in the public recommendation window, and establish a two-way index association between each style anchor and the marketing events in its event cluster.

[0037] It should be noted that the public recommendation window is a vector storage structure with a fixed window capacity. The first storage location is the starting address of the window, and subsequent marketing events are stored in adjacent addresses in the order of their generation. The sliding window is a dynamic window with a variable capacity that starts from the starting address, and the preset step size is the length of the storage unit occupied by a marketing event in the public recommendation window. Topic similarity is calculated by taking the cosine similarity between each pair of marketing event representation vectors within the current sliding window, and then averaging all the cosine similarities as the topic similarity of the marketing events within the window. The clustering conditions are: the topic similarity of marketing events within the sliding window is greater than the preset similarity threshold and the number of marketing events within the sliding window reaches the preset minimum number. An event cluster is a set of all marketing events within the current sliding window after satisfying the clustering condition. The cluster center of an event cluster is the location of the marketing event whose sum of distances from the representation vectors of each marketing event within the cluster to the representation vectors of other marketing events within the cluster is the smallest. The semantic vector principal component is obtained by performing principal component analysis on all marketing event representation vectors within the event cluster, extracting the projection component along the first principal component direction, sorting the projection components along the first principal component direction from largest to smallest, and retaining a predetermined number of component values ​​to obtain a low-dimensional feature vector. The bidirectional index association includes a forward index pointing from a style anchor to a marketing event, and a reverse index pointing from a marketing event to a style anchor. The forward index is used to retrieve all marketing events within the event cluster corresponding to a style anchor, while the reverse index is used to query the style anchor corresponding to the event cluster to which the current marketing event belongs.

[0038] It should be further explained that, in the specific implementation process, step S3, which involves establishing a marketing demand pool, connecting all style anchors to the marketing demand pool, having customers submit their respective demands, adapting them to corresponding style anchors to form an anchor set, and retrieving and generating corresponding marketing content that meets their needs based on the anchor set, includes: Step S31: Establish a marketing demand pool, which is a database that stores the mapping relationship between customer marketing demand data and style anchors to be matched; Step S32: Connect all style anchors to the marketing demand pool according to their respective marketing theme baseline and event cluster affiliation, and configure a demand matching interface for each style anchor. Step S33: Obtain the customer's input requirement description text, perform semantic parsing on the requirement description text, extract the requirement intent vector, calculate the similarity between the requirement intent vector and the feature vector of each style anchor point, sort according to similarity, and adapt style anchor points with similarity exceeding a preset threshold to form an anchor point set. Step S34: Using the marketing events corresponding to each style anchor in the anchor set as search conditions, retrieve the corresponding marketing content fragments from the public recommendation window, splice each marketing content fragment according to the narrative logic in the requirement description text, automatically adjust the content structure based on preset scene tags, fine-tune the content in the annotation and review interface, and generate marketing content that meets the customer's needs.

[0039] It should be noted that the marketing demand pool is a relational database deployed on a cloud server. This relational database contains a demand record table and an anchor mapping table. The demand record table is used to store customer demand description text and its parsed demand intent vector. The anchor mapping table is used to store style anchor identifiers and their associated marketing event identifiers. The demand record table and the anchor mapping table establish a many-to-many mapping relationship through the demand-anchor association field. Style anchors are grouped into primary groups based on the baseline label of the marketing theme baseline they belong to, and then into secondary groups based on the cluster identifier of the event cluster they belong to. Each style anchor after the secondary grouping is connected to the marketing demand pool through an interface registration method. The semantic parsing of the demand description text is performed using a large language model. The demand intent vector is extracted by taking the output vector corresponding to the marker in the last hidden layer of the large language model. The similarity calculation uses the cosine similarity formula, with a preset threshold of 0.7. Style anchors with similarity exceeding the preset threshold are combined into an anchor set. The narrative logic is a causal or temporal sequence extracted from the requirements description text. The content fragments are arranged in the order in which events occur in the narrative logic sequence. The scene tags are extracted from the requirements description text through keyword matching.

[0040] It should be further explained that, in the specific implementation process, the step S32, which involves connecting all the style anchors to the marketing demand pool according to their respective marketing theme baselines and event clusters, and configuring a demand matching interface for each style anchor, includes: Step S321: Traverse all event clusters in the public recommendation window, and extract the corresponding style anchor and the baseline label of the marketing theme baseline where the corresponding style anchor is located for each event cluster; Step S322: Organize each style anchor point and its corresponding baseline label, event cluster identifier, and modal grouping label into a demand matching index tuple, and write all demand matching index tuples into the demand matching index table of the marketing demand pool; Step S323: Configure a corresponding demand matching interface for each style anchor point. The demand matching interface is used to receive the demand intent vector from external input and output the matching score between the current style anchor point and the current customer demand.

[0041] When traversing all event clusters in the public recommendation window, the cluster identifier of the event cluster is used as the traversal index to extract the style anchor vector corresponding to each event cluster and its associated marketing theme baseline label. The demand matching index tuple is a four-tuple data structure, consisting of a style anchor identifier, a baseline label, an event cluster identifier, and a modal grouping label. The demand matching index table uses the style anchor identifier as the primary key and the baseline label, event cluster identifier, and modal grouping label as composite index fields. It supports query matching through any combination of fields. The input parameter of the demand matching interface is the customer demand intent vector. The interface internally performs similarity calculation between the demand intent vector and the current style anchor feature vector. The interface output value is a similarity score between 0 and 1. The demand matching interface for each style anchor is deployed as an independent service instance, and the interfaces are isolated from each other.

[0042] It should be further explained that, in the specific implementation process, the step S34, which uses the marketing events corresponding to each style anchor point in the anchor point set as search conditions, retrieves the corresponding marketing content fragments from the public recommendation window, splices each marketing content fragment according to the narrative logic in the demand description text, automatically adjusts the content structure based on preset scene tags, and performs fine-tuning of the content in the annotation and review interface to generate marketing content that meets the customer's needs, includes: Step S341: Using the marketing events of each style anchor in the anchor set as search conditions, retrieve the corresponding marketing content fragments from the public recommendation window, extract the narrative logic sequence, determine the splicing order of each marketing content fragment according to the narrative logic sequence, and generate preliminary spliced ​​content. Step S342: Obtain the scene tags associated with customer needs, match the corresponding content structure template, and adaptively adjust the arrangement and fusion weight of each modal fragment in the initial spliced ​​content according to the content structure template to obtain structure-adapted content; Step S343: Input the structural adaptation content into the annotation review interface, and in response to the reviewer's fine-tuning instruction for any modal segment, recall semantically similar alternative segments from the public recommendation window and replace them; Step S344: After all marketing content fragments are confirmed, render and output multimodal marketing content, and store the generated content and its corresponding style anchors, narrative logic identifiers and scene tags in the marketing demand pool.

[0043] It should be noted that after retrieving the corresponding marketing content fragments from the public recommendation window, the narrative logic identifiers associated with each marketing content fragment when it was stored in the public recommendation window are extracted. The splicing order of each marketing content fragment is determined according to the causal order or time order corresponding to the narrative logic identifiers. The marketing content fragments are spliced ​​together in sequence, and transitional statements are inserted between adjacent marketing content fragments to generate preliminary spliced ​​content. Extract scenario tags from the requirements description text. Scenario tags include various tags such as technical roadshow scenarios, customer visit scenarios, industry exhibition scenarios, and online communication scenarios. Match the corresponding content structure template from the preset multi-scenario narrative template library based on the scenario tags. Each content structure template has a pre-set arrangement order of each content module, weight distribution of technical parameters, case presentation method, and call to action design format. According to the arrangement order of each module in the content structure template, rearrange the order of each modal segment in the initial spliced ​​content, adjust the display density and display position of the technical parameter content segments according to the weight distribution of technical parameters, adjust the display format of the case-type content segments according to the case presentation method, and adjust the content and layout of the closing introductory text according to the call to action design format to obtain structurally adapted content. The structure-adapted content is pushed to the annotation review interface, which is a web-based visual editing interface. The annotation review interface displays the preview area and editing tool area of ​​each content module in sections. It responds to the modification operation triggered by the reviewer on any content module in the annotation review interface, records the modification point, the content before modification, the content after modification, and the modification timestamp, and generates modification trace data. The modification trace data is parsed to extract the semantic difference vector of the content before and after the modification. This semantic difference vector is used as the supervision signal for LoRA fine-tuning. LoRA fine-tuning is performed on the current large language model to update the model parameters. The fine-tuned large language model is then used to retrieve the corresponding marketing content fragments from the anchor set and regenerate the structurally adapted content. This content is then pushed to the annotation and review interface again until the reviewers confirm and approve it.

[0044] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent generation of multimodal marketing content for enterprise brands based on a large language model, characterized in that: Includes the following steps: Step S1: Perform modeling based on historical marketing data to obtain a large language model that represents all marketing methods. Perform semantic analysis on all marketing data related to the enterprise brand through the large language model to determine the brand tone vectors of multiple modal types. Step S2: Determine multiple marketing theme baselines, embed all brand tone vectors into different word distance positions of the marketing theme baselines to obtain multiple marketing events, establish a public recommendation window to store all marketing events, and set the style anchor points of the public recommendation window accordingly. Step S3: Establish a marketing demand pool, connect all style anchors to the marketing demand pool, allow customers to put forward their own demands, and adapt the corresponding style anchors to form an anchor set, and retrieve and generate corresponding marketing content that meets their needs based on the anchor set.

2. The intelligent generation method for enterprise brand multimodal marketing content based on a large language model according to claim 1, characterized in that, The step of determining the brand tone vectors for multiple modal types includes: Identify multiple marketing themes in marketing activities based on historical marketing data, analyze the marketing data corresponding to the marketing themes in marketing activities, use them as marketing tools, perform convolution and pooling processing on the marketing tools, determine the corresponding model layers for each, and integrate the model layers to obtain a large language model; By adapting multiple modal types using a large language model, the model analysis domain corresponding to each modal type is determined, and marketing data related to the enterprise brand from each model analysis domain is integrated to determine the corresponding model analysis layer for each model analysis domain. The model analysis layer is enabled to perform multidimensional semantic analysis on marketing data. The analysis results are then vectorized to obtain brand tone vectors that match the marketing theme of the enterprise brand. The brand tone vectors are grouped according to modality type.

3. The intelligent generation method for enterprise brand multimodal marketing content based on a large language model according to claim 2, characterized in that, The step of determining the corresponding model analysis layer for each model analysis domain includes: Obtain all modalities involved in the enterprise's brand marketing, including text modalities, image modalities, and audio modalities; Using the underlying shared parameters of the large language model as a common base, a corresponding modality encoder is instantiated for each modality type. Each modality encoder receives the original marketing data of the corresponding modality type and extracts modality-specific features in its own feature space. The extracted modality-specific features are mapped to the unified semantic space of the large language model. Combined with the enterprise brand identity information, the model analysis domain boundary corresponding to each modality type is defined in the unified semantic space, and an independent model analysis layer is assigned to each model analysis domain.

4. The intelligent generation method for enterprise brand multimodal marketing content based on a large language model according to claim 3, characterized in that, The steps for grouping the brand tone vector based on modal type include: Enable the model analysis layer corresponding to each modality type to perform multidimensional semantic analysis on the input marketing data of each modality, including sentiment polarity analysis, style feature analysis and rhetorical device analysis; The semantic analysis results of each dimension are mapped to numerical vector components, and the numerical vector components are combined to obtain the original semantic vector of the corresponding modality type. For each modality type, the original semantic vector is oriented and normalized to generate a brand tone vector that matches the marketing theme of the enterprise brand. Brand tone vectors belonging to the same modality type are grouped into the same vector group, and each vector group is labeled with a modality label.

5. The intelligent generation method for enterprise brand multimodal marketing content based on a large language model according to claim 4, characterized in that, The step of setting the style anchor point of the public recommendation window includes: Identify multiple marketing theme baselines involved in corporate brand marketing. These marketing theme baselines are semantic baseline vector sequences covering different marketing narrative frameworks. The brand tone vectors under each modality group are semantically embedded into the preset word distance positions of each marketing theme baseline according to the modality correspondence. During the embedding process, the brand tone vectors in adjacent positions are smoothed through an attention mechanism to generate multiple marketing events carrying modality fusion information. All marketing events are stored in a public recommendation window according to their generation sequence or theme similarity. For marketing events located at the cluster center in the public recommendation window, the corresponding key semantic features are extracted and associated as style anchors for the corresponding public recommendation window.

6. The intelligent generation method for enterprise brand multimodal marketing content based on a large language model according to claim 5, characterized in that, The step of generating multiple marketing events carrying modal fusion information includes: Extract multiple pre-defined word spacing slots from the baseline of each marketing theme, with each word spacing slot associated with a location code and a modality type label; Based on the modal grouping tags to which each brand tone vector belongs, the brand tone vector is matched to word distance slots with tags of the same modal type, and the initial embedding is completed by vector summation. For the adjacent word distance position vectors after the initial embedding is completed, a multi-head self-attention mechanism is used to calculate the semantic association weight between each position. Based on the association weight, a weighted smoothing process is performed on the adjacent position vectors to obtain a marketing event representation that integrates contextual semantics.

7. The intelligent generation method for enterprise brand multimodal marketing content based on a large language model according to claim 6, characterized in that, The step of extracting the corresponding key semantic features and associating them as style anchors for the corresponding public recommendation window includes: Using the first entry position of the public recommendation window as the starting point of the sliding window, subsequent marketing events are sequentially merged into the current sliding window according to the preset step size until the theme similarity of the marketing events in the sliding window meets the clustering condition to form an event cluster. Determine the marketing event corresponding to the cluster center of the event cluster, extract the principal components of the semantic vector of the corresponding marketing event, compress the principal components of the semantic vector into a low-dimensional feature vector, and use the low-dimensional feature vector as the style anchor point corresponding to the current event cluster. Set corresponding style anchors for all event clusters within the public recommendation window, and establish a two-way index association between each style anchor and the marketing events within its event cluster.

8. The intelligent generation method for enterprise brand multimodal marketing content based on a large language model according to claim 7, characterized in that, The step of generating corresponding marketing content that meets the requirements based on anchor point set retrieval includes: Establish a marketing demand pool, which is a database that stores the mapping relationship between customer marketing demand data and style anchors to be matched; All style anchors are connected to the marketing demand pool according to their respective marketing theme baseline and event cluster affiliation, and a demand matching interface is configured for each style anchor. Obtain the customer's input requirement description text, perform semantic parsing on the requirement description text, extract the requirement intent vector, calculate the similarity between the requirement intent vector and the feature vector of each style anchor point, sort according to similarity, and adapt style anchor points with similarity exceeding a preset threshold to form an anchor point set. Using marketing events corresponding to each style anchor in the anchor set as search criteria, the system retrieves corresponding marketing content fragments from the public recommendation window, splices each marketing content fragment according to the narrative logic in the requirement description text, automatically adjusts the content structure based on preset scene tags, and makes fine-tuning of the content in the annotation and review interface to generate marketing content that meets the client's needs.

9. The intelligent generation method for enterprise brand multimodal marketing content based on a large language model according to claim 8, characterized in that, The step of configuring a requirement matching interface for each style anchor point includes: Iterate through all event clusters in the public recommendation window, and extract the corresponding style anchor and the baseline label of the marketing theme baseline where the corresponding style anchor is located for each event cluster; Organize each style anchor point and its corresponding baseline label, event cluster identifier, and modal grouping label into a demand matching index tuple, and write all demand matching index tuples into the demand matching index table of the marketing demand pool. Configure a corresponding demand matching interface for each style anchor point to receive the demand intent vector from external input and output the matching score between the current style anchor point and the current customer demand.

10. The intelligent generation method for enterprise brand multimodal marketing content based on a large language model according to claim 9, characterized in that, The steps for generating marketing content that meets customer needs include: Using marketing events of each style anchor in the anchor set as search criteria, the corresponding marketing content fragments are retrieved from the public recommendation window, the narrative logic sequence is extracted, the splicing order of each marketing content fragment is determined according to the narrative logic sequence, and preliminary spliced ​​content is generated. Obtain scene tags associated with customer needs, match the corresponding content structure template, and adaptively adjust the arrangement and fusion weight of each modal fragment in the initial spliced ​​content according to the content structure template to obtain structure-adapted content; Input the structural adaptation content into the annotation and review interface. In response to the reviewer's fine-tuning instructions for any modal segment, recall semantically similar alternative segments from the public recommendation window and replace them. Once all marketing content fragments are confirmed, multimodal marketing content is rendered and output, and the generated content, along with its corresponding style anchors, narrative logic identifiers, and scene tags, is stored in the marketing demand pool.

Images