A multi-platform e-commerce marketing copy generation and delivery effect tracking system

Abstract

The application discloses a multi-platform e-commerce marketing copy generation and delivery effect tracking system, and particularly relates to the technical field of e-commerce intelligent marketing and language processing, which comprises a commodity information analysis module, a cross-platform copy generation module, a multi-platform delivery module, an effect tracking module and a strategy feedback adjustment module.

Description

Technical Field

[0001] This invention relates to the field of e-commerce intelligent marketing and language processing technology, and more specifically, to a multi-platform e-commerce marketing copy generation and placement effect tracking system. Background Technology

[0002] With the rapid development of e-commerce, merchants need to conduct marketing activities simultaneously on multiple e-commerce platforms. These platforms differ significantly in their user base, language style, and content tone, placing high demands on the differentiated generation of marketing copy. Traditional methods relying on manual writing of copy for multiple platforms are inefficient and costly, failing to meet the business needs of e-commerce companies that require frequent new product launches and simultaneous promotion across multiple platforms.

[0003] Currently, the automatic generation of e-commerce marketing copy mainly employs two technical approaches: template-based generation and direct generation based on large language models. Template-based generation involves manually designing copy templates for each platform and filling product attributes into fixed slots to generate the copy. Large language model-based generation, on the other hand, uses product information and platform instructions as input, leveraging the text generation capabilities of a pre-trained model to directly output the copy. During the campaign delivery phase, existing systems typically distribute the generated copy to various e-commerce platforms after attaching tracking tags. In the performance tracking phase, metrics such as exposure, clicks, and conversions are extracted from log data returned by each platform for performance evaluation.

[0004] However, in practical use, it still has some drawbacks. For example, the copy generated by the template-filling scheme lacks flexibility and diversity; the copy for different platforms only differs in template structure, making it difficult to truly reflect the unique language style and expression habits of each platform. The large language model direct generation scheme mixes product content information with platform style information into the model, resulting in a tight coupling between content description and style expression. Adjusting the style of a particular platform can easily affect the accurate expression of product information. Secondly, existing systems lack an effective closed-loop feedback mechanism between performance tracking and copy generation. Performance data is only used for manual reference or offline analysis, making it difficult to guide the automatic optimization of copy generation strategies online. Furthermore, the contributions of different performance factors to the final marketing effect are interdependent, and existing technologies struggle to independently evaluate and specifically adjust each factor. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a multi-platform e-commerce marketing copy generation and placement effect tracking system, which solves the problems mentioned in the background art through the following solutions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a multi-platform e-commerce marketing copy generation and placement effect tracking system, including a product information parsing module: inputting product data for entity recognition and attribute extraction, obtaining a structured product attribute set and splitting it into a general attribute subset and a promotion-sensitive attribute subset;

[0007] Cross-platform copywriting generation module: includes a platform style mapper and a text decoder. The platform style mapper stores the content fidelity matrix, platform style matrix and intensity adjustment matrix corresponding to the e-commerce platform.

[0008] After receiving the structured product attribute set and the target platform identifier, the content fidelity matrix maps the general attribute subset to the platform's general content latent vector, the platform style matrix maps the general attribute subset to the platform's exclusive style offset, and the intensity adjustment matrix maps the promotion-sensitive attribute subset to the promotion intensity latent vector.

[0009] The content latent vector is superimposed with the platform-specific style offset to obtain the platform-based latent vector; the text decoder uses the platform-based latent vector and the promotional intensity latent vector as conditions to decode and generate conditional copy text.

[0010] Multi-platform delivery module: Obtain conditional copy text, attach a unique tracking identifier, encapsulate it according to the target platform's message format, and send it to the corresponding e-commerce platform's marketing release interface;

[0011] Performance tracking module: Using a unique tracking identifier, it extracts exposure, click, and conversion events from log data returned by various e-commerce platforms to construct performance feature vectors, specifically including platform dimension, time window dimension, and promotion intensity dimension;

[0012] Strategy Feedback Adjustment Module: Based on the difference between the effect feature vector and the preset target effect vector, calculate the decoupled adjustment gradient, independently update the platform style matrix and intensity adjustment matrix parameters in reverse, and indicate when the difference between the subsequent generated conditional copywriting text delivery effect and the preset target effect is lower than the threshold.

[0013] Preferably, the product information parsing module uses a sequence labeling architecture based on a pre-trained language model combined with a conditional random field layer to construct an entity recognizer. It defines entity tags including brand, model, specification name, specification value, product category, promotion type, promotion intensity value, and time limit. Based on dependency syntax rules, it performs post-processing on the recognition results to form the entity relationship into a set of attribute-value pairs, thus obtaining a structured product attribute set.

[0014] Preferably, in the product information parsing module, attribute subset splitting is achieved by a hybrid discriminator combining heuristic rules and neural networks. Specifically, this includes: maintaining a keyword whitelist and blacklist; attributes that hit the whitelist are classified as a general attribute subset, and attributes that hit the blacklist are classified as a promotional sensitive attribute subset; for attributes that do not hit the whitelist or blacklist, a finely tuned classifier is used to jointly discriminate the attribute name, attribute value, and the context of the sentence in which they are located, and outputs a binary classification result of general or promotional.

[0015] Preferably, when the content fidelity matrix maps the subset of general attributes to the platform-wide latent content vector, the information bottleneck constraint ensures that the latent content vector retains all the necessary information to restore the subset of general attributes; when the platform style matrix maps the subset of general attributes to the platform-specific style offset, each target platform corresponds to an independent platform style matrix, and the style offsets generated by different platform style matrices are independent of each other, respectively corresponding to the linguistic direction and commonly used expression features of different e-commerce platforms.

[0016] Preferably, the text decoder adopts a generation architecture based on conditional layer normalization and gated cross-attention. The platformized latent vector is transformed into adaptive scaling factors and adaptive bias factors of each layer of the decoder through a conditional feature network, adjusting the normalization process of the hidden states of each layer of the decoder. The promotion intensity latent vector is linearly transformed through the promotion attention gating module and added to the decoder attention score to enhance the attention weight of promotion-related words.

[0017] Preferably, the cross-platform copywriting generation module introduces multi-task decoupling pre-training objectives during the training phase, including content fidelity auxiliary loss, which constrains the reconstruction of a subset of general attributes from the content latent vector; style adversarial loss, which trains the platform discriminator to distinguish the platform source corresponding to the platform-specific style shift, while minimizing the mutual information between the content latent vector and the platform-specific style shift; and promotion intensity reconstruction loss, which constrains the reconstruction of key values ​​in the promotion intensity latent vector within the subset of promotion-sensitive attributes.

[0018] Preferably, the multi-platform delivery module includes a tracking identifier generator and a platform message adapter; the tracking identifier generator creates a globally unique tracking identifier and simultaneously converts it into short URL parameters and pass-through fields returned by each e-commerce platform; the platform message adapter adopts the adapter design pattern, encapsulates the format of the copy publishing interface of different e-commerce platform application interfaces, performs field mapping and character length verification, and then calls the marketing publishing interface of the corresponding e-commerce platform to send the delivery request.

[0019] Preferably, the construction method of the effect feature vector is as follows: the platform dimension adopts a platform one-hot encoding vector; the time window dimension sets multiple backtracking windows, counts the exposure, clicks and conversions in each window, calculates the click-through rate, conversion rate and add-to-cart rate of each window, takes the transformation value of each window indicator, and calculates the decay coefficient of adjacent windows as derived features, and concatenates the features of each window in sequence to form a time-series feature segment; the promotion intensity dimension reuses the promotion intensity latent vector and, after dimensionality reduction mapping, concatenates it with the one-hot encoding of discrete intensity levels to form a promotion intensity representation; the features of the above dimensions are concatenated along the feature dimensions to form the effect feature vector.

[0020] Preferably, the calculation method of the decoupled adjustment gradient is as follows: construct an effect decomposer, and decouple the effect feature vector into content base contribution latent variables, platform style gain latent variables, and promotion intensity gain latent variables based on a variational autoencoder architecture, and introduce mutual information minimization constraints and maximum mean difference alignment loss; after the effect decomposer is trained, extract the platform style gain latent variable components and promotion intensity gain latent variable components of the actual effect feature vector and the target effect vector respectively, and construct style loss and promotion loss.

[0021] Preferably, the strategy feedback adjustment module includes a style gain predictor and a promotion gain predictor, which are used to calculate the update gradients of the platform style matrix and the intensity adjustment matrix, respectively, and the gradients are used to update the corresponding matrices after orthogonal projection decoupling.

[0022] The technical effects and advantages of this invention are as follows:

[0023] 1. This invention decouples and expresses general product attributes, platform style preferences, and promotional intensity information in a latent vector space by independently setting a content fidelity matrix, a platform style matrix, and an intensity adjustment matrix within the cross-platform copywriting generation module. The content latent vector carries inherent product factual information, the platform-specific style offset independently encodes the linguistic direction and common expression features of each platform, and the promotional intensity latent vector encodes promotional strength and type abstraction factors. These three are combined through overlay and conditional injection and then input into the text decoder. Adjusting any factor will not interfere with the others. When adapting to a new platform or adjusting the style, only the style matrix parameters for the corresponding platform need to be updated, without retraining the entire generation model, significantly reducing the deployment and maintenance costs of cross-platform copywriting generation.

[0024] 2. This invention constructs a multi-dimensional effect feature vector encompassing platform, time window, and promotional intensity dimensions through an effect tracking module. It introduces an effect decomposer based on a variational autoencoder architecture, decoupling the overall effect features into three approximately independent latent variable components: content-based contribution, platform style gain, and promotional intensity gain. Compared to existing technologies that only provide overall effect metrics without distinguishing the contributions of individual factors, this invention accurately quantifies the independent impact of different factors on marketing effectiveness, providing fine-grained decision-making support for strategy adjustments and significantly improving the interpretability and relevance of effect analysis.

[0025] 3. This invention, through the style gain predictor and promotion gain predictor in the strategy feedback adjustment module, connects each gain component obtained from effect decomposition with the corresponding generation-side matrix parameters, and projects the update gradient to a direction orthogonal to the column space of the content fidelity matrix based on the orthogonal projection operator, thereby achieving independent online updates of the platform style matrix and intensity adjustment matrix. Compared with the existing technology where effect data is only used for offline manual analysis, this invention achieves online adaptive adjustment of copywriting generation strategies, significantly improving the intelligence level and optimization efficiency of multi-platform marketing campaigns. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall structure of the present invention;

[0027] Figure 2 This is a schematic diagram of entity recognition and attribute decomposition according to the present invention;

[0028] Figure 3 This is a schematic diagram of the latent vector mapping and decoding of the present invention;

[0029] Figure 4 This is a schematic diagram illustrating the construction of the effect feature vector of the present invention;

[0030] Figure 5 This is a schematic diagram of the decoupled gradient update of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] like Figure 1As shown, a multi-platform e-commerce marketing copy generation and campaign performance tracking system includes a product information parsing module: inputting product data for entity recognition and attribute extraction, obtaining a structured product attribute set and splitting it into a general attribute subset and a promotion-sensitive attribute subset.

[0033] like Figure 2 As shown, it should be specifically noted that the product information parsing module uses a sequence labeling architecture based on a pre-trained language model combined with a conditional random field layer to construct an entity recognizer. The entity labels are defined to include brand, model, specification name, specification value, product category, promotion type, promotion strength value, and time limit. The recognition results are post-processed based on dependency syntax rules to form the entity relationship into a set of attribute-value pairs, resulting in a structured product attribute set.

[0034] In the product information parsing module, attribute subset splitting is achieved by a hybrid discriminator combining heuristic rules and neural networks. Specifically, it includes maintaining a keyword whitelist and blacklist. Attributes that hit the whitelist are classified as general attribute subsets, and attributes that hit the blacklist are classified as promotion-sensitive attribute subsets. For attributes that do not hit the whitelist or blacklist, a fine-tuned classifier is used to jointly discriminate based on the attribute name, attribute value, and the context of the sentence, and outputs a binary classification result of general or promotion.

[0035] It should be further explained that the product information parsing module receives unstructured or semi-structured raw product data in the form of JSON documents, database records, or operational form text. Typical fields include product title, detailed description, specifications, price, promotional tags, and activity duration.

[0036] In attribute extraction and entity recognition, this embodiment adopts a sequence labeling architecture based on the large-scale pre-trained language model BERT-Base-Chinese, which takes into account both generalization ability and domain-specific terminology. A bidirectional long short-term memory network and a conditional random field layer are added to the top layer of the pre-trained model to construct an entity recognizer of "BERT+BiLSTM+CRF".

[0037] Defined entity labels include: "BRAND", "MODEL", and "SPEC". NAME (Specification name), "SPEC" VALUE (Specification value), "CATEGORY" (Product category), "PROMO" TYPE (Promotion type), "PROMO" VAL (Promotional strength value) and "TIME" TAG(Time-limited phrase). For example, if you input "XX brand 50-inch 4K smart TV 2026 summer new product launch with a direct price reduction of 800 yuan", the model extracts: "XX brand" as BRAND, "50-inch" and "4K" as SPECIAL. VALUE These correspond to the implicit specification names "size" and "resolution," respectively; "Smart TV" is CATEGORY; and "2026 Summer New Product Launch" is TIME. TAG In the phrase "direct price reduction of 800 yuan", "direct price reduction" is actually a promotion. TYPE "800 yuan" is a PROMO VAL .

[0038] Based on dependency syntax-based rule post-processing, entity relations are formed into a set of attribute-value pairs, such as "Brand: XX, Size: 50 inches, Resolution: 4K, Promotion Type: Direct Price Reduction, Promotion Amount: 800 yuan", which is a structured product attribute set.

[0039] In attribute subset splitting, the system incorporates an attribute classification mapper that divides attributes into a general attribute subset and a promotion-sensitive attribute subset. Specifically, this is achieved using a hybrid discriminator combining heuristic rules and a lightweight neural network. For most standardized product attributes, classification can be completed using a maintained keyword whitelist and blacklist. The whitelist includes inherent product attribute keywords unrelated to promotional activities, such as "brand," "model," "material," and "weight." Attributes matching the whitelist are directly classified as the general attribute subset. The blacklist includes promotion-sensitive keywords that directly indicate price changes, discount formats, or time limits, such as "direct price reduction," "spend more to save," "limited time offer," "flash sale price," and "coupon." Attributes matching the blacklist are directly classified as the promotion-sensitive attribute subset. For attributes with ambiguous boundaries, a fine-tuned bidirectional encoder is used to represent the classifier. The input is the attribute name, attribute value, and the sentence context; the output is a binary classification probability of [general, promotion].

[0040] After decomposition, the general attribute subset includes inherent product information that remains unchanged with time and marketing campaigns, representing the core content that cross-platform copywriting must accurately reflect. The promotion-sensitive attribute subset includes all variables related to price fluctuations, discount strength, time windows, and reward mechanisms; these attributes directly affect consumers' perception of the strength of the discount. These two subsets are denoted as Ag and Ap, respectively, and serve as input to the cross-platform copywriting generation module.

[0041] like Figure 3 As shown, the cross-platform copywriting generation module includes a platform style mapper and a text decoder. The platform style mapper stores the content fidelity matrix, platform style matrix, and intensity adjustment matrix corresponding to the e-commerce platform.

[0042] After receiving the structured product attribute set and the target platform identifier, the content fidelity matrix maps the general attribute subset to the platform's general content latent vector, the platform style matrix maps the general attribute subset to the platform's exclusive style offset, and the intensity adjustment matrix maps the promotion-sensitive attribute subset to the promotion intensity latent vector.

[0043] The content latent vector is superimposed with the platform-specific style offset to obtain the platform-specific latent vector; the text decoder uses the platform-specific latent vector and the promotional intensity latent vector as conditions to decode and generate conditional copy text.

[0044] It should be specifically noted that when the content fidelity matrix maps the subset of general attributes to the platform-wide latent content vector, the information bottleneck constraint ensures that the latent content vector retains all the necessary information to restore the subset of general attributes; when the platform style matrix maps the subset of general attributes to the platform-specific style offset, each target platform corresponds to an independent platform style matrix, and the style offsets generated by different platform style matrices are independent of each other, respectively corresponding to the linguistic direction and commonly used expression features of different e-commerce platforms.

[0045] The text decoder adopts a generation architecture based on conditional layer normalization and gated cross-attention. The platformized latent vector is transformed into adaptive scaling factors and adaptive bias factors of each layer of the decoder through a conditional feature network, which adjusts the normalization process of the hidden state of each layer of the decoder. The promotion intensity latent vector is linearly transformed through the promotion attention gating module and added to the decoder attention score to enhance the attention weight of promotion-related words.

[0046] The cross-platform copywriting generation module introduces multi-task decoupling pre-training objectives during the training phase, including content fidelity auxiliary loss, which constrains the reconstruction of a subset of general attributes from the content latent vector; style adversarial loss, which trains the platform discriminator to distinguish the platform source corresponding to the platform-specific style shift, while minimizing the mutual information between the content latent vector and the platform-specific style shift; and promotion intensity reconstruction loss, which constrains the reconstruction of key values ​​in the promotion intensity latent vector within the promotion-sensitive attribute subset.

[0047] It should be further noted that the platform style mapper stores three sets of independently trained and updated parameter matrices, specifically: a content fidelity matrix. A collection of platform style matrices specific to each target platform Where K represents the total number of target e-commerce platforms, and the promotional intensity adjustment matrix. The aforementioned matrices collectively constitute a decoupled control panel in the copywriting generation process. Specifically, this includes a platform style matrix set. Platform style matrix corresponding to each target e-commerce platform The set consists of a set of platform style matrices, each of which is a dedicated parameter matrix for that platform, stored and updated independently. When generating copy for any target platform, the set of platform style matrices retrieves the platform style matrix corresponding to the platform identifier for calculation.

[0048] That is, the platform style matrix set Platform style matrix as a container for all parameters A single parameter instance in the container that is associated with a specific platform.

[0049] Ag and Ap are respectively fed into two attribute encoders. and The role of an attribute encoder is to transform discrete attribute key-value pairs into dense vectors of fixed dimensions. Specifically, using... For example, for each attribute pair, the attribute name and attribute value are embedded through a word embedding layer and a character convolutional network, respectively. After concatenation, they are passed through a Transformer encoder layer for information exchange. The output vectors of all attribute pairs are aggregated into a single general attribute representation vector through attention-based pooling operations. , dimension .same, Output promotion attribute representation vector , dimension The two encoders share partial word embedding layer parameters, but their top layers are independent to maintain the specificity of the semantic space.

[0050] Content Fidelity Matrix It is a size of The trainable matrix serves to... Linear transformation to implicit content vector , specifically, Here, T is the content fidelity matrix, and T is the transpose symbol. This is a bias term, which can be omitted or incorporated into the affine transformation, depending on the situation. Within the content fidelity subspace, it encodes the factual semantics of "what this product is." During training, information bottlenecks and reconstruction constraints are imposed to ensure... This includes restoring all necessary information from the common attribute set, regardless of subsequent style and promotion changes. The objective description of the products carried should not drift, that is, it should have cross-platform copywriting consistency.

[0051] For the kth e-commerce platform (e.g., in this embodiment, k=1 corresponds to Taobao, k=2 corresponds to JD.com, and k=3 corresponds to Pinduoduo), there exists a platform style matrix. The size is We can set ds=dc so that they can be added together. Similarly As input, generate platform-specific style offsets. , This is the platform style matrix, where T stands for transpose. It should be noted that... Not a simple adjustment Instead of focusing on a single dimension, it extracts highly malleable stylistic elements from the general attribute representation. For example, for the general attribute of "XX brand headphones", Taobao's style may shift towards the tone of "dear friends" and "must-buy items", JD.com's style may shift towards the direction of "quality assurance" and "211 express delivery", and Pinduoduo's style may shift towards the direction of "buying together is more cost-effective" and "10 billion subsidies".

[0052] The content latent vector and style offset are added together in the same semantic space to obtain the platform-based latent vector. .

[0053] Promotional Sensitive Attribute Characterization Through the force adjustment matrix The size is Mapped to the latent vector of promotional intensity , This is a promotional intensity adjustment matrix, where T is the transpose symbol. It does not directly include promotional terms, but rather encodes abstract factors including the "intensity" and "type" of the promotion, such as high intensity, medium intensity, and tiered discount types.

[0054] Text decoder with and To generate conditional copy, this embodiment employs a Transformer decoder architecture, utilizing two conditions: conditional layer normalization and gated cross-attention injection. Specifically, it will... This is transformed into an adaptive gain that is normalized for each decoder layer through a conditional feature network. and bias That is, for the first Hidden state of layer ,conduct:

[0055]

[0056] In the formula, For conditional layer normalization operations, Let be the hidden state vector of a certain layer of the decoder, i.e., the activation value to be normalized. for The mean of the vector The arithmetic mean of the values ​​of each dimension for The standard deviation of the vector, i.e. Standard deviation of each dimension value For conditional scaling factor, by The adaptive gain vector generated by mapping through a conditional feature network (small fully connected layer) has dimensions of... same. For conditional bias factors, by The adaptive bias vector generated by the conditional feature network mapping has dimensions and... The same applies. This platform-based latent vector influences generation from both structural and overall semantic levels, including the promotional intensity latent vector. This is achieved through a promotional attention gating module, which adds a promotional bias term to the decoder's cross-attention layer (which can be converted to self-attention adjustment when there is no external encoder). When calculating the attention score, [the following is used]. After a linear transformation, the bias value is multiplied by the query vector and added to the original attention logits, thus systematically enhancing the attention to promotional-related terms such as "instant discount," "discount," "limited-time offer," and "frenzy." Decoding employs an autoregressive approach, with a special start symbol as the initial input. Each step outputs a probability distribution from the vocabulary, and the final conditional text G is generated through beam search or kernel sampling.

[0057] Specifically, this invention designs a multi-task decoupled pre-training objective and introduces a content-fidelity auxiliary loss: requiring from It can reconstruct the general attribute set Ag; it introduces style adversarial loss: training the platform discriminator, from Differentiate platforms, while minimizing Mutual information with style offset; introducing promotional strength reconstruction loss: making It can reconstruct key values ​​in the promotion-sensitive attribute set.

[0058] Multi-platform delivery module: Obtain conditional copy text, attach a unique tracking identifier, encapsulate it according to the target platform's message format, and send it to the corresponding e-commerce platform's marketing release interface.

[0059] It should be specifically noted that the multi-platform delivery module includes a tracking identifier generator and a platform message adapter; the tracking identifier generator creates a globally unique tracking identifier and simultaneously converts it into short URL parameters and pass-through fields returned by each e-commerce platform; the platform message adapter adopts the adapter design pattern, encapsulates the format of the copy publishing interface of different e-commerce platform application interfaces, performs field mapping and character length verification, and then calls the marketing publishing interface of the corresponding e-commerce platform to send the delivery request.

[0060] It should be further explained that the multi-platform delivery module obtains the generated copy G, the target platform identifier, and the product ID, and executes the delivery task. Specifically, the core sub-modules of the multi-platform delivery module are the tracking identifier generator and the platform message adapter.

[0061] The tracking identifier generator creates a globally unique tracking identifier, which is a fixed-length string composed of a platform code, a timestamp, and a random sequence number. The identifier is simultaneously converted into a short URL parameter and a pass-through field for the open platform's backhaul. To ensure cross-platform consistency, this module generates channel-specific tracking codes based on the identifier and uses consistent hashing to ensure that all data is routed to the same identifier.

[0062] The platform message adapter adopts the adapter design pattern to encapsulate the copywriting publishing interface of different e-commerce platform APIs in terms of format. The adapter is responsible for field mapping, character length truncation and validation, and image material splicing. Then, it calls the corresponding platform's marketing publishing API through the HTTPS protocol to send the delivery request. After successful delivery, it records logs including tracking identifiers, copywriting text, target platform, and delivery timestamp for performance tracking.

[0063] like Figure 4 As shown, the performance tracking module extracts exposure, click, and conversion events from log data returned by various e-commerce platforms using a unique tracking identifier, and constructs a performance feature vector, specifically including platform dimension, time window dimension, and promotion intensity dimension.

[0064] It should be specifically explained that the construction method of the effect feature vector is as follows: the platform dimension adopts the platform one-hot encoding vector; the time window dimension sets multiple backtracking windows, counts the exposure, clicks and conversions in each window, calculates the click-through rate, conversion rate and add-to-cart rate of each window, takes the transformation value of each window indicator, and calculates the decay coefficient of adjacent windows as derived features, and concatenates the features of each window in sequence to form a time-series feature segment; the promotion intensity dimension reuses the promotion intensity latent vector and after dimensionality reduction mapping, it is concatenated with the one-hot encoding of discrete intensity levels to form the promotion intensity representation; the features of the above dimensions are concatenated along the feature dimensions to form the effect feature vector.

[0065] It should be further explained that the performance tracking module accesses event logs from various e-commerce platforms in real time, including exposure, clicks, add-to-cart, order placement, and conversion, through event tracking and data pipelines. Platform logs are pushed via server-side callbacks or data subscription services. The performance tracking module aggregates event streams according to unique tracking identifiers, forming a time-series sequence of behaviors with the tracking identifier as the key value.

[0066] Based on aggregated data, the module constructs a high-dimensional feature vector. , The structure integrates platform, time window, and promotional intensity dimensions, and is constructed as follows:

[0067] Platform dimension: K-dimensional one-hot vectors are used to identify the advertising platform, such as [1,0,0] representing Taobao.

[0068] Time window dimension: Considering the time decay characteristic of marketing effectiveness, multiple retrospective windows are set, such as 15 minutes, 1 hour, 6 hours, 24 hours, 72 hours, and 7 days after the campaign. For each window, the number of add-to-cart events within the corresponding time window is counted. Exposure Click volume Conversion volume And calculate the click-through rate. Conversion rate Add-to-cart rate :

[0069]

[0070]

[0071]

[0072] In the formula, It is an extremely small positive number, usually taken as 1e. -6 or 1e -8 This term, added to the denominator as a smoothing / zero-prevention term, prevents calculation anomalies caused by dividing by zero when exposures or clicks are zero, while not affecting the ratio calculation results under normal data. It eliminates the influence of absolute value units and highlights trends. Specifically, it takes the log transformation value of each window's indicator, calculates the CTR decay coefficient of adjacent windows as a derived feature, and finally flattens and splices the features of all windows in sequence to form a time-series feature segment. .

[0073] Promotional strength dimension: Reusing the information required by the strategy feedback, the promotional strength latent vector hp (used in the current campaign) is extracted after stopping gradient tracing, and then reduced to a promotional strength embedding vector through a mapping network. Simultaneously, the one-hot encoding of discrete strength levels is extracted, and the two are concatenated to form the promotional strength representation. .

[0074] Shallow text feature segments: This part is optional, such as text length, whether it contains exclamation marks, and whether it contains numbers. It is used to assist in analysis but does not participate in the gradient backpropagation of the generative model.

[0075] By concatenating the above sets of features along their respective features, we obtain the final high-dimensional feature vector. , For the set of real numbers, This refers to the dimension of the vector space. Additionally, operators can set preset target effect vectors based on different marketing campaigns. For example, when the goal is to maximize click-through rate (CTR), the CTR-related dimensions in the target performance vector can be set to high values, while other dimensions remain at normal levels; if the goal is high conversion, then the CVR dimension target should be increased. This preset target vector is related to... The dimensional structures are completely identical and have undergone the same standardization process.

[0076] like Figure 5 As shown, the strategy feedback adjustment module is based on the effect feature vector and The difference is calculated, the gradient is decoupled and adjusted, and the platform style matrix and intensity adjustment matrix parameters are updated independently in reverse. The update completion indicator is that the difference between the subsequent generated conditional copywriting text delivery effect and the preset target effect is lower than the threshold.

[0077] It should be specifically explained that the calculation method of the decoupling adjustment gradient is as follows: an effect decomposer is constructed, and the effect feature vector is decoupled into content-based contribution latent variables, platform style gain latent variables, and promotion intensity gain latent variables based on a variational autoencoder architecture. Mutual information minimization constraints and maximum mean difference alignment loss are introduced. After the effect decomposer is trained, the platform style gain latent variable components and promotion intensity gain latent variable components of the actual effect feature vector and the target effect vector are extracted respectively to construct style loss and promotion loss.

[0078] The strategy feedback adjustment module includes a style gain predictor and a promotion gain predictor, which are used to calculate the update gradients of the platform style matrix and the intensity adjustment matrix, respectively. The gradients are decoupled by orthogonal projection and then used to update the corresponding matrices.

[0079] It should be further explained that the strategy feedback adjustment module realizes online adaptive optimization of the copywriting generation strategy, and introduces an effect decomposer and a decoupled adjustment gradient algorithm based on orthogonal projection.

[0080] To achieve accurate separation of the contributions of different factors in the effect vector, this module constructs and trains an effect decomposer network, Decomp. Decomp is based on a β-variable autoencoder and mutual information minimization constraints. The encoder receives... .

[0081] Output the posterior distribution parameters of the three independent latent variables: content-based contribution latent variable. Platform style gain hidden variables Promotional efforts gain hidden variables , Represents a normal distribution. , , The mean vector of the contribution to the content base, the mean vector of the platform style gain, and the mean vector of the promotional intensity gain. , , The standard deviation vectors of the content base contribution, platform style gain, and promotional intensity gain. The decoder will then use the sampled... Reorganization and Reconstruction Training the loss function involves three steps:

[0082] Step 1: Reconstructing the Loss of Effect ,in, Represents the mean squared error function. The actual input feature vector. The reconstructed feature vector;

[0083] Step 2: Sum the KL divergences of the distributions of each latent variable and the standard normal prior, and weight the coefficient β>1 to encourage decoupling;

[0084] Step 3: Introduce an adversarial mutual information estimator, using the estimated mutual information between each pair of latent variables as a loss penalty term. Utilize the latent vectors of the content recorded during delivery. Style offset and the implicit vector of promotional efforts As a weakly supervised signal, an additional maximum mean difference loss is applied to align with... and Statistical distribution and Distribution and The distribution of . After training, the effect decomposer can dissect any effect vector into three approximately independent components, for Similarly, the target component can be obtained by decomposition. , , For the actual Decomposition yields actual components , , .

[0085] Since the link from copywriting generation to effect feedback involves discrete text sampling, directly calculating the gradient through the text decoder suffers from high variance and non-differentiability issues. Therefore, this invention adds two lightweight gain prediction networks: a style gain predictor Fs and a promotion gain predictor Fp. Fs takes the style offset as input. Output the predicted style gain vector, and train to narrow the gap between the predicted style gain vector and the actual decomposed style gain vector. The mean square error between them. Fp takes the promotional strength latent vector hp as input, predicts the promotional gain vector, and compares it with... Alignment.

[0086] Each closed-loop adjustment cycle, comparison and The Euclidean distance, if || - ||>ε, where ε is a preset threshold, triggers the update process, updating the platform style matrix only for the target platform k. and global intensity adjustment matrix Content Fidelity Matrix The account has been frozen. The specific steps are as follows:

[0087] Will and Input the effects into the decomposer separately and extract the style gain components. and Define style loss Similarly, extract the promotional gain component. and Define promotional losses .

[0088] Connected via style gain predictor Fs and ,calculate right The gradient, and then through , Using the transpose notation, the initial style gradient Gs is obtained through a linear mapping. To avoid style updates interfering with the content fidelity subspace, an orthogonal projection operator Pr is introduced, based on the content fidelity matrix. Column space construction: Calculate the projection matrix , The identity matrix is ​​a square matrix with all diagonal elements being 1 and all off-diagonal elements being 0. Its dimension is 1 / 2. The number of rows is the same. The transpose symbol is used if If the rank is not full, a pseudo-inverse is used. Projecting Gs onto a direction orthogonal to the content subspace yields the decoupled adjustment gradient. Then use the optimizer to renew .

[0089] Adjustment matrix for promotional intensity ,calculate The gradient of the promotion gain predictor Fp with respect to hp, and combined with The initial promotion gradient Gp is ​​obtained. To ensure that promotion adjustments do not interfere with style, Gp is ​​projected onto the orthogonal complement space of the platform style matrix column space, while also avoiding impact on content. Specifically, a joint orthogonal projection is constructed, with the projection matrix approximating as... ,in For including Column vectors and various platforms The concatenated basis matrix of the principal components of column vectors. The transpose symbol is used to obtain the result after projection. Then use the optimizer to renew .

[0090] After the update, the newly generated copy will have the adjusted style and promotional strength. The performance tracking module will continue to monitor the new batch of campaign data and calculate new... Compare again. Iterate repeatedly until two consecutive detection cycles pass || - If ||≤ε, it is determined that the difference between the delivery effect and the preset target is lower than the threshold. The update completion indicator is triggered, and the system maintains the current matrix parameters and only performs light monitoring.

[0091] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments of this disclosure. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0092] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-platform e-commerce marketing copy generation and placement performance tracking system, characterized in that, include: Product information parsing module: Input product data, perform entity recognition and attribute extraction, obtain a structured product attribute set and split it into a general attribute subset and a promotion-sensitive attribute subset; Cross-platform copywriting generation module: includes a platform style mapper and a text decoder. The platform style mapper stores the content fidelity matrix, platform style matrix and intensity adjustment matrix corresponding to the e-commerce platform. After receiving the structured product attribute set and the target platform identifier, the content fidelity matrix maps the general attribute subset to the platform's general content latent vector, the platform style matrix maps the general attribute subset to the platform's exclusive style offset, and the intensity adjustment matrix maps the promotion-sensitive attribute subset to the promotion intensity latent vector. The content latent vector is superimposed with the platform-specific style offset to obtain the platform-based latent vector; the text decoder uses the platform-based latent vector and the promotional intensity latent vector as conditions to decode and generate conditional copy text. Multi-platform delivery module: Obtain conditional copy text, attach a unique tracking identifier, encapsulate it according to the target platform's message format, and send it to the corresponding e-commerce platform's marketing release interface; Performance tracking module: Using a unique tracking identifier, it extracts exposure, click, and conversion events from log data returned by various e-commerce platforms to construct performance feature vectors, specifically including platform dimension, time window dimension, and promotion intensity dimension; Strategy Feedback Adjustment Module: Based on the difference between the effect feature vector and the preset target effect vector, calculate the decoupled adjustment gradient, independently update the platform style matrix and intensity adjustment matrix parameters in reverse, and indicate when the difference between the subsequent generated conditional copywriting text delivery effect and the preset target effect is lower than the threshold.

2. The multi-platform e-commerce marketing copy generation and placement effect tracking system according to claim 1, characterized in that: The product information parsing module uses a sequence labeling architecture based on a pre-trained language model combined with a conditional random field layer to construct an entity recognizer. It defines entity tags including brand, model, specification name, specification value, product category, promotion type, promotion intensity value, and time limit. Based on dependency syntax rules, it performs post-processing on the recognition results to form entity relationships into a set of attribute-value pairs, thus obtaining a structured product attribute set.

3. The multi-platform e-commerce marketing copy generation and placement effect tracking system according to claim 1, characterized in that: In the product information parsing module, attribute subset splitting is achieved by a hybrid discriminator combining heuristic rules and neural networks. Specifically, it includes maintaining a keyword whitelist and blacklist. Attributes that hit the whitelist are classified as general attribute subsets, and attributes that hit the blacklist are classified as promotion-sensitive attribute subsets. For attributes that do not hit the whitelist or blacklist, a fine-tuned classifier is used to jointly discriminate based on the attribute name, attribute value, and the context of the sentence, and outputs a binary classification result of general or promotion.

4. The multi-platform e-commerce marketing copy generation and placement effect tracking system according to claim 1, characterized in that: When the content fidelity matrix maps a subset of general attributes to a platform-wide content latent vector, it ensures, through information bottleneck constraints, that the content latent vector retains all the necessary information to restore the subset of general attributes. When the platform style matrix maps a subset of general attributes to platform-specific style offsets, each target platform corresponds to an independent platform style matrix. The style offsets generated by different platform style matrices are independent of each other and correspond to the linguistic direction and commonly used expression features of different e-commerce platforms.

5. The multi-platform e-commerce marketing copy generation and placement effect tracking system according to claim 1, characterized in that: The text decoder adopts a generation architecture based on conditional layer normalization and gated cross-attention. The platformized latent vector is transformed into adaptive scaling factors and adaptive bias factors of each layer of the decoder through a conditional feature network, which adjusts the normalization process of the hidden state of each layer of the decoder. The promotion intensity latent vector is linearly transformed through the promotion attention gating module and added to the decoder attention score to enhance the attention weight of promotion-related words.

6. The multi-platform e-commerce marketing copy generation and placement effect tracking system according to claim 1, characterized in that: The cross-platform copywriting generation module introduces multi-task decoupling pre-training objectives during the training phase, including content fidelity auxiliary loss, which constrains the reconstruction of a subset of general attributes from the content latent vector; style adversarial loss, which trains the platform discriminator to distinguish the platform source corresponding to the platform-specific style shift, while minimizing the mutual information between the content latent vector and the platform-specific style shift; and promotion intensity reconstruction loss, which constrains the reconstruction of key values ​​in the promotion intensity latent vector within the promotion-sensitive attribute subset.

7. The multi-platform e-commerce marketing copy generation and placement effect tracking system according to claim 1, characterized in that: The multi-platform delivery module includes a tracking identifier generator and a platform message adapter; The tracking identifier generator creates a globally unique tracking identifier and simultaneously converts it into short URL parameters and pass-through fields returned by various e-commerce platforms. The platform message adapter adopts the adapter design pattern, which encapsulates the format of the copy publishing interface of different e-commerce platform application interfaces, performs field mapping and character length validation, and then calls the marketing publishing interface of the corresponding e-commerce platform to send the delivery request.

8. The multi-platform e-commerce marketing copy generation and placement effect tracking system according to claim 1, characterized in that: The construction method of the effect feature vector is as follows: the platform dimension adopts the platform one-hot encoding vector; the time window dimension sets multiple backtracking windows, counts the exposure, clicks and conversions in each window, calculates the click-through rate, conversion rate and add-to-cart rate of each window, takes the transformation value of each window indicator, and calculates the decay coefficient of adjacent windows as derived features, and concatenates the features of each window in sequence to form a time-series feature segment; the promotion intensity dimension reuses the promotion intensity latent vector and after dimensionality reduction mapping, it is concatenated with the one-hot encoding of discrete intensity levels to form the promotion intensity representation; the features of the above dimensions are concatenated along the feature dimensions to form the effect feature vector.

9. The multi-platform e-commerce marketing copy generation and placement effect tracking system according to claim 1, characterized in that: The decoupling adjustment gradient is calculated as follows: an effect decomposer is constructed, and the effect feature vector is decoupled into content-based contribution latent variables, platform style gain latent variables, and promotional strength gain latent variables based on a variational autoencoder architecture. Mutual information minimization constraints and maximum mean difference alignment losses are introduced. After the effect decomposer is trained, the platform style gain latent variable components and promotional strength gain latent variable components of the actual effect feature vector and the target effect vector are extracted respectively to construct style loss and promotional loss.

10. The multi-platform e-commerce marketing copy generation and placement effect tracking system according to claim 1, characterized in that: The strategy feedback adjustment module includes a style gain predictor and a promotion gain predictor, which are used to calculate the update gradients of the platform style matrix and the intensity adjustment matrix, respectively. The gradients are decoupled by orthogonal projection and then used to update the corresponding matrices.

Images