AI marketing creative generation method driven by multi-source information fusion

Abstract

The present invention discloses an AI marketing creative generation method driven by multi-source information fusion, which relates to the field of marketing creative generation technology and includes the following steps: extracting the core concept keywords of the event for a preset target event, and establishing a unified keyword mapping dictionary in multiple social media platforms based on the core concepts; synchronously crawling high-frequency text data related to the target event from multiple social media platforms within a preset time window; extracting the semantic center vector and emotional tendency label corresponding to the same keyword in the context of different platforms. The method of the present invention realizes multi-platform semantic standardization extraction by constructing a keyword mapping dictionary, time window collection and corpus filtering; combining semantic fingerprint modeling with deviation measurement, integrating user interaction density for dynamic weighting, and finally regulating the semantic fusion weight based on the beta function, thereby improving the emotional fit and contextual accuracy of the creative generated content and enhancing the brand communication effect and security.

Description

Technical Field

[0001] The present invention relates to the technical field of marketing creativity generation, and in particular to an AI marketing creativity generation method driven by multi-source information fusion. Background Art

[0002] AI-powered marketing creative generation driven by multi-source information fusion integrates data from multiple heterogeneous sources (such as user behavior data, social media content, market trends, product reviews, and competitor updates), leveraging AI models for semantic understanding, sentiment analysis, and user profiling. This approach drives the automatic generation and personalized recommendation of creative content. This approach not only improves the relevance and diversity of creative generation but also dynamically adjusts creative direction based on real-time data, achieving precise reach and efficient conversion of target users. Its core lies in combining big data mining with generative AI technologies, enabling creative marketing to be data-driven, adaptively optimized, and iterative at high frequencies.

[0003] The existing technology has the following deficiencies:

[0004] In AI-powered marketing creative generation driven by multi-source information fusion, processing dynamic social semantics often faces the challenge of insufficient recognition of contextual shifts. Different platforms (such as Weibo, Douyin, and Zhihu) exhibit significant differences in semantic expression, sentiment, and user perception of the same event. If the fusion engine ignores this cross-platform contextual heterogeneity when capturing user sentiment and integrating hashtags, it may misjudge sentiment. For example, the phrase "lying flat" is often neutral and self-deprecating on Bilibili, but often carries a distinctly negative tone on Weibo. If the system fails to discern this difference, it may inject the wrong sentiment into the automatically generated marketing copy, leading to a misaligned brand stance, user misunderstandings, public opinion controversy, and even reputational risk.

[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0006] The purpose of the present invention is to provide an AI marketing creative generation method driven by multi-source information fusion, and to achieve standardized extraction of multi-platform semantic information by constructing a unified keyword mapping dictionary, combined with time window control and topic filtering to ensure the timeliness and relevance of the corpus. The semantic fingerprints of each platform are extracted using word vector embedding and sentiment modeling, and the semantic-sentiment deviation matrix is generated by integrating cosine distance and KL divergence, and the offset weighted model is constructed by introducing user interaction density to effectively correct sample distortion and participation imbalance. Finally, through normalization and beta function variable weight mechanism, the platform semantic weight is dynamically adjusted, the interference of platforms with high offset is suppressed, the semantic consistency and emotional fit of the generated content are improved, and the brand communication effect and security are enhanced to solve the problems in the above-mentioned background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for generating AI marketing ideas driven by multi-source information fusion, comprising the following steps:

[0008] For the preset target event, we extract the core concept keywords of the event and establish a unified keyword mapping dictionary based on the core concepts across multiple social media platforms, including Weibo, Douyin, and Zhihu. This is used to control the consistency of the corpus source and ensure the comparability of the context comparison in the subsequent data collection stage.

[0009] Within a preset time window, we use a keyword dictionary to simultaneously crawl high-frequency text materials related to the target event across multiple social media platforms, ensuring that the samples collected from each platform have temporal consistency and content consistency.

[0010] We perform word embedding and sentiment label modeling on text corpora collected from multiple social media platforms, extracting the semantic center vectors and sentiment labels corresponding to the same keywords in the context of different platforms, and forming a semantic fingerprint model that reflects the semantic expression characteristics of the platforms.

[0011] The cosine distance difference between the semantic center vectors of the same keyword across multiple social media platforms and the Kullback-Leibler divergence between the corresponding sentiment label distributions are calculated and combined to generate a semantic-sentiment deviation matrix that measures the intensity of expression differences across multiple platforms in the context of the target event.

[0012] Correlation analysis is performed between the semantic-emotional deviation matrix and the user interaction density data of the corresponding social media platform. A contextual deviation weighting factor model is established to correct sample deviation distortion and platform participation imbalance, and the fused deviation value is output.

[0013] The fusion offset value is normalized to obtain the context offset index of the target event within the time window. The semantic fusion weights of multiple social media platforms are nonlinearly and dynamically adjusted according to the context offset index. The Beta function variable weight model is used to assign inhibitory fusion weights to platforms with large offset intensity to control their semantic participation in the generation of artificial intelligence marketing creativity.

[0014] Preferably, the steps of constructing a cross-platform keyword dictionary include:

[0015] Based on the original text description of the preset target event, natural language processing technology is used to perform event semantic analysis to extract core semantic units closely related to the event theme, including but not limited to event keywords, contextual verbs, subject entities, and related emotional modifiers. The initial keyword set is then verified for semantic integrity and redundancy removed through contextual relationship mining based on dependency syntactic analysis to generate a candidate set of core keywords.

[0016] For the core keyword candidate set, we perform keyword expansion tasks on platforms such as Weibo, Douyin, and Zhihu. We use a word embedding vector space model and a hot topic context retrieval mechanism to mine platform-specific keyword expansion sets that are highly semantically relevant to the core keywords and have high usage frequency in the context of each platform. We also record platform-specific expressions and their semantic mapping relationships.

[0017] The keyword sets expanded from multiple social media platforms are semantically normalized to construct a unified keyword mapping dictionary between platforms. Each core keyword corresponds to a mapping entry containing multiple platform keyword variants. The keyword mapping dictionary is stored as a structured vector index and used for retrieval consistency control and context comparison comparability assurance in the subsequent multi-platform corpus collection process.

[0018] Preferably, in order to mine the keyword extension set with high semantic relevance and high usage frequency to the core keyword in each social media platform, a threshold comparison method is adopted, which specifically includes the following operations:

[0019] For each core keyword, use a pre-trained word embedding model (such as Word2Vec, FastText, or BERT embedding) to calculate its word vector representation in the target platform corpus. Then, among the high-frequency words collected from the corresponding platform, calculate the semantic similarity between these words and the core keyword (semantic similarity can be Euclidean distance similarity, Manhattan distance, etc.);

[0020] A semantic relevance threshold is set. When the semantic similarity between a candidate word and a core keyword is greater than or equal to the semantic relevance threshold, the candidate word is determined to be semantically related to the core keyword. At the same time, the usage frequency of each candidate word within the target time window is counted, and a usage frequency threshold is set. Only when the usage frequency of the candidate word is greater than or equal to the usage frequency threshold will it be retained in the keyword expansion set within the platform.

[0021] By integrating keywords that meet the dual threshold conditions, we construct a keyword expansion set that is both semantically relevant and representative in the current platform context, achieving dual guarantees for the accuracy of keyword expansion and platform adaptability.

[0022] Preferably, the step of simultaneously crawling high-frequency text corpora related to the target event in multiple social media platforms includes:

[0023] Set a unified time window, which includes the start time and end time. According to the outbreak time node and propagation cycle of the target event, dynamically adjust the length of the time window to cover the main peak stage and high interaction period of the event propagation, so as to improve the timeliness and representativeness of the sampling.

[0024] Based on the constructed keyword mapping dictionary, parallel keyword search requests are initiated to platforms such as Weibo, Douyin, and Zhihu within a time window. The open interfaces of the corresponding platforms or self-developed crawlers are called to collect various text forms such as original user posts, comment data, and forwarded statements containing keywords. The collected samples are then timestamped and labeled with platform identification tags.

[0025] For the collected original corpus data, a unified content topic screening operation is performed, and a text clustering method based on the topic model, such as the Latent Dirichlet Allocation model, is used to remove redundant information, cross-topic noise or non-event-related text, thereby ensuring that the samples collected from different social media platforms are highly consistent in both the time dimension and the content topic dimension, providing a basic guarantee for subsequent semantic alignment and sentiment modeling.

[0026] Preferably, the step of extracting the semantic center vector and emotional tendency label corresponding to the same keyword in different platform contexts includes:

[0027] For the text corpus collected by each platform, we use a pre-trained word vector model (such as Word2Vec, FastText, or BERT) to embed the text corpus into word vectors and generate a word vector representation for each word. The word vector can reflect the semantic information of the word in the context and its relationship with other words.

[0028] Based on the generated word vectors, sentiment analysis models (such as LSTM, BERT sentiment classifier, or VADER) are used to model sentiment labels for platform texts. This identifies the emotional tendencies in the texts, such as positive, negative, or neutral, and assigns corresponding sentiment labels to each keyword, ensuring that sentiment analysis can adapt to the expression methods of different platforms.

[0029] Combining the obtained word vectors with the sentiment tag data, the semantic center vector of each keyword in the context of each platform is calculated. Specifically, the semantic center vector representation of the keyword is obtained by calculating the mean or weighted average of the word vectors of the keyword in different contexts to reflect the differences in the semantic features of the keyword on different platforms.

[0030] Based on the semantic center vector and emotional tendency label, a platform semantic fingerprint model is formed. The semantic vectors of the same keywords on each platform are combined with the emotional label to create a platform-specific semantic fingerprint for subsequent context understanding and creative generation, ensuring the correct identification and integration of cross-platform emotions and semantics.

[0031] Preferably, the step of generating a semantic-emotional deviation matrix comprises:

[0032] For each keyword, we obtained the semantic center vector of the keyword on each platform from multiple social platforms and calculated the cosine distance difference between the platforms. That is, we calculated the angle difference between the semantic center vectors of the same keyword on different platforms to measure the degree of difference in semantic understanding of the keyword between the platforms.

[0033] Based on the sentiment analysis results, we extract the distribution of sentiment labels for the keyword on each platform (e.g., positive, negative, neutral, etc.) and calculate the Kullback-Leibler divergence between the sentiment label distributions on each platform. This divergence measures the distribution differences in sentiment tendencies for the same keyword across different platforms.

[0034] The cosine distance difference value and Kullback-Leibler divergence value are combined to obtain a comprehensive semantic-emotional deviation metric. This value can fully reflect the degree of deviation between the semantic and emotional expressions of the same keyword on different platforms and reflect the intensity of contextual differences across platforms.

[0035] Based on the semantic-emotion deviation measurement value, a semantic-emotion deviation matrix is constructed. As a key component of multi-platform data fusion, the semantic-emotion deviation matrix is further used for cross-platform semantic optimization in subsequent creative generation, dynamically adjusting the fusion weight of the semantics of each platform to enhance the accuracy and emotional consistency of the final creative content.

[0036] Preferably, the steps of fusing the cosine distance difference value and the Kullback-Leibler divergence value to obtain a comprehensive semantic-emotion deviation metric are as follows:

[0037] The cosine distance difference between the semantic center vectors of the same keyword and the Kullback-Leibler divergence of the sentiment label distribution are calculated across each platform. These two measures reflect the semantic and sentiment differences between platforms, respectively. Next, to comprehensively account for differences in both semantic and sentiment dimensions, a preset weighting coefficient is set to assign corresponding weights to the cosine distance difference and Kullback-Leibler divergence. These weights are typically determined through historical data analysis or experimental validation to ensure that they accurately reflect the contribution of semantics and sentiment to the overall deviation in specific applications. The cosine distance difference and Kullback-Leibler divergence are then weighted and summed to produce a comprehensive deviation metric. This value comprehensively reflects not only the strength of semantic differences but also the deviation in sentiment expression, ultimately generating a unified semantic-sentiment deviation metric. This metric effectively describes the overall degree of contextual variation across platforms, providing a more accurate basis for contextual adjustment in subsequent multi-platform data fusion and creative generation.

[0038] Preferably, the steps of establishing a context offset weighting factor model for correcting sample offset distortion and platform participation imbalance, and outputting a fused offset value include:

[0039] Based on the semantic-sentiment deviation matrix of each platform, we extract the user interaction density data of the platform. The interaction density includes but is not limited to the frequency of likes, comments, reposts, and shares. This data reflects the popularity and user engagement of keywords or topics on the platform.

[0040] We correlated the semantic-emotional deviation matrix with the platform's user interaction density data and weighted the semantic deviation based on the interaction density of different platforms. In particular, we considered that on platforms with lower interaction density, the contextual deviation of keywords may be distorted due to a smaller sample size and thus require correction with a higher weight.

[0041] Based on the results of the correlation analysis, a context shift weighting factor model was established. This model uses the platform's interaction density information as a correction parameter to dynamically adjust the semantic-emotional deviation metric to ensure that context shift on low-interaction platforms is appropriately suppressed, while giving higher weight to high-interaction platforms, thereby avoiding shift distortion caused by differences in platform engagement.

[0042] Through the contextual offset weighting factor model, a weighted and corrected fusion offset value is output. This value comprehensively reflects the semantic and emotional expression deviations of different platforms on the target event, and is reasonably adjusted according to the differences in platform interaction density, providing accurate weighted support for multi-platform data fusion in subsequent creative generation.

[0043] Preferably, the semantic fusion weights of multiple social media platforms are nonlinearly and dynamically adjusted according to the context shift index. The specific steps are as follows:

[0044] The fusion offset values of each platform are mapped to a unified scale, and the context offset index is calculated to measure the intensity of the semantic emotion difference between each platform and other platforms in the current time window. The calculation expression of the context offset index is: ,in: For the platform i The context shift index, normalized to scope; For the platform i The fusion offset value has been weighted by the semantic-emotional difference and interaction density; and are the minimum and maximum fusion offset values for all platforms; It is a decimal offset to prevent the normalized result from being zero. The recommended value is , used to enhance the stability of subsequent beta functions;

[0045] By utilizing the nonlinear shape characteristics of the beta function, the context shift index is mapped into a weighted suppression function for the platform semantic participation, achieving the regulation goal that the greater the shift, the lower the participation. The mapping expression of the context shift index is: ,in: For the platform i The initial semantic fusion weight of and is the Beta distribution parameter, used to control the shape of the weight curve. The recommended setting is:

[0046] If you need to suppress the high offset platform, you can set , , right skewed;

[0047] If neutral adjustment is required, you can set , , evenly distributed;

[0048] is the normalization constant of the standard beta function to ensure that the sum of weights is controllable;

[0049] After obtaining the initial semantic fusion weight of each platform based on the context shift index, the initial semantic fusion weights of all platforms are normalized to calculate the final actual semantic fusion weight. The calculation expression of the actual semantic fusion weight is as follows: ,in: Representation Platform i The final actual semantic fusion weight (i.e. the final semantic engagement weight) is used to control the platform iThe proportion of semantic integration in the creative generation process; n The total number of social media platforms participating in the integration; represents the weighted sum of all platform fusion weights, j is a number from 1 to n The integer index represents the numbers of all social media platforms participating in semantic fusion.

[0050] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0051] This paper constructs a unified keyword mapping dictionary to achieve standardized extraction of multi-platform semantic information, ensuring semantic alignment across corpora across platforms. It also combines time window control with corpus topic filtering to further ensure the timeliness and content relevance of collected data. Subsequently, word vector embedding and sentiment tag modeling techniques are used to extract semantic fingerprint information unique to each social media platform, quantifying the semantic differences and sentiment deviation of the same keyword in different platform contexts. The cosine distance and Kullback-Leibler divergence are then combined to construct a semantic-sentiment deviation matrix that comprehensively reflects the intensity of contextual differences. Furthermore, the deviation matrix is correlated with platform user interaction density data to construct an offset weighting factor model, dynamically correcting sample offset distortion and platform participation imbalance, thereby improving the accuracy and robustness of the contextual offset metric. Finally, by normalizing the fused offset value and combining it with a beta function variable weighting mechanism, the fusion weights of each platform's semantics are nonlinearly adjusted, effectively suppressing semantic interference from platforms with high offset intensity and increasing engagement on platforms with high semantic consistency. This significantly enhances the emotional relevance, expression accuracy, and user acceptance of the generated content, comprehensively improving brand communication effectiveness and public relations security. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction to the drawings required for use in the embodiments will be given below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0053] Figure 1 This is a flow chart of the method for generating AI marketing ideas driven by multi-source information fusion of the present invention. DETAILED DESCRIPTION

[0054] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of this disclosure will be thorough and complete and will fully convey the concepts of the example embodiments to those skilled in the art.

[0055] The present invention provides Figure 1 The AI marketing creative generation method driven by multi-source information fusion shown includes the following steps:

[0056] S101. Extract key words related to the target event. Based on these key words, establish a unified keyword mapping dictionary across multiple social media platforms, including Weibo, TikTok, and Zhihu. This will be used to ensure consistency of corpus sources and comparability of context comparisons during the subsequent data collection phase.

[0057] The steps to build a cross-platform keyword dictionary include:

[0058] Based on the original text description of the preset target event, natural language processing technology is used to perform event semantic analysis to extract core semantic units closely related to the event theme, including but not limited to event keywords, contextual verbs, subject entities, and related emotional modifiers. The initial keyword set is then verified for semantic integrity and redundancy removed through contextual relationship mining based on dependency syntactic analysis to generate a candidate set of core keywords.

[0059] For the core keyword candidate set, we perform keyword expansion tasks on platforms such as Weibo, Douyin, and Zhihu. We use a word embedding vector space model and a hot topic context retrieval mechanism to mine platform-specific keyword expansion sets that are highly semantically relevant to the core keywords and have high usage frequency in the context of each platform. We also record platform-specific expressions and their semantic mapping relationships.

[0060] The keyword sets expanded from multiple social media platforms are semantically normalized to construct a unified keyword mapping dictionary between platforms. Each core keyword corresponds to a mapping entry containing multiple platform keyword variants. The keyword mapping dictionary is stored as a structured vector index and used for retrieval consistency control and context comparison comparability assurance in the subsequent multi-platform corpus collection process.

[0061] In order to mine the extended set of keywords that are highly semantically relevant to the core keywords and frequently used in various social media platforms, a threshold comparison method is used, which specifically includes the following operations:

[0062] For each core keyword, use a pre-trained word embedding model (such as Word2Vec, FastText, or BERT embedding) to calculate its word vector representation in the target platform corpus. Then, among the high-frequency words collected from the corresponding platform, calculate the semantic similarity between these words and the core keyword (semantic similarity can be Euclidean distance similarity, Manhattan distance, etc.);

[0063] A semantic relevance threshold is set. When the semantic similarity between a candidate word and a core keyword is greater than or equal to the semantic relevance threshold, the candidate word is determined to be semantically related to the core keyword. At the same time, the usage frequency of each candidate word within the target time window is counted, and a usage frequency threshold is set. Only when the usage frequency of the candidate word is greater than or equal to the usage frequency threshold will it be retained in the keyword expansion set within the platform.

[0064] By integrating keywords that meet the dual threshold conditions, we construct a keyword expansion set that is both semantically relevant and representative in the current platform context, achieving dual guarantees for the accuracy of keyword expansion and platform adaptability.

[0065] This step aims to establish a unified semantic framework for the collection and analysis of event-related corpus from multi-source social media platforms, ensuring corpus consistency and comparability of contextual comparisons in the subsequent data fusion process. On different social platforms, even when discussing the same event, users may use significantly different language styles, keyword expressions, topic tags, and even emotional tendencies. For example, on Weibo, trending words and simplified colloquial tags are often used to express emotions, while the Zhihu platform tends to favor rational discussions, long sentence structures, and professional terminology, while Douyin mainly uses video titles and short phrase tags. This difference in expression can lead to semantic confusion and biased judgments when directly performing sentiment modeling or keyword statistics on multi-platform corpora without standardization.

[0066] Therefore, by pre-extracting the core concept keywords of the target event and constructing a keyword mapping dictionary with cross-platform semantic alignment capabilities, the semantic expressions of the same event on different platforms can be summarized into unified concept categories, achieving alignment at the data structure level. This mapping dictionary not only serves as a set of keyword matching rules during data collection, improving the relevance and coverage of data acquisition, but also establishes a unified semantic standard for subsequent semantic fingerprint extraction, sentiment modeling, and cross-platform context shift calculation. It can also be used to control the structural noise caused by keyword differences between platforms, enhancing the stability of data fusion and the contextual rationality of generated content.

[0067] S102. Within a preset time window, using a keyword dictionary, simultaneously crawl high-frequency text corpora related to the target event across multiple social media platforms to ensure that the samples collected from each platform have temporal consistency and content consistency;

[0068] The steps for simultaneously crawling high-frequency text corpora related to target events from multiple social media platforms include:

[0069] Set a unified time window, which includes the start time and end time. According to the outbreak time node and propagation cycle of the target event, dynamically adjust the length of the time window to cover the main peak stage and high interaction period of the event propagation, so as to improve the timeliness and representativeness of the sampling.

[0070] Based on the constructed keyword mapping dictionary, parallel keyword search requests are initiated to platforms such as Weibo, Douyin, and Zhihu within a time window. The open interfaces of the corresponding platforms or self-developed crawlers are called to collect various text forms such as original user posts, comment data, and forwarded statements containing keywords. The collected samples are then timestamped and labeled with platform identification tags.

[0071] For the collected original corpus data, a unified content topic screening operation is performed, and a text clustering method based on the topic model, such as the Latent Dirichlet Allocation model, is used to remove redundant information, cross-topic noise or non-event-related text, thereby ensuring that the samples collected from different social media platforms are highly consistent in both the time dimension and the content topic dimension, providing a basic guarantee for subsequent semantic alignment and sentiment modeling.

[0072] The core function of this step is to ensure that the text data collected from multiple social media platforms is highly comparable in terms of timeliness and topic relevance through the control of time consistency and content theme consistency, laying a solid foundation for subsequent multi-platform data fusion and analysis. Specifically, time consistency refers to ensuring that the collected samples from all social platforms are generated within the same time period within a preset time window, avoiding delayed or incomplete event information due to large differences in collection time. This is especially true in the social media environment where information disseminates rapidly, and the popularity of events, user emotions, and discussion topics will change rapidly over time. Only by ensuring that data is captured within the same time window can we ensure that the event is fully captured and accurately reflected.

[0073] On the other hand, content theme consistency means that the data collected on different platforms must be consistent with the core theme of the target event to avoid irrelevant or cluttered information. For example, although the same event may be discussed on platforms such as Weibo, Douyin, and Zhihu, the discussion methods, topic focus, and user feedback may vary. If the data collection lacks theme consistency, irrelevant content may be introduced, affecting subsequent semantic analysis and sentiment modeling. Therefore, ensuring that the text content captured on each platform is strictly filtered around the core of the event can ensure the relevance and representativeness of the data, thereby providing high-quality input data for subsequent keyword expansion, sentiment analysis, and creative generation.

[0074] This step not only improves the efficiency of multi-platform data collection, but also effectively avoids information noise and bias, ensuring the accuracy and reliability of subsequent analysis results.

[0075] S103, performing word vector embedding and sentiment label modeling operations on the text corpora collected from multiple social media platforms, extracting the semantic center vectors and sentiment tendency labels corresponding to the same keywords in the context of different platforms, and forming a semantic fingerprint model that reflects the semantic expression characteristics of the platforms;

[0076] The steps for extracting the semantic center vector and sentiment labels corresponding to the same keyword in different platform contexts include:

[0077] For the text corpus collected by each platform, we use a pre-trained word vector model (such as Word2Vec, FastText, or BERT) to embed the text corpus into word vectors and generate a word vector representation for each word. The word vector can reflect the semantic information of the word in the context and its relationship with other words.

[0078] Based on the generated word vectors, sentiment analysis models (such as LSTM, BERT sentiment classifier, or VADER) are used to model sentiment labels for platform texts. This identifies the emotional tendencies in the texts, such as positive, negative, or neutral, and assigns corresponding sentiment labels to each keyword, ensuring that sentiment analysis can adapt to the expression methods of different platforms.

[0079] Combining the obtained word vectors with the sentiment tag data, the semantic center vector of each keyword in the context of each platform is calculated. Specifically, the semantic center vector representation of the keyword is obtained by calculating the mean or weighted average of the word vectors of the keyword in different contexts to reflect the differences in the semantic features of the keyword on different platforms.

[0080] Based on the semantic center vector and emotional tendency label, a platform semantic fingerprint model is formed. The semantic vectors of the same keywords on each platform are combined with the emotional label to create a platform-specific semantic fingerprint for subsequent context understanding and creative generation, ensuring the correct identification and integration of cross-platform emotions and semantics.

[0081] This step involves embedding words and modeling sentiment labels for text corpora collected from multiple social media platforms. This process extracts the semantic center vectors and sentiment labels for the same keyword in the context of different platforms, thereby forming a semantic fingerprint model that reflects the unique semantic expression characteristics of each platform. Different social media platforms (such as Weibo, Douyin, and Zhihu) have their own unique language styles, expressions, and user groups. This can lead to significant differences in the context and sentiment of the same event or keyword across different platforms. For example, certain words may have a more entertaining and lighthearted tone on Douyin, while they may convey more emotional and political leanings on Weibo. Therefore, relying solely on sentiment analysis and semantic understanding from a single platform can lead to misjudgments of sentiment and affect the effectiveness of marketing creative generation.

[0082] By using word embedding technologies (such as Word2Vec and BERT), text content can be converted into digital representations of high-dimensional vectors that contain the deep semantic information of the words. Based on this, sentiment labeling modeling techniques (such as LSTM and BERT sentiment classification models) are used to analyze the sentiment of the text and assign sentiment labels to each keyword. Subsequently, by calculating the semantic center vectors and sentiment labels for the same keyword across different platforms, the unique semantic expression characteristics and sentiment trends of each platform can be accurately captured. The resulting semantic fingerprint model not only helps identify contextual differences across platforms but also provides accurate data support for subsequent creative generation, sentiment manipulation, and personalized content recommendations. This step plays a crucial role in multi-platform data integration, ensuring accurate cross-platform contextual understanding and enabling brands to provide more refined and personalized marketing creatives.

[0083] S104, calculating the cosine distance difference between the semantic center vectors of the same keyword across multiple social media platforms and the Kullback-Leibler divergence value between the corresponding sentiment tag distributions, and combining the two to generate a semantic-sentiment deviation matrix for measuring the intensity of expression differences across multiple platforms in the context of the target event;

[0084] The steps to generate the semantic-sentiment deviation matrix include:

[0085] For each keyword, we obtained the semantic center vector of the keyword on each platform from multiple social platforms and calculated the cosine distance difference between the platforms. That is, we calculated the angle difference between the semantic center vectors of the same keyword on different platforms to measure the degree of difference in semantic understanding of the keyword between the platforms.

[0086] Based on the sentiment analysis results, we extract the distribution of sentiment labels for the keyword on each platform (e.g., positive, negative, neutral, etc.) and calculate the Kullback-Leibler divergence between the sentiment label distributions on each platform. This divergence measures the distribution differences in sentiment tendencies for the same keyword across different platforms.

[0087] The cosine distance difference value and Kullback-Leibler divergence value are combined to obtain a comprehensive semantic-emotional deviation metric. This value can fully reflect the degree of deviation between the semantic and emotional expressions of the same keyword on different platforms and reflect the intensity of contextual differences across platforms.

[0088] Based on the semantic-emotion deviation measurement value, a semantic-emotion deviation matrix is constructed. As a key component of multi-platform data fusion, the semantic-emotion deviation matrix is further used for cross-platform semantic optimization in subsequent creative generation, dynamically adjusting the fusion weight of the semantics of each platform to enhance the accuracy and emotional consistency of the final creative content.

[0089] The steps for fusing the cosine distance difference value and the Kullback-Leibler divergence value to obtain a comprehensive semantic-sentiment deviation measure are as follows:

[0090] The cosine distance difference between the semantic center vectors of the same keyword and the Kullback-Leibler divergence of the sentiment label distribution are calculated across each platform. These two measures reflect the semantic and sentiment differences between platforms, respectively. Next, to comprehensively account for differences in both semantic and sentiment dimensions, a preset weighting coefficient is set to assign corresponding weights to the cosine distance difference and Kullback-Leibler divergence. These weights are typically determined through historical data analysis or experimental validation to ensure that they accurately reflect the contribution of semantics and sentiment to the overall deviation in specific applications. The cosine distance difference and Kullback-Leibler divergence are then weighted and summed to produce a comprehensive deviation metric. This value comprehensively reflects not only the strength of semantic differences but also the deviation in sentiment expression, ultimately generating a unified semantic-sentiment deviation metric. This metric effectively describes the overall degree of contextual variation across platforms, providing a more accurate basis for contextual adjustment in subsequent multi-platform data fusion and creative generation.

[0091] This step precisely quantifies the differences in semantic understanding and emotional expression of the same keyword across multiple social media platforms, thereby constructing a semantic-emotional deviation matrix to measure the intensity of differences in the contextual expression of target events across platforms, providing key support for subsequent multi-platform data fusion, context optimization, and creative generation.

[0092] First, when discussing the same event or topic on social media platforms like Weibo, Douyin, and Zhihu, users vary in their language styles, sentiment, and hashtag usage. Factors such as the user base, discussion patterns, and emotional expressions of these platforms can lead to significant differences in the semantics and sentiment expressed for the same keyword across different platforms. To accurately capture these differences, it is necessary to independently model the semantics and sentiment expressed for the same keyword on each platform and conduct comparative analysis of these models.

[0093] In this step, we first calculate the cosine distance difference between the semantic center vectors of the same keyword on each platform to quantify the differences in semantic understanding between different platforms. Cosine similarity measures the angle between two word vectors. A smaller angle indicates closer semantics, while a smaller angle indicates greater semantic divergence. This metric can help us understand whether the core semantics of the same keyword are consistent or different across platforms.

[0094] Next, we calculated the Kullback-Leibler divergence between the distributions of sentiment labels for each platform to further quantify the deviations in sentiment expression across platforms. Kullback-Leibler divergence measures the difference between two probability distributions and accurately describes the distribution differences in sentiment labels (such as positive, negative, and neutral) across platforms. This divergence can be used to determine whether the sentiment tendencies of the same keyword differ significantly across platforms, thereby revealing the deviations in sentiment expression across platforms.

[0095] Finally, a weighted fusion of semantic differences (cosine distance) and sentiment differences (Kullback-Leibler divergence) is used to generate a comprehensive semantic-sentiment deviation metric, which is then used to construct a semantic-sentiment deviation matrix. This matrix effectively measures the intensity of contextual differences across platforms, reflecting the diverse expressions of the same event across different platforms. Using this matrix, the system can identify differences in semantic and sentimental expression across platforms and automatically adjust the fusion weights of data from different platforms during the subsequent creative generation process, ensuring that the resulting creative content better meets the cognitive and emotional expectations of users across multiple platforms.

[0096] In summary, this step plays a crucial role in the entire creative generation process driven by multi-source information fusion. It not only helps understand and quantify cross-platform contextual differences, but also provides solid data support for accurate creative generation, personalized recommendations, and public opinion analysis.

[0097] S105, performing correlation analysis on the semantic-emotional deviation matrix and the user interaction density data of the corresponding social media platform, establishing a contextual deviation weighting factor model for correcting sample deviation distortion and platform participation imbalance, and outputting a fusion deviation value;

[0098] The steps of establishing a context offset weighting factor model for correcting sample offset distortion and platform participation imbalance and outputting a fused offset value include:

[0099] Based on the semantic-sentiment deviation matrix of each platform, we extract the user interaction density data of the platform. The interaction density includes but is not limited to the frequency of likes, comments, reposts, and shares. This data reflects the popularity and user engagement of keywords or topics on the platform.

[0100] We correlated the semantic-emotional deviation matrix with the platform's user interaction density data and weighted the semantic deviation based on the interaction density of different platforms. In particular, we considered that on platforms with lower interaction density, the contextual deviation of keywords may be distorted due to a smaller sample size and thus require correction with a higher weight.

[0101] Based on the results of the correlation analysis, a context shift weighting factor model was established. This model uses the platform's interaction density information as a correction parameter to dynamically adjust the semantic-emotional deviation metric to ensure that context shift on low-interaction platforms is appropriately suppressed, while giving higher weight to high-interaction platforms, thereby avoiding shift distortion caused by differences in platform engagement.

[0102] Through the contextual offset weighting factor model, a weighted and corrected fusion offset value is output. This value comprehensively reflects the semantic and emotional expression deviations of different platforms on the target event, and is reasonably adjusted according to the differences in platform interaction density, providing accurate weighted support for multi-platform data fusion in subsequent creative generation.

[0103] Suppose we are analyzing multi-platform data on an "environmental advocacy" campaign. We collected keyword data related to "environmental advocacy" from three social platforms: Weibo, Douyin, and Zhihu. We then calculated the semantic-sentiment deviation matrix for these keywords on each platform. For example, on Weibo, the keywords "environmental protection" and "green living" tend to be sentimentally negative, reflecting some users' dissatisfaction with environmental policies. Semantically, they may question environmental measures, resulting in a high degree of sentiment deviation. On Zhihu, however, the keyword "environmental protection" carries a positive sentiment, with users expressing support and advocacy for environmental policies, and the semantics tend to favor rational discussion and proposals. Therefore, the calculated sentiment deviation values for Weibo and Zhihu might be 0.8 and 0.2, respectively, representing the difference in sentiment between the two platforms.

[0104] Next, we obtained the user interaction density data of each platform. Assuming that the interaction density (likes, comments, reposts, etc.) of the Weibo platform is relatively high, reaching an average of 500 interactions per post, while the Douyin platform is lower, with only 100 interactions, and the interaction density of the Zhihu platform is 300 interactions. Through correlation analysis, we found that although the emotional expression on the Zhihu platform is relatively positive, due to the low interaction density and small sample size, there may be a risk of emotional bias distortion. Therefore, in the weighting process, the semantic-emotional deviation value of Zhihu will be adjusted to 0.25 (after weight suppression) based on its low interaction density, while the semantic deviation value of the Weibo platform is maintained at 0.8, and the Douyin platform is further adjusted according to the interaction density. Finally, through the contextual offset weighting factor model, the fusion offset value obtained may be:

[0105] Final weighted value for Weibo platform: 0.75 (original value 0.8, minus interaction density adjustment);

[0106] Final weighted value of Zhihu platform: 0.25 (original value 0.2, with added weighting for interaction density);

[0107] Final weighted value of the TikTok platform: 0.5 (weighted based on interaction density);

[0108] Through these weighted fusion offset values, we can precisely adjust the fusion weight of platform data in creative generation, ensuring that content on high-interaction platforms such as Weibo has more say in creative generation, while low-interaction platforms such as Zhihu can avoid excessively affecting the final creative expression effect based on weighted correction.

[0109] This step constructs a contextual deviation weighting factor model with dynamic correction capabilities by correlating the semantic-emotional deviation matrix with the user interaction density data of various social media platforms. This solves the problems of contextual distortion and platform participation imbalance caused by uneven sample numbers and different interactive activity levels in cross-platform data fusion. In actual multi-source social media environments, different platforms often have deviations in their semantic understanding and emotional expression of the same event, and these deviations are not always reliable or representative. For example, due to low user activity on a platform, only a few people may participate in the discussion. However, if the semantic and emotional deviations of these few samples are high, directly including them in the analysis will cause excessive weighting, affecting the overall deviation determination results, and thus interfering with the accuracy and emotional consistency of subsequent creative content generation.

[0110] To address this issue, this step first associates the semantic-emotional deviation matrix calculated for each platform with the user interaction density indicators of that platform, such as the average number of likes, comments, reposts, and views. Interaction density essentially reflects the level of attention a keyword or event receives on the platform and the depth of user participation, and has a high representative weight value. Then, by fitting and modeling the relationship between interaction density and the degree of semantic deviation, a contextual deviation weighting factor model is established. This model can be used to dynamically adjust the influence of the semantic-emotional deviation value of each platform, suppress the distortion problem of high deviation values on low-participation platforms, and improve the representativeness of data on high-participation platforms. Ultimately, the fused deviation value output by the model can provide a more realistic and balanced contextual portrayal result that has been corrected for the intensity of user participation for subsequent content generation.

[0111] In summary, the significance of this step lies in introducing a "participation-aware" weight adjustment mechanism for multi-platform data fusion. This not only enhances the model's robustness to small sample anomalies, but also improves the generated content's fit with mainstream user sentiment trends. It is one of the key links in achieving high-credibility creative generation.

[0112] S106. Normalize the fusion offset value to obtain the context offset index of the target event within the time window, and perform nonlinear dynamic adjustment on the semantic fusion weights of multiple social media platforms based on the context offset index. Use a beta function variable weight model to assign inhibitory fusion weights to platforms with large offset intensity to control their semantic participation in the generation of AI marketing creativity.

[0113] The semantic fusion weights of multiple social media platforms are nonlinearly and dynamically adjusted based on the context shift index. The specific steps are as follows:

[0114] The fusion offset values of each platform are mapped to a unified scale, and the context offset index is calculated to measure the intensity of the semantic emotion difference between each platform and other platforms in the current time window. The calculation expression of the context offset index is: ,in: For the platform i The context shift index, normalized to scope; For the platform i The fusion offset value has been weighted by the semantic-emotional difference and interaction density; and are the minimum and maximum fusion offset values for all platforms; It is a decimal offset to prevent the normalized result from being zero. The recommended value is , used to enhance the stability of subsequent beta functions;

[0115] Through normalization, the offset values of different scales are mapped to a unified interval, eliminating dimensional differences, so that the context offset index can be used as the core control parameter for subsequent weight adjustment.

[0116] By utilizing the nonlinear shape characteristics of the beta function, the context shift index is mapped into a weighted suppression function for the platform semantic participation, achieving the regulation goal that the greater the shift, the lower the participation. The mapping expression of the context shift index is: ,in: For the platform i Initial semantic fusion weight of variable x It is a formal variable used to define the mathematical integration process of a function. It does not represent a specific platform or a specific value, but an abstract, independent variable symbol used to represent The main function of this variable is to construct a standard Beta function integral expression so that it can be used for different parameters. and The function curve under is normalized, which is used as the normalization factor of the probability density function. In practical applications, for example, the context shift index When mapping to platform fusion weights, we will x The position is offset by the actual context index Instead, substitute the Beta distribution density function to obtain the participation weight of the platform under the current context shift. However, it should be noted that the integral expression x It is just a symbol used for modeling and calculation, and has no actual meaning of specific platform or semantic indicators; and is the Beta distribution parameter, used to control the shape of the weight curve. The recommended setting is:

[0117] If you need to suppress the high offset platform, you can set , , right skewed;

[0118] If neutral adjustment is required, you can set , , evenly distributed;

[0119] is the normalization constant of the standard beta function to ensure that the sum of weights is controllable;

[0120] Compared with the linear suppression strategy, the beta function is more flexible and adjustable, allowing the weight distribution strategy to be dynamically adapted according to the degree of platform offset, thereby improving the responsiveness of the model and the contextual fit of the generated content.

[0121] After obtaining the initial semantic fusion weight of each platform based on the context shift index, the initial semantic fusion weights of all platforms are normalized to calculate the final actual semantic fusion weight. The calculation expression of the actual semantic fusion weight is as follows: ,in: Representation Platform i The final actual semantic fusion weight (i.e. the final semantic engagement weight) is used to control the platform i The proportion of semantic integration in the creative generation process; n The total number of social media platforms participating in the integration; represents the weighted sum of all platform fusion weights, j is a number from 1 to n The integer index of represents the numbers of all social media platforms participating in semantic fusion;

[0122] Through standardization, a probabilistic weighted input is formed, so that the proportion of semantic content of each platform in creative generation is dynamically controlled by the degree of its contextual deviation, ensuring that the influence of high-deviation platforms decreases and low-deviation platforms dominate the semantic generation structure, thereby achieving high-quality content generation that is more in line with user expectations and reduces contextual distortion.

[0123] This step dynamically adjusts the semantic contribution of each social media platform to AI-powered marketing creative generation by normalizing the fusion offset and constructing a beta function variable weighting model. This ensures that, during the multi-source semantic fusion process, interference from platforms with significant contextual deviation is effectively suppressed, while platforms with similar expressions and high user consensus are prioritized, thereby achieving contextual consistency control and optimizing generation quality. Specifically, when discussing the same target event, different social platforms (such as Weibo, Zhihu, and Douyin) often exhibit significant differences in semantic expression, emotional tendencies, and public opinion context. Without quantifying and weighting these differences, direct fusion can lead to semantic conflict, emotional imbalance, and even deviations from users' true perceptions, impacting the effectiveness and acceptance of marketing creatives. Therefore, the fusion offset values must first be normalized to obtain a unified contextual deviation index. This index reflects the degree of semantic and emotional deviation of a platform relative to other platforms.

[0124] By taking the context shift index as the input variable, the beta function variable weight model with nonlinear mapping characteristics is introduced to build a regulation mechanism for the platform's semantic contribution. The beta function can adjust the semantic contribution of the platform through the parameter according to the different values of the context shift index. and The setting of can realize the curve's "right-biased", "left-biased" or "symmetrical" shape, so that Large platforms exhibit natural inhibition in weight calculations, while Smaller platforms receive higher engagement. This adjustment method is more flexible than linear weighting and can adaptively adjust the influence of platform context, meeting the semantic fusion requirements of actual contexts where high deviations need to be suppressed and low deviations should be amplified.

[0125] Finally, by normalizing the raw weights output by the beta function, we generate probabilistic platform fusion weights that can be directly applied to the semantic generation system, ensuring that semantics have a clear proportion and dynamic controllability when participating in content generation. This step effectively prevents low-interaction platforms and platforms with high semantic bias from interfering with creative output, improving the emotional consistency, semantic logic, and user recognition of generated content across multiple platforms. This is one of the key mechanisms of this invention in handling multi-source contextual conflicts and achieving refined content generation.

[0126] This multi-source information fusion-driven AI marketing creative generation method effectively addresses the existing issues of sentiment misjudgment, semantic misalignment, and content inaccuracy caused by a lack of contextual shift identification mechanisms. This method constructs a unified keyword mapping dictionary to achieve consistent semantic extraction across platforms. Time window control and corpus topic filtering ensure data timeliness and relevance. Furthermore, word embedding and sentiment modeling are used to extract platform-specific semantic fingerprints, quantify the degree of semantic and sentiment deviation, and construct a semantic-sentiment deviation matrix by integrating cosine difference and KL divergence. Furthermore, a deviation weighting factor model is established based on user interaction density to effectively correct for sample bias caused by low-engagement platforms and improve the accuracy of deviation assessment. Finally, through normalization and a beta function nonlinear weighting mechanism, dynamic suppression and enhancement control of semantic engagement across multiple platforms is achieved. This ensures that platform semantics with strong contextual consistency and low deviation dominate creative generation, significantly improving the emotional relevance, expressive accuracy, and user identification of generated content. This significantly enhances the effectiveness and safety of brand communication, demonstrating promising practical value and engineering application prospects.

[0127] The above description is merely illustrative of certain exemplary embodiments of the present invention. It goes without saying that those skilled in the art will be able to modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and description are illustrative in nature and should not be construed as limiting the scope of protection of the claims.

Claims

1. An AI marketing creative generation method driven by multi-source information fusion, characterized by: The following steps are involved: S101. For a preset target event, extract the core concept keywords of the event, and based on the core concepts, establish a unified keyword mapping dictionary across multiple social media platforms; S102, within a preset time window, synchronously crawling high-frequency text data related to the target event from multiple social media platforms; S103, extracting the semantic center vectors and sentiment labels corresponding to the same keyword in different platform contexts to form a semantic fingerprint model that reflects the semantic expression characteristics of the platform; S104, generating a semantic-emotion deviation matrix for measuring the intensity of expression differences in the target event context across multiple platforms based on the differences between the semantic center vectors of the same keywords across multiple social media platforms and the divergence between the distributions of corresponding emotion tags; S105: performing correlation analysis on the semantic-emotional deviation matrix and the user interaction density data of the corresponding social media platform, establishing a contextual deviation weighting factor model for correcting sample deviation distortion and platform participation imbalance, and outputting a fused deviation value; S106 , normalizing the fusion offset value to generate a context offset index, and performing nonlinear dynamic adjustment on the semantic fusion weights of multiple social media platforms based on the context offset index.

2. The AI marketing creative generation method driven by multi-source information fusion according to claim 1 is characterized in that: The steps to build a cross-platform keyword dictionary include: Based on the original text description of the preset target event, the core semantic units closely related to the event theme are extracted, and the semantic integrity and redundancy removal of the preliminary keyword set are performed to generate a core keyword candidate set; For the core keyword candidate set, we use the word embedding vector space model and hot topic context retrieval mechanism to mine the extended set of keywords within the platform that are highly semantically relevant to the core keywords and have high usage frequency in the context of each platform; The keyword sets expanded from multiple social media platforms are semantically normalized to construct a unified keyword mapping dictionary between platforms, where each core keyword corresponds to a mapping entry containing multiple platform keyword variants.

3. The AI marketing creative generation method driven by multi-source information fusion according to claim 2 is characterized in that: A threshold comparison method is used to mine keyword extension sets with high semantic relevance and high usage frequency to core keywords on various social media platforms. The specific operations include the following: For each core keyword, use the pre-trained word embedding model to calculate its word vector representation in the target platform corpus; then, in the high-frequency word set collected by the corresponding platform, calculate the semantic similarity between each candidate keyword and the core keyword; A semantic relevance threshold is set. When the semantic similarity between a candidate word and a core keyword is greater than or equal to the semantic relevance threshold, the candidate word is determined to be semantically related to the core keyword. At the same time, the usage frequency of each candidate word within the target time window is counted, and a usage frequency threshold is set. Only when the usage frequency of the candidate word is greater than or equal to the usage frequency threshold will it be retained in the keyword expansion set within the platform. Keywords that meet the dual threshold conditions are integrated to construct an extended set of keywords that are both semantically relevant and representative in the current platform context.

4. The AI marketing creative generation method driven by multi-source information fusion according to claim 1 is characterized in that: The steps for simultaneously crawling high-frequency text corpora related to target events from multiple social media platforms include: Set a unified time window, which includes the start time and end time, and dynamically adjust the time window length according to the outbreak time node and propagation cycle of the target event; Based on the constructed keyword mapping dictionary, various text forms containing keywords are collected, and timestamps and platform identification tags are added to the collected samples; A unified content topic screening operation is performed on the collected original corpus data, and a text clustering method based on the topic model is used to remove redundant information, cross-topic noise or non-event-related text.

5. The AI marketing creative generation method driven by multi-source information fusion according to claim 1 is characterized in that: The steps for extracting the semantic center vector and sentiment labels corresponding to the same keyword in different platform contexts include: For the text corpus collected by each platform, the pre-trained word vector model is used to perform word vector embedding processing on the text corpus to generate each word vector representation; Based on the generated word vectors, the sentiment analysis model is used to model the sentiment labels of the platform text, identify the emotional tendencies in the text, and attach corresponding sentiment labels to each keyword; Combine the obtained word vectors with the sentiment tag data to calculate the semantic center vector of each keyword in the context of each platform; Based on the semantic center vector and sentiment tendency label, a platform semantic fingerprint model is formed.

6. The AI marketing creative generation method driven by multi-source information fusion according to claim 1 is characterized in that: The steps to generate the semantic-sentiment deviation matrix include: For each keyword, we obtain the semantic center vector of the keyword on each platform from multiple social platforms and calculate the cosine distance difference between the platforms, that is, calculate the angle difference between the semantic center vectors of the same keyword on different platforms; Based on the sentiment analysis results, extract the sentiment label distribution of the keyword on each platform and calculate the Kullback-Leibler divergence value between the sentiment label distributions of each platform; The cosine distance difference value and Kullback-Leibler divergence value are combined to obtain a comprehensive semantic-emotional deviation measure; Based on the semantic-emotional deviation measure, a semantic-emotional deviation matrix is constructed.

7. The AI marketing creative generation method driven by multi-source information fusion according to claim 6 is characterized in that: The steps for fusing the cosine distance difference value and the Kullback-Leibler divergence value to obtain a comprehensive semantic-sentiment deviation measure are as follows: Calculate the cosine distance difference of the semantic center vector of the same keyword between each platform and the Kullback-Leibler divergence value of the sentiment label distribution; Set a preset weighting coefficient to assign corresponding weights to the cosine distance difference value and the Kullback-Leibler divergence value respectively; The cosine distance difference value and the Kullback-Leibler divergence value are weighted and summed according to the weights to obtain the deviation metric value.

8. The AI marketing creative generation method driven by multi-source information fusion according to claim 1 is characterized in that: The steps of establishing a context offset weighting factor model for correcting sample offset distortion and platform participation imbalance and outputting a fused offset value include: Based on the semantic-sentiment deviation matrix of each platform, extract the user interaction density data of the platform; Correlate the semantic-emotional deviation matrix with the user interaction density data of the platform, and weight the semantic deviation according to the interaction density of different platforms; Based on the results of association analysis, a contextual deviation weighting factor model is established, which uses the platform's interaction density information as a correction parameter to dynamically adjust the semantic-emotional deviation metric; The context offset weighting factor model is used to output the weighted and corrected fusion offset value.

9. The AI marketing creative generation method driven by multi-source information fusion according to claim 1 is characterized in that: The semantic fusion weights of multiple social media platforms are nonlinearly and dynamically adjusted based on the context shift index. The specific steps are as follows: The fusion offset values of each platform are mapped to a unified scale, and the context offset index is calculated. The calculation expression of the context offset index is: , where: For the platform i Context shift index; For the platform i The fusion offset value of and are the minimum and maximum fusion offset values for all platforms; is the decimal offset; By utilizing the nonlinear shape characteristics of the beta function, the context shift index is mapped into a weighted suppression function of the platform semantic participation. The mapping expression of the context shift index is: ,in: For the platform i The initial semantic fusion weight of and is the Beta distribution parameter, which is used to control the shape of the weight curve; is the normalization constant of the standard beta function; The initial semantic fusion weights of all platforms are normalized to calculate the final actual semantic fusion weight. The calculation expression of the actual semantic fusion weight is as follows: ,in: Representation Platform i The final actual semantic fusion weight is used to control the platform i The proportion of semantic integration in the creative generation process; n The total number of social media platforms participating in the integration; represents the weighted sum of all platform fusion weights, j is a number from 1 to n The integer index represents the numbers of all social media platforms participating in semantic fusion.

Images