Short message content compliance generation method and system based on fusion semantic large model

Abstract

The application relates to the technical field of communication, and provides a short message content compliance generation method and system fusing a semantic large model. The method comprises the following steps: performing coarse identification on short message content to be processed by a service provider, and obtaining a short message type and user demand; inputting the short message content into a semantic large model for deep semantic analysis, and outputting short message semantic analysis information; performing multi-dimensional compliance comparison on the short message semantic analysis information according to the short message type and the user demand, and identifying risk description content; and according to the short message type and the user demand, optimizing and adjusting the risk description content according to semantic analysis information of the risk content, and generating target short message content. The application solves the technical problem that short message content generation in communication services relies on templates and cannot dynamically identify and avoid compliance risks according to user demand, realizes deep semantic analysis and multi-dimensional risk comparison through a semantic large model, and improves the intelligent level and compliance guarantee capability of short message content generation in communication services.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of communication, in particular to a short message content compliance generation method and system fusing a semantic large model. BACKGROUND

[0002] As an important information transmission method in communication services, short messages are widely used in many industries such as business, finance, medical treatment, and government affairs. In these industries, short message content often involves sensitive content such as user privacy, financial information, and medical data. Therefore, it is crucial to ensure the compliance of short message content. Currently, short message content review mainly relies on manual experience and rule library matching. However, in the face of massive short message processing needs, manual review is inefficient and subjective, and it is difficult to accurately identify risks in complex scenarios such as semantic ambiguity and context association. For example, in financial industry marketing short messages, inducement expressions such as "high returns without risk" may evade basic rule library detection through simple keyword replacement, and manual review may miss judgment due to insufficient semantic understanding. In medical short messages, hidden expressions related to efficacy promises are often difficult to be captured by traditional automated systems due to complex context logic. At the same time, communication service regulations require dynamic adjustment according to industry policies, such as frequent updates of Internet finance advertising banned word lists. Traditional static rule libraries are difficult to adapt in time, leading to compliance review lagging behind regulatory changes and increasing enterprise operational risks. In addition, there is often a conflict between user demand for short message expression effect (such as marketing copy appeal and notification content clarity) and compliance requirements, and there is a lack of intelligent dynamic balancing mechanism, which may cause excessive review and damage service efficiency. Therefore, how to ensure short message content compliance while improving the intelligent level of content generation and user experience has become a technical problem to be solved. SUMMARY

[0003] The present application provides a short message content compliance generation method and system fusing a semantic large model, aiming to solve the technical problem that short message content generation in communication services relies on templates and cannot dynamically identify and avoid compliance risks according to user demand.

[0004] The first aspect of the present application provides a short message content compliance generation method fusing a semantic large model, which comprises: performing coarse identification on short message content to be processed by a service provider to obtain short message type and user demand; inputting the short message content into a semantic large model for deep semantic analysis to output short message semantic analysis information; performing multi-dimensional compliance comparison on the short message semantic analysis information according to the short message type and user demand to identify risk description content; and optimizing and adjusting the risk description content according to the short message type and user demand according to the semantic analysis information of the risk content to generate target short message content.

[0005] In another aspect of the present application, a short message content compliance generation system based on a semantic large model is provided, which comprises: a coarse identification unit that coarsely identifies short message content to be processed by a service provider to obtain a short message type and user demand; a deep semantic analysis unit that inputs the short message content into a semantic large model for deep semantic analysis to output short message semantic analysis information; a multi-dimensional compliance comparison unit that compares the short message semantic analysis information in multiple dimensions according to the short message type and user demand to identify risk description content; and an optimization adjustment unit that optimizes and adjusts the risk description content according to the short message type and user demand based on semantic analysis information of the risk content to generate target short message content.

[0006] The one or more technical solutions provided in the present application have at least the following technical effects or advantages:

[0007] The short message content compliance generation method based on a semantic large model described above first analyzes the short message content to be processed submitted by a service provider to identify which category (such as marketing, notification, etc.) the short message belongs to and the specific expression demand of the user. Then, the short message text is input into a trained semantic large model for deep semantic understanding to extract semantic information expressed in the short message. Subsequently, the semantic information analyzed is compared in multiple dimensions for compliance according to the identified short message type and user demand to detect whether there is risk expression that may violate industry standards or laws and regulations. Finally, the risk content identified is combined with the semantic information for targeted content optimization and adjustment to ensure that the user's expression intention is met while generating compliant, safe, and sendable short message content.

[0008] The above description is only a summary of the technical solutions of the present application. In order to more clearly understand the technical means of the present application, the content of the specification can be implemented, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following specific embodiments of the present application are described. BRIEF DESCRIPTION OF DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0010] Figure 1 A flowchart of a short message content compliance generation method based on a semantic large model in an embodiment.

[0011] Figure 2 A short message content compliance generation system architecture diagram based on a semantic large model in an embodiment.

[0012] Figure labeling: Coarse identification unit 11, deep semantic parsing unit 12, multi-dimensional compliance comparison unit 13, optimization and adjustment unit 14. Detailed Implementation

[0013] This application provides a method and system for generating compliant SMS content based on a fused semantic big model, thereby solving the technical problem that SMS content generation in communication services relies on templates and cannot dynamically identify and avoid compliance risks based on user needs.

[0014] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0015] It should be noted that the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such process, method, product, or device.

[0016] Example 1, as Figure 1 As shown, this application provides a method for compliant SMS content generation based on a fusion semantic big model, the method comprising:

[0017] The system performs a preliminary identification of the text messages to be processed by the service provider to obtain the text message type and user needs.

[0018] In this embodiment, the original SMS content submitted by the service provider undergoes preliminary classification and intent recognition. Specifically, the system first receives the SMS text to be processed by the service provider and extracts key features, such as keywords and sentence structure. Subsequently, combining the service scenario corresponding to the SMS (such as finance, healthcare, government affairs, etc.) and the purpose of the SMS (such as marketing promotion, user notification, service reminder, etc.), a pre-trained logistic regression model is used for coarse identification, i.e., preliminary classification and intent recognition, outputting the specific type of the SMS and the user needs it reflects. In this way, clear scenario labels and user intent basis can be provided for subsequent semantic analysis and compliance processing, which helps to improve the targeting and accuracy of compliance processing.

[0019] Furthermore, this application provides a preliminary identification of the SMS content to be processed by the service provider to obtain the SMS type and user needs, including:

[0020] An industry type and a short message demand type for short message compliance auditing are obtained; data labels are made according to the industry type and the short message demand type, and training data are constructed; a logical regression model is trained by using the training data, and a semantic coarse recognition channel is constructed, wherein the logical regression model is used to identify and output the short message type and user demand, the short message type corresponds to the industry type, and the user demand corresponds to the short message demand type.

[0021] Preferably, the industry type and the short message demand type for short message compliance auditing are obtained based on current communication service compliance standards, industry supervision regulations and short message business access range, wherein the industry type is an industry classification sorted according to objects and content sources served by a short message platform, for example, finance, medical treatment, government affairs, education, e-commerce, transportation and the like; and the short message demand type is classified into types such as marketing (such as advertisement promotion), notification (such as verification code, bill reminder), service (such as progress query, service reservation) and the like according to a functional positioning of the short message in a business process. Subsequently, a plurality of short message samples are collected, and each short message sample is labeled with the industry type and the short message demand type, thereby constructing training data. After text preprocessing of the training data, that is, performing word segmentation, removing stop words, special character cleaning and the like on the short message samples, the preprocessed short message samples are converted into numerical feature vectors by using TF-IDF (Term Frequency-Inverse Document Frequency), Bag-of-Words and the like, so as to enable the logical regression model to receive and learn. Subsequently, a multinomial logistic regression model is selected as a core classifier to construct the semantic coarse recognition channel, and key parameters such as task type (multinomial), optimizer (lbfgs), regularization strength (1.0) and the like are initialized, and the training data are divided into a training set and a validation set. The numerical feature vectors are used as input features, and the corresponding labels (industry type, short message demand type) are used as target variables, and training of the semantic coarse recognition channel is carried out. In the training process, the semantic coarse recognition channel optimizes its weight parameters by iteration to minimize a loss function (cross-entropy loss) between a predicted result and a true label, thereby gradually improving classification accuracy of the channel until convergence or a preset maximum number of iterations is reached. After the training is completed, the model is evaluated by using the validation set, and classification accuracy, recall rate, F1-score and the like are calculated to ensure generalization ability of the model on unseen data, and model hyperparameters (such as a regularization parameter) are further adjusted according to the evaluation result, and finally accurate classification of the industry type and the demand type of the short message is realized, wherein the industry type identified by the semantic coarse recognition channel is output as the short message type, and the short message demand type identified by the semantic coarse recognition channel is output as the user demand, which is used to provide a pre-input condition for subsequent deep semantic analysis and compliance judgment, and helps to perform more accurate semantic analysis and risk comparison.

[0022] Further, the application provides that the industry type at least includes: financial industry, medical industry, government announcement, and the short message demand type at least includes marketing, notification, service.

[0023] Optionally, the industry type refers to the business field involved in the short message, which is used to reflect the service provider background or business scope of the short message, and at least includes financial industry, medical industry, and government announcement. The financial industry involves banking, insurance, securities, loans, etc., such as credit card repayment reminders, loan approval notifications, investment and financial planning promotion, etc. The medical industry covers hospitals, physical examination centers, health management platforms, etc., such as appointment notification, physical examination report reminder, and vaccination reminder. The government announcement is a notification type short message sent by a government agency or a government service platform, such as community announcement, epidemic prevention propaganda, and policy change notification. The short message demand type refers to the actual use or intention of the short message, which is used to determine the scene or purpose of the short message, and at least includes marketing, notification, and service. The marketing type is a short message used to promote products or services, such as merchant promotion activities, time-limited discounts, and member discounts. The notification type is a short message used to convey important information to users, such as verification code, bill reminder, and appointment success notification. The service type is a short message used to provide auxiliary service support, such as customer service follow-up, satisfaction survey, and system upgrade prompt. By analyzing the short message text, the industry type and demand type to which it belongs can be identified, thereby providing accurate context information for subsequent semantic analysis and compliance processing.

[0024] The short message content is input into a semantic large model for deep semantic analysis, and short message semantic analysis information is output.

[0025] In one embodiment, when a short message content to be processed is obtained, a semantic large model is used to deeply understand and analyze the semantic structure of the short message content. Specifically, the preprocessed short message text is input into the constructed semantic large model, which is usually based on a deep neural network architecture and includes multiple layers such as input layer, encoding layer, context understanding layer, and output layer. By training the semantic large model, a deep semantic analysis channel can be built, which can not only identify explicit information in the short message, but also deeply mine implicit semantic relationships, providing accurate semantic support for subsequent compliance comparison and content optimization. Finally, the deep semantic analysis channel will perform deep semantic analysis on the input short message content based on the learned knowledge, thereby outputting short message semantic analysis information to help determine whether the content is compliant and which parts may have risks, thereby improving the intelligent level and compliance guarantee capability of short message content generation in communication services.

[0026] Further, the application provides inputting the short message content into the semantic large model for deep semantic analysis, and outputting short message semantic analysis information, including:

[0027] The large model framework is constructed, including an input layer, an encoder layer, a context understanding layer, a semantic analysis layer, and an output layer. The large model framework is trained and converged by collecting the established training data set, and the semantic large model is obtained. A deep semantic analysis channel is built based on the semantic large model, which is used for deep semantic analysis of short message content and outputs the short message semantic analysis information.

[0028] Optionally, the semantic large model framework designed and built based on a deep neural network includes five core structure layers, i.e., an input layer, an encoder layer, a context understanding layer, a semantic analysis layer, and an output layer. The input layer is responsible for receiving preprocessed short message texts. The encoder layer supports a multi-head attention mechanism, enabling the model to simultaneously focus on information at different positions in the text. The context understanding layer can determine whether a word has a risk meaning in different contexts, providing a basis for subsequent processing. The semantic analysis layer further deepens the understanding of the overall semantics of the short message, extracts deep semantic features, including logical relationships, expression intentions, emotional tendencies, etc. The output layer integrates the features extracted by the previous layers and outputs the semantic analysis information of the short message, which is used for downstream compliance comparison and optimization processing. After completing the model structure design, the model training stage is entered. Specifically, training data sets covering multiple industries and multiple types of short messages are collected and constructed. Each short message is semantically annotated, such as sensitive word position, semantic relationship label, and potential risk level. The annotated training data is input into the constructed large model framework, and supervised learning is used for training. In this process, the training data is input into the model for forward propagation, and the predicted short message semantic analysis information is calculated through layer-by-layer transmission. The error between the predicted result and the true label is calculated through the cross-entropy loss function, and the model parameters are iteratively adjusted combined with optimization algorithms such as Adam or SGD until the loss function converges, the model performance is stable, and the accuracy reaches the set standard. When the training is completed and verified by the validation set, the trained semantic large model is deployed as a deep semantic analysis channel. In practical applications, the channel receives new short message texts, performs semantic decomposition, context evaluation, risk identification, and other operations, and outputs structured short message semantic analysis information, providing semantic support for subsequent multi-dimensional compliance comparison and content optimization.

[0029] Further, the application provides constructing a large model framework, including:

[0030] The input layer is used to process text embedding; the encoding layer adopts a self-attention mechanism to capture the dependency relationship of the text, and through multi-head attention, the model can simultaneously focus on different parts of the text; the context understanding layer is used to detect sensitive words and identify information related to sensitive words in the local context; the semantic analysis layer uses a multi-layer encoder to extract deep semantic features and further analyze the deep association between sensitive words and global context; and the output layer is used to integrate sensitive words, deep semantic features and context association, and output SMS semantic analysis information.

[0031] Optionally, after preprocessing (such as removing stop words, word segmentation, and cleaning special characters), the SMS text is input into the input layer of the model. The task of the input layer is to convert the original text information into a numerical vector form that the model can recognize. The specific method is the same as described above, thereby forming a complete text vector sequence that retains the basic semantic features of the vocabulary and serves as the basis for subsequent in-depth analysis.

[0032] The encoding layer is used to model the dependency relationship between words and adopts a self-attention mechanism to capture the relative importance of each word in the context. Through the multi-head attention mechanism, the model can learn the multi-dimensional relationship between words in parallel from multiple subspaces, such as syntax structure, syntax dependency, and entity association, thereby enhancing the model's understanding of semantic diversity. The output of the encoding layer is a set of context-dependent word vectors that represent the semantic position of each word in the entire SMS text.

[0033] After obtaining the context-aware encoding result, the context understanding layer is responsible for detecting and analyzing sensitive words. By quickly matching the SMS content with a pre-set risk word library, the model can locate possible risky words. Subsequently, based on the dependency relationship output by the encoding layer, the model can identify the local context centered on these sensitive words, helping it determine whether the sensitive words truly have risk meanings in a specific context and avoiding misjudgment caused by "semantic ambiguity".

[0034] The semantic analysis layer further constructs the deep semantic structure of the SMS based on the context understanding result. Semantic analysis usually consists of multiple layers of encoders that perform more in-depth abstraction and feature extraction on the entire SMS content. Through the semantic analysis layer, the model can output highly abstract semantic labels such as risk description, compliance expression, and ambiguous intent, providing a structural basis for subsequent processing.

[0035] Finally, the output layer integrates the information from the encoding layer, the context understanding layer, and the semantic parsing layer to form a set of structured output content, including the detected sensitive words and their positions in the text, the context semantic environment and dependency relationship of the sensitive words, the overall deep semantic label of the short message (such as intent, expression risk level), risk intensity score, or type judgment of potential violation risk, etc. These output results constitute the final short message semantic parsing information, which serves as the input basis for subsequent multi-dimensional compliance comparison and content optimization adjustment.

[0036] In summary, this semantic large model has a clear hierarchical structure, from bottom-level vocabulary understanding to high-level semantic reasoning, ensuring that short message content can be fully parsed at different levels and potential compliance risks can be effectively identified, providing a solid foundation for intelligent compliance generation.

[0037] Further, the present application provides inputting the short message content into the semantic large model for deep semantic parsing, and outputting short message semantic parsing information, including:

[0038] The short message content is preprocessed, including cleaning and removing irrelevant characters, and inputting into the semantic large model; the input layer converts the preprocessed short message content into an embedding vector; the encoding layer captures the dependency relationship between each word in the short message content based on the self-attention mechanism, and assigns each word a weight representing its importance in the current context; the context understanding layer identifies sensitive words by matching the short message content with a sensitive word list, and analyzes the detected sensitive words based on the dependency relationship output by the encoding layer to determine the meaning and risk of the sensitive words in the current context; based on the identified sensitive words and local context relationships, the semantic parsing layer uses a multi-layer encoder structure to extract deep semantic features of the short message content and further analyze the deep association relationship between the sensitive words in the global context; the identified sensitive words, local context relationships, deep semantic features, and deep association relationships are connected to output the short message semantic parsing information.

[0039] Optionally, first, the input short message content is cleaned and preprocessed, which mainly includes removing irrelevant characters (such as punctuation marks, special symbols), and standardizing the text format (such as uniform case, removing stop words), so that the processed text is convenient for subsequent analysis and ensures the purity of the information. On this basis, through word segmentation and other operations, the text is converted into a form that can be processed by the model, ready to be input into the semantic large model for analysis.

[0040] The preprocessed SMS content is input into the input layer, which primarily functions to convert the textual content into numerical embedding vectors. In this way, each word in the SMS is represented as a high-dimensional vector, allowing the model to understand the semantics of the words based on these vectors. The output of this stage is a complete sequence of SMS vectors, representing the overall information structure of the SMS.

[0041] In the encoding layer, the model captures the dependencies between words in the SMS content by employing self-attention mechanisms and multi-head attention mechanisms. The self-attention mechanism allows the model to dynamically adjust the weight of each word based on its relationship with other words when processing each word. Specifically, the model calculates the relevance of each word to other words through a "Query- Key" scoring mechanism in the attention mechanism, generating a set of weights that reflect the importance of different words in the current context. Through multi-head attention, the model allows simultaneous attention to different parts of the text from multiple perspectives, with each attention head independently calculating a set of weights, enabling the model to understand different semantic levels of the SMS in multiple subspaces. This mechanism effectively improves the model's ability to understand complex semantic structures in the text, ensuring that the importance of each word in the current context is accurately assessed.

[0042] In the context understanding layer, the model matches the SMS content with a list of sensitive words extracted from the risk word library, identifying potential sensitive words in the SMS. Once the sensitive words are identified, the model further analyzes their local context in the text based on the dependencies output by the encoding layer. Specifically, the model extracts surrounding words within the context of these sensitive words, analyzes their dependency relationships and co-occurrence frequencies through a sliding window mechanism, and determines the specific meaning of the sensitive words in the current context and their potential risks. This step is a critical risk identification phase that ensures the model can identify potential violations in the SMS.

[0043] The semantic analysis layer uses a multi-layer encoder structure to conduct deeper semantic analysis on the identified sensitive words and local context relationships, extracting deep semantic features of the SMS. In particular, this layer not only focuses on local context but also further analyzes the association between sensitive words and global context. Through multi-layer encoders, the model can deeply understand the semantic structure of the entire SMS and infer the specific meaning and risks of sensitive words in the global context, such as determining whether words like "transfer," "winning," and "privacy" are in a real service context or a marketing-induced context. The output of this layer is highly abstract semantic features that can fully reflect the potential intent and compliance risks of the SMS content.

[0044] Finally, at the output layer, the model integrates the identified sensitive words, local context relationships, deep semantic features, and global correlation relationships to generate structured SMS semantic analysis information, including sensitive words and their positions in the text, local context information (semantic associations between sensitive words and surrounding words), deep semantic features (intent, sentiment, potential risks, etc.), and risk assessment of sensitive words in the global context. Through these information, compliance judgment can be made on the SMS content, providing a basis for subsequent risk comparison and optimization adjustment, thereby achieving accurate semantic analysis and compliance analysis of SMS content.

[0045] Further, the application provides a method for determining the meaning and risk of a sensitive word in a current context, comprising:

[0046] According to the weight of the sensitive word, the size of the moving window is configured; based on the moving window, the local context of the sensitive word is selected, the co-occurrence frequency of the sensitive word and the directly associated words within the moving window size is identified, the grammatical relationship between the words is analyzed according to the dependency relationship between the words, and the meaning and risk of the directly associated word combination are identified; the window size is adjusted according to the risk within the moving window and the similarity of the words at the edge of the window, and based on the adjusted moving window, the indirect association between the sensitive word and the intermediate words is identified, the meaning and risk of the indirectly associated words are analyzed according to the dependency relationship between the words; the meaning and risk of the directly associated word combination and the meaning and risk of the indirectly associated words are comprehensively analyzed, and the meaning and risk of the sensitive word in the current context are output.

[0047] Optionally, the size of the moving window for analysis is first determined, the window can be a fixed size, for example, 5 words before and after, or a dynamically adjusted size, depending on the specific context and the importance of the sensitive word. In the preliminary processing, a fixed size window is selected for local text analysis, and in the subsequent steps, the size of the moving window can be dynamically adjusted according to the weight and semantic importance of the sensitive word (the calculation method is: divide the weight of the sensitive word by the standard weight, and then multiply the quotient by the fixed moving window size), for example, the more complex or important the context of the sensitive word is, the window range may be appropriately expanded to ensure that more context information is captured.

[0048] After the mobile window size is determined, the first-order association analysis begins. In this phase, the model starts to make local context selection within this window and checks the co-occurrence frequency between the sensitive word and the directly associated words (i.e., the words that are grammatically closely connected to the sensitive word). That is, the model calculates the frequency of the sensitive word co-occurring with each relevant word (such as adjectives, adverbs, or verbs, etc.). For example, if the sensitive word is "loan", the model will focus on other words in the window that are related to "loan", such as "approval", "interest rate", "credit limit", or "application", etc. By calculating the frequency of these words co-occurring with "loan" in the window, the model can assess whether these words frequently co-occur with the sensitive word, thereby determining the semantic association strength between them. Subsequently, the model will analyze the dependency relationships between the words based on their grammatical structure. The model will identify the grammatical dependency relationships between the sensitive word and other words through syntactic analysis (such as dependency syntactic analysis). For example, in "loan application", "loan" as a noun has a dependency relationship with "application", indicating that "application" is a further description of the sensitive word "loan". By analyzing these dependency relationships, the model can determine the role and meaning of each word in the grammatical layer, thereby more accurately understanding their semantic connection with the sensitive word. After identifying the co-occurrence frequency of the sensitive word and the directly associated words and their grammatical dependency relationships, the combination of these directly associated words will be subjected to semantic analysis to determine their meaning in the current context. For example, the combination of the words "loan" and "approval" may indicate a formal loan approval with low risk, while the combination of "loan" and "high interest rate" may indicate a loan with high risk. Based on the contextual meaning of the combination, the risk coefficient of the combination is obtained, and the risk coefficient is multiplied by the corresponding co-occurrence frequency to obtain the risk of the combination of directly associated words.

[0049] After the first-order analysis, the second-order association analysis phase begins. In this phase, the model will dynamically adjust the size of the window based on the similarity of the edge words. The edge words are the words located on the boundary of the window, which may affect the semantic connection and risk assessment between the sensitive word and other words in the window. The similarity is obtained by calculating the cosine similarity between the word vectors of the edge words and the sensitive word. Based on the similarity analysis, the window size will be adaptively adjusted. If the model detects that some edge words are highly important or strongly related, the window will be expanded. If the edge words are unrelated or have large semantic differences, the window will be reduced. In this way, the model can flexibly select an appropriate window range to capture more meaningful context information. Subsequently, the dependency relationship and risk analysis will be performed in the adjusted mobile window in the same way as described above to obtain the meaning and risk of the indirectly associated words.

[0050] Finally, the risk of all direct association vocabulary obtained is combined and compared with the corresponding preset risk threshold. If greater than or equal to the risk threshold, the meaning and corresponding risk of the direct association vocabulary combination are extracted. Similarly, the same operation is performed on the indirect association vocabulary, and the meaning and risk of the indirect association vocabulary greater than or equal to the corresponding risk threshold are extracted. By dividing all the meanings extracted according to the sensitive words, the meaning of each sensitive word in the current context is obtained, and the risk of the sensitive word in the current context is obtained by weighting the risk belonging to the same sensitive word. In summary, this process ensures that the sensitive words in the SMS content can be dynamically identified, and the meaning and potential risk of the sensitive words in a specific context can be accurately judged through moving window, association analysis and risk assessment.

[0051] According to the SMS type and user demand, the SMS semantic analysis information is compared in multiple dimensions to identify risk description content.

[0052] In one embodiment, first, according to the SMS type and user demand, it is determined which category it belongs to. This step is crucial because different types of SMS need to follow different compliance rules. For example, marketing SMS needs to comply with the relevant provisions of the Advertising Law, while notification SMS may need to focus on privacy protection and data security issues. Subsequently, based on the type of SMS and user demand, the SMS semantic analysis information is compared with the preset compliance rules such as industry standards, laws and regulations, and company policies. These rules can be textual descriptions or model-based rule libraries, covering multiple aspects such as the expression of SMS, word compliance, and sensitive information disclosure. In addition, the grammar and format of the SMS semantic analysis information are also checked to analyze potential risk descriptions. For example, marketing SMS may involve misleading advertising, false advertising, or unauthorized promotional content; service SMS may involve user privacy leakage or notification without user consent. Through multi-dimensional comparison, it can be determined whether the SMS content contains risk points that violate compliance requirements. During the comparison process, specific risk description content can be identified, such as non-compliant advertising language (exaggerated marketing statements, false promotional information, etc.), excessive exposure of sensitive information (leakage of personal information, unauthorized promotion, etc.), lack of explicit consent for marketing activities (advertising or promotion without user authorization), and other compliance gaps (SMS does not provide unsubscribe methods or customer service contact information in the required format, etc.). Through this process, not only can it be determined whether the SMS is compliant according to the type and user demand, but also the potential risk content in the SMS can be accurately identified, providing a basis for subsequent compliance review and optimization.

[0053] According to the SMS type and user demand, the risk description content is adjusted and optimized according to the semantic analysis information of the risk content to generate target SMS content.

[0054] In one embodiment, first, the corresponding industry compliance constraints are obtained according to the type of short message (such as marketing, notification, service, etc.), and the expression characteristics of the short message are analyzed according to the user demand, to ensure that the information transmission meets the user's expectations. Subsequently, using these industry compliance constraints and user demand characteristics, an fitness evaluation function is constructed to quantify the matching degree between the short message content and the compliance standards and user demand. Based on this evaluation function, the short message content will be adjusted to obtain the short message content that best meets the compliance requirements and user demand, thereby generating the final target short message content, which meets the relevant laws and regulations and industry standards, avoids potential legal risks, and better meets the user demand, ensuring that the user's experience and rights and interests are not harmed.

[0055] Further, the application provides that the risk description content is optimized and adjusted according to the type of the short message and the user demand, and the semantic analysis information of the risk content is used to generate the target short message content, including:

[0056] According to the type of the short message, the industry compliance constraints are obtained; according to the user demand, the user sentence expression demand characteristics are obtained; according to the industry compliance constraints and the user sentence expression demand characteristics, an fitness evaluation function is constructed, and based on the fitness evaluation function, an optimization strategy analysis space is constructed, which is used to search for the short message content that meets the industry compliance constraints and the user sentence expression demand characteristics evaluation result the largest, to generate the target short message content.

[0057] Optionally, the industry compliance constraints related to the short message are determined according to the short message type. Different industries (such as finance, medical care, government affairs, etc.) have different compliance requirements for short message content. For example, the finance industry requires clear risk warnings to users, the medical industry requires avoidance of false propaganda, and the government affairs industry requires information accuracy and transparency. Therefore, the corresponding compliance rules need to be obtained according to the industry type of the short message. According to user needs, the user's sentence expression demand characteristics are obtained, and specific demands such as coherence and advertising effect are identified. For example, advertising short messages need to be smooth and attractive, avoiding long or abrupt sentences to ensure clear and persuasive information transmission. In addition, key information such as discounts and promotions should be highlighted to clearly show the selling points (such as discounts, gifts, etc.) that attract users. Government short messages require accurate information content to avoid ambiguity and misunderstanding. These user demand characteristics can help the system understand the user's intended tone, style, and key content, and ensure that the user's expectations are met during short message generation. Subsequently, based on the industry compliance constraints and user sentence expression demand characteristics, an adaptability evaluation function is constructed. The main goal of this function is to measure the performance of the short message content in meeting industry compliance requirements and user demand expression. This is achieved by weighting the compliance score, coherence score, and advertising effect score. The compliance score is obtained by performing the aforementioned sensitive word identification on the optimized short message content, obtaining the number of sensitive words, dividing this number by the maximum allowed sensitive threshold for the industry, and then subtracting the quotient from 1. The coherence score is obtained by inputting the optimized short message content into an existing language model (such as GPT, OpenAI's GPT-3, etc.). The advertising effect score is obtained by dividing the density of marketing keywords by the recommended density threshold, and then comparing the calculated quotient with 1 to obtain the minimum value. Based on the adaptability evaluation function, an optimization strategy analysis space is constructed. This analysis space considers various possible short message content generation strategies, such as multiple expression methods (different tones, information organization structures, etc.), compliance adjustments (adjusting the wording of the short message content or adding necessary compliance statements), and effect optimization (adjusting the order of advertising information or highlighting selling points), with the goal of searching for short message content that meets industry compliance constraints and maximizes user demand expression.By searching in this space, the optimized SMS content is obtained, and the fitness evaluation function is used to evaluate the optimized SMS content, calculate the fitness value of the optimized SMS content, and compare the fitness with the preset fitness. If the fitness is not greater than the preset fitness, the optimized SMS content is discarded, and new SMS content is generated from the optimization strategy analysis space. Otherwise, the current optimized SMS content is taken as the target SMS content, which not only meets the compliance requirements of the industry, but also maximizes the satisfaction of the user's demand, and can improve the intelligent level and compliance guarantee capability of the SMS content generation in the communication service.

[0058] Further, the application provides a target SMS content generation method, which further comprises:

[0059] A multi-dimensional risk word library is established, and the multi-dimensional risk word library is updated in multiple dimensions according to the sensitive words of the SMS type and the compliance rule update frequency. The semantic coarse recognition channel and the deep semantic analysis channel are combined to build a double semantic understanding channel, and the double semantic understanding channel is integrated with the optimization strategy analysis space and the multi-dimensional risk word library to build an SMS integrated generation module. The SMS integrated generation module is used for integrated processing of semantic analysis, strategy optimization and compliance conversion of the SMS content to be processed, and the target SMS content is generated.

[0060] Optionally, a multi-dimensional risk library is pre-established, which is used to store all sensitive words and risk words related to the content of the short message. The library is updated according to the sensitive words of the short message type, the compliance rule update frequency, that is, the risk library is regularly updated according to the new sensitive words of the short message type and the new industry compliance rules. The update frequency depends on the changes of industry regulations and the addition or deletion of sensitive words. For example, some advertising words may be disabled as "misleading advertising", so it is necessary to follow up the relevant regulations in real time and update the sensitive word library. In this way, a comprehensive and dynamically updated multi-dimensional risk library is maintained, which can help subsequent compliance checking and risk assessment. Subsequently, the semantic coarse recognition channel and the deep semantic analysis channel are combined to form a double semantic understanding channel. The semantic coarse recognition channel classifies the content of the short message through a plurality of logistic regression models to identify the industry type and user demand type of the short message. The deep semantic analysis channel uses a deep neural network model to perform more detailed semantic analysis on the content of the short message to capture complex semantic structures and context-dependent relationships in the text and identify potential risks and compliance issues. Through this double semantic understanding channel, the system can analyze the compliance and risk of the short message from the industry level and the text level at the same time, ensuring that the content of the short message complies with industry standards and has no implicit illegal problems in semantics. Then, the double semantic understanding channel is integrated with the optimization strategy analysis space and the multi-dimensional risk library to form a unified short message integrated generation module. The optimization strategy analysis space is used to optimize the content of the short message in multiple dimensions according to user demand and industry compliance requirements. The multi-dimensional risk library detects sensitive words in the short message in real time to ensure that the short message does not contain illegal words. Through this short message integrated generation module, the system can automatically optimize and adjust the content of the short message after semantic analysis, risk assessment and compliance detection. Then, under the support of the short message integrated generation module, the content of the short message to be processed is subjected to semantic analysis, strategy optimization, compliance conversion and other operations, so as to generate target short message content that meets all industry compliance requirements and can accurately express user demand, ensuring its best performance in compliance and marketing effect, thereby improving the efficiency and accuracy of short message generation.

[0061] In summary, the embodiments of the present application have at least the following technical effects:

[0062] The embodiment of the application first coarsely identifies the short message content to be processed by the service provider, obtains the short message type and user demand; then inputs the short message content into a semantic large model for deep semantic analysis, and outputs short message semantic analysis information; thereafter, multi-dimensional compliance comparison is performed on the short message semantic analysis information according to the short message type and user demand, and risk description content is identified; finally, the target short message content is generated by optimizing and adjusting the risk description content according to the short message type and user demand and according to the semantic analysis information of the risk content. These technical effects jointly solve the technical problems that the generation of short message content in communication services relies on templates and cannot dynamically identify and avoid compliance risks according to user demand, and achieve the technical effects of deep semantic analysis and multi-dimensional risk comparison through a semantic large model, thereby improving the intelligent level and compliance guarantee capability of short message content generation in communication services.

[0063] Embodiment two, based on the same inventive concept as the short message content compliance generation method of the foregoing embodiments, as shown in Figure 2 The application provides a short message content compliance generation system based on a semantic large model, which comprises: a coarse identification unit 11 for coarsely identifying the short message content to be processed by the service provider, obtaining the short message type and user demand; a deep semantic analysis unit 12 for inputting the short message content into a semantic large model for deep semantic analysis, and outputting short message semantic analysis information; a multi-dimensional compliance comparison unit 13 for performing multi-dimensional compliance comparison on the short message semantic analysis information according to the short message type and user demand, and identifying risk description content; and an optimization and adjustment unit 14 for optimizing and adjusting the risk description content according to the short message type and user demand and according to the semantic analysis information of the risk content, and generating target short message content.

[0064] Further, the coarse identification unit 11 is further configured to perform the following method:

[0065] obtain the industry type and short message demand type of short message compliance review; perform data labeling according to the industry type and short message demand type, and construct training data; perform logistic regression model training using the training data, and construct a semantic coarse identification channel, wherein the logistic regression model is used to identify and output the short message type and user demand, the short message type corresponds to the industry type, and the user demand corresponds to the short message demand type.

[0066] Further, the coarse identification unit 11 is further configured to perform the following method:

[0067] The industry type at least includes the financial industry, the medical industry, and government announcements, and the short message demand type at least includes marketing, notification, and service.

[0068] Further, the deep semantic analysis unit 12 is further configured to perform the following method:

[0069] The large model framework is constructed, including an input layer, an encoder layer, a context understanding layer, a semantic analysis layer, and an output layer. The large model framework is trained and converged through collection of a training data set, and a semantic large model is obtained. A deep semantic analysis channel is built based on the semantic large model, and is used for deep semantic analysis of short message content to output short message semantic analysis information.

[0070] Further, the deep semantic analysis unit 12 is further used to execute the following method:

[0071] The input layer is used for processing text embedding. The encoder layer adopts a self-attention mechanism to capture the dependency relationship of the text, and allows the model to simultaneously focus on different parts of the text through multi-head attention. The context understanding layer is used to detect sensitive words and identify information related to sensitive words in the local context. The semantic analysis layer uses a multi-layer encoder to extract deep semantic features and further analyze the deep correlation between sensitive words and the global context. The output layer is used to integrate sensitive words, deep semantic features, and context correlation, and output short message semantic analysis information.

[0072] Further, the deep semantic analysis unit 12 is further used to execute the following method:

[0073] The short message content is preprocessed, including cleaning and removing irrelevant characters, and is input into the semantic large model. The input layer converts the preprocessed short message content into an embedding vector. The encoder layer captures the dependency relationship between each word in the short message content based on the self-attention mechanism, and assigns each word a weight representing its importance in the current context. The context understanding layer identifies sensitive words by matching the short message content with a sensitive word list, and analyzes the surrounding local context information of the detected sensitive words based on the dependency relationship output by the encoder layer to determine the meaning and risk of the sensitive words in the current context. Based on the identified sensitive words and local context relationships, the semantic analysis layer uses a multi-layer encoder structure to extract deep semantic features of the short message content and further analyze the deep correlation between sensitive words in the global context. The identified sensitive words, local context relationships, deep semantic features, and deep correlation are connected to output the short message semantic analysis information.

[0074] Further, the deep semantic analysis unit 12 is further used to execute the following method:

[0075] According to the weight of the sensitive word, the size of the moving window is configured; based on the moving window, local context selection of the sensitive word is performed, the co-occurrence frequency of the sensitive word and directly associated words is identified within the range of the size of the moving window, the grammatical relationship between the words is analyzed according to the dependency relationship between the words, the meaning and risk of the directly associated word combination are identified; the size of the window is adjusted according to the risk in the moving window and the similarity of the words at the edge of the window, based on the adjusted moving window, the indirect association between the sensitive word and the intermediate word is identified, the meaning and risk of the indirectly associated words are analyzed according to the dependency relationship between the words; the meaning and risk of the directly associated word combination and the meaning and risk of the indirectly associated words are comprehensively analyzed, and the meaning and risk of the sensitive word in the current context are output.

[0076] Further, the optimization adjustment unit 14 is further used to perform the following method:

[0077] According to the short message type, an industry compliance constraint is obtained; according to the user demand, a user sentence expression demand feature is obtained; an adaptability evaluation function is constructed according to the industry compliance constraint and the user sentence expression demand feature, and an optimization strategy analysis space is constructed based on the adaptability evaluation function, which is used to search for a short message content that satisfies the industry compliance constraint and the user sentence expression demand feature evaluation result, and generate the target short message content.

[0078] Further, the optimization adjustment unit 14 is further used to perform the following method:

[0079] A multi-dimensional risk word library is established, and the multi-dimensional risk word library is synchronously updated according to the sensitive word of the short message type and the compliance rule update frequency; the semantic coarse recognition channel and the deep semantic analysis channel are combined to construct a double semantic understanding channel, and the double semantic understanding channel, the optimization strategy analysis space and the multi-dimensional risk word library are integrated to construct a short message integrated generation module; the short message integrated generation module is used for integrated processing of semantic analysis, strategy optimization and compliance conversion on the short message content to be processed, and the target short message content is generated.

[0080] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments. The above describes a specific embodiment of the present application. The processes depicted in the drawings do not necessarily require the specific order and continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or may be advantageous.

[0081] The above only describes the preferred embodiments of the present application and does not limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0082] The specification and drawings are, of course, to be regarded in an illustrative rather than a restrictive sense. It is to be understood that any such modifications, variations, combinations or equivalents that fall within the scope of the application are intended to be embraced herein.

Claims

1. A method for short message content compliance generation of a fusion semantic large model, characterized in that, The method comprises the following steps: Coarse identification is performed on the short message content to be processed by the service provider to obtain the short message type and user demand; The short message content is input into a semantic large model for deep semantic analysis, and short message semantic analysis information is output; According to the short message type and user demand, the short message semantic analysis information is compared in multiple dimensions to identify the risk description content; According to the short message type and user demand, the risk description content is optimized and adjusted according to the semantic analysis information of the risk content to generate target short message content; The coarse identification of the short message content to be processed by the service provider obtains the short message type and user demand, which comprises the following steps: Obtain the industry type and short message demand type of short message compliance audit; According to the industry type and short message demand type, data labels are constructed to build training data; A logistic regression model is trained using the training data to build a semantic coarse identification channel, wherein the logistic regression model is used to identify and output the short message type and user demand, and the short message type corresponds to the industry type and the user demand corresponds to the short message demand type; The short message content is input into a semantic large model for deep semantic analysis, and short message semantic analysis information is output, which comprises the following steps: A large model framework is constructed, including an input layer, an encoder layer, a context understanding layer, a semantic analysis layer, and an output layer; The large model framework is trained and converged by collecting the established training data set to obtain the semantic large model, and a deep semantic analysis channel is built based on the semantic large model to perform deep semantic analysis on the short message content and output the short message semantic analysis information; According to the short message type, the industry compliance constraints are obtained; According to the user demand, the user statement expression demand characteristics are obtained; According to the industry compliance constraints and user statement expression demand characteristics, an adaptability evaluation function is constructed, and an optimization strategy analysis space is constructed based on the adaptability evaluation function, which is used to search for short message content that satisfies the industry compliance constraints and user statement expression demand characteristics evaluation results, and generate the target short message content. The industry type at least includes: financial industry, medical industry, government announcement, and the short message demand type at least includes marketing, notification, and service.

2. The method of claim 1, wherein the method further comprises: The large model framework is constructed, including: 3.The method of claim 1, wherein, The input layer is used for processing text embedding; The coding layer adopts a self-attention mechanism to capture the dependency relationship of the text, and allows the model to focus on different parts of the text at the same time through multi-head attention; The context understanding layer is used to detect sensitive words and identify information related to sensitive words in the local context; The semantic analysis layer uses a multi-layer encoder to extract deep semantic features and further analyze the deep correlation between sensitive words and the global context; The output layer is used to integrate sensitive words, deep semantic features and context correlation to output short message semantic analysis information. The short message content is input into a semantic large model for deep semantic analysis, and short message semantic analysis information is output, which comprises the following steps:

4. The method of claim 3, wherein the method further comprises: ​ The short message content is preprocessed, including cleaning and removing irrelevant characters, and inputting the semantic large model; The input layer converts the preprocessed short message content into an embedding vector; The encoding layer captures the dependency relationship between each word in the short message content based on the self-attention mechanism and assigns each word a weight representing its importance in the current context; The context understanding layer identifies sensitive words by matching the short message content with a sensitive word list and analyzes the detected sensitive words based on the dependency relationship output by the encoding layer to determine their meaning and risk in the current context; Based on the identified sensitive words and local context relationships, the semantic analysis layer uses a multi-layer encoder structure to extract deep semantic features of the short message content and further analyze the deep association relationships between sensitive words in the global context; The identified sensitive words, local context relationships, deep semantic features, and deep association relationships are connected to output the short message semantic analysis information. 5.The method of claim 4, wherein, Determining the meaning and risk of sensitive words in the current context includes: Configuring a moving window size based on the weight of the sensitive word; Based on the moving window, the local context of the sensitive word is selected, the co-occurrence frequency of the sensitive word and the directly associated words within the moving window size is identified, the grammar relationship between the words is analyzed based on the dependency relationship between the words, and the meaning and risk of the directly associated word combination are identified; Adjusting the window size based on the risk within the moving window and the similarity of the window edge words, identifying the indirect association between the sensitive word and the intermediate word based on the adjusted moving window, and analyzing the meaning and risk of the indirectly associated words based on the dependency relationship between the words; Comprehensive analysis of the meaning and risk of the directly associated word combination and the meaning and risk of the indirectly associated words outputs the meaning and risk of the sensitive word in the current context. 6.The method of claim 1, wherein, Generating target short message content also includes: Establishing a multi-dimensional risk word library and updating the multi-dimensional risk word library based on the sensitive words of the short message type and the compliance rule update frequency; Combining the semantic coarse recognition channel and the deep semantic analysis channel to build a dual semantic understanding channel, integrating the dual semantic understanding channel with the optimization strategy analysis space and the multi-dimensional risk word library to build a short message integrated generation module; The short message integrated generation module performs semantic analysis, strategy optimization, and compliance conversion on the short message content to be processed to generate the target short message content.

7. A short message content compliance generation system of a fusion semantic large model, characterized in that, The system is used to perform the short message content compliance generation method of the fused semantic large model according to any one of claims 1-6, including: Coarse recognition unit: coarsely recognizing the short message content to be processed by the service provider to obtain the short message type and user demand; Deep semantic analysis unit: inputting the short message content into the semantic large model for deep semantic analysis to output short message semantic analysis information; Multi-dimensional compliance comparison unit: comparing the short message semantic analysis information in multiple dimensions based on the short message type and user demand to identify risk description content; Optimization adjustment unit: optimizing and adjusting the risk description content according to the short message type and user demand based on the semantic analysis information of the risk content to generate target short message content.

Images