Intelligent marketing copywriting generation and effect evaluation method driven by large language model

Abstract

The invention provides an intelligent marketing copywriting generation and effect evaluation method driven by a large language model, relates to the technical field of language models, and comprises the steps of constructing a hierarchical cross-modal knowledge graph and establishing a knowledge retrieval index. Semantic analysis is performed based on a bidirectional attention mechanism, related knowledge is retrieved by using query vectors, and an initial marketing copywriting is generated. Performing knowledge consistency verification to generate a knowledge-enhanced marketing copywriting, and generating a candidate copywriting set by adopting a diversity sampling strategy; and performing knowledge coverage, creativity and expected effect scoring on the candidate copywriting by using a multi-target evaluation network, screening an optimal copywriting through a Pareto optimization algorithm, and taking a generated path of the optimal copywriting as a positive sample to update a knowledge graph and decoder network parameters. Through knowledge graph enhancement and multi-target evaluation optimization, high-quality and high-matching-degree marketing copywriting can be generated, and the marketing effect is improved.

Description

Technical Field

[0001] The present invention relates to language model technology, and in particular to a method for generating and evaluating intelligent marketing copy driven by a large language model. Background Art

[0002] Marketing copy serves as a bridge between products and users. Excellent marketing copy can effectively boost product awareness and sales. With the rapid development of the internet, traditional methods of creating marketing copy are no longer able to meet the growing market demand. Therefore, how to efficiently generate high-quality, creative marketing copy has become a focus for many companies and marketers.

[0003] Currently, some natural language processing and machine learning-based technologies have been applied to the automatic generation of marketing copy, such as template-based copy generation, rule-based copy generation, and deep learning-based copy generation. However, these methods still have some flaws and shortcomings: Difficulty in fully utilizing multimodal marketing data: Most existing methods focus only on text data, while ignoring other types of marketing data such as images and videos, resulting in the generated copy lacking rich expressiveness and appeal. Lack of knowledge-guided copy generation: The copy generated by existing methods often lacks knowledge support related to products, users, and scenarios, resulting in empty content and lack of persuasiveness. Single copy evaluation method: Existing methods usually only consider indicators such as the fluency and grammatical correctness of the copy, while ignoring more important evaluation dimensions such as the creativity, knowledge coverage, and expected effects of the copy, making it difficult to effectively screen out truly excellent marketing copy. Summary of the Invention

[0004] The embodiments of the present invention provide a large language model-driven intelligent marketing copy generation and effect evaluation method, which can solve the problems in the existing technology.

[0005] According to a first aspect of the embodiments of the present invention, Provides a large language model-driven intelligent marketing copywriting generation and effect evaluation method, including: A multi-channel feature extraction network is used to perform hierarchical feature extraction on multimodal marketing data collected in real time to obtain a feature vector matrix; hierarchical entity nodes and relationship edges are constructed based on the feature vector matrix; the hierarchical entity nodes and relationship edges are input into a graph neural network to generate an initial cross-modal knowledge graph; a contrastive learning method is used to perform semantic vectorization on the initial cross-modal knowledge graph to obtain a knowledge graph vector representation, and a knowledge retrieval index is established; Receive marketing needs, perform semantic analysis on the marketing needs based on a bidirectional attention mechanism to generate a query vector; use the query vector to retrieve relevant knowledge graph vector representations in the knowledge retrieval index; construct a decoder network based on the knowledge graph vector representation, input the marketing needs into the decoder network, and generate an initial marketing copy; verify the semantic match between the initial marketing copy and the knowledge graph vector representation to generate a knowledge-enhanced marketing copy; and use a diversity sampling strategy to expand the knowledge-enhanced marketing copy to generate a set of candidate marketing copies; The candidate marketing copy set is input into a preset multi-objective evaluation network to obtain an evaluation score matrix; based on the evaluation score matrix, the optimal marketing copy is screened out using a Pareto optimization algorithm; the generation path of the optimal marketing copy is used as a positive sample to optimize the knowledge retrieval index.

[0006] Retrieving relevant knowledge graph vector representations in the knowledge retrieval index using the query vector; constructing a decoder network based on the knowledge graph vector representation, inputting the marketing demand into the decoder network, and generating initial marketing copy including: A locality-sensitive hashing algorithm is used to partition the high-dimensional vector space into subspaces, construct a fast retrieval index of the knowledge graph vector, set a dynamic retrieval threshold based on cosine similarity calculation, and adaptively adjust the dynamic retrieval threshold using an annealing algorithm; the query vector is input into the fast retrieval index to obtain a set of first-order knowledge nodes, and a graph attention network is constructed based on the first-order knowledge node set. The attention weight distribution of adjacent nodes is calculated, and the relationship path with the highest attention weight is selected for multi-hop retrieval to obtain a multi-hop knowledge subgraph; Organizing the knowledge graph vectors in the multi-hop knowledge subgraph into a knowledge memory matrix, constructing a knowledge selection attention layer and a semantic fusion attention layer, wherein the knowledge selection attention layer obtains attention distribution based on the product operation of the decoding state and the knowledge memory matrix, and the semantic fusion attention layer fuses the knowledge representation and the decoding state through a feedforward neural network, adopts residual connection and layer normalization for information transfer, and generates a knowledge-enhanced decoder network; The marketing demand and the knowledge context vector are concatenated as the initial state of the knowledge-enhanced decoder network. The beam search algorithm is used to maintain the candidate sequence. The probability distribution of the next word is generated through decoder state update, vocabulary distribution calculation and knowledge constraints. A decoding strategy based on kernel sampling is introduced, and the sampling temperature parameters are dynamically adjusted according to the probability distribution. The generation process is controlled based on the length perception mechanism to output the initial marketing copy.

[0007] Verify the semantic matching degree between the initial marketing copy and the knowledge graph vector representation to generate a knowledge-enhanced marketing copy; adopt a diversity sampling strategy to expand the knowledge-enhanced marketing copy to generate a set of candidate marketing copies including: Obtain word-level semantic representations for the initial marketing copy through a character-enhanced word vector model, perform semantic dependency analysis on the word-level semantic representations to obtain phrase-level semantic representations, and fuse the word-level semantic representations and the phrase-level semantic representations to generate a hierarchical semantic representation; Calculating a soft alignment matrix between the hierarchical semantic representation and the knowledge graph vector representation based on a preset bidirectional interactive attention network, performing cross-modal information fusion based on the soft alignment matrix through a gating mechanism to obtain a fusion feature, calculating a matching score for the fusion feature, and generating a knowledge consistency verification result; Determining knowledge-missing regions and semantic deviation regions based on the matching scores in the knowledge consistency verification results, calculating deviation vectors of the knowledge-missing regions and the semantic deviation regions using contrastive learning, adjusting the content based on the deviation vectors through gradient guidance, and reorganizing the supplemented and adjusted content using a chapter structure template to generate knowledge-enhanced marketing copy; The knowledge-enhanced marketing copy is variationally sampled in each semantic cluster through a conditional variational autoencoder to obtain copy variants, and the semantic distance matrix between the copy variants is calculated; a dynamic difference threshold is set based on the semantic distance matrix to construct a maximization marginal relevance criterion, and a greedy selection algorithm is used to screen differential copy that meets the maximization marginal relevance criterion from the copy variants; the differential copy is input into a multi-objective optimization function for comprehensive evaluation, and the copy that meets the evaluation requirements is selected to form a candidate marketing copy set.

[0008] The knowledge-enhanced marketing copy is sampled through a conditional variational autoencoder in each semantic cluster to obtain copy variants, and a semantic distance matrix between the copy variants is calculated; a dynamic difference threshold is set based on the semantic distance matrix to construct a maximum marginal relevance criterion, and a greedy selection algorithm is used to select different copies that meet the maximum marginal relevance criterion from the copy variants, including: The knowledge-enhanced marketing copy is subjected to variational sampling in each semantic cluster through a conditional variational autoencoder. The conditional variational autoencoder inputs the knowledge-enhanced marketing copy into a content encoder to obtain a mean vector and a variance vector, inputs the corresponding semantic cluster features into a semantic cluster encoder to obtain a conditional vector, performs reparameterized sampling based on the mean vector and the variance vector to obtain latent variables, and fuses the latent variables with the conditional vector to obtain multiple copy variants through a decoder; Extracting contextual semantic features of the copy variants using a BERT-based text representation model, calculating word mover distances based on the contextual semantic features to obtain lexical difference values, calculating syntactic dependency tree edit distances to obtain structural difference values, calculating topic distribution JS divergence to obtain topic difference values, performing weighted fusion on the lexical difference values, the structural difference values, and the topic difference values to calculate semantic distances between the copy variants, and obtaining a semantic distance matrix; A dynamic difference threshold is set based on the weighted sum of the standard deviation of the semantic distance matrix and a basic threshold, and a maximization marginal relevance criterion is constructed that integrates the quality scoring item and the similarity scoring item. The quality scoring item includes a weighted combination of the language fluency score, the knowledge relevance score, and the marketing goal fit score. The similarity scoring item calculates the maximum similarity between the candidate copy and the selected copy based on the semantic distance matrix; a greedy selection algorithm is used to screen different copy that meets the maximization marginal relevance criterion from the copy variants.

[0009] The candidate marketing copy set is input into a preset multi-objective evaluation network to obtain an evaluation score matrix; based on the evaluation score matrix, the optimal marketing copy is screened using a Pareto optimization algorithm, including: The candidate marketing copy set is input into a preset multi-objective evaluation network for evaluation, and each marketing copy is scored for knowledge coverage, creativity, and expected effect.

[0010] The knowledge coverage score, the creativity score and the expected effect score are combined into an evaluation score matrix. Based on the evaluation score matrix, the candidate marketing copy is divided into multiple Pareto fronts using non-dominated sorting. The optimal copy is selected within the optimal Pareto front in combination with the crowding sorting, wherein the crowding sorting simultaneously considers the target space distance and the decision space distance. The target space distance is calculated based on the evaluation score matrix, and the decision space distance is calculated based on the semantic representation of the copy, so as to obtain the optimal marketing copy that is balanced and differentiated in multiple evaluation dimensions.

[0011] Dividing the candidate marketing copy into multiple Pareto fronts using non-dominated sorting based on the evaluation score matrix, and selecting the optimal copy within the optimal Pareto front in combination with the congestion sorting includes: performing non-dominated sorting on the candidate marketing copy based on the evaluation score matrix, counting the number of times each candidate marketing copy is dominated by other candidate marketing copies, and placing candidate marketing copies with zero dominated times into the first Pareto frontier; reducing the number of times the candidate marketing copy in the first Pareto frontier is dominated by the other candidate marketing copy, and placing the candidate marketing copy whose number of domination becomes zero after the update into the second Pareto frontier, and repeating the above process until all candidate marketing copies are placed into the corresponding Pareto frontier; For each candidate marketing copy in the first Pareto front, calculate the crowding distance of each candidate marketing copy, where the crowding distance is calculated based on the difference in scores of the candidate marketing copy and its adjacent candidate marketing copies on each evaluation indicator; The candidate marketing copy in the first Pareto front is sorted in descending order according to the crowding distance, and the candidate marketing copy with the largest crowding distance is selected as the optimal marketing copy.

[0012] Taking the generation path of the optimal marketing copy as a positive sample, optimizing the knowledge retrieval index includes: Obtaining a generation path for an optimal marketing copy, the generation path including a search sequence of a knowledge search index during the generation of the optimal marketing copy and the knowledge content obtained during each search; Constructing the search sequence and the corresponding knowledge content in the generation path into positive sample training data, wherein the positive sample training data is used to characterize the knowledge search pattern in the process of generating high-quality marketing copy; The knowledge retrieval index is optimized based on the positive sample training data, and the retrieval priority of the knowledge content matching the retrieval sequence is improved by adjusting the association weights of the knowledge content in the knowledge retrieval index.

[0013] The beneficial effects of this application are as follows: 1. Improving the quality of marketing copy generation: This method builds a cross-modal knowledge graph based on multimodal marketing data and combines it with a knowledge-enhanced decoder network to generate marketing copy. This fully leverages multimodal information, ensuring that the generated copy is rich in content, accurate in knowledge, and highly aligned with marketing needs. Knowledge consistency verification and diverse sampling strategies further enhance the quality and diversity of the copy.

[0014] 2. Achieve multi-objective optimization: This invention constructs a multi-objective evaluation network that can simultaneously evaluate the knowledge coverage, creativity and expected effect of the copy, and uses the Pareto optimization algorithm to screen the optimal copy, thereby achieving a balance between multiple objectives and generating high-quality copy that better meets actual needs.

[0015] 3. Support continuous learning and optimization: The present invention uses the optimal copy generation path to update the knowledge graph vector representation and knowledge retrieval index, and updates the decoder network parameters through online learning, so that the system can continuously learn from historical data and continuously improve the quality and efficiency of copy generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flowchart of a method for generating and evaluating intelligent marketing copy driven by a large language model according to an embodiment of the present invention; Figure 2 This is a comparison chart of marketing copy quality assessment based on the conditional variational autoencoder according to an embodiment of the present invention; Figure 3 This is a diagram of dimensional semantic difference evaluation and comparative analysis in an embodiment of the present invention; Figure 4 This is a trend diagram of copywriting diversity evaluation based on the maximization marginal relevance criterion in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0018] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0019] Figure 1 Schematic diagram of the process of the method for generating and evaluating the effect of intelligent marketing copy driven by a large language model according to an embodiment of the present invention. Figure 1 As shown, the method includes: A multi-channel feature extraction network is used to perform hierarchical feature extraction on multimodal marketing data collected in real time to obtain a feature vector matrix; hierarchical entity nodes and relationship edges are constructed based on the feature vector matrix; the hierarchical entity nodes and relationship edges are input into a graph neural network to generate an initial cross-modal knowledge graph; a contrastive learning method is used to perform semantic vectorization on the initial cross-modal knowledge graph to obtain a knowledge graph vector representation, and a knowledge retrieval index is established; Receive marketing needs, perform semantic analysis on the marketing needs based on a bidirectional attention mechanism to generate a query vector; use the query vector to retrieve relevant knowledge graph vector representations in the knowledge retrieval index; construct a decoder network based on the knowledge graph vector representation, input the marketing needs into the decoder network, and generate an initial marketing copy; verify the semantic match between the initial marketing copy and the knowledge graph vector representation to generate a knowledge-enhanced marketing copy; and use a diversity sampling strategy to expand the knowledge-enhanced marketing copy to generate a set of candidate marketing copies; The candidate marketing copy set is input into a preset multi-objective evaluation network to obtain an evaluation score matrix; based on the evaluation score matrix, the optimal marketing copy is screened out using a Pareto optimization algorithm; the generation path of the optimal marketing copy is used as a positive sample to optimize the knowledge retrieval index.

[0020] In an optional embodiment, using the query vector to retrieve relevant knowledge graph vector representations in the knowledge retrieval index; constructing a knowledge-enhanced decoder network based on the knowledge graph vector representation; inputting the marketing demand into the knowledge-enhanced decoder network; and generating an initial marketing copy includes: A locality-sensitive hashing algorithm is used to perform subspace partitioning on the high-dimensional vector space, construct a fast retrieval index of the knowledge graph vector, set a dynamic retrieval threshold based on cosine similarity calculation, and adaptively adjust the dynamic retrieval threshold using an annealing algorithm; the query vector is input into the fast retrieval index to obtain a first-order knowledge node set, a graph attention network is constructed based on the first-order knowledge node set, the attention weight distribution of adjacent nodes is calculated, and the relationship path with the highest attention weight is selected for multi-hop retrieval to obtain a multi-hop knowledge subgraph; Organizing the knowledge graph vectors in the multi-hop knowledge subgraph into a knowledge memory matrix, constructing a knowledge selection attention layer and a semantic fusion attention layer, wherein the knowledge selection attention layer obtains attention distribution based on the product operation of the decoding state and the knowledge memory matrix, and the semantic fusion attention layer fuses the knowledge representation and the decoding state through a feedforward neural network, adopts residual connection and layer normalization for information transfer, and generates a knowledge-enhanced decoder network; The marketing demand and the knowledge context vector are concatenated as the initial state of the knowledge-enhanced decoder network. The beam search algorithm is used to maintain the candidate sequence. The probability distribution of the next word is generated through decoder state update, vocabulary distribution calculation and knowledge constraints. A decoding strategy based on kernel sampling is introduced, and the sampling temperature parameters are dynamically adjusted according to the probability distribution. The generation process is controlled based on the length perception mechanism to output the initial marketing copy.

[0021] First, a knowledge retrieval index is constructed. The locality-sensitive hashing (LSH) algorithm is used to partition the high-dimensional vector space into multiple subspaces. The vector representations in the knowledge graph are stored in the corresponding subspaces, thereby constructing a fast retrieval index. A dynamic retrieval threshold based on cosine similarity is set, and the threshold is adaptively adjusted using an annealing algorithm to ensure retrieval accuracy and efficiency. For example, the dimension of the knowledge graph vector representation is set to 512, and the LSH algorithm is used to partition it into 1024 subspaces. The initial retrieval threshold is set to 0.8, the initial temperature of the annealing algorithm is set to 1, and the cooling coefficient is set to 0.95. The threshold is adjusted after each iteration based on the quality of the retrieval results.

[0022] Secondly, retrieve relevant knowledge. Represent the marketing demand as a query vector and input it into the constructed fast retrieval index. According to the dynamic retrieval threshold, obtain a set of first-order knowledge nodes whose cosine similarity with the query vector is higher than the threshold. For example, for the demand of "promoting new sports shoes", first-order knowledge nodes such as "sports shoes", "running", and "fitness" can be retrieved. Next, construct a graph attention network based on the set of first-order knowledge nodes. Calculate the attention weight distribution between adjacent nodes, and select the path with the highest attention weight for multi-hop retrieval to obtain a knowledge subgraph containing more relevant information. For example, starting from the "sports shoes" node, the "running" and "fitness" nodes can be reached through the relationship "applicable to". By calculating the attention weight, the "applicable to running" path with a higher weight is selected for multi-hop retrieval to further obtain nodes such as "marathon" and "sports injuries", and finally obtain a knowledge subgraph containing multiple related nodes.

[0023] Next, a knowledge-enhanced decoder network is constructed. The knowledge graph vectors in the retrieved multi-hop knowledge subgraph are organized into a knowledge memory matrix. A knowledge selection attention layer and a semantic fusion attention layer are constructed. The knowledge selection attention layer obtains an attention distribution by multiplying the decoding state with the knowledge memory matrix, which is used to select knowledge relevant to the current decoding state. The semantic fusion attention layer fuses the knowledge representation with the decoding state through a feedforward neural network, and uses residual connections and layer normalization for information transfer to generate a knowledge-enhanced decoder network.

[0024] Finally, marketing copy is generated. The marketing demand and knowledge context vectors are concatenated as the initial state of the decoder network. A beam search algorithm is used to maintain the candidate sequence. The probability distribution of the next word is generated through decoder state updates, vocabulary distribution calculation, and knowledge constraints. A decoding strategy based on kernel sampling is introduced, dynamically adjusting the sampling temperature parameter based on the probability distribution. A length-aware mechanism is used to control the generation process, ultimately outputting the initial marketing copy. For example, for the demand for "promoting new sports shoes," the generated copy might be "New sports shoes, lightweight and breathable, help you run faster and farther, and avoid sports injuries."

[0025] The solution of this application can: Improving the relevance of marketing copy: By introducing knowledge graphs, we can provide rich background knowledge for the generation of marketing copy, ensuring that the generated copy is highly relevant to the target product or service, and avoiding irrelevant content. Enhancing the information content of marketing copy: Knowledge graphs contain a large number of entities, concepts, and relationships, which can provide richer information for marketing copy, making it more attractive and persuasive. Improving the efficiency of marketing copy generation: By building a knowledge retrieval index and a knowledge-enhanced decoder network, we can effectively improve the efficiency of marketing copy generation and reduce the cost of manual creation.

[0026] In an optional embodiment, verifying the semantic matching between the initial marketing copy and the knowledge graph vector representation to generate a knowledge-enhanced marketing copy; and using a diversity sampling strategy to expand the knowledge-enhanced marketing copy to generate a set of candidate marketing copies includes: Obtain word-level semantic representations for the initial marketing copy through a character-enhanced word vector model, perform semantic dependency analysis on the word-level semantic representations to obtain phrase-level semantic representations, and fuse the word-level semantic representations and the phrase-level semantic representations to generate a hierarchical semantic representation; Calculating a soft alignment matrix between the hierarchical semantic representation and the knowledge graph vector representation based on a preset bidirectional interactive attention network, performing cross-modal information fusion based on the soft alignment matrix through a gating mechanism to obtain a fusion feature, calculating a matching score for the fusion feature, and generating a knowledge consistency verification result; Determining knowledge-missing regions and semantic deviation regions based on the matching scores in the knowledge consistency verification results, calculating deviation vectors of the knowledge-missing regions and the semantic deviation regions using contrastive learning, adjusting the content based on the deviation vectors through gradient guidance, and reorganizing the supplemented and adjusted content using a chapter structure template to generate knowledge-enhanced marketing copy; The marketing copy enhanced by the knowledge is sampled through a conditional variational autoencoder in each semantic cluster to obtain copy variants, and the semantic distance matrix between the copy variants is calculated; a dynamic difference threshold is set based on the semantic distance matrix to construct a maximization marginal relevance criterion, and a greedy selection algorithm is used to screen differential copy that meets the maximization marginal relevance criterion from the copy variants; the differential copy is input into a multi-objective optimization function for comprehensive evaluation, and the copy that meets the evaluation requirements is selected to form a candidate marketing copy set First, a hierarchical semantic representation is constructed for the input initial marketing copy. Using a character-enhanced word vector model, each word in the initial copy is converted into a word vector to obtain a word-level semantic representation. Semantic dependency analysis is then performed on the word-level semantic representation to identify semantic relationships between words, such as subject-verb and verb-object relationships, thereby obtaining a phrase-level semantic representation. Finally, the phrase-level semantic representation is input into a self-attention network to capture the semantic connections between different phrases in the sentence and generate a sentence-level semantic representation. The word-level, phrase-level, and sentence-level semantic representations are fused to ultimately obtain a hierarchical semantic representation of the initial marketing copy.

[0027] Next, we construct a vector representation of the knowledge graph. A knowledge graph is composed of entities and relationships, which can be converted into a vector representation for easier computation. For example, we can use knowledge graph embedding techniques to map the entities and relationships in the knowledge graph into a low-dimensional vector space.

[0028] Then, knowledge consistency verification is performed. A bidirectional interactive attention network is constructed, and the hierarchical semantic representation of the initial marketing copy and the vector representation of the knowledge graph are input into the network. The network calculates a soft alignment matrix between the two, which represents the degree of association between different parts of the copy and different concepts in the knowledge graph. Based on the soft alignment matrix, the information of the two modalities is fused through a gating mechanism to obtain fusion features. Then, the matching score is calculated based on three dimensions: semantic similarity, knowledge coverage, and logical consistency. Semantic similarity refers to the semantic proximity between the copy content and related knowledge concepts; knowledge coverage refers to the coverage of the copy content on related knowledge; and logical consistency refers to whether there is a logical contradiction between the copy content and the knowledge in the knowledge graph. Finally, the knowledge consistency verification results are generated, including the matching score and the specific scores of each dimension.

[0029] For example, suppose the initial marketing copy is "This phone has powerful camera capabilities and exceptionally long battery life." The knowledge graph contains concepts such as "phone," "camera," and "battery life." Calculations show that the copy has high semantic similarity with the concepts "camera" and "battery life," good knowledge coverage, and logical consistency, resulting in a high match score.

[0030] Based on the knowledge consistency verification results, we identify knowledge-missing areas and semantic deviation areas. For example, if the copy lacks information about the phone's processor, we can mark it as a knowledge-missing area. If the description of battery capacity in the copy is inconsistent with the information in the knowledge graph, we can mark it as a semantic deviation area.

[0031] For areas with missing knowledge, relevant knowledge nodes are retrieved from the knowledge graph to supplement the content. For example, knowledge related to mobile phone processors, such as processor models and performance parameters, can be retrieved and incorporated into the copy. For areas with semantic deviation, contrastive learning is used to calculate the deviation vector between the copy and the standard knowledge representation. Based on this deviation vector, the content is adjusted using gradient guidance. For example, the description of battery capacity in the copy can be revised based on the standard description of battery capacity in the knowledge graph.

[0032] Use a chapter structure template to restructure the supplemented and adjusted content to generate knowledge-enhanced marketing copy. For example, you can add supplementary processor information and revised battery capacity information to the corresponding locations in the copy according to a pre-defined template.

[0033] Finally, the knowledge-enhanced marketing copy is sampled and expanded for diversity to generate a set of candidate marketing copy. The knowledge graph vector space is divided into semantic clusters. For example, the knowledge graph can be divided into different semantic clusters based on different product characteristics. The knowledge-enhanced marketing copy is variationally sampled within each semantic cluster using a conditional variational autoencoder to obtain multiple copy variants. The semantic distance matrix between the copy variants is calculated, and a dynamic difference threshold is set based on this matrix to construct a criterion for maximizing marginal relevance. A greedy selection algorithm is used to screen differential copy from the copy variants that meets the criterion for maximizing marginal relevance. The screened differential copy is input into a multi-objective optimization function, which integrates the knowledge relevance score, expression fluency index, and marketing goal adaptability to comprehensively evaluate the copy. The copy that meets the evaluation requirements is selected to form a set of candidate marketing copy.

[0034] The solution of this application can: Improve the knowledge consistency of marketing copy: By comparing and verifying with the knowledge graph, ensure that the content of the marketing copy is consistent with the existing knowledge system, avoid knowledge errors or logical contradictions, and improve the credibility and professionalism of the copy. Enrich the content of marketing copy: By supplementing relevant knowledge from the knowledge graph, the marketing copy is made more complete and detailed, which can better showcase the advantages and characteristics of the product and enhance user cognition and understanding. Enhance the diversity of marketing copy: Through a diverse sampling strategy, multiple candidate copies with different expressions are generated, providing marketers with more options and enabling them to select the most appropriate copy based on different target audiences and marketing scenarios, thereby improving marketing effectiveness.

[0035] In an optional embodiment, the knowledge-enhanced marketing copy is subjected to variational sampling within each semantic cluster using a conditional variational autoencoder to obtain copy variants, and a semantic distance matrix between the copy variants is calculated; a dynamic difference threshold is set based on the semantic distance matrix to construct a maximization marginal relevance criterion, and a greedy selection algorithm is used to select differential copy that meets the maximization marginal relevance criterion from the copy variants, including: The knowledge-enhanced marketing copy is subjected to variational sampling in each semantic cluster through a conditional variational autoencoder. The conditional variational autoencoder inputs the knowledge-enhanced marketing copy into a content encoder to obtain a mean vector and a variance vector, inputs the corresponding semantic cluster features into a semantic cluster encoder to obtain a conditional vector, performs reparameterized sampling based on the mean vector and the variance vector to obtain a latent variable, and fuses the latent variable with the conditional vector to obtain a copy variant through a decoder; Extracting contextual semantic features of the copy variants using a BERT-based text representation model, calculating word mover distances based on the contextual semantic features to obtain lexical difference values, calculating syntactic dependency tree edit distances to obtain structural difference values, calculating topic distribution JS divergence to obtain topic difference values, performing weighted fusion on the lexical difference values, the structural difference values, and the topic difference values to calculate semantic distances between the copy variants, and obtaining a semantic distance matrix; A dynamic difference threshold is set based on the weighted sum of the standard deviation of the semantic distance matrix and a basic threshold, and a maximization marginal relevance criterion is constructed that integrates the quality scoring item and the similarity scoring item. The quality scoring item includes a weighted combination of the language fluency score, the knowledge relevance score, and the marketing goal fit score. The similarity scoring item calculates the maximum similarity between the candidate copy and the selected copy based on the semantic distance matrix; a greedy selection algorithm is used to screen different copy that meets the maximization marginal relevance criterion from the copy variants.

[0036] First, semantic clustering is performed on the input knowledge-enhanced marketing copy, grouping semantically similar copy into the same semantic cluster. For example, if the input marketing copy is about a new smartphone, the copy can be divided into different semantic clusters based on the copy's focus, such as "camera function," "performance," or "appearance design."

[0037] Then, within each semantic cluster, a conditional variational autoencoder (CVA) is used to generate multiple copy variants. The CVA consists of a content encoder, a semantic cluster encoder, and a decoder. The content encoder encodes the input copy into a mean vector and a variance vector, while the semantic cluster encoder encodes the semantic cluster features into a conditional vector. The mean and variance vectors are then reparameterized and sampled to generate latent variables. These latent variables are then fused with the conditional vector and fed into the decoder to generate copy variants. For example, the conditional vector for the "camera function" semantic cluster could be a combination of word vectors for keywords like "high pixel," "night shot," and "anti-shake." Assuming the mean vector is [0.5, 0.2, 0.8] and the variance vector is [0.1, 0.05, 0.2], sampling yields the latent variable [0.55, 0.22, 0.9]. Fusing this with the conditional vector and feeding it into the decoder generates a copy variant like "This phone boasts ultra-high pixel count, stunning night shots, and powerful anti-shake capabilities, allowing you to easily capture every moment."

[0038] Next, the semantic distance matrix between the copy variants is calculated. The BERT-based text representation model is used to extract contextual semantic features of the copy variants. Based on these features, lexical, structural, and topical differences are calculated. Lexical differences are calculated by calculating the word mover distance. For example, the lexical difference between "high-definition photography" and "ultra-clear photography" is small. Structural differences are calculated by calculating the syntactic dependency tree edit distance. For example, the structural difference between "this phone takes very clear photos" and "this phone that takes clear photos" is small. The topical difference is calculated by calculating the JS divergence of the topic distribution. For example, the topical difference between two copies that both emphasize the "night photography function" is small. The lexical, structural, and topical differences are weighted and fused to determine the semantic distance between the copy variants, ultimately forming a semantic distance matrix. For example, the semantic distance between copy A and copy B is 0.2, and the semantic distance between copy A and copy C is 0.8.

[0039] Next, a dynamic difference threshold is set based on the semantic distance matrix. The dynamic difference threshold is a weighted sum of the standard deviation of the semantic distance matrix and the base threshold. For example, if the standard deviation of the semantic distance matrix is 0.1, the base threshold is 0.5, and the weights are 0.3 and 0.7, respectively, then the dynamic difference threshold is 0.3 * 0.1 + 0.7 * 0.5 = 0.38.

[0040] A maximizing marginal relevance criterion was constructed, combining quality and similarity scoring. The quality scoring factor comprises a weighted combination of language fluency, knowledge relevance, and marketing goal alignment. For example, a copy with fluent and natural language, including product-related knowledge points, and appealing to target users would receive a high quality score. The similarity scoring factor calculates the maximum similarity between the candidate copy and the selected copy based on the semantic distance matrix.

[0041] Finally, a greedy selection algorithm is used to select differentiating copywriting from the copywriting variants that meets the criterion of maximizing marginal relevance. The copywriting with the highest quality score is selected first. Then, copywriting with a similarity below the dynamic difference threshold and the highest possible quality score is selected, until the preset number of copies is met or no copies meet the criteria. For example, assuming the dynamic difference threshold is 0.38, the similarities between selected copywriting A and candidate copies B and C and A are 0.2 and 0.8, respectively, and the quality scores of copies B and C are 0.9 and 0.7, respectively. Since copywriting B has a similarity below 0.38 and the highest quality score, copywriting B is selected.

[0042] Figure 2 This is a comparison chart of marketing copy quality assessment based on the conditional variational autoencoder according to an embodiment of the present invention: In this chart, the horizontal axis represents the experiment number (ranging from 0 to 5), and the vertical axis represents the overall marketing copy score (ranging from 60 to 100). Squares and circles are used to distinguish between these two shapes. Specifically, this technical solution, represented by a square in the figure, represents a copy generation method based on a combination of conditional variational autoencoders and knowledge augmentation. This method uses a content encoder to construct mean and variance vectors, and combines these with the conditional vectors to generate multiple copy variants. A greedy selection algorithm then selects copy with high linguistic fluency, knowledge relevance, and marketing goal alignment. The scores for this technical solution were 80, 85, 90, 88, 92, and 95, respectively, demonstrating a trend of continuous improvement and optimization. In the legend, squares represent this novel technical solution, highlighting its efficient optimization effect in marketing copy generation. On the other hand, the existing technologies, represented by circles in the figure, represent traditional marketing copy generation methods or template matching methods. These methods may rely solely on simpler language models or rule bases to generate copy, resulting in scores of only 70, 72, 75, 73, 76, and 78, with overall low scores and a small fluctuation range. In the legend, the existing technologies represented by circles reflect the limitations of traditional methods in terms of language quality and marketing fit. By comparing the two technical solutions, the specific values in the figure (such as 80 points versus 70 points in experiment number 0, and 95 points versus 78 points in experiment number 5) intuitively demonstrate the significant advantage of this technical solution in copy scoring.

[0043] Figure 3 This is a diagram of the semantic difference evaluation and comparative analysis of the embodiment of the present invention: The chart uses the horizontal axis to represent the experimental sample number (0 to 4) and the vertical axis to represent the calculated semantic distance value, ranging from 0.0 to 0.1. Two different icons are used for annotation in the figure: triangles are used to represent this technical solution, and diamonds are used to represent existing technologies. In this technology, this technical solution represents a method that uses the BERT model to extract text context semantic features and calculates the semantic distance through a weighted fusion of word movers, syntactic dependency tree edit distance, and topic distribution JS divergence. The figure shows that the data points of this technical solution are 0.05, 0.04, 0.03, 0.035 and 0.03 respectively. From the overall point of view, as the sample number increases, the semantic distance shows a gradual downward trend, reaching a minimum of 0.03, which shows that this solution can more accurately capture the subtle semantic relationships between texts. In the legend, the triangle clearly represents the advanced technology that uses deep semantic processing, conditional variational sampling and context fusion strategies. In contrast, existing technologies, such as those relying on traditional text processing methods or shallow feature extraction, have higher semantic distances of 0.08, 0.09, 0.085, 0.08, and 0.078, respectively, indicating significant semantic differences and lower text consistency. In the legend, the diamonds represent semantic similarity calculated by existing technologies, such as those based on simple statistics or traditional natural language processing. This chart, through specific numerical comparisons (e.g., the difference between 0.03 and 0.08), clearly demonstrates the significant advantages of adopting this technical solution in ensuring semantic consistency across copy variants and reducing redundancy.

[0044] Figure 4 This is a trend diagram of the copywriting diversity evaluation based on the maximization marginal relevance criterion according to an embodiment of the present invention: The horizontal axis in this chart also represents the experiment number (0 to 5), and the vertical axis represents the maximum marginal relevance score, ranging from 0.5 to 0.8. In this figure, the five-pointed star icon represents the present invention, while the cross icon represents the existing technology. The present invention in the figure represents an optimization strategy that integrates language fluency, knowledge relevance, and marketing goal fit scores. This strategy generates multiple copywriting variants using a conditional variational autoencoder and uses a greedy selection algorithm to select the copywriting that best meets the maximum marginal relevance criterion. The score gradually improves from 0.65, 0.68, 0.70, 0.72, and 0.75 to 0.78. This numerical improvement from 0.65 to 0.78 fully demonstrates the advantages of this new approach in optimizing copywriting performance by integrating multi-dimensional scoring. In the legend, the five-pointed star representing the present invention illustrates the practical application of this method's advanced model, dynamic difference threshold construction, and maximum marginal relevance criterion screening strategy. In contrast, the existing technology is represented by a cross in the figure, which represents the traditional scoring and screening method, which can only reach the range of 0.55, 0.57, 0.58, 0.6, 0.62, and 0.63. The overall score value is low and the improvement is small. In the legend, the existing technology marked with a cross usually refers to the process of re-screening copy variants based on traditional rules or simple statistical methods. The comparison of specific values in the figure, such as the gap between 0.72 and 0.6 and 0.78 and 0.63, clearly shows the significant technical improvement of this technical solution in maximizing marginal relevance screening and final copy optimization, which has a higher guarantee for the fit with marketing goals and the quality of copy.

[0045] The solution of this application can: Improving copywriting diversity: By performing variational sampling and differential screening within the semantic space, we generate copywriting combinations with richer semantics and expressions, avoiding copywriting homogeneity and reaching a wider user base. Ensuring copywriting quality: The method incorporates a quality scoring mechanism to ensure that the selected copywriting reaches a high level in terms of language fluency, knowledge relevance, and alignment with marketing goals, effectively improving marketing effectiveness. Improving screening efficiency: Using a greedy selection algorithm, while ensuring copywriting quality and differentiation, we quickly select the optimal combination from a large number of candidate copies, saving labor costs and improving marketing efficiency.

[0046] In an optional embodiment, the candidate marketing copy set is input into a preset multi-objective evaluation network to obtain an evaluation score matrix; based on the evaluation score matrix, the optimal marketing copy is screened using a Pareto optimization algorithm, including: The candidate marketing copy set is input into a preset multi-objective evaluation network for evaluation, and each marketing copy is scored for knowledge coverage, creativity, and expected effect.

[0047] The knowledge coverage score, the creativity score and the expected effect score are combined into an evaluation score matrix. Based on the evaluation score matrix, the candidate marketing copy is divided into multiple Pareto fronts using non-dominated sorting. The optimal copy is selected within the optimal Pareto front in combination with the crowding sorting, wherein the crowding sorting simultaneously considers the target space distance and the decision space distance. The target space distance is calculated based on the evaluation score matrix, and the decision space distance is calculated based on the semantic representation of the copy, so as to obtain the optimal marketing copy that is balanced and differentiated in multiple evaluation dimensions.

[0048] The multi-objective evaluation network includes modules for knowledge coverage assessment, creativity assessment, and expected effect assessment. The knowledge coverage assessment module uses semantic similarity to calculate the degree of match between marketing copy and relevant knowledge in the knowledge base. A weight coefficient is assigned to each knowledge point, with the higher the importance of the knowledge point, the greater the weight coefficient. The sum of the weighted scores of all matching knowledge points serves as the knowledge coverage score. For example, a marketing copy for a certain brand of mobile phone stated, "The ultra-large aperture main camera, combined with a new generation of image processing chips, allows for crystal-clear photos even in low-light conditions." The semantic similarity between this copy and the knowledge point "large aperture camera technology" in the knowledge base is 0.85, and the semantic similarity with the knowledge point "image processing chip" is 0.92. Given the weights of these two knowledge points being 0.6 and 0.4, respectively, the knowledge coverage score for this copy is 0.878.

[0049] The creativity assessment module evaluates the creative level of copywriting based on three dimensions: lexical novelty, structural uniqueness, and semantic innovation. Lexical novelty is determined by calculating the degree of difference between the copywriting's vocabulary and a commonly used marketing vocabulary library. Structural uniqueness is calculated based on the edit distance between the copywriting's syntactic dependency tree and a template library. Semantic innovation is determined by extracting copywriting features using a deep text representation model and calculating the semantic distance with a historical copywriting library. Taking the mobile phone copy mentioned above as an example, its lexical novelty score is 0.76, its structural uniqueness score is 0.82, and its semantic innovation score is 0.79. The weighted average gives a creativity score of 0.79.

[0050] The expected performance evaluation module is based on a pre-trained performance prediction model trained on large-scale historical marketing data. It takes as input the semantic features of the copy, the characteristics of the target audience, and the delivery scenario, and outputs predicted values for performance metrics such as click-through rate and conversion rate. For the example copy, the predicted click-through rate is 3.2% and the conversion rate is 0.8% for a young user group and social media delivery scenario, resulting in an expected performance score of 0.85.

[0051] After inputting a set of candidate copywriting into a multi-objective evaluation network, an evaluation score matrix is generated. Taking 10 candidate copywritings as an example, each copywriting is scored based on three dimensions: knowledge coverage, creativity, and expected effect, forming a 10×3 evaluation score matrix. Based on this matrix, a non-dominated sort is performed to determine the dominance relationship between the candidate copywritings. For any two copywritings, if the first copywriting is not inferior to the second copywriting in all evaluation dimensions and is superior to the second copywriting in at least one dimension, the first copywriting is considered to dominate the second copywriting.

[0052] We count the number of times each document is dominated by other documents and place documents that are never dominated in the first Pareto front. In this example, three documents are included in the first front, with scores of [0.878, 0.79, 0.85], [0.85, 0.83, 0.82], and [0.86, 0.78, 0.87], respectively. Within the first front, we calculate the crowding of documents, taking into account both target space distance and decision space distance. The target space distance is calculated based on the evaluation score, while the decision space distance is calculated by calculating the cosine distance after extracting the semantic representation of the document using the BERT model.

[0053] In an optional embodiment, dividing the candidate marketing copy into multiple Pareto fronts using non-dominated sorting based on the evaluation score matrix, and selecting the optimal copy within the optimal Pareto front in combination with the congestion sorting includes: performing non-dominated sorting on the candidate marketing copy based on the evaluation score matrix, counting the number of times each candidate marketing copy is dominated by other candidate marketing copies, and placing candidate marketing copies with zero dominated times into the first Pareto frontier; reducing the number of times the candidate marketing copy in the first Pareto frontier is dominated by the other candidate marketing copy, and placing the candidate marketing copy whose number of domination becomes zero after the update into the second Pareto frontier, and repeating the above process until all candidate marketing copies are placed into the corresponding Pareto frontier; For each candidate marketing copy in the first Pareto front, calculate the crowding distance of each candidate marketing copy, where the crowding distance is calculated based on the difference in scores of the candidate marketing copy and its adjacent candidate marketing copies on each evaluation indicator; The candidate marketing copy in the first Pareto front is sorted in descending order according to the crowding distance, and the candidate marketing copy with the largest crowding distance is selected as the optimal marketing copy.

[0054] When performing a non-dominant ranking of candidate marketing copy, the dominance relationship between the candidate copy is determined based on the evaluation score matrix. The evaluation score matrix contains the scores of each candidate copy across the three dimensions of knowledge coverage, creativity, and expected effect. To determine the dominance relationship, the scores of the two candidate copy to be compared across the three evaluation dimensions are compared one by one. When the first candidate copy scores at least as high as the second candidate copy across all evaluation dimensions and strictly exceeds the second candidate copy in at least one evaluation dimension, the first candidate copy is determined to dominate the second candidate copy.

[0055] After determining the dominance relationship, we count the total number of times each candidate is dominated by other candidates, i.e., the domination count. We then traverse all candidate copies and identify those with a domination count of zero. These candidates are not dominated by any other candidate copies and are placed on the first Pareto front. The candidate copies on the first Pareto front represent a set of copies that perform well across multiple evaluation dimensions.

[0056] For each candidate already included in the first Pareto front, the algorithm iterates through the other candidate proposals it dominates and reduces the domination count of each dominated candidate proposal by one. This updated domination count reflects the change in dominance between the remaining candidate proposals after removing the first Pareto front proposal. After updating the domination count, the algorithm examines all candidate proposals not already included in the front and assigns any candidate proposals with a domination count of zero after the update to the second Pareto front. This domination count update and frontier partitioning process is repeated until all candidate proposals are included in the corresponding Pareto front.

[0057] After obtaining the Pareto front partitioning results, focus on the candidate documents in the first Pareto front. The diversity of these candidate documents is assessed by calculating the crowding distance. This distance is calculated based on the difference in scores between the candidate document and its neighbors on each evaluation dimension. For each evaluation dimension, the candidate documents in the first Pareto front are sorted by their scores on that dimension, and the neighbors of the target candidate document on that dimension are identified. The difference in scores between the target candidate document and its neighbors on that dimension is calculated. A larger score difference indicates a more unique candidate document on that dimension.

[0058] When calculating the crowding distance, it's important to consider the dimensional differences across different evaluation dimensions. Normalization is used to convert the score differences across each evaluation dimension to the same scale. A weighted summation of these normalized score differences is used to calculate the crowding distance of a candidate copy. A larger crowding distance indicates a greater degree of difference between the candidate and other copywriters, and a higher degree of uniqueness.

[0059] Based on the calculated crowding distance, candidate copywriting on the first Pareto front is sorted in descending order. This crowding distance sorting considers the distribution of candidate copywriting across the evaluation metric space, helping to maintain diversity in the final selection results. The candidate copywriting with the largest crowding distance is selected as the optimal marketing copywriting. This copywriting not only exhibits balanced and excellent performance across multiple evaluation dimensions but also maintains significant differentiation from the other candidate copywriting.

[0060] By combining non-dominated sorting with crowding distance, this approach ensures both quality and diversity, providing a reliable technical solution for the screening of marketing copy. This method is applicable to copy evaluation and screening in various marketing scenarios, demonstrating its versatility and practical value.

[0061] In an optional implementation, taking the generation path of the optimal marketing copy as a positive sample, optimizing the knowledge retrieval index includes: Obtaining a generation path for an optimal marketing copy, the generation path including a search sequence of a knowledge search index during the generation of the optimal marketing copy and the knowledge content obtained during each search; Constructing the search sequence and the corresponding knowledge content in the generation path into positive sample training data, wherein the positive sample training data is used to characterize the knowledge search pattern in the process of generating high-quality marketing copy; The knowledge retrieval index is optimized based on the positive sample training data, and the retrieval priority of the knowledge content matching the retrieval sequence is improved by adjusting the association weights of the knowledge content in the knowledge retrieval index.

[0062] To determine the optimal marketing copy generation path, it's necessary to record the complete interaction with the knowledge retrieval index during the generation process. This path involves multiple rounds of retrieval operations, each of which includes the search keyword, search timestamp, search results, and their application in copy generation. For each piece of knowledge in the knowledge retrieval index, its knowledge number, knowledge type, knowledge text, associated tags, and current weight are recorded.

[0063] The search sequence reflects the order and logical relationships of knowledge acquisition during the copywriting process. By analyzing the search sequence, typical patterns of knowledge retrieval can be identified, such as the search path from basic product attributes to technological advantages and then to application scenarios. The knowledge content generated in each round of search operations is annotated with how it is applied in the final copywriting, including direct citation, synonymous paraphrase, combination and fusion, and other forms.

[0064] When converting the generated paths into positive training data, a mapping relationship between the search sequence and the knowledge content needs to be established. Semantic expansion is performed on each search term in the search sequence, and synonyms and near-synonyms are included in the search scope. Key concepts and attribute features of the retrieved knowledge content are extracted, and a correlation graph is constructed between the knowledge points. Based on this correlation graph, the coverage and depth of the knowledge retrieval are analyzed to assess the completeness and accuracy of the acquired knowledge.

[0065] Positive training data is annotated with templates to record search pattern information. This annotation includes elements such as search trigger conditions, search term selection strategies, knowledge content screening rules, and knowledge application methods. By analyzing the generation paths of multiple high-quality documents, common features are extracted to form standardized search pattern descriptions. These search patterns reflect best practices in knowledge application and can guide the optimization of knowledge retrieval indexing.

[0066] When optimizing knowledge retrieval indexes, we focus on adjusting the relevance weights of knowledge content. Weight adjustments follow the following principles: increasing the weight of knowledge content with high retrieval frequency and good application results; strengthening the connection strength between closely related knowledge content; and balancing the coverage of different types of knowledge. Weight adjustments are performed iteratively, and the degree of improvement in retrieval results is evaluated after each round of adjustments.

[0067] The association weight in the knowledge retrieval index is reflected in two aspects: the correlation between knowledge content and search terms, and the correlation between knowledge content. The higher the correlation, the higher the ranking position in the corresponding search scenario. By analyzing the search sequences in positive samples, we identify frequently occurring combinations of search terms and knowledge content and increase the association weight between them. For knowledge content pairs that frequently appear together in the generated paths, we increase the association weight between them.

[0068] Retrieval priority is improved through dynamic sorting. When performing a search, the relevance weight is used as a key ranking factor, which, along with the underlying relevance of the knowledge content, determines the order of the search results. Knowledge content that closely matches the current search sequence pattern is given a higher ranking boost. By optimizing the sorting strategy, commonly used knowledge retrieval paths in the process of generating high-quality content are more easily replicated.

[0069] Establish a knowledge retrieval effectiveness evaluation mechanism, verifying optimization results by comparing metrics such as search result relevance, knowledge coverage, and search path similarity before and after optimization. Continuously collect new, high-quality copywriting paths and enrich positive sample training data to achieve dynamic optimization and iterative updates of the knowledge retrieval index. This optimization method can effectively improve the accuracy and efficiency of knowledge retrieval, providing better knowledge support for marketing copywriting generation.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A large language model-driven intelligent marketing copywriting generation and effect evaluation method, characterized by: include: A multi-channel feature extraction network is used to perform hierarchical feature extraction on multimodal marketing data collected in real time to obtain a feature vector matrix; Constructing hierarchical entity nodes and relationship edges based on the feature vector matrix; inputting the hierarchical entity nodes and the relationship edges into a graph neural network to generate an initial cross-modal knowledge graph; performing semantic vectorization processing on the initial cross-modal knowledge graph using a contrastive learning method to obtain a knowledge graph vector representation, and establishing a knowledge retrieval index; receiving marketing demands, performing semantic analysis on the marketing demands based on a bidirectional attention mechanism to generate a query vector; and using the query vector to retrieve relevant knowledge graph vector representations in the knowledge retrieval index; Building a decoder network based on the knowledge graph vector representation, inputting the marketing demand into the decoder network, and generating an initial marketing copy; Verifying the semantic matching degree between the initial marketing copy and the knowledge graph vector representation to generate a knowledge-enhanced marketing copy; using a diversity sampling strategy to expand the knowledge-enhanced marketing copy to generate a set of candidate marketing copies; The candidate marketing copy set is input into a preset multi-objective evaluation network to obtain an evaluation score matrix; based on the evaluation score matrix, the optimal marketing copy is screened out using a Pareto optimization algorithm; the generation path of the optimal marketing copy is used as a positive sample to optimize the knowledge retrieval index.

2. The method according to claim 1, characterized in that Retrieving relevant knowledge graph vector representations in the knowledge retrieval index using the query vector; Constructing a decoder network based on the knowledge graph vector representation, inputting the marketing demand into the decoder network, and generating an initial marketing copy includes: A locality-sensitive hashing algorithm is used to partition the high-dimensional vector space into subspaces, construct a fast retrieval index of the knowledge graph vector, set a dynamic retrieval threshold based on cosine similarity calculation, and adaptively adjust the dynamic retrieval threshold using an annealing algorithm; the query vector is input into the fast retrieval index to obtain a set of first-order knowledge nodes, and a graph attention network is constructed based on the first-order knowledge node set. The attention weight distribution of adjacent nodes is calculated, and the relationship path with the highest attention weight is selected for multi-hop retrieval to obtain a multi-hop knowledge subgraph; Organizing the knowledge graph vectors in the multi-hop knowledge subgraph into a knowledge memory matrix, constructing a knowledge selection attention layer and a semantic fusion attention layer, wherein the knowledge selection attention layer obtains attention distribution based on the product operation of the decoding state and the knowledge memory matrix, and the semantic fusion attention layer fuses the knowledge representation and the decoding state through a feedforward neural network, adopts residual connection and layer normalization for information transfer, and generates a knowledge-enhanced decoder network; The marketing demand and the knowledge context vector are concatenated as the initial state of the knowledge-enhanced decoder network. The beam search algorithm is used to maintain the candidate sequence. The probability distribution of the next word is generated through decoder state update, vocabulary distribution calculation and knowledge constraints. A decoding strategy based on kernel sampling is introduced, and the sampling temperature parameters are dynamically adjusted according to the probability distribution. The generation process is controlled based on the length perception mechanism to output the initial marketing copy.

3. The method according to claim 1, characterized in that Verify the semantic matching degree between the initial marketing copy and the knowledge graph vector representation to generate a knowledge-enhanced marketing copy; adopt a diversity sampling strategy to expand the knowledge-enhanced marketing copy to generate a set of candidate marketing copies including: Obtain word-level semantic representations for the initial marketing copy through a character-enhanced word vector model, perform semantic dependency analysis on the word-level semantic representations to obtain phrase-level semantic representations, and fuse the word-level semantic representations and the phrase-level semantic representations to generate a hierarchical semantic representation; Calculating a soft alignment matrix between the hierarchical semantic representation and the knowledge graph vector representation based on a preset bidirectional interactive attention network, performing cross-modal information fusion based on the soft alignment matrix through a gating mechanism to obtain a fusion feature, calculating a matching score for the fusion feature, and generating a knowledge consistency verification result; Determining knowledge-missing regions and semantic deviation regions based on the matching scores in the knowledge consistency verification results, calculating deviation vectors of the knowledge-missing regions and the semantic deviation regions using contrastive learning, adjusting the content based on the deviation vectors through gradient guidance, and reorganizing the supplemented and adjusted content using a chapter structure template to generate knowledge-enhanced marketing copy; The knowledge-enhanced marketing copy is variationally sampled in each semantic cluster through a conditional variational autoencoder to obtain copy variants, and the semantic distance matrix between the copy variants is calculated; a dynamic difference threshold is set based on the semantic distance matrix to construct a maximization marginal relevance criterion, and a greedy selection algorithm is used to screen differential copy that meets the maximization marginal relevance criterion from the copy variants; the differential copy is input into a multi-objective optimization function for comprehensive evaluation, and the copy that meets the evaluation requirements is selected to form a candidate marketing copy set.

4. The method according to claim 3, characterized in that The knowledge-enhanced marketing copy is sampled through a conditional variational autoencoder in each semantic cluster to obtain copy variants, and a semantic distance matrix between the copy variants is calculated; a dynamic difference threshold is set based on the semantic distance matrix to construct a maximum marginal relevance criterion, and a greedy selection algorithm is used to select different copies that meet the maximum marginal relevance criterion from the copy variants, including: The knowledge-enhanced marketing copy is subjected to variational sampling in each semantic cluster through a conditional variational autoencoder. The conditional variational autoencoder inputs the knowledge-enhanced marketing copy into a content encoder to obtain a mean vector and a variance vector, inputs the corresponding semantic cluster features into a semantic cluster encoder to obtain a conditional vector, performs reparameterized sampling based on the mean vector and the variance vector to obtain latent variables, and fuses the latent variables with the conditional vector to obtain multiple copy variants through a decoder; Extracting contextual semantic features of the copy variants using a BERT-based text representation model, calculating word mover distances based on the contextual semantic features to obtain lexical difference values, calculating syntactic dependency tree edit distances to obtain structural difference values, calculating topic distribution JS divergence to obtain topic difference values, performing weighted fusion on the lexical difference values, the structural difference values, and the topic difference values to calculate semantic distances between the copy variants, and obtaining a semantic distance matrix; A dynamic difference threshold is set based on the weighted sum of the standard deviation of the semantic distance matrix and a basic threshold, and a maximization marginal relevance criterion is constructed that integrates the quality scoring item and the similarity scoring item. The quality scoring item includes a weighted combination of the language fluency score, the knowledge relevance score, and the marketing goal fit score. The similarity scoring item calculates the maximum similarity between the candidate copy and the selected copy based on the semantic distance matrix; a greedy selection algorithm is used to screen different copy that meets the maximization marginal relevance criterion from the copy variants.

5. The method according to claim 1, wherein Inputting the candidate marketing copy set into a preset multi-objective evaluation network to obtain an evaluation score matrix; Based on the evaluation score matrix, the Pareto optimization algorithm is used to screen out the optimal marketing copy, including: Input the candidate marketing copy set into a preset multi-objective evaluation network for evaluation, and perform a knowledge coverage score, a creativity score, and an expected effect score on each marketing copy; The knowledge coverage score, the creativity score and the expected effect score are combined into an evaluation score matrix. Based on the evaluation score matrix, the candidate marketing copy is divided into multiple Pareto fronts using non-dominated sorting. The optimal copy is selected within the optimal Pareto front in combination with the crowding sorting, wherein the crowding sorting simultaneously considers the target space distance and the decision space distance. The target space distance is calculated based on the evaluation score matrix, and the decision space distance is calculated based on the semantic representation of the copy, so as to obtain the optimal marketing copy that is balanced and differentiated in multiple evaluation dimensions.

6. The method according to claim 5, characterized in that Dividing the candidate marketing copy into multiple Pareto fronts using non-dominated sorting based on the evaluation score matrix, and selecting the optimal copy within the optimal Pareto front in combination with the congestion sorting includes: performing non-dominated sorting on the candidate marketing copy based on the evaluation score matrix, counting the number of times each candidate marketing copy is dominated by other candidate marketing copies, and placing candidate marketing copies with zero dominated times into the first Pareto frontier; reducing the number of times the candidate marketing copy in the first Pareto frontier is dominated by the other candidate marketing copy, and placing the candidate marketing copy whose number of domination becomes zero after the update into the second Pareto frontier, and repeating the above process until all candidate marketing copies are placed into the corresponding Pareto frontier; For each candidate marketing copy in the first Pareto front, calculate the crowding distance of each candidate marketing copy, where the crowding distance is calculated based on the difference in scores of the candidate marketing copy and its adjacent candidate marketing copies on each evaluation indicator; The candidate marketing copy in the first Pareto front is sorted in descending order according to the crowding distance, and the candidate marketing copy with the largest crowding distance is selected as the optimal marketing copy.

7. The method according to claim 1, characterized in that Taking the generation path of the optimal marketing copy as a positive sample, optimizing the knowledge retrieval index includes: Obtaining a generation path for an optimal marketing copy, the generation path including a search sequence of a knowledge search index during the generation of the optimal marketing copy and the knowledge content obtained during each search; Constructing the search sequence and the corresponding knowledge content in the generation path into positive sample training data, wherein the positive sample training data is used to characterize the knowledge search pattern in the process of generating high-quality marketing copy; The knowledge retrieval index is optimized based on the positive sample training data, and the retrieval priority of the knowledge content matching the retrieval sequence is improved by adjusting the association weights of the knowledge content in the knowledge retrieval index.

Images