A large-scale model-driven method and system for dynamic detection and interception of marketing content compliance

Abstract

This invention discloses a large-model-driven method and system for dynamic detection and interception of marketing content compliance. Addressing the core issue of potential negative emotional impact and contextual misunderstanding risks in marketing promotional texts, it proposes a complete solution. The method extracts text features through a semantic processing model, combines this with a pre-trained sentiment analysis model for multi-dimensional sentiment classification, identifies negative sentiment scores, and constructs a dependency graph when thresholds are exceeded. This graph is then matched with a contextual knowledge base to uncover hidden risk tags, generating a comprehensive compliance scoring matrix and calculating the overall risk score. For high-risk content, an adaptive rule adjustment algorithm is used, integrating sentiment vectors and risk tags for iterative correction, ultimately outputting a compliant version of the text. This invention combines sentiment analysis, contextual understanding, and dynamic optimization to achieve real-time detection and accurate interception of marketing content, effectively reducing compliance risks and improving content quality and brand image.

Description

Technical Field

[0001] This invention relates to the field of dynamic detection and interception technology for marketing content compliance, and more specifically, to a large-model-driven method and system for dynamic detection and interception of marketing content compliance. Background Technology

[0002] In today's digital marketing field, content compliance testing and management has become a crucial aspect of safeguarding corporate image and legal security. This area directly relates to brand reputation and user trust, and is a critical part of business operations that cannot be ignored. With the diversification of marketing content and the increasing complexity of dissemination channels, ensuring the compliance of content before and after its release has become a major issue that the industry urgently needs to address.

[0003] However, existing methods often struggle to adapt to the variability and real-time nature of marketing content when addressing this challenge. Many solutions lack sufficient flexibility and adaptability when dealing with content across different platforms and contexts, resulting in inaccurate detection results or even misjudgments or omissions. This limitation makes it difficult for companies to effectively avoid potential compliance risks in a rapidly changing market environment. A deeper technical challenge lies in recognizing the emotional tone and semantic implicit information of marketing content, which becomes a core obstacle. Marketing copy often contains complex emotional nuances, such as irony and exaggeration. These emotional tones may be misunderstood in different cultures or contexts, thus triggering compliance issues.

[0004] The identification of sentiment further complicates the mining of semantic implicit information. While content may appear compliant on the surface, its hidden meanings may touch upon sensitive topics or contain inappropriate innuendos. This progressive connection from sentiment to implicit information makes compliance detection technically exceptionally complex. For example, on a social media platform, an advertisement might use humor to mention a social issue, seemingly lighthearted and fun, but in reality, it may provoke user dissatisfaction or regulatory attention due to its implicit controversial viewpoints. Therefore, accurately capturing the sentiment in content and deeply analyzing its implicit meaning in a dynamically changing marketing environment has become a key issue in ensuring compliance. Solving this problem is not only about technological breakthroughs but also directly affects the compliant operation and brand safety of enterprises in complex markets. Summary of the Invention

[0005] The purpose of this invention is to provide a large-model-driven method and system for dynamic detection and interception of marketing content compliance, in order to solve the problem mentioned in the background art of how to accurately capture the emotional tendencies in content and deeply analyze its implicit meaning in a dynamically changing marketing environment.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A large-scale model-driven method for dynamic detection and interception of marketing content compliance includes: The marketing promotion text data to be detected is obtained from the marketing content database, and the semantic processing model is used to extract features from the marketing promotion text data to obtain an initial semantic representation vector; Based on the initial semantic representation vector, the pre-trained sentiment analysis model is input for multi-dimensional sentiment classification to obtain sentiment vectors, which include positive sentiment score distribution, negative sentiment score distribution and neutral sentiment score distribution. If the negative sentiment score in the sentiment vector exceeds a preset threshold, then the sentence structure in the marketing promotion text data is mined from the initial semantic representation vector to obtain a dependency graph of potential implicit meanings. By matching and comparing the dependency graph with the pre-set contextual knowledge base, it is determined whether there is an implicit meaning of contextual misunderstanding, and the implicit risk label sequence is obtained. Based on the implicit risk label sequence, a comprehensive compliance scoring matrix is ​​generated by combining the sentiment vector. The overall compliance risk score is obtained by calculating the risk weight of each label through weighted calculation. If the overall compliance risk score is higher than the preset warning threshold, a real-time feedback mechanism will be used to mark the marketing and promotion text data as high-risk types and determine the list of high-risk content segments. By applying an adaptive rule adjustment algorithm to the list of content fragments and iteratively correcting it by integrating sentiment vectors and implicit risk label sequences, the final compliant version of the marketing promotion text content is obtained.

[0007] A preferred technical solution of this application: The semantic processing model extracts features from marketing promotion text data to obtain an initial semantic representation vector as follows: The marketing promotion text data is segmented into words to obtain word sequences. , To determine the effective word segmentation of marketing promotion text data, a pre-trained language encoder is used to obtain the basic semantic vectors of each word. ; in, For the first The basic semantic vector of each word For language encoders, The first data in the marketing promotion text Each word, ; Construct a weighted coefficient that integrates location weight, marketing attribute weight, and length normalization factor. : ; in, , and The preset hyperparameters satisfy the following conditions: , For text position weights, For marketing attribute weight, This is the text length normalization factor; We obtain the initial semantic representation vector by performing a weighted summation and modulus normalization on the basic semantic vectors. : ; in, This indicates the segmentation of the marketing promotion text data from the first word to the second word. Each word segment is summed, and the basic semantic vector of each word is calculated sequentially. The combined weighting coefficient of the word segmentation Add the product results together. Represents the basic semantic vector The L2 module length.

[0008] A preferred technical solution of this application: The initial semantic representation vector is input into a pre-trained sentiment analysis model for multi-dimensional sentiment classification processing to obtain the sentiment vector, as follows: Obtain the number of multi-dimensional features contained in the initial semantic representation vector; For each dimension of the multi-dimensional features, extract the corresponding negative sentiment probability; The negative sentiment probabilities of all extracted dimensions are aggregated and calculated to obtain the negative sentiment score; Extract the positive sentiment probability and neutral sentiment probability from the multi-dimensional feature count; The positive sentiment score distribution is obtained by aggregating and calculating the probabilities of positive sentiment. The distribution of neutral sentiment scores is obtained by aggregating and calculating the probabilities of neutral sentiment. A sentiment vector is constructed by combining the distributions of negative sentiment scores, positive sentiment scores, and neutral sentiment scores. Compare the negative sentiment score with the preset negative sentiment score threshold; If the negative sentiment score is greater than the negative sentiment score threshold, sentence structure mining processing is triggered on the initial semantic representation vector; The negative sentiment score is calculated by aggregation using the following formula: ; in, This indicates the number of multi-dimensional features in marketing and promotional text data. Indicates the first Dimensioning the probability of negative emotions This indicates the score for negative emotions.

[0009] A preferred technical solution of this application: the step of mining sentence structure in marketing promotion text data based on initial semantic representation vectors to obtain a dependency graph of potential implicit meanings includes: Extract the lexical feature set and the syntactic feature set from the initial semantic representation vector; Based on the lexical feature set, determine the sequence of entity nodes in the marketing promotion text data; Based on the set of syntactic features, determine the dependency relationships between entity nodes in the entity node sequence. Combine entity node sequences with dependency relationships to construct an initial sentence structure topology graph; Extract the hidden semantic nodes from the initial sentence structure topology graph; Based on the implicit semantic nodes, a dependency graph vector with potential implicit meanings is generated; Obtain the vector of the preset context knowledge base; Calculate the matching similarity between the dependency graph vector of potential implicit meaning and the vector of the pre-defined context knowledge base; Compare the matching similarity with the preset similarity threshold; If the matching similarity is greater than the similarity threshold, extract high-risk semantic features from the dependency graph vector with potential implicit meanings; Based on the high-risk semantic features, the corresponding risk tags are retrieved from the preset context knowledge base vector; The retrieved risk tags are sorted according to the order of the entity node sequence to generate a sequence of hidden risk tags; Combine the dependency graph of implicit risk label sequences and potential implicit meanings to complete the construction of the dependency graph; Extract contextual features from marketing and promotional text data; Based on contextual association features, the dependency relationship connections in the initial sentence structure topology graph are weighted to obtain weighted dependency relationship connections. Based on weighted dependency connections, core dependency paths with weight values ​​greater than a preset connection threshold are selected. The core dependency path is defined as the input feature of the dependency graph vector with potential implicit meaning; The input features are dimensionality reduced to obtain standardized dependency graph vectors with latent meanings. When calculating the matching similarity between the dependency graph vector with potential implicit meaning and the vector of the preset context knowledge base, the cosine similarity algorithm is used to calculate the specific numerical value of the matching similarity.

[0010] A preferred technical solution of this application: the matching similarity between the dependency graph of the implicit meaning and the preset context knowledge base is calculated using the cosine similarity algorithm. ; in, This indicates the similarity between the dependency graph and the contextual knowledge base. Dependency graph vectors representing potential implicit meanings. Representing vectors transpose, Represents a predefined context knowledge base vector, if If the threshold is exceeded, a hidden risk label sequence is generated.

[0011] A preferred technical solution of this application is as follows: A comprehensive compliance scoring matrix is ​​generated based on the implicit risk label sequence and combined with sentiment vectors. The overall compliance risk score is obtained by weighting the risk weights of each label. The weighting calculation formula is as follows: ; in, This indicates the overall compliance risk score. The weights representing the negative scores in the sentiment vector. Indicates the first Risk weights for implicit risk labels Indicates the first A quantitative score for each implicit risk label. This represents the total number of hidden risk labels.

[0012] A preferred technical solution of this application: the step of using a real-time feedback mechanism to mark marketing promotion text data as high-risk type and determining a list of content fragments of high-risk type includes: Obtain the specific value of the preset warning threshold; Compare the overall compliance risk score with the specific values ​​of the preset warning threshold; If the overall compliance risk score exceeds the preset warning threshold, a real-time feedback mechanism will be triggered. High-risk type markers are written into the metadata of marketing and promotion text data through a real-time feedback mechanism; Based on the high-risk type markers, extract the text fragments in the marketing promotion text data that correspond to the implicit risk label sequences; Record the start and end positions of text segments within the marketing and promotional text data; Extract the corresponding text content based on the start and end positions; Summarize the extracted text content and generate a list of high-risk content fragments.

[0013] A preferred technical solution of this application: the marketing promotion text content is obtained by applying an adaptive rule adjustment algorithm to the content fragment list, fusing sentiment vectors and implicit risk label sequences for iterative correction, and including: Obtain initial marketing and promotional text data; We perform preliminary extraction of sentiment vectors and implicit risk information contained in the initial marketing and promotion text data to obtain a basic set of sentiment tendencies and risk tags; An adaptive rule adjustment algorithm is used to analyze the extracted basic sentiment and risk label sets. Detect whether the basic sentiment tendency deviates from the preset sentiment tendency threshold; Detect whether the number of tags in the risk tag set exceeds the preset safety range; If the basic sentiment tendency is detected to deviate from the preset sentiment tendency threshold, or if the number of tags in the risk tag set exceeds the preset safety range, the adjustment mechanism will be triggered. Based on the adjustment mechanism, identify the content segments to be corrected from the list of content segments; A fusion-iteration method is used to match sentiment vectors with implicit risk label sequences layer by layer; Extract the deviation from the layer-by-layer matching results; For the deviations, the text expression of the content segments to be corrected is gradually revised to obtain the preliminary optimized text content; Based on the preliminary optimized text content and in conjunction with the standard requirements for compliant text, extract the compliance features from the preliminary optimized text content; Compare compliance features with the standard requirements of compliance documents; Determine whether there are any non-standard expressions in the initially optimized text content; If there are expressions that do not conform to the specifications, extract the specific location information of the non-conformities; By using text optimization techniques, local adjustments are made to the initially optimized text content based on specific location information; Based on the results of the local adjustments, a second-processed text version is generated. Obtain the text version after secondary processing; A deep scan was performed on the remaining risk control points in the second-processed text version; Obtain depth scan results; Determine if the depth scan results indicate any potential risks; If the depth scan results show potential problems, the text version after secondary processing is further optimized by an adaptive rule adjustment algorithm to obtain a text structure that meets the requirements. Perform a comprehensive validation of the text structure that meets the requirements; Based on the goals and needs of marketing and promotion, extract the emotional expression features and content presentation features from the text structure that meets the requirements; Determine whether the characteristics of emotional expression and content presentation meet the expected emotional expression and content presentation; If the expected emotional expression and content presentation are met, the output text structure that meets the requirements will be the final compliant version of the marketing promotion text content.

[0014] A preferred technical solution of this application: The calculation process of the adaptive rule adjustment algorithm is as follows: No. The overall compliance score for each iteration is obtained by weighted fusion of the sentiment score and the risk label normalized score: ; in, For the first Overall compliance score, For the first In terms of emotional weight, Score the sentiment of the text. Normalize the score for the text risk label; The sentiment weights are dynamically updated based on sentiment bias and risk bias. ; in, The iteration step size, For emotional threshold, This is the risk safety threshold; The iteration stops when the overall compliance score converges and both the sentiment bias and risk label meet the requirements. ,and ,in, This is the convergence accuracy threshold; The correction strength of each text segment is calculated based on the sentiment weights after iterative convergence, guiding local text rewriting: ; in, For the first The strength of correction for each content segment, The sentiment weights after iterative convergence. For the first The emotional score of each content segment. For the first The risk score for each content segment.

[0015] This invention also provides a large-model-driven dynamic detection and interception system for marketing content compliance, the system comprising: The text acquisition module is used to retrieve marketing promotion text data to be detected from the marketing content database; The semantic extraction module, equipped with a semantic processing model, is used to extract features from marketing and promotional text data and output an initial semantic representation vector. The sentiment analysis module is equipped with a pre-trained sentiment analysis model. It receives the initial semantic representation vector, performs multi-dimensional sentiment classification and aggregation calculation, and outputs a sentiment vector containing positive sentiment score distribution, negative sentiment score distribution and neutral sentiment score distribution. It also compares the negative sentiment score with a preset threshold. The semantic mining module is used to mine the sentence structure in the marketing and promotion text data based on the initial semantic representation vector when the negative sentiment score in the sentiment vector exceeds a preset threshold. It constructs a dependency graph of potential implicit meanings, generates dependency graph vectors and matches them with preset context knowledge base vectors, and outputs a dependency graph of potential implicit meanings and a sequence of implicit risk labels. The compliance scoring module is used to generate a comprehensive compliance scoring matrix based on the implicit risk label sequence and the sentiment vector. It calculates the risk weight of each label by weighting and outputs the overall compliance risk score. The real-time feedback module is used to trigger the real-time feedback mechanism when the overall compliance risk score is higher than the preset warning threshold. It marks the marketing and promotion text data as high-risk, extracts the corresponding text fragments, and generates a list of high-risk content fragments. The iterative correction module, equipped with an adaptive rule adjustment algorithm, applies the adaptive rule adjustment algorithm to the list of content fragments, integrates sentiment vectors and implicit risk label sequences for iterative correction, and outputs the final compliant version of the marketing promotion text content; The context knowledge base module is used to store preset context knowledge base vectors, providing data support for matching similarity calculation in the semantic mining module; The parameter configuration module is used to preset and store all preset hyperparameters, thresholds, weight coefficients and iteration-related parameters in the method, providing parameter support for the operation of each module.

[0016] Compared with the prior art, the technical solution provided by this invention has the following advantages: This invention addresses the core issue of potential negative emotional impact and contextual misunderstanding risks in marketing and promotional texts by proposing a complete solution. It extracts text features through a semantic processing model, combines this with a pre-trained sentiment analysis model for multi-dimensional sentiment classification, identifies negative sentiment scores, and constructs a dependency graph when thresholds are exceeded. This graph is then matched with a contextual knowledge base to uncover hidden risk tags, generating a comprehensive compliance scoring matrix and calculating the overall risk score. For high-risk content, an adaptive rule adjustment algorithm is employed, iteratively correcting the content by integrating sentiment vectors and risk tags, ultimately outputting a compliant version of the text. This invention combines sentiment analysis, contextual understanding, and dynamic optimization to achieve real-time detection and precise interception of marketing content, effectively reducing compliance risks and improving content quality and brand image. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] 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. The present invention will be further described below with reference to the embodiments.

[0019] Please see Figure 1 The embodiments of this application provide a large model-driven method for dynamic detection and interception of marketing content compliance, including: The marketing promotion text data to be detected is obtained from the marketing content database, and the semantic processing model is used to extract features from the marketing promotion text data to obtain an initial semantic representation vector; In this embodiment, the semantic processing model extracts features from the marketing promotion text data to obtain the initial semantic representation vector as follows: The marketing promotion text data is segmented into words to obtain word sequences. , To determine the effective word segmentation of marketing promotion text data, a pre-trained language encoder is used to obtain the basic semantic vectors of each word. ; in, For the first The basic semantic vector of each word For language encoders, a pre-trained language encoder based on the Transformer architecture. The first data in the marketing promotion text Each word, ; Construct a weighted coefficient that integrates location weight, marketing attribute weight, and length normalization factor. : ; in, , and The preset hyperparameters satisfy the following conditions: , For text position weights, For marketing attribute weight, This is the text length normalization factor, which is calculated based on the total text length. Adaptively generated normalization coefficients are used to eliminate the impact of text length differences; We obtain the initial semantic representation vector by performing a weighted summation and modulus normalization on the basic semantic vectors. : ; in, This indicates the segmentation of the marketing promotion text data from the first word to the second word. Each word segment is summed, and the basic semantic vector of each word is calculated sequentially. The combined weighting coefficient of the word segmentation Add the product results together. Represents the basic semantic vector The L2 modulus of the denominator is the square root of the sum of the squares of the weighted semantic lengths of all word segments, which is the overall L2 modulus of the weighted semantic vector. It is used to normalize the numerator so that the semantic representation of texts of different lengths is unified to the same numerical range.

[0020] Based on the initial semantic representation vector, the pre-trained sentiment analysis model is input for multi-dimensional sentiment classification to obtain sentiment vectors, which include positive sentiment score distribution, negative sentiment score distribution and neutral sentiment score distribution. In this embodiment, the initial semantic representation vector is input into a pre-trained sentiment analysis model for multi-dimensional sentiment classification processing to obtain the sentiment vector. The process is as follows: Obtain the number of multi-dimensional features contained in the initial semantic representation vector; For each dimension of the multi-dimensional features, extract the corresponding negative sentiment probability; The negative sentiment probabilities of all extracted dimensions are aggregated and calculated to obtain the negative sentiment score; Extract the positive sentiment probability and neutral sentiment probability from the multi-dimensional feature count; The positive sentiment score distribution is obtained by aggregating and calculating the probabilities of positive sentiment. The distribution of neutral sentiment scores is obtained by aggregating and calculating the probabilities of neutral sentiment. A sentiment vector is constructed by combining the distributions of negative sentiment scores, positive sentiment scores, and neutral sentiment scores. Compare the negative sentiment score with the preset negative sentiment score threshold; If the negative sentiment score is greater than the negative sentiment score threshold, sentence structure mining processing is triggered on the initial semantic representation vector.

[0021] The negative sentiment score is calculated using the following formula: ; in, This indicates the number of multi-dimensional features in marketing and promotional text data. Indicates the first Dimensioning the probability of negative emotions This indicates the score for negative emotions.

[0022] If the negative sentiment score in the sentiment vector exceeds a preset threshold, then the sentence structure in the marketing promotion text data is mined from the initial semantic representation vector to obtain a dependency graph of potential implicit meanings. In this embodiment, sentence structure is mined from the initial semantic representation vector in the marketing promotion text data to obtain a dependency graph with potential implicit meanings, including: Extract the lexical feature set and the syntactic feature set from the initial semantic representation vector; Based on the lexical feature set, determine the sequence of entity nodes in the marketing promotion text data; Based on the set of syntactic features, determine the dependency relationships between entity nodes in the entity node sequence. Combine entity node sequences with dependency relationships to construct an initial sentence structure topology graph; Extract the hidden semantic nodes from the initial sentence structure topology graph; Based on the implicit semantic nodes, a dependency graph vector with potential implicit meanings is generated; Obtain the vector of the preset context knowledge base; Calculate the matching similarity between the dependency graph vector of potential implicit meaning and the vector of the pre-defined context knowledge base; Compare the matching similarity with the preset similarity threshold; If the matching similarity is greater than the similarity threshold, extract high-risk semantic features from the dependency graph vector with potential implicit meanings; Based on the high-risk semantic features, the corresponding risk tags are retrieved from the preset context knowledge base vector; The retrieved risk tags are sorted according to the order of the entity node sequence to generate a sequence of hidden risk tags; Combine the dependency graph of implicit risk label sequences and potential implicit meanings to complete the construction of the dependency graph; Extract contextual features from marketing and promotional text data; Among them, the contextual association feature is the strength of the semantic association between words in the text; Based on contextual association features, the dependency relationship connections in the initial sentence structure topology graph are weighted to obtain weighted dependency relationship connections. Based on weighted dependency connections, core dependency paths with weight values ​​greater than a preset connection threshold are selected. Among them, the core dependency path is the key syntactic path that represents the main semantic logic of the text; The core dependency path is defined as the input feature of the dependency graph vector with potential implicit meaning; The input features are dimensionality reduced to obtain standardized dependency graph vectors with latent meanings. The dimensionality reduction process is implemented using PCA or global average pooling. When calculating the matching similarity between the dependency graph vector with potential implicit meaning and the vector of the preset context knowledge base, the cosine similarity algorithm is used to calculate the specific numerical value of the matching similarity.

[0023] The matching similarity between the dependency graph with implicit meaning and the pre-set contextual knowledge base is calculated using the cosine similarity algorithm. ; in, This represents the similarity between the dependency graph and the contextual knowledge base, with values ​​ranging from [0,1]. Higher values ​​indicate greater semantic similarity. Dependency graph vectors representing potential implicit meanings. Representing vectors transpose, Represents a predefined context knowledge base vector, if If the threshold is exceeded, a hidden risk label sequence is generated.

[0024] By matching and comparing the dependency graph with the pre-set contextual knowledge base, it is determined whether there is an implicit meaning of contextual misunderstanding, and the implicit risk label sequence is obtained. Based on the implicit risk label sequence, a comprehensive compliance scoring matrix is ​​generated by combining the sentiment vector. The overall compliance risk score is obtained by calculating the risk weight of each label through weighted calculation. In this embodiment, a comprehensive compliance scoring matrix is ​​generated based on the implicit risk label sequence and the sentiment vector. The overall compliance risk score is obtained by calculating the risk weight of each label using a weighted average. The weighted average calculation formula is as follows: ; in, This indicates the overall compliance risk score. The weights representing the negative scores in the sentiment vector. Indicates the first Risk weights for implicit risk labels Indicates the first A quantitative score for each implicit risk label. This represents the total number of hidden risk labels.

[0025] If the overall compliance risk score is higher than the preset warning threshold, a real-time feedback mechanism will be used to mark the marketing and promotion text data as high-risk types and determine the list of high-risk content segments. In this embodiment, a real-time feedback mechanism is used to mark marketing promotion text data as high-risk types, and a list of high-risk content fragments is determined, including: Obtain the specific value of the preset warning threshold; Compare the overall compliance risk score with the specific values ​​of the preset warning threshold; If the overall compliance risk score exceeds the preset warning threshold, a real-time feedback mechanism will be triggered. High-risk type markers are written into the metadata of marketing and promotion text data through a real-time feedback mechanism; The real-time feedback mechanism refers to the risk response and location mechanism that the system immediately triggers and executes without delay when the overall compliance risk score is higher than the preset warning threshold. Specifically, it includes: marking high-risk types in real time, locating text fragments corresponding to the hidden risk label sequence in real time, recording the fragment position and extracting the text content in real time, and generating a list of content fragments that need further optimization in real time, so as to realize the immediate identification, dynamic interception and rapid correction of compliance risks in marketing content. Metadata consists of additional attribute information used to identify the risk status of text; Based on the high-risk type markers, extract the text fragments in the marketing promotion text data that correspond to the implicit risk label sequences; Record the start and end positions of text segments within the marketing and promotional text data; The start and end positions are the character offset positions in the marketing promotion text data; Extract the corresponding text content based on the start and end positions; Summarize the extracted text content and generate a list of high-risk content fragments.

[0026] By applying an adaptive rule adjustment algorithm to the list of content fragments and iteratively correcting it by integrating sentiment vectors and implicit risk label sequences, the final compliant version of the marketing promotion text content is obtained. In this embodiment, by applying an adaptive rule adjustment algorithm to the list of content fragments and iteratively correcting it by fusing sentiment vectors and implicit risk label sequences, the final compliant version of the marketing promotion text content is obtained, including: Obtain initial marketing and promotional text data; We perform preliminary extraction of sentiment vectors and implicit risk information contained in the initial marketing and promotion text data to obtain a basic set of sentiment tendencies and risk tags; An adaptive rule adjustment algorithm is used to analyze the extracted basic sentiment and risk label sets. Detect whether the basic sentiment tendency deviates from the preset sentiment tendency threshold; Detect whether the number of tags in the risk tag set exceeds the preset safety range; If the basic sentiment tendency is detected to deviate from the preset sentiment tendency threshold, or if the number of tags in the risk tag set exceeds the preset safety range, the adjustment mechanism will be triggered. The adjustment mechanism is a rule that automatically selects the segment to be corrected based on emotional bias and risk level; Based on the adjustment mechanism, identify the content segments to be corrected from the list of content segments; A fusion-iteration method is used to match sentiment vectors with implicit risk label sequences layer by layer; Extract the deviation from the layer-by-layer matching results; For the deviations, the text expression of the content segments to be corrected is gradually revised to obtain the preliminary optimized text content; Based on the preliminary optimized text content and in conjunction with the standard requirements for compliant text, extract the compliance features from the preliminary optimized text content; Compare compliance features with the standard requirements of compliance documents; Determine whether there are any non-standard expressions in the initially optimized text content; If there are expressions that do not conform to the specifications, extract the specific location information of the non-conformities; By using text optimization techniques, local adjustments are made to the initially optimized text content based on specific location information; Text optimization techniques include sentence replacement, weakening of expression, and rewriting of risky words; Based on the results of the local adjustments, a second-processed text version is generated. Obtain the text version after secondary processing; A deep scan was performed on the remaining risk control points in the second-processed text version; Among them, deep scanning involves performing secondary matching of risk features on each segment of the text; Obtain depth scan results; Determine if the depth scan results indicate any potential risks; If the depth scan results show potential problems, the text version after secondary processing is further optimized by an adaptive rule adjustment algorithm to obtain a text structure that meets the requirements. Perform a comprehensive validation of the text structure that meets the requirements; Among them, the comprehensive verification is to verify whether both emotional compliance and risk controllability meet the requirements; Based on the goals and needs of marketing and promotion, extract the emotional expression features and content presentation features from the text structure that meets the requirements; Determine whether the characteristics of emotional expression and content presentation meet the expected emotional expression and content presentation; If the expected emotional expression and content presentation are met, the output text structure that meets the requirements will be the final compliant version of the marketing promotion text content.

[0027] The calculation process of the adaptive rule adjustment algorithm is as follows: No. The overall compliance score for each iteration is obtained by weighted fusion of the sentiment score and the risk label normalized score: ; in, For the first Overall compliance score, For the first In terms of emotional weight, Score the sentiment of the text. The normalized score for the text risk label is the result of normalizing the risk score to [0,1]. The sentiment weights are dynamically updated based on sentiment bias and risk bias. ; in, The iteration step size, For emotional threshold, This is the risk safety threshold; The iteration stops when the overall compliance score converges and both the sentiment bias and risk label meet the requirements. ,and ,in, This is the convergence accuracy threshold; The correction strength of each text segment is calculated based on the sentiment weights after iterative convergence, guiding local text rewriting: ; in, For the first The strength of correction for each content segment, The sentiment weights after iterative convergence. For the first The emotional score of each content segment. For the first The risk score of each content segment To normalize the denominator, the correction strength is... It falls within the interval [0,1].

[0028] By applying an adaptive rule adjustment algorithm to the list of content fragments and iteratively correcting it by integrating sentiment vectors and implicit risk label sequences, a compliant version of marketing promotion text content is finally generated. The specific implementation method is as follows: First, assuming the input list of marketing text fragments contains 10 short sentences, each between 20 and 50 characters in length, the system uses an adaptive rule adjustment algorithm to initially filter these fragments. The rule threshold is set to a sentiment score greater than 0.6, with a maximum score of 1.0. Natural language processing technology is used to calculate the sentiment score of each fragment. For example, if fragment 1 scores 0.75 and fragment 2 scores 0.45, then fragment 2 is eliminated, and the remaining fragments proceed to the next stage. Next, for the retained segments, the system extracts sentiment vectors. Assuming that the sentiment vector for each segment is a 5-dimensional array, such as the vector for segment 1 being [0.8, 0.3, 0.1, 0.5, 0.2], representing the distribution of positive, negative, neutral, surprised, and angry emotions, the system also combines the implicit risk label sequence. Assuming that the risk label for segment 1 is "low risk" with a risk value of 0.1, and the risk value ranges from 0 to 1, while the risk value for segment 3 is 0.7, it is marked as "high risk". The system adjusts the sentiment vectors according to the risk value, reducing the positive sentiment weight of high-risk segments, for example, reducing the positive sentiment value of segment 3 from 0.7 to 0.4. Subsequently, the system entered the iterative correction phase, adjusting the weighted fusion ratio of sentiment vector and risk value in increments of 0.01. The initial ratio was 0.6:0.4, and after three rounds of iteration, the optimal ratio was adjusted to 0.55:0.45, ensuring that the overall sentiment score of the text remained stable above 0.65 while the risk value was controlled below 0.3. Finally, compliant text content was generated. For example, the sentences in fragments 1 and 5 were reorganized into "Our products have won trust with their superior quality. Welcome to experience them," ensuring that the sentiment was positive and there was no compliance risk. The entire process is completed automatically through algorithms. High-risk content is intercepted based on sentiment analysis and risk assessment. If there is no similar content in the original text, the system can deduce similar sentiment distribution and risk tags based on historical marketing data to ensure that the generated text meets business objectives, such as increasing the user click-through rate to more than 5%. By analyzing the positive correlation between historical click-through rate and sentiment score, the correlation coefficient of the positive correlation is 0.82, further optimizing the text expression.

[0029] This invention also provides a large-model-driven dynamic detection and interception system for marketing content compliance, the system comprising: The text acquisition module is used to retrieve marketing promotion text data to be detected from the marketing content database; The semantic extraction module, equipped with a semantic processing model, is used to extract features from marketing and promotional text data and output an initial semantic representation vector. The sentiment analysis module is equipped with a pre-trained sentiment analysis model. It receives the initial semantic representation vector, performs multi-dimensional sentiment classification and aggregation calculation, and outputs a sentiment vector containing positive sentiment score distribution, negative sentiment score distribution and neutral sentiment score distribution. It also compares the negative sentiment score with a preset threshold. The semantic mining module is used to mine the sentence structure in the marketing and promotion text data based on the initial semantic representation vector when the negative sentiment score in the sentiment vector exceeds a preset threshold. It constructs a dependency graph of potential implicit meanings, generates dependency graph vectors and matches them with preset context knowledge base vectors, and outputs a dependency graph of potential implicit meanings and a sequence of implicit risk labels. The compliance scoring module is used to generate a comprehensive compliance scoring matrix based on the implicit risk label sequence and the sentiment vector. It calculates the risk weight of each label by weighting and outputs the overall compliance risk score. The real-time feedback module is used to trigger the real-time feedback mechanism when the overall compliance risk score is higher than the preset warning threshold. It marks the marketing and promotion text data as high-risk, extracts the corresponding text fragments, and generates a list of high-risk content fragments. The iterative correction module, equipped with an adaptive rule adjustment algorithm, applies the adaptive rule adjustment algorithm to the list of content fragments, integrates sentiment vectors and implicit risk label sequences for iterative correction, and outputs the final compliant version of the marketing promotion text content; The context knowledge base module is used to store preset context knowledge base vectors, providing data support for matching similarity calculation in the semantic mining module; The parameter configuration module is used to preset and store all preset hyperparameters and thresholds in the method, including negative sentiment score threshold, similarity threshold, warning threshold, weight coefficients and iteration-related parameters, providing parameter support for the operation of each module.

[0030] Based on the above methods and systems, text features are extracted through the semantic processing model in the embodiments, and multi-dimensional sentiment classification is performed by combining a pre-trained sentiment analysis model to identify negative sentiment scores. When the threshold is exceeded, a dependency graph is constructed and matched with the contextual knowledge base to mine implicit risk tags. Then, a comprehensive compliance scoring matrix is ​​generated, and the overall risk score is calculated. For high-risk content, an adaptive rule adjustment algorithm is adopted to iteratively correct by integrating sentiment vectors and risk tags. Finally, a compliant version of the text is output, which effectively solves the core problem of potential negative emotional impact and contextual misunderstanding risk in marketing and promotional texts. This invention combines sentiment analysis, contextual understanding, and dynamic optimization to achieve real-time detection and accurate interception of marketing content, effectively reducing compliance risks and improving content quality and brand image.

[0031] The present invention and its embodiments have been described above illustratively. This description is not restrictive, and the figures shown are only one embodiment of the present invention; the actual structure is not limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

[0032] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only independent technical solutions. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A large-model-driven method for dynamic detection and interception of marketing content compliance, characterized in that, include: The marketing promotion text data to be detected is obtained from the marketing content database, and the semantic processing model is used to extract features from the marketing promotion text data to obtain an initial semantic representation vector; Based on the initial semantic representation vector, the pre-trained sentiment analysis model is input for multi-dimensional sentiment classification to obtain sentiment vectors, which include positive sentiment score distribution, negative sentiment score distribution and neutral sentiment score distribution. If the negative sentiment score in the sentiment vector exceeds a preset threshold, then the sentence structure in the marketing promotion text data is mined from the initial semantic representation vector to obtain a dependency graph of potential implicit meanings. By matching and comparing the dependency graph with the pre-set contextual knowledge base, it is determined whether there is an implicit meaning of contextual misunderstanding, and the implicit risk label sequence is obtained. Based on the implicit risk label sequence, a comprehensive compliance scoring matrix is ​​generated by combining the sentiment vector. The overall compliance risk score is obtained by calculating the risk weight of each label through weighted calculation. If the overall compliance risk score is higher than the preset warning threshold, a real-time feedback mechanism will be used to mark the marketing and promotion text data as high-risk types and determine the list of high-risk content segments. By applying an adaptive rule adjustment algorithm to the list of content fragments and iteratively correcting it by integrating sentiment vectors and implicit risk label sequences, the final compliant version of the marketing promotion text content is obtained.

2. The method for dynamic detection and interception of marketing content compliance driven by a large model according to claim 1, characterized in that, The semantic processing model extracts features from marketing promotional text data to obtain an initial semantic representation vector as follows: The marketing promotion text data is segmented into words to obtain word sequences. , To determine the effective word segmentation of marketing promotion text data, a pre-trained language encoder is used to obtain the basic semantic vectors of each word. ； in, For the first The basic semantic vector of each word For language encoders, The first data in the marketing promotion text Each word, ; Construct a weighted coefficient that integrates location weight, marketing attribute weight, and length normalization factor. : ； in, , and The preset hyperparameters satisfy the following conditions: , For text position weights, For marketing attribute weight, This is the text length normalization factor; We obtain the initial semantic representation vector by performing a weighted summation and modulus normalization on the basic semantic vectors. : ； in, This indicates the segmentation of the marketing promotion text data from the first word to the second word. Each word segment is summed, and the basic semantic vector of each word is calculated sequentially. The combined weighting coefficient of the word segmentation Add the product results together. Represents the basic semantic vector The L2 module length.

3. The method for dynamic detection and interception of marketing content compliance driven by a large model according to claim 1, characterized in that, The initial semantic representation vector is input into a pre-trained sentiment analysis model for multi-dimensional sentiment classification, and the process of obtaining the sentiment vector is as follows: Obtain the number of multi-dimensional features contained in the initial semantic representation vector; For each dimension of the multi-dimensional features, extract the corresponding negative sentiment probability; The negative sentiment probabilities of all extracted dimensions are aggregated and calculated to obtain the negative sentiment score; Extract the positive sentiment probability and neutral sentiment probability from the multi-dimensional feature count; The positive sentiment score distribution is obtained by aggregating and calculating the probabilities of positive sentiment. The distribution of neutral sentiment scores is obtained by aggregating and calculating the probabilities of neutral sentiment. A sentiment vector is constructed by combining the distributions of negative sentiment scores, positive sentiment scores, and neutral sentiment scores. Compare the negative sentiment score with the preset negative sentiment score threshold; If the negative sentiment score is greater than the negative sentiment score threshold, sentence structure mining processing is triggered on the initial semantic representation vector; The negative sentiment score is calculated by aggregation using the following formula: ； in, This indicates the number of multi-dimensional features in marketing and promotional text data. Indicates the first Dimensioning the probability of negative emotions This indicates the score for negative emotions.

4. A method for dynamic detection and interception of marketing content compliance driven by a large model, as described in claim 1, is characterized in that... The process of mining sentence structure from initial semantic representation vectors in marketing promotion text data to obtain a dependency graph with latent meanings includes: Extract the lexical feature set and the syntactic feature set from the initial semantic representation vector; Based on the lexical feature set, determine the sequence of entity nodes in the marketing promotion text data; Based on the set of syntactic features, determine the dependency relationships between entity nodes in the entity node sequence. Combine entity node sequences with dependency relationships to construct an initial sentence structure topology graph; Extract the hidden semantic nodes from the initial sentence structure topology graph; Based on the implicit semantic nodes, a dependency graph vector with potential implicit meanings is generated; Obtain the vector of the preset context knowledge base; Calculate the matching similarity between the dependency graph vector of potential implicit meaning and the vector of the pre-defined context knowledge base; Compare the matching similarity with the preset similarity threshold; If the matching similarity is greater than the similarity threshold, extract high-risk semantic features from the dependency graph vector with potential implicit meanings; Based on the high-risk semantic features, the corresponding risk tags are retrieved from the preset context knowledge base vector; The retrieved risk tags are sorted according to the order of the entity node sequence to generate a sequence of hidden risk tags; Combine the dependency graph of implicit risk label sequences and potential implicit meanings to complete the construction of the dependency graph; Extract contextual features from marketing and promotional text data; Based on contextual association features, the dependency relationship connections in the initial sentence structure topology graph are weighted to obtain weighted dependency relationship connections. Based on weighted dependency connections, core dependency paths with weight values ​​greater than a preset connection threshold are selected. The core dependency path is defined as the input feature of the dependency graph vector with potential implicit meaning; The input features are dimensionality reduced to obtain standardized dependency graph vectors with latent meanings. When calculating the matching similarity between the dependency graph vector with potential implicit meaning and the vector of the preset context knowledge base, the cosine similarity algorithm is used to calculate the specific numerical value of the matching similarity.

5. A method for dynamic detection and interception of marketing content compliance driven by a large model according to claim 4, characterized in that, The similarity between the dependency graph of the implicit meaning and the preset context knowledge base is calculated using the cosine similarity algorithm: ； in, This indicates the similarity between the dependency graph and the contextual knowledge base. Dependency graph vectors representing potential implicit meanings. Representing vectors transpose, Represents a predefined context knowledge base vector, if If the threshold is exceeded, a hidden risk label sequence is generated.

6. A method for dynamic detection and interception of marketing content compliance driven by a large model according to claim 1, characterized in that, The process involves generating a comprehensive compliance scoring matrix based on the implicit risk label sequence and combining it with sentiment vectors. The overall compliance risk score is then obtained by weighting each label according to its risk weight. The weighted calculation formula is as follows: ； in, This indicates the overall compliance risk score. The weights representing the negative scores in the sentiment vector. Indicates the first Risk weights for implicit risk labels Indicates the first A quantitative score for each implicit risk label. This represents the total number of hidden risk labels.

7. A method for dynamic detection and interception of marketing content compliance driven by a large model according to claim 1, characterized in that, The method employs a real-time feedback mechanism to mark marketing promotion text data as high-risk, determining a list of high-risk content segments, including: Obtain the specific value of the preset warning threshold; Compare the overall compliance risk score with the specific values ​​of the preset warning threshold; If the overall compliance risk score exceeds the preset warning threshold, a real-time feedback mechanism will be triggered. High-risk type markers are written into the metadata of marketing and promotion text data through a real-time feedback mechanism; Based on the high-risk type markers, extract the text fragments in the marketing promotion text data that correspond to the implicit risk label sequences; Record the start and end positions of text segments within the marketing and promotional text data; Extract the corresponding text content based on the start and end positions; Summarize the extracted text content and generate a list of high-risk content fragments.

8. A method for dynamic detection and interception of marketing content compliance driven by a large model according to claim 1, characterized in that, The process involves applying an adaptive rule adjustment algorithm to the list of content fragments, fusing sentiment vectors and implicit risk label sequences for iterative correction, to obtain the final compliant version of the marketing promotion text content, including: Obtain initial marketing and promotional text data; We perform preliminary extraction of sentiment vectors and implicit risk information contained in the initial marketing and promotion text data to obtain a basic set of sentiment tendencies and risk tags; An adaptive rule adjustment algorithm is used to analyze the extracted basic sentiment and risk label sets. Detect whether the basic sentiment tendency deviates from the preset sentiment tendency threshold; Detect whether the number of tags in the risk tag set exceeds the preset safety range; If the basic sentiment tendency is detected to deviate from the preset sentiment tendency threshold, or if the number of tags in the risk tag set exceeds the preset safety range, the adjustment mechanism will be triggered. Based on the adjustment mechanism, identify the content segments to be corrected from the list of content segments; A fusion-iteration method is used to match sentiment vectors with implicit risk label sequences layer by layer; Extract the deviation from the layer-by-layer matching results; For the deviations, the text expression of the content segments to be corrected is gradually revised to obtain the preliminary optimized text content; Based on the preliminary optimized text content and in conjunction with the standard requirements for compliant text, extract the compliance features from the preliminary optimized text content; Compare compliance features with the standard requirements of compliance documents; Determine whether there are any non-standard expressions in the initially optimized text content; If there are expressions that do not conform to the specifications, extract the specific location information of the non-conformities; By using text optimization techniques, local adjustments are made to the initially optimized text content based on specific location information; Based on the results of the local adjustments, a second-processed text version is generated. Obtain the text version after secondary processing; A deep scan was performed on the remaining risk control points in the second-processed text version; Obtain depth scan results; Determine if the depth scan results indicate any potential risks; If the depth scan results show potential problems, the text version after secondary processing is further optimized by an adaptive rule adjustment algorithm to obtain a text structure that meets the requirements. Perform a comprehensive validation of the text structure that meets the requirements; Based on the goals and needs of marketing and promotion, extract the emotional expression features and content presentation features from the text structure that meets the requirements; Determine whether the characteristics of emotional expression and content presentation meet the expected emotional expression and content presentation; If the expected emotional expression and content presentation are met, the output text structure that meets the requirements will be the final compliant version of the marketing promotion text content.

9. A method for dynamic detection and interception of marketing content compliance driven by a large model according to claim 8, characterized in that, The calculation process of the adaptive rule adjustment algorithm is as follows: No. The overall compliance score for each iteration is obtained by weighted fusion of the sentiment score and the risk label normalized score: ； in, For the first Overall compliance score, For the first In terms of emotional weight, Score the sentiment of the text. Normalize the score for the text risk label; The sentiment weights are dynamically updated based on sentiment bias and risk bias. ； in, The iteration step size, For emotional threshold, This is the risk safety threshold; The iteration stops when the overall compliance score converges and both the sentiment bias and risk label meet the requirements. ,and ,in, This is the convergence accuracy threshold; The correction strength of each text segment is calculated based on the sentiment weights after iterative convergence, guiding local text rewriting: ； in, For the first The strength of correction for each content segment, The sentiment weights after iterative convergence. For the first The emotional score of each content segment. For the first The risk score for each content segment.

10. A large-scale model-driven dynamic detection and interception system for marketing content compliance, characterized in that, The system, applied to the large-model-driven dynamic detection and interception method for marketing content compliance as described in any one of claims 1-9, comprises: The text acquisition module is used to retrieve marketing promotion text data to be detected from the marketing content database; The semantic extraction module, equipped with a semantic processing model, is used to extract features from marketing and promotional text data and output an initial semantic representation vector. The sentiment analysis module is equipped with a pre-trained sentiment analysis model. It receives the initial semantic representation vector, performs multi-dimensional sentiment classification and aggregation calculation, and outputs a sentiment vector containing positive sentiment score distribution, negative sentiment score distribution and neutral sentiment score distribution. It also compares the negative sentiment score with a preset threshold. The semantic mining module is used to mine the sentence structure in the marketing and promotion text data based on the initial semantic representation vector when the negative sentiment score in the sentiment vector exceeds a preset threshold. It constructs a dependency graph of potential implicit meanings, generates dependency graph vectors and matches them with preset context knowledge base vectors, and outputs a dependency graph of potential implicit meanings and a sequence of implicit risk labels. The compliance scoring module is used to generate a comprehensive compliance scoring matrix based on the implicit risk label sequence and the sentiment vector. It calculates the risk weight of each label by weighting and outputs the overall compliance risk score. The real-time feedback module is used to trigger the real-time feedback mechanism when the overall compliance risk score is higher than the preset warning threshold. It marks the marketing and promotion text data as high-risk, extracts the corresponding text fragments, and generates a list of high-risk content fragments. The iterative correction module, equipped with an adaptive rule adjustment algorithm, applies the adaptive rule adjustment algorithm to the list of content fragments, integrates sentiment vectors and implicit risk label sequences for iterative correction, and outputs the final compliant version of the marketing promotion text content; The context knowledge base module is used to store preset context knowledge base vectors, providing data support for matching similarity calculation in the semantic mining module; The parameter configuration module is used to preset and store all preset hyperparameters, thresholds, weight coefficients and iteration-related parameters in the method, providing parameter support for the operation of each module.

Images