Creative content generation method and apparatus, device, and storage medium

Abstract

The application relates to the technical field of natural language processing, can be applied to the creation content generation scene of financial technology and medical health, and discloses a creation content generation method, device, equipment and storage medium. A user initial creation requirement is received, and a structured core creation demand is extracted based on a preconfigured understanding instruction; a content first draft, a guiding question and a multi-modal prompt word are synchronously generated based on a preconfigured generation instruction, the content first draft and the guiding question are pushed to the user; user feedback information is received, knowledge demand is analyzed based on a preconfigured analysis instruction and an optimization instruction, and matched professional information is called from a preset knowledge base, a revised text is generated in combination with the feedback information and the content first draft; multi-modal content with semantic alignment is generated based on the multi-modal prompt word and the revised text, and the revised text and the multi-modal content are integrated and pushed to the user. The application realizes high-quality, personalized and automatic generation of creation content through preconfigured exclusive instructions in the whole process.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology and can be applied to creative content generation scenarios in fintech and healthcare. Specifically, it relates to a creative content generation method, apparatus, device, and storage medium. Background Technology

[0002] AI-assisted content creation tools represent a significant application of AI technology in content creation. Currently, mainstream AI-assisted content creation tools (such as text generation tools, graphic design tools, and video generation tools) all employ a user-driven, one-way generation model. This means that the user provides the tool with complete and precise creative prompts, and the tool generates corresponding content based on these prompts.

[0003] In practical applications, user-generated prompts have inherent flaws: On the one hand, ordinary users lack professional prompt engineering capabilities, making it difficult to clearly, completely, and systematically express key information such as creative intent, target audience, style requirements, and professional dimensions. This often results in vague prompts, missing elements, and logical inconsistencies, directly leading to significant deviations between generated content and actual needs. On the other hand, different users have significantly different language habits, expression abilities, and professional backgrounds, resulting in a wide variety of prompts with inconsistent standards. Tools struggle to provide unified analysis and accurate responses, and even repeated generation often fails to achieve the desired effect, greatly increasing users' trial-and-error costs and creation time. For example, in the fintech field, when writing promotional copy for wealth management products, users need to clearly define all professional information such as product yield, risk level, and purchase rules at once; omissions result in content that does not meet compliance requirements. Similarly, in the healthcare field, when creating popular science copy, details such as disease classification, medication guidelines, and treatment guidelines must be clearly stated at once; otherwise, professional biases are likely to occur.

[0004] Furthermore, existing technologies for generating multimodal content such as text, images, and videos operate independently, lacking collaborative logic that unifies constraints through pre-configured prompts from the same source. This often results in semantic inconsistencies between AI-generated image descriptions and text content, and between video scripts and visual content, leading to poor coherence and matching across modalities. For example, in the fintech field, when creating promotional materials for fund products, existing technologies may generate text content that mentions data, but the accompanying illustrations may not match the corresponding data visualization charts. Similarly, in the healthcare field, when creating disease education videos, existing technologies may generate scripts that mention professional treatment procedures, but the accompanying visuals may consist of other treatment-related materials, resulting in a mismatch of professional information and rendering the content useless.

[0005] Furthermore, existing technologies cannot accurately analyze users' professional knowledge needs when generating content, nor can they retrieve matching professional information from pre-set knowledge bases. This results in insufficient professionalism and credibility of the generated content, making it difficult to meet the creative needs of professional fields. For example, creating industry analysis reports in the fintech field requires access to the latest market data, policy documents, product net asset values, and other information; or, creating interpretations of clinical guidelines in the healthcare field requires citing the latest treatment guidelines, drug instructions, clinical trial data, and other information. Traditional tools, lacking the support of professional knowledge bases, generate mostly general information, failing to meet the requirements of professional creation.

[0006] There is currently no effective solution to the aforementioned technical problems. There is an urgent need for an AI-assisted creation method that can achieve multimodal collaborative generation and personalized adaptation to improve the efficiency and accuracy of content creation. Summary of the Invention

[0007] To address the aforementioned issues, this application provides a method, apparatus, device, and storage medium for generating creative content. Driven by pre-configured exclusive prompts throughout the entire process, it solves the technical problems of traditional AI-driven content creation, which is driven by user prompts, suffers from multimodal fragmentation, and has low integration of professional information. This enables high-quality, personalized, and automated generation of creative content.

[0008] The technical solution adopted in this application is as follows: Firstly, this application provides a method for generating creative content, including: Receive initial creative requirements from the user, understand the initial creative requirements based on pre-configured understanding instructions, and extract structured core creative needs; Based on pre-configured generation instructions, a draft of the content is generated according to the core creative requirements, guiding questions to help users supplement the creative requirements, and multimodal prompts that match the core creative requirements. The draft of the content and guiding questions are pushed to the user simultaneously. Receive user feedback, parse the knowledge requirements in the feedback based on pre-configured parsing instructions, call professional information matching the knowledge requirements from the preset knowledge base, and generate revised text based on the professional information, feedback information and the initial draft of the content according to pre-configured optimization instructions; Based on multimodal prompts and revised text, semantically aligned multimodal content is generated, and the revised text and multimodal content are integrated into the created content and pushed to users.

[0009] Secondly, this application also provides a creative content generation device, comprising: The requirement parsing unit is used to receive the initial creation requirements input by the user, understand the initial creation requirements based on the pre-configured understanding instructions, and extract the structured core creation requirements. The draft generation unit is used to generate a draft of the content based on the core creative requirements according to the pre-configured generation instructions, and to provide guiding questions and multimodal prompts that match the core creative requirements to guide users to supplement the creative requirements. The draft of the content and the guiding questions are pushed to the user at the same time. The text revision unit is used to receive user feedback information, parse the knowledge requirements in the feedback information based on pre-configured parsing instructions, call professional information that matches the knowledge requirements from the preset knowledge base, and generate revised text based on the professional information, feedback information and the initial draft of the content based on pre-configured optimization instructions. The content creation unit is used to generate semantically aligned multimodal content based on multimodal prompts and revised text, and integrates the revised text and multimodal content into created content to be pushed to users.

[0010] Thirdly, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described creative content generation method.

[0011] Fourthly, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described content creation method.

[0012] The above-mentioned technical solution adopted in this application can achieve the following beneficial effects: The aforementioned content generation method, device, computer equipment, and storage medium perform structured analysis of the user's initial creative requirements based on pre-configured understanding instructions to extract core creative needs. While generating a first draft based on pre-configured generation instructions, it guides the user to supplement the draft through guiding questions. Then, based on user feedback, it retrieves professional knowledge and generates revised text based on pre-configured analysis instructions and pre-configured optimization instructions. Finally, it achieves semantic alignment and integration of multimodal content.

[0013] This application breaks away from the traditional one-way generation model driven by user prompts in AI-assisted content creation tools. It pre-configures exclusive prompts throughout the entire process—from requirements analysis and initial draft generation to feedback analysis, text optimization, and multimodal generation—eliminating excessive reliance on user-generated prompts and effectively addressing the instability of generated results caused by vague, non-standard, and inconsistent user prompts. By uniformly driving model execution at each stage with exclusive prompts, it significantly improves the accuracy of requirements understanding, the standardization of content generation, and the integration of professional information, reducing the user's learning curve and the cost of repeated trial and error. Simultaneously, relying on pre-configured prompts enables text and multimodal content to be generated from the same source and semantically aligned, avoiding multimodal fragmentation and ensuring content coherence and consistency. It can be widely applied to various AI-assisted content creation scenarios such as copywriting, video script creation, graphic design, and report writing, and is particularly well-suited to meet the high demands of professional content creation in the fintech and healthcare sectors. Attached Figure Description

[0014] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A schematic diagram illustrating the application environment of a creative content generation method according to an embodiment of this application is shown; Figure 2 A flowchart illustrating a method for generating creative content according to an embodiment of this application is shown; Figure 3 A schematic diagram of the structure of a creative content generation apparatus according to an embodiment of this application is shown; Figure 4 A schematic diagram of the structure of an electronic device according to an embodiment of this application is shown. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0016] The content creation method provided in this embodiment of the invention can be applied to, for example... Figure 1In the application environment, the user terminal communicates with the server through the network. The server can receive the user's initial creation requirements through the user terminal, perform structured analysis of the core creation requirements, generate a draft content, guiding questions, and multimodal prompt words, and push them to the user terminal; after receiving the feedback information from the user terminal, the server can analyze the knowledge requirements, retrieve professional information from the preset knowledge base, generate a revised text, and then generate multimodal content semantically aligned with the revised text based on the multimodal prompt words, and push the integrated content to the user terminal.

[0017] The user terminal can be, but is not limited to, various personal computers, laptop computers, smartphones, tablet computers, intelligent creation terminals, etc. The server can be implemented by an independent server or a server cluster composed of multiple servers.

[0018] Figure 2 It shows a schematic flow chart of a creation content generation method according to an embodiment of the present application. Referring to Figure 2 As shown, this embodiment includes steps S210 to S240: Step S210, receive the initial creation requirements input by the user, and understand and extract the structured core creation requirements based on the pre-configured understanding instructions.

[0019] The creation content generation method provided by the present application can be applied to various artificial intelligence-assisted creation scenarios such as copywriting writing, video script creation, graphic design, report writing, etc., and is especially suitable for professional content creation in the fields of fintech and healthcare. The server can receive the initial creation requirements input by the user through the user terminal in real time in forms such as text and voice. For example, in the fintech field, the user inputs "Write a promotion copy for a fund product, including risk warnings" based on the user terminal; or, in the healthcare field, the user inputs "Create diabetes popular science graphics for the elderly population, including medication specifications and dietary advice" based on the user terminal. After receiving the initial creation requirements, the server performs structured analysis on them based on the pre-configured understanding instructions and exclusive prompt words, guides the model to complete the extraction of core creation requirements, and realizes standardized and unbiased requirement understanding from the source, so as to provide accurate basis for subsequent content generation.

[0020] In some optional implementations, step S210, receiving the initial creative requirements input by the user, and understanding the initial creative requirements and extracting structured core creative needs based on pre-configured understanding instructions, includes: receiving the initial creative requirements input by the user and loading pre-configured understanding instructions to guide the model in parsing the initial creative requirements; using a natural language understanding model, an intent classification model, and an entity recognition model respectively, combined with the initial creative requirements and understanding instructions, to complete the extraction of core needs, core intents, and core entities; and structurally integrating the extracted core needs, core intents, and core entities to obtain structured core creative requirements.

[0021] Understanding instructions are pre-configured model understanding rules, set by technical professionals, and can be pre-stored on the server in the form of understanding instruction prompts. Understanding instructions may include, but are not limited to, requirements parsing dimensions, intent classification criteria, and entity extraction ranges, used to guide each model to accurately understand the initial creation requirements and avoid comprehension biases. Requirements parsing dimensions can guide natural language understanding models to extract core requirements from the initial creation requirements; intent classification criteria can guide intent classification models to extract core intents from the initial creation requirements; and entity extraction ranges can guide entity recognition models to extract core entities from the initial creation requirements.

[0022] After receiving the user's initial creation request, the server can first perform basic preprocessing on the initial creation request, such as speech-to-text conversion, text standardization, word segmentation, etc., before loading the understanding instructions.

[0023] Natural language understanding models are used to semantically parse initial creative requirements based on understood instructions, extracting the user's core needs, such as creative theme, audience, form, style, and domain. Natural language understanding models can employ pre-trained models such as DeepSeek and BERT, or other similar mature models.

[0024] Intent classification models are used to categorize initial creative requirements and identify core intents, such as writing, producing, designing, polishing, and revising. Intent classification models can employ existing technologies like SVM classification models or other similar mature models.

[0025] Entity recognition models are used to extract key entities from the initial creation requirements, i.e., nouns with practical meaning. For example, specific terms in the fintech field such as "fund products" and "yield rates"; or professional terms in the healthcare field such as "diabetes" and "treatment procedures." Entity recognition models can utilize the entity extraction capabilities of existing technologies like DeepSeek-NER, or other similar mature models.

[0026] The server fills in the extracted core requirements, core intentions, and core entities according to a preset structured template, and then obtains a structured core creation requirement. For example, continuing with the above example in the medical and health field of "creating diabetes popular science graphics and texts for middle-aged and elderly people, which need to include medication specifications and dietary suggestions", the structured and integrated core creation requirement is: "Creation theme": "Diabetes popular science", "Target audience": "Middle-aged and elderly people", "Content form": "Graphics and texts", "Creation style": "Easy to understand", "Professional field": "Medical and health", "Core intention": "Create (graphics and texts)", "Core entities": "Diabetes", "Middle-aged and elderly people", "Medication specifications", "Dietary suggestions", "Graphics and texts".

[0027] Through the core logic of pre-configured understanding instructions, multi-model hierarchical understanding, and structured integration, this application accurately understands the user's initial creation requirements and extracts structured core creation requirements. Compared with the traditional artificial intelligence-assisted creation tool that directly generates content based on the original input, it solves the technical problems of demand understanding deviation, incomplete extraction of core information, and poor adaptability of professional field requirements from the source of core creation requirement understanding.

[0028] Step S220, based on the pre-configured generation instructions, generate a content draft, guiding questions for guiding the user to supplement the creation requirements, and multi-modal prompt words matching the core creation requirements, and push the content draft and guiding questions to the user synchronously.

[0029] After obtaining the structured core creation requirement, the server synchronously generates a content draft, guiding questions, and multi-modal prompt words according to the core creation requirement with exclusive prompt words of the pre-configured generation instructions, ensuring that the multi-modal prompt words are naturally matched with the text content. The guiding questions are used to guide the user to supplement creation details, clarify personalized requirements and professional requirements, and solve the problem that traditional tools cannot refine requirements due to the one-way drive of user prompt words. The multi-modal prompt words provide a basis for the generation of subsequent multi-modal content such as images and videos. The server pushes the content draft and guiding questions to the user, and the multi-modal prompt words are cached by the server for subsequent multi-modal content generation.

[0030] In some optional embodiments, in step S220, according to a pre-configured generation instruction, a preliminary content draft, guiding questions for guiding the user to supplement the creation requirements, and multimodal prompt words matching the core creation requirements are generated, and the preliminary content draft and the guiding questions are synchronously pushed to the user, including: loading the pre-configured generation instruction for guiding the model to parse the core creation requirements; respectively using a text generation large model, a question and answer model, and a prompt word generation model, combining the core creation requirements and the generation instruction, and successively completing the generation of the preliminary content draft, the generation of the guiding questions, and the generation of the multimodal prompt words; integrating the preliminary content draft and the guiding questions and then synchronously pushing them to the user.

[0031] The generation instruction is a pre-configured model generation rule set by professional technicians and can be pre-stored on the server in the form of generation instruction prompt words. The generation instruction can but is not limited to including content such as text generation specifications, question generation dimensions, prompt word generation standards, etc., and is used to guide each model to generate draft content, guiding questions, and multimodal prompt words that meet the requirements according to the core creation requirements.

[0032] For example, according to "content form = graphic and text, professional field = medical and health" in the structured core creation requirements, the server loads the graphic and text generation instruction for the medical and health field that matches.

[0033] The text generation large model is used to generate a text preliminary draft according to the core creation requirements and the generation instruction. The text generation large model can adopt GPT-4o, Qwen, etc. in the prior art, or other similar mature models. Continuing with the above example of "producing diabetes popular science graphic and text for the middle-aged and elderly population, which needs to include medication specifications and dietary advice" in the medical and health field, the text preliminary draft can be presented in layers according to "disease awareness - medication specifications - dietary advice - precautions".

[0034] The question and answer model is used to generate guiding questions for the unclarified details according to the core creation requirements and the generation instruction to guide the user to supplement the requirements. The question and answer model can adopt the general QG model in the prior art, or other similar mature models. Continuing with the above example of "producing diabetes popular science graphic and text for the middle-aged and elderly population, which needs to include medication specifications and dietary advice" in the medical and health field, the guiding questions can include: whether it is necessary to clarify the specific hypoglycemic drugs and dosage ranges, whether the image style of the graphic and text is desired to be in a cartoon style or a realistic scene style, whether it is necessary to add additional popular science modules, etc.

[0035] The prompt generation model is used to generate multimodal prompts that match core creative needs. Multimodal prompts can include, but are not limited to, prompts that adapt to the input requirements of the subsequent multimodal generation model, such as content theme, style requirements, professional details, and format standards. The prompt generation model can employ existing technologies such as GPT-4o, or other similar mature models. During multimodal prompt generation, core intent and core extraction can also be introduced as conditional variables to precisely control the generation direction.

[0036] The server integrates the generated draft content and guiding questions and pushes them to the user's terminal for viewing. Simultaneously, the server caches the associated data of the draft content, guiding questions, and multimodal prompts to facilitate rapid matching upon receiving subsequent feedback.

[0037] This application utilizes pre-configured generation instructions and multi-model collaborative generation to simultaneously generate initial content drafts, guiding questions, and multimodal prompts based on structured core creative requirements. Through parallel operation of multiple models, while generating the initial content draft, it automatically uncovers unclear details within the core creative requirements and generates guiding questions, thus advancing the output of the initial content draft and the collection of potential requirements simultaneously. The multimodal prompts and the initial content draft are generated based on the same structured core creative requirements and share pre-configured generation instructions, ensuring a match between the multimodal prompts and the initial content.

[0038] Step S230: Receive user feedback information, parse the knowledge requirements in the feedback information based on pre-configured parsing instructions, call professional information matching the knowledge requirements from the preset knowledge base, and generate revised text based on the professional information, feedback information, and initial content draft according to pre-configured optimization instructions.

[0039] Users send feedback to the server via their devices based on the initial draft of the pushed content and guiding questions. Feedback may include, but is not limited to, requests for modifications to the initial draft and answers to the guiding questions. For example, user feedback might request clarification of the specific dosage ranges for metformin and gliclazide, a realistic scene style for images, the addition of a blood glucose monitoring module, and modifications to some text descriptions. Upon receiving the feedback, the server analyzes the knowledge requirements based on pre-configured parsing prompts and retrieves relevant professional information. Combining the feedback and the initial draft with pre-configured optimization prompts, it then generates a revised text.

[0040] In some optional implementations, step S230, receiving user feedback information, parsing the knowledge requirements in the feedback information based on pre-configured parsing instructions, calling professional information matching the knowledge requirements from a preset knowledge base, and generating revised text based on the professional information, feedback information, and initial draft content based on pre-configured optimization instructions, includes: receiving user feedback information and loading pre-configured parsing instructions for guiding the model to parse the feedback information and pre-configured optimization instructions for guiding the model to optimize the text content; using an intent recognition model combined with the feedback information and parsing instructions to identify feedback content belonging to knowledge requirements; using an entity extraction model combined with the feedback content and parsing instructions to extract key retrieval entities from the feedback content to form a set of retrieval keywords; searching the preset knowledge base based on the set of retrieval keywords to match professional information that matches the knowledge requirements; and using a large text generation model, combined with the initial draft content, professional information, feedback information, and optimization instructions, to complete the generation of revised text.

[0041] Parsing instructions are pre-configured rules for parsing feedback information, set by technical professionals, and can be pre-stored on the server in the form of parsing instruction prompts. Parsing instructions may include, but are not limited to, feedback type identification criteria, knowledge requirement dimensions, key entity extraction rules, etc., to guide the model in accurately parsing feedback information.

[0042] Optimization instructions are pre-configured text content optimization rules set by technical professionals and can be pre-stored on the server in the form of optimization instruction prompts. Optimization instructions may include, but are not limited to, text modification specifications, professional information fusion standards, and content logic adjustment requirements, used to guide the model in optimizing and modifying text content.

[0043] After receiving user feedback, the server can first perform basic preprocessing on the feedback, such as speech-to-text conversion, text normalization, and word segmentation, before loading parsing and optimization instructions.

[0044] The intent recognition model is used to identify the intent in feedback information based on parsed instructions, determining knowledge needs, basic modification needs, and personalized preference needs within the feedback information. Knowledge needs refer to requirements in the feedback information that require retrieving professional information from a pre-defined knowledge base to satisfy. Examples include data needs, case study needs, policy document needs, and professional standard needs. The intent recognition model can employ pre-trained models such as DeepSeek, or other similar mature models. For instance, regarding the user feedback information mentioned above, "clarifying the specific dosage range of metformin and gliclazide" is identified as a knowledge need, "selecting a realistic scene style for the image" as a basic modification need, and "adding a blood glucose monitoring module" as a personalized preference need.

[0045] The entity recognition model, based on the identified knowledge requirements and combined with the key entity extraction rules of the parsing instructions, extracts the key retrieval entities, namely the core keywords used to retrieve professional information from a preset knowledge base, forming a retrieval keyword set. For example, the retrieval keyword set {metformin, gliclazide, dosage range} is extracted.

[0046] The server inputs a set of search keywords into a pre-defined knowledge base and uses keyword retrieval to filter out matching professional knowledge from the knowledge base. The pre-defined knowledge base can employ retrieval architectures such as Elasticsearch or vector databases. It can also deploy general and specialized knowledge bases depending on the scenario. For example, in the fintech field, it could include market data, policy documents, product information, and compliance regulations; or in the healthcare field, it could include treatment guidelines, drug instructions, clinical trial data, and disease information.

[0047] The large text generation model is used again to integrate professional information into the initial draft based on optimization instructions. The draft is then adjusted again based on feedback to generate revised text. Compared to the initial draft, the revised text is more professional, logically coherent, and stylistically more aligned with user preferences.

[0048] This application, by loading pre-configured parsing and optimization instructions and combining intent recognition and entity extraction models, filters out knowledge needs from feedback information. Based on a set of search keywords, it performs targeted searches in a pre-defined knowledge base, matching relevant professional information and significantly improving the credibility and applicability of the generated content. A single user feedback based on a guiding question triggers a knowledge base search without further follow-up questions. The retrieved professional information is automatically integrated into the initial draft, and a large text generation model completes logical integration and language optimization, greatly shortening the creation cycle.

[0049] Step S240: Generate semantically aligned multimodal content based on multimodal prompts and revised text, and integrate the revised text and multimodal content into created content to push to the user.

[0050] After generating the revised text, the server uses the previously generated multimodal prompts to generate multimodal content based on the revised text. Cross-modal alignment ensures semantic consistency between the multimodal content and the revised text, resolving the issue of fragmented multimodal content in traditional tools. Finally, the two are integrated into the created content and pushed to the user. Multimodal content generation can include the following two scenarios: Image-text matching: Generating matching illustrations, data visualization charts, and other image content, ensuring automatic alignment between image descriptions and revised text content; Video script linkage: Generating storyboards, accompanying voice-overs, and other audio-visual content, adjusting camera logic (such as slow motion and shot transitions) to achieve synchronized generation of scripts and visuals.

[0051] In some optional implementations, step S240, generating semantically aligned multimodal content based on multimodal cue words and revised text, and integrating the revised text and multimodal content into created content to be pushed to the user, includes: generating initial multimodal content using a multimodal generation model combined with multimodal cue words; semantically aligning the initial multimodal content with the revised text using a cross-modal alignment model to obtain multimodal content; and logically integrating the revised text and multimodal content to form created content to be pushed to the user.

[0052] Multimodal generative models are artificial intelligence models capable of generating non-textual content such as images, videos, and audio. Based on the content format, multimodal generative models can include existing technologies such as Stable Diffusion, DALL E3, or other similar mature models.

[0053] The server selects a matching multimodal generation model based on the content format in the core creation requirements, inputs the previously generated multimodal prompts into the multimodal generation model, and then generates initial multimodal content. Initial multimodal content refers to multimodal content generated solely based on multimodal prompts and not semantically aligned with the revised text.

[0054] Cross-modal alignment models refer to models that can achieve semantic matching between different modalities of content such as text, images, and videos. They can employ existing CLIP models or other similar mature models. Their core function is to ensure consistency between multimodal content and revised text, avoiding semantic fragmentation across modalities.

[0055] For example, if the revised text is "Metformin is taken once daily, 0.5g each time", while the initial multimodal content (illustration) shows "twice daily, 0.5g each time", the cross-modal alignment model identifies the inconsistency and feeds it back to the multimodal generation model to regenerate the multimodal content until it matches the revised text.

[0056] The server will logically integrate and format the revised text and multimodal content to ensure the coherence and readability of the content before pushing it to the user's terminal.

[0057] In this application, the multimodal generation model generates initial multimodal content based on multimodal cue words that are from the same source as the initial draft content. Then, the cross-modal alignment model extracts the semantic features of the revised text and the initial multimodal content and maps them to the same space to accurately verify consistency. After the initial multimodal content is generated, content that does not match the semantics of the text can be quickly filtered out, thus preventing unqualified multimodal content from entering the final created content.

[0058] Through the above technical solution, this application has the following beneficial effects compared with the prior art: The content generation method proposed in this application fundamentally solves the shortcomings of traditional AI-assisted creation tools, such as user prompt-driven one-way generation, multimodal fragmentation, insufficient professional adaptation, and lack of personalization, through a closed-loop design of structured demand analysis, collaborative generation of initial draft + guiding questions + multimodal prompts, feedback information-driven professional information fusion, and multimodal semantic alignment integration.

[0059] The method in this application fundamentally eliminates the over-reliance on user-generated prompts by configuring pre-configured exclusive prompts throughout the entire process of requirements analysis, draft generation, feedback analysis, text optimization, and multimodal generation. This effectively solves the problem of unstable generation results caused by vague, non-standard, and highly varied user prompts.

[0060] In this application, multimodal prompts and initial content drafts are generated synchronously based on the same core creative requirements. Subsequently, a cross-modal alignment model is used to ensure accurate semantic matching between multimodal content and revised text, achieving a high degree of synergy between revised text and multimodal content, avoiding logical breaks in cross-modal content, and reducing the cost of later integration and modification.

[0061] Based on the technical solution proposed in this application, users only need to input initial creation requirements and optimize based on guidance feedback to obtain integrated complete creative content, which greatly reduces the creation threshold and allows non-professional creators to quickly produce high-quality, professional creative content.

[0062] In some optional implementations, the method further includes: determining whether the user confirms the created content; if not confirmed, returning to the step of receiving feedback information for multiple rounds of iterative optimization; if confirmed, completing the content creation process.

[0063] After the server pushes the created content to the user, it determines whether the user confirms the created content. If the user confirms, the content creation process is completed; if the user does not confirm, it returns to step S230 to receive new feedback information from the user and performs multiple rounds of iterative optimization until the user confirms the created content.

[0064] In some optional implementations, the method further includes: maintaining historical interaction context using a dialogue state tracking model and compressing historical interaction context using a summary model; making targeted adjustments to multimodal prompts, pre-configured parsing instructions, and optimization instructions based on feedback information of unconfirmed user creation content and in conjunction with historical interaction context; and retaining adjustment records for each round of iteration after completion to form an adjustment record library adapted to users' personalized creation needs.

[0065] The dialogue state tracking model is used to maintain the complete historical interaction context between the user and the server, including initial creation requirements, feedback information in each round, and generated versions of content. Its core function is to ensure that context is not lost during subsequent iterations. The dialogue state tracking model can adopt the existing DST model or other similar mature models.

[0066] The summarization model is used to compress historical interaction context, avoiding information overload while retaining key information. The summarization model can employ the existing T5 model or other similar mature models.

[0067] If the user does not confirm the content they create, the server analyzes the user's creative preferences, modification patterns, and professional requirements based on new user feedback and historical interaction context. It then makes targeted adjustments to the multimodal prompts, parsing instructions, and optimization instructions to ensure that the generated content better meets the user's needs. Adjustments can be made incrementally, modifying only prompts and / or instructions relevant to user feedback. Further adjustments can also be made based on a reinforcement learning framework. For example, if the user requests "add professional data" multiple times, the weight of professional information in the generated results can be increased.

[0068] After each iteration, the server stores the adjustments made in that round in the format of User ID-Adjustment Item-Adjustment Content, including instruction adjustments, prompt word adjustments, and content modifications, forming an adjustment record library. This allows for the binding of personalized needs with users. The more times a user creates content, the richer the adjustment records become, and the more accurate the personalized adaptation.

[0069] This application incorporates user feedback into a closed-loop optimization process, performing multiple iterations even before the user confirms the content creation. Each iteration is precisely adjusted based on the content deficiencies of the previous round and new user needs, ensuring that each iteration solves specific problems and gradually approaches the user's final requirements. A dialogue state tracking model fully maintains the historical interaction context, and a summary model compresses core information, avoiding information overload while ensuring that key needs are not overlooked. The adjustment records from each iteration are continuously accumulated in an adjustment record library, forming a rich, scenario-based, and personalized dataset. This dataset can be used for subsequent fine-tuning and optimization of each model.

[0070] Figure 3 A schematic diagram of the structure of a content creation apparatus according to an embodiment of this application is shown. (Refer to...) Figure 3 As shown, the content creation device 300 includes: The requirement parsing unit 310 is used to receive the initial creation requirements input by the user, understand the initial creation requirements based on the pre-configured understanding instructions, and extract the structured core creation requirements. The draft generation unit 320 is used to generate a draft of the content based on the core creative requirements according to the pre-configured generation instructions, and to guide users to supplement the creative requirements with guiding questions and multimodal prompts that match the core creative requirements. The draft of the content and the guiding questions are pushed to the user simultaneously. The text revision unit 330 is used to receive user feedback information, parse the knowledge requirements in the feedback information based on pre-configured parsing instructions, call professional information matching the knowledge requirements from the preset knowledge base, and generate revised text based on the professional information, feedback information and the initial draft of the content based on pre-configured optimization instructions. Content creation unit 340 is used to generate semantically aligned multimodal content based on multimodal prompts and revised text, and integrates the revised text and multimodal content into created content to be pushed to users.

[0071] In some optional embodiments, in the above-described apparatus, the requirement parsing unit 310 is used to: receive the initial creation requirements input by the user and load pre-configured understanding instructions for guiding the model to parse the initial creation requirements; use a natural language understanding model, an intent classification model, and an entity recognition model respectively, combined with the initial creation requirements and understanding instructions, to complete the extraction of core requirements, core intents, and core entities; and integrate the extracted core requirements, core intents, and core entities in a structured manner to obtain structured core creation requirements.

[0072] In some optional implementations, in the above-described apparatus, the draft generation unit 320 is used to: load pre-configured generation instructions for guiding the model to parse core creative requirements; use a large text generation model, a question-and-answer model, and a prompt word generation model respectively, combined with core creative requirements and generation instructions, to sequentially complete the generation of the content draft, the generation of guiding questions, and the generation of multimodal prompt words; and integrate the content draft and guiding questions and push them to the user simultaneously.

[0073] In some optional embodiments, in the above-described apparatus, the text revision unit 330 is configured to: receive user feedback information and load pre-configured parsing instructions for guiding the model to parse the feedback information and pre-configured optimization instructions for guiding the model to optimize the text content; use an intent recognition model to combine the feedback information and parsing instructions to identify feedback content belonging to knowledge needs in the feedback information; use an entity extraction model to combine the feedback content and parsing instructions to extract key retrieval entities from the feedback content to form a set of retrieval keywords; search a preset knowledge base based on the set of retrieval keywords to match professional information that matches the knowledge needs; and use a large text generation model to combine the initial draft of the content, professional information, feedback information, and optimization instructions to complete the generation of revised text.

[0074] In some optional implementations, in the above-described apparatus, the content creation unit 340 is used to: generate initial multimodal content using a multimodal generation model combined with multimodal prompts; semantically align the initial multimodal content with the revised text using a cross-modal alignment model to obtain multimodal content; and logically integrate the revised text and the multimodal content to form created content that is pushed to the user.

[0075] In some optional implementations, the above-mentioned device further includes an iteration unit for: determining whether the user confirms the created content; if not confirmed, returning to the step of receiving feedback information for multiple rounds of iterative optimization; if confirmed, completing the created content generation process.

[0076] In some optional embodiments, the above-mentioned device further includes an adjustment unit, used for: maintaining historical interaction context using a dialogue state tracking model, compressing historical interaction context using a summary model; making targeted adjustments to multimodal prompts, pre-configured parsing instructions and optimization instructions based on feedback information of unconfirmed user creation content and in combination with historical interaction context; and retaining adjustment records for each round after each iteration to form an adjustment record library adapted to the user's personalized creation needs.

[0077] It should be noted that the aforementioned creative content generation device 300 can realize all of the aforementioned creative content generation methods, which will not be elaborated further.

[0078] Figure 4 This invention illustrates a schematic diagram of the structure of an electronic device according to an embodiment of the present application. Figure 4 As shown, the electronic device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used for communication with external devices via a network connection. When the computer program is executed by the processor, it implements the functions or steps of the content creation method.

[0079] In one embodiment, the electronic device provided in this application includes a memory and a processor. The memory stores a database and a computer program that can run on the processor. When the processor executes the computer program, it implements the steps of the aforementioned creative content generation method.

[0080] The above is as stated in this application. Figure 3The method executed by the content creation device disclosed in the illustrated embodiment can be applied to a processor or implemented by a processor. During implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by software instructions. The processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The steps of the method disclosed in the embodiments of this application can be directly embodied as being executed by a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0081] In one embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the aforementioned content creation method.

[0082] It should be noted that the functions or steps that the above-mentioned electronic devices or computer-readable storage media can achieve can be referred to the relevant descriptions in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0083] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0084] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above.

[0085] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0086] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for generating creative content, characterized in that, include: Receive initial creative requirements from the user, understand the initial creative requirements based on pre-configured understanding instructions, and extract structured core creative needs; Based on pre-configured generation instructions, a draft of the content is generated according to the core creative requirements, guiding questions to help users supplement the creative requirements, and multimodal prompts that match the core creative requirements. The draft of the content and guiding questions are pushed to the user simultaneously. Receive user feedback, parse the knowledge requirements in the feedback based on pre-configured parsing instructions, call professional information matching the knowledge requirements from the preset knowledge base, and generate revised text based on the professional information, feedback information, and initial content draft based on pre-configured optimization instructions; Based on multimodal prompts and revised text, semantically aligned multimodal content is generated, and the revised text and multimodal content are integrated into the created content and pushed to users.

2. The method according to claim 1, characterized in that, The process of receiving initial creative requirements from the user, understanding these requirements based on pre-configured understanding instructions, and extracting structured core creative needs includes: Receive initial creation requirements from the user and load pre-configured understanding instructions to guide the model in parsing the initial creation requirements; Natural language understanding model, intent classification model and entity recognition model are used respectively, combined with initial creation requirements and understanding instructions, to complete the extraction of core needs, core intent and core entities; The extracted core needs, core intentions, and core entities are structurally integrated to obtain structured core creative needs.

3. The method according to claim 1, characterized in that, The pre-configured generation instructions generate a draft of content based on core creative needs, guiding questions to help users supplement creative requirements, and multimodal prompts matching the core creative needs. The draft content and guiding questions are then simultaneously pushed to the user, including: Load pre-configured generation instructions to guide the model in resolving core creative requirements; The system employs a large text generation model, a question-and-answer model, and a prompt word generation model, respectively, and combines core creative needs and generation instructions to complete the generation of initial content drafts, guiding questions, and multimodal prompt words. The initial draft of the content and the guiding questions were integrated and pushed to users simultaneously.

4. The method according to claim 1, characterized in that, The process of receiving user feedback involves parsing the knowledge requirements within the feedback information based on pre-configured parsing instructions, retrieving relevant professional information from a pre-defined knowledge base, and generating revised text based on the professional information, feedback information, and initial draft content using pre-configured optimization instructions. This includes: Receive user feedback and load pre-configured parsing instructions to guide the model in parsing the feedback and pre-configured optimization instructions to guide the model in optimizing the text content; An intent recognition model is used in conjunction with feedback information and parsing instructions to identify feedback content that belongs to knowledge needs. By employing an entity extraction model that combines feedback content and parsing instructions, key search entities are extracted from the feedback content to form a set of search keywords. The search is performed on a pre-defined knowledge base based on a set of search keywords to match professional information that meets the knowledge needs. A large-scale text generation model is used, combining initial drafts, professional information, feedback, and optimization instructions to generate revised text.

5. The method according to claim 1, characterized in that, The process of generating semantically aligned multimodal content based on multimodal prompts and revised text, and integrating the revised text and multimodal content into created content pushed to users, includes: Initial multimodal content is generated using a multimodal generation model combined with multimodal cue words; The initial multimodal content is semantically aligned with the revised text using a cross-modal alignment model to obtain the multimodal content. The revised text and multimodal content are logically integrated to form the created content that is pushed to users.

6. The method according to claim 1, characterized in that, The method further includes: Determine whether the user has confirmed the created content; If not confirmed, return to the step of receiving feedback information for multiple rounds of iterative optimization; If confirmed, the content creation process is complete.

7. The method according to claim 6, characterized in that, The method further includes: A dialogue state tracking model is used to maintain the historical interaction context, and a summary model is used to compress the historical interaction context. Based on feedback from users who have not confirmed their creations, and combined with the historical interaction context, we make targeted adjustments to multimodal prompts, pre-configured parsing instructions, and optimization instructions. After each iteration is completed, the adjustment records for each iteration are retained to form an adjustment record library that adapts to users' personalized creative requirements.

8. A creative content generation device, characterized in that, include: The requirement parsing unit is used to receive the initial creation requirements input by the user, understand the initial creation requirements based on the pre-configured understanding instructions, and extract the structured core creation requirements. The draft generation unit is used to generate a draft of the content based on the core creative requirements according to the pre-configured generation instructions, and to provide guiding questions and multimodal prompts that match the core creative requirements to guide users to supplement the creative requirements. The draft of the content and the guiding questions are pushed to the user at the same time. The text revision unit is used to receive user feedback information, parse the knowledge requirements in the feedback information based on pre-configured parsing instructions, call professional information that matches the knowledge requirements from the preset knowledge base, and generate revised text based on the professional information, feedback information and the initial draft of the content based on pre-configured optimization instructions. The content creation unit is used to generate semantically aligned multimodal content based on multimodal prompts and revised text, and integrates the revised text and multimodal content into created content to be pushed to users.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the content creation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the content creation method as described in any one of claims 1 to 7.

Images