New media AI marketing content creation method and device, equipment and medium

Abstract

The invention relates to a new media AI marketing content creation method and device, equipment and a medium. The method comprises the steps of obtaining user demand configuration, and obtaining a creation intention vector through natural language analysis; based on the platform characteristic knowledge base, extracting a corresponding structure specification and a propagation mechanism according to the target platform, and encoding to obtain a platform characteristic vector; obtaining historical content interaction data corresponding to the audience group, and generating a user-content interaction vector by adopting collaborative filtering and a label similarity algorithm in combination with the content keyword; and calling an artificial intelligence large model, and performing content generation according to the creation intention vector, the platform feature vector and the user-content interaction vector to obtain content creation data. By adopting the method, the goal of automatic, high-quality and personalized new media marketing content creation in a multi-platform environment can be realized by means of natural language analysis, knowledge structure extraction, large model generation constraint and the like.

Description

Technical Field

[0001] The present invention belongs to the field of data processing, and in particular relates to a new media AI marketing content creation method, device, equipment and medium. Background Art

[0002] With the development of new media platform technology, new media platform content marketing technology has emerged. It usually uses user behavior data, platform characteristics and content features to generate marketing content adapted to different platforms and audiences to improve user engagement and conversion rate.

[0003] Traditional technologies rely on manual editing or simple template filling, lacking a deep understanding of user interests and platform specifications. This often results in unattractive and untargeted content. Processing multimodal content like images, text, and video requires multiple tools to work together, which is inefficient and difficult to ensure content consistency and quality. Consequently, methods have emerged to generate marketing content using existing large-scale AI models.

[0004] However, the current method of generating marketing content using existing large AI models still uses manual editing requirements to instruct AI models to produce marketing content. This method does not have a deep understanding of user needs, making it difficult to generate highly personalized content, and is not adaptable to the platform, resulting in poor performance of content on different platforms. Summary of the Invention

[0005] Based on this, it is necessary to provide a new media AI marketing content creation method, device, equipment and medium that can deeply understand user needs and adapt to the specifications of different new media platforms to address the above technical problems.

[0006] In the first aspect, this application provides a new media AI marketing content creation method, including:

[0007] Obtain user demand configuration and derive creative intent vectors through natural language analysis; user demand configuration includes target platform, audience group, content keywords, and content format preferences;

[0008] Based on the platform feature knowledge base, the corresponding structural specifications and propagation mechanisms are extracted according to the target platform and encoded to obtain the platform feature vector;

[0009] Obtain historical content interaction data corresponding to the audience group, and use collaborative filtering and tag similarity algorithms combined with content keywords to generate user-content interaction vectors; user-content interaction vectors include user preferences and user habits;

[0010] Call the artificial intelligence big model to generate content based on the creative intent vector, platform feature vector and user-content interaction vector to obtain content creation data.

[0011] In one embodiment, obtaining a user requirement configuration and obtaining a creative intent vector through natural language parsing includes:

[0012] Perform keyword recognition and syntactic dependency analysis on content keywords to obtain semantic themes and keyword meanings;

[0013] We build a style semantic vector space based on cross-platform, highly interactive corpora, and use a lightweight semantic matching model to perform style mapping on content keywords and content form preferences, thereby obtaining emotional tone and content style.

[0014] Construct creative intent vectors based on semantic themes, keyword meanings, emotional tone, and content style.

[0015] In one embodiment, historical content interaction data corresponding to an audience group is obtained, and collaborative filtering and tag similarity algorithms are used in combination with content keywords to generate a user-content interaction vector, including:

[0016] Build a user interaction matrix based on historical content interaction data;

[0017] Collaborative filtering algorithm is used to extract similar user preference vectors in the user interaction matrix;

[0018] Perform label vector similarity matching between content keywords and similar user preferences to generate a label fusion preference vector;

[0019] The user-content interaction vector is constructed based on the tag fusion preference vector and the user interaction matrix.

[0020] In one embodiment, a large artificial intelligence model is called to generate content based on a creative intent vector, a platform feature vector, and a user-content interaction vector to obtain content creation data, including:

[0021] Construct a multi-dimensional prompt word template based on the creative intention vector and the user-content interaction vector;

[0022] Determine a candidate template set based on the platform feature vector;

[0023] Semantically adjust the candidate template set according to the emotional tone and content style to obtain the structural content template;

[0024] Call the artificial intelligence big model, create content according to the multi-dimensional prompt word template under the constraints of the structural content template, and obtain content creation data; content creation data includes title, text and corresponding marketing form; marketing form includes graphic form and video form.

[0025] In one embodiment, determining a candidate template set based on a platform feature vector includes:

[0026] The platform feature vector is parsed to obtain the constraint parameters of the structural template; the constraint parameters include title character truncation, content template pool, paragraph organization, platform style and time mode;

[0027] The templates in the content structure template library are screened for platform feature adaptation using constraint parameters to obtain a set of candidate templates.

[0028] In one embodiment, a large artificial intelligence model is called to create content based on a multi-dimensional prompt word template under the constraints of a structured content template, and content creation data is obtained, including:

[0029] Based on the keyword meanings and user preferences of the multi-dimensional prompt word template, a large language model is used to generate multiple alternative titles and their corresponding attractiveness prediction values ​​and platform specification matching degrees;

[0030] Based on the structural content template and the multi-dimensional prompt word template, the language model is guided to generate the text that meets the alternative title, emotional tone and content style;

[0031] If the marketing format is text and images, use the image generation model to generate image materials based on the text, and generate a mixed text and image layout strategy based on user habits. The image materials and mixed text and image layout strategy are output together with the title and text in the form of content creation data.

[0032] If the marketing form is video, the shot script and voice copy are generated according to the main text based on the preset video strategy, and the video generation engine is used to generate shot materials and voice synthesis data according to the shot script and voice copy; the shot materials and voice synthesis data are output together with the title and main text in the form of content creation data.

[0033] In one embodiment, the method further comprises:

[0034] Obtain dissemination data based on content creation data after it is published on the target platform;

[0035] Optimize the feature extraction weight parameters for user-content interaction vector generation based on the propagation data.

[0036] Secondly, this application also provides a new media AI marketing content creation device, including:

[0037] The demand analysis module is used to obtain user demand configuration and obtain the creative intent vector through natural language analysis;

[0038] The marketing platform analysis module is used to extract the corresponding structural specifications and dissemination mechanisms according to the target platform based on the platform characteristic knowledge base, and encode them to obtain the platform feature vector;

[0039] The platform user portrait module is used to obtain historical content interaction data corresponding to the audience group and generate user-content interaction vectors using collaborative filtering and tag similarity algorithms;

[0040] The AI ​​model scheduling module is used to call the artificial intelligence large model to generate content based on the creative intent vector, platform feature vector and user-content interaction vector to obtain content creation data.

[0041] In a third aspect, the present application also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any of the above-mentioned new media AI marketing content creation methods when executing the computer program.

[0042] In a fourth aspect, the present application also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the steps of any of the above-mentioned new media AI marketing content creation methods are implemented.

[0043] The aforementioned new media AI marketing content creation method, apparatus, device, and medium establish a content generation logic framework through the collaborative drive of creative intent vectors, platform feature vectors, and user-content interaction vectors. This avoids the traditional AI content generation methods' reliance on single keyword prompts or generic templates, achieving more precise, tailored, and multi-dimensionally controlled content creation. User demand configuration items include platform, audience, keyword, and format preferences. The content generation logic can be dynamically adjusted based on different configurations, making it suitable for different platforms, multiple audience types, and diverse marketing scenarios, thus possessing excellent versatility and adaptability. Incorporating the audience's real-world interactive behavior into the vector generation process reflects a user-centric generation approach, enhancing conversion capabilities and marketing value in practical applications. By using structured semantic vectors as control conditions for large models, this method avoids the style drift and logical instability of general large models in marketing content generation, effectively scheduling and adapting the general model's capabilities to specific scenarios, and improving the controllability and consistency of content generation quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0045] Figure 1 This is a flowchart of the new media AI marketing content creation method of the present invention;

[0046] Figure 2Schematic diagram of the step-by-step process of step S104;

[0047] Figure 3 Schematic diagram of the step-by-step process of step S204;

[0048] Figure 4 This is a structural diagram of the new media AI marketing content creation device of the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0050] In one embodiment, Figure 1 As shown, a new media AI marketing content creation method is provided. This embodiment uses the method applied to a terminal as an example for illustration. It is understandable that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0051] S101. Obtain user demand configuration and obtain a creative intent vector through natural language analysis; the user demand configuration includes target platform, audience group, content keywords, and content format preferences.

[0052] In principle, user demand configuration refers to the personalized needs of users at the stage of initiating content creation tasks, including but not limited to target platforms, audience groups, content keywords, and content form preferences. Optionally, the creation goal selection guidance can be clearly given through the interactive interface, or the user's natural language description can be received. Specifically, keyword recognition and syntactic dependency analysis techniques are used to perform structured extraction of input content keywords, identify subject words and their hyponymous and hyponymous semantic structures; further, with the help of a lightweight semantic matching network, the style implied by the content form preference is determined, and the analysis results are encoded into a high-dimensional creation intention vector, covering keyword semantics, style intentions, emotional tone, and target user expectations.

[0053] S102. Based on the platform characteristic knowledge base, extract the corresponding structural specifications and propagation mechanisms according to the target platform, and encode them to obtain the platform characteristic vector.

[0054] In order to adapt the generated content to the specifications and content preferences of different new media platforms, the structural specifications and dissemination mechanisms corresponding to the target platform are extracted from the platform feature knowledge base and encoded as a platform feature vector. Schematically, the platform feature knowledge base is a pre-built structured data set that can be generated through the platform's open API, developer documentation, or statistical analysis based on crawled sample content. It records the content presentation, title length truncation rules, recommendation algorithm preferences, and style preferences of mainstream new media platforms. Among them, the content presentation includes supported video length, cover image ratio, paragraph number limit, etc. Specifically, structural specifications are mainly used to constrain content format, while dissemination mechanisms more affect the expression focus of the content generation stage, such as reversal endings, problem orientation, and argumentation process. Optionally, the structural specifications and dissemination mechanisms are standardized through rule annotation and statistical modeling, and finally encoded into a multi-dimensional vector.

[0055] S103. Obtain historical content interaction data corresponding to the audience group, and use collaborative filtering and tag similarity algorithms combined with content keywords to generate a user-content interaction vector; the user-content interaction vector includes user preferences and user habits.

[0056] Indicatively, the user portrait data is used to obtain the audience's historical interactive behavior data on the platform, including behavioral sequences such as likes, comments, reposts, and dwell time, and the interest factor matrix is ​​established in combination with the user's social tags. Furthermore, in combination with the extracted content keywords, the collaborative filtering algorithm and the tag semantic similarity algorithm are used to conduct correlation analysis on the historical content to explore the audience's preference patterns for specific themes, styles or media forms. For example, the target audience is the "new mother" group, and the keyword is "baby food". It can be found from the historical content that this group has a high interaction rate for content such as "graphic step-by-step instructions and real baby feedback", and has a significantly improved trust in content such as "video explanations and professional doctor endorsements". Such behavioral preferences are encoded as user-content interaction vectors, which include user preferences for content structure, emotional tendencies, etc., as well as user habits such as browsing time periods and interaction rhythms.

[0057] S104: Call the artificial intelligence big model to generate content based on the creative intent vector, platform feature vector and user-content interaction vector to obtain content creation data.

[0058] Schematically, an artificial intelligence (AI) model is called to generate content to obtain the final content creation data. Specifically, based on the structural template library determined by the platform feature vector, a set of candidate templates that match the current platform specifications are screened out, and then based on the style and tone information in the creative intent vector, the candidate templates are semantically fine-tuned to adjust the title structure, copy paragraph length or tone expression, and finally form a structural content template. Furthermore, the structural content template is used as the structural skeleton for content generation to guide AI language models such as GPT and ERNIE to perform style control and logical arrangement when filling in content. Specifically, during the creation process, the model is guided to input subject words, keyword meanings and emotional tone information into the text generation module, and generate multiple sets of title options based on user preference data. The optimal title is selected through the attractiveness scoring model and platform specification matching prediction. Under the constraints of the template, the text content is filled in by paragraphs while maintaining language coherence and attractiveness.

[0059] For example, the target platform is Xiaohongshu, and the user demand is to generate a graphic and text content promoting healthy yogurt for young white-collar workers. The platform feature vector will be limited to 6 pictures and paragraph-style pictures and texts, including a preference for real experience tone and product recommendations. Based on this, the content format of the title, product scenario story, user feedback, and mixed text and pictures is generated, and the AI ​​model is called to generate copy and visual materials that match the style, completing the one-click generation of graphic and text content.

[0060] The aforementioned new media AI marketing content creation method semantically analyzes user input for target platforms, audience groups, content keywords, and content format preferences. This transforms previously unstructured user needs into creative intent vectors with clear semantic dimensions. This helps subsequent modules accurately utilize corresponding rules and model resources, significantly improving AI content generation's ability to understand complex and diverse user needs, thereby enhancing the targeted and controllable nature of content generation. By extracting structural norms and dissemination mechanisms from a platform-specific knowledge base and encoding them into platform feature vectors, the generated content is highly aligned with platform-specific requirements in terms of structure, layout, language style, paragraph length, and release time, avoiding platform release restrictions or content format mismatches, thereby improving the success rate and dissemination efficiency of content releases. By analyzing historical content interaction data from audiences and combining collaborative filtering with tag similarity algorithms to introduce the two dimensions of user preferences and user habits, a user-content interaction vector is generated. This vector understands the preferences of the target group in their actual content consumption behavior, enhances the content generation module's adaptability to the audience, and effectively improves the user relevance and acceptance of the generated content. Based on three types of vectors of creative intent, platform specifications and user preferences as input conditions, guiding the artificial intelligence large model to generate content helps to achieve multi-dimensional constraints on content style, structure and communication characteristics while ensuring semantic coherence and language naturalness, thereby improving the overall coordination of the content and making the generated results more in line with marketing goals and user expectations.

[0061] In one embodiment, obtaining a user requirement configuration and obtaining a creative intent vector through natural language parsing includes:

[0062] S11. Perform keyword recognition and syntactic dependency analysis on content keywords to obtain semantic themes and keyword meanings.

[0063] In an illustrative manner, keyword recognition and syntactic dependency analysis are performed on the content keywords provided by the user to clarify their potential semantic themes and specific meanings. Among them, content keywords usually include core noun phrases in the user input, which often present hierarchical, modifier and modified relationships in natural language, and need to be structurally parsed through syntactic dependency relationships. Specifically, based on the part-of-speech tagging and dependency relationships of new media corpus, combined with named entity recognition, the keywords are divided into semantic roles. For example, in "Recommending portable air fryers for students", "students" are identified as the target group, "portable" is an adjective modifier, and "air fryer" is the main core entity. "Portable air fryer" is further extracted as a complete semantic unit, and "students" are used as semantic audience labels. Through this process, the core expression object and its semantic scope can be clearly identified from the user input, forming a structured semantic theme and keyword meaning.

[0064] S12. Construct a style semantic vector space based on cross-platform highly interactive corpus, and use a lightweight semantic matching model to perform style mapping on content keywords and content form preferences to obtain emotional tone and content style.

[0065] Furthermore, we identify the user's desired expression style and content tone to ensure that the subsequently generated content not only matches the theme but also aligns its presentation with the communication context. For example, based on a massive cross-platform corpus of highly interactive content, we pre-construct a style semantic vector space. This includes multiple mainstream style dimensions, such as knowledge popularization, lightheartedness and fun, enthusiastic recommendation, and authentic experience. Each style is embedded in a semantic space, combining the distribution of style words, sentiment polarity, and sentence structure features extracted from a large number of typical text samples.

[0066] For example, the content keywords and form preferences in the user input are taken as dual inputs, and their projections in the style semantic space are calculated through lightweight semantic matching models such as a small style classification network fine-tuned based on BERT (Bidirectional Encoder Representations from Transformers, a natural language processing model based on the Transformer architecture) or RoBERTa (improved BERT), to obtain the emotional tonality and content style. The emotional tonality is used to determine the emotional tone of the language expression; the content style is used to constrain the structural framework and tone of the generated text.

[0067] S13. Construct a creative intent vector based on semantic themes, keyword meanings, emotional tone, and content style.

[0068] Schematically, a 512-dimensional or 768-dimensional floating-point vector dense embedding structure is used to form a topic representation sub-vector with semantic themes and keywords. The tonality and expression style in the style semantic subspace are encoded as emotional channel and structural channel embeddings respectively. The two are then fused through an attention weighted mechanism to form a semantic vector of the overall creative intention. It has the ability to recognize content targets and has a style guidance function, and can be directly used to generate the input condition control of the model.

[0069] The above method, which captures user demand configuration and generates a creative intent vector, involves structured semantic deconstruction and stylistic intent modeling of user language input. Through multi-level processing, including syntactic analysis, style projection, and semantic fusion, this mechanism preserves the core of user expression while converting it into a high-dimensional semantic representation suitable for generative models. This lays the foundation for highly consistent and highly disseminable AI-generated content.

[0070] In one embodiment, historical content interaction data corresponding to an audience group is obtained, and collaborative filtering and tag similarity algorithms are used in combination with content keywords to generate a user-content interaction vector, including:

[0071] S21. Construct a user interaction matrix based on historical content interaction data.

[0072] In principle, a user interaction matrix is ​​constructed based on content interaction data recorded on existing platforms. Interaction data includes, but is not limited to, behavior types such as likes, comments, forwarding, favorites, and length of stay, which can be recorded in a structured manner by the new media platform. For example, if a user likes a certain type of short video three times in a row and stays on it for longer than the average time, he or she may be given a high-weighted score. By defining a set of interaction scoring rules, the interaction relationship between all audience groups, that is, potential content recipients, and the content they have been exposed to on the target platform is encoded into a two-dimensional sparse matrix, where rows represent user identifiers, columns represent content identifiers, and the elements in the matrix are the user's interaction scores for the content.

[0073] S22. Use collaborative filtering algorithm to extract similar user preference vectors in the user interaction matrix.

[0074] Schematically, the user interaction matrix is ​​processed using a collaborative filtering algorithm to extract user preference vectors that are similar to the behavior patterns of the current audience group. Among them, collaborative filtering, as a recommendation mechanism based on the similarity hypothesis, can discover groups with similar preferences only through user behavior statistics without clear labels or content understanding. Exemplarily, a user-user collaborative filtering strategy can be adopted to identify a set of highly similar users that are closest to the current audience characteristics by calculating the cosine similarity or Pearson correlation coefficient between each user in the user interaction matrix. Furthermore, the content topics that similar users have frequently interacted with in history are aggregated and weighted to form a preliminary vector representing potential preferences, namely the similar user preference vector.

[0075] S23: Perform label vector similarity matching on the content keywords and similar user preference quantities to generate a label fusion preference vector.

[0076] Indicatively, in order to avoid problems such as cold start or semantic misalignment caused by relying solely on historical behavior, content keywords are semantically fused with similar user preference vectors. Specifically, a content label vector space is constructed, which can be composed of classification labels attached to the platform's historical content when it is uploaded and topic labels generated by automatic annotation. Furthermore, each label is represented by a dense vector in the vector space. After the content keywords are converted into semantic vectors through the word vector model, cosine similarity is calculated with the label vectors involved in similar user preferences to evaluate the matching degree between the keywords and each type of label. The label distribution is adjusted and reweighted according to the matching score to obtain a label fusion preference vector that integrates the current creative semantics.

[0077] S24. Construct a user-content interaction vector based on the tag fusion preference vector and the user interaction matrix.

[0078] The tag fusion preference vector and the user interaction matrix are combined to construct the final user-content interaction vector. This vector not only captures historical user preference trends but also incorporates the potential semantic focus points of the current creative theme that may attract attention within the target audience. For example, a weighted concatenation or attention mechanism is used to fuse and encode these two types of features. The resulting user-content interaction vector can be used as a user-level personalized adjustment variable in the subsequent AI model generation process. This influences vocabulary selection, tone of voice, and even paragraph organization during content generation, significantly increasing the content's appeal to the target audience.

[0079] For example, if the current creative goal is to recommend air fryers suitable for home use to women over 30 years old, the historical content interaction records of female users over 30 years old on the platform will be extracted from the audience group, and through collaborative filtering, it will be found that their preferences are concentrated on light cooking and energy-saving and health-related topics; then, combined with the keyword air fryer, it is matched with tags such as kitchen appliances, healthy cooking, and family rhythm in the tag library, and it is found that healthy cooking is highly correlated with kitchen appliances. Based on this, the weight of the corresponding tag is enhanced to generate a fusion preference vector; the user-content interaction vector finally formed by the fusion will be used to guide the generation of a recommendation content that emphasizes health and efficiency, and has a language style that is close to maturity and practicality.

[0080] In one embodiment, Figure 2 As shown, the artificial intelligence model is called to generate content based on the creative intent vector, platform feature vector, and user-content interaction vector to obtain content creation data, including:

[0081] S201: Construct a multi-dimensional prompt word template according to the creative intention vector and the user-content interaction vector.

[0082] Schematically, the creative intent vector encompasses abstract semantic information such as content theme, keyword semantics, emotional tone, and stylistic intent. The user-content interaction vector reflects the target audience's preferences for content details such as tone of voice, rhythm of expression, and information density. Specifically, a multi-layered prompt framework is constructed to maximize the direction of large-scale model generation. The prompt template is divided into two dimensions: structural prompts and semantic prompts. Structural prompts clarify the organization of content elements. For example, the structural prompt is "title + introduction + three key points + closing call to action." Semantic prompts describe the target expression through natural language. Optionally, predefined template slots and dynamic filling methods can be used to construct the final prompt text through search, interpolation, or a template language framework.

[0083] S202: Determine a candidate template set according to the platform feature vector.

[0084] To ensure that the structure of the content generated conforms to the technical specifications and communication characteristics of the target platform, the preset template library needs to be screened according to the platform feature vector to obtain a set of candidate templates. The platform feature vector is encoded by the target platform's structural specifications such as word limit, paragraph structure, whether illustrations / videos are supported, as well as communication mechanisms such as algorithm recommendation preferences, hot word distribution, and user interaction tendencies. For example, the templates of graphic and text platforms usually include a short title, horizontal comparison points, and a summary recommendation structure, while video platforms may prefer a content format that starts with emotional stimulation, cuts into scenes, and presents step-by-step content, and ends with a spoken announcement.

[0085] S203: semantically adjust the candidate template set according to the emotional tone and content style to obtain a structural content template.

[0086] In principle, emotional tone may include expression tones such as positive motivation, rational analysis, and light-hearted ridicule, while content style covers language style, narrative method, and cultural context adaptation. This type of information is mapped to the semantic adjustment rule library, and style adaptation at the semantic level is achieved by fine-tuning the template's opening words, conjunctions, information arrangement, etc.

[0087] S204. Call the artificial intelligence big model, create content according to the multi-dimensional prompt word template under the constraints of the structural content template, and obtain content creation data; the content creation data includes the title, text and corresponding marketing form; the marketing form includes graphic form and video form.

[0088] Schematically, the artificial intelligence big model is called to generate content, and the constructed structural content template and multi-dimensional prompt word template are used as context input. The big model generates complete content creation data based on these inputs. Specifically, a structure-guided content generation strategy is adopted when the model is generated, that is, the structural template is used as a hard constraint and the prompt word is used as a soft guide, so that the content generated by the model is controllable in terms of logical structure, personalized at the semantic level, and has the target audience's acceptance tendency. Optionally, the generated content creation data includes the title, text and corresponding marketing form, and the marketing form may include graphic content or video content. Optionally, the content form can be automatically selected according to the platform characteristics and user content preferences. On video-based platforms, video scripts and supporting instructions are generated first, and graphic content that conforms to paragraph specifications is generated in graphic-based scenes.

[0089] The working mechanism of the above method fully mobilizes the three vector modules of intent understanding, platform constraints and user interests. Through the template mechanism, the information constraints are uniformly converted into structural and semantic information that can be processed by large models. Then, through the structure-guided generation strategy, it effectively solves the common problems in the existing AI content generation process, such as lack of structural sense, content style not suitable for media platforms and inconsistent language tone, thereby providing a highly adaptable, efficient and highly disseminable solution for new media content creation.

[0090] In one embodiment, determining a candidate template set based on a platform feature vector includes:

[0091] S31. Analyze the platform feature vector to obtain constraint parameters of the structural template; the constraint parameters include title character truncation, content template pool, paragraph organization, platform style and time mode.

[0092] Schematically, the platform feature vector is parsed to extract and generate the constraint parameters required for the structural template. The constraint parameters mainly include title character truncation, content template pool, paragraph organization, platform style and time mode. Among them, title character truncation is used to limit the maximum length or recommended length range of title generation to ensure the complete presentation of content on the platform cover or recommended position; the content template pool is used to limit the types of templates available for screening, including whether it supports the combination of pictures and texts, whether it contains interactive components such as voting, and whether it contains script instructions; the paragraph organization method specifies the functional layout of each part of the content structure; the platform style includes language style preferences, tone tendencies, etc., which are used to guide style matching adjustments; the time mode includes the release time, and for video platforms, specifies the duration distribution and transition rhythm of the content.

[0093] S32. Using constraint parameters, perform platform feature adaptation screening on templates in the content structure template library to obtain a set of candidate templates.

[0094] Schematically, the content structure template library is a collection of templates that include a variety of content presentation forms, paragraph structures, and logical organization methods, which can be managed by labeling according to dimensions such as industry, platform, and expression intention. The core content of the template can include the organization of the structural skeleton, as well as some semantic fragments such as predefined emotional starter sentences, transition connectors, and summary sentences. Specifically, the screening process adopts a parameter matching mechanism to compare the structural indicators in the template tags one by one to see if they meet the constraint parameters of the current platform, and screen out incompatible templates. Furthermore, secondary filtering can be performed based on the implicit rules in the platform dissemination mechanism, and the template matching degree can be further improved based on the platform style tags. Optionally, a template matching scoring mechanism is introduced to perform weighted sorting of each template according to the content effect in the historical release data, so that the candidate set is not only structurally compliant, but also has actual dissemination advantages.

[0095] The above method, through the structural template screening mechanism driven by the platform feature vector, can significantly improve the publishing adaptability and user acceptance of AI-generated content, ensuring that the content not only conforms to user intentions at the language and semantic levels, but also strictly meets the platform operation and dissemination requirements at the form and structure levels.

[0096] In one embodiment, Figure 3 As shown, the artificial intelligence model is called to create content according to the multi-dimensional prompt word template under the constraints of the structured content template to obtain content creation data, including:

[0097] S301. Based on the keyword meanings and user preferences of the multi-dimensional prompt word template, a large language model is used to generate multiple candidate titles and corresponding attractiveness prediction values ​​and platform specification matching degrees.

[0098] Using a multi-dimensional prompt word template as input, AI is used to create titles, combining the semantic meaning of keywords with user preferences. Schematically, during the title generation process, the AI ​​outputs multiple candidate titles based on the prompt word. Each title not only corresponds to a different expression style and information angle, but also calculates the title's predicted attractiveness and platform specification match. The predicted attractiveness is derived from an evaluation network constructed using audience profiles and historically highly interactive content models to assess the title's click potential. Platform specification match, on the other hand, evaluates whether the title meets the platform's sensitive word rules and presentation requirements.

[0099] S302: Based on the structure content template and the multi-dimensional prompt word template, the language model is guided to generate a text that meets the candidate titles, emotional tone and content style.

[0100] The language generation model is then used to generate the main content based on the structural content template and the prompt word template. The structural content template provides constraints such as paragraph distribution, semantic organization order, and emotional progression nodes. The prompt word template provides the specific language elements required to complete the paragraphs. Specifically, guiding prompts are injected into each paragraph according to the order set in the template, and the language style is dynamically adjusted during the generation process to match the emotional tone and content style, making the overall content more personalized, semantically complete, and persuasive.

[0101] S303. If the marketing format is text and image format, use the image generation model to generate image materials based on the text, and generate a mixed text and image layout strategy based on user habits; the image materials and the mixed text and image layout strategy are output together with the title and text in the form of content creation data.

[0102] Schematically, when the target content is in the form of text and images, an image generation model such as Stable Diffusion is called, and the key scene descriptions, product feature descriptions, or user behavior guidance in the main text are used as semantic inputs for image generation to generate supporting image materials. Specifically, the image generation process includes extracting candidate image description sentences from the main text, using semantic enhancement to complete the language-image prompt mapping, and controlling the tonality of the generated image through style constraints to be consistent with the platform. Optionally, the user-content interaction vector is combined to analyze the user's past text and image reading habits, generate a mixed text and image layout strategy, and output unified text and image mixed content creation data.

[0103] S304. If the marketing form is video, generate shot scripts and voice copy according to the main text based on the preset video strategy, and use the video generation engine to generate shot materials and voice synthesis data according to the shot scripts and voice copy; the shot materials and voice synthesis data are output together with the title and main text in the form of content creation data.

[0104] When the target content is in the form of a video type, the main text content is split into lenses and a voice script is written according to the preset video strategy. Schematically, the lens strategy usually includes the number of lenses, duration control, transition rhythm, visual guide sequence, etc. Based on the preset video content preference model of the media platform, the main text is broken down into multiple lens nodes, each node corresponds to a visual narrative target, and a lens script is generated, including the scene description, subject behavior and visual focus of each lens. Furthermore, a matching voice copy is generated to ensure that the video dubbing is consistent with the text content and the tone is natural. The generated lens script and voice copy are input into the video generation engine to generate lens materials and voice synthesis data. The video generation engine may include TTS (Text To Speech, from text to speech) voice synthesis, video material scheduling or automatic animation engine modules, generate short video clips based on semantics, and support auxiliary functions such as automatic insertion of subtitles, lens transitions and background music configuration.

[0105] The above method calls on a large artificial intelligence model and combines multi-dimensional prompt words and structured content templates to develop a content generation strategy. It can efficiently and accurately transform abstract creative intentions and platform specifications into publishable structured content results. It not only realizes natural generation at the language level, but also expands to adaptive creation in multimodal forms such as images and videos, thereby comprehensively improving the dissemination and expressiveness of new media marketing content in multi-platform and multi-population environments.

[0106] In one embodiment, the method further comprises:

[0107] S41. Obtain dissemination data after the content creation data is published on the target platform.

[0108] Illustratively, after content creation data is published, data on the content's spread on the target platform is obtained through interface calls or platform monitoring mechanisms. This data includes, but is not limited to, key behavioral indicators such as number of views, forwarding volume, number of likes, comment popularity, dwell time, click-through rate, and conversion rate. Exemplarily, the data interface provided by the target platform is called to periodically pull metrics, which are then stored based on the content's unique identifier and time window to reflect the actual audience response.

[0109] S42. Optimize feature extraction weight parameters generated by user-content interaction vectors based on the propagation data.

[0110] In principle, using communication data as the target signal, the weight distribution of various feature dimensions used in interaction vector generation can be reversely optimized through regression models, reinforcement learning, or weighted supervised learning. For example, if it is found that audiences are more likely to engage deeply with content with titles containing words such as "trend," "2025," and "quickly implement," the weight of prompt words semantically related to these words can be increased; or if a certain structural template does not spread effectively among users on a certain platform, the priority of the paragraph organization method corresponding to this structural template in subsequent content generation can be lowered.

[0111] Alternatively, we can use a weight learning network based on an attention mechanism, or embed an adjustment mechanism through a back-propagation model, to fine-tune the generation process of the user-content interaction vector, making it more sensitive to the characteristics of content with historically successful dissemination. Furthermore, by clustering user groups, we can make the weight adjustment strategy differentiated for different user clusters, achieving personalized optimization of clustering driven by dissemination data.

[0112] The aforementioned method, combining dissemination data acquisition and interaction vector optimization modules, not only establishes a direct link between content creation and actual performance but also forms a feedback mechanism, enabling the entire content creation process to be self-learning and evolving. By dynamically integrating dissemination feedback features with user interest modeling, it effectively overcomes issues such as cold starts and sample drift, continuously improving content matching and marketing conversion rates, and thus empowering the system with greater intelligent adaptability and commercial effectiveness.

[0113] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0114] Based on the same inventive concept, the embodiment of the present application also provides a new media AI marketing content creation device for implementing the new media AI marketing content creation method involved above. The implementation solution provided by the device is similar to the implementation solution described in the above method, so the specific limitations in the one or more new media AI marketing content creation device embodiments provided below can be found in the above limitations on the new media AI marketing content creation method, and will not be repeated here.

[0115] In an exemplary embodiment, Figure 4 As shown, a new media AI marketing content creation device is provided, including:

[0116] Demand analysis module 401, used to obtain user demand configuration and obtain a creative intent vector through natural language analysis;

[0117] The marketing platform analysis module 402 is used to extract the corresponding structural specifications and propagation mechanisms according to the target platform based on the platform characteristic knowledge base, and encode them to obtain the platform feature vector;

[0118] The platform user portrait module 403 is used to obtain historical content interaction data corresponding to the audience group and generate user-content interaction vectors using collaborative filtering and tag similarity algorithms;

[0119] The AI ​​model scheduling module 404 is used to call the artificial intelligence large model to generate content according to the creative intent vector, the platform feature vector and the user-content interaction vector to obtain content creation data.

[0120] In one embodiment, it further includes:

[0121] The language processing module is used to perform keyword recognition and syntactic dependency analysis on content keywords to obtain semantic themes and keyword meanings;

[0122] The semantic understanding module is used to construct a style semantic vector space based on cross-platform highly interactive corpus, and use a lightweight semantic matching model to perform style mapping on content keywords and content form preferences to obtain emotional tone and content style;

[0123] The demand analysis module 401 is also used to construct a creative intention vector based on semantic themes, keyword meanings, emotional tone and content style.

[0124] In one embodiment, it further includes:

[0125] Data analysis module, used to build a user interaction matrix based on historical content interaction data;

[0126] A filter operator module is used to extract similar user preference vectors from the user interaction matrix using a collaborative filtering algorithm;

[0127] Similarity algorithm module, used to perform label vector similarity matching between content keywords and similar user preferences to generate a label fusion preference vector;

[0128] The platform user portrait module 403 is further configured to construct a user-content interaction vector based on the tag fusion preference vector and the user interaction matrix.

[0129] In one embodiment, it further includes:

[0130] A guidance module, used to construct a multi-dimensional prompt word template based on the creative intention vector and the user-content interaction vector;

[0131] A constraint module, used to determine a candidate template set based on the platform feature vector;

[0132] An adjustment module is used to semantically adjust the candidate template set according to the emotional tone and content style to obtain a structural content template;

[0133] The AI ​​model scheduling module 404 is also used to call the artificial intelligence large model, create content according to the multi-dimensional prompt word template under the constraints of the structural content template, and obtain content creation data.

[0134] In one embodiment, the constraint module is further used to parse the platform feature vector to obtain constraint parameters of the structure template;

[0135] The constraint module is also used to use constraint parameters to perform platform feature adaptation screening on templates in the content structure template library to obtain a set of candidate templates.

[0136] In one embodiment, it further includes:

[0137] The title module is used to generate multiple candidate titles based on the keyword meanings and user preferences of the multi-dimensional prompt word template using a large language model, along with corresponding attractiveness prediction values ​​and platform specification matching degrees;

[0138] The main text module is used to guide the language model to generate the main text that meets the alternative titles, emotional tone and content style based on the structural content template and the multi-dimensional prompt word template;

[0139] The image and text module is used to generate image materials based on the main text when the marketing format is image and text, and to generate an image and text mixing strategy based on user habits;

[0140] The video module is used to generate shot scripts and voice copy based on the main text based on the preset video strategy if the marketing form is video, and use the video generation engine to generate shot materials and voice synthesis data based on the shot scripts and voice copy.

[0141] In one embodiment, it further includes:

[0142] Feedback module, used to obtain dissemination data after the content creation data is published on the target platform;

[0143] An optimization module is used to optimize the feature extraction weight parameters generated by the user-content interaction vector based on the propagation data.

[0144] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0145] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0146] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely illustrative, wherein the components described as separate parts may or may not be physically separated, and the parts displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the disclosed solution. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0147] The above-described embodiments merely represent several implementation methods of the embodiments of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the concept of the embodiments of the present application, and these modifications and improvements fall within the scope of protection of the embodiments of the present application.

Claims

1. A new media AI marketing content creation method, characterized by: The method comprises: Obtaining user demand configuration and obtaining a creative intent vector through natural language analysis; the user demand configuration includes target platform, audience group, content keywords, and content format preferences; Based on the platform characteristic knowledge base, extract the corresponding structural specifications and propagation mechanisms according to the target platform, and encode them to obtain the platform feature vector; Obtaining historical content interaction data corresponding to the audience group, and using collaborative filtering and tag similarity algorithms in combination with the content keywords to generate a user-content interaction vector; the user-content interaction vector includes user preferences and user habits; The artificial intelligence big model is called to generate content according to the creation intention vector, the platform feature vector and the user-content interaction vector to obtain content creation data.

2. The method according to claim 1, characterized in that The step of obtaining user requirement configuration and obtaining a creative intent vector through natural language analysis includes: Perform keyword recognition and syntactic dependency analysis on the content keywords to obtain semantic themes and keyword meanings; Constructing a style semantic vector space based on cross-platform highly interactive corpus, and using a lightweight semantic matching model to perform style mapping on the content keywords and the content form preferences to obtain the emotional tone and content style; The creative intention vector is constructed according to the semantic theme, the keyword meaning, the emotional tone and the content style.

3. The method according to claim 1, characterized in that The acquiring of historical content interaction data corresponding to the audience group and generating a user-content interaction vector using collaborative filtering and tag similarity algorithms in combination with the content keywords includes: constructing a user interaction matrix based on the historical content interaction data; Using collaborative filtering algorithm to extract similar user preference vectors in the user interaction matrix; Perform tag vector similarity matching on the content keywords and the similar user preferences to generate a tag fusion preference vector; A user-content interaction vector is constructed according to the tag fusion preference vector and the user interaction matrix.

4. The method according to claim 2, characterized in that The calling of the artificial intelligence big model to generate content according to the creation intention vector, the platform feature vector, and the user-content interaction vector to obtain content creation data includes: Constructing a multi-dimensional prompt word template according to the creative intention vector and the user-content interaction vector; Determining a candidate template set according to the platform feature vector; semantically adjusting the candidate template set according to the emotional tonality and the content style to obtain a structural content template; The artificial intelligence big model is called to create content according to the multi-dimensional prompt word template under the constraints of the structural content template to obtain the content creation data; the content creation data includes the title, text and corresponding marketing form; the marketing form includes graphic and text form and video form.

5. The method according to claim 4, characterized in that The determining of a candidate template set according to the platform feature vector includes: Parsing the platform feature vector to obtain constraint parameters of the structural template; the constraint parameters include title character truncation, content template pool, paragraph organization, platform style and time mode; The templates in the content structure template library are screened for platform feature adaptation using the constraint parameters to obtain the candidate template set.

6. The method according to claim 4, characterized in that The calling of the artificial intelligence big model and performing content creation according to the multi-dimensional prompt word template under the constraints of the structured content template to obtain the content creation data includes: Based on the keyword meanings of the multi-dimensional prompt word template and the user preferences, a language model is used to generate multiple candidate titles and corresponding attractiveness prediction values ​​and platform specification matching degrees; Based on the structure content template and the multi-dimensional prompt word template, the language model is guided to generate a text that conforms to the candidate title, the emotional tone and the content style; If the marketing format is text and image format, an image generation model is used to generate image materials based on the text, and a text and image mixing strategy is generated based on the user's habits; the image materials and the text and image mixing strategy are output together with the title and the text in the form of the content creation data; If the marketing form is video form, a shot script and voice copy are generated according to the text based on a preset video strategy, and a video generation engine is used to generate shot materials and voice synthesis data according to the shot script and the voice copy; the shot materials and the voice synthesis data are output together with the title and the text in the form of the content creation data.

7. The method according to any one of claims 1 to 6, characterized in that The method further comprises: Obtaining dissemination data after the content creation data is published on the target platform; The feature extraction weight parameters generated by the user-content interaction vector are optimized based on the propagation data.

8. A new media AI marketing content creation device, characterized by: The device comprises: The demand analysis module is used to obtain user demand configuration and obtain the creative intent vector through natural language analysis; The marketing platform analysis module is used to extract the corresponding structural specifications and dissemination mechanisms according to the target platform based on the platform characteristic knowledge base, and encode them to obtain the platform feature vector; The platform user portrait module is used to obtain historical content interaction data corresponding to the audience group and generate user-content interaction vectors using collaborative filtering and tag similarity fusion; The AI ​​model scheduling module is used to call the artificial intelligence large model, generate content according to the creative intention vector, the platform feature vector and the user-content interaction vector, and obtain content creation data.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images