Multi-objective optimized short message content generation type artificial intelligence method and system

Abstract

The invention discloses a multi-objective optimized short message content generation type AI method and system, and the method comprises the steps: firstly, constructing a multi-dimensional user portrait, and carrying out the quantization of the multi-dimensional user portrait, and generating a multi-objective weight vector; then, the short message content is structured into five functional components of triggering, main body, numerical value, timeliness and action, and the functional components are respectively generated in combination with a large-scale language model, a structured database and a rule engine; then, assembling the components into candidate short messages according to a preset grammar rule matrix; performing bidirectional semantic verification, performing verification by calculating semantic deviation between the original business intention and the candidate short messages, and returning to reassembly if verification fails; and finally, based on a time sequence difference optimization framework, carrying out iterative optimization on the short message sequences of a plurality of time windows in a marketing period so as to balance an instant effect and a long-term value, and generating an optimal short message sequence. According to the method, the individuation degree and the semantic accuracy of the short message content and the overall return on investment of marketing activities are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and marketing automation technology, specifically relating to an AI method and system for generating SMS content based on multi-objective optimization. Background Technology

[0002] With the widespread adoption of mobile internet and the rapid development of digital marketing, SMS marketing, as a direct and relatively low-cost marketing method, has been widely used across various industries. SMS marketing platforms possess highly automated and customizable capabilities, adapting to information dissemination needs of various scales, and play an indispensable role in industries such as finance, e-commerce, education, and healthcare. However, existing SMS marketing technologies have many limitations in content generation and optimization. Current SMS marketing solutions on the market mainly rely on template-based content generation; however, these methods are essentially still based on simple filling and replacing of preset templates, lacking a deep understanding and optimization of marketing objectives.

[0003] Multi-objective optimization techniques have been widely used in recommendation ranking systems, primarily addressing the comprehensive ranking problem of multiple business objectives such as CTR (click-through rate) and CVR (conversion rate). Currently commonly used algorithms, such as ESMM and MMOE, are optimized based on the equivalence or linear correlation of objective importance. However, existing multi-objective optimization methods face technical challenges such as gradient conflicts and parameter distortion when dealing with marketing content generation.

[0004] Reinforcement learning has begun to be applied in scenarios such as targeted coupon marketing, learning optimal strategies through the interaction between the agent and the environment. Temporal difference algorithms, as an important method in reinforcement learning, do not require a complete sequence of states; they can solve problems using two consecutive states and their corresponding rewards. However, the application of these techniques in optimizing marketing content sequences is still in the exploratory stage.

[0005] In the field of natural language generation, existing evaluation metrics mainly rely on n-gram matching methods such as BLEU and ROUGE. These metrics assess accuracy by comparing the overlap of n-grams between the generated and reference texts. Traditional automated evaluation methods are primarily based on "formal matching," which suffers from neglecting semantics, relying on reference texts, and failing to capture subtle differences between different tasks. Although model-based evaluation methods such as BERTScore and GPTScore have emerged, ensuring the generated content aligns with the original business intent remains a challenge. Figure 1 There are still shortcomings in terms of consistency.

[0006] Through in-depth analysis of existing technologies, the following main technical problems have been identified in the field of SMS marketing content generation: 1. Limitations of Single-Objective Optimization. Existing SMS content generation methods typically focus on a single metric, such as click-through rate or conversion rate, failing to simultaneously balance multiple conflicting marketing objectives such as conversion rate, content conciseness, compliance, and personalization. This results in generated SMS content that excels in one dimension but performs poorly in overall marketing effectiveness.

[0007] 2. Lack of structured content generation mechanism. While traditional template-based methods ensure content standardization, they lack flexibility and creativity. Existing large language models, although possessing powerful creative generation capabilities, often generate content that lacks structured constraints, making it difficult to meet the specific format requirements of marketing SMS messages.

[0008] 3. Insufficient semantic consistency verification. Existing technologies lack effective mechanisms to verify whether the generated SMS content accurately conveys the original business intent. Evaluation methods that rely solely on form matching cannot capture semantic biases, potentially leading to generated content that does not align with marketing objectives.

[0009] 4. Ignoring temporal relevance and long-term effects. Marketing campaigns typically involve multiple time windows, with varying marketing focuses and user receptiveness at different stages. Current technology lacks holistic optimization of SMS sequences and cannot consider the impact of current SMS messages on subsequent marketing effectiveness.

[0010] 5. Insufficient depth in utilizing user profiles. While existing systems can acquire basic user behavior data, they fall short in translating user characteristics into specific content generation strategies, thus failing to achieve truly personalized marketing.

[0011] Therefore, there is an urgent need for an intelligent SMS marketing solution that can simultaneously address issues such as multi-objective optimization, structured content generation, semantic verification, and timing optimization. Summary of the Invention

[0012] This invention addresses the limitations of existing technologies, such as single-objective optimization, lack of structured content generation mechanisms, insufficient semantic consistency verification, neglect of temporal correlation and long-term effects, and insufficient depth of user profiling utilization. It provides a multi-objective optimization-based artificial intelligence method and system for SMS content generation.

[0013] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: In a first aspect, the present invention provides a multi-objective optimized artificial intelligence method for generating SMS content, the method comprising: S1: Based on the differentiated optimization priorities within each time window of the marketing campaign cycle, construct a multi-dimensional user profile, and quantify the optimization priorities into a multi-objective weight vector by combining the business scenario with the user profile; S2: Adopt a functional component-based generation method to structure the SMS content to be generated into five types of content components, including trigger components, body components, numerical components, time-sensitive components, and action components. The trigger and action components are generated using a large language model, the body and numerical components are directly extracted from a structured database, and the time-sensitive component is generated by a rule engine based on the marketing strategy; S3: Assemble the five types of content components into candidate SMS content according to a preset grammar rule matrix; S4: Implement two-way semantic verification, encoding the original business intent into a first language... The semantic feature matrix is ​​generated by inversely decoding the candidate SMS content using an analytical model that differs from the large language model. The semantic deviation between the first and second semantic feature matrices is calculated. If the semantic deviation meets the preset verification criteria, the verification is successful; otherwise, the process returns to step S3 for reassembly. S5: For multiple time windows in the marketing campaign cycle, steps S2 to S4 are repeated to generate multiple verified candidate SMS messages corresponding to the optimization focus of each time window. The candidate SMS messages are combined to form an SMS sequence. Based on the temporal difference optimization framework, the SMS sequence is iteratively optimized using a temporal value function so that each SMS message in the sequence achieves the optimal goal of the current time window while laying the groundwork for the effect of subsequent SMS messages.

[0014] Secondly, the present invention provides a multi-objective optimized artificial intelligence system for generating SMS content, the system comprising: The user profile and weighting module is used to construct multi-dimensional user profiles based on the differentiated optimization priorities in each time window of the marketing campaign cycle, and quantify the optimization priorities into a multi-objective weight vector by combining the business scenario with the user profiles. The content component generation module is used to structure the SMS content to be generated into five types of content components: trigger components, main components, numerical components, time-sensitive components, and action components, using a functional component-based generation method. The content assembly module is used to assemble the five types of content components into candidate SMS content according to a preset syntax rule matrix. The semantic verification module is used to implement two-way semantic verification, and to verify the candidate SMS content by calculating the semantic deviation between the original business intent and the candidate SMS content. The sequence optimization module is used to control the content component generation module, content assembly module, and semantic verification module to generate multiple verified candidate SMS messages corresponding to the optimization priorities of each time window for multiple time windows in the marketing campaign cycle, combine the candidate SMS messages to form an SMS sequence, and iteratively optimize the SMS sequence through a time-series value function based on a time-series difference optimization framework to achieve a balance between the optimal goal of the current time window and the preparation for the effect of subsequent SMS messages.

[0015] The beneficial effects of this invention are: 1. By constructing a multi-objective weight vector, it simultaneously optimizes multiple conflicting marketing objectives such as conversion rate, content simplicity, compliance, and personalization, avoiding the limitations of traditional single-objective optimization and achieving an overall improvement in marketing effectiveness.

[0016] 2. A decomposition method based on functional components is adopted to break down the complex SMS content generation task into five controllable functional components. This not only ensures the structured and standardized nature of the content, but also fully leverages the creative generation capabilities of the large language model, solving the problem of the lack of flexibility in traditional template filling methods.

[0017] 3. By establishing a two-way semantic mapping between the original business intent and the generated content, the semantic consistency verification between the generated content and the marketing objectives is realized, effectively avoiding marketing failures caused by semantic deviations and improving the accuracy and reliability of content generation.

[0018] 4. Based on the temporal difference optimization framework, the correlation and long-term effects of different time windows in the marketing campaign are considered, and a balance is achieved between the immediate effect of a single SMS message and the long-term value of the overall sequence, thereby improving the overall return on investment of the marketing campaign.

[0019] 5. By constructing multi-dimensional user profiles, we can deeply explore users' historical behavioral characteristics, preference patterns, and lifecycle stages, thereby achieving precise personalized content generation and improving user engagement and conversion rates.

[0020] In summary, this invention achieves intelligent generation of SMS content with multi-objective collaborative optimization through steps such as multi-dimensional user profile construction, functional component decomposition and generation, grammatical rule assembly, two-way semantic verification, and temporal differential optimization. This improves the personalization and semantic accuracy of marketing content, enhances the overall effectiveness of marketing activities, and increases long-term return on investment. In particular, it demonstrates significant technical advantages and application value when dealing with multi-objective conflicts and temporal correlation optimization in complex marketing scenarios. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a multi-objective optimization-based artificial intelligence method for generating SMS content.

[0022] Figure 2 This is a schematic diagram of the structure of a multi-objective optimization AI system for generating SMS content. Detailed Implementation

[0023] 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.

[0024] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. A feature described by "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0025] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment or design that is described herein as "for example" should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the term "for example" is intended to present the concept in a specific manner.

[0026] Example 1: like Figure 1 As shown, this embodiment of the invention provides a multi-objective optimized artificial intelligence method for SMS content generation, the method comprising: Step S1: Construct a multi-dimensional user profile and a multi-objective weight vector Specifically, step S1 constructs a multi-dimensional user profile based on the differentiated optimization priorities within each time window of the marketing campaign cycle. Then, combining the business scenario with the user profile, the optimization priorities are quantified into a multi-objective weight vector. This step achieves precise modeling of marketing objectives and quantitative expression of personalized strategies, providing accurate optimization guidance for subsequent content generation and effectively improving the targeting and predictability of marketing results.

[0027] Step S110: Extract historical user behavior features. Specifically, extract historical user behavior features, which include at least click-through rate (CTR), conversion rate, and active time periods. CTR feature extraction involves statistically analyzing user click behavior on different types of SMS content within a preset time period, calculating the CTR for specific product categories, price ranges, and promotion types. Preferably, a sliding window algorithm is used, with a 7-day basic statistical period, a recent behavior weight coefficient of 0.7, and a historical behavior weight coefficient of 0.3. Conversion rate feature extraction involves tracking the complete conversion path from clicking an SMS message to completing a purchase. Specifically, conversion rate features include: the conversion rate CR1 from SMS click to page visit, the conversion rate CR2 from page visit to adding to cart, and the conversion rate CR3 from adding to cart to order payment. Preferably, a funnel analysis model and logistic regression algorithm are used to model and predict conversion probabilities. Active time period feature extraction involves analyzing the user's behavioral distribution patterns across different time periods within a 24-hour period, identifying the user's high-activity time periods. Preferably, a clustering algorithm is used to categorize user active time periods into four types: morning, midday, evening, and nighttime.

[0028] The beneficial effect of this step is that by accurately extracting multi-dimensional behavioral characteristics, a quantitative representation of user behavior is established, which significantly improves the accuracy of matching marketing content with user needs.

[0029] Step S120: Analyze user preference patterns. Specifically, analyze user preference patterns, which include at least content length preference, promotion sensitivity, and interaction frequency. Content length preference analysis involves statistically analyzing user responses to SMS messages of different lengths, categorizing SMS lengths into short (20-50 characters), medium (51-80 characters), and long (81-120 characters). Specifically, a length preference model is established by analyzing user click-through rates and dwell time. Preferably, a Gaussian mixture model is used to model user length preferences. Promotion sensitivity analysis quantifies the intensity of user responses to different promotional strategies. Preferably, to understand the degree of user response to promotional activities, this method evaluates the conversion effect of different promotional types (such as discounts or full reductions), multiplies the conversion rate of each promotion by its corresponding importance coefficient, sums them to obtain a weighted total, and then compares this total with the user's usual conversion rate to obtain their promotion sensitivity. Interaction frequency analysis involves statistically analyzing the user's tolerance for SMS receiving frequency and the optimal reach interval. Specifically, by analyzing user complaint rates, unsubscription rates, and response decay curves, a personalized sending frequency strategy is determined. Preferably, a survival analysis model is used to establish a user fatigue prediction function.

[0030] Where F(t) is the user's fatigue probability function at time t, t is the time variable, representing the time interval since the last SMS was sent, λ is the fatigue rate parameter, which controls the rate at which the user's fatigue increases over time, and e is the base of the natural logarithm.

[0031] The benefit of this step lies in its in-depth analysis of users' content preferences and behavioral patterns, providing a scientific basis for the development of personalized content strategies and effectively improving user experience and marketing conversion rates.

[0032] Step S130: Identify the user's lifecycle stages. Specifically, identify the user's lifecycle stages, which include at least four key stages: new customer, active, dormant, and churn warning. The new customer stage is identified for users who have registered for less than 30 days and made fewer than 3 transactions, analyzing their browsing behavior, dwell time, page navigation paths, and other indicators. Preferably, the k-means clustering algorithm is used to divide new customers into three subcategories: high-intent, indifferent, and low-intent. The active stage is identified for users who have made purchases in the past 30 days and whose transaction amount exceeds the average level, calculating an activity score. Preferably, for active users, their activity level is measured by a comprehensive score that integrates four types of user behavior data and assigns different levels of importance: purchase frequency is the most important, followed by spending amount, then browsing time, and finally sharing behavior. These standardized behavioral data are weighted and summed to obtain the final activity score. The dormant stage is identified for users who have not made purchases in the past 60-180 days but still have some browsing activity, identified using an RFM model. Specifically, the criteria for identifying dormant users are an R score < 0.3 and an F score < 0.02. Preferably, the churn warning stage uses a machine learning classification algorithm to predict churn risk based on user behavior decay patterns. Preferably, a random forest algorithm is used to train the churn prediction model, and a churn warning is triggered when the predicted probability exceeds 0.75.

[0033] The benefit of this step is that it enables differentiated deployment of marketing strategies through accurate lifecycle stage identification, significantly improving the efficiency of marketing resource allocation.

[0034] Step S140: Construct a multi-objective weight vector. Specifically, a multi-objective weight vector is constructed, and the dimensions of the multi-objective weight vector include at least four core dimensions: conversion rate, content conciseness, compliance, and personalization. The conversion rate weight is calculated based on historical marketing data and user conversion probabilities, using a logistic regression model. Preferably, the conversion rate weight is not a single indicator, but is dynamically generated by combining three factors. First, the user's historical conversion performance is considered; second, the model's prediction of the user's future conversion probability is combined; and third, the attractiveness of the current promotional activity compared to competitors is evaluated. These three factors are combined according to a preset importance to obtain a comprehensive conversion rate weight value. Preferably, the weight values ​​can be α=0.5, β=0.3, and γ=0.2. The content conciseness weight is calculated by dynamically adjusting the conciseness weight based on the user's content length preference and information delivery needs. Preferably, the importance of content conciseness is dynamically adjusted. It is determined by first assessing the user's preference for long texts and the complexity of the SMS content itself. The conciseness weight is inversely proportional to the combined effect of these two factors; that is, the stronger the user's preference for long texts or the more complex the information, the lower the requirement for conciseness, and the lower the weight. Compliance weight is calculated based on regulatory requirements and corporate risk control strategies. Specifically, the compliance weight is adjusted from a base value. Preferably, this method starts with a basic compliance requirement, then increases this base value according to the specific risk level of the industry and the user's past complaint records, ultimately forming a compliance weight that fits the current scenario. Personalization weight is calculated based on the richness of the user profile and the intensity of personalization needs. Specifically, the personalization weight is determined by comprehensively considering three key factors. Preferably, these three factors are the completeness of the currently available user profile information, the intensity of the user's potential demand for personalized content, and the suitability of personalized recommendations at the current stage of the marketing campaign. The evaluation results of these three factors jointly determine the final weight of the personalization objective.

[0035] The beneficial effect of this step is that it enables the scientific quantification and weight allocation of multiple marketing objectives, solves the problem of conflict between objectives in traditional single-objective optimization, and significantly improves the scientific nature of marketing decisions.

[0036] Furthermore, the specific implementation process of constructing a multi-dimensional user profile in step S1 includes: Step S111: User behavior data collection and preprocessing. Preferably, user behavior data is collected from multiple data sources, and an ETL data processing workflow is used to clean, standardize, and deduplicate the raw data.

[0037] Step S112: Feature Engineering and Dimension Construction. Preferably, feature engineering techniques are used to convert the raw behavioral data into feature vectors that can be used for machine learning. This includes temporal feature encoding, behavioral sequence encoding, and statistical feature construction.

[0038] Furthermore, the dynamic adjustment mechanism of the multi-objective weight vector in step S1 includes: Step S141: Time-Window Differentiated Weight Allocation. The multi-objective weight vector is dynamically adjusted according to different stages of the marketing campaign. Preferably, the multi-armed slot machine algorithm in reinforcement learning is used for online weight optimization.

[0039] Step S142: Weight Vector Normalization and Verification. The constructed multi-objective weight vector is normalized to ensure that the sum of the weights in each dimension equals 1. Specifically, the normalization formula is:

[0040] in, For the normalized i-th weight value, This represents the i-th original weight value before normalization. The sum of all original weight values, where i and j are indices of the weight vector.

[0041] Preferably, the effectiveness of the weight settings is verified through A / B testing.

[0042] Through the above steps, this invention achieves the accurate construction of multi-dimensional user profiles and the scientific quantification of multi-objective weight vectors, providing personalized optimization guidance for subsequent content component generation and assembly, and ensuring the relevance and effectiveness of SMS content generation.

[0043] Step S2: Use a functional component-based generation method to generate the SMS content in a structured manner.

[0044] Specifically, this step innovatively deconstructs a complete SMS content generation task into the generation process of five independent and functionally distinct content components: trigger components, main components, numerical components, time-sensitive components, and action components. By matching the optimal generation technology to different types of components, fine-grained control over multiple dimensions of SMS content, including structure, creativity, accuracy, and timeliness, is achieved. For trigger and action components, which require high creativity and personalization, large-scale language models are used for dynamic generation; for main and numerical components, which require absolute accuracy, information is securely extracted from a structured database; and for time-sensitive components with strong business logic and timeliness, a rules engine precisely generates them based on preset marketing strategies.

[0045] The beneficial effect of this step is that by breaking down the complex task of generating SMS content into five controllable functional components, it not only utilizes the powerful creative generation capabilities of large language models to enhance the attractiveness and personalization of the content, but also ensures the accuracy of core information and the rigor of business logic through database extraction and rule engine. This solves the contradiction between the lack of flexibility in traditional template filling methods and the poor controllability of content generated directly by large models, and significantly improves the quality and efficiency of content generation.

[0046] Furthermore, the method for generating functional components described in step S2 specifically includes the following sub-steps: Step S210: Generate Trigger Components. Specifically, this step uses a large-scale language model to generate trigger components for the opening of the SMS content, based on the user profile and multi-objective weight vector generated in step S1. This component aims to attract the user's attention immediately and establish an emotional connection through a personalized greeting or address. Preferably, a generative large-scale language model with fine-tuned instructions is used (e.g., a model based on the Transformer architecture with at least 7 billion parameters). The model's input is a structured prompt, which includes at least: user profile tags (e.g., "high-value - price-sensitive"), lifecycle stage (e.g., "churn warning"), the core objective of this marketing campaign (e.g., "increase conversion rate"), and the dimension with the highest weight in the multi-objective weight vector (e.g., "personalization"). The model is instructed to generate N (e.g., N=5) candidate trigger components. For example, for "pre-churn" users, the generated component might be: "Long time no see, we've missed you terribly, we've prepared a special return gift for you."

[0047] Step S220: Generate main components and numerical components. Specifically, this step extracts the main components and numerical components that constitute the core information of the SMS message by querying an internal structured database. The main components primarily contain specific product or service information, while the numerical components contain explicit numerical information such as price, discount, and coupon amount. Preferably, the structured database is a relational database (e.g., MySQL), which contains a product information table and a marketing activity table. Once the business intent is determined (e.g., promoting a product with product ID P001), a Structured Query Language (SQL) instruction is generated to extract the product name and description as the main components. Simultaneously, another query instruction is executed to extract the discount rate and coupon code as the numerical components. This direct extraction method eliminates the possibility of illusions or errors in the generation of core business information, ensuring the absolute accuracy of the information.

[0048] Step S230: Generate Time-Limited Components. Specifically, this step uses a rule engine to generate time-limited components containing the campaign's validity period or time limit, based on the current marketing campaign's strategy and stage. These components aim to create a sense of urgency, encouraging users to take action as soon as possible. Preferably, the rule engine uses a mature business rule management system such as Drools. The marketing strategy is predefined as a series of "WHEN-THEN" rules. For example, a rule can be defined as: "WHEN The marketing campaign status is 'last day' AND the current time is after 18:00 THEN Generate time-limited component content 'Ends at 24:00 tonight'". The engine's input is the current timestamp and campaign ID, automatically matching and executing the corresponding rules to output accurate time-limited information text. This method ensures that the generation of time-limited information strictly adheres to business logic and can automatically update as the marketing progresses, without manual intervention.

[0049] Step S240: Generate Action Component. Specifically, this step is similar to generating the trigger component. A large language model is used to generate an action component for the end of the SMS message. Its content mainly includes a guiding call-to-action phrase and a final redirect link. Preferably, the prompt input to the large language model will additionally include a marketing objective (e.g., "guide users to add to cart" or "guide users to share"), and instruct the model to generate highly persuasive copy. For example, when the objective is "guide users to add to cart," the model might generate "Click [LINK] now to take it home!" Here, [LINK] is a standard link placeholder that will be replaced with a real, encrypted, and tracked short link in the subsequent SMS sending process. By leveraging the copywriting capabilities of the large language model, more diverse and persuasive calls to action can be generated than traditional templates, thereby effectively increasing click-through rates.

[0050] Step S250: Standardized Encapsulation of Components. Preferably, after all components have been generated, a standardized data structure encapsulation is performed to ensure the standardization and efficiency of subsequent processing. Specifically, each generated content component (whether it is a trigger, subject, value, time-sensitive, or action component) is packaged into a structured object containing multiple attributes. This object not only records the final text content of the component but also contains metadata describing its characteristics, such as tags that clearly identify its component type (e.g., "trigger component") and identifiers recording its generation source (e.g., "large language model generation" or "database extraction").

[0051] This standardized encapsulation process provides a consistent and standardized input data format for the content assembly module in the subsequent S3 step. This not only greatly reduces the complexity of the assembly process and makes the assembly logic clearer and more robust, but also facilitates the traceability, analysis, and quality monitoring of the entire content generation process.

[0052] Step S3: Based on the preset syntax rule matrix, perform structured assembly of content components. Specifically, this step receives the five types of content component objects generated in step S2 and standardized and encapsulated as input. Then, based on a preset syntax rule matrix, these independent components are connected according to a predetermined logical order and syntax rules, ultimately generating a candidate SMS message with a complete structure and fluent sentences. This candidate SMS message will be submitted as output to the semantic verification module in step S4 for consistency verification. The beneficial effect of this step is that by separating content generation (step S2) from structure assembly (step S3), a standardized and reproducible SMS message construction process is established. Using the syntax rule matrix not only ensures the structural consistency and professionalism of all generated SMS messages but also greatly improves the flexibility and automation level of content arrangement. More importantly, this mechanism provides a clear and operable path for subsequent targeted corrections based on verification feedback (such as replacing components or adjusting connection methods), effectively improving the closed-loop optimization efficiency from generation to verification and ensuring the quality of the final output content.

[0053] Furthermore, the assembly process based on the preset syntax rule matrix described in step S3 specifically includes: Step S310: Define the basic structural framework of the SMS content. Specifically, this step first establishes the macro-arrangement order of five types of content components in the SMS to form a logically clear and marketing-compliant complete information flow. The opening section will use a trigger component generated by a large language model, containing a personalized greeting or user name, as the starting part of the SMS content, aiming to attract the user's attention immediately. The core information body will combine the main component containing product or service information extracted from the structured database and the numerical component containing price or discount information with the time-sensitive component generated by the rules engine, containing the activity's validity period or time limit, to jointly constitute the core value proposition of the SMS. The closing section will use an action component generated by a large language model, containing explicit call-to-action terms or jump links, as the conclusion of the SMS content, directly guiding the user to the next step.

[0054] Step S320: Query the syntax rule matrix to connect components. Specifically, after determining the macro-order of the components, this step precisely defines the connection method between any two adjacent content components by querying a preset syntax rule matrix. The syntax rule matrix is ​​a data structure that maps content component type pairs to a corresponding assembly rule. During assembly, each pair of adjacent components is processed sequentially to obtain and apply the corresponding assembly rule. This rule includes at least the following three types: relative position constraints, defining a fixed sequential relationship between two components; connectors, words or phrases used to semantically connect two components, such as "only needed" or "discount available"; punctuation marks, punctuation marks used to syntactically separate two components, such as commas, periods, and exclamation marks. Preferably, the syntax rule matrix can be implemented as a hash table, where the key is a tuple representing a component type pair (e.g., (main component, numeric component)) and the value is an object containing the above assembly rule (e.g., {connector: only available, punctuation mark: !}).

[0055] Step S330: Implement a reassembly strategy based on verification feedback. Specifically, when the semantic verification in step S4 fails and returns an error message, this step will be triggered again to perform reassembly. This reassembly process adjusts the assembly strategy according to the specific semantic elements that exceed the preset verification standard, as fed back from step S4. First, the identified problematic semantic elements are mapped to the specific content component type that generated that semantic in step S2. Then, during the reassembly process, a preset correction constraint is applied to the selection of this specific content component type. The correction constraint includes at least the following two methods: Content component replacement involves selecting another unused content component from the multiple candidate contents generated for this component type in step S2. For example, if the initially generated "action component" is judged to be insufficiently guiding in semantic verification, the system will automatically replace it with another action component from the candidate pool (e.g., replacing "Click to learn more" with "Buy Now") and reassemble it.

[0056] Syntax rule adjustments are made to the syntax rules used to connect the content components. For example, if the combination of "body component" and "time-sensitive component" is determined to be semantically incoherent, the system can select an alternative connection rule for this pair of components from the syntax rule matrix during reassembly, such as changing to a more suitable conjunction or punctuation mark.

[0057] Preferably, a priority for the correction strategy can be set. For example, "content component replacement" can be tried first, because replacing the content itself usually brings more significant semantic changes; if the verification still fails after replacing all candidate components, then a more subtle correction, "syntactic rule adjustment," can be tried. Through this closed-loop reassembly mechanism with feedback and correction, the success rate of the generated content finally passing semantic verification is significantly improved.

[0058] Step S4: Implement two-way semantic verification to check the consistency of candidate SMS content.

[0059] Specifically, this step, as the core verification step for content generation quality, receives the candidate SMS content assembled in step S3 as input. By establishing a bidirectional mapping between "forward encoding" from the original business intent to the generated content and "reverse decoding" from the generated content to semantic elements, the semantic deviation between the two is accurately calculated, and the semantic accuracy of the candidate SMS content is adjudicated according to preset verification standards. When the verification passes, the qualified candidate SMS is output to step S5 for timing optimization; when the verification fails, an instruction containing specific deviation information is generated and returned to step S3 to initiate the targeted correction and reassembly process.

[0060] The beneficial effect of this step is that by establishing a two-way semantic mapping between the original business intent and the generated content, the semantic consistency verification between the generated content and the marketing objectives is achieved, effectively avoiding marketing failures caused by semantic deviations and improving the accuracy and reliability of content generation.

[0061] Furthermore, the specific implementation process of the bidirectional semantic verification described in step S4 includes: Step S410: Encode the original business intent and generate a first semantic feature matrix. Specifically, this step parses the original business intent into a set of key semantic elements composed of preset types, and encodes this set of elements into a first semantic feature matrix with preset dimensions. Each dimension of this matrix corresponds to a semantic element of a preset type, thereby transforming the abstract business requirement into a standardized mathematical expression that can be compared by a machine. The types of the key semantic elements include at least product information for identifying products, discount information for defining offers, and timeliness information for indicating urgency.

[0062] Preferably, the original business intent can be defined by a structured JSON object containing all the core parameters of the marketing campaign. The parsing process is accomplished by a rule-based parser that iterates through the JSON object, extracting the values ​​of preset fields as key semantic elements. For example, "product_id: 89757" is mapped to a product information element, "discount_rate: 0.8" to a discount information element, and "end_time: 2025-12-31 23:59:59" to a time-sensitive information element. The resulting first semantic feature matrix can be a fixed-length vector, where different index ranges represent different semantic elements through numerical values, one-hot encoding, or embedding.

[0063] Step S420: Reverse decode the candidate SMS messages to generate a second semantic feature matrix. Specifically, this step uses a parsing model distinct from the large language model used in step S2 to extract key semantic elements of the same type as those preset in step S410 from the natural language text of the candidate SMS content. Subsequently, the extracted elements are encoded into a second semantic feature matrix with the same dimensions and structure as the first semantic feature matrix, ensuring that the two matrices can be directly compared dimension-by-dimensionally. Employing an independent parsing model aims to avoid confirmation biases that may arise from the "self-verification" of the generation model, thereby ensuring the objectivity and reliability of the verification process. Preferably, this parsing model can be a lightweight language model fine-tuned for a specific task (such as Named Entity Recognition, NER) (e.g., a distilled version of BERT or RoBERTa). This model is trained to accurately identify and extract key information such as product name, price, discount, date, and time from SMS text. For example, for the candidate SMS message "

Brand A

[0064] Step S430: Calculate semantic deviation and perform threshold verification. Specifically, this step compares the key semantic elements corresponding to the first and second semantic feature matrices one by one to generate a semantic deviation vector that can characterize the degree of deviation of each semantic element. Then, each degree of deviation in the semantic deviation vector is compared with a preset threshold vector that can set different verification standards for each key semantic element type. When all deviations do not exceed their corresponding verification standards, the verification is deemed successful. This threshold vector-based verification mechanism allows for setting differentiated tolerances for semantic elements of different importance. For example, a zero-tolerance verification standard can be set for core business information such as product ID and discount amount; while for some descriptive text, a certain semantic fluctuation can be allowed. Preferably, the calculation method of semantic deviation depends on the element type: for numerical elements (such as price), the deviation can be the absolute value of the difference between the two; for categorical elements (such as product ID), the deviation can be a binary value of 0 (match) or 1 (mismatch); for textual elements, the deviation can be the difference between the cosine similarity of the two text embedding vectors and 1. An exemplary threshold vector could be set as {Product ID deviation: 0, Discount amount deviation: 0, Activity expiration date deviation: 0, Slogan similarity deviation: 0.3}, which means that the core business information must be completely consistent, while the semantic similarity of the slogan is acceptable if it is not less than 0.7.

[0065] Step S440: Generate verification results and trigger reassembly. Specifically, when the comparison result of step S430 shows that any deviation exceeds its corresponding verification standard, the verification fails. At this time, this step will identify the specific semantic element that exceeds the standard and use the identification result as structured error feedback information. This feedback information will clearly indicate which semantic element (such as "discount information") failed the verification and pass it back to step S3. Step S3 will adjust its content component assembly strategy based on the feedback information to correct the deviation in a targeted manner. If all deviations meet the preset verification standards, the verification passes, and the candidate SMS content is passed to the subsequent step S5. Preferably, the feedback information returned to step S3 can be designed as a structured object containing "problem element type" and "current deviation value". For example, {error_element: discount information, deviation_value: 0.1}. This allows step S3 to accurately locate the problem in the stage of generating "numerical components" and prioritize selecting another content from the candidate pool of that component for replacement, thereby achieving efficient and accurate closed-loop correction.

[0066] Step S5: Iteratively optimize the SMS sequence based on the temporal difference optimization framework. Specifically, this step treats the entire marketing campaign as a multi-stage decision-making process, aiming to address the problem of neglecting temporal correlation and long-term effects in existing technologies. This step receives a set of candidate SMS messages, all semantically validated and generated for multiple time windows within the marketing campaign period in previous steps, as input. By introducing a temporal difference optimization framework, a temporal value function is established that can simultaneously evaluate the immediate reward of a single SMS message and its future value in laying the groundwork for subsequent effects. Subsequently, this function is updated through an iterative optimization process until convergence, ultimately decoding an optimal SMS message sequence that maximizes the expected total return throughout the entire marketing campaign period, and outputting this sequence as the final result.

[0067] The beneficial effect of this step is that by globally optimizing the SMS sequence throughout the entire marketing campaign cycle based on a temporal difference optimization framework, a dynamic balance is achieved between the immediate effect of a single SMS message and the long-term value of the entire sequence. This method overcomes the limitation of traditional optimization strategies that only focus on the optimal state at the current point in time while ignoring its impact on the future, making marketing strategies more forward-looking and strategic, thereby significantly improving the overall return on investment of the marketing campaign throughout its entire lifecycle.

[0068] Furthermore, the iterative optimization process described in step S5 specifically includes the following sub-steps: Step S510: Constructing the Decision Sequence. Specifically, this step decomposes a complete marketing campaign cycle into multiple ordered, connected time windows based on its inherent business rhythm (e.g., pre-launch, peak, and encore periods), and abstracts each time window as a decision state. Simultaneously, the multiple candidate SMS messages generated for each time window in steps S2 to S4, which have passed bidirectional semantic verification, constitute the action set available in that state. Thus, the complex SMS sequence generation problem is successfully transformed into a Markov Decision Process (MDP) with a defined state and action space that can be solved using reinforcement learning methods.

[0069] Step S520: Establish a Time-Series Value Function. Specifically, this step establishes a time-series value function for the decision-making process to quantify the expected total return of any SMS message sequence. The mathematical expression of the function defines the value of an SMS message sequence as the weighted sum of the values ​​generated by each SMS message in the sequence. Specifically, within any time window (state s...),... t Select and send a text message (action a) tThe value generated by the SMS message consists of two parts: the first part is the immediate reward (R) generated by the SMS message within the current time window to achieve the optimization objective of that window (defined by the multi-objective weight vector in step S1). t+1 The second part is that the sending of this text message will transition the status to the next time window (status s). t+1 The expected future value (E[V(s)) generated after laying the groundwork for the effects of all subsequent SMS messages. t+1 At the same time, a discount factor (γ) is introduced to adjust the future value, balancing short-term and long-term interests.

[0070] Step S530: Perform temporal difference iterative optimization. Specifically, this step involves an iterative optimization process to update the temporal value function established in step S520 until it converges. In each iteration, for any time window (state s)... t ) and any of the candidate text messages (action a) t Its value assessment is updated based on the immediate reward R obtained by jumping from this time window to the next time window. t+1 The update is performed based on the latest value assessment results for the next time window. This update method does not require waiting for the entire sequence to end; it only needs to utilize the information from the next state, reflecting the core idea of ​​temporal difference learning. Preferably, the Q-Learning algorithm is used as the specific temporal difference method, and its value function update formula is:

[0071] in, Indicates the state Take action below The value of α is the learning rate, which controls the step size of each update; γ is the discount factor, which balances the importance of immediate rewards and long-term value. Indicates the next state The maximum future value that can be obtained. The iterative process will continue until the change in Q-value of all state-action pairs is less than a preset convergence threshold (e.g., 10). 4).

[0072] Step S540: Select the optimal SMS sequence. Specifically, after the iterative optimization process in step S530 converges and a stable time-series value function (i.e., the Q-value table) is obtained, this step will decode the optimal SMS sequence from this function. The decoding process starts from the start time window of the marketing campaign (initial state s0), selecting candidate SMS messages (actions) with the largest Q-value in this state. This is the first text message in the sequence. Then, it proceeds to... The next time window (state s1) is then selected, and the candidate SMS a1 with the largest Q value in that state is chosen again. This greedy strategy is repeated, selecting the currently optimal SMS in each time window until the marketing campaign ends. Finally, all the selected SMS messages are arranged in chronological order to form a unique sequence of SMS messages with the highest expected total return, which is the final generated result.

[0073] Example 2: like Figure 2 As shown, based on the same inventive concept as the multi-objective optimized SMS content generation artificial intelligence method provided in Embodiment 1, this embodiment of the invention also provides a multi-objective optimized SMS content generation artificial intelligence system, the system comprising: The user profile and weight module is used to construct multi-dimensional user profiles based on the differentiated optimization priorities in each time window of the marketing campaign cycle, and to quantify the optimization priorities into a multi-objective weight vector by combining the business scenario with the user profiles. The content component generation module is used to structure the SMS content to be generated into five types of content components: trigger components, main components, numerical components, time-sensitive components, and action components, using a functional component-based generation method. The content assembly module is used to assemble the five types of content components into candidate SMS content according to a preset syntax rule matrix. The semantic verification module is used to implement two-way semantic verification, which verifies the candidate SMS content by calculating the semantic deviation between the original business intent and the candidate SMS content; The sequence optimization module is used to control the content component generation module, content assembly module, and semantic verification module to generate multiple verified candidate SMS messages corresponding to the optimization focus of each time window for multiple time windows in the marketing campaign cycle. The candidate SMS messages are combined to form an SMS sequence, and the SMS sequence is iteratively optimized through a time-series value function based on the temporal differential optimization framework to achieve a balance between the optimal goal of the current time window and the preparation for the effect of subsequent SMS messages.

[0074] The user profile and weighting module is specifically used to execute step S1 of Embodiment 1. The core function of this module is to achieve precise modeling of marketing objectives and quantitative expression of personalized strategies, providing accurate optimization guidance for the subsequent content generation module. Through this module, the scientific quantification and weight allocation of multiple marketing objectives are realized, solving the problem of conflicting objectives in traditional single-objective optimization and significantly improving the scientific nature of marketing decisions. Furthermore, the user profile and weighting module in this embodiment of the invention is also used to execute the following steps: Specifically, the module extracts users' historical behavioral characteristics, which include at least click-through rate, conversion rate, and active time periods; analyzes users' preference patterns, which include at least content length preference, promotional sensitivity, and interaction frequency; and identifies users' lifecycle stages, which include at least new customer, active, dormant, and churn warning stages. Specifically, based on the above user profile, the module constructs a multi-objective weight vector, whose dimensions include at least four core dimensions: conversion rate, content simplicity, compliance, and personalization, and dynamically adjusts the multi-objective weight vector according to different stages of the marketing campaign. Preferably, when identifying user lifecycle stages, this module uses the k-means clustering algorithm to stratify new customers and the RFM model to identify dormant users. When dynamically adjusting weights, the multi-armed slot machine algorithm from reinforcement learning is used for online weight optimization, and the effectiveness of the weight settings is verified through A / B testing.

[0075] Content Component Generation Module. Specifically, the content component generation module is used to execute step S2 of Embodiment 1. This module deconstructs a complete SMS content generation task into the generation process of five independent and functionally defined content components. The beneficial effect of this module is that by decomposing the complex SMS content generation task into five controllable functional components, it utilizes the powerful creative generation capabilities of a large language model to enhance the attractiveness and personalization of the content, while ensuring the accuracy of core information and the rigor of business logic through database extraction and rule engine. This solves the contradiction between the lack of flexibility in traditional template filling methods and the poor controllability of content directly generated by large models, significantly improving the quality and efficiency of content generation. Furthermore, the content component generation module of this embodiment of the invention is also used to execute the following steps: Specifically, this module uses a large language model to generate a trigger component for the beginning of the SMS content and an action component for the end of the SMS content based on the user profile and multi-objective weight vector output by the profile and weight module. Specifically, this module extracts the main components (including product or service information) and numerical components (including price, discount, and other numerical information) that constitute the core information of the SMS by querying the internal structured database. Specifically, this module uses a rules engine to generate time-sensitive components that include the campaign's validity period or time limits, based on the current marketing campaign's strategy and stage. Preferably, after all components are generated, the module performs a standardized data structure encapsulation, packaging each component into a structured object containing text content and metadata (such as component type and generation source), providing a consistent and standardized input data format for subsequent content assembly modules.

[0076] Content Assembly Module. Specifically, the content assembly module is used to execute step S3 of Embodiment 1. This module receives five types of content component objects generated by the content component generation module, connects them according to a preset syntax rule matrix, and generates candidate SMS content with complete structure and fluent sentences. The beneficial effect of this module is that by separating content generation from structure assembly, a standardized and reproducible SMS construction process is established. Using the syntax rule matrix, not only is the consistency and professionalism of all generated SMS messages in terms of structure guaranteed, but the flexibility and automation level of content arrangement are also greatly improved, and a clear and operable path is provided for subsequent targeted correction based on verification feedback, effectively improving the efficiency of closed-loop optimization. Furthermore, the content assembly module of this embodiment of the invention is also used to execute the following steps: Specifically, the module first establishes the basic structural framework of the SMS content, that is, the trigger component is used as the opening, the body, numerical and time-sensitive components constitute the core information, and the action component is used as the ending. Specifically, the module precisely defines the connection method between any two adjacent content components by querying the preset syntax rule matrix. The assembly rules include at least the relative position constraints between the two, the conjunctions used for semantic connection, and the punctuation marks used for grammatical separation. Specifically, when a verification failure feedback is received from the semantic verification module, the module will adjust the assembly strategy based on the specific semantic elements indicated in the feedback and initiate a reassembly process. The adjustment strategy includes at least selecting another candidate content component of the same type for replacement, or adjusting the syntax rules used to connect the content components.

[0077] Semantic verification module. Specifically, the semantic verification module is used to execute step S4 of embodiment one. As the core verification link of content generation quality, this module establishes a bidirectional mapping between the "forward encoding" from the original business intent to the generated content and the "reverse decoding" from the generated content to semantic elements, accurately calculates the semantic deviation between the two, and adjudicates the semantic accuracy of the candidate SMS content. The beneficial effect of this module is that by establishing a bidirectional semantic mapping between the original business intent and the generated content, it realizes the semantic consistency verification between the generated content and the marketing objectives, effectively avoids the marketing failure problem caused by semantic deviation, and improves the accuracy and reliability of content generation. Further, the semantic verification module of this embodiment of the invention is also used to execute the following steps: Specifically, the module parses and encodes the original business intent into a first semantic feature matrix with a preset dimension; and uses a parsing model that is different from the large language model used by the content component generation module to reverse extract key semantic elements from the candidate SMS content and encode them into a second semantic feature matrix with the same dimension and structure as the first semantic feature matrix. The types of the key semantic elements include at least product information, discount information, and timeliness information. Specifically, this module compares the key semantic elements in the two semantic feature matrices one by one, generating a semantic deviation vector representing the degree of deviation of each element. Then, it compares each deviation degree in this vector with a preset threshold vector that sets different verification standards for each key semantic element type. Specifically, when any deviation degree exceeds its corresponding verification standard, the module determines that the verification has failed, identifies the specific semantic element that exceeded the standard, generates structured error feedback information, and returns it to the content assembly module to trigger reassembly.

[0078] Sequence Optimization Module. Specifically, the sequence optimization module is used to execute step S5 of Embodiment 1. This module treats the entire marketing campaign as a multi-stage decision-making process, aiming to solve the problem of neglecting temporal correlation and long-term effects in the prior art. The beneficial effect of this module is that by globally optimizing the SMS sequence of the entire marketing campaign cycle based on the temporal differential optimization framework, a dynamic balance is achieved between the immediate effect of a single SMS and the overall long-term value of the sequence. This method overcomes the limitation of traditional optimization strategies that only focus on the current point in time and ignore its impact on the future, making the marketing strategy more forward-looking and strategic, thereby significantly improving the overall return on investment of the marketing campaign throughout its entire life cycle. Further, the sequence optimization module of this embodiment of the invention is also used to execute the following steps: Specifically, the module decomposes a complete marketing campaign cycle into multiple ordered time windows, and generates a candidate SMS set consisting of multiple verified candidate SMS messages for each time window, thereby constructing a decision sequence. Specifically, the module establishes a temporal value function, which defines the expected total return of any SMS sequence as the sum of the immediate reward and future value generated by each SMS in the sequence, and balances short-term and long-term interests through a discount factor. Specifically, this module updates the time-series value function through an iterative optimization process until it converges. Specifically, after the evaluation result converges, the module selects the SMS messages with the highest expected return in each time window, starting from the initial time window, ultimately constructing the optimal SMS sequence as the output. Preferably, this module employs the Q-Learning algorithm as the specific time-series differencing method.

[0079] The above-described embodiments are merely preferred embodiments of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The modules and steps of the system (Embodiment 2) and method (Embodiment 1) described herein can be referred to in correspondence with each other. Those skilled in the art should understand that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention, based on the technical solutions disclosed in the present invention, without creative effort, should be included within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the scope of protection set forth in the claims.

Claims

1. A multi-objective optimization-based artificial intelligence method for SMS content generation, characterized in that, Includes the following steps: S1: Based on the differentiated optimization priorities in each time window of the marketing campaign cycle, construct a multi-dimensional user profile, and combine the business scenario with the user profile to quantify the optimization priorities into a multi-objective weight vector; S2: A functional component-based generation method is adopted to structure the SMS content to be generated into five types of content components, including trigger components, main components, numerical components, time-sensitive components and action components. The trigger components and action components are generated using a large language model, the main components and numerical components are directly extracted from a structured database, and the time-sensitive components are generated by a rule engine based on marketing strategies. S3: According to the preset syntax rule matrix, assemble the five types of content components into candidate SMS content; S4: Implement bidirectional semantic verification, encode the original business intent into a first semantic feature matrix, use a parsing model that is different from the large language model to reverse decode the candidate SMS content, generate a second semantic feature matrix, calculate the semantic deviation between the first semantic feature matrix and the second semantic feature matrix, and pass the verification when the semantic deviation meets the preset verification standard; otherwise, return to step S3 to reassemble. S5: For multiple time windows in the marketing campaign cycle, repeat steps S2 to S4 to generate multiple verified candidate SMS messages corresponding to the optimization focus of each time window. Combine the candidate SMS messages to form an SMS sequence. Then, based on the time-series differential optimization framework, iteratively optimize the SMS sequence through the time-series value function so that each SMS message in the sequence achieves the optimal goal of the current time window while laying the groundwork for the effect of subsequent SMS messages.

2. The multi-objective optimization-based artificial intelligence method for SMS content generation according to claim 1, characterized in that, In step S1, constructing the multi-dimensional user profile and multi-objective weight vector includes: Extract users' historical behavioral characteristics, which include at least click-through rate, conversion rate, and active time periods; Analyze user preference patterns, which include at least content length preference, promotion sensitivity, and interaction frequency; Identify user lifecycle stages, which include at least new customer, active, dormant, and churn warning stages; The dimensions of the multi-objective weight vector include at least conversion rate, content simplicity, compliance, and personalization.

3. The multi-objective optimization-based artificial intelligence method for SMS content generation according to claim 1, characterized in that, The method for generating functional components in step S2 specifically includes: The trigger component, which contains a greeting or user address and is generated by the large language model, is used as the opening part of the SMS content; The main component containing product or service information and the numerical component containing price or discount information extracted from the structured database, together with the timeliness component containing the activity validity period or time limit generated by the rule engine according to the rules preset for the current activity stage based on the marketing strategy, constitute the core information body of the SMS content; The action component, which is generated by the large language model and contains call-to-action terms or jump links, is used as the end part of the SMS content.

4. The multi-objective optimization-based artificial intelligence method for SMS content generation according to claim 1, characterized in that, In step S3, the preset syntax rule matrix is ​​a data structure that maps content component type pairs to a corresponding assembly rule. The assembly obtains the assembly rule corresponding to any two adjacent content component types by querying the data structure, and connects the content components according to the rule. The assembly rule is used to define the connection relationship between the two content component types. The connection relationship includes at least the relative position constraints between the two, the connectors used to connect the two semantically, and the punctuation marks used to separate the two grammatically.

5. The multi-objective optimization-based artificial intelligence method for SMS content generation according to claim 1, characterized in that, The bidirectional semantic verification in step S4 includes: The original business intent is parsed into a set of key semantic elements composed of preset types, and the set of elements is encoded into a first semantic feature matrix with preset dimensions, wherein each dimension of the matrix corresponds to a semantic element of a preset type. Using the parsing model that is distinct from the large language model, key semantic elements of the same type as the preset type are extracted from the candidate SMS content and encoded into a second semantic feature matrix with the same dimension and structure as the first semantic feature matrix; The types of the key semantic elements include at least product information for identifying products, discount information for defining offers, and timeliness information for indicating urgency.

6. The multi-objective optimization-based artificial intelligence method for SMS content generation according to claim 1, characterized in that, Step S4 further includes: The evaluation of the semantic deviation specifically includes comparing the key semantic elements corresponding to the first semantic feature matrix and the second semantic feature matrix one by one to generate a semantic deviation vector that can characterize the degree of deviation of each semantic element. The semantic deviation meets the preset verification standard, which specifically means that each degree of deviation in the semantic deviation vector is compared with a preset threshold vector that can set different verification standards for each key semantic element type. When all degrees of deviation do not exceed their corresponding verification standards, the verification is passed. The return step S3 reassembly involves identifying the specific semantic element that exceeds the corresponding verification standard when any deviation exceeds the standard, and adjusting the content component assembly strategy adopted in step S3 based on the identification result to correct the deviation in a targeted manner.

7. The multi-objective optimization-based artificial intelligence method for SMS content generation according to claim 6, characterized in that, The content component assembly strategy adopted in the adjustment step S3 includes: mapping the identified specific semantic elements to the specific content component type corresponding to the generation of the semantic elements in step S2; and applying a preset correction constraint to the selection of the specific content component type during the reassembly process in step S3. The correction constraint includes at least selecting another content component from the candidate content components of the same type for replacement, or adjusting the syntax rules used to connect the content components.

8. The multi-objective optimization-based artificial intelligence method for SMS content generation according to claim 1, characterized in that, The iterative optimization in step S5 includes the following steps: Construct a decision sequence, break down the entire marketing campaign cycle into multiple ordered time windows; and for each time window, generate a set of candidate SMS messages consisting of multiple verified candidate SMS messages; A time-series value function is established, which defines the expected total return of any SMS sequence as the sum of the values ​​generated by each SMS in the sequence. The value generated by each SMS consists of two parts: the first part is the immediate reward generated by the SMS within its time window for achieving the optimization goal of that window; the second part is the future value generated by the SMS for laying the groundwork for the effects of other SMS in subsequent time windows. The value is adjusted by a weight used to balance short-term and long-term interests. Temporal difference iterative optimization is performed, and the temporal value function is updated until it converges through an iterative optimization process. In each iteration, the value assessment of any time window is updated based on the immediate reward obtained from the transition from the current time window to the next time window and the latest value assessment result of the next time window. After the evaluation results converge, the SMS sequence with the highest expected total return is selected as the optimal SMS sequence.

9. A multi-objective optimization-based artificial intelligence system for generating SMS content, characterized in that, include: The user profile and weight module is used to construct multi-dimensional user profiles based on the differentiated optimization priorities in each time window of the marketing campaign cycle, and to quantify the optimization priorities into a multi-objective weight vector by combining the business scenario with the user profiles. The content component generation module is used to structure the SMS content to be generated into five types of content components: trigger components, main components, numerical components, time-sensitive components, and action components, using a functional component-based generation method. The content assembly module is used to assemble the five types of content components into candidate SMS content according to a preset syntax rule matrix. The semantic verification module is used to implement two-way semantic verification, which verifies the candidate SMS content by calculating the semantic deviation between the original business intent and the candidate SMS content; The sequence optimization module is used to control the content component generation module, content assembly module, and semantic verification module to generate multiple verified candidate SMS messages corresponding to the optimization focus of each time window for multiple time windows in the marketing campaign cycle. The candidate SMS messages are combined to form an SMS sequence, and the SMS sequence is iteratively optimized through a time-series value function based on the temporal differential optimization framework to achieve a balance between the optimal goal of the current time window and the preparation for the effect of subsequent SMS messages.

Images