Content matrix collaborative management method and system for full-media multi-terminal

Abstract

The application provides a content matrix collaborative management method and system for full-media multi-terminal, and belongs to the field of digital content distribution; solves the problem of low content adaptation efficiency of full-platform; and specifically as follows: a content matrix is established, the preferred theme of a user is determined, and an initial manuscript is screened according to the preferred theme; the style specifications of each platform are extracted, and the initial manuscript is processed to generate derivative manuscripts of different platforms; the optimal publishing time is allocated for the derivative manuscripts of each platform; the manuscript publishing proportion of each platform account in the same or different content partitions is adjusted; the change of the corresponding browsing volume of each platform and each content partition in the content matrix is monitored in real time, and new manuscripts are supplemented for the manuscript library of an enterprise or unit; the application monitors and adapts the flow and manuscript materials of full-platform, obtains derivative manuscripts adapted to the styles of each platform, and has the effect of improving the full-platform content collaborative operation efficiency and the propagation benefit.

Description

Technical Field

[0001] This invention relates to a content matrix collaborative management method and system for all-media multi-terminals, belonging to the field of digital content distribution. Background Technology

[0002] Existing methods or systems for content recommendation across various platforms have the following shortcomings: Enterprises operate multiple accounts across platforms in a fragmented manner, lacking a unified content matrix for overall coordination: Currently, many enterprises or organizations operate accounts independently on multiple platforms. The operators on each platform operate independently, and the data is isolated from each other. There is no unified, comprehensive content matrix framework. Operators rely solely on partial data from the backend of a single platform to judge user preferences. They can only see the browsing performance of a single account and cannot make horizontal comparisons of browsing data for all accounts or different content tracks. It is difficult to objectively distinguish which content belongs to popular tracks, ordinary tracks, and niche tracks.

[0003] The high labor costs of adapting articles for multiple platforms: Under the traditional media operation model, if the same topic is to be published on multiple platforms, the operators need to write copy, adjust the cover, and reconstruct the title independently according to the requirements of each platform. An original article needs to be manually disassembled, rewritten, and formatted multiple times. The amount of repetitive work is large, the content production cycle is long, and manual adaptation is also prone to inconsistencies in the core viewpoints and propaganda of articles on different platforms, resulting in information conflicts and increasing the pressure of content review.

[0004] The lack of time-series view count data results in low view count utilization efficiency: In traditional content publishing, operators mostly rely on their existing experience to publish content at fixed times, without dynamically adjusting the publishing window based on the actual activity patterns of users in different sections and platforms; there are significant differences in the active time periods of users in different content sections, and the peak viewing times of users in popular and unpopular tracks do not overlap. Fixed publishing times will cause high-quality content to miss the view count window, and the exposure effect of a large amount of content will not meet expectations. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide a content matrix collaborative management method and system for all media and multiple terminals, aiming to solve the problem of low efficiency in automated content adaptation across all platforms.

[0006] To achieve the above objectives, the present invention provides a content matrix collaborative management method for all-media multi-terminals, comprising: Establish a content matrix; analyze the changes in page views of accounts across the entire platform in different content sections, and divide all content sections into popular content, ordinary content, and niche content; Based on the content categories corresponding to popular, ordinary, and unpopular content, we determine the user's preferred themes and select initial drafts according to these themes. We also extract the style guidelines of each platform, automatically rewrite the initial drafts, and generate derivative drafts adapted to different platforms. Based on the content matrix, we analyzed the rise and fall patterns of page views in different time periods and content sections. Combining this with the peak activity periods of users on each platform and in each section, we allocated the optimal publication time for the corresponding derivative articles on each platform and adjusted the publication ratio of articles from accounts on each platform in the same or different content sections. Real-time monitoring of pageview changes across platforms and content sections within the content matrix allows for the replenishment of new articles into the enterprise's or organization's article library.

[0007] Furthermore, the processing steps for content partitioning are as follows: Establish a content matrix by using media platforms as row elements and content themes as column elements; We obtained the daily pageview changes for each platform account across different content categories over the past month, and calculated the preference scores for each platform account across different content categories: Let the platform have There is a content section Extract the column matrix of the account corresponding to the first platform from the content matrix, and statistically analyze the daily pageview changes of different content sections on the first platform over the past month to obtain the matrix. For the matrix The data in the matrix is ​​standardized to obtain a matrix. ; Calculate matrix The covariance matrix is ​​obtained by obtaining the matrix. ; Using the matrix Perform eigenvalue decomposition to obtain nonnegative eigenvalues ~ and eigenvalues ~ The corresponding unit orthogonal eigenvectors ~ ; Will ~ Sort in descending order to get ~ ; Will ~ The corresponding feature vectors are concatenated horizontally to obtain a matrix. ; Calculate matrix Right multiplication of a matrix , to obtain the matrix ; Calculate the matrix The matrix is ​​obtained by averaging the elements of each corresponding column. ; matrix The transposed values ​​are used as the preference scores of the first platform account in different content categories and filled into the column matrix corresponding to the first platform account in the content matrix; Calculate the preference scores of other platform accounts in different content categories and fill them into the content matrix; Calculate the average value of the corresponding elements in each column of the content matrix to obtain the preference score for different content sections; record the update time of the content matrix.

[0008] Furthermore, the content partitioning process also includes the following steps: Obtain two content matrices with consecutive update times, denoted as matrix. With matrix ; Calculate matrix The preference scores for all content sections are obtained. ~ ; Calculate matrix The preference scores for all content sections are obtained. ~ ; according to ~ and ~ Calculate the growth rate of preference scores for different content categories to obtain... ~ ; Set a first threshold for preference scores. In other words, content sections with a preference score below this threshold are considered niche content. exist ~ The preference score in the middle of the selection is less than the threshold. The content section is designated as niche content; ~ In the middle, the preference score is greater than or equal to the threshold. The content section is designated as non-niche content; Calculate the average preference score for non-niche content as the second threshold. ; Among the preference scores corresponding to non-niche content, those with preference scores greater than or equal to a threshold are selected. Content sections with a positive preference score growth rate are considered popular content. In all content sections, those sections containing neither popular nor unpopular content will be treated as regular content.

[0009] Furthermore, the processing steps for determining preferred topics are as follows: The system analyzes the click-through rate, completion rate, and interaction rate of users viewing content across various content categories on different platforms; it also calculates the behavioral weights for each content category. ~ ;in, Indicates the number of content partitions; Encode the content tags corresponding to each content partition to obtain the semantic vector of each content partition; Based on the behavioral weights corresponding to each content tag, the semantic vectors of each content partition are weighted once to obtain the first semantic vector of each content partition. A second weight is set for popular content, ordinary content, and unpopular content. The first semantic vectors corresponding to popular content, ordinary content, and unpopular content are then weighted twice to obtain the second semantic vectors for each content partition. Set the number of clusters, cluster the second semantic vectors of each content partition, and obtain several interest clusters; The centroid vectors corresponding to each interest cluster are converted into natural language topic labels, and the confidence scores of each topic label are output simultaneously. Calculate the preference score for each centroid vector, and take the topic tag with the highest preference score as the user's preferred topic.

[0010] Furthermore, the steps for extracting style guidelines from each platform are as follows: Obtain template texts and extract text style specifications from various platforms, including text style features and subtitle timing features; Extracting text style features: Preprocess the template text to obtain standardized template text; input the standardized template text of each platform into the language model to output the semantic representation vector of each text; train the style classification head using the text to be style-labeled, input the semantic representation vector of each text into the style classification head to obtain the style feature set corresponding to each platform; The structure of the style category header includes: Input layer: used to receive initial semantic features; Dimensionality reduction projection layer: The initial semantic features are reduced in dimension, and the reduced semantic features are nonlinearly transformed using an activation function to obtain effective semantic features; Regularization layer: Sets the inactivation probability, and performs regularization processing on the effective semantic features according to the inactivation probability to obtain optimized semantic features; Classification output layer: The optimized semantic features are subjected to three parallel fully connected operations and normalization processing to obtain the probability distributions of tone style, person style and narrative style respectively; Based on the probability distribution of tone style, person style and narrative style in the template manuscripts, the corresponding tags of tone style, person style and narrative style of the platform are selected; In the regularization layer of the style classification head, the optimized semantic features of tone style tags, person style tags and narrative style tags corresponding to each platform are extracted and concatenated into text style feature vectors for each platform.

[0011] Furthermore, the steps for extracting subtitle timing features are as follows: Select template videos, extract frames from template videos on various platforms, identify the subtitle text in the corresponding frames of each template video, record the start timestamp, end timestamp, and number of characters in each subtitle, calculate the temporal index of the subtitles corresponding to each template video, and obtain the temporal characteristics of the subtitles on each platform. Visual style feature extraction: Different visual style tags are preset, and a multimodal model is used to encode the template videos and preset visual style tags of each platform to obtain the visual feature vector of the template video and the tag feature vector of the preset visual style tag. On the same platform, calculate the cosine similarity between the visual feature vector corresponding to each template video and the feature vector of each tag, and select the tag with the highest cosine similarity as the visual style attribute corresponding to each template video; on the same platform, select the visual feature vector with the most visual style attributes as the visual style feature vector of the platform. The visual parameter standards of template videos corresponding to each platform are extracted using a computer vision library. The non-feature vector parts of the text and image style specifications corresponding to each platform are combined with the platform specifications of each platform and organized into structured text instructions corresponding to the style of each platform. By aggregating the structured text instructions, text style feature vectors, and visual style feature vectors from various platforms, a style feature set corresponding to each platform is obtained, and the initial manuscript is transformed into a derivative manuscript adapted to each platform.

[0012] Furthermore, the transformation steps for the initial manuscript are as follows: Rewrite the initial draft based on the stylistic characteristics of the first platform: The structured text instructions of the first platform are denoted as... The initial draft is recorded as ; The encoder uses a deep learning model to... and Tokenization is performed to obtain the following results sequentially. token sequence and token sequence ; Concatenate token sequences and The total token sequence is obtained. ; Total token sequence Input the token embedding layer and the position embedding layer of the deep learning model to obtain the token embedding vector and position embedding vector of each token, respectively. The text style feature vector of the first platform is mapped to a feature vector of the same dimension as the token embedding vector using a fully connected layer to obtain the style guidance vector; The token embedding vector, position embedding vector, and style guidance vector are added element by element to obtain the fused vector sequence; The fused vector sequence is input into the decoder of the deep learning model to generate the first platform's derived text. The text loss is calculated once a derived text is generated. Set a threshold for the text loss of the derived text; determine whether the text loss of the derived text is less than the threshold; if it is less than or equal to the threshold, generate the title of the derived text according to the title specifications of the first platform; if it is greater than the threshold, regenerate the derived text.

[0013] Furthermore, the subsequent steps in rewriting the initial draft are as follows: Each image in the image resource library is converted into an image feature vector, which is denoted as the candidate image vector. Calculate the vectors of each candidate image and the vector... The cosine similarity is used, and the image with the highest cosine similarity is used as the cover image for the derived text; where, vector Represents the theme vector of the derived text; Extract the visual style feature vector from the first platform, denoted as vector. The cover image is processed using an image encoder to obtain a vector. Feature vectors of the same dimension are denoted as vectors. ; vector with vector Cosine similarity is used as the image loss; An image optimizer is used to iterate using the cover image as a sample. During the iteration, the encoded image features are continuously updated, and the image loss is calculated until the image loss converges or the maximum number of iterations is reached, thus obtaining the final latent variables of the cover image. ; Latent variables Input the VAE decoder and perform upsampling pixel reconstruction to obtain a cover image adapted to the visual style of the first platform; If the initial submission is a short video, a style transfer algorithm is used to transfer the style of the initial submission, resulting in a derivative submission for the first platform. Repeat the steps of obtaining the derivative manuscripts corresponding to the first platform, and generate derivative manuscripts adapted to the styles of different platforms based on the text style feature vectors, image style feature vectors, and structured text instructions corresponding to other platforms.

[0014] Furthermore, text loss includes: Core semantic loss and style similarity loss; The calculation process for the core semantic loss is as follows: Extract the main idea vector from the initial manuscript, denoted as vector. ; Extract the main idea vector from the derived text, denoted as vector. ; Calculate vectors with vector Cosine similarity serves as the core semantic loss of the derived text; Style similarity loss: Let the text style feature vector of the first platform be denoted as vector. Repeated vectors The process of obtaining the text style feature vector of the derived text is as follows: ; Calculate vectors with vector Cosine similarity serves as the core semantic loss of the derived text; The sum of the core semantic loss and style similarity loss of the derived text is the text loss of the core semantics.

[0015] The content matrix collaborative management system for all media and multiple terminals includes: User Preference Analysis Module: Used to build a content matrix; analyze the changes in page views of accounts across the entire platform in different content sections, and divide all content sections into popular content, ordinary content, and niche content; Manuscript adaptation module: It is used to determine the user's preferred themes based on the content categories corresponding to popular, ordinary and unpopular content, and to filter the initial manuscripts based on the preferred themes; it extracts the style specifications of each platform, automatically rewrites the initial original manuscripts, and generates derivative manuscripts adapted to different platforms; Article publishing module: Based on the content matrix, it analyzes the rise and fall patterns of page views in different time periods and content sections, combines the peak activity times of users on each platform and in each section, and allocates the optimal publishing time for corresponding derivative articles on each platform; it also adjusts the publishing ratio of articles from accounts on each platform in the same or different content sections. Update monitoring module: Used to monitor changes in page views for each platform and content section in the content matrix in real time, and to add new articles to the manuscript library of enterprises or organizations.

[0016] Compared with the prior art, the beneficial effects of the present invention are: Constructing a unified content matrix across the entire domain avoids the problem of one-sided data from a single platform: This invention gathers all the operating accounts of an enterprise or organization on various platforms, builds an integrated content matrix, analyzes the changes in page views across the entire platform based on the content matrix, determines the main user preferences, and automatically classifies popular, ordinary, and niche content categories. It breaks away from the limitations of the traditional operating model that relies solely on the backend data of a single account and the subjective experience of operators to judge user preferences, breaks down the operational barriers of independent accounts and fragmented data across different platforms, reduces the probability of content selection errors, and improves the overall matching degree between distributed content and user preferences.

[0017] Demand-driven manuscript selection: This invention first identifies user-preferred themes through content matrix view count data analysis, then retrieves the most relevant manuscripts from the company's manuscript library as initial manuscripts, achieving user-demand-driven material selection. Simultaneously, for multi-platform distribution scenarios of the same theme, this invention extracts the specific style requirements of each platform, automatically rewrites the manuscripts, and batch-generates derivative manuscripts adapted to the dissemination characteristics of different platforms. While ensuring information consistency, it also considers the reading habits of users on different platforms, achieving efficient distribution of a single manuscript across multiple channels.

[0018] By fully utilizing the browsing volume windows of various platforms, this invention enhances content exposure efficiency: It continuously collects browsing time-series data from all platforms and content sections within the content matrix, uncovers the patterns of browsing volume fluctuations at different times, distinguishes between three categories of browsing volume sections: rising, stable, and declining popularity, captures peak user activity for each track and platform, and automatically allocates the optimal release time for derivative articles. Simultaneously, this invention also enables dynamic allocation of content matrix resources, flexibly adjusting content placement layout based on browsing trends. For sections with rising popularity, it increases the number of accounts and the scale of content supply; for sections with declining popularity, it reduces content placement, minimizing ineffective resource investment and improving the overall dissemination effectiveness of the content. Attached Figure Description

[0019] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the management method of the present invention; Figure 2 This is a schematic diagram of the management system of the present invention; Figure 3 This is a schematic diagram of the functional flow of the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more readily understood, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] Please see Figure 1 and Figure 3 The content matrix collaborative management method for all media and multiple terminals includes: Step S1: Analyze user preferences; Obtain all accounts of the enterprise or organization on various platforms and establish a content matrix; analyze the changes in page views of all accounts across different content sections, and divide all content sections into popular content, ordinary content, and niche content; The specific steps of step S1 are as follows: Obtain all accounts of an enterprise or organization on various platforms, and build a content matrix with media platforms as row elements and content themes as column elements; the intersection of each row element and column element in the content matrix represents the content section of different accounts on each platform. It should be noted that the aforementioned "content zoning" refers to virtual containers used by various media platforms to categorize, isolate, and display various articles according to their internal regulations, based on the review and style of the articles on each platform.

[0022] The values ​​in the content matrix are tentatively set as the daily pageviews of each platform account in the corresponding content category (only pageviews lasting at least 10 seconds are counted in the pageview statistics). The changes in daily pageviews of each platform account in different content categories over the past month are obtained, and the preference scores of each platform account in different content categories are calculated: Let the platform have There is a content section Extract the column matrix of the account corresponding to the first platform from the content matrix, and statistically analyze the daily pageview changes of different content sections on the first platform over the past month to obtain the matrix. ;matrix The dimension is ( × );in, Indicates the number of days in the past month; For matrix The data in the matrix is ​​Z-score standardized to obtain the matrix. ; Calculate matrix The covariance matrix is ​​obtained by obtaining the matrix. ;matrix The dimension is ( × ); Use the Jacobi iteration method to process the matrix Perform eigenvalue decomposition to obtain nonnegative eigenvalues ~ and eigenvalues ~ The corresponding unit orthogonal eigenvectors ~ The dimensions of ; are all ( ×1); Will ~ Sort in descending order to get ~ ; Will ~ The corresponding feature vectors are concatenated horizontally to obtain a matrix. ;matrix The dimensions are all ( × ); Calculate matrix Right multiplication of a matrix , to obtain the matrix ;matrix The dimension is ( × ); Calculate matrix The matrix is ​​obtained by averaging the elements of each corresponding column. ;matrix The dimension is (1× ); matrix The transposed values ​​are used as the preference scores of the first platform account in different content categories and filled into the column matrix corresponding to the first platform account in the content matrix; Repeatedly calculate the matrix The steps are as follows: calculate the preference scores of other platform accounts in different content categories, and fill them into the content matrix; Calculate the average value of the corresponding elements in each column of the content matrix to obtain the preference score for different content sections; record the update time of the content matrix. Obtain two content matrices with consecutive update times, denoted as matrix. With matrix ; where the matrix The update time is earlier than the matrix Update time; Calculate matrix The preference scores for all content sections are obtained. ~ ; Calculate matrix The preference scores for all content sections are obtained. ~ ; according to ~ and ~ Calculate the growth rate of preference scores for different content categories to obtain... ~ ; in, The formula for calculation is: ; And so on, The formula for calculation is: ; Set a first threshold for preference scores. In other words, content sections with a preference score below this threshold are considered niche content. exist ~ The preference score in the middle of the selection is less than the threshold. The content section is designated as niche content; ~ In the middle, the preference score is greater than or equal to the threshold. The content section is designated as non-niche content; Calculate the average preference score for non-niche content as the second threshold. ; Among the preference scores corresponding to non-niche content, those with preference scores greater than or equal to a threshold are selected. Content sections with a positive preference score growth rate are considered popular content. In all content sections, those sections containing neither popular nor unpopular content will be treated as regular content.

[0023] Step S2: Based on the content categories corresponding to popular, ordinary, and unpopular content, determine the user's preferred themes, and select the articles with the highest matching degree to the preferred themes from the company's or organization's article library as the initial articles; Then, extract the style guidelines and layout requirements of each platform (such as Douyin, WeChat Official Accounts, Xiaohongshu, etc.), automatically rewrite the selected original manuscript, crop the image, reconstruct the title and adapt the layout to generate derivative manuscripts adapted to different platforms; The style guidelines for each platform include text style guidelines and image style guidelines.

[0024] The specific steps to determine a user's preferred topics based on content categories corresponding to popular, regular, and less popular content are as follows: Statistics are compiled on the click-through rate, completion rate, and interaction rate of users watching articles in various content categories across different platforms; for each content category corresponding to popular, ordinary, and unpopular content, the average click-through rate, average completion rate, and average interaction rate of all users across the entire platform are calculated. Based on the average click-through rate, average completion rate, and average interaction rate for each content section, the behavioral weight for each content section is calculated. ~ ;in, Indicates the number of content partitions; The formula for calculation is: = Click-through rate weight · Average click-through rate of the first content section + Completion rate weight · Average completion rate of the first content section + Interaction rate weight · Average interaction rate of the first content section; And so on, The formula for calculation is: = Click-through rate weighting · number The weight of average click-through rate + completion rate for each content section. The average completion rate + interaction rate weight of each content section · The average interaction rate of each content section; where the values ​​of click-through rate weight, completion rate weight and interaction rate weight are all in the range of 0 to 1; the sum of click-through rate weight, completion rate weight and interaction rate weight is 1, and the values ​​of click-through rate weight, completion rate weight and interaction rate weight can be obtained through A / B testing; The content tags corresponding to each content section are encoded using a Chinese-optimized general vector model (such as GTE (General Text Embeddings)) to obtain the semantic vector of each content section; the semantic vector has 512 dimensions. Based on the behavioral weights corresponding to each content tag, the semantic vectors of each content partition are weighted once to obtain the first semantic vector of each content partition. A second weight is set for popular content, ordinary content, and unpopular content. The first semantic vectors corresponding to popular content, ordinary content, and unpopular content are then weighted twice to obtain the second semantic vector of each content partition. The values ​​of the second weights for popular content, ordinary content, and unpopular content are all in the range of 0 to 1. The sum of the second weights for popular content, ordinary content, and unpopular content is 1, and the second weight of popular content > the second weight of ordinary content > the second weight of unpopular content. Set the number of clusters, perform K-Means clustering on the second semantic vector of each content partition, and obtain several interest clusters; Using topic modeling tools (such as the BERTopic model), the centroid vectors corresponding to each interest cluster are transformed into natural language topic labels; the topic modeling tools (such as the BERTopic model) also output the confidence scores of each topic label simultaneously; Calculate the preference score for each centroid vector, and take the topic tag with the highest preference score as the user's preferred topic; The formula for calculating preference score is: The centroid vector's preference score = the number of elements in the interest cluster corresponding to the centroid vector + the average weight of the interest cluster corresponding to the centroid vector + the confidence score of the topic label corresponding to the centroid vector; The average weight of the interest cluster corresponding to the centroid vector is: the average of the product of the behavior weights and the second weights of all semantic vectors in the interest cluster corresponding to the centroid vector; It should be noted that the aforementioned "content categorization" is based on the article review and style of each media platform. A single video may be categorized into multiple content categories. For example, "XXX Home Appliance Review" could be categorized into both the "Home Appliances" and "Money-Saving Tips" categories. Therefore, the content categorization by platforms naturally results in some similarities or overlaps. Thus, when determining user preferences, this invention considers popular, common, and niche content simultaneously to extract topics that users are truly interested in. Sub-step S21: Collect standardized style rules for each platform channel: Extract the style specifications of text and images for each platform (such as word count, tone of voice, screen style, subtitle duration, paragraph format, etc.) and establish a style constraint rule library for each platform. From the articles on each platform, extract the top 30% of articles by monthly page views as template articles for each platform; Extract text style guidelines from various platforms: Extract the text style features and subtitle timing features of the corresponding templates from each platform; Extracting text style features: The text portion of the template manuscript is preprocessed by cleaning, denoising, and removing irrelevant symbols to obtain standardized template texts for each platform; the standardized template texts for each platform are input into a BERT-like pre-trained language model to output the semantic representation vector of each text; the style classification head is trained using the text to be style-labeled, and the semantic representation vector of each text is input into the style classification head to obtain the style feature set corresponding to each platform. The structure of the style category header includes: Input layer: Initial semantic features used to receive the output of the BERT model; Dimensionality reduction projection layer: The initial semantic features output by the BERT model are reduced in dimension, and then the ReLU activation function is used to perform a nonlinear transformation on the reduced semantic features to obtain effective semantic features; Regularization layer: Set the inactivation probability of the regularization layer (the inactivation probability is usually set to 0.3). Based on the inactivation probability, perform regularization processing on the effective semantic features by randomly inactivating neurons to obtain optimized semantic features that suppress overfitting. Classification output layer: Three independent fully connected sub-classification heads are connected in parallel to perform three-way parallel fully connected operations and Softmax normalization classification processing on the optimized semantic features to obtain the probability distributions of tone style, person style and narrative style respectively; Fully connected subclass headers include: Tone and style classification header: Used to output the probability of various language style tags such as formal written, colloquial, popular science interpretation, and news broadcast; Person style classification head: Used to output the probability of first person, third person, and no person narrative person tags; Narrative Style Classification Header: Used to output the probability of various narrative style tags such as knowledge listing, story narration, and policy interpretation; The style tags with the highest probability in the three dimensions of tone style, person style and narrative style (i.e., the language style tag, narrative person tag and narrative style tag mentioned above) in the single template manuscripts are used as the alternative style tags for the single template manuscripts. All template articles corresponding to a platform are fed into a BERT-like pre-trained language model and a style classification head for processing, resulting in candidate style tags for tone style, person style, and narrative style for all template articles corresponding to that platform. Among the candidate style tags for tone style, the one with the most occurrences is selected as the tone style tag for the platform. Similarly, the candidate style tag with the most occurrences for person style is selected as the person style tag for the platform, and the candidate style tag with the most occurrences for narrative style is selected as the narrative style tag for the platform. In the regularization layer of the style classification head, the optimized semantic features of tone style tags, person style tags and narrative style tags corresponding to each platform are extracted and concatenated into text style feature vectors for each platform. Extracting subtitle timing features: The template videos are the top 30% of short videos in terms of monthly views on each platform; Set a fixed frame rate (e.g., 1 frame per second), extract frames from template videos on various platforms, use an OCR engine to recognize the subtitle text in the corresponding frames of each template video, and record the start timestamp, end timestamp, and number of characters in each subtitle. Calculate the timing metrics for the subtitles corresponding to each template video. The timing metrics include: total subtitle duration, subtitle switching frequency, average number of characters in the subtitles, and average duration of subtitle display. Users or relevant technical personnel can add other timing metrics as needed. The timing indicators of the subtitles corresponding to each platform are used as the timing characteristics of the subtitles on each platform; Extracting image style guidelines from various platforms: Extract the visual style features and visual parameter standards of each platform from the template videos of each platform: Visual style feature extraction based on multimodal models (e.g., CLIP model): Preset visual style tags for different styles such as cool color tone, warm color tone, and minimalist real scene. Use a multimodal model (such as CLIP model) to encode the template videos of each platform and the preset visual style tags to obtain the visual feature vector of the template video and the tag feature vector of the preset visual style tags. On the same platform, calculate the cosine similarity between the visual feature vector corresponding to each template video and the feature vector of each tag, and select the tag with the highest cosine similarity as the visual style attribute corresponding to each template video; on the same platform, select the visual feature vector with the most visual style attributes as the visual style feature vector of the platform. Use computer vision libraries (such as OpenCV) to extract the visual parameter standards of template videos for each platform: Use computer vision libraries (such as OpenCV) to extract the image format parameters (such as pixel width and height, frame size, resolution, aspect ratio, etc.) of the cover image and video frames corresponding to each template video. Each template video is frame-sampling at a preset fixed frame rate. The frame difference method is used to calculate the image difference between adjacent frames of the same template video. The image difference between adjacent frames of all template videos on the same platform is summarized, and the average image difference of all template videos on the same platform is calculated to obtain the average transition interval of the platform. Based on OCR detection, the text area and image area of ​​the cover image corresponding to each template video are identified, and the average cover text ratio of all template videos on each platform is calculated (cover text ratio = text image area in the cover / total image area of ​​the cover). By summarizing the image format parameters, average transition intervals, and average cover text ratios of various platforms, the visual parameter standards for each platform are obtained. It should be noted that the three types of visual parameter standards listed here are only the parameter standards recommended by this invention. Users or relevant technical personnel can add other visual parameter standards according to their actual needs. Obtain the platform specifications for each platform; combine the non-feature vector parts of the text and image style specifications for each platform with the platform specifications for each platform, and organize them into structured text instructions (i.e., structured Prompt) for the style of each platform. By summarizing the structured text instructions, text style feature vectors, and visual style feature vectors from various platforms, we can obtain the style feature set corresponding to each platform. Sub-step S22: Multi-platform differentiated conversion of the initial manuscript: The system selects the articles that best match the preferred theme from the company's or organization's manuscript library as initial articles; based on the selected initial articles, it automatically rewrites the copy, trims the materials, and generates matching titles and covers according to the style rules of each platform to obtain derivative articles exclusive to each platform. Rewrite the initial draft based on the stylistic characteristics of the first platform: Sub-step S221: Rewrite the text using a deep learning model (e.g., the Transformer model): The structured text instructions of the first platform are denoted as... The initial draft is recorded as ; The encoder uses a deep learning model to... and Tokenization is performed to obtain the following results sequentially. token sequence and token sequence ; Concatenate token sequences and The total token sequence is obtained. ; Total token sequence Input the token embedding layer and the position embedding layer of the deep learning model to obtain the token embedding vector and the position embedding vector of each token respectively; the token embedding vector and the position embedding vector of each token have the same dimension. Using the fully connected layer of a deep learning model, the text style feature vector of the first platform is mapped to a feature vector of the same dimension as the token embedding vector to obtain the style guidance vector; The token embedding vector, position embedding vector, and style guidance vector are added element by element to obtain the fused vector sequence; The fused vector sequence is input into the decoder of the deep learning model to generate the first platform's derived text. The text loss is calculated once a derived text is generated. Text loss includes: core semantic loss and style similarity loss; The calculation process for the core semantic loss is as follows: The initial manuscript is processed using the BERT model, and the token is extracted as a feature vector of CLS. This vector is used as the main theme vector of the initial manuscript and is denoted as vector. Among them, "CLS" is a special token pre-set at the beginning of the input sequence of the BERT model, which represents the global semantics of the file processed by the BERT model; The BERT model is used to process the derived text generated by the deep learning model, resulting in the topic vector of the derived text, denoted as vector. ; Calculate vectors with vector Cosine similarity serves as the core semantic loss of the derived text; Style similarity loss: Let the text style feature vector of the first platform be denoted as vector. Repeated vectors The process of obtaining the text style feature vector of the derived text is as follows: ; Calculate vectors with vector Cosine similarity serves as the core semantic loss of the derived text; The sum of the core semantic loss and style similarity loss of the derived text is the text loss of the core semantics; Set a threshold for the text loss of the derived text; determine whether the text loss of the derived text is less than the threshold; if it is less than or equal to the threshold, generate the title of the derived text according to the title specifications of the first platform; if it is greater than the threshold, regenerate the derived text. Load an enterprise or organization's image resource library, and use a multimodal model to convert each image in the library into a vector. Image feature vectors of the same dimension are denoted as candidate image vectors; Calculate the vectors of each candidate image and the vector... The image with the highest cosine similarity is used as the cover image for the derived text. Extract the visual style feature vector from the first platform, denoted as vector. The cover image was processed using the CLIP image encoder to obtain a vector image. Feature vectors of the same dimension are denoted as vectors. ; vector with vector Cosine similarity is used as the image loss; Using an image optimizer (such as a Diffusion model or VAE image generator), the cover image is used as a sample for iterative processing. During the iteration process, the encoded image features are continuously updated and the image loss is calculated until the image loss converges or the maximum number of iterations is reached. During the iteration of the cover image, the backbone parameters of the image optimizer (Diffusion model) are fixed in each iteration, and only the latent variables are updated by gradient backwards. The cover image is first processed by a VAE decoder to obtain latent variables. ; hidden variables Input CLIP image encoder to extract latent variables Based on the corresponding image feature vectors, recalculate the image loss for this iteration; based on the image loss of the current round, adjust the latent variables using the inverse gradient update algorithm. Distribution parameters, reducing latent variables Image features and vectors The feature distance is used to obtain new latent variables. ; Latent variables The image loss is recalculated and the latent variables are updated in the CLIP image encoder until the image loss of the latent variables corresponding to the cover image converges to a preset threshold, thus obtaining the final latent variables of the cover image. ; Latent variables Input the VAE decoder and perform upsampling pixel reconstruction to obtain a cover image adapted to the visual style of the first platform; If the initial submission is a short video, the AdaIN (Adaptive Instance Normalization) style transfer algorithm is used to perform style transfer on the initial submission based on the image style feature vector corresponding to the first platform, resulting in a derivative submission for the first platform. Repeat the steps of obtaining the derivative manuscripts corresponding to the first platform, and generate derivative manuscripts adapted to the styles of different platforms based on the text style feature vectors, image style feature vectors, and structured text instructions corresponding to other platforms. Sub-step S23: Platform-wide consistency verification: By comparing the core data and viewpoints of all derivative articles on various platforms with the original articles, we can identify and correct textual inconsistencies and errors in the main idea of ​​the derivative articles, ensuring that the core content of the derivative articles on multiple platforms is consistent, with only differences in the external channel style.

[0025] Step S3: Continuously obtain the effective browsing time sequence of each platform and content section in the content matrix, statistically analyze the rise and fall of browsing volume in different time periods and different content sections, divide the browsing volume into three categories: rising popularity, stable, and declining popularity, and combine the peak activity time of users in each platform and section to allocate the optimal release time for the corresponding derivative articles of each platform. Dynamically adjust the posting ratio of the same or different content across different platform accounts, increase the number of accounts and the amount of content supplied in the sections with rising popularity and declining popularity, and reduce the number of articles published in the sections with declining popularity. Sub-step S31: Aggregate the time-series browsing dataset across all platforms Retrieve valid browsing data from all platforms and all content sections within the matrix for consecutive periods, archive it by date, time period, and content track, remove abnormal browsing volume (such as: platform traffic throttling, short-term sudden hot spots, etc.), and form a standardized browsing volume time-series ledger. Sub-step S32: Analyze the types of pageview changes and user activity periods: The content matrix is ​​used to record the changes in pageviews across platforms and content sections, analyze the increase in pageviews for each section, and distinguish between three categories of sections: rising popularity, stable, and continuous decline; the daily and weekly high-frequency activity time periods of users on each platform are extracted simultaneously. The content matrix is ​​used to record the real-time pageview data of different content sections on each platform. The autoregressive integral moving average model (i.e., ARIMA(p,d,q) model; where p represents the autoregressive order, d represents the difference order, and q represents the moving average order) is used to fit the real-time pageview changes of different content sections on each platform and find the peak time with the largest pageview for each content section on different platforms. Calculate the weekly total pageview change rate for each content category across the entire platform, select content categories with a positive total pageview change rate as rising popularity categories, select content categories with a zero total pageview change rate as stable categories, and select content categories with a negative total pageview change rate as declining categories. Sub-step S33: Select the optimal publication time for the derivative article: Based on the content category and target platform of a single article, the article is automatically published by binding it to the peak time of user browsing on that channel, avoiding low-browning and unpopular periods. Sub-step S34: Dynamic allocation of overall content distribution: Based on the popularity of each content category, adjust the weighting of each content category across all accounts on the platform: Distribute more accounts to popular sections and push more articles of the same type; The popularity of the zones remains stable and the status quo is maintained. For sections experiencing declining page views, reduce the number of accounts that publish content and decrease the proportion of similar content produced. Sub-step S35: Verify the optimization effect: After adjusting the release time, re-record the changes in page views for each content section, determine whether the page view changes meet the standards, mark the platforms and content sections whose page view changes do not meet the standards, and analyze the reasons for the failure to meet the standards from three dimensions: "product strength, page view strength, and operational strength" as lessons learned; if the standards are met, no action is taken.

[0026] Product strength is indicated by whether the derivative articles for each platform address the needs of the users on that platform, i.e., whether the semantic theme vector of the derivative articles for each platform is consistent with the semantic feature vector of the corresponding trending section on each platform. Pageview strength indicates whether the corresponding derivative articles on each platform are correctly distributed by the platform's recommendation system. It mainly checks whether the content of the corresponding derivative articles on each platform matches the algorithm weight of each platform, and whether there are any violations that trigger the platform's traffic restriction mechanism. Operational capabilities indicate whether the release time of derivative articles on each platform is appropriate for the user activity time of the corresponding platform, and whether the derivative articles on each platform have achieved targeted adaptation to the different platform styles.

[0027] Step S4: Continuously update the data; Real-time monitoring of pageview changes across platforms and content sections within the content matrix; real-time updates of content sections categorizing popular, ordinary, and less popular content; and replenishment of new content to an organization's or institution's content library.

[0028] When replenishing a company's or organization's manuscript library, prioritize replenishing manuscripts related to popular content, followed by ordinary content, and lastly, less popular content.

[0029] Please see Figure 2 The content matrix collaborative management system for all media and multiple terminals includes: User Preference Analysis Module: Used to obtain all accounts of an enterprise or organization on various platforms, establish a content matrix; analyze the changes in page views of accounts across the entire platform in different content sections, and divide all content sections into popular content, ordinary content, and niche content; Manuscript Adaptation Module: This module determines the user's preferred themes based on content categories such as popular, common, and unpopular content. It then selects the manuscripts with the highest matching degree to the preferred themes from the company's or organization's manuscript library as the initial manuscripts. Next, it extracts the style guidelines and layout requirements of each platform and automatically rewrites, crops, restructures the titles, and adapts the layout of the selected initial manuscripts to generate derivative manuscripts adapted to different platforms. The article publishing module is used to continuously acquire the effective browsing time sequence of each platform and content section in the content matrix, statistically analyze the rise and fall patterns of browsing volume in different time periods and different content sections, divide the browsing volume into three categories: rising popularity, stable, and declining popularity, and allocate the optimal publishing time for the corresponding derivative articles of each platform based on the peak activity time of users in each section; dynamically adjust the publishing ratio of each platform account for the same or different content, increase the number of accounts and the content supply in the browsing volume section with rising popularity, and reduce the distribution quota in the browsing volume section with declining popularity. Update monitoring module: Used to monitor the changes in page views for each platform and content section in the content matrix in real time, update the content section division of popular content, ordinary content and unpopular content in real time, and supplement the manuscript library of enterprises or organizations with new manuscripts.

[0030] Compared to the problems described in the background technology, this invention fully integrates account data from various platforms under the omni-media matrix, content category browsing characteristics, and differentiated platform content style specifications. It utilizes operational strategies such as content popularity stratification filtering, automatic cross-platform article adaptation, and browsing time sequence analysis to achieve full-link collaborative management of articles. This effectively compensates for the shortcomings of the traditional model where content is manually produced separately for multiple platforms and original articles are directly distributed without adaptation. It can cater to the preferences of different users across all platforms and improve the audience matching degree of content across multiple platforms.

[0031] Furthermore, this invention can continuously collect browsing time-series data from various platforms and content sections, dynamically adapt the optimal timing for article placement based on the rise and fall of browsing volume, and dynamically adjust the content supply and account placement ratio for different content sections. This breaks the limitations of traditional fixed-time release and unchanging content placement scale, effectively improves the adaptability of content distribution strategies to the browsing cycles of various platforms, accurately tilts content resources towards high-potential browsing volume sections, thereby optimizing the overall dissemination effect of the multimedia content matrix, expanding content exposure, and enhancing the stability and return on investment of the entire content operation.

[0032] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to specific implementation methods. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The present invention is limited only to the technical solutions and their full scope and equivalents.

Claims

1. A content matrix collaborative management method for all-media multi-terminals, characterized in that: include: Establish a content matrix; Analyze the changes in pageviews across all accounts on the platform in different content categories, and divide all content categories into popular content, ordinary content, and niche content; Based on the content categories corresponding to popular, ordinary, and unpopular content, we determine the user's preferred themes and select initial drafts according to these themes. We also extract the style guidelines of each platform, automatically rewrite the initial drafts, and generate derivative drafts adapted to different platforms. Based on the content matrix, we analyzed the rise and fall patterns of page views in different time periods and content sections. Combining this with the peak activity periods of users on each platform and in each section, we allocated the optimal publication time for the corresponding derivative articles on each platform and adjusted the publication ratio of articles from accounts on each platform in the same or different content sections. Real-time monitoring of pageview changes across platforms and content sections within the content matrix, supplementing the manuscript library with new articles.

2. The content matrix collaborative management method for all-media multi-terminals according to claim 1, characterized in that, The steps for content partitioning are as follows: Establish a content matrix by using media platforms as row elements and content themes as column elements; We obtained the daily pageview changes for each platform account across different content categories over the past month, and calculated the preference scores for each platform account across different content categories: Let the platform have There is a content section Extract the column matrix of the account corresponding to the first platform from the content matrix, and statistically analyze the daily pageview changes of different content sections on the first platform over the past month to obtain the matrix. For the matrix The data in the matrix is ​​standardized to obtain a matrix. ; Calculate matrix The covariance matrix is ​​obtained by obtaining the matrix. ; Use the matrix Perform eigenvalue decomposition to obtain eigenvalues. ~ and the unit orthogonal eigenvectors corresponding to the eigenvalues. ~ ; Will ~ Sort in descending order to get ~ ; Will ~ The corresponding feature vectors are concatenated horizontally to obtain a matrix. ; Calculate matrix Right multiplication of a matrix , to obtain the matrix ; Calculate the matrix The matrix is ​​obtained by averaging the elements of each corresponding column. ; matrix The transposed values ​​are used as the preference scores of the first platform account in different content categories and filled into the column matrix corresponding to the first platform account in the content matrix; Calculate the preference scores of other platform accounts in different content categories and fill them into the content matrix; Calculate the average value of the corresponding elements in each column of the content matrix to obtain the preference score for different content sections; record the update time of the content matrix.

3. The content matrix collaborative management method for all-media multi-terminals according to claim 2, characterized in that, The content partitioning process also includes: Obtain two content matrices with consecutive update times, denoted as matrix. With matrix ; Calculate matrix The preference scores for all content sections are obtained. ~ ; Calculate matrix The preference scores for all content sections are obtained. ~ ; according to ~ and ~ Calculate the growth rate of preference scores for different content categories to obtain... ~ ; Set a first threshold for preference scores. ,exist ~ The preference score in the middle of the selection is less than the threshold. The content section is designated as niche content; ~ In the middle, the preference score is greater than or equal to the threshold. The content section is designated as non-niche content; Calculate the average preference score for non-niche content as the second threshold. ; Among the preference scores corresponding to non-niche content, those with preference scores greater than or equal to a threshold are selected. Content sections with a positive preference score growth rate are considered popular content. In all content sections, those sections containing neither popular nor unpopular content will be treated as regular content.

4. The content matrix collaborative management method for all-media multi-terminals according to claim 1, characterized in that, The steps for determining preferred topics are as follows: Calculate the behavior weight corresponding to each content partition to obtain ~ ; Encode the content tags corresponding to each content partition to obtain the semantic vector of each content partition; Based on the behavioral weights corresponding to each content tag, the semantic vectors of each content partition are weighted once to obtain the first semantic vector of each content partition. A second weight is set for popular content, ordinary content, and unpopular content. The first semantic vectors corresponding to popular content, ordinary content, and unpopular content are then weighted twice to obtain the second semantic vectors for each content partition. Set the number of clusters, cluster the second semantic vectors of each content partition, and obtain several interest clusters; The centroid vectors corresponding to each interest cluster are converted into natural language topic labels, and the confidence scores of each topic label are output simultaneously. Calculate the preference score for each centroid vector, and take the topic tag with the highest preference score as the user's preferred topic.

5. The content matrix collaborative management method for all-media multi-terminals according to claim 1, characterized in that, The steps for extracting style guidelines from each platform are as follows: Obtain template manuscripts, extract text style specifications and image style specifications according to the style specifications of each platform, and extract text style features and subtitle timing features according to the text style specifications; Extracting text style features: Preprocess the template text to obtain standardized template text; input the standardized template text of each platform into the language model to output the semantic representation vector of each text; train the style classification head using the text to be style-labeled, input the semantic representation vector of each text into the style classification head to obtain the style feature set corresponding to each platform; The structure of the style category header includes: Input layer: used to receive initial semantic features; Dimensionality reduction projection layer: The initial semantic features are reduced in dimension, and the reduced semantic features are nonlinearly transformed using an activation function to obtain effective semantic features; Regularization layer: Sets the inactivation probability, and performs regularization processing on the effective semantic features according to the inactivation probability to obtain optimized semantic features; Classification output layer: The optimized semantic features are subjected to three parallel fully connected operations and normalization processing to obtain the probability distributions of tone style, person style and narrative style respectively; Based on the probability distribution of tone style, person style and narrative style in the template manuscripts, the corresponding tags of tone style, person style and narrative style of the platform are selected; In the regularization layer of the style classification head, the optimized semantic features of tone style tags, person style tags and narrative style tags corresponding to each platform are extracted and concatenated into text style feature vectors for each platform.

6. The content matrix collaborative management method for all-media multi-terminals according to claim 5, characterized in that, The steps for extracting subtitle timing features are as follows: Select template videos, extract frames from template videos on various platforms, identify the subtitle text in the corresponding frames of each template video, record the start timestamp, end timestamp, and number of characters in each subtitle, calculate the temporal index of the subtitles corresponding to each template video, and obtain the temporal characteristics of the subtitles on each platform. Visual style feature extraction: Different visual style tags are preset, and a multimodal model is used to encode the template videos and preset visual style tags of each platform to obtain the visual feature vector of the template video and the tag feature vector of the preset visual style tag. On the same platform, calculate the cosine similarity between the visual feature vector corresponding to each template video and the feature vector of each tag, and select the tag with the highest cosine similarity as the visual style attribute corresponding to each template video; on the same platform, select the visual feature vector with the most visual style attributes as the visual style feature vector of the platform. The visual parameter standards of template videos corresponding to each platform are extracted using a computer vision library. The non-feature vector parts of the text and image style specifications corresponding to each platform are combined with the platform specifications of each platform and organized into structured text instructions corresponding to the style of each platform. By aggregating the structured text instructions, text style feature vectors, and visual style feature vectors from various platforms, a style feature set corresponding to each platform is obtained, and the initial manuscript is transformed into a derivative manuscript adapted to each platform.

7. The content matrix collaborative management method for all-media multi-terminals according to claim 6, characterized in that, The conversion steps for the initial manuscript are as follows: Rewrite the initial draft based on the stylistic characteristics of the first platform: The structured text instructions of the first platform are denoted as... The initial draft is recorded as ; The encoder uses a deep learning model to... and Tokenization is performed to obtain the following results sequentially. token sequence and token sequence ; Concatenate token sequences and The total token sequence is obtained. ; Total token sequence Input the token embedding layer and the position embedding layer of the deep learning model to obtain the token embedding vector and position embedding vector of each token, respectively. The text style feature vector of the first platform is mapped to a feature vector of the same dimension as the token embedding vector using a fully connected layer to obtain the style guidance vector; The token embedding vector, position embedding vector, and style guidance vector are added element by element to obtain the fused vector sequence; The fused vector sequence is input into the decoder of the deep learning model to generate the first platform's derived text. The text loss is calculated once a derived text is generated. Set a threshold for acceptable text loss in derived text; Determine whether the text loss of the derived text is less than the acceptable threshold; If the title is less than or equal to the title of the first platform, the title of the derived text will be generated according to the title specifications of the first platform; if the title is greater than the title of the first platform, the derived text will be regenerated.

8. The content matrix collaborative management method for all-media multi-terminals according to claim 1, characterized in that, The following are the subsequent steps in rewriting the initial draft: Each image in the image resource library is converted into an image feature vector, which is denoted as the candidate image vector. Calculate the vectors of each candidate image and the vector... The cosine similarity is used, and the image with the highest cosine similarity is used as the cover image for the derived text; where, vector Represents the theme vector of the derived text; Extract the visual style feature vector from the first platform, denoted as vector. The cover image is processed using an image encoder to obtain a vector. Feature vectors of the same dimension are denoted as vectors. ; vector with vector Cosine similarity is used as the image loss; An image optimizer is used to iterate using the cover image as a sample. During the iteration, the encoded image features are continuously updated, and the image loss is calculated until the image loss converges or the maximum number of iterations is reached, thus obtaining the final latent variables of the cover image. ; Latent variables Input the VAE decoder and perform upsampling pixel reconstruction to obtain a cover image adapted to the visual style of the first platform; If the initial submission is a short video, a style transfer algorithm is used to transfer the style of the initial submission, resulting in a derivative submission for the first platform. Repeat the steps of obtaining the derivative manuscripts corresponding to the first platform, and generate derivative manuscripts adapted to the styles of different platforms based on the text style feature vectors, image style feature vectors, and structured text instructions corresponding to other platforms.

9. The content matrix collaborative management method for all-media multi-terminals according to claim 7, characterized in that, Text loss includes: Core semantic loss and style similarity loss; The calculation process for the core semantic loss is as follows: Extract the main idea vector from the initial manuscript, denoted as vector. ; Extract the main idea vector from the derived text, denoted as vector. ; Calculate vectors with vector Cosine similarity serves as the core semantic loss of the derived text; Style similarity loss: Let the text style feature vector of the first platform be denoted as vector. Repeated vectors The process of obtaining the text style feature vector of the derived text is as follows: ; Calculate vectors with vector Cosine similarity serves as the core semantic loss of the derived text; The sum of the core semantic loss and style similarity loss of the derived text is the text loss of the core semantics.

10. A content matrix collaborative management system for all-media multi-terminals, applicable to any one of the content matrix collaborative management methods for all-media multi-terminals according to claims 1-9, characterized in that, include: User preference analysis module: used to build a content matrix; Analyze the changes in pageviews across all accounts on the platform in different content categories, and divide all content categories into popular content, ordinary content, and niche content; Manuscript adaptation module: It is used to determine the user's preferred themes based on the content categories corresponding to popular, ordinary and unpopular content, and to filter the initial manuscripts based on the preferred themes; it extracts the style specifications of each platform, automatically rewrites the initial original manuscripts, and generates derivative manuscripts adapted to different platforms; Article publishing module: Based on the content matrix, it analyzes the rise and fall patterns of page views in different time periods and content sections, combines the peak activity times of users on each platform and in each section, and allocates the optimal publishing time for corresponding derivative articles on each platform; it also adjusts the publishing ratio of articles from accounts on each platform in the same or different content sections. Update monitoring module: Used to monitor changes in page views for each platform and content section in the content matrix in real time, and to add new articles to the manuscript library of enterprises or organizations.

Images