A large model-based all-media content intelligent generation review and closed-loop iteration method and system

Abstract

The application discloses a kind of based on big model's full media content intelligent generation review and closed loop iteration method and system, belong to electric digital data processing technical field, this method includes: obtaining the historical manuscript data of target author and the generation full media content first draft of content task demand;Recognize media type and load different scoring templates, reconstruct the content data after review through channel, generate online digital channel publishing data and traditional offline channel publishing data, and perform de-identification competitive product benchmarking evaluation;Collect the limited interaction data and operation briefing data of online publishing channel backflow, obtain relative utility index;Attribute mapping is carried out to relative utility index, style characteristics, scoring matrix and dynamic benchmarking results, style preference weight atlas is generated, and online style template library and traditional paper media style constraint library are updated, to realize the generation quality, publishing security and cross-media optimization efficiency of full media content.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing, and in particular to a method and system for intelligent generation, review, and closed-loop iteration of multimedia content based on a large model. Background Technology

[0002] Media convergence organizations face the following challenges in their efforts to achieve intelligent and comprehensive transformation: First, there is a disconnect in the quality and safety evaluation of automated editing and proofreading across multiple categories. The reporting formats of converged media organizations cover five major categories: print media, new media, video, photography, and comics. Existing AI proofreading systems are mostly limited to typo checking and cannot perform differentiated, in-depth value assessments and quantitative scoring of core competitiveness for these five types of manuscripts with vastly different quality standards (e.g., print media emphasizes guidance and logical rigor, new media emphasizes online appeal and timeliness, video emphasizes storyboarding and pacing, and photography and comics emphasize visual creativity). Traditional multi-level manual proofreading is cumbersome and has a high error rate.

[0003] Second, there is a disconnect between automated generation and cross-platform distribution. Existing automated writing systems produce mechanical and impersonal text that cannot automatically adapt to the modal requirements of different channels. In actual multi-channel distribution operations, a significant amount of manpower is still required for rewriting, layout adjustments, or storyboard editing, which can easily lead to deviations in the core message of the news due to editors' personal interpretations.

[0004] Third, there is a disconnect between traditional print media and the online digital closed loop. When establishing a "generation-distribution-optimization" closed-loop mechanism, external third-party online platforms, due to data security and commercial barriers, strictly limit the opening of interfaces for in-depth interactive data, only providing coarse-grained feedback. Traditional print media, due to physical limitations, cannot directly feed back its distribution and evaluation data. Existing automatic generation models face a "black box" optimization dilemma, especially preventing traditional print media channels from enjoying the self-evolutionary benefits of data feedback. Summary of the Invention

[0005] This invention provides a method and system for intelligent generation, review, and closed-loop iteration of multimedia content based on a large model. This addresses the problems in existing multimedia content production, such as the gap in quality and safety evaluation of automated editing and review across multiple categories, the homogenization of cross-platform distribution of generated content, and the disconnect between traditional print media and online digital closed loops. These issues result in inconsistent content generation quality, low efficiency in cross-media adaptation, and the inability to achieve closed-loop iterative optimization in data-constrained environments.

[0006] In a first aspect, embodiments of the present invention provide a method for intelligent generation, review, and closed-loop iteration of multimedia content based on a large model, including: Obtain the task requirements for the content to be generated and the historical manuscript data of the target creator, extract the writing style features from the historical manuscript data, construct a historical writing style feature vector, and generate a first draft of multimedia content based on the historical writing style feature vector and the task requirements. Identify the media type corresponding to the initial draft of the multimedia content, load the corresponding differentiated scoring template based on the media type, perform quantitative scoring and compliance audit on the initial draft of the multimedia content, obtain a structured review report, and determine the revision results and release decision based on the structured review report; Based on the target distribution channel attributes, the audited content is reconstructed across channels in a multimodal manner, and online digital channel release data and traditional offline channel publication data are generated respectively. At the same time, an initial benchmarking matrix is ​​generated based on the content of competing products on the same theme. After the online digital channel data is published for a preset period, restricted interaction data and operation briefing data returned from each online publishing channel are collected. The restricted interaction data is dynamically baseline aligned to obtain a relative utility index. The initial benchmarking matrix is ​​updated based on the operation briefing data to obtain a dynamic benchmarking evaluation result. Using the relative utility index as the dependent variable, and the historical writing style feature vector, the scoring matrix in the structured review report, and the dynamic benchmarking evaluation results as the concatenated and aligned fine-grained style feature matrix as the independent variable, the independent variable and the dependent variable are input into the feature association evaluation model for attribution mapping analysis to generate a style preference weight map. The online style template library and the traditional print media writing style constraint library are updated based on the style preference weight map.

[0007] Optionally, the task requirements for the content to be generated and the historical manuscript data of the target creator are obtained; stylistic features are extracted from the historical manuscript data to construct a historical stylistic feature vector; and a preliminary draft of multimedia content is generated based on the historical stylistic feature vector and the task requirements, including: Obtain historical manuscript data of the target creator within a preset time range, and preprocess the historical manuscript data; Historical writing style feature vectors are extracted from the processed historical manuscript data. These historical writing style feature vectors include at least: sentence length distribution, title structure features, vocabulary preference features, rhetorical features, emotional tone features, paragraph organization features, and theme expression features. Using the historical writing style feature vector and the task requirements of the content to be generated as constraints, a first draft of multimedia content that retains the writing style features of the target creator is generated.

[0008] Optionally, the media type corresponding to the initial draft of the multimedia content is identified, a corresponding differentiated scoring template is loaded based on the media type, the initial draft of the multimedia content is quantitatively scored and audited for compliance, a structured review report is obtained, and the revision results and release decisions are determined based on the structured review report, including: Identify the media type of the initial draft of the multimedia content, wherein the media type includes at least one of print media, new media, video, photos, and comics; Based on the media type, load the corresponding differentiated scoring template, use a combination of base score and cumulative deduction to quantitatively score the initial draft of the multimedia content, conduct a compliance audit on the initial draft of the multimedia content, and output a structured review report; If the structured review report contains high-risk items, the release decision will be to prohibit release or require manual review. If the structured review report does not contain any high-risk items and the overall score is less than the preset release threshold, the initial draft of the multimedia content will be revised according to the modification suggestions, and the revised content will be re-evaluated and subjected to quantitative scoring and compliance audit. If the structured review report does not contain any high-risk items and the overall score reaches the preset release threshold, then the release decision is determined to allow release or allow printing.

[0009] Optionally, based on the target distribution channel attributes, the audited content is reconstructed across channels in a multimodal manner, generating online digital channel distribution data and traditional offline channel publication data respectively. Simultaneously, an initial benchmarking matrix is ​​generated based on competitor content on the same theme, including: Obtain content data and target distribution channel attributes after review and approval; When the target publishing channel is an online digital channel, the content data is reconstructed into one or more of the following based on the target publishing channel attributes: short text, long text, title group, topic tag group, cover prompt, video storyboard, image description, or short video script. When the target distribution channel is a traditional offline channel, the content data is converted into structured page data or digital proof files based on the print media layout template; Obtain competitor content with the same theme or event as the content data, and perform source de-identification processing on the content data and the competitor content; Based on a pre-set benchmarking evaluation template, the de-identified competitor content is evaluated to obtain an initial benchmarking matrix.

[0010] Optionally, after the online digital channel data release runs for a preset period, restricted interaction data and operational briefing data from each online release channel are collected. Dynamic baseline alignment processing is performed on the restricted interaction data to obtain a relative utility index. The initial benchmarking matrix is ​​then updated based on the operational briefing data to obtain a dynamic benchmarking evaluation result, including: After the data is published on the online digital channels, restricted interactive data that flows back from each online publishing channel is collected according to a preset collection cycle. Obtain online new media operation briefing data from this organization and its competitors, and extract dissemination performance, content structure, user feedback, and competitor dissemination characteristics from the online new media operation briefing data; Using historical baseline data from various online publishing channels, average data of similar content, or average data from platforms within the same time window as dynamic baselines, the restricted interaction data is normalized, denoised, and processed for outliers. The relative utility index is calculated based on the processed restricted interaction data. The relative utility index is used to characterize the content dissemination effect after excluding differences in traffic allocation across different platforms. The initial benchmarking matrix is ​​updated based on the operational briefing data and the relative utility index to obtain dynamic benchmarking evaluation results.

[0011] Optionally, the relative utility index is used as the dependent variable, and the historical writing style feature vector, the scoring matrix in the structured review report, and the dynamic benchmarking evaluation results are concatenated and aligned into a fine-grained style feature matrix as the independent variable. The independent variable and the dependent variable are input into a feature association evaluation model for attribution mapping analysis to generate a style preference weight map, including: The historical writing style feature vector, the scores of each dimension in the structured review report, the deduction details, the risk markers, the modification suggestions, and the benchmarking scores in the dynamic benchmarking evaluation results are aligned by fields and time windows to obtain a fine-grained style feature matrix. Using the relative utility index as the dependent variable and the fine-grained style feature matrix as the independent variable, attribution mapping analysis is performed through a feature association evaluation model. Based on the results of the attribution mapping analysis, the correlation weights between different writing style elements, rating dimensions, benchmarking dimensions and dissemination effects are determined, and a style preference weight map is generated.

[0012] Optionally, updating the online style template library and the traditional print media writing style constraint library based on the style preference weight graph includes: Update the cue word constraints in the online style template library based on the style preference weight map, and adjust the multimodal reconstruction strategy of the online digital channels; The weights related to print media topic selection, title, lead, structure, and expression style in the style preference weight map are mapped to the traditional print media writing style constraint library, and the prompt word parameters when generating traditional print media manuscripts are adjusted.

[0013] Secondly, embodiments of the present invention provide a large-model-based intelligent generation, review, and closed-loop iteration system for multimedia content. The system is used to execute the large-model-based intelligent generation, review, and closed-loop iteration method for multimedia content as described in any embodiment of the present invention, including: The acquisition module is used to acquire the task requirements of the content to be generated and the historical manuscript data of the target creator, extract the writing style features of the historical manuscript data, construct the historical writing style feature vector, and generate the first draft of multimedia content based on the historical writing style feature vector and the task requirements. The determination module is used to identify the media type corresponding to the initial draft of the multimedia content, load the corresponding differentiated scoring template based on the media type, perform quantitative scoring and compliance audit on the initial draft of the multimedia content, obtain a structured review report, and determine the correction results and release decision based on the structured review report; The generation module is used to perform cross-channel multimodal reconstruction of audited content based on the target publishing channel attributes, and generate online digital channel publishing data and traditional offline channel publishing data respectively. At the same time, it generates an initial benchmarking matrix based on the content of competitors on the same theme. The data collection module is used to collect restricted interaction data and operational briefing data from each online publishing channel after the online digital channel publishes data for a preset period of time, perform dynamic baseline alignment processing on the restricted interaction data to obtain a relative utility index, and update the initial benchmarking matrix based on the operational briefing data to obtain a dynamic benchmarking evaluation result. The analysis module is used to take the relative utility index as the dependent variable, and to align the historical writing style feature vector, the scoring matrix in the structured review report, and the dynamic benchmarking evaluation results into a fine-grained style feature matrix as the independent variable. The independent variable and the dependent variable are input into the feature association evaluation model for attribution mapping analysis to generate a style preference weight map. The online style template library and the traditional print media writing style constraint library are updated according to the style preference weight map.

[0014] Thirdly, embodiments of the present invention provide an electronic device, the electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to execute the intelligent generation, review, and closed-loop iteration method for multimedia content based on a large model as described in any embodiment of the present invention.

[0015] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing computer instructions, which are used to cause a processor to execute the intelligent generation, review, and closed-loop iteration method for multimedia content based on a large model as described in any embodiment of the present invention.

[0016] This invention extracts the creator's historical writing style feature vectors and combines them with task requirements to generate personalized initial drafts. Then, it loads differentiated scoring templates based on media type for quantitative scoring and compliance auditing. Subsequently, it reconstructs the approved content across channels in a multimodal manner and generates a competitor benchmarking matrix. Furthermore, it collects limited interactive data from online publishing channels and calculates a relative utility index through dynamic baseline alignment. Finally, it concatenates this index with writing style features, the scoring matrix, and the benchmarking results into a fine-grained feature matrix, which is then input into a feature association evaluation model for attribution mapping analysis to generate a style preference weight map. Based on this, it updates the online style template library and the traditional print media writing style constraint library. This constitutes a complete closed loop from generation, review, distribution, collection to adaptive updates of the template library. It overcomes the shortcomings of existing technologies, such as AI proofreading being limited to typo checking and unable to provide differentiated in-depth evaluation, automated writing systems being mechanical and unable to adapt to multimodal channels, and the limitation of third-party platform data preventing traditional print media from enjoying the self-evolutionary benefits of data feedback. This invention achieves the beneficial effects of significantly improving the quality of multimedia content generation, publishing security, and cross-media optimization efficiency in a data-constrained environment.

[0017] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 A flowchart of a method for intelligent generation, review, and closed-loop iteration of multimedia content based on a large model, provided in Embodiment 1 of the present invention; Figure 2 This is a framework diagram of a multimedia content intelligent generation, review, and closed-loop iterative system based on a large model, provided in Embodiment 3 of the present invention. Figure 3 A schematic diagram of an electronic device that can be used to implement embodiments of the present invention is shown. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0020] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus. Example

[0021] Figure 1 This is a flowchart of a method for intelligent generation, review, and closed-loop iteration of multimedia content based on a large model, provided in Embodiment 1 of the present invention. This embodiment is applicable to the intelligent generation, review, and closed-loop iteration of media content, and the method can be executed by a system for intelligent generation, review, and closed-loop iteration of multimedia content based on a large model. Figure 1 As shown, the method includes: S110. Obtain the task requirements for the content to be generated and the historical manuscript data of the target creator, extract the writing style features from the historical manuscript data, construct a historical writing style feature vector, and generate a first draft of multimedia content based on the historical writing style feature vector and the task requirements.

[0022] S120. Identify the media type corresponding to the initial draft of the multimedia content, load the corresponding differentiated scoring template based on the media type, perform quantitative scoring and compliance audit on the initial draft of the multimedia content, obtain a structured review report, and determine the correction results and release decision based on the structured review report.

[0023] S130. Based on the target release channel attributes, the audited content is reconstructed across channels in a multimodal manner, and online digital channel release data and traditional offline channel publication data are generated respectively. At the same time, an initial benchmarking matrix is ​​generated based on the content of competing products on the same theme.

[0024] S140. After the online digital channel publishes data for a preset period, collect the restricted interaction data and operation briefing data returned from each online publishing channel, perform dynamic baseline alignment processing on the restricted interaction data to obtain the relative utility index, and update the initial benchmarking matrix based on the operation briefing data to obtain the dynamic benchmarking evaluation result.

[0025] S150. Using the relative utility index as the dependent variable, and the historical writing style feature vector, the scoring matrix in the structured review report, and the dynamic benchmarking evaluation results spliced ​​and aligned into a fine-grained style feature matrix as the independent variable, the independent variable and the dependent variable are input into the feature association evaluation model for attribution mapping analysis to generate a style preference weight map. The online style template library and the traditional print media writing style constraint library are updated according to the style preference weight map.

[0026] In this embodiment, personalized drafts are generated by extracting the creator's historical writing style feature vectors and combining them with task requirements. Then, differentiated scoring templates are loaded according to media type for quantitative scoring and compliance auditing. Subsequently, the content that passes the review is reconstructed across channels in a multimodal manner to generate a competitor benchmarking matrix. Then, limited interaction data from online publishing channels is collected and the relative utility index is calculated through dynamic baseline alignment. Finally, this index is concatenated with writing style features, scoring matrix and benchmarking results to form a fine-grained feature matrix and input into a feature association evaluation model for attribution mapping analysis to generate a style preference weight map. Based on this, the online style template library and the traditional print media writing style constraint library are updated. This constitutes a complete closed loop from generation, review, distribution, collection to adaptive updating of the template library. It overcomes the shortcomings of existing technologies, such as AI proofreading being limited to typo checking and unable to conduct differentiated in-depth evaluation, automatic writing systems having mechanical text and being unable to adapt to multimodal channels, and the limitation of third-party platform data preventing traditional print media from enjoying the self-evolution dividend of data feedback. It achieves the beneficial effect of significantly improving the quality of multimedia content generation, publishing security and cross-media optimization efficiency in a data-constrained environment. Example

[0027] The technical solution in this embodiment is a further refinement based on the above embodiments.

[0028] In step S110, the task requirements for the content to be generated and the historical manuscript data of the target creator are obtained. The writing style features of the historical manuscript data are extracted to construct a historical writing style feature vector. Based on the historical writing style feature vector and the task requirements, a preliminary draft of the multimedia content is generated, including: Obtain historical manuscript data of the target creator within a preset time range, and preprocess the historical manuscript data; The target creator can refer to journalists, editors, commentators, column teams, content studios, official accounts of institutions, or other content producers with a stable writing style. Historical manuscript data refers to text, titles, leads, scripts, graphic descriptions, video narration, comic scripts, and other content data that the target creator has published, released, or internally reviewed within a certain period. The preset time range can be set according to business needs, such as the most recent three months, six months, one year, or three years, to limit the selection scope of historical manuscripts. Preprocessing refers to a series of operations on historical manuscript data, such as cleaning, deduplication, segmentation, anonymization, and format standardization, to make it a standard format suitable for feature extraction.

[0029] Specifically, the process involves retrieving all historical manuscript data of the target creator within a preset time range from the organization's internal database, content management system, or file storage. The preset time range can be set according to business needs, such as the most recent three months or six months. The historical manuscript data is then cleaned, removing garbled characters, residual advertisements, invalid links, duplicate paragraphs, editorial annotations, template headers and footers, and other useless information. The historical manuscript data is deduplicated, merging or removing completely duplicate, reprinted, and highly similar manuscripts. The historical manuscript data is then segmented, breaking down manuscripts into structural segments such as titles, introductions, body text, quotations, conclusions, image captions, and video narration, facilitating feature extraction by segment type. The historical manuscript data is anonymized, deleting or replacing personal identification information, contact information, internal approval information, and unpublished information that should not be included in the modeling process to ensure data compliance. Finally, the historical manuscript data is formatted uniformly, converting files from different sources (such as Word, PDF, HTML, etc.) into plain text, JSON, Markdown, or other machine-readable formats.

[0030] Historical writing style feature vectors are extracted from the processed historical manuscript data. These historical writing style feature vectors include at least: sentence length distribution, title structure features, vocabulary preference features, rhetorical features, emotional tone features, paragraph organization features, and theme expression features. Among them, the historical writing style feature vector can refer to the vector obtained by numerically encoding various extracted writing style features (such as sentence length, title structure, etc.), which is used to represent the stable expression habits of the target creator.

[0031] Specifically, the processed historical manuscript data is used to extract sentence length distribution, title structure features, lexical preference features, rhetorical features, emotional tone features, paragraph organization features, and theme expression features. Sentence length distribution refers to the distribution characteristics obtained by statistically analyzing average sentence length, median sentence length, proportion of long sentences, proportion of short sentences, and frequency of punctuation usage. Title structure features can be determined by analyzing title length, whether numbers are used, whether interrogative sentences are used, whether parallel structures are used, whether main and sub-headings are used, and whether the subject and result of the event are included. Lexical preference features can be determined by statistically analyzing high-frequency content words, common verbs, common conjunctions, common institutional expressions, and the density of domain-specific terminology. Rhetorical features can be determined by identifying the frequency of rhetorical devices such as parallelism, rhetorical questions, metaphors, contrasts, and quotations. Emotional tone features can be determined by using a sentiment analysis model to determine whether the overall expression is formal, restrained, passionate, colloquial, commentary, or narrative. Sentiment analysis models can be developed by collecting large amounts of text data labeled with sentiment categories, dividing the labeled data into training, validation, and test sets, and using the cross-entropy loss function with a small learning rate (e.g., 2e-5) to fine-tune the model (using a pre-trained language model, such as BERT or RoBERTa, with a classification head added on top) to optimize the parameters of the classification head. Data augmentation (such as synonym replacement) can be used during training to improve robustness. Accuracy, F1 score, and other metrics are evaluated on the test set, and the model is deployed to the system once these metrics are met. Paragraph organization features can be determined by analyzing the length of the introduction, the order of factual development, the location of background information, the proportion of quoted paragraphs, and the way the conclusion is summarized. Thematic expression features can be determined by analyzing the organization of elements such as characters, events, data, scenes, conflicts, and solutions.

[0032] Features of different dimensions are processed through normalized encoding. Continuous numerical features (such as sentence length) can be normalized using min-max normalization or Z-score standardization; categorical features (such as sentiment tone) can be encoded using one-hot encoding, label encoding, or embedding encoding; and textual semantic features can be encoded using a pre-trained language model to obtain semantic vectors. The normalized sub-vectors are then concatenated or weighted according to preset weights to form the final historical writing style feature vector V. If a weighted fusion method is used, the weights corresponding to each sub-vector can be set according to different institutions or media types, or can be automatically updated based on historical generation results. In addition, to avoid one type of feature excessively dominating the writing style vector, the values ​​of each weight can range from 0.05 to 0.30, and the weights of at least three sub-vectors are greater than 0.10 to ensure that writing style modeling does not rely solely on a single feature.

[0033] Using the historical writing style feature vector and the task requirements of the content to be generated as constraints, a first draft of multimedia content that retains the writing style features of the target creator is generated.

[0034] The task requirements can refer to constraints such as the theme, audience, channel, length, and timeliness of the content to be generated. The initial draft of multimedia content can refer to the initial content generated by a large model based on historical writing style feature vectors and task requirements, which retains the writing style characteristics of the target creator. This can include text, scripts, titles, illustrations, video storyboards, etc.

[0035] Specifically, when generating the initial draft of multimedia content, historical writing style feature vectors can be converted into cue word constraints, style control parameters, or directly injected as part of the retrieval enhancement context (RAG) into cue words that the large model can accept. For example, the following style control information can be generated based on historical writing style feature vectors: titles are somewhat formal but result-oriented; the lead prioritizes the time, place, subject, and core event; each paragraph of the body text is controlled to be between 80 and 150 words; and key facts are connected using common short sentences. Integrating task requirements, the theme, audience, channel, length, and timeliness requirements of the content to be generated are added as additional constraints and input along with the large model's cue words.

[0036] The large model is invoked, employing a generative model that supports style control, such as a Transformer-based language model. The prompt explicitly requires "preserving the target creator's writing style characteristics" and provides specific style constraints. Based on these constraints and task requirements, the model generates a preliminary draft of multimedia content that meets the requirements. This draft can be text, scripts, titles, image captions, video storyboards, etc., with the specific form depending on the channel and media type specified in the task requirements.

[0037] The input data and training dataset used in this embodiment are all from compliant sources. Historical manuscript data can come from copyrighted content owned by the organization, legally authorized publicly available content, or business data anonymized by the user. Personal information, internal approval information, contact information, or undisclosed sensitive information involved in the data has been anonymized before entering the style modeling process. The purpose of style modeling is to extract expressive structure and style statistical features, rather than to identify sensitive identity attributes of natural persons or to create violation profiles.

[0038] In this embodiment, through historical manuscript preprocessing, multi-dimensional style feature extraction, normalized encoding, and vector fusion, the text style that is difficult to call directly is transformed into a calculable, injectable, and updatable historical style feature vector, so that the initial draft of subsequent multimedia content can not only meet the task requirements but also maintain style consistency.

[0039] In step S120, the media type corresponding to the initial draft of the multimedia content is identified, a corresponding differentiated scoring template is loaded based on the media type, and the initial draft of the multimedia content is quantitatively scored and audited for compliance to obtain a structured review report. The revision results and release decisions are then determined based on the structured review report, including: Identify the media type of the initial draft of the multimedia content, wherein the media type includes at least one of print media, new media, video, photos, and comics; Among them, media type refers to the specific form of expression or product to which the content belongs.

[0040] Specifically, media type can be directly determined through the channel field in the task requirements, or it can be automatically identified through content structure. For example, when content includes fields such as title, introduction, body text, and layout columns, it can be identified as print media; when content includes fields such as cover title, body text, tags, and interactive prompts, it can be identified as new media; when content includes fields such as shot number, image, narration, subtitles, and duration, it can be identified as video; when content includes fields such as image description, shooting theme, composition description, and visual tags, it can be identified as photography; when content includes fields such as panels, dialogue, image description, and onomatopoeia, it can be identified as comics. For content containing multiple structures, it can be identified as a composite media type, and corresponding scoring templates can be loaded for each.

[0041] Based on the media type, a corresponding differentiated scoring template is loaded. A combination of base score and cumulative deduction is used to quantitatively score the initial draft of the multimedia content. The initial draft of the multimedia content is then subject to compliance audit, and a structured review report is output. The structured review report includes at least bonus points, scores for each dimension, details of deductions, risk markers, modification suggestions, and release decisions. The differentiated scoring templates are pre-configured templates for different media types, specifying scoring dimensions, deduction rules, risk items, thresholds, and output formats. Compliance audits conduct legal, ethical, and platform rule-based security checks on initial drafts of multimedia content, aiming to identify and prevent potentially risky content such as false information and copyright infringement. The structured review report is a machine-readable, clearly formatted report, typically in JSON format, containing at least the scores for each dimension, reasons for deductions, risk points, suggested modifications, and the final release decision.

[0042] Specifically, the system first uses AI to identify the media type of the manuscript, categorizing it into one of five major categories: print media, new media, video, photography, and comics. Each media type corresponds to an independent Prompt scoring template. The core differences between these templates lie in their differentiated scoring dimensions: Print media templates emphasize logic, accuracy of factual data, and relevance of the title to the text; while the weight given to innovation may be lower. New media templates emphasize timeliness, storytelling / internet appeal, and attractive titles; while the requirements for logical structure rigor may be relatively lenient. Video scoring dimensions include: storyboard rhythm, composition, sound effects and background music, editing smoothness, and visual impact. Photo scoring dimensions include: compositional aesthetics, use of light and shadow, color matching, thematic expression, and emotional tension. Comic scoring dimensions include: panel narrative, consistency of art style, dialogue, creative expression, and visual rhythm.

[0043] All parameters for each template, including scoring dimensions, deduction details, weighting coefficients, deduction limits, and output formats, are loaded into memory for subsequent use by the scoring engine. The specific scoring model formula is: Final Score = Base Score (10 points) + Subtotal of Scoring Dimensions - Total Deductions; where the base score is fixed at 10 points, representing a flawless manuscript. The subtotal of scoring dimensions is the sum of scores from all scoring dimensions, with a maximum total of 10 points. The total deductions are the sum of accumulated deductions from all deduction items, but each category has an upper limit; if the total deductions exceed 10 points, the final score is 0. Taking the template for print media or new media as an example, it has 5 scoring dimensions and 7 deduction items. Using a large language model, the initial draft of multimedia content is scored across these 5 dimensions: Timeliness (2 points): Primarily judges whether it's a first release / closely follows a trending topic. For example, first release on the entire internet: 2 points; delayed release: 0 points. Newsworthiness (3 points): Evaluates the value and social significance of the topic. For example, a topic that sparks public discussion: 3 points; a topic that raises questions: 0 points. Storytelling (3 points): Evaluates narrative techniques and emotional resonance. For example, textbook-level narrative: 3 points; pure document copying: 0 points. Innovation (1 point): Evaluates the use of technology and visual design. For example, using AI generation, data visualization, or other new technologies: 1 point; outdated format: 0 points. Logic (1 point): Evaluates structure and writing style. For example, a rigorous structure and concise, precise writing: 1 point; illogical writing: 0 points. The above rules are compared against a large language model, and a score and corresponding comments are given for each dimension. The analysis of the article's performance in each dimension is presented in at least 80 words. Highlights are specifically pointed out, and shortcomings are expressed tactfully.

[0044] The manuscript was thoroughly checked through seven categories of deductions. For each issue found, points were deducted according to the deduction criteria and judgment points. The categories are: Language and Writing Standards (misspellings, grammatical errors, incorrect punctuation, and grammatical mistakes), with a maximum deduction of 3 points per item. News Elements (missing or incorrectly stating basic news elements such as time, place, people, cause, process, and result of an event), with a maximum deduction of 3 points per item. Content Quality (empty, repetitive, off-topic, insufficient information, or lack of in-depth analysis), with a maximum deduction of 4 points per item. Logical Structure (disordered paragraph transitions, broken causal relationships, weak argumentation, or unclear hierarchy), with a maximum deduction of 4 points per item. Factual data category: This refers to instances where the data, cases, citations, historical facts, etc., cited in the manuscript contain errors, inaccuracies, or have not been verified. A maximum of 4 points will be deducted for each category, with a maximum deduction of 4 points. Title / Text category: This refers to issues where the manuscript's title does not match the content, is exaggerated, ambiguous, unattractive, or violates title guidelines. A maximum of 3 points will be deducted for each category, with a maximum deduction of 3 points. Timeliness category: This refers to instances where the events or information covered in the manuscript are outdated, do not reflect the latest developments, or do not meet the requirements for current publication. A maximum of 3 points will be deducted for each category, with a maximum deduction of 3 points. Deduction accumulation rules: For the same category, multiple issues within the same category will accumulate deductions, not exceeding the maximum deduction for that category; for cross-category issues, deductions from different categories will be added together after each category is deducted.

[0045] For example, a manuscript might receive the following scores: Timeliness 1.5 (good timeliness, but not first publication); Newsworthiness 2.5 (unique topic, potential for virality); Storytelling 2 (emotional impact); Innovation 0.5 (some design sense); Logic 1 (rigorous structure); totaling 7.5 points. Two grammatical errors result in a 1-point deduction for language; a lack of a time element results in a 1-point deduction for news elements; and content duplication results in a 1-point deduction for content quality. The total deduction is 3 points. The final score = base score 10 points + 7.5 points - 3 points deducted = 14.5 points. Additionally, there is a rule that if the total deduction exceeds 10 points, the final score is 0. Other templates' corresponding scoring dimensions, deduction rules, weighting coefficients, deduction limits, and output formats can be developed using domain expert experience, historical data statistics, industry standards, and business testing; this example does not impose such restrictions.

[0046] The compliance audit scans each submission for violations of guidance, ethics, and platform rules. For guidance-related issues, any problems or errors in direction trigger a veto, resulting in submission termination and a final score of 0. A "3-5 point deduction" is added to the risk record for internal accountability statistics and does not participate in quantitative scoring. For infringement issues, the audit checks for plagiarism, misappropriation of others' work, infringement of portrait rights, reputation rights, and privacy rights. Plagiarism triggers a veto, resulting in submission termination and a final score of 0; other infringements incur a deduction of 2-4 points. For content violations, the review will examine whether there are prohibited visual materials, dangerous scenes, or other violations, and whether the content complies with public order and good morals. If prohibited visual materials are found, the final score for the submission will be 0, and an additional 3 to 5 points will be deducted from the internal risk statistics depending on the severity of the violation. For serious dangerous scenes, a veto will be triggered, the submission will be terminated, the final score will be 0, and "additional deduction of 3-5 points" will be marked in the risk record. For minor dangerous scenes, 1-2 points will be deducted, and a veto will not be triggered. For platform rules violations, the submission content must comply with the red lines explicitly prohibited by the target platform, such as infringing on the rights of others, inappropriate speech, or illegal marketing. Serious violations will result in a final score of 0, while minor violations will be handled according to the standard deduction rules, with a deduction of 1 to 2 points.

[0047] Output a structured review report in JSON format, including: Manuscript title: AAA; Basic information {Article type: New media, Source: ABC, Publication time: BBBB-CC-DD}; Compliance audit results {Audit conclusion: Pass, Risk issue list: []}; Bonus evaluation: {Timeliness: {Score: 1.5, Comment:...}, Newsworthiness: {Score: 2.5, Comment:...}, Storytelling: {Score: 2.0, Comment:...}, Innovation: {Score: 0.5, Comment:...}, Logicality: {Score: 1.0, Comment:...}, Subtotal of bonus dimensions: 7.5}; Deduction details {Problem category: Language and writing standardization, Deduction situation: ≥2 grammatical errors, Problem description: 2 grammatical errors in paragraph 3, Deduction: 1, Modification suggestion: It is recommended to change “AA” to “CC”}; Subtotal of deduction dimensions: 1; Final score: 16.5; Overall evaluation: The manuscript is of excellent overall quality, with a novel topic and a strong story. There are only minor flaws in terms of language and writing style. It is recommended to revise it before publication. Publication suggestion: Revise and publish.

[0048] If the structured review report contains high-risk items, the release decision will be to prohibit release or require manual review. If the structured review report does not contain any high-risk items and the overall score is less than the preset release threshold, the initial draft of the multimedia content will be revised according to the modification suggestions, and the revised content will be re-evaluated and subjected to quantitative scoring and compliance audit. If the structured review report does not contain any high-risk items and the overall score reaches the preset release threshold, then the release decision is determined to allow release or allow printing.

[0049] The "correction result" refers to the new version obtained after revising the initial draft based on the modification suggestions in the review report. The "release decision" refers to the final action instruction issued by the system based on the review report, which includes four states: approved, revised and released, manually reviewed, and prohibited from release. High-risk issues are those discovered during the compliance audit process that have irreversible harm or extremely high directional risk. Once identified, the system determines that the risk cannot be eliminated through simple modifications and must directly prohibit release or submit to in-depth manual review. High-risk issues include the following four categories: Directional bias (veto); Infringement (plagiarism); Illegal content (prohibited visual materials and seriously dangerous scenes); Platform rules (serious violations). Compliance audit issues that meet any of the above criteria are considered high-risk issues, and the system directly outputs a decision to prohibit release or require manual review.

[0050] If the structured review report contains high-risk items, the release decision will be to prohibit release or require manual review. If the structured review report does not contain any high-risk items and the overall score is less than the preset release threshold, the initial draft of the multimedia content will be revised according to the modification suggestions, and the revised content will be re-evaluated and subjected to quantitative scoring and compliance audit. If the structured review report does not contain any high-risk items and the overall score reaches the preset release threshold, then the release decision is determined to allow release or allow printing.

[0051] The "correction result" refers to the new version obtained after modifying the initial draft based on the modification suggestions in the review report. The "release decision" refers to the final action instruction issued by the system based on the review report, which includes four states: approved, revised and released, manually reviewed, and prohibited from release.

[0052] Specifically, if the structured review report contains high-risk items, the publication decision is determined to be either a prohibition on publication or a manual review, regardless of the score. If there are no high-risk items, the overall score is checked. If the overall score is less than the preset publication threshold (e.g., 8.0), the initial draft of the multimedia content is revised according to the suggested modifications. The revised content is then re-evaluated using quantitative scoring and compliance auditing. This process is iterated until the revised content meets the pass standard, i.e., the overall score reaches the preset publication threshold, or it is determined to be subject to manual review or a prohibition on publication. If the overall score is greater than or equal to the preset publication threshold, the publication decision is determined to be either permitted for publication or permitted for printing.

[0053] In this embodiment, by identifying media types and loading corresponding differentiated scoring templates, a combination of basic scores and cumulative deductions is used for quantitative scoring and compliance auditing. Then, based on high-risk and comprehensive scoring thresholds, a release decision or a correction iteration is triggered. This achieves refined and automated quality and security assessment for various media types such as print media, new media, and video. It overcomes the shortcomings of existing AI proofreading, which is limited to typo checking and cannot perform differentiated in-depth scoring, and significantly improves the accuracy, efficiency, and security of the review and evaluation.

[0054] In step S130, based on the target distribution channel attributes, the audited content is reconstructed across channels in a multimodal manner, generating online digital channel distribution data and traditional offline channel publication data respectively. Simultaneously, an initial benchmarking matrix is ​​generated based on competing content on the same theme, including: Obtain content data and target distribution channel attributes after review and approval; Target distribution channels can refer to the platforms / media where the content is prepared for distribution or publication, such as mobile apps, social media platforms, self-media platforms, short video platforms, newspapers, periodicals, brochures, etc. Audited content refers to content that has passed AI-driven review and scoring, guidance audits, factual checks, and language quality checks, meeting the requirements for publication or printing. Content data includes not only the main text but also structured information such as titles, leads, body text, image captions, interviewees, news elements, factual data, sources, author information, review scores, revision suggestions, and relevant tags.

[0055] Specifically, the process involves acquiring the content data after review and approval, including: title, body text, summary, images, video footage, interview information, news elements, review report, and revised version. The channel name is parsed from the task requirements, and all attribute tags for that channel are loaded from a pre-defined channel attribute library, such as channel type, word count requirements, style requirements, title length, cover image specifications, whether hashtags are needed, and whether video scripts are required.

[0056] When the target publishing channel is an online digital channel, the content data is reconstructed into one or more of the following based on the target publishing channel attributes: short text, long text, title group, topic tag group, cover prompt, video storyboard, image description, or short video script. Cross-channel multimodal reconstruction refers to transforming the same source material into different media formats based on the characteristics of different channels. For example, the same press release can be reconstructed into a long article on social media platforms, a short post, a product recommendation graphic, a short video script, or a newspaper article. Online digital channels refer to channels distributed via the internet, such as mobile clients, public platforms, social media platforms, and self-media platforms.

[0057] Specifically, the system first parses the approved content data, such as the main text, title, abstract, core facts, and source materials, into a structured, platform-independent intermediate representation. This intermediate representation includes core factual elements, the main theme and stance, source materials, and style constraints. Core factual elements include: time, place, people, events, causes, results, data, and quotations. The main theme and stance include: core viewpoints and directions that must be maintained consistently. Source materials include: metadata for images, video clips, audio, charts, etc. Style constraints include: constraints derived from historical writing style feature vectors and the review report.

[0058] The system will identify the attributes of the target publishing channel, including: platform name; content genre preference, such as long articles with images, short texts, product recommendation notes, short videos, live streaming scripts, etc.; word / duration limits, such as 140 characters, 1000 characters, 60 seconds for videos, etc.; expression style preference, such as public platforms preferring depth and online appeal, social media platforms preferring brevity and trending topics, and self-media platforms preferring lifestyle and visual appeal, etc.; and functional features, such as hashtags, cover images, and comment section interaction, etc.

[0059] For example, if the target channel's restructuring logic is to generate a well-structured tweet, the output would be: a long text containing an engaging introduction, segmented body text (e.g., introducing exhibition highlights, key artifacts, curator interviews, and a well-structured conclusion). The title group could include a main title and a subtitle, such as "Breaking News! A Major Annual Exhibition 'B' Opens Today – Don't Miss These 5 Treasures!" The cover image would be a high-resolution, impactful picture of an artifact, accompanied by textual descriptions. Image descriptions would generate detailed explanatory text for each image in the text, including the artifact's name, era, and symbolic meaning.

[0060] If the target channel's restructuring logic is to extract the most essential and shareable information, then the output should be a short text containing the core information within 140 characters, such as "[#New Exhibition at Museum A#] Exhibition 'B' opens today! C is on display for the first time—it's amazing! Come check it out!" The hashtag groups are: #MuseumA# #ExhibitionD# #CCulturalRelics#. Image description: Select 1-2 of the most eye-catching images of cultural relics, along with brief descriptions.

[0061] If the target channel's restructuring logic emphasizes "product recommendation" and a lifestyle perspective, then the output should be short / long text, narrated from a first-person or visitor's perspective, recounting the exhibition experience, such as "Sisters! This new AA exhibition is so photogenic! These artifacts are breathtaking!" The title should be something like, "BB's Weekend Getaway | AA's New Exhibition YYDS! Photo Tips Included." The cover image should feature a beautifully composed, atmospheric photo of the exhibition, along with phrases like "Photo Tips" or "Must-See." The hashtags should include: #BBTravel #Museum #ExhibitionOutfit #PhotoSpots If the target channel's restructuring logic is to transform textual content into audiovisual language, then the output video storyboard script is: a table containing "shot number, shot type, scene description, narration / subtitles, and duration." Shot 1 (close-up, 2 seconds): A close-up of the exquisite details of an artifact, narration: "Have you ever seen a 'DEF' from two thousand years ago?" Shot 2 (medium shot, 5 seconds): The camera pans across the entire exhibition hall, narration: "At the AA Museum, this new exhibition takes you on a journey through time." Shot 3 (close-up, 3 seconds): The audience's amazed expression, narration: "After watching, all I can say is, the aesthetics are amazing!" Short video script: A complete video script, including an opening, climax, ending, and interactive prompts.

[0062] When the target distribution channel is a traditional offline channel, the content data is converted into structured page data or digital proof files based on the print media layout template; Traditional offline channels refer to those primarily based on physical publishing or traditional typesetting, such as newspapers, periodicals, print special issues, and brochures. Print media layout templates refer to the format templates used before print publication, specifying page layout, number of columns, title font size, text area, image area, lead text location, and author signature location. Structured page layout data refers to data structures recognizable by typesetting systems, such as page numbering, manuscript blocks, title blocks, image blocks, coordinates, fonts, font sizes, line spacing, columns, and image captions. Digital proof files refer to electronic proof files used in traditional publishing, typically in the form of PDFs, page preview files, or pre-print confirmation files.

[0063] Specifically, a print media layout template is a pre-defined, structured configuration file that defines various physical and style parameters for the newspaper or periodical layout, such as: Page size: A3, A4, double-page spread, etc. Number of columns: e.g., 5 columns, 6 columns. Column width: Width of each column (mm). Heading font size and style: Font size and style for main title, subtitle, and introduction, such as SimSun or Heiti. Body text font size and style: Font size and style for body text, figure captions, and author signatures. Image position and size: Reserved area for the image, whether it can be cropped, and its wrapping style with text. Column name: The page it belongs to, such as the name and style for news, culture, and people's livelihood.

[0064] The content data is broken down into basic elements such as titles, introductions, body paragraphs, images, and captions. Based on the column widths and font sizes defined in the template, the system calculates the number of lines and space occupied by each paragraph and each title. For example, how many lines would an 800-word body text occupy in a 5-column, 10-point font template? Intelligent typesetting is implemented, including: title layout (automatic centering, left alignment, or column placement of main and subtitles); body text flow (text automatically flows from the bottom of one column to the top of the next); image and text wrapping (images are inserted in specified positions according to template rules, and text automatically wraps around them); and pagination and column division (if the content exceeds one page, the system automatically performs pagination or column division).

[0065] The output is structured layout data, in a machine-readable format (such as XML or JSON), which details the position, size, content, font, etc. of each element on the page. This data can be read and further refined by professional typesetting software, such as Adobe InDesign. The digital proof file is usually in PDF format, which already has the final typesetting effect and can be directly sent to the printing press for CTP (computer-to-plate) and printing.

[0066] Obtain competitor content with the same theme or event as the content data, and perform source de-identification processing on the content data and the competitor content; Competitive content can refer to other media content that competes with our content on the same theme or event for dissemination effectiveness. Content on the same theme can refer to competing articles discussing the same topic, such as both reporting on "low-altitude economy," "urban renewal," or "cultural tourism activities." Competitive content on the same event can refer to competing articles reporting on the same news event, such as the same press conference or the same breaking news. Source de-identification refers to hiding or replacing identity information such as the source, media name, author, column, and platform of the article before evaluation to avoid bias from the model or reviewers due to brand recognition.

[0067] Specifically, based on the theme, keywords, subject of the event, and time frame of the articles submitted by the organization, a search is conducted within a pre-defined competitor media database. This can be achieved through public platform scraping, using legitimate APIs to obtain content on the same topic from media platforms, public platforms, news websites, and other public channels. The internal database is a competitor article repository maintained by the organization itself.

[0068] The content data and competitor content underwent source de-identification processing: Media names / logos were removed, such as "AA Daily." Author byline, including reporter names and pen names. Account IDs / avatars were removed, such as public platform IDs, platform nicknames, and platform accounts. Brand identifiers were removed, including unique column names and brand slogans found in the articles. Follower / readership information was removed, i.e., any data that could reflect the account's influence.

[0069] Replace "According to our reporter ZZZ" with "According to the reporter's report" using text replacement. Convert all articles to plain text or a uniform HTML format, removing the original layout styles. Assign each article a unique, random ID, such as article A_001, article B002, instead of using the original name.

[0070] The final result is a folder or list containing multiple anonymous manuscripts. At this point, neither human reviewers nor AI models can determine from the manuscript itself whether it comes from a major, authoritative newspaper or a local tabloid. A "double-blind test" environment is created, applying the same treatment to both the author's own manuscripts and those of all competitors, removing all information that could suggest the source.

[0071] Based on a pre-set benchmarking evaluation template, the de-identified competitor content is evaluated to obtain an initial benchmarking matrix.

[0072] The pre-set benchmarking evaluation template refers to a standardized evaluation template used for scoring competitor comparisons, specifying comparison dimensions, scoring levels, output format, and evaluation criteria. This template includes dimensions such as timeliness and trending topics, newsworthiness and guidance, storytelling and expression, innovation and approach, and dissemination effect. The initial benchmarking matrix refers to the tabular scoring results after a static comparison of the organization's content with competitor content. It is typically presented in a matrix format of "article × evaluation dimension," such as the score, level, and brief review for each article in dimensions such as timeliness, newsworthiness, storytelling, innovation, and dissemination potential.

[0073] Specifically, through expert-level AI review, a pre-set evaluation template is loaded, clearly defining the evaluation objectives, dimensions, and standards. The AI ​​model or human experts, in a double-blind manner, independently evaluate each anonymous manuscript according to each dimension in the template, providing a rating and comments. An initial benchmarking matrix is ​​generated: the evaluation results are summarized into a structured table, i.e., the initial benchmarking matrix.

[0074] For example, the rating levels can be S, A+, A, A-, B+, B, corresponding to scores from 5.0 to 2.5, as shown in Table 1 below.

[0075] Table 1

[0076] The initial benchmarking matrix above visually demonstrates the gap between our organization's articles and those of our competitors across various dimensions. It clearly identifies the weaknesses in our articles. For example, our organization lags behind competitor 1 in "Timeliness and Hot Topics," but slightly surpasses it in "Newsworthiness and Guiding Principles." Our scores are low in "Innovation and Approach."

[0077] In this embodiment, by performing cross-channel multimodal reconstruction of the approved content based on the target distribution channel attributes, it automatically generates short texts, video scripts, and other forms of content adapted to online digital channels, as well as structured layout data for traditional offline channels. At the same time, it generates an initial benchmarking matrix based on de-identified competitor content on the same theme. This achieves one-click intelligent adaptation of content from the original draft to multiple platforms and multiple modalities, and objective benchmarking against competitors. It overcomes the shortcomings of existing technologies that require a lot of manual rewriting and typesetting, which can easily lead to a deviation in the main theme. This significantly improves cross-media distribution efficiency, content consistency, and competitive insight capabilities.

[0078] In step S140, after the online digital channel data release runs for a preset period, restricted interaction data and operational briefing data from each online release channel are collected. Dynamic baseline alignment processing is performed on the restricted interaction data to obtain a relative utility index. The initial benchmarking matrix is ​​then updated based on the operational briefing data to obtain a dynamic benchmarking evaluation result, including: After the data is published on the online digital channels, restricted interactive data that flows back from each online publishing channel is collected according to a preset collection cycle. Online digital channel release data can refer to the specific content finally published on different online platforms after reconstructing the same master draft according to their different characteristics. Examples include: short text, long text, title groups, hashtag groups, cover prompts, video storyboards, image descriptions, or short video scripts (one or more of these). Preset collection cycles refer to a fixed time window for waiting for data feedback. For example, 1 hour, 6 hours, 24 hours, or 72 hours after publication, depending on the content type; news flashes have shorter cycles, while in-depth reports have longer cycles. Restricted interaction data refers to coarse-grained metrics that third-party platforms are only willing to feed back due to data security and commercial barriers. Examples include: page views, likes, comments, shares, completion rate, read completion rate, follower conversion rate, and negative feedback.

[0079] Specifically, the collection cycle is configured according to the content type. For example, for breaking news, data is collected 10 minutes, 30 minutes, and 1 hour after publication. For in-depth reports, data is collected 6 hours, 24 hours, and 72 hours after publication. For platforms with open APIs, the system periodically calls the data interface to obtain data such as views and likes for specified content after OAuth2.0 authentication. For platforms that do not provide real-time APIs (such as some mini-programs), the system can periodically download Excel reports manually exported by operations personnel. The collected restricted data is usually a JSON or CSV file containing the following fields: views, likes, comments, favorites, shares, completion rate, viewing duration, conversion rate (number of new followers due to content), and negative feedback (such as dislikes, reports, and blocks).

[0080] Obtain online new media operation briefing data from this organization and its competitors, and extract dissemination performance, content structure, user feedback, and competitor dissemination characteristics from the online new media operation briefing data; The operational report data refers to the operational reports regularly compiled by this organization or competitors, typically including summaries of dissemination performance in tabular or text format, such as daily operational reports, topic reviews, platform rankings, and competitor observation reports. Dissemination performance can refer to macro-level data recorded in the operational report, such as readership, interaction rate, and fan growth. Content structure can refer to structural information that may be included in the report, such as title type, body length, image-to-text ratio, and video length. User feedback can refer to popular opinions, sentiments, and questions raised by users in the comments section. Competitor dissemination characteristics can refer to the unique strategies adopted by competitors in this release, such as release time, title style, and interaction methods.

[0081] Specifically, the system connects to the organization's internal daily operations report system, automatically reading the .docx or .xlsx files generated that day. Competitive analysis documents can be submitted through public channels or purchased from third-party monitoring services, or manually uploaded by operations staff.

[0082] The dissemination performance, content structure, user feedback, and competitor dissemination characteristics are extracted from the online new media operation report data. For structured data, such as tables, cells are directly parsed and mapped to predefined fields. For example, if the table columns are named "Title," "Read Count," and "Likes," they are directly extracted. For unstructured text, such as paragraph descriptions, information extraction is performed using large-scale models or traditional NLP techniques. Dissemination performance can be determined by extracting numerical values ​​such as "This article has over 100,000 reads and a 5% interaction rate." Content structure can be determined by extracting descriptions such as "The title uses a question format, and the body contains three sub-arguments." User feedback can be determined by extracting information such as "The focus of discussion in the comments section is on XX, with 70% positive sentiment." Competitor dissemination characteristics can be determined by extracting information such as "Competitor A published at 8 AM, the title contains numbers, and it used a short video format."

[0083] Using historical baseline data from various online publishing channels, average data of similar content, or average data from platforms within the same time window as dynamic baselines, the restricted interaction data is normalized, denoised, and processed for outliers. Among these, dynamic baseline refers to a dynamically calculated reference value used to eliminate inherent traffic differences across platforms. Historical baseline data refers to the average performance of the same platform, account, and content type over a period of time, such as 30 days. Average performance data for similar content refers to the average performance of other content on the same platform with a similar theme to the current article. Average performance data for the same platform within a similar time window refers to the average performance of all content on the same platform within a similar publication time window, used to eliminate the influence of time factors, such as holidays and peak hours.

[0084] Specifically, calculate the dynamic baseline: the historical baseline, which is the average readership of all similar content from the same account over the past 30 days, such as news about people's livelihood. and standard deviation Average of similar content: The average number of views for all content on the same topic on the same platform on that day. The average reading time across platforms during the same time window is calculated as the average reading time of all content on the same platform between 8 and 9 AM on the same day. .

[0085] The restricted interactive data can be normalized by ratio normalization or Z-score standardization.

[0086] Ratio normalization is given by R = X / B, where X is the original value and B is the selected baseline. For example, if the number of reads is 15,000 and the historical baseline is 10,000, then R = 1.5, indicating that the reading volume is 50% higher than the baseline. Z-score standardization involves first taking the natural logarithm of the original data X, resulting in log(1+X), and adding 1 to prevent the logarithm from being unavailable when X=0. The mean μ and standard deviation σ of this batch of logarithmic data are then calculated. Finally, the formula is applied... The Z-score is calculated for each data point. Z=0 indicates that the data is exactly at the average level. Z>0 indicates that it is above the average level; for example, Z=1.5 means it is 1.5 standard deviations above the average level. Z<0 indicates that it is below the average level. By standardizing the Z-score, previously vastly different pageviews (e.g., 100 and 10000) become comparable relative values, no longer affected by the platform's overall traffic volume. Significantly abnormal data points in the restricted interaction data are removed. For example, if a platform suddenly experiences a 100-fold surge in pageviews due to bots, the system will detect the abrupt change using a sliding window and mark it as noise. A quantile truncation method is used to replace values ​​above the 99.5th percentile or below the 0.5th percentile in the restricted interaction data with the corresponding quantile value, preventing extreme values ​​from distorting the overall analysis.

[0087] The relative utility index is calculated based on the processed restricted interaction data. The relative utility index is used to characterize the content dissemination effect after excluding differences in traffic allocation across different platforms. The relative utility index is a comprehensive, standardized metric used to measure the actual dissemination effect of content after excluding differences in platform traffic allocation. A higher value indicates a better effect from the content itself.

[0088] Specifically, the system assigns an importance weight to each dissemination metric. For example, for short video platforms, the completion rate might be weighted at 0.3, likes at 0.2, shares at 0.2, comments at 0.15, and negative feedback at 0.15. These weights can be derived from business experience or regression analysis of historical data.

[0089] To calculate the score for a single platform, for content on the same platform, multiply each indicator value (after baseline alignment, normalization, and denoising) by its respective weight, sum them, and then subtract the weighted sum of negative indicators (negative feedback numbers) to obtain a score for that platform. This score represents the component of the relative utility index on that platform.

[0090] If the content is published on multiple platforms, the importance of each platform also needs to be considered. For example, the weight of a public platform is 0.4, a self-media platform is 0.3, and a social media platform is 0.3. Multiply the relative utility index of each platform by its corresponding channel weight, and then sum them up to obtain the final total relative utility index.

[0091] The initial benchmarking matrix is ​​updated based on the operational briefing data and the relative utility index to obtain dynamic benchmarking evaluation results.

[0092] Among them, the dynamic benchmarking evaluation results can refer to the benchmarking matrix that is updated after integrating the actual dissemination data after release, namely the relative utility index and the operational briefing data, and is more in line with the real market performance.

[0093] Specifically, the calculated relative utility index is added to the initial benchmarking matrix as a new evaluation dimension. Simultaneously, user feedback data extracted from operational reports, such as the proportion of positive sentiment in the comments section, the volume of discussion, and competitor dissemination characteristics (e.g., whether competitors used video formats and whether their release times were more optimal), are also included as new dimensions.

[0094] Adjusting existing dimension scores: Some static dimensions (such as "title attractiveness") may differ from actual click-through rates. If an article's title has a low static score but a high actual click-through rate, the system will adjust the score for that dimension accordingly, and vice versa. The adjustment amount depends on the gap between the static score and the dynamic data, as well as the sample size.

[0095] The initial static matrix and the newly added dynamic propagation matrix are merged in a certain proportion. For example, when the data is first fed back, the reliability of the dynamic data is not high, with static scores accounting for 80% and dynamic scores accounting for 20%; as data accumulates, the proportion of dynamic data gradually increases to 60% or even higher.

[0096] Generating Dynamic Benchmarking Evaluation Results: The fused matrix is ​​a dynamic benchmarking evaluation result. It includes both static judgments of "how well this article is written" and dynamic verifications of "how widely this article actually circulated." For example, the matrix might show that our organization's article scored 4.0 in "title attractiveness" statically, but its dynamic click-through rate was as high as 1.3 (30% higher than the baseline), indicating that the title was actually very effective; while the competitor scored 4.5 in "content depth" statically, but its completion rate was only 0.8, indicating that although the content was deep, users did not have the patience to read it all.

[0097] In this embodiment, limited interactive data and operational briefing data from online publishing channels are collected, and the relative utility index is calculated after normalization, noise reduction, and outlier processing using historical baselines, similar averages, or platform averages as dynamic baselines. This eliminates the interference of differences in traffic distribution across different platforms on the measurement of dissemination effects, and overcomes the shortcomings of existing technologies where third-party platforms only return coarse-grained data and traditional print media cannot return data, thus making it impossible to objectively assess the true effects. This achieves accurate quantification of the actual dissemination effectiveness of the content, and based on this, the initial benchmarking matrix is ​​updated to obtain dynamic benchmarking evaluation results, providing reliable empirical evidence for subsequent attribution analysis and template iteration.

[0098] In step S150, the relative utility index is used as the dependent variable, and the historical writing style feature vector, the scoring matrix in the structured review report, and the dynamic benchmarking evaluation results are concatenated and aligned into a fine-grained style feature matrix as the independent variable. The independent variable and the dependent variable are input into the feature association evaluation model for attribution mapping analysis to generate a style preference weight map. The online style template library and the traditional print media writing style constraint library are updated according to the style preference weight map, including: The historical writing style feature vector, the scores of each dimension in the structured review report, the deduction details, the risk markers, the modification suggestions, and the benchmarking scores in the dynamic benchmarking evaluation results are aligned by fields and time windows to obtain a fine-grained style feature matrix. The fine-grained style feature matrix refers to a two-dimensional table formed by aligning the writing style vector, review matrix, and benchmarking results according to the same content, version, and time window. Each row represents a published version of the content, and each column represents a specific style or rating feature.

[0099] Specifically, the system obtains the following data sources: historical writing style feature vectors: N-dimensional floating-point arrays for single manuscripts (vocabulary preference, sentence structure, sentiment, narrative structure); structured review reports: multi-field JSON (dimensional scores, deduction details, risk markers, revision suggestions, manuscript type tags); and dynamic benchmarking evaluation results: one-dimensional score arrays for competitor benchmarking (timeliness score, content depth, creativity score).

[0100] A globally unique primary key, the manuscript ID, is added as the core for all data association; unified category tags and publication channel tags are used based on manuscript type; all scores are uniformly mapped to the standard range of 0-10, and the vector dimensions are uniformly floating-point precision; risk markers are uniformly coded (0 no risk / 1 mild / 2 significant public reaction / 3 risk to guide direction), and the reasons for deductions are converted into category codes. Using the manuscript ID as the association key, the three types of data are horizontally concatenated, with one row representing one manuscript, and the columns being: all dimensions of the writing style vector, scores for each review dimension, deduction item codes, risk codes, modification suggestion feature tags, and competitor benchmark scores, eliminating field heterogeneity.

[0101] Set a fixed statistical window, such as 7 days after the publication of a single article as a complete observation period; use the publication timestamp of the article as the benchmark to filter all the preceding feature data of the article within the same time window; remove mismatched data across windows, such as if an article is revised and re-examined in the middle, and only retain the data of the final approved version that matches the corresponding period of dissemination data; attach a matrix to the time tag uniformly to distinguish the short-term / long-term dissemination effect features later.

[0102] After alignment, a two-dimensional structured matrix is ​​generated, where the rows are individual multimedia articles (samples); the columns are all subdivided features (writing style subdivisions, review items, deduction tags, risk levels, and competitor benchmarking items); the granularity is refined to the smallest unit, no longer a general total score, and can accurately locate the impact of "a certain word usage habit, a certain deduction, or a certain competitor weakness" on dissemination.

[0103] Using the relative utility index as the dependent variable and the fine-grained style feature matrix as the independent variable, attribution mapping analysis is performed through a feature association evaluation model. Among these, feature association assessment models refer to mathematical models used to calculate the relationship between independent and dependent variables, such as linear regression, gradient boosting trees, and causal forests, which can output the direction and magnitude of each independent variable's contribution to the dependent variable. Attribution mapping analysis refers to calculating the direction and degree of influence of each independent variable on the dependent variable.

[0104] Specifically, the aligned fine-grained style feature matrix is ​​used as the independent variable X; the relative utility index dependent variable Y, calculated for each manuscript within the same time window, is bound to the features in the Nth row of the matrix, corresponding to the dissemination utility index of the Nth manuscript.

[0105] Unlike simple correlation statistics, a large-scale causal inference model is used to complete the mapping: First, the positive and negative correlation between each sub-feature and the relative utility index is calculated; irrelevant interference factors such as channel traffic and release time are eliminated to avoid misjudging "false associations"; the fluctuation range of the dissemination effect (relative utility index) when a single feature changes alone is calculated to distinguish between features that "only appear synchronously" and features that "truly affect dissemination"; a causal mapping relationship table is output to record the positive / negative effect logic of each style / review / benchmarking feature on the dissemination effect.

[0106] Based on the results of the attribution mapping analysis, the correlation weights between different writing style elements, scoring dimensions, benchmarking dimensions and dissemination effects are determined, and a style preference weight map is generated. Among them, the style preference weight map is a result of the attribution analysis. It is a structured map where nodes are various style elements, rating dimensions, and benchmarking dimensions, and edges are the correlation weights between them and the dissemination effect. Positive values ​​indicate positive promotion, and negative values ​​indicate negative inhibition.

[0107] Specifically, based on the causal effect value output by the model, a weight coefficient of 0 to 1 is generated through normalization: the higher the weight, the more significantly the element enhances the dissemination effect (such as short video titles, emotional expression on self-media platforms, and in-depth investigative content); a negative weight indicates that the element reduces the dissemination effect (such as lengthy introductions in print media, stiff official language, and factual deductions). The weights are divided into three major weight pools: writing style element weights, including sentence structure, vocabulary, emotion, and narrative structure; review dimension weights, including compliance score, logic score, visual score, and negative impact of deductions; and benchmarking dimension weights, including timeliness advantage, content depth advantage, and creative gap weights.

[0108] Generate a dual-format style preference weight map: a structured data map (machine-readable in the backend) and a standardized weight data table containing: feature name, category, weight value, positive and negative impact, and applicable channel tags, for the system to automatically read and update the template library. A visual representation of the map (for manual operation viewing) includes: Multi-dimensional radar chart: weight distribution of each dimension of a single channel; heat map: color blocks distinguishing high and low feature weights; correlation network diagram: showing the mutual influence between writing style, rating, and competitor indicators; intuitively presenting which content features users currently prefer on each channel and which defects will significantly reduce dissemination data.

[0109] Update the cue word constraints in the online style template library based on the style preference weight map, and adjust the multimodal reconstruction strategy of the online digital channels; The online style template library stores prompt word templates, formatting rules, and style parameters used when generating content for different online platforms. Prompt word constraints refer to the rule-based Prompt instructions input to the large model, limiting the style, length, tone, compliance red lines, and channel adaptation requirements of the generated content, ensuring the model output does not deviate from channel standards. The multimodal reconstruction strategy refers to the standardized rules for automatically converting the master draft for online channels, including text splitting, short video script generation, image tagging, layout formatting, and title rewriting logic, dynamically adjusted according to channel characteristics.

[0110] Specifically, the constraints on prompts in the online style template library are updated, and the weights of the channels are split: High-weight positive features specific to each online channel, such as social media platforms, self-media platforms, and short video platforms, are extracted; existing Prompt constraints are rewritten: High-weight positive features are reinforced in the prompts (e.g., for short video channels, "titles should be limited to 15 characters, and the opening 3 seconds should directly address the hot topic" has a high weight, so this constraint is forcibly added to the short video Prompt); Low-weight negative features are subject to restrictions and prohibitions (e.g., lengthy written language has a negative weight, so a new constraint "avoid long, written sentences" is added). Channel-specific updates are implemented: each new media template library iterates independently, without sharing the same set of prompt rules.

[0111] Multimodal reconstruction is the rule for automatically converting master drafts into short videos, images and text, and short copy. Based on a weighted graph, the conversion logic is modified as follows: Text reconstruction: High-weight short sentences and internet-savvy words are automatically extracted first, while low-weight long paragraphs are automatically deleted; Video scene reconstruction: If the "first 30 seconds of conflict shots" has a high weight, the scene splitting rules are adjusted to prioritize placing the core conflict at the beginning; Image and text tag reconstruction: High-popularity emotional tags and trending tags are automatically matched first, and inefficient tags are replaced; Distribution adaptation rule iteration: High-weight content formats are automatically given a higher distribution weight, while low-efficiency formats are given a lower automatic generation ratio.

[0112] The weights related to print media topic selection, title, lead, structure, and expression style in the style preference weight map are mapped to the traditional print media writing style constraint library, and the prompt word parameters when generating traditional print media manuscripts are adjusted.

[0113] The traditional print media writing style constraint library can refer to the storage of prompt word parameters used when generating print media manuscripts, including constraints such as headline specifications, lead structure, body text density, and language style. Prompt word parameters can refer to adjustable quantitative parameters within the underlying Prompt of the print media model, including adjustable numerical configurations such as sentence formality, proportion of in-depth discussion, headline rigor level, lead information density, and material selection rules.

[0114] Specifically, the system filters and matches five key features of print media from a complete weighted graph: topic selection, headline, lead, article structure, and formal written expression, while removing online-specific style weights. Online dissemination data reflects the true preferences of the general public, and the high-weight, high-quality online content features are transformed into rules that align with the rigorous publishing standards of print media. For example, the high weight of "short leads that get straight to the point" online corresponds to print media constraints: leads should condense redundant preambles and directly state the core facts at the outset; the good dissemination effect of in-depth online analysis content corresponds to the weight of print media topic selection, increasing the priority of generating in-depth investigative topics.

[0115] The underlying prompt parameters of the print media model were adjusted, and the print media writing style constraint library stores quantifiable Prompt parameters. Numerical configurations were modified based on mapping weights: Topic selection parameters: Increase the probability of generating topics with high dissemination depth and reduce the weight of vague and shallow topics; Title parameters: Balance formality and information density, strengthening high-weight, concise, and information-rich titles; Lead / Structure parameters: Adjust paragraph length and the proportion of argumentation; Expression style parameters: Fine-tune the degree of formality and the proportion of material citations, avoiding stiff and lengthy expressions with negative online weights; When the system generates print media master drafts, the underlying Prompt has already absorbed the real dissemination preferences of online users, so print media articles no longer need to rely on offline paper reading data, but rely on new media distribution feedback to achieve model self-iterative optimization.

[0116] In this embodiment, the relative utility index is used as the dependent variable, and the historical writing style feature vector, the review and scoring matrix, and the dynamic benchmarking results are concatenated into a fine-grained style feature matrix as the independent variable. This matrix is ​​then input into the feature association evaluation model for attribution mapping analysis to generate a style preference weight map. Based on this map, the online style template library and the traditional print media writing style constraint library are updated respectively. This achieves an intelligent closed-loop iteration that accurately attributes the actual dissemination effect to specific writing style elements and scoring dimensions. This overcomes the defect in the existing technology that the content generation model cannot automatically evolve based on real user feedback. It enables the multimodal reconstruction strategy of online channels and the generation prompts of print media to be continuously optimized based on data-driven preference weights, significantly improving the adaptability of multimedia content production and the cross-media dissemination effect. Example

[0117] Figure 2 This is a framework diagram of a large-model-based intelligent generation, review, and closed-loop iteration system for multimedia content, provided in Embodiment 3 of the present invention. The system is used to execute the large-model-based intelligent generation, review, and closed-loop iteration method for multimedia content as described in any embodiment of the present invention. Figure 2 As shown, the system includes: The acquisition module 210 is used to acquire the task requirements of the content to be generated and the historical manuscript data of the target creator, extract the writing style features of the historical manuscript data, construct the historical writing style feature vector, and generate the first draft of multimedia content based on the historical writing style feature vector and the task requirements. The determination module 220 is used to identify the media type corresponding to the initial draft of the multimedia content, load the corresponding differentiated scoring template based on the media type, perform quantitative scoring and compliance audit on the initial draft of the multimedia content, obtain a structured review report, and determine the correction results and release decision based on the structured review report; The generation module 230 is used to perform cross-channel multimodal reconstruction of audited content based on the target publishing channel attributes, and generate online digital channel publishing data and traditional offline channel publishing data respectively, while generating an initial benchmarking matrix based on the content of competing products on the same theme; The data collection module 240 is used to collect restricted interaction data and operation briefing data from each online publishing channel after the online digital channel data has been running for a preset period, perform dynamic baseline alignment processing on the restricted interaction data to obtain a relative utility index, and update the initial benchmarking matrix based on the operation briefing data to obtain a dynamic benchmarking evaluation result. The analysis module 250 is used to take the relative utility index as the dependent variable, and the historical writing style feature vector, the scoring matrix in the structured review report and the dynamic benchmarking evaluation results as the fine-grained style feature matrix as the independent variable. The independent variable and the dependent variable are input into the feature association evaluation model to perform attribution mapping analysis, generate a style preference weight map, and update the online style template library and the traditional print media writing style constraint library according to the style preference weight map.

[0118] The intelligent generation, review, and closed-loop iteration system for multimedia content based on a large model provided in this invention can execute the intelligent generation, review, and closed-loop iteration method for multimedia content based on a large model provided in any embodiment of this invention, and has the corresponding functional modules and beneficial effects of the execution method. Example

[0119] Figure 3 A schematic diagram of an electronic device that can be used to implement embodiments of the present invention is shown. The electronic device 10 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0120] like Figure 3 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12, a random access memory (RAM) 13, etc., which is communicatively connected to the at least one processor 11. The memory stores a computer program that can be executed by the at least one processor 11, and the computer program is executed by the at least one processor 11 to enable the at least one processor 11 to perform the method provided by the present invention.

[0121] The processor 11 can perform various appropriate actions and processes based on a computer program stored in the read-only memory (ROM) 12 or a computer program loaded from the storage unit 18 into the random access memory (RAM) 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0122] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0123] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 executes the various methods and processes described above, such as a large-model-based intelligent generation, review, and closed-loop iterative method for multimedia content.

[0124] In some embodiments, the large-model-based intelligent generation, review, and closed-loop iterative method for multimedia content can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the large-model-based intelligent generation, review, and closed-loop iterative method for multimedia content described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to execute the large-model-based intelligent generation, review, and closed-loop iterative method for multimedia content by any other suitable means (e.g., by means of firmware).

[0125] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard parts (ASSPs), systems-on-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a memory system, at least one input device, and at least one output device, and transmitting data and instructions to the memory system, the at least one input device, and the at least one output device.

[0126] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0127] In the context of this invention, a computer-readable storage medium stores computer instructions that, when executed by a processor, implement the large-model-based intelligent generation, review, and closed-loop iterative method for multimedia content provided by this invention. The computer-readable storage medium can be a tangible medium that may contain or store computer programs for use by or in conjunction with an instruction execution system, apparatus, or device. The computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. Alternatively, the computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0128] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device for displaying information to the user (e.g., a cathode ray tube (CRT) or a liquid crystal display (LCD monitor)); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0129] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0130] A computing system can include clients and servers. Clients and servers are generally geographically separated and typically interact via communication networks. The client-server relationship is established by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system. It addresses the shortcomings of traditional physical hosts and Virtual Private Server (VPS) services, such as high management difficulty and weak business scalability.

[0131] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0132] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for intelligent generation, review, and closed-loop iteration of multimedia content based on a large model, characterized in that, include: Obtain the task requirements for the content to be generated and the historical manuscript data of the target creator, extract the writing style features from the historical manuscript data, construct a historical writing style feature vector, and generate a first draft of multimedia content based on the historical writing style feature vector and the task requirements. Identify the media type corresponding to the initial draft of the multimedia content, load the corresponding differentiated scoring template based on the media type, perform quantitative scoring and compliance audit on the initial draft of the multimedia content, obtain a structured review report, and determine the revision results and release decision based on the structured review report; Based on the target distribution channel attributes, the audited content is reconstructed across channels in a multimodal manner, and online digital channel release data and traditional offline channel publication data are generated respectively. At the same time, an initial benchmarking matrix is ​​generated based on the content of competing products on the same theme. After the online digital channel data is published for a preset period, restricted interaction data and operation briefing data returned from each online publishing channel are collected. The restricted interaction data is dynamically baseline aligned to obtain a relative utility index. The initial benchmarking matrix is ​​updated based on the operation briefing data to obtain a dynamic benchmarking evaluation result. Using the relative utility index as the dependent variable, and the historical writing style feature vector, the scoring matrix in the structured review report, and the dynamic benchmarking evaluation results as the concatenated and aligned fine-grained style feature matrix as the independent variable, the independent variable and the dependent variable are input into the feature association evaluation model for attribution mapping analysis to generate a style preference weight map. The online style template library and the traditional print media writing style constraint library are updated based on the style preference weight map.

2. The method according to claim 1, characterized in that, The process involves acquiring the task requirements for the content to be generated and the historical manuscript data of the target creator; extracting stylistic features from the historical manuscript data to construct a historical stylistic feature vector; and generating a preliminary draft of multimedia content based on the historical stylistic feature vector and the task requirements, including: Obtain historical manuscript data of the target creator within a preset time range, and preprocess the historical manuscript data; Historical writing style feature vectors are extracted from the processed historical manuscript data. These historical writing style feature vectors include at least: sentence length distribution, title structure features, vocabulary preference features, rhetorical features, emotional tone features, paragraph organization features, and theme expression features. Using the historical writing style feature vector and the task requirements of the content to be generated as constraints, a first draft of multimedia content that retains the writing style features of the target creator is generated.

3. The method according to claim 1, characterized in that, Identify the media type corresponding to the initial draft of the multimedia content, load the corresponding differentiated scoring template based on the media type, perform quantitative scoring and compliance auditing on the initial draft of the multimedia content, obtain a structured review report, and determine the revision results and release decision based on the structured review report, including: Identify the media type of the initial draft of the multimedia content, wherein the media type includes at least one of print media, new media, video, photos, and comics; Based on the media type, load the corresponding differentiated scoring template, use a combination of base score and cumulative deduction to quantitatively score the initial draft of the multimedia content, conduct a compliance audit on the initial draft of the multimedia content, and output a structured review report; If the structured review report contains high-risk items, the release decision will be to prohibit release or require manual review. If the structured review report does not contain any high-risk items and the overall score is less than the preset release threshold, the initial draft of the multimedia content will be revised according to the modification suggestions, and the revised content will be re-evaluated and subjected to quantitative scoring and compliance audit. If the structured review report does not contain any high-risk items and the overall score reaches the preset release threshold, then the release decision is determined to allow release or allow printing.

4. The method according to claim 1, characterized in that, Based on the target distribution channel attributes, the audited content is reconstructed across channels in a multimodal manner, generating online digital channel distribution data and traditional offline channel publication data respectively. Simultaneously, an initial benchmarking matrix is ​​generated based on competitor content on the same theme, including: Obtain content data and target distribution channel attributes after review and approval; When the target publishing channel is an online digital channel, the content data is reconstructed into one or more of the following based on the target publishing channel attributes: short text, long text, title group, topic tag group, cover prompt, video storyboard, image description, or short video script. When the target distribution channel is a traditional offline channel, the content data is converted into structured page data or digital proof files based on the print media layout template; Obtain competitor content with the same theme or event as the content data, and perform source de-identification processing on the content data and the competitor content; Based on a pre-set benchmarking evaluation template, the de-identified competitor content is evaluated to obtain an initial benchmarking matrix.

5. A method according to claim 1, characterized in that, After a preset period of data release on the online digital channels, restricted interaction data and operational briefing data from each online release channel are collected. Dynamic baseline alignment processing is performed on the restricted interaction data to obtain a relative utility index. The initial benchmarking matrix is ​​then updated based on the operational briefing data to obtain dynamic benchmarking evaluation results, including: After the data is published on the online digital channels, restricted interactive data that flows back from each online publishing channel is collected according to a preset collection cycle. Obtain online new media operation briefing data from this organization and its competitors, and extract dissemination performance, content structure, user feedback, and competitor dissemination characteristics from the online new media operation briefing data; Using historical baseline data from various online publishing channels, average data of similar content, or average data from platforms within the same time window as dynamic baselines, the restricted interaction data is normalized, denoised, and processed for outliers. The relative utility index is calculated based on the processed restricted interaction data. The relative utility index is used to characterize the content dissemination effect after excluding differences in traffic allocation across different platforms. The initial benchmarking matrix is ​​updated based on the operational briefing data and the relative utility index to obtain dynamic benchmarking evaluation results.

6. The method according to claim 1, characterized in that, Using the relative utility index as the dependent variable, and concatenating and aligning the historical writing style feature vector, the scoring matrix in the structured review report, and the dynamic benchmarking evaluation results into a fine-grained style feature matrix as the independent variable, the independent variable and the dependent variable are input into a feature association evaluation model for attribution mapping analysis to generate a style preference weight map, including: The historical writing style feature vector, the scores of each dimension in the structured review report, the deduction details, the risk markers, the modification suggestions, and the benchmarking scores in the dynamic benchmarking evaluation results are aligned by fields and time windows to obtain a fine-grained style feature matrix. Using the relative utility index as the dependent variable and the fine-grained style feature matrix as the independent variable, attribution mapping analysis is performed through a feature association evaluation model. Based on the results of the attribution mapping analysis, the correlation weights between different writing style elements, rating dimensions, benchmarking dimensions and dissemination effects are determined, and a style preference weight map is generated.

7. The method according to claim 1, characterized in that, The step of updating the online style template library and the traditional print media writing style constraint library based on the style preference weight map includes: Update the cue word constraints in the online style template library based on the style preference weight map, and adjust the multimodal reconstruction strategy of the online digital channels; The weights related to print media topic selection, title, lead, structure, and expression style in the style preference weight map are mapped to the traditional print media writing style constraint library, and the prompt word parameters when generating traditional print media manuscripts are adjusted.

8. A multimedia content intelligent generation, review, and closed-loop iterative system based on a large model, characterized in that: The system is used to execute the intelligent generation, review, and closed-loop iteration method for multimedia content based on a large model, as described in any one of claims 1-7, including: The acquisition module is used to acquire the task requirements of the content to be generated and the historical manuscript data of the target creator, extract the writing style features of the historical manuscript data, construct the historical writing style feature vector, and generate the first draft of multimedia content based on the historical writing style feature vector and the task requirements. The determination module is used to identify the media type corresponding to the initial draft of the multimedia content, load the corresponding differentiated scoring template based on the media type, perform quantitative scoring and compliance audit on the initial draft of the multimedia content, obtain a structured review report, and determine the correction results and release decision based on the structured review report; The generation module is used to perform cross-channel multimodal reconstruction of audited content based on the target publishing channel attributes, and generate online digital channel publishing data and traditional offline channel publishing data respectively. At the same time, it generates an initial benchmarking matrix based on the content of competitors on the same theme. The data collection module is used to collect restricted interaction data and operational briefing data from each online publishing channel after the online digital channel publishes data for a preset period of time, perform dynamic baseline alignment processing on the restricted interaction data to obtain a relative utility index, and update the initial benchmarking matrix based on the operational briefing data to obtain a dynamic benchmarking evaluation result. The analysis module is used to take the relative utility index as the dependent variable, and to align the historical writing style feature vector, the scoring matrix in the structured review report, and the dynamic benchmarking evaluation results into a fine-grained style feature matrix as the independent variable. The independent variable and the dependent variable are input into the feature association evaluation model for attribution mapping analysis to generate a style preference weight map. The online style template library and the traditional print media writing style constraint library are updated according to the style preference weight map.

9. An electronic device, characterized in that, include: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to execute the intelligent generation, review and closed-loop iteration method for multimedia content based on a large model, as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute the intelligent generation, review, and closed-loop iterative method for multimedia content based on a large model, as described in any one of claims 1-7.

Images