Marketing e-mail delivery system and method

Abstract

The invention relates to a marketing e-mail delivery system and method, and relates to the technical field of e-mail delivery, and the delivery system comprises an e-mail content generation module, an e-mail delivery module, an e-mail prediction module, an acquisition and analysis module and a client feedback module. According to the invention, personalized E-mail content generation and multi-modal E-mail content detection are realized through the E-mail content generation module, and the cost of manually writing the E-mail template content and manually checking the E-mail content is reduced. The mail delivery module and the mail prediction module are adopted to predict the content elements of the mail before delivery, and the decision of delivery or delivery canceling is made for the current mail according to the prediction score, so that the overall sending success rate of mail delivery is improved. Through the acquisition and analysis module, continuous optimization of mail content and delivery success rate and effective supplement of a local junk mail feature library are realized. And the complaint rate is reduced through the customer feedback module, and the customer experience is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of email delivery, and in particular to a marketing email delivery system and method. Background Art

[0002] With the acceleration of digital transformation, financial institutions and organizations such as banks, insurance companies, and securities companies are increasingly trying to adopt a variety of messaging methods to inform customers of information in a timely manner and conduct business marketing and other activities. Email, as an indispensable messaging channel in people's daily life and work, is often used by banks to send credit card statements, online marketing activities, institutional announcements and other notifications to customers, pushing content and information that they are interested in.

[0003] At present, the existing email delivery methods in the banking industry have some shortcomings. First, there is a lack of personalized email content, that is, most of the email content is manually written and formulated by the institution, which cannot adapt to the rapidly changing market environment and customized customer needs. Once the customer is not interested in the email content, the institution's sender email address is easily blacklisted by the customer, resulting in the inability to deliver emails to the relevant customers in the future. Secondly, the delivery strategy and email processing mode are single and fixed, and the delivered emails are likely to trigger the security or risk control strategy of the email service operator, resulting in a large number of emails delivered to customers being intercepted or failed to be sent in a short period of time. At the same time, there are many formats for email attachments and the email content lacks regular verification. Some email features are easily detected as spam, which poses compliance risks.

[0004] Therefore, it is necessary to provide a marketing email delivery system and method to improve the email delivery success rate and marketing activity business conversion rate, improve email marketing effects and email service customer experience. Summary of the invention

[0005] The purpose of the present invention is to provide a marketing email delivery system and method to improve the email delivery success rate and marketing activity business conversion rate, improve email marketing effect and email service customer experience.

[0006] In order to solve the problems existing in the prior art, the present invention provides a marketing email delivery system, comprising: The email content generation module is configured to load the email template and the corresponding content of the prompt word from the database as the input of the natural language big model, and perform semantic splicing according to the context of the location of the email template variable and the corresponding content of the prompt word. The natural language big model outputs the corresponding text content, and the generated text content fills the placeholders and variables in the email template; The mail delivery module includes a multimodal processing layer, a word segmentation layer, an embedding layer, a RoBERTa pre-trained model and an optimization unit. The RoBERTa pre-trained model consists of a multi-layer bidirectional Transformer encoder. The RoBERTa pre-trained model uses the Sigmoid function as the activation function of the output layer neurons for binary classification prediction, and maps the output vector to the range of [0,1]. The optimization unit is configured to optimize the RoBERTa pre-trained model by using the binary cross entropy function as the Loss loss function. The Loss loss function is as follows: Where n is the number of layers of the Transformer encoder, i is any layer of the Transformer encoder, y is a binary label value 0 or 1, and p(y) refers to the probability of belonging to the y label; The email prediction module is configured to integrate the outputs of multiple large models through ensemble learning. The multiple large models include the RoBERTa pre-trained model. The outputs of each large model are weighted averaged to obtain the email prediction score, and a decision to deliver or cancel the current email is made based on the prediction score.

[0007] Optionally, in the marketing email delivery system, the email content generation module is further configured to perform content review and rule review.

[0008] Optionally, in the marketing email delivery system, The multimodal processing layer is configured to separate the multimodal fusion input consisting of the email title, email body and email attachment into three separate inputs; The word segmentation layer is configured to split all content into multiple language units, process multiple language units and various special characters, and convert all content of the original email into a vector sequence; The embedding layer is configured to map the vector sequence into a dense vector of fixed dimension. The dense vector output by the embedding layer will be used as the input of the Transformer encoder after position encoding. The Transformer encoder of the RoBERTa pre-trained model is configured to extract the input content using a multi-head attention mechanism; perform residual connection after fusing the context semantics to retain the basic features of the input; normalize the results of the residual connection to ensure the stability and convergence speed of the RoBERTa pre-trained model during training; input the normalized input content into the feedforward neural network; and map the output results of the feedforward neural network processing to a standard normal distribution through secondary normalization; The Transformer decoder of the RoBERTa pre-trained model is configured to transform the complex semantic features output by the Transformer encoder into an interpretable probability distribution, and the Transformer decoder includes three layers of different linear transformations and a Sigmoid activation function.

[0009] Optionally, in the marketing email delivery system, When the email prediction module predicts, it queries the ID of the email template to be predicted from the database, and obtains the list of detectable large models and corresponding prompt words associated with this email template ID. The prompt words and the content of the email to be detected are used as inputs of each large model, and each large model makes predictions.

[0010] Optionally, in the marketing email delivery system, the delivery system also includes a collection and analysis module, which is configured to collect data based on the received sending results and the log information printed by the mail service middleware, and associate it with the delivery details of each email attempted to be delivered, and then infer the delivery result of each email attempted to be delivered, and resend it based on the inferred delivery result and the resending rule set.

[0011] Optionally, in the marketing email delivery system, the configuration interface of the email template is optimized based on the estimated delivery result and by comparing the new and old prompt words.

[0012] Optionally, in the marketing email delivery system, the delivery system also includes a customer feedback module, which is configured to provide a clear unsubscribe link at the bottom of the email and regularly collect user feedback in the form of questionnaires and feedback forms.

[0013] The present invention also provides a marketing email delivery method, using the delivery system, the method comprising the following steps: S1: Load the email template and the corresponding content of the prompt word from the database as the input of the natural language model, and perform semantic splicing based on the context of the location of the email template variable and the corresponding content of the prompt word. The natural language model outputs the corresponding text content, and the generated text content fills the placeholders and variables in the email template; S21: divide the multimodal fusion input consisting of the email title, email body and email attachment into three separate inputs; S22: split all the content into multiple language units, process the multiple language units and multiple special characters, and convert all the original email content into a vector sequence; S23: Map the vector sequence into a dense vector of fixed dimension. The dense vector output by the embedding layer will be used as the input of the Transformer encoder after position encoding. S24: The Transformer encoder of the RoBERTa pre-trained model uses a multi-head attention mechanism to extract input content; after fusing the context semantics, residual connections are performed to retain the basic features of the input; the results of the residual connections are normalized to ensure the stability and convergence speed of the RoBERTa pre-trained model during training; the normalized input content is input into the feedforward neural network; the output results of the feedforward neural network are mapped to a standard normal distribution through secondary normalization; S25: The Transformer decoder of the RoBERTa pre-trained model converts the complex semantic features output by the Transformer encoder into an interpretable probability distribution. The Transformer decoder includes three layers of different linear transformations and a Sigmoid activation function. The Sigmoid function is used as the activation function of the output layer neurons for binary classification prediction, and the output vector is mapped to the range of [0,1]. S26: The RoBERTa pre-trained model is optimized using the binary cross entropy function as the Loss loss function. The Loss loss function is as follows: Where n is the number of layers of the Transformer encoder, i is any layer of the Transformer encoder, y is a binary label value 0 or 1, and p(y) refers to the probability of belonging to the y label; S3: The outputs of multiple large models are integrated through ensemble learning. The multiple large models include the RoBERTa pre-trained model. The outputs of each large model are weighted averaged to obtain the email prediction score. The decision to deliver or cancel the current email is made based on the prediction score.

[0014] Optionally, the marketing email delivery method further includes the following steps: S4: Collect the received sending results and the log information printed by the mail service middleware, and associate it with the delivery details of each mail attempted to be delivered, and then calculate the delivery result of each mail attempted to be delivered, and resend it according to the calculated delivery result and the resending rule set.

[0015] Optionally, the marketing email delivery method further includes the following steps: S5: Provide a clear unsubscribe link at the bottom of your emails and collect user feedback regularly through questionnaires and feedback forms.

[0016] Compared with the prior art, the present invention has the following advantages: (1) The email content generation module realizes personalized email content generation and multi-modal email content detection, reducing the cost of manual writing of email templates and manual review of email content.

[0017] (2) The email delivery module and email prediction module are used to predict the email content elements before delivery, and a decision to deliver or cancel the current email is made based on the prediction score, thereby improving the overall success rate of email delivery.

[0018] (3) The collection and analysis module collects email receipts from different Internet email operators to obtain delivery results, and generates email features for emails with low delivery success rates and large numbers of interceptions, thereby achieving continuous optimization of email content and delivery success rates and effectively supplementing the local spam feature library; The email delivery strategy is associated with the delivery status, and the delivery strategy is dynamically configured and continuously optimized to avoid a single strategy triggering the email service provider's interception strategy and resulting in delivery failure, thereby improving the overall delivery success rate.

[0019] (4) The customer feedback module optimizes the customer behavior processing model after email delivery, provides a clear unsubscribe process and questionnaire management process, reduces customer opinion survey costs, email customer blacklisting rate and overall customer complaint rate, and improves customer experience.

[0020] (5) It reduced the development and management costs of email delivery to customers by relevant organizations and institutions, optimized email delivery strategies and customer feedback processing models, increased email delivery success rates and marketing campaign business conversion rates, and improved email marketing effectiveness and email service customer experience.

[0021] (6) It presents statistical reports on email sending results and delivery details inquiries, optimizing business marketing results and system operation and maintenance capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A module diagram of a marketing email delivery system provided by an embodiment of the present invention; Figure 2 A flowchart of mail delivery and prediction provided by an embodiment of the present invention; Figure 3 A schematic diagram of the structure of the RoBERTa pre-training model provided in an embodiment of the present invention; Figure 4 A schematic diagram of prediction of multiple large models provided by an embodiment of the present invention; Figure 5 A schematic diagram of the mail display effect provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0023] The specific implementation of the present invention will be described in more detail below in conjunction with the schematic diagram. The advantages and features of the present invention will become clearer based on the following description. It should be noted that the drawings are all in a very simplified form and are not in exact proportions, and are only used to facilitate and clearly assist in explaining the purpose of the embodiments of the present invention.

[0024] Hereinafter, if the method described herein includes a series of steps, the order in which the steps are presented herein is not necessarily the only order in which the steps may be performed, and some of the steps described may be omitted and / or some other steps not described herein may be added to the method.

[0025] At present, the existing email delivery methods in the banking industry have some shortcomings. First, there is a lack of personalized email content, that is, most of the email content is manually written and formulated by the institution, which cannot adapt to the rapidly changing market environment and customized customer needs. Once the customer is not interested in the email content, the institution's sender email address is easily blacklisted by the customer, resulting in the inability to deliver emails to the relevant customers in the future. Secondly, the delivery strategy and email processing mode are single and fixed, and the delivered emails are likely to trigger the security or risk control strategy of the email service operator, resulting in a large number of emails delivered to customers being intercepted or failed to be sent in a short period of time. At the same time, there are many formats for email attachments and the email content lacks regular verification. Some email features are easily detected as spam, which poses compliance risks.

[0026] In order to solve the problems existing in the prior art, the present invention provides a marketing email delivery system, such as Figure 1 As shown, including: In order to achieve personalized customer experience for each individual and improve email marketing effectiveness and business efficiency, after completing the email variable configuration, image configuration, and template configuration through the front-end configuration interface developed by Vue or React, for content that requires customized writing, such as holiday greetings, event introductions, product recommendation copy, etc., the natural language model trained by artificial intelligence technology can be used to generate the corresponding copy content.

[0027] The email content generation module is configured to load the email template and the corresponding content of the prompt word from the database as the input of the natural language model, and perform semantic splicing according to the context of the location of the email template variable and the corresponding content of the prompt word. The natural language model outputs the corresponding text content (i.e., AI content generation), and the generated text content fills the placeholders and variables in the email template (i.e., email variable filling); The prompt words include email subject, customer group information, demand clues and keywords. Email subject and customer group information are required, while demand clues and keywords are optional. Email subject can be the email subject of birthday card email, credit card statement email, credit card application activity email, financial product recommendation email, etc. Customer group information can be customer group label information generated by the ECIF (Enterprise Customer Information Facility, enterprise customer information management) system to divide customers, such as pension payment customer group, debit card new card customer group, designated financial product holding customer group, etc. Demand clues can be demand information identified by data such as user preferences and historical interaction behaviors under the corresponding customer groups, such as potential demand for home purchase and renovation loans identified by behavioral data such as searches for "mortgage interest rates" or "decoration loan interest rates" in the App, and records of large changes in provident fund deposit account balances, etc., demand for foreign exchange services identified from data such as frequent cross-border consumption details of foreign currency credit cards and foreign exchange settlement and sales records, and demand for wealth management and asset allocation optimization identified from behaviors such as clicking to browse money funds and short-term wealth management funds and calling customer service to inquire about low-risk preference wealth management products. Keywords can be keywords related to the subject and content of the email, such as the name of the recommended wealth management product, the name of the credit card opening event, the event time, and the event requirements.

[0028] Specifically, enter the email subject, customer information, demand clues, keywords and other prompt words in the email template configuration interface, associate the corresponding template ID, bind the variable sequence that needs to be replaced in the template, enter and persist it in the database. When the email template triggers the sending task, the natural language model will first load the email template content and the corresponding content of the prompt words from the database as the input of the natural language model according to the pre-defined workflow process, and perform semantic splicing based on the context of the location of the email template variables and the corresponding content of the prompt words, and output the corresponding text content. The generated content will replace the placeholders and variables in the form of "${variable}" in the original email template. The interaction process between modules is as follows: Figure 1 shown.

[0029] Preferably, the email content generation module is also configured to perform content audit and rule audit. The specific work of the email content generation module is as follows: the background management program organizes the email content results, sending program logs, and template variables after the natural language large model is filled into a JSON string, and uses the email content value in the key-value pair as the input of the content audit large model. The content audit large model will query the audit rules corresponding to the email template from the Prompt library, predict the probability of email interception, and stop delivering emails that exceed the set probability threshold. And judge whether the email content meets the requirements of the regulatory agency, such as data privacy protection, common email format inspection, sensitive information review, and misleading statement verification, to ensure that the email content complies with the format constraints and regulatory compliance requirements of the corresponding email template. When the content audit large model is detected and passed, it will also pass through the rule library for rule verification to determine whether the email to be delivered meets the current delivery strategy, prevent frequent email sending from accidentally triggering the interception strategy of the email service provider, and improve the success rate of email delivery.

[0030] After the rule library is reviewed and approved, the email will be delivered to the email service provider through the SpringBoot program, and the delivery process details and abnormal situation data will be persisted. At the same time, the email opening rate, activity link click rate, marketing conversion funnel data, etc. after the email is delivered will be collected in real time to quantify the impact of prompt words and model parameters on the marketing reach effect. This allows the subsequent analysis based on the email delivery data in the collection and analysis module, and by comparing the new and old prompt words, continuous optimization can be performed in the email template configuration interface, such as adjusting the keyword weight, context modifier, model input content, email template content, etc., to form a continuously optimized email delivery link, thereby achieving continuously improved delivery effects and customer experience.

[0031] Mail delivery modules, such as Figure 2 and 3 As shown in the figure, it includes a multimodal processing layer, a word segmentation layer, an embedding layer, a RoBERTa pre-training model and an optimization unit. The RoBERTa pre-training model consists of a multi-layer bidirectional Transformer encoder and is optimized in the pre-training stage to achieve better prediction results.

[0032] Specifically, the original email is split into email title, email body and email attachment. The multimodal processing layer is configured to divide the multimodal fusion input composed of email title, email body and email attachment into three separate inputs, and multimodal processing is performed on the email title, email body and email attachment content; when processing image type files, OCR recognition and text extraction are performed through the pytesseract library to extract the text data involved from image files of png, jpg, jpeg and other types; when processing PDF type files, PDF text information is extracted through the pdfminer library; when processing Excel type files, the pandas library is used for processing; when processing Word type files, the python-doc and python-docx libraries are used for processing. After program processing, the email title text, email body and email attachment are extracted from the email to be detected as inputs for the word segmentation layer.

[0033] The word segmentation layer is configured to split all content into multiple language units (such as splitting a complete sentence into words or subwords), and process multiple language units and various special characters to convert all the original email content into a vector sequence; if the text content A to be detected is extracted from the email content and consists of n words, the RoBERTa pre-trained model can convert the email text A into tokens through the word segmentation layer i The vector sequence is as follows: RoBERTa(A)={token 1 ,…,token i ,…token n}, The embedding layer is configured to map the vector sequence into a dense vector of fixed dimension. For example, a word in the token vector sequence may be converted into a semantically rich 768-dimensional vector, so that the RoBERTa pre-trained model can capture the semantic relationship between words and better understand the intention of sending emails. For example, "balance" and "statement" may be closer in the vector space because they often appear in the same email content. The dense vector output by the embedding layer will be used as the input of the Transformer encoder after position encoding; refer to Figure 3 , a schematic diagram of the structure of the RoBERTa pre-training model, where E N and X Nare the input vector and output vector respectively, T is the Transformer encoder, C is the classification label output by the RoBERTa pre-trained model, and CLS is the input information aggregation vector. The Transformer encoder of the RoBERTa pre-trained model is configured to use a multi-head attention mechanism to extract the input content, so as to achieve parallel attention to different subspaces of the input sequence to improve computational efficiency; after fusing the contextual semantics, residual connections are performed to retain the basic features of the input, so that when the sub-layer learning effect is poor, a large amount of key information is avoided from being lost; the results of the residual connections are normalized to ensure the stability and convergence speed of the RoBERTa pre-trained model during training, so that the RoBERTa pre-trained model can maintain good performance even in small sample conditions, and the hidden layer is normalized to a standard normal distribution; the normalized input content is input into the Feed Forward Network (FFN) to prevent local details that may be ignored by the attention mechanism, and the FFN output may cause numerical changes due to parameter updates, so the output results after the feedforward neural network processing are mapped to a standard normal distribution through secondary normalization to avoid the decoder from losing stability due to input scale differences; The Transformer decoder of the RoBERTa pre-trained model is configured to convert the complex semantic features output by the Transformer encoder into an interpretable probability distribution to adapt to downstream task requirements. The Transformer decoder includes three layers of different linear transformations and Sigmoid activation functions. After three layers of linear transformations, the feature dimensions are gradually compressed. At the same time, the three-layer weight distribution can be adjusted according to different content, enhancing the fine control and flexibility of information flow and improving the accuracy of the final decision result.

[0034] Preferably, the RoBERTa pre-trained model uses the Sigmoid function as the activation function of the output layer neurons for binary classification prediction, and maps the output vector to the range of [0,1], indicating the probability of the email being detected as spam.

[0035] Furthermore, the optimization unit is configured to optimize the RoBERTa pre-trained model by using a binary cross entropy function as a Loss loss function, and the Loss loss function is as follows: Where n is the number of layers of the Transformer encoder, i is any layer of the Transformer encoder, y is a binary label value of 0 or 1, and p(y) refers to the probability of belonging to the y label. Because in the email interception prediction scenario, the positive example is the intercepted email, and the negative example is the unintercepted email. The binary cross entropy can be used to compare the probability after Sigmoid activation with the true label, measure the gap between the result probability value predicted by the RoBERTa pre-trained model and the true value, and adjust the RoBERTa pre-trained model parameters through back propagation after calculating the loss to improve the prediction accuracy, so as to improve the classification accuracy and training effect of the RoBERTa pre-trained model.

[0036] like Figure 4 As shown, the email prediction module is configured to integrate the outputs of multiple large models (such as the DeepSeek-V3 large model, LlaMA-3 large model, Qwen2.5-VL large model, etc.) through ensemble learning. The multiple large models include the RoBERTa pre-trained model. The outputs of the large models are weighted averaged to obtain the email prediction score, and the decision to deliver or cancel the current email is made based on the prediction score. In one embodiment, when a total of N large models are deployed in the organization, the weight of each large model is recorded as K. i , each large model predicts the current email and obtains the result P i (A) The predicted results are weighted averaged to obtain the final expected result E RES , the formula is as follows: For example, when the prediction score is ≥95%, it is classified as spam and the delivery is cancelled to prevent repeated triggering of the interception rules of the mail operator service provider, which leads to the failure of normal mail delivery. When the prediction score is <95%, after normal mail delivery, the status report returned by the mail operator service provider is collected to obtain the final result of mail delivery, and compared with the prediction result of the mail prediction module, so as to continuously optimize the prediction accuracy of this delivery system.

[0037] The emails predicted by various models are used to record email feature information to facilitate the subsequent collection of delivery details and summarize them into statistical reports of corresponding email templates for data analysis. When the success rate of sending corresponding email templates is low, the corresponding email features are extracted to expand the local spam feature library as one of the ways to supplement the risk control rules for receiving emails.

[0038] Preferably, when the email prediction module predicts, it queries the email template ID to be predicted from the database, and obtains the list of detectable large models associated with this email template ID (such as RoBERTa pre-trained model, DeepSeek-V3 large model combined with private deployment, LlaMA-3 large model, Qwen2.5-VL large model, etc.) and corresponding Prompt prompt words, and uses the Prompt prompt words and the content of the email to be detected as inputs of each large model, and each large model performs prediction. After multiple large models predict, the prediction results are comprehensively calculated to improve the accuracy and reliability of detection.

[0039] In the process of email delivery, since customer mailboxes often come from different email operation service providers, different service providers use different versions of email servers and different types of middleware, the sender receives different delivery result receipts and log information after delivering the email, which makes it impossible to know the email delivery status in time. Therefore, the delivery system also includes a collection and analysis module, which is configured to collect data based on the received delivery results and the log information printed by the email service middleware, and associate it with the delivery details of each email attempted to be delivered, and then calculate the delivery result of each email attempted to be delivered, and resend it according to the calculated delivery result and the resend rule set. For example, if the calculated delivery result is a delivery failure, and the email meets the resend rule, it will be resent; if it does not meet the resend rule, it will not be resent. The present invention also implements a dynamic delivery strategy. When a scheduled monitoring task finds that a large number of emails trigger a delivery failure scenario, it first records the interception information and sends an early warning notification to the system administrator, and then automatically adjusts the delivery strategy according to the configured rules, thereby reducing the risk of subsequent email delivery failures. The delivery strategy can also be continuously optimized based on subsequent email transactions with operators.

[0040] Optionally, in the marketing email delivery system, the delivery system also includes a customer feedback module, which is configured to provide a clear unsubscribe link at the bottom of the email and regularly collect user feedback in the form of questionnaires and feedback forms. With the help of data such as email read rate, the user's interest preferences and true intentions are investigated in detail, and the user's feedback on the content and frequency of the email is understood, and corresponding adjustments are made. In order to simplify customer service operations, ensure that the email content remains intuitive and clear, and the unsubscribe operation and process are simple and easy to understand, so that users can quickly complete operations such as browsing email content or unsubscribing, analyze the unsubscribe data, questionnaires and feedback form results, thereby reducing the complaint rate and improving customer experience. The final email display effect is as follows: Figure 5 As shown, Figure 5 Contains examples of email layouts and sample effects.

[0041] The present invention generates customer group information based on the consumption behavior, interest preferences, interaction history and other dimensions of the target customers, and combines the demand clue data of the corresponding customer group in the CRM (Customer Relationship Management) system and the email marketing keywords as the input of the natural language big model. Under the structured content framework, the natural language big model outputs personalized email dynamic content such as product introduction words, marketing activity recommendation words, and multimedia content elements. And improves from multiple aspects such as configuration management before email delivery, feature detection and content optimization during delivery, status collection and customer feedback after delivery, and observable indicator analysis, so as to improve the overall sending efficiency, delivery success rate and marketing conversion rate of customer emails, improve customer experience, and increase customer satisfaction.

[0042] The present invention also provides a marketing email delivery method, using the delivery system, the method comprising the following steps: S1: Load the email template and the corresponding content of the prompt word from the database as the input of the natural language model, and perform semantic splicing based on the context of the location of the email template variable and the corresponding content of the prompt word. The natural language model outputs the corresponding text content (i.e., AI content generation), and the generated text content fills the placeholders and variables in the email template (i.e., email variable filling); The prompt words include email subject, customer group information, demand clues and keywords. Email subject and customer group information are required, while demand clues and keywords are optional. Email subject can be the email subject of birthday card email, credit card statement email, credit card application activity email, financial product recommendation email, etc. Customer group information can be customer group label information generated by the ECIF (Enterprise Customer Information Facility, enterprise customer information management) system to divide customers, such as pension payment customer group, debit card new card customer group, designated financial product holding customer group, etc. Demand clues can be demand information identified by data such as user preferences and historical interaction behaviors under the corresponding customer groups, such as potential demand for home purchase and renovation loans identified by behavioral data such as searches for "mortgage interest rates" or "decoration loan interest rates" in the App, and records of large changes in provident fund deposit account balances, etc., demand for foreign exchange services identified from data such as frequent cross-border consumption details of foreign currency credit cards and foreign exchange settlement and sales records, and demand for wealth management and asset allocation optimization identified from behaviors such as clicking to browse money funds and short-term wealth management funds and calling customer service to inquire about low-risk preference wealth management products. Keywords can be keywords related to the subject and content of the email, such as the name of the recommended wealth management product, the name of the credit card opening event, the event time, and the event requirements.

[0043] Specifically, enter the email subject, customer information, demand clues, keywords and other prompt words in the email template configuration interface, associate the corresponding template ID, bind the variable sequence that needs to be replaced in the template, enter and persist it in the database. When the email template triggers the sending task, the natural language model will first load the email template content and the corresponding content of the prompt words from the database as the input of the natural language model according to the pre-defined workflow process, and perform semantic splicing based on the context of the location of the email template variables and the corresponding content of the prompt words, and output the corresponding text content. The generated content will replace the placeholders and variables in the form of "${variable}" in the original email template. The interaction process between modules is as follows: Figure 1 shown.

[0044] Preferably, the email content generation module is also configured to perform content audit and rule audit. The specific work of the email content generation module is as follows: the background management program organizes the email content results, sending program logs, and template variables after the natural language large model is filled into a JSON string, and uses the email content value in the key-value pair as the input of the content audit large model. The content audit large model will query the audit rules corresponding to the email template from the Prompt library, predict the probability of email interception, and stop delivering emails that exceed the set probability threshold. And judge whether the email content meets the requirements of the regulatory agency, such as data privacy protection, common email format inspection, sensitive information review, and misleading statement verification, to ensure that the email content complies with the format constraints and regulatory compliance requirements of the corresponding email template. When the content audit large model is detected and passed, it will also pass through the rule library for rule verification to determine whether the email to be delivered meets the current delivery strategy, prevent frequent email sending from accidentally triggering the interception strategy of the email service provider, and improve the success rate of email delivery.

[0045] After the rule library is reviewed and approved, the email will be delivered to the email service provider through the SpringBoot program, and the delivery process details and abnormal situation data will be persisted. At the same time, the email opening rate, activity link click rate, marketing conversion funnel data, etc. after the email is delivered will be collected in real time to quantify the impact of prompt words and model parameters on the marketing reach effect. This allows the subsequent analysis based on the email delivery data in the collection and analysis module, and by comparing the new and old prompt words, continuous optimization can be performed in the email template configuration interface, such as adjusting the keyword weight, context modifier, model input content, email template content, etc., to form a continuously optimized email delivery link, thereby achieving continuously improved delivery effects and customer experience.

[0046] like Figure 2 As shown, S21: Split the original email into email title, email body and email attachment, split the multimodal fusion input consisting of email title, email body and email attachment into three separate inputs, and perform multimodal processing on the email title, email body and email attachment content; when processing image type files, use the pytesseract library to perform OCR recognition and text extraction, and extract the text data involved from image files of png, jpg, jpeg and other types; when processing PDF type files, use the pdfminer library to extract PDF text information; when processing Excel type files, use the pandas library for processing; when processing Word type files, use the python-doc and python-docx libraries for processing. After program processing, extract the email title text, email body, and email attachment from the email to be detected as word segmentation layer input; S22: Split all content into multiple language units (e.g. split a complete sentence into words or subwords), process multiple language units and various special characters, and convert all original email content into vector sequences; if the text content A to be detected is composed of n words, the RoBERTa pre-trained model can convert the email text A into tokens through the word segmentation layer i The vector sequence is as follows: RoBERTa(A)={token 1 ,…,token i ,…token n}; S23: Map the vector sequence into a dense vector of fixed dimension. For example, a word in the token vector sequence may be converted into a semantically rich 768-dimensional vector, so that the RoBERTa pre-trained model can capture the semantic relationship between words and better understand the intention of sending emails. For example, "balance" and "statement" may be closer in the vector space because they often appear in the same email content. The dense vector output by the embedding layer will be used as the input of the Transformer encoder after position encoding; S24: Reference Figure 3 , a schematic diagram of the structure of the RoBERTa pre-training model, where E N and X Nare the input vector and output vector respectively, T is the Transformer encoder, C is the classification label output by the RoBERTa pre-trained model, and CLS is the input information aggregation vector. The Transformer encoder of the RoBERTa pre-trained model uses a multi-head attention mechanism to extract the input content, so as to achieve parallel attention to different subspaces of the input sequence to improve computational efficiency; after fusing the contextual semantics, residual connection is performed to retain the basic features of the input, so that when the sub-layer learning effect is poor, a large amount of key information is avoided from being lost; the results of the residual connection are normalized to ensure the stability and convergence speed of the RoBERTa pre-trained model during training, so that the RoBERTa pre-trained model can maintain good performance even in the case of small samples, and the hidden layer is normalized to a standard normal distribution; the normalized input content is input into the Feed Forward Network (FFN) to prevent local details that may be ignored by the attention mechanism, and the FFN output may cause numerical changes due to parameter updates, so the output results after the feedforward neural network processing are mapped to a standard normal distribution through secondary normalization to avoid the decoder from losing stability due to input scale differences; S25: The Transformer decoder of the RoBERTa pre-trained model converts the complex semantic features output by the Transformer encoder into an interpretable probability distribution to adapt to downstream task requirements. The Transformer decoder includes three layers of different linear transformations and Sigmoid activation functions. After three layers of linear transformations, the feature dimensions are gradually compressed. At the same time, the three-layer weight distribution can be adjusted according to different content, enhancing the fine control and flexibility of information flow and improving the accuracy of the final decision result.

[0047] Preferably, the RoBERTa pre-trained model uses the Sigmoid function as the activation function of the output layer neurons for binary classification prediction, and maps the output vector to the range of [0,1], indicating the probability of the email being detected as spam.

[0048] S26: The RoBERTa pre-trained model is optimized using the binary cross entropy function as the Loss loss function. The Loss loss function is as follows: Where n is the number of layers of the Transformer encoder, i is any layer of the Transformer encoder, y is a binary label value of 0 or 1, and p(y) refers to the probability of belonging to the y label. Because in the email interception prediction scenario, the positive example is the intercepted email, and the negative example is the unintercepted email. The binary cross entropy can be used to compare the probability after Sigmoid activation with the true label, measure the gap between the result probability value predicted by the RoBERTa pre-trained model and the true value, and adjust the RoBERTa pre-trained model parameters through back propagation after calculating the loss to improve the prediction accuracy, so as to improve the classification accuracy and training effect of the RoBERTa pre-trained model.

[0049] S3: The outputs of multiple large models (such as the DeepSeek-V3 large model, LlaMA-3 large model, Qwen2.5-VL large model, etc.) are integrated through ensemble learning. The multiple large models include the RoBERTa pre-trained model. The outputs of the large models are weighted averaged to obtain the email prediction score. The decision to deliver or cancel the current email is made based on the prediction score. In one embodiment, when a total of N large models are deployed in the organization, the weight of each large model is K. i , each large model predicts the current email and obtains the result P i (A) The predicted results are weighted averaged to obtain the final expected result E RES , the formula is as follows: For example, when the prediction score is ≥95%, it is classified as spam and the delivery is cancelled to prevent repeated triggering of the interception rules of the mail operator service provider, which leads to the failure of normal mail delivery. When the prediction score is <95%, after normal mail delivery, the status report returned by the mail operator service provider is collected to obtain the final result of mail delivery, and compared with the prediction result of the mail prediction module, so as to continuously optimize the prediction accuracy of this delivery system.

[0050] The emails predicted by various models are used to record email feature information to facilitate the subsequent collection of delivery details and summarize them into statistical reports of corresponding email templates for data analysis. When the success rate of sending corresponding email templates is low, the corresponding email features are extracted to expand the local spam feature library as one of the ways to supplement the risk control rules for receiving emails.

[0051] Preferably, when the email prediction module predicts, it queries the email template ID to be predicted from the database, and obtains the list of detectable large models associated with this email template ID (such as RoBERTa pre-trained model, DeepSeek-V3 large model combined with private deployment, LlaMA-3 large model, Qwen2.5-VL large model, etc.) and corresponding Prompt prompt words, and uses the Prompt prompt words and the content of the email to be detected as inputs of each large model, and each large model performs prediction. After multiple large models predict, the prediction results are comprehensively calculated to improve the accuracy and reliability of detection.

[0052] S4: Collect information based on the received sending results and the log information printed by the mail service middleware, and associate it with the delivery details of each mail that is attempted to be delivered, and then infer the delivery result of each mail that is attempted to be delivered, and reissue it according to the inferred delivery result and the reissue rule set. For example, if the inferred delivery result is a delivery failure, and the mail meets the reissue rule, it will be reissued; if it does not meet the reissue rule, it will not be reissued. The present invention also implements a dynamic delivery strategy. When a scheduled monitoring task finds that a large number of mails trigger a delivery failure scenario, it first records the interception information and sends a warning notification to the system administrator, and then automatically adjusts the delivery strategy according to the configured rules, thereby reducing the risk of subsequent mail delivery failures, and can continuously optimize the delivery strategy based on subsequent mail transactions with operators.

[0053] S5: Provide a clear unsubscribe link at the bottom of the email, and regularly collect user feedback through questionnaires and feedback forms. Use data such as email read rates to conduct detailed surveys of user interests, preferences and true intentions, understand user feedback on email content and frequency, and make corresponding adjustments. Simplify customer service operations, ensure that email content remains intuitive and clear, and that unsubscribe operations and processes are simple and easy to understand, so that users can quickly browse email content or unsubscribe. Analyze unsubscribe data, questionnaires and feedback form results, thereby reducing complaints and improving customer experience. The final email display effect is as follows: Figure 5 As shown, Figure 5 Contains examples of email layouts and sample effects.

[0054] In summary, compared with the prior art, the present invention has the following advantages: (1) The email content generation module realizes personalized email content generation and multi-modal email content detection, reducing the cost of manual writing of email templates and manual review of email content.

[0055] (2) The email delivery module and email prediction module are used to predict the email content elements before delivery, and a decision to deliver or cancel the current email is made based on the prediction score, thereby improving the overall success rate of email delivery.

[0056] (3) The collection and analysis module collects email receipts from different Internet email operators to obtain delivery results, and generates email features for emails with low delivery success rates and large numbers of interceptions, thereby achieving continuous optimization of email content and delivery success rates and effectively supplementing the local spam feature library; The email delivery strategy is associated with the delivery status, and the delivery strategy is dynamically configured and continuously optimized to avoid a single strategy triggering the email service provider's interception strategy and resulting in delivery failure, thereby improving the overall delivery success rate.

[0057] (4) The customer feedback module optimizes the customer behavior processing model after email delivery, provides a clear unsubscribe process and questionnaire management process, reduces customer opinion survey costs, email customer blacklisting rate and overall customer complaint rate, and improves customer experience.

[0058] (5) It reduced the development and management costs of email delivery to customers by relevant organizations and institutions, optimized email delivery strategies and customer feedback processing models, increased email delivery success rates and marketing campaign business conversion rates, and improved email marketing effectiveness and email service customer experience.

[0059] (6) It presents statistical reports on email sending results and delivery details inquiries, optimizing business marketing results and system operation and maintenance capabilities.

[0060] The above is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any technician in the relevant technical field, without departing from the scope of the technical solution of the present invention, makes any form of equivalent replacement or modification to the technical solution and technical content disclosed in the present invention, which does not depart from the content of the technical solution of the present invention and still falls within the protection scope of the present invention.

Claims

1. A marketing email delivery system, characterized in that: include: The email content generation module is configured to load the email template and the corresponding content of the prompt word from the database as the input of the natural language big model, and perform semantic splicing according to the context of the location of the email template variable and the corresponding content of the prompt word. The natural language big model outputs the corresponding text content, and the generated text content fills the placeholders and variables in the email template; The mail delivery module includes a multimodal processing layer, a word segmentation layer, an embedding layer, a RoBERTa pre-trained model and an optimization unit. The RoBERTa pre-trained model consists of a multi-layer bidirectional Transformer encoder. The RoBERTa pre-trained model uses the Sigmoid function as the activation function of the output layer neurons for binary classification prediction, and maps the output vector to the range of [0,1]. The optimization unit is configured to optimize the RoBERTa pre-trained model by using the binary cross entropy function as the Loss loss function. The Loss loss function is as follows: Where n is the number of layers of the Transformer encoder, i is any layer of the Transformer encoder, y is a binary label value 0 or 1, and p(y) refers to the probability of belonging to the y label; The email prediction module is configured to integrate the outputs of multiple large models through ensemble learning. The multiple large models include the RoBERTa pre-trained model. The outputs of each large model are weighted averaged to obtain the email prediction score, and a decision to deliver or cancel the current email is made based on the prediction score.

2. The marketing email delivery system according to claim 1, characterized in that: The email content generation module is also configured to perform content review and rule review.

3. The marketing email delivery system according to claim 1, characterized in that: The multimodal processing layer is configured to separate the multimodal fusion input consisting of the email title, email body and email attachment into three separate inputs; The word segmentation layer is configured to split all the contents of the email into multiple language units, process multiple language units and various special characters, and convert all the contents of the original email into a vector sequence; The embedding layer is configured to map the vector sequence into a dense vector of fixed dimension. The dense vector output by the embedding layer will be used as the input of the Transformer encoder after position encoding. The Transformer encoder of the RoBERTa pre-trained model is configured to extract the input content using a multi-head attention mechanism; after fusing the contextual semantics, a residual connection is performed to retain the basic features of the input; Normalize the results of residual connections to ensure the stability and convergence speed of the RoBERTa pre-training model during training; The normalized input is fed into the feedforward neural network; The output of the feedforward neural network is mapped to a standard normal distribution through quadratic normalization; The Transformer decoder of the RoBERTa pre-trained model is configured to transform the complex semantic features output by the Transformer encoder into an interpretable probability distribution, and the Transformer decoder includes three layers of different linear transformations and a Sigmoid activation function.

4. The marketing email delivery system according to claim 1, characterized in that: When the email prediction module predicts, it queries the ID of the email template to be predicted from the database, and obtains the list of detectable large models and corresponding prompt words associated with this email template ID. The prompt words and the content of the email to be detected are used as inputs of each large model, and each large model makes predictions.

5. The marketing email delivery system according to claim 1, characterized in that: The delivery system also includes a collection and analysis module, which is configured to collect data based on the received sending results and the log information printed by the mail service middleware, and associate it with the delivery details of each mail attempted to be delivered, and then infer the delivery result of each mail attempted to be delivered, and resend it based on the inferred delivery result and the resending rule set.

6. The marketing email delivery system according to claim 5, characterized in that: Based on the estimated delivery results and by comparing the old and new prompts, the configuration interface of the email template is optimized.

7. The marketing email delivery system according to claim 1, characterized in that: The delivery system also includes a customer feedback module, which is configured to provide a clear unsubscribe link at the bottom of the email and regularly collect user feedback in the form of questionnaires and feedback forms.

8. A marketing email delivery method, characterized in that: Using the delivery system according to any one of claims 1 to 7, the method comprises the following steps: S1: Load the email template and the corresponding content of the prompt word from the database as the input of the natural language model, and perform semantic splicing based on the context of the location of the email template variable and the corresponding content of the prompt word. The natural language model outputs the corresponding text content, and the generated text content fills the placeholders and variables in the email template; S21: divide the multimodal fusion input consisting of the email title, email body and email attachment into three separate inputs; S22: split all the content into multiple language units, process the multiple language units and multiple special characters, and convert all the original email content into a vector sequence; S23: Map the vector sequence into a dense vector of fixed dimension. The dense vector output by the embedding layer will be used as the input of the Transformer encoder after position encoding. S24: The Transformer encoder of the RoBERTa pre-trained model uses a multi-head attention mechanism to extract input content; after fusing the context semantics, residual connections are performed to retain the basic features of the input; the results of the residual connections are normalized to ensure the stability and convergence speed of the RoBERTa pre-trained model during training; the normalized input content is input into the feedforward neural network; the output results of the feedforward neural network are mapped to a standard normal distribution through secondary normalization; S25: The Transformer decoder of the RoBERTa pre-trained model converts the complex semantic features output by the Transformer encoder into an interpretable probability distribution. The Transformer decoder includes three layers of different linear transformations and a Sigmoid activation function. The Sigmoid function is used as the activation function of the output layer neurons for binary classification prediction, and the output vector is mapped to the range of [0,1]. S26: The RoBERTa pre-trained model is optimized using the binary cross entropy function as the Loss loss function. The Loss loss function is as follows: Where n is the number of layers of the Transformer encoder, i is any layer of the Transformer encoder, y is a binary label value 0 or 1, and p(y) refers to the probability of belonging to the y label; S3: The outputs of multiple large models are integrated through ensemble learning. The multiple large models include the RoBERTa pre-trained model. The outputs of each large model are weighted averaged to obtain the email prediction score. The decision to deliver or cancel the current email is made based on the prediction score.

9. The marketing email delivery method according to claim 8, characterized in that: The following steps are also included: S4: Collect the received sending results and the log information printed by the mail service middleware, and associate it with the delivery details of each mail attempted to be delivered, and then calculate the delivery result of each mail attempted to be delivered, and resend it according to the calculated delivery result and the resending rule set.

10. The marketing email delivery method according to claim 8, characterized in that: The following steps are also included: S5: Provide a clear unsubscribe link at the bottom of your emails and collect user feedback regularly through questionnaires and feedback forms.

Images