Marketing document automatic generation method based on deep learning

Abstract

The invention discloses a marketing copywriting automatic generation method based on deep learning, and belongs to the technical field of marketing copywriting generation. The method comprises the following steps: extracting visual features, functional attributes and text description of a product through a multi-modal fusion model to obtain a comprehensive product representation vector; calculating weight distribution among the features by adopting an attention mechanism to obtain a weighted feature set; mapping the features to a brand style inheritance space through an embedded layer to obtain an embedded representation with a consistent style; carrying out sequence decoding on the basis to generate a preliminary copywriting draft; filtering the efficacy selling point phrases through an operation strategy module to form refined document fragments; and carrying out matching verification by adopting an iterative optimization loop, carrying out backtracking adjustment through consistency verification, and finally generating a marketing copywriting conforming to a brand style and an operation target.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of marketing copy generation, and particularly relates to a marketing copy automatic generation method based on deep learning. BACKGROUND

[0002] As the core driving force of digital business, marketing copy creation directly affects brand communication effectiveness and business conversion rate, and its importance is increasingly prominent in the e-commerce era. With the rapid growth of consumers' demand for personalized content, traditional marketing copy generation methods are facing unprecedented challenges. The current copy generation technology generally has the fundamental defect of insufficient information understanding dimension. Existing methods can usually only handle single type of product information and cannot simultaneously deeply understand the internal correlation between the visual features, functional attributes and textual descriptions of the product. This limitation of information understanding directly leads to the problem of brand style inheritance breaking. Due to the lack of comprehensive grasp of multi-dimensional product information, the system cannot accurately capture and continue the unique expression and language style of the brand, resulting in obvious inconsistency in the brand tone of the generated copy. For example, when generating promotion copy for a new product on an e-commerce platform, if the system only bases on the text description and ignores the visual elements of the product packaging and the brand's established color language, it will produce copy content that does not conform to the brand image.

[0003] The lack of brand style inheritance further leads to the technical problem of insufficient controllability in the generation process. When the system cannot accurately understand the complete information of the product and maintain brand consistency, the operation personnel lose the ability to effectively guide the direction of copy generation. For example, when a high-end skin care brand needs to generate copy for a new product, the system not only needs to understand the delicate visual elements of the product packaging, but also needs to grasp the brand's consistent elegant expression style, and at the same time needs to highlight specific efficacy selling points according to the operation strategy, but existing technologies often produce style confusion or deviate from the expected copy content under such complex multiple constraints. This technical limitation makes it difficult for operation personnel to ensure that the generated copy can reflect the product characteristics, comply with the brand tone, and meet the immediate marketing strategy needs through system settings. Therefore, how to accurately inherit the brand style and precisely control the generation process on the basis of deeply integrating multi-dimensional product information has become a key problem that needs to be broken through in the field of intelligent creation of marketing copy. SUMMARY

[0004] To solve the above technical problems, the application provides a marketing copy automatic generation method based on deep learning to solve the problems existing in the prior art.

[0005] In the first aspect, to achieve the above object, the application provides a marketing copy automatic generation method based on deep learning, comprising the following steps: S1, extracting visual features, functional attributes and textual descriptions from product data through a multi-modal fusion model to obtain a comprehensive product representation vector; S2, calculating a weight distribution between the visual features and the functional attributes according to the comprehensive product representation vector using an attention mechanism to determine a weighted feature set; S3, if the similarity of the weighted feature set exceeds a preset threshold, mapping to a brand style inheritance space through an embedding layer to obtain an embedded representation consistent in style; S4, after obtaining the embedded representation consistent in style, performing sequence decoding to determine a continuous sequence of language style continuation to obtain a preliminary draft of the script; S5, filtering phrases related to the emphasis of the selling points from the preliminary draft of the script through an operation strategy highlighting module to determine a refined script segment; S6, matching and verifying the refined script segment with multi-dimensional information using an iterative optimization cycle to obtain a complete script output; S7, performing consistency verification on the complete script output, and if the verification result shows deviation, backtracking to S2 to adjust the weight distribution.

[0006] In a second aspect, the application further provides a marketing script automatic generation system based on deep learning, which is used to implement a marketing script automatic generation method based on deep learning, and the system comprises: a multi-modal fusion module for extracting visual features, functional attributes and textual descriptions from product data to obtain a comprehensive product representation vector; an attention weighting module for calculating a weight distribution between the visual features and the functional attributes according to the comprehensive product representation vector using an attention mechanism to determine a weighted feature set; a brand style embedding module for, when the similarity of the weighted feature set exceeds a preset threshold, mapping to a brand style inheritance space through an embedding layer to obtain an embedded representation consistent in style; a sequence decoding module for, after obtaining the embedded representation consistent in style, performing sequence decoding to determine a continuous sequence of language style continuation to obtain a preliminary draft of the script; an operation strategy highlighting module for filtering phrases related to the emphasis of the selling points from the preliminary draft of the script to determine a refined script segment; an iterative optimization module for matching and verifying the refined script segment with multi-dimensional information using an iterative optimization cycle to obtain a complete script output; a consistency verification module for performing consistency verification on the complete script output, and when the verification result shows deviation, triggering backtracking to the attention weighting module to adjust the weight distribution.

[0007] In a third aspect, the present application further provides a computer terminal device, comprising: one or more processors; a memory coupled to the processors, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the deep learning-based marketing copy automatic generation method in the first aspect.

[0008] In a fourth aspect, the present application further provides a computer readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the deep learning-based marketing copy automatic generation method in the first aspect.

[0009] In a fifth aspect, the present application further provides a computer program product comprising a computer program, wherein the computer program, when executed by a processor, implements the steps of the deep learning-based marketing copy automatic generation method in the first aspect.

[0010] Compared with the prior art, the present application has the following advantages and technical effects: The deep learning-based marketing copy automatic generation method provided by the present application comprehensively extracts and represents product visual features, functional attributes and textual descriptions through a multi-modal fusion model, forming a comprehensive vector that can deeply reflect the internal information of the product; based on an attention mechanism, the weight distribution between visual features and functional attributes is calculated to strengthen the relevance of key features and generate a weighted feature set with internal logic; through an embedding layer, the weighted features are mapped to a brand style inheritance space to ensure that the generated copy is consistent with the brand historical content in terms of language style; sequence decoding technology is used to capture the continuity of language style, generate a preliminary copy draft, and combine the operation strategy to highlight the filtering and reconstruction of the efficacy selling points, forming a refined copy that meets the marketing orientation; through an iterative optimization cycle, the copy content is subjected to multi-dimensional matching verification and style consistency evaluation, and when a deviation is detected, the feature weight is adjusted, and finally a complete marketing copy with accurate content, unified style and consistent operation target is output. BRIEF DESCRIPTION OF DRAWINGS

[0011] The accompanying drawings, which form a part of the present application, are intended to provide further understanding of the present application, and the illustrative embodiments of the present application and their description serve the purpose of explaining the present application. The accompanying drawings should not be construed as an improper limitation of the present application. In the drawings: Figure 1 A flowchart of the deep learning-based marketing copy automatic generation method of the embodiment of the present application; Figure 2 A feature weighting and brand style embedding schematic diagram of the embodiment of the present application; Figure 3This is a schematic diagram illustrating the text generation, optimization, and verification process in an embodiment of the present invention. Detailed Implementation

[0012] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0013] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0014] Example 1 like Figure 1 As shown, this embodiment provides a method for automatically generating marketing copy based on deep learning, including: S1. Visual features, functional attributes, and textual descriptions are extracted from product data using a multimodal fusion model to obtain a comprehensive product representation vector; S2. Based on the comprehensive product representation vector, use an attention mechanism to calculate the weight distribution between visual features and functional attributes, and determine the weighted feature set; S3. If the similarity of the weighted feature set exceeds a preset threshold, then the embedding layer is mapped to the brand style inheritance space to obtain a style-consistent embedding representation. S4. After obtaining the style-consistent embedded representation, perform sequence decoding to determine the continuous sequence of language style to obtain a preliminary draft of the text. S5. Filter relevant phrases emphasizing efficacy and selling points from the preliminary copy draft through the operational strategy highlighting module, and determine the refined copy fragments; S6. Use an iterative optimization loop to match and verify the refined copy fragments with multi-dimensional information to obtain complete copy output; S7. Perform a consistency check on the complete text output. If the check result shows a deviation, backtrack to S2 to adjust the weight distribution.

[0015] As one implementation method in this embodiment, in S1, the process of extracting visual features, functional attributes, and textual descriptions from product data through a multimodal fusion model includes: Visual feature vectors are obtained by extracting color distribution features and texture features from product images using convolutional neural networks. Natural language processing tools are used to perform lexical segmentation and semantic parsing on the product text description to obtain text feature vectors. The visual feature vector and the text feature vector are merged according to a preset weight ratio by a feature fusion algorithm to obtain a comprehensive feature vector. The effectiveness of the comprehensive feature vector is verified by cosine similarity calculation, and if the similarity result meets the preset range, it is saved as the comprehensive product representation vector.

[0016] Step S101, the visual features, functional attributes and text descriptions are extracted from the product data by a pre-established multi-modal fusion model to obtain a comprehensive product representation vector.

[0017] 1. Color distribution features and texture features are extracted from product database images by a convolutional neural network. If the image resolution is lower than the preset threshold, an interpolation algorithm is used to improve image quality, and the visual feature vector group is obtained. 2. The product text description is segmented and semantically analyzed by a natural language processing tool. If the text length exceeds the preset character limit, the key description paragraph is truncated, and the text feature vector is obtained. 3. The visual feature vector and the text feature vector are merged according to a preset weight ratio by a feature fusion algorithm. If the dimensions do not match, adjust through a dimension transformation matrix to determine the comprehensive feature vector. 4. The effectiveness of the comprehensive feature vector is verified by cosine similarity calculation, and if the similarity result meets the preset range, it is saved to obtain the final product representation vector.

[0018] For example, when processing clothing product images on e-commerce platforms, color distribution features and texture features can be extracted by a convolutional neural network.

[0019] Specifically, the convolutional neural network is a deep learning model that captures local patterns in images through multiple convolutional and pooling layers. For example, first use a convolution kernel to scan the image to capture edges, and then gradually abstract high-level features such as color histogram representation distribution and Gabor filter extracted texture details. If the image resolution is lower than the preset threshold, such as 512x512 pixels, a bilinear interpolation algorithm is used to improve the quality, and the'red' color feature and'silk' texture feature extracted will be integrated into the subsequent copy generation as key information, such as to generate a descriptive phrase like'surprising red, showing silky smooth texture'.

[0020] In an embodiment, for the processing of product text description, natural language processing tools such as BERT model can be used for lexical segmentation and semantic parsing. Lexical segmentation is the process of splitting text into words or phrases, for example, identifying keywords in “red silk dress” through tokenization algorithms, while semantic parsing involves understanding contextual meanings, such as calculating word vectors using attention mechanisms. If the text length exceeds a preset character limit, such as 500 characters, key description paragraphs are truncated, for example, by prioritizing sections containing product specifications and materials, resulting in a text feature vector, which effectively reduces data processing volume and avoids information redundancy in business.

[0021] For example, continuing the scenario of the above-mentioned clothing product, the visual feature vector and the text feature vector are combined according to a preset weight ratio, such as 0.6:0.4, using a feature fusion algorithm.

[0022] Specifically, the fusion algorithm can be a simple linear combination, multiplying the vector elements by the weights and then adding them together. If the dimensions do not match, such as a 512-dimensional visual vector and a 768-dimensional text, a dimension transformation matrix, such as principal component analysis, is used to adjust to a unified dimension to determine the comprehensive feature vector, which can integrate multi-modal information and improve the comprehensiveness of product search in business.

[0023] In an embodiment, the effectiveness of the comprehensive feature vector is verified by cosine similarity calculation, for example, calculating the cosine value between two product vectors, and if the result is within a preset range of 0.8 to 1.0, it is saved to obtain the final product representation vector, which ensures the reliability and consistency of the vector.

[0024] As an implementation of the embodiment, in S2, the process of determining the weighted feature set according to the comprehensive product representation vector using an attention mechanism to calculate the weight distribution between visual features and functional attributes includes: extracting visual feature data and functional attribute data from the comprehensive product representation vector; calculating the weight distribution between the visual features and the functional attributes through an attention mechanism to obtain a set of weighted feature pairs; grouping the set of weighted feature pairs using a clustering algorithm to determine a feature association set; for each group of feature pairs in the feature association set, obtaining an intrinsic association strength value, and if the association strength value exceeds a preset threshold, determining that the group of feature pairs has effective association; performing context logic analysis on the feature association set with effective association using an attribute matching method to generate the weighted feature set.

[0025] Step S102: Based on the comprehensive product representation vector, use the attention mechanism to calculate the weight distribution between visual features and functional attributes, and determine the weighted feature set after grasping the intrinsic relationship.

[0026] like Figure 2 As shown, visual feature data and functional attribute data are obtained from the comprehensive product representation vector. The weight distribution between the visual features and functional attributes is calculated using an attention mechanism to obtain a weighted feature pair set. Data fusion processing is performed on the weighted feature pair set, and a clustering algorithm is used to group the feature pairs to determine the grouped feature association set. For each feature pair in the feature association set, the intrinsic association strength value is obtained. If the association strength value exceeds a preset threshold, the feature pair is determined to have effective association. Contextual logic analysis is performed on the feature association set with effective association using an attribute matching method to generate the final weighted feature set, resulting in the comprehensive feature vector of the product representation.

[0027] For example, when processing the product representation vector of a high-end skincare product, visual feature data, such as the 'golden hue' and 'streamlined bottle design' of the packaging, and functional attribute data, such as '24-hour long-lasting hydration' and 'repairing the skin barrier', are first extracted from the comprehensive vector. By calculating the weight distribution among these features using an attention mechanism, it is found that the visual feature of 'golden hue' is highly associated with the brand style description of 'luxury and exclusivity,' while the functional attribute of 'repairing the skin barrier' is strongly correlated with the copywriting phrase 'restores healthy radiance.' This weight allocation ensures that these key feature pairs with marketing value are highlighted when generating copy.

[0028] For example, by pairing the visual feature of screen resolution with the functional attribute of processor speed, the mechanism assigns weights based on relevance. For instance, if a high-resolution screen is often associated with a powerful processor, it receives a higher weight, thus forming multiple weighted feature pairs. For example, the weight of the "high resolution - high-efficiency processor" pair is 0.8, ultimately resulting in a set of weighted feature pairs. This process involves the principle of attention mechanisms, which normalize the attention scores between features using a softmax function, emphasizing more relevant pairings while ignoring weakly correlated terms.

[0029] In one possible implementation, the set of weighted features is then subjected to data fusion processing.

[0030] For example, the "color distribution-battery capacity" pair and the "texture feature-processor speed" pair of a smartphone can be fused together, and the K-means clustering algorithm can be used to group them.

[0031] Specifically, the clustering algorithm first initializes centroids, then iteratively calculates the Euclidean distance from each feature pair to the centroids, assigning similar pairs to the same group. For example, all durable-related pairs are clustered into one group, while another group focuses on performance-related pairs, thus determining the set of associated features after grouping. This grouping helps capture potential patterns between features, and the algorithm's principle lies in minimizing within-group variance to ensure that feature pairs within each set are highly similar.

[0032] For example, the process of obtaining the intrinsic correlation strength value for each feature pair in the feature association set can be implemented as follows: for the "body texture-battery capacity" pair in the durability group of smartphones, the intensity value is calculated using the Pearson correlation coefficient. If the result is 0.75 and exceeds the preset threshold of 0.6, the group is determined to have a valid correlation.

[0033] It should be noted that the intrinsic correlation strength value is obtained based on statistical methods. The principle is to quantify the linear relationship strength by dividing the covariance by the standard deviation. This ensures that only strongly correlated pairs are retained. The threshold setting depends on domain experience. For example, 0.6 represents a significant correlation in electronic products.

[0034] In one possible implementation, the attribute matching method is finally used to perform contextual logic analysis on the feature association sets of these valid associations.

[0035] For example, in the context of smartphones, feature pairs from the durability group are matched with those from the performance group. The analysis logic, such as matching a high-efficiency processor to balance power consumption if the battery capacity is high, generates the final weighted feature set, thereby obtaining a comprehensive feature vector representing the product.

[0036] Specifically, the principle of the attribute matching method involves rule-based reasoning, checking the compatibility between sets one by one, and adjusting the weights to form a cohesive vector. This analysis ensures the contextual consistency of the feature set.

[0037] As one implementation method in this embodiment, in S3, the process of mapping the similarity of the weighted feature set to the brand style inheritance space through the embedding layer to obtain a style-consistent embedding representation if the similarity exceeds a preset threshold includes: Obtain the visual element feature vector of the target brand, wherein the visual elements include brand logo, color scheme, font style and pattern texture; The cosine similarity calculation method is used to quantitatively evaluate the similarity between the visual element feature vector and the brand benchmark feature. If the similarity value exceeds a preset threshold, the feature matching result is obtained. The feature matching results are mapped to the brand style inheritance space through a neural network embedding layer to obtain an embedding representation vector. Brand style consistency constraints are constructed based on the embedded representation vector, and the style-consistent embedded representation is obtained through vector normalization.

[0038] Step S103: If the similarity of the weighted feature set exceeds a preset threshold, then the embedding layer is mapped to the brand style inheritance space to obtain a style-consistent embedding representation.

[0039] The visual element feature vectors of the target brand are obtained. A feature extractor is used to numerically process the brand logo, color scheme, font style, and pattern texture to generate a first weighted feature set containing weight coefficients. A cosine similarity calculation method is used to quantitatively evaluate the similarity of the first weighted feature set. If the similarity value between the current design scheme and the brand's baseline features exceeds a preset threshold, a second feature matching result that meets the brand inheritance requirements is obtained. The second feature matching result is mapped to the brand style inheritance space through a neural network embedding layer. The feature vectors are then spatially transformed according to the brand coding rules to obtain a first embedded representation vector with brand gene identification. Brand style consistency constraints are constructed based on the first embedded representation vector. Vector normalization is used to determine the final brand style inheritance representation, resulting in a second standardized embedded representation with consistent style.

[0040] For example, when processing the visual elements of a well-known coffee brand, the first step is to obtain the feature vector of its visual elements.

[0041] Specifically, a feature extractor quantifies brand logo elements such as the classic green mermaid pattern, color schemes such as the dark green and white main color scheme, font styles such as rounded sans-serif fonts, and texture patterns such as wavy coffee bean textures. This process involves inputting image data into a pre-trained convolutional neural network model, such as using a ResNet architecture to extract high-dimensional features, and then assigning weight coefficients to each element, for example, logo weight 0.4, color weight 0.3, font weight 0.2, and texture weight 0.1, thereby generating a first weighted feature set. These weights are trained based on historical brand data to ensure that the set captures the brand's visual core.

[0042] In one possible implementation, the cosine similarity method is then used to quantitatively evaluate the similarity of this first weighted feature set. Cosine similarity measures similarity by calculating the cosine of the angle between two vectors; the formula simplifies to the vector dot product divided by the product of their magnitudes. Specifically, the weighted feature vectors of the current design scheme are compared with the brand's baseline feature vectors, such as vectors from the brand's standard manual. If the calculated similarity value, such as 0.85, exceeds the preset threshold of 0.7, it is determined to meet the brand heritage requirements, thus obtaining the second feature matching result. This result includes a list of matched sub-features, helping designers quickly verify the consistency of the scheme.

[0043] For example, continuing with the coffee brand example, the second feature matching result is mapped to the brand style heritage space through a neural network embedding layer. This embedding layer is typically a multilayer perceptron structure, used to reduce the dimensionality of high-dimensional features and project them into a custom space. Based on brand encoding rules, such as defining style dimensions as traditional, innovative, and warm, the feature vectors undergo spatial transformation, for example, by adjusting the vector coordinates through matrix multiplication, to obtain the first embedded representation vector containing the brand's genetic identity. This vector not only preserves the original visual information but also incorporates the brand's historical evolution encoding, ensuring that subsequent processing reflects its heritage.

[0044] In one possible implementation, brand style consistency constraints are constructed based on the first embedded representation vector. This involves setting an optimization objective, such as minimizing the distance between the vector and the ideal style vector, and then determining the final brand style inheritance representation by normalizing the vector, i.e., dividing the vector by its L2 norm to make its length 1, thus obtaining a second normalized embedded representation with consistent style.

[0045] For example, in actual business operations, this representation can be used to generate new product packaging designs, ensuring that the new design is visually consistent with the brand's DNA, thereby enhancing market recognition and maintaining brand loyalty.

[0046] As one implementation method in this embodiment, in S4, the process of obtaining a style-consistent embedded representation, performing sequence decoding, determining a continuous sequence of language style continuation, and obtaining a preliminary draft includes: The input text is segmented using a pre-defined word vector library. The weight distribution of each word is calculated using a word frequency statistical method. The embedding vector matrix is ​​then calculated based on semantic similarity. If the semantic similarity of the word embedding vectors in the embedding vector matrix exceeds a preset threshold, the syntactic structure recognition program is activated to extract features of the text's rhetorical devices and obtain grammatical pattern tags. Temporal modeling is performed based on the grammatical pattern markers, and the probability distribution of word selection at the current position is determined by the preceding context information to identify the candidate word sequence. The candidate word sequence is received and its content logic consistency is checked. If the semantic correlation between adjacent words in the sequence meets the preset conditions, the sequence is marked as a valid segment and spliced ​​together to generate the preliminary draft.

[0047] Step S104: After obtaining the style-consistent embedded representation, perform sequence decoding on the expression mode capture, determine the continuous sequence of language style continuation, and obtain the preliminary copy draft.

[0048] The system receives input text and performs word segmentation using a preset word vector library. The word segmentation process uses a word frequency statistical method to calculate the weight distribution of each word and calculates a first embedding vector matrix based on semantic similarity. If the semantic similarity of the word embedding vectors in the first embedding vector matrix exceeds a preset threshold, a syntactic structure recognition program is initiated to extract features of the text's rhetorical devices and obtain grammatical pattern tags for the expressions. Temporal modeling is performed based on the grammatical pattern tags, and the probability distribution of word selection at the current position is determined by using the preceding context information to identify candidate word sequences that conform to the language continuity rules. The candidate word sequences are received and their content logic consistency is checked. If the semantic correlation between adjacent words in the sequence meets a preset condition, the sequence is marked as a valid segment and spliced ​​together to generate a preliminary draft.

[0049] For example, when processing input text about brand promotion, the system first receives short phrases such as "Explore the beauty of nature, embrace green living," and then performs word segmentation using a pre-defined word vector library, such as the Word2Vec model. This word vector library is essentially a set of pre-trained vector representations that can map words into a high-dimensional space.

[0050] Specifically, word segmentation breaks down the text into lexical units, such as "explore," "nature," "beauty," "embrace," "green," and "life." Then, a word frequency (TF-IDF) method is used to calculate the weight distribution of each word. This TF-IDF method assigns weights by calculating the frequency of a word in a document and its inverse document frequency in the entire corpus. For example, "green" has a high frequency in environmentally themed texts and might have a weight of 0.8, while "explore," being a common word, might have a lower weight of 0.3. Subsequently, a first embedding vector matrix is ​​calculated based on semantic similarity. This semantic similarity typically uses the cosine similarity formula, quantifying the similarity by comparing the angle between vectors. For example, if the similarity between the vectors of "nature" and "green" is 0.75, it indicates that they are semantically related. The final result is a matrix where each row represents the embedding vector of a word, with the matrix dimension being the number of words multiplied by the vector dimension.

[0051] In one embodiment, if the semantic similarity of the word embedding vectors in the first embedding vector matrix exceeds a preset threshold, such as 0.7, the system will initiate a syntactic structure recognition program to extract features of the text's rhetorical devices. This syntactic structure recognition program can be a dependency parsing-based tool, such as Stanford Parser, which analyzes the subject-verb-object structure and modification relations in the sentence to extract grammatical pattern markers of the expressions.

[0052] For example, in the input text “Explore the beauty of nature and embrace green living”, the program identifies “explore” as a parallel structure dominated by a verb, marks rhetorical devices such as metaphor or parallelism, and obtains pattern tags such as “verb-object + parallel”.

[0053] For example, continuing to perform temporal modeling based on the aforementioned tags is achieved through models such as LSTM networks, which use preceding context information to determine the probability distribution of word selection at the current position.

[0054] Specifically, temporal modeling inputs grammatical pattern tags into the sequence model. The model remembers the preceding word sequence. For example, when generating a continuation, if the preceding text is "explore nature", the probability of the next position "the beauty" is calculated to be 0.6, while the probability of alternatives such as "journey" is 0.4. This determines the candidate word sequence that conforms to the language continuation rule, such as "the beauty, embrace green life" as the priority sequence.

[0055] In one embodiment, after receiving the candidate word sequence, the content logic consistency check is performed by checking the semantic relevance of adjacent words. Here, the semantic relevance can be calculated based on the dot product of the embedded vectors. If the preset condition is met, such as the relevance is greater than 0.5, it is marked as a valid fragment and spliced ​​to generate a preliminary draft of the copy.

[0056] For example, in brand promotion, if the sequence "Embrace Green Living" is highly related to the preceding text, the system will piece together a complete draft, "Explore the beauty of nature, embrace green living, and embark on a sustainable journey." This helps ensure that the copy is logically coherent and supports the continuous dissemination of brand information.

[0057] As one implementation method in this embodiment, in S5, the process of filtering relevant phrases emphasizing efficacy and selling points from the preliminary copy draft through the operational strategy highlighting module to determine the refined copy fragments includes: The initial draft copy is segmented and semantically annotated using natural language processing tools. If the copy contains product efficacy-related words, semantic feature vectors are extracted. Core selling point phrases are matched according to a preset module filtering rule library to obtain a candidate phrase set. The matching degree between the candidate phrase set and the operational strategy keyword library is evaluated using a text similarity calculation method. If the similarity exceeds a preset threshold, the phrase is judged to meet the operational orientation requirements, and the importance ranking of each phrase in the original text is determined. Based on the marketing conversion rules, the high-weight phrases in the importance ranking are combined and reconstructed. If the semantic integrity score of the phrase combination reaches the standard value, the combination is retained. A syntactic analyzer is used to perform grammatical checks on the reconstructed phrases to obtain refined copywriting candidate fragments. The refined copywriting candidate segments are scored using a content quality assessment algorithm to predict their marketing effectiveness. If the score meets the conversion target parameters set in the operation strategy, the segment is determined to be the refined copywriting segment.

[0058] Step S105: Filter relevant phrases emphasizing efficacy and selling points from the initial draft copy through the operational strategy highlighting module, and determine the refined copy fragments under process control.

[0059] The system receives a preliminary draft of the copy and performs word segmentation and semantic annotation on it using natural language processing tools. If the copy contains product efficacy-related words, semantic feature vectors are extracted. Core selling point phrases are matched against a pre-defined module filtering rule base to obtain a first candidate phrase set with semantic weights. A text similarity calculation method is used to evaluate the matching degree between the first candidate phrase set and the operational strategy keyword library. If the similarity exceeds a preset threshold, the phrase is determined to meet operational guidance requirements. The importance ranking of each phrase in the original copy is determined, resulting in a second candidate phrase set. High-weight phrases in the second candidate phrase set are combined and reconstructed according to marketing conversion rules. If the semantic integrity score of the combined phrases reaches a standard value, the combination is retained. A syntactic analyzer is used to perform grammatical checks on the reconstructed phrases to obtain refined copy candidate fragments. A content quality assessment algorithm is used to predict the marketing effect of the candidate fragments. If the score meets the conversion target parameters set in the operational strategy, the fragment is determined as the final output content.

[0060] For example, when processing initial draft copy, the first step is to use natural language processing tools for word segmentation and semantic annotation. Suppose the copy describes a skin cream, such as "This cream deeply moisturizes the skin, improves dryness, and enhances radiance." The tool will segment the text into words like "deeply moisturizes," "skin," "improves dryness," and "enhances radiance," and then annotate its semantics, for example, labeling "deeply moisturizes" as an efficacy category. If the copy contains product efficacy-related words, such as "improves dryness," its semantic feature vector is extracted. This involves using a pre-trained model like BERT to generate vector representations that capture the semantic context of the words. Next, core selling point phrases are matched against a pre-defined module filtering rule base. This rule base might include keywords like "moisturizes" and "radiance." Matching is calculated using cosine similarity to obtain a first set of candidate phrases, such as "deeply moisturizes the skin" and "enhances radiance," which are weighted phrases.

[0061] In one possible implementation, the text similarity calculation method is used to evaluate the matching degree between the first candidate phrase set and the operational strategy keyword library. The operational strategy keyword library may contain leading words such as "high conversion rate" and "user attraction." For example, the similarity between the phrase "deeply moisturizes the skin" and "moisturizing effect" in the library is calculated. If it exceeds a threshold such as 0.8, it is judged to meet the operational guidance requirements. Then, the importance ranking of each phrase in the original copy is determined, and the frequency and scarcity are evaluated using the TF-IDF algorithm. After ranking, the second candidate phrase set is obtained. For example, "enhance gloss" is ranked first because it matches the strategy library better.

[0062] Specifically, high-weight phrases in the second candidate phrase set are combined and reconstructed according to marketing conversion rules.

[0063] For example, combining "deeply moisturizes skin" and "enhances radiance" into "deeply moisturizes skin and enhances natural radiance" assesses semantic integrity. Semantic coherence is checked, and logical flow is ensured using dependency parsing. If the score reaches a standard value, such as 90 points, it is retained. Next, a syntactic analyzer performs a grammatical check on the reconstructed phrase. Analyzers like Stanford Parser verify the correctness of the subject-verb-object structure, obtaining refined candidate phrases, such as the complete sentence "Use this cream to deeply moisturize skin and enhance radiance."

[0064] In one possible implementation, a content quality assessment algorithm is used to predict and score the marketing effectiveness of candidate snippets. The algorithm may integrate machine learning models, such as regression models, with input features including phrase length, sentiment score, and keyword density, and predict conversion target parameters such as expected click-through rate. If the score meets the parameters set in the operational strategy, such as a conversion rate greater than 5%, the snippet is determined as the final output content. This evaluation can optimize copywriting performance on e-commerce platforms and increase user purchase intent.

[0065] like Figure 3 The diagram illustrates the copy generation, optimization, and validation process. Steps 106 and 107 introduce... Figure 3 The detailed process of the intermediate steps is as follows: Step S106: Use an iterative optimization loop to match and verify the refined copy fragments with multi-dimensional information to obtain the final complete copy output that inherits the style.

[0066] The original text content is segmented using regular expressions. Based on a preset text segment length threshold, the original text content is further segmented. If the number of characters in a text segment exceeds a preset upper limit, a natural language processing tool is used to perform a second segmentation, resulting in a set of standard text segments. A vectorized encoder extracts semantic features from each text segment in the set of standard text segments. Feature vectors for style, sentiment, and intonation dimensions are obtained from a pre-established multi-dimensional information database. If the cosine similarity between the semantic feature vector and the multi-dimensional information feature vector of a text segment is greater than a preset matching threshold, the text segment is considered to have passed preliminary verification. A recurrent neural network is used to evaluate the style consistency of the text segments that have passed preliminary verification. The style deviation value of the text segment is calculated based on the feature parameters of the target style template. If the style deviation value exceeds the tolerance range, word replacement and sentence structure adjustment are performed on the text segment to obtain an optimized text segment. The optimized text fragments are reassembled in their original order using a text splicing algorithm. The semantic coherence of the splicing nodes is verified according to the context coherence detection rules. If the semantic relevance between adjacent text fragments is lower than a preset coherence threshold, transitional words are inserted at the splicing position to determine the final complete style inheritance text.

[0067] For example, when processing original copy for product promotion, regular expressions are first used to identify and segment words. For instance, the copy "This mask can deeply nourish your skin and give you a new lease on life" is segmented into words such as "this," "mask," "can," "deeply," "nourish," "skin," "give," "you," "regain," and "new lease on life." Next, based on a preset segment length threshold, such as no more than 200 characters per segment, the entire copy is initially segmented. If a segment, such as the part introducing product efficacy, exceeds the limit (e.g., 250 characters), natural language processing tools such as sentence segmentation algorithms are used for secondary segmentation, breaking the long sentence into smaller logical units. This ultimately forms multiple sets of standard copy segments, each retaining its independent meaning for easy subsequent analysis.

[0068] In one embodiment, for these standard text fragment sets, each fragment is converted into a semantic feature vector using a vectorized encoder such as the Word2Vec model. For example, "deeply nourish the skin" is encoded as a high-dimensional vector to represent its meaning. Feature vectors for style dimensions (formal or lively), emotion dimensions (positive or neutral), and tone dimensions (enthusiastic or calm) are extracted from a pre-established multi-dimensional information database. If the cosine similarity between the fragment vector and the vectors in the database is calculated, for example, by using the formula cosθ=(A·B) / (|A||B|), and the value is greater than a preset threshold of 0.8, then the fragment is confirmed to have passed the initial verification. This ensures the initial matching of the text in multiple dimensions.

[0069] For example, the style consistency of the verified segments can be evaluated using a recurrent neural network such as an LSTM model. The network processes the segments word by word, generates hidden states to capture sequence dependencies, and calculates the deviation value based on the feature parameters of the target style template, such as word frequency and sentence complexity. If the deviation value, such as 0.15, exceeds the tolerance range of 0.1, the words are replaced, for example, "reborn" is changed to "revitalized", and the sentence structure is modified from a statement to a question to enhance interactivity, thereby obtaining an optimized copywriting segment.

[0070] In one embodiment, a text splicing algorithm is used to reassemble these optimized fragments in their original order, for example, connecting product introduction fragments with efficacy description fragments. Then, the splicing nodes are verified according to contextual coherence detection rules such as semantic graph models. If the semantic relevance between adjacent fragments, such as 0.6, is lower than the threshold of 0.7, transitional words such as "not only that" or "in addition" are inserted to ensure that the final copy flows smoothly and maintains the style, forming a complete promotional text.

[0071] For example, in actual business, this logical progression from initial word segmentation to final concatenation ensures that each step supports the other, forming a consistent copywriting optimization path and ultimately outputting high-quality brand marketing copy.

[0072] In one embodiment, for complex copy such as e-commerce page descriptions, the core is a logical progression from initial word segmentation to final concatenation, ensuring that each step supports each other and forms a consistent copy optimization path.

[0073] Step S107: Perform a consistency check on the complete copy output. If the check result shows a deviation, backtrack to the weighted feature set to adjust the weight distribution and determine whether the corrected copy meets the requirement of grasping the internal relevance.

[0074] The system receives complete text content, which is then semantically parsed using a natural language processor to obtain the relevance values ​​between paragraphs. If the relevance values ​​are lower than a preset threshold, deviation marker coordinates are generated to identify logical breaks in the text. A corresponding weighted feature set is extracted based on the deviation marker coordinates, and a corrected weight parameter is obtained by adjusting the weight distribution matrix in the weighted feature set. A new text fragment is generated using the corrected weight parameter, and a similarity score is calculated between the fragment and the original text. An intrinsic relevance verification is performed on the similarity score. If the similarity score meets a preset standard and the logical relationship coverage meets the requirements, the text is judged to meet the intrinsic relevance grasp standard.

[0075] For example, after receiving a complete copy generated for a young, trendy brand, a natural language processor performs semantic analysis on it. The processor specifically analyzes the intonation feature vectors (such as whether the brand's commonly used words like "cool" or "trendy") and sentiment scores of each paragraph in the copy. If the analysis finds that the tone suddenly becomes formal, the sentiment score drops, and the similarity value with the brand's benchmark falls below a preset threshold in a user case paragraph, a deviation marker is generated. The system then backtracks and adjusts the feature weights, prioritizing words and sentence structures that reflect the "cool" style for regeneration to ensure that the overall style is consistent with the brand's tone.

[0076] In one embodiment, if the relevance score is lower than a preset threshold, the system generates deviation marker coordinates. These coordinates can be understood as specific location markers in the text, such as paragraph numbers or sentence indexes, used to accurately indicate logical breaks.

[0077] For example, in the marketing copy mentioned above, if the relevance between the advantage description paragraph and the user case paragraph is 0.5, which is below the threshold, the system may mark the coordinates as "from the end of paragraph 2 to the beginning of paragraph 3". This indicates the logical jump position from product advantages to actual cases, which facilitates subsequent targeted adjustments.

[0078] For example, a weighted feature set is extracted based on the coordinates of the deviation marker positions. This weighted feature set refers to the set of elements such as keywords, sentence structures, and sentiment tags extracted from the vicinity of the marker positions. The weight distribution matrix is ​​then adjusted to obtain corrected weight parameters. A weight distribution matrix is ​​a structure representing the importance of features; for example, rows in a matrix represent feature types, and columns represent weight values. The adjustment process involves iterative optimization. For instance, initially, the keyword weight in the matrix is ​​0.4, and the sentence structure weight is 0.3. If the analysis shows logical breaks due to incoherent sentence structures, the sentence structure weight is increased to 0.5, generating corrected parameters. In the marketing copy example, the extracted feature set might include "innovative technology" as a keyword and "declarative sentence" as a sentence structure, and then the matrix is ​​adjusted to emphasize the weight of transitional sentence structures.

[0079] In one embodiment, the copy fragment is regenerated using modified weight parameters. This step employs a generative model such as a variant of GPT, taking the modified parameters as input to create a new fragment, and then calculating a similarity score with the original copy. Similarity calculations are typically based on cosine similarity, such as the cosine value of the angle between the new fragment vector and the original fragment vector. If, in the copy, the regenerated fragment describes the transition between product advantages and case studies, a similarity score of 0.85 indicates that the original meaning is preserved while improving coherence.

[0080] For example, an intrinsic correlation verification is performed on the similarity score. This verification involves checking the coverage of logical relationships, such as the completeness of causal and sequential relationships. If the score meets the 0.8 standard and the coverage reaches 90%, the copy is considered to meet the standard. In marketing copy, if the verification shows that the new segment covers the causal logic from advantages to case studies and the score is satisfactory, the entire copy is considered logically sound, thus ensuring the overall flow of the content. From a business perspective, this helps improve the persuasiveness of the copy and avoids confusing readers due to logical gaps.

[0081] Example 2 In this embodiment, a computer terminal device is provided, including: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the above-described deep learning-based marketing copy automatic generation method.

[0082] In this embodiment, a computer-readable storage medium is also provided, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the above-described method for automatically generating marketing copy based on deep learning.

[0083] In this embodiment, an electronic device is also provided, including a memory and a processor. The memory stores a computer program, and the processor is configured to run the computer program to perform the steps of the above-described deep learning-based marketing copy automatic generation method.

[0084] In this embodiment, a computer program product is also provided, including a computer program that, when executed by a processor, implements the steps of the above-described deep learning-based marketing copy automatic generation method.

[0085] The aforementioned program can run on a processor or be stored in memory (or a computer-readable medium). Computer-readable media includes both permanent and non-permanent, removable and non-removable media, and information storage can be achieved by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random-access memory (SRAM), dynamic random-access memory (DRAM), other types of random-access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0086] These computer programs may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes can be implemented by different modules for different steps.

[0087] This embodiment provides such an apparatus or system. The system, referred to as a deep learning-based automatic marketing copy generation system, includes: The multimodal fusion module is used to extract visual features, functional attributes, and textual descriptions from product data to obtain a comprehensive product representation vector; The attention weighting module is used to calculate the weight distribution between visual features and functional attributes based on the comprehensive product representation vector using an attention mechanism, and to determine the weighted feature set. The brand style embedding module is used to map the similarity of the weighted feature set to the brand style inheritance space through the embedding layer to obtain a style-consistent embedding representation when the similarity exceeds a preset threshold. The sequence decoding module is used to obtain the style-consistent embedded representation and then perform sequence decoding to determine the continuous sequence of language style to obtain a preliminary draft of the copy. The operational strategy highlighting module is used to filter relevant phrases emphasizing efficacy and selling points from the initial copy draft and determine refined copy fragments. The iterative optimization module is used to match and verify the refined copy fragments with multi-dimensional information using an iterative optimization loop to obtain complete copy output; The consistency check module is used to perform consistency checks on the complete text output, and when the check results show a deviation, it triggers a backtracking to the attention weighting module to adjust the weight distribution.

[0088] The system or apparatus is used to implement the functions of the methods in the above embodiments. Each module in the system or apparatus corresponds to each step in the method, as has been described in the method and will not be repeated here.

[0089] The above implementation method solves the problem of automatic generation of marketing copy based on deep learning in related technologies, thereby ensuring that the problems existing in the prior art are resolved.

[0090] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention 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 the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for automatically generating marketing copy based on deep learning, characterized in that, Includes the following steps: S1. Visual features, functional attributes, and textual descriptions are extracted from product data using a multimodal fusion model to obtain a comprehensive product representation vector; S2. Based on the comprehensive product representation vector, use an attention mechanism to calculate the weight distribution between visual features and functional attributes, and determine the weighted feature set; S3. If the similarity of the weighted feature set exceeds a preset threshold, then the embedding layer is mapped to the brand style inheritance space to obtain a style-consistent embedding representation. S4. After obtaining the style-consistent embedded representation, perform sequence decoding to determine the continuous sequence of language style to obtain a preliminary draft of the text. S5. Filter relevant phrases emphasizing efficacy and selling points from the preliminary copy draft through the operational strategy highlighting module, and determine the refined copy fragments; S6. Use an iterative optimization loop to match and verify the refined copy fragments with multi-dimensional information to obtain complete copy output; S7. Perform a consistency check on the complete text output. If the check result shows a deviation, backtrack to S2 to adjust the weight distribution.

2. The method according to claim 1, characterized in that, In S1, the process of extracting visual features, functional attributes, and textual descriptions from product data using a multimodal fusion model includes: Visual feature vectors are obtained by extracting color distribution features and texture features from product images using convolutional neural networks. Natural language processing tools are used to perform lexical segmentation and semantic parsing on the product text description to obtain text feature vectors. A feature fusion algorithm is used to weight and merge the visual feature vector and the text feature vector according to a preset weight ratio to obtain a comprehensive feature vector. The effectiveness of the comprehensive feature vector is verified by cosine similarity calculation. If the similarity result meets the preset range, it is saved as the comprehensive product representation vector.

3. The method according to claim 1, characterized in that, In S2, the process of calculating the weight distribution between visual features and functional attributes using an attention mechanism based on the comprehensive product representation vector, and determining the weighted feature set, includes: Visual feature data and functional attribute data are extracted from the comprehensive product representation vector; The weight distribution between the visual features and functional attributes is calculated using an attention mechanism to obtain a set of weighted feature pairs. Clustering algorithms are used to group the weighted feature pairs to determine feature association sets; For each feature pair in the feature association set, the intrinsic association strength value is obtained. If the association strength value exceeds a preset threshold, the feature pair is determined to have effective association. The weighted feature set is generated by performing contextual logic analysis on the feature association set with effective correlation using the attribute matching method.

4. The method according to claim 1, characterized in that, In S3, if the similarity of the weighted feature sets exceeds a preset threshold, the process of mapping the embedding layer to the brand style inheritance space to obtain a style-consistent embedding representation includes: Obtain the visual element feature vector of the target brand, wherein the visual elements include brand logo, color scheme, font style and pattern texture; The cosine similarity calculation method is used to quantitatively evaluate the similarity between the visual element feature vector and the brand benchmark feature. If the similarity value exceeds a preset threshold, the feature matching result is obtained. The feature matching results are mapped to the brand style inheritance space through a neural network embedding layer to obtain an embedding representation vector. Brand style consistency constraints are constructed based on the embedded representation vector, and the style-consistent embedded representation is obtained through vector normalization.

5. The method according to claim 1, characterized in that, In S4, the process of obtaining a style-consistent embedded representation, performing sequence decoding, determining a continuous sequence of language style, and obtaining a preliminary draft copy includes: The input text is segmented using a pre-defined word vector library. The weight distribution of each word is calculated using a word frequency statistical method. The embedding vector matrix is ​​then calculated based on semantic similarity. If the semantic similarity of the word embedding vectors in the embedding vector matrix exceeds a preset threshold, the syntactic structure recognition program is activated to extract features of the text's rhetorical devices and obtain grammatical pattern tags. Temporal modeling is performed based on the grammatical pattern markers, and the probability distribution of word selection at the current position is determined by the preceding context information to identify the candidate word sequence. The candidate word sequence is received and its content logic consistency is checked. If the semantic correlation between adjacent words in the sequence meets the preset conditions, the sequence is marked as a valid segment and spliced ​​together to generate the preliminary draft.

6. The method according to claim 1, characterized in that, In S5, the process of filtering relevant phrases emphasizing efficacy and selling points from the initial draft copy based on operational strategies, and determining the refined copy snippets, includes: The initial draft copy is segmented and semantically annotated using natural language processing tools. If the copy contains product efficacy-related words, semantic feature vectors are extracted. Core selling point phrases are matched according to a preset module filtering rule library to obtain a candidate phrase set. The matching degree between the candidate phrase set and the operational strategy keyword library is evaluated using a text similarity calculation method. If the similarity exceeds a preset threshold, the phrase is judged to meet the operational orientation requirements, and the importance ranking of each phrase in the original text is determined. Based on the marketing conversion rules, the high-weight phrases in the importance ranking are combined and reconstructed. If the semantic integrity score of the phrase combination reaches the standard value, the combination is retained. A syntactic analyzer is used to perform grammatical checks on the reconstructed phrases to obtain refined copywriting candidate fragments. The refined copywriting candidate segments are scored using a content quality assessment algorithm to predict their marketing effectiveness. If the score meets the conversion target parameters set in the operation strategy, the segment is determined to be the refined copywriting segment.

7. A marketing copy automatic generation system based on deep learning, characterized in that, The system for implementing the method of any one of claims 1-6 comprises: The multimodal fusion module is used to extract visual features, functional attributes, and textual descriptions from product data to obtain a comprehensive product representation vector; The attention weighting module is used to calculate the weight distribution between visual features and functional attributes based on the comprehensive product representation vector using an attention mechanism, and to determine the weighted feature set. The brand style embedding module is used to map the similarity of the weighted feature set to the brand style inheritance space through the embedding layer to obtain a style-consistent embedding representation when the similarity exceeds a preset threshold. The sequence decoding module is used to obtain the style-consistent embedded representation and then perform sequence decoding to determine the continuous sequence of language style to obtain a preliminary draft of the copy. The operational strategy highlighting module is used to filter relevant phrases emphasizing efficacy and selling points from the initial copy draft and determine refined copy fragments. The iterative optimization module is used to match and verify the refined copy fragments with multi-dimensional information using an iterative optimization loop to obtain complete copy output; The consistency check module is used to perform consistency checks on the complete text output, and when the check results show a deviation, it triggers a backtracking to the attention weighting module to adjust the weight distribution.

8. A computer terminal device, characterized in that, include: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors perform the steps of the method as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-6.

Images