News short video automatic clipping and multi-platform differentiated dissemination strategy generation system

Abstract

This invention discloses a system for automatically editing short news videos and generating differentiated dissemination strategies for multiple platforms, belonging to the field of intelligent dissemination technology. The system includes: an intelligent news material analysis module, an automatic short video editing module, a multi-platform differentiated strategy generation module, a dissemination effect simulation and prediction module, a strategy dynamic review and optimization module, a user interaction and decision-making module, and a data storage and management module. This invention deeply analyzes news materials using natural language processing and computer vision technologies, automatically generating differentiated dissemination strategies and effect predictions for different platforms based on machine learning models. Simultaneously, by collecting actual dissemination data and dynamically comparing and reviewing the prediction results, it continuously optimizes the strategy knowledge base using an incremental learning mechanism based on case-based reasoning. This achieves full-process automation and intelligence from content understanding, intelligent editing, strategy generation, effect prediction to dynamic optimization, improving the efficiency and accuracy of cross-platform dissemination of short news videos.

Description

Technical Field

[0001] This invention relates to the field of intelligent communication technology, and in particular to a system for automatically editing short news videos and generating differentiated communication strategies across multiple platforms. Background Technology

[0002] Automatic news short video editing and multi-platform differentiated dissemination strategy generation refers to using artificial intelligence technology to automatically and quickly edit long news videos into short videos, and intelligently generate customized titles, formats, tags, and other publishing schemes based on the characteristics of different social media platforms. Its purpose is to improve the timeliness and scale of news production, enhance the accuracy and effectiveness of cross-platform dissemination through content adaptation to local conditions, and ultimately maximize the news's influence.

[0003] Currently, the focus is usually on automating a single step, lacking a deep understanding and correlation analysis of news semantics and visual content. This results in automatically generated video content not being well-aligned with the core news theme. When distributing across multiple platforms, the same strategies are often used, or only simple format conversions are performed. It is impossible to generate truly differentiated titles, covers, and publishing strategies based on platform characteristics, making it difficult to maximize the dissemination potential of each platform. In addition, existing solutions generally lack the ability to simulate forward-looking dissemination effects and provide risk warnings based on historical data and machine learning. Editorial decisions mainly rely on personal experience, resulting in a high degree of blindness in strategy formulation.

[0004] Therefore, a system for automatically editing short news videos and generating differentiated dissemination strategies across multiple platforms is proposed to address the aforementioned issues. Summary of the Invention

[0005] The main objective of this invention is to provide a system for automatically editing short news videos and generating differentiated dissemination strategies across multiple platforms, in order to solve the problems mentioned in the background above.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is: a system for automatically editing short news videos and generating differentiated dissemination strategies across multiple platforms, the system comprising: The intelligent analysis module for news materials extracts keywords from news texts to generate short video tags, identifies entities to match portrait materials, generates summaries as the basis for subtitles, identifies targets, scenes and focus in video content, and performs multimodal sentiment analysis and public opinion sensitivity scanning and labeling. The short video automatic editing module automatically recommends candidate materials based on the importance of the news, automatically generates subtitle drafts and supports manual proofreading, and recommends matching background music and sound effects based on sentiment analysis results to help users efficiently complete the initial production of videos to adapt to different titles, covers and publishing rules; Multi-platform differentiation strategy generation module: used to adapt to the publishing specifications of different platforms, automatically generate multiple style title suggestions, and recommend cover images that are adapted to the size and style of each platform, so as to generate a draft of a differentiated communication strategy that conforms to the characteristics of multiple platforms; The dissemination effect simulation and prediction module uses an ensemble learning model trained on historical data to perform multi-dimensional feature fusion and multi-indicator correlation analysis on the input news features and strategy drafts. It outputs the potential dissemination indicator ranges and risk indicators for each platform and generates a visual simulation analysis report to predict the dissemination effect and provide risk warnings. The news features include summary, people, scenes, sentiment, and public opinion sensitivity. The strategy dynamic review and optimization module collects actual dissemination data after content release, compares and analyzes it with the initial goals and historical cases, and extracts effective strategy patterns and updates the strategy knowledge base through an incremental learning mechanism based on case reasoning, so as to continuously improve the accuracy and adaptability of strategy generation. User interaction and decision-making module: It provides a centralized editing workbench to display and modify video drafts, strategy suggestions and simulation reports, and provides an approval interface for authorized users to complete review and one-click multi-platform publishing; Data storage and management module: It supports data access during the editing process through a real-time database, forms a traceable case library through a historical archiving unit, and implements role-based access control through a permission management unit.

[0007] Preferably, the news material intelligent analysis module includes a text analysis unit, a visual analysis unit, and an emotion recognition unit; The text analysis unit extracts news keywords using the TF-IDF algorithm, performs entity recognition using a model based on the BERT architecture, and generates summaries using the TextRank algorithm. The visual analysis unit identifies people, scenes, actions, and visual focus areas in the video through object detection and scene classification models. The emotion recognition unit determines the emotional tendency and public opinion sensitivity of news content through multimodal emotion extraction and fusion and sensitive word scanning, and provides annotation prompts.

[0008] Preferably, the short video automatic editing module includes a material recommendation unit, a subtitle assistance unit, and a sound effect recommendation unit; The material recommendation unit is used to automatically recommend candidate video clips and image sequences based on the importance of the news. The subtitle assistance unit is used to automatically generate subtitle text drafts and supports manual proofreading and synchronization. The sound effect recommendation unit recommends background music and sound effect library options that match the sentiment of the news.

[0009] Preferably, the multi-platform differentiation strategy generation module includes a platform rule adaptation unit, a title suggestion generation unit, and a cover suggestion optimization unit; The platform rule adaptation unit stores the publishing specifications of each platform and verifies the compliance of the generated content; The title suggestion generation unit is used to generate multiple title suggestions with different style preferences for editors to choose from or modify; The cover suggestion optimization unit is used to automatically generate or recommend multiple cover images from the materials that are suitable for the size and style of each platform for the editor to select.

[0010] Preferably, the propagation effect simulation and prediction module includes a data training unit, a model simulation unit, and a report generation unit; The data training unit trains a multi-platform propagation effect simulation and analysis model based on historical propagation data; When the model simulation unit is input with the current news characteristics and strategy package draft, it outputs the potential dissemination indicator range and risk indicators for each platform. The report generation unit is used to generate visual simulation analysis reports and risk warnings for personnel to edit and reference.

[0011] Preferably, the model simulation unit employs an ensemble learning model to perform multi-indicator correlation analysis, specifically implemented as follows: Feature Fusion Layer: News features, platform features, and strategy features are fused in multiple dimensions using a multi-head self-attention network; Multi-indicator correlation analysis layer: Outputs the correlation trends and confidence intervals of key indicators; Risk scanning step: Based on the sensitive word database and historical public opinion pattern matching, identify risk points in the content that may cause controversy.

[0012] Preferably, the strategy dynamic review and optimization module includes a data collection unit, a comparative analysis unit, and a knowledge base update unit; The data collection unit collects actual dissemination data at a set period after the content is published. The comparative analysis unit compares the actual data with the project's initial goals and the performance of similar content in the past to analyze the gains and losses of the strategy; The knowledge base update unit stores the effective strategy patterns and experiences derived from the review into the strategy knowledge base to improve the quality of future strategy recommendations.

[0013] Preferably, the knowledge base update unit adopts an incremental learning mechanism based on case-based reasoning to dynamically optimize the strategy knowledge base. Specific implementation methods include: Strategy pattern extraction and quantification stage: Based on the comparative analysis report of the dissemination effect obtained from the review, extract the effective combination of differentiated strategy parameters and transform them into quantifiable strategy feature vectors; Similar case retrieval and fusion stage: Based on the core features of the current news theme and sentiment, retrieve the strategy vectors of historical similar cases from the strategy knowledge base, and fuse the new effective strategy patterns with the historical patterns through weighted averaging or clustering methods; Knowledge base iteration and update phase: The integrated optimization strategy pattern is stored in the strategy knowledge base as a new case, and effect tags and confidence scores are added to it to improve the recommendation quality and adaptability of the subsequent strategy generation module.

[0014] Preferably, the user interaction and decision-making module includes an editing workbench unit and a publishing approval unit; The editing workbench unit is used to centrally display and edit video drafts, strategy suggestions, and simulation reports generated by the system; The publication approval unit provides a review interface before final publication and a one-click multi-platform submission function, and all publication operations must be completed through an authorized user account.

[0015] Preferably, the data storage and management module includes a real-time database unit, a historical archiving unit, and a permission management unit; The real-time database unit is used to support real-time access during the editing process; The historical archiving unit is used to archive news materials, approved strategy versions, and dissemination records to form a case library; The permission management unit implements role-based content access, editing, and publishing permission control.

[0016] The present invention has the following beneficial effects: 1. In this invention, the multi-platform differentiated strategy generation module does not simply perform format conversion. Instead, based on a deep understanding of user profiles, content preferences, and publishing rules of each platform, it generates customized strategy suggestions for the same news content in dimensions such as title style, cover visuals, description tags, and publishing time. Combined with the automatic verification of the platform rule adaptation unit and the intelligent cropping and style adaptation of the cover suggestion optimization unit, it ensures that the generated content not only conforms to platform specifications but also matches the preferences of platform users. This solves the pain points of poor content adaptability and poor dissemination effect in existing distribution models, and can tap the dissemination potential of the same content on different platforms, thereby comprehensively improving the overall cross-platform dissemination efficiency and precise reach of news short videos.

[0017] 2. In this invention, through the dissemination effect simulation and prediction module, an ensemble learning model is used to deeply integrate and correlate news features, platform features, and strategy features. This enables the output of multi-platform dissemination effect predictions and public opinion risk indicators with confidence intervals. This provides editors with quantitative and forward-looking decision-making references before content publication, allowing them to assess the potential effects and risks of different strategy combinations in advance, thereby making better choices. This solves the problem of blind editorial decision-making and trial-and-error based solely on experience in practice, effectively avoiding potential dissemination risks, optimizing resource allocation, and reducing the trial-and-error costs and potential reputational losses of multi-platform content operation as a whole.

[0018] 3. In this invention, the strategy dynamic review and optimization module collects actual dissemination data after short video release on a regular basis and automatically compares and analyzes it with the initial project goals, the performance of similar historical content, and the simulation prediction report before release to generate an objective dissemination effect review report. Its knowledge base update unit adopts an incremental learning mechanism based on case reasoning, which can automatically extract the verified and effective differentiated strategy parameter combinations from the review report, transform them into quantifiable strategy feature vectors, and then dynamically update them to the strategy knowledge base in the form of new cases with confidence scores after merging them with similar historical cases. This process realizes the automatic storage and iterative optimization of system strategy experience, enabling subsequent strategy generation to improve itself based on continuously enriched practical data, thereby significantly improving the accuracy and adaptability of the entire system's strategy recommendation and solving the defects of traditional methods that rely on manual experience for strategy adjustment and lack data-driven closed loops. Attached Figure Description

[0019] Figure 1 This is a framework diagram of the news short video automatic editing and multi-platform differentiated dissemination strategy generation system of the present invention; Figure 2 This is a schematic diagram illustrating the implementation framework of the model simulation unit of the news short video automatic editing and multi-platform differentiated dissemination strategy generation system of the present invention; Figure 3 This is a schematic diagram illustrating the implementation framework of the knowledge base update unit of the news short video automatic editing and multi-platform differentiated dissemination strategy generation system of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0021] Please see Figures 1-3This invention provides a technical solution: a system for automatically editing short news videos and generating differentiated dissemination strategies across multiple platforms, the system comprising: The intelligent analysis module for news materials extracts keywords from news texts to generate short video tags, identifies entities to match portrait materials, generates summaries as the basis for subtitles, identifies targets, scenes and focus in video content, and performs multimodal sentiment analysis and public opinion sensitivity scanning and labeling. The short video automatic editing module automatically recommends candidate materials based on the importance of the news, automatically generates subtitle drafts and supports manual proofreading, and recommends matching background music and sound effects based on sentiment analysis results to help users efficiently complete the initial production of videos to adapt to different titles, covers and publishing rules; Multi-platform differentiation strategy generation module: used to adapt to the publishing specifications of different platforms, automatically generate multiple style title suggestions, and recommend cover images that are adapted to the size and style of each platform, so as to generate a draft of a differentiated communication strategy that conforms to the characteristics of multiple platforms; The dissemination effect simulation and prediction module uses an ensemble learning model trained on historical data to perform multi-dimensional feature fusion and multi-indicator correlation analysis on the input news features and strategy drafts. It outputs the potential dissemination indicator ranges and risk indicators for each platform and generates a visual simulation analysis report to predict the dissemination effect and provide risk warnings. The news features include summary, people, scenes, sentiment, and public opinion sensitivity. The strategy dynamic review and optimization module collects actual dissemination data after content release, compares and analyzes it with the initial goals and historical cases, and extracts effective strategy patterns and updates the strategy knowledge base through an incremental learning mechanism based on case reasoning, so as to continuously improve the accuracy and adaptability of strategy generation. User interaction and decision-making module: It provides a centralized editing workbench to display and modify video drafts, strategy suggestions and simulation reports, and provides an approval interface for authorized users to complete review and one-click multi-platform publishing; Data storage and management module: It supports data access during the editing process through a real-time database, forms a traceable case library through a historical archiving unit, and implements role-based access control through a permission management unit.

[0022] The intelligent analysis module for news materials includes a text analysis unit, a visual analysis unit, and a sentiment recognition unit; The text analysis unit extracts news keywords using the TF-IDF algorithm, performs entity recognition based on a BERT-based model, and generates summaries using the TextRank algorithm, including the following steps: Text preprocessing and keyword extraction: First, the input news text is encoded, cleaned, and segmented, and stop words are removed. Then, the TF-IDF (Term Frequency-Inverse Document Frequency) algorithm is applied to quantify the importance of each term; the formula is: ; This calculation extracts high-weight keywords, which not only summarize the news theme but also serve as a direct basis for generating short video tags (#) and categorizing auxiliary materials. Named Entity Recognition and Structuring: The system calls a BERT-based model to classify the segmented text word by word, recognizing and labeling entities such as people's names, place names, organization names, and time. All recognized entities and their categories are structured into a list. People entities are used to retrieve matching portrait materials, place entities can be used for localized push strategies, and time entities are used for timeliness judgment. Automatic summary generation and core information extraction: To generate concise news summaries, the system uses a TextRank-based algorithm to treat sentences in the text as nodes. It iteratively calculates the similarity weights between sentences to assess their importance. The calculation formula is as follows: ; in It is the score of sentence i. The damping coefficient is... Pointing to node The set of nodes, It is the similarity contribution of sentence j to sentence i. For the node The specified set of nodes, It is sentence j to sentence The similarity contribution; the sentence with the highest score is selected and combined to form a concise text summary, which will serve as the core text basis for automatically generating subtitles in short videos, or be transformed into the first draft of the voice-over script; The visual analysis unit identifies people, scenes, actions, and visual focus areas in the video using object detection and scene classification models, including the following steps: Adaptive keyframe sampling and scene segmentation: To efficiently analyze video content, the system calculates the color histogram differences between consecutive frames (e.g., using Bach distance or chi-square distance formulas). It detects sudden scene changes; when the difference value exceeds a preset threshold, it is determined to be a scene transition point, and higher density sampling is performed before and after this point. Multi-granularity visual object and scene recognition: For the sampled keyframes, the system runs the object detection model and the scene classification model in parallel. The object detection model outputs the bounding boxes, categories and confidence scores of all salient objects in the image. At the same time, the scene classification model assigns a macro-label to the entire image. For example, for an outdoor event scene, the system will count the objects that appear frequently and have high confidence scores in the entire video and define them as core visual elements. This information is directly used to guide the automatic selection of the cover image, prioritizing the selection of frames that contain core visual elements and have clear composition as cover candidates. Dynamic analysis and visual focus tracking: For video segments containing people or important dynamic events, the system uses optical flow to analyze the motion trajectory of objects; by calculating the displacement of specific feature points between adjacent frames, the motion vector and direction of objects can be estimated; the dynamic analysis results are combined with static recognition results to support the automatic editing module in selecting the most expressive segments and determining the rhythm of shots. The sentiment recognition unit determines the sentiment tendency and public opinion sensitivity of news content through multimodal sentiment extraction and fusion and sensitive word scanning, and provides annotation prompts, including the following steps: Multimodal sentiment feature extraction: On the text side, deep semantic understanding is performed on the main text and potential caption text to output a continuous sentiment polarity score. On the visual side, the scene category and dominant color output by the visual analysis unit are quantified into an emotion modulation coefficient. ; Comprehensive sentiment analysis: The system will score the sentiment of the text. Visual Emotion Modulation Coefficient Perform weighted fusion to calculate the overall sentiment tendency score: (in and (Weights set according to modal confidence). The values ​​are discretized and categorized as positive, neutral, or negative, which will directly affect the automatic recommendation of background music and sound effects, and provide guidance for the tone of copy in the communication strategy. Sentiment Sensitivity Scanning and Risk Labeling: First, the AC automaton algorithm is used for high-speed pattern matching of the text, compared with a dynamically updated sensitive word database; any direct match will trigger an alert. Second, combined with named entity recognition results, it is checked whether specific high-attention entities are involved. Finally, the sentiment indicators obtained in the previous steps are... Incorporate topic categories into the risk assessment model to calculate a comprehensive risk probability. Ultimately, a risk label with a classification and explanation of the cause is generated. This label is prominently pushed to the editing and review interface and serves as one of the key input parameters, affecting the prediction results of the public opinion risk by the dissemination effect simulation and prediction module.

[0023] The short video automatic editing module includes a material recommendation unit, a subtitle assistance unit, and a sound effect recommendation unit; The material recommendation unit automatically recommends candidate video clips and image sequences based on news importance, including the following steps: Multimodal Importance Score: First, the system receives key entities, sentiment, and visual focus from the news material intelligent analysis module; then, it calculates a comprehensive importance score for each original video clip. The score is calculated by weighting text relevance, visual prominence, and emotional intensity, using the following formula: ; in, As adjustable weighting coefficients, this formula allows for a more quantitative evaluation of the system; for example, a speaker announcing a key policy segment will score much higher than a segment with empty audience footage. Timeline analysis and climax extraction: based on the calculated timeline of each segment. The system plots an importance score curve along the video timeline, and then applies a peak detection algorithm to identify scores exceeding a global threshold. The system extracts several climactic segments; at the same time, it ensures that the extracted segments meet the duration constraints of short videos, and optimizes between importance, total duration and segment coherence to generate an optimal candidate segment sequence as the main material for editing. Diversity control and image sequence recommendation: Overly similar segments are categorized, and then representative segments are selected from each major category to ensure that the recommended materials have sufficient visual variation and information coverage in terms of both imagery and content. For static images, keyframes from climax segments are selected... The frame with the highest score and best composition quality is recommended. The subtitle assistance unit is used to automatically generate subtitle text drafts and supports manual proofreading and synchronization; The sound effects recommendation unit recommends background music and sound effects library options that match the sentiment of the news, including the following steps: Mapping of Emotional Tags to Musical Features: The system has a built-in mapping table between emotional tags and musical features. This table quantifies emotional tags into a series of calculable target values ​​for musical audio features. For example, positive / exhilarating emotions may be mapped to high tempo (BPM>120), major key, and bright timbre; while negative / serious emotions are mapped to low tempo (BPM<90), minor key, and steady melody. This mapping relationship can be represented as a function: ; in For the target rhythm, For target tone, The target spectrum center; Music Library Feature Extraction and Index Construction: The system performs offline analysis on all audio files in the background music library, extracting rhythm features (such as BPM), tonality features (such as major / minor probability), timbre features, and emotion tags. These features constitute a high-dimensional vector. And store it in an index database that supports fast similarity retrieval; Feature matching and contextualized recommendation: When processing specific news, the system recommends based on... Obtain the target feature vector Subsequently, a nearest neighbor search is performed in the music library index to calculate... With each The similarity distance, for example, using the weighted cosine similarity formula: ; in For weighted cosine similarity, The total dimension of the feature vector. For the index variable of the feature dimension, The system assigns weights to different features and returns the 3-5 most similar tracks as recommendations; at the same time, for specific news types, tracks with corresponding tags are given priority.

[0024] The multi-platform differentiation strategy generation module includes a platform rule adaptation unit, a title suggestion generation unit, and a cover suggestion optimization unit; The platform rules adaptation unit stores the publishing specifications of each platform and verifies the compliance of the generated content; The construction and structured storage of the platform specification knowledge base: This unit first needs to maintain a dynamically updated platform rule knowledge graph. Each rule is structured and stored as a five-tuple (platform, content type, constraint field, constraint operator, threshold), for example (Douyin, video, duration, ≤, 60 seconds). The system updates this knowledge base through a combination of API monitoring and manual input, and uses a version management mechanism to track rule changes to ensure that policy generation is based on the latest specifications. Real-time content compliance verification and quantitative assessment: The verification process is essentially rule matching and condition judgment; for example, the verification of title length can be described as: The system will generate a compliance score matrix for each draft strategy. ,in Representing the platform, This represents a specific rule, with a value of 1 (passed), 0 (violated), or 0.5 (a warning is required, such as if sensitive words are involved). Automated correction suggestion generation: For violations found during validation, predefined correction strategies are invoked based on the violation type. For example, when the title is too long, the system uses a text summarization algorithm (such as extracting core clauses based on TextRank) to generate a shortened version; when the cover image size is mismatched, an image cropping algorithm is invoked, guided by a visual saliency map, based on the target aspect ratio. Calculate the optimal clipping window: ; in For optimal clipping window, To obtain the parameter corresponding to the maximum value, For pixels, For candidate cropping window, coordinates Visual salience values ​​at each location; ensure that the most important image content is retained, and submit all suggested revisions to the editorial team along with the original draft for a decision-making process; The title suggestion generation unit generates multiple title suggestions with different style preferences for editors to choose from or modify; The cover suggestion optimization unit is used to automatically generate or recommend multiple cover images from the materials that are suitable for various platform sizes and styles for editors to choose from.

[0025] The propagation effect simulation and prediction module includes a data training unit, a model simulation unit, and a report generation unit; The data training unit trains a multi-platform propagation effect simulation and analysis model based on historical propagation data, including the following steps: Multi-source historical data fusion and feature engineering: This unit first extracts complete data packages of historical news short video projects from the data storage module, including news feature vectors, the final adopted strategy package, actual dissemination indicators of each platform, and environmental factors. The system cleans and denoises this data. Subsequently, large-scale feature engineering is performed, where non-numerical features are one-hot encoded or embedded, and then normalized and concatenated with numerical features (such as video duration and title length) to form a unified, high-dimensional training feature matrix. The corresponding labels are multi-dimensional propagation indicator vectors for each platform. ; Multi-task ensemble learning model training and validation: The system employs an ensemble model within a multi-task learning framework (such as Stacking ensemble based on XGBoost or LightGBM) for training. This model can simultaneously predict multiple relevant propagation metrics, capturing the intrinsic correlations between the metrics. For input, output pair The predicted values ​​are obtained by adjusting hyperparameters through cross-validation during training, and evaluating the model's performance on the validation set using metrics such as mean absolute percentage error. The calculation formula is as follows: When the MAPE is lower than a preset threshold, the model is considered to have completed training. Model version management and dynamic update mechanism: The trained model is assigned a version number and stored in the model repository. The system regularly collects new propagation data. When the accumulated data reaches the threshold or the prediction error of the model on new data continues to rise, the incremental training or full retraining process is automatically triggered to ensure that the prediction model can adapt to the dynamic changes in the content ecosystem and user preferences. When the model simulation unit is input with the current news characteristics and strategy package draft, it outputs the potential dissemination indicator range and risk indicators for each platform, including the following steps: Real-time feature construction and strategy package encoding: After the editor completes the strategy draft adjustment for a news short video, the system runs in real time. The model simulation unit receives the current news features and multi-platform strategy package draft from the upstream module. The unit first encodes the strategy draft into a strategy feature vector consistent with the training phase; then, it concatenates it with the news feature vector and real-time environment features to generate a real-time simulation input feature vector that is completely aligned with the format of historical training data. ; Interval prediction of propagation index based on uncertainty estimation: Input a pre-trained multi-task ensemble model, and the model outputs point estimates of various propagation metrics for each platform. The prediction differences among the base learners in the ensemble model (e.g., calculating the standard deviation of the predicted values) are utilized. Alternatively, quantile regression techniques can be used to output the prediction interval for each indicator; for example, the prediction interval for play count can be expressed as... ,in The system simultaneously invokes a parallel risk classifier, based on the required confidence level, to calculate the probability of triggering negative public opinion based on the textual and visual content and sentiment in the policy draft. And mark high-risk points; Inter-platform comparative analysis and comprehensive potential assessment: The unit conducts a horizontal comparative analysis of the prediction results of the same news on different platforms, and calculates a comprehensive dissemination potential index. ; in The weights are set according to the operational objectives, and This index is used to visually compare the expected performance of different platforms and to provide a basis for ranking and highlighting key points for the report generation unit; The report generation unit is used to generate visual simulation analysis reports and risk warnings for editors to reference.

[0026] The model simulation unit employs an ensemble learning model to perform multi-indicator correlation analysis. Specific implementation methods include: Feature Fusion Layer: News features, platform features, and strategy features are fused in multiple dimensions using a multi-head self-attention network; Heterogeneous feature alignment and standardization preprocessing: This layer first receives three categories of heterogeneous features from the upstream module: news features, strategy features, and platform features. Since these features have different dimensions and distributions, the system first performs Z-score standardization on numerical features and embedding encoding on categorical features to ensure that all features are in a comparable numerical space, laying the foundation for subsequent fusion. Feature Dynamic Weighted Fusion Based on Multi-Head Attention Mechanism: The system employs a multi-head self-attention network to achieve deep interaction and fusion of features. This mechanism allows the model to learn the relationships between features in parallel from different subspaces. Specifically, the standardized feature sequence is used as input, and through querying ( ),key( ),value( A linear transformation of ) is used to calculate the attention weights for the ) The size, and its attention output is: ; in Key matrix transpose, This is the dimension of the key vector, used to scale the dot product. Then, the outputs of all the heads are concatenated and linearly transformed again to form the final fused feature vector. This process enables the model to dynamically identify and reinforce feature combinations that contribute more to the prediction of propagation effects. Feature Dimensionality Reduction and Enhancement: To mitigate the curse of dimensionality that high-dimensional features may cause and improve the model's generalization ability, the system performs feature dimensionality reduction and enhancement. Dimensionality reduction is performed using principal component analysis or a lightweight autoencoder to extract the most informative low-dimensional representation. Simultaneously, to preserve specific important information from the original features, the system concatenates the dimensionality-reduced features with key original features selected by expert rules, forming an enhanced fused feature vector. , serving as the unified input for the next layer of multi-indicator correlation analysis; Multi-indicator correlation analysis layer: Outputs the correlation trends and confidence intervals of key indicators (such as play count and interaction rate), including the following steps: Multi-task learning based on gradient boosting ensemble model: This layer uses a multi-task gradient boosting decision tree model (such as using the multi-task learning function of LightGBM) as the core predictor; the model outputs from the fusion layer. As input, the model learns to predict multiple propagation metrics simultaneously; during training, the model optimizes by minimizing a joint loss function. ; in For the model to the first The predicted value for each task, For the first The true label value of each task It is the first The loss function for each task These are task weights, reflecting the importance of different metrics; Confidence Interval Estimation Based on Model Uncertainty: To quantify the uncertainty of the prediction and provide the prediction interval, the system utilizes the prediction discrepancies of all base learners (decision trees) in the ensemble model. For each sample, it collects the indicators of all tree pairs. Predicted value set Final point estimation of the forecast Take the mean of this set; then calculate the standard deviation of these predictions. Based on the assumption of normal distribution, a 95% confidence interval can be constructed as follows: ; Risk scanning step: Based on a sensitive word database and historical public opinion pattern matching, identify potential controversial risk points in the content, including the following steps: Multi-level sensitive content matching and initial risk screening: First, the AC automaton algorithm is used for precise matching against a core sensitive word library containing dimensions such as politics, violence, and discrimination; any match is directly marked as high risk. Second, fuzzy matching and semantic similarity models (such as Sentence-BERT) are used to compare against an expanded risk expression pattern library to identify risk content with similar semantics but different expressions. Each matched word or pattern is assigned a preset severity level. and context weights Accumulate an initial risk score ; Historical public opinion pattern matching and scenario-based risk assessment: The system performs similarity retrieval between the feature vectors (topic, entity, sentiment) of the current news and the features of cases in the historical public opinion case database; it then uses the cosine similarity formula to find the Top-K most similar historical cases, as follows: ; If a certain percentage of these similar cases have triggered negative public opinion, an alert will be triggered, and the system will adjust the alert based on the average negative impact of similar cases. Combined with the current news's potential for dissemination, an additional risk score based on historical patterns is calculated: ; Comprehensive Risk Score Calculation and Classification: The system aggregates risk signals from different channels; the final comprehensive risk score... Calculated using a weighted aggregation model: ; in It is a risk classification model (such as one based on fusion features) The probability output by the logistic regression model. To harmonize the weights, and ,according to The system maps the value to three risk levels: high, medium, and low, and clearly identifies the news item in the forecast report.

[0027] The strategy dynamic review and optimization module includes a data collection unit, a comparative analysis unit, and a knowledge base update unit; The data collection unit collects actual dissemination data at set intervals after content release. The raw data obtained may vary in structure depending on the platform and may contain noise. The system first maps the data fields of each platform to internal standard fields according to a preset data mapping rule table. Subsequently, the time-series data is smoothed, for example, using a sliding window mean filter. The calculation formula is as follows: ; in For a moment The smoothing value, For window size, Half width, For the original sequence at time... The system eliminates short-term fluctuations by setting the value of the data. At the same time, the system identifies and marks abnormal data points that deviate significantly from historical trends or platform averages, placing them in a separate queue for subsequent analysis to avoid polluting the core analysis dataset. The comparative analysis unit compares actual data with the project's initial goals and the performance of similar content in the past to analyze the gains and losses of the strategies, including the following steps: Multi-benchmark data alignment and comparison framework construction: After the unit starts, three sets of benchmark data are first extracted from the data storage module: The initial goals of the project, i.e., the vector of key performance indicator (KPI) targets set before release. ; Historical prediction reports are reports output by the propagation effect simulation and prediction module that include the predicted median and confidence intervals. ; Based on the feature vector of the current news, the system uses cosine similarity to retrieve the N most similar historical cases from the case library and calculates the average of their various metrics. , as a historical baseline; Quantitative bias calculation and attribution analysis: The system will use the actual collected data A point-by-point comparison is made with the three benchmarks mentioned above; the core is to calculate a series of quantitative deviation indicators. For example, calculating the degree of target achievement. ; Calculate prediction deviation ,in Indicates the median of the forecast; calculates the historical transcendence. Simultaneously, the system performs strategy-level attribution, analyzing the differences in the final outcome caused by different strategy choices using either the controlled variable method or a lightweight regression model. Based on the contribution level, a preliminary assessment of the strategy's gains and losses can be made; Generate a structured debriefing report: Based on the above quantitative analysis and attribution results, the system automatically generates a structured debriefing report, which includes a core summary, a detailed data comparison table for each platform, key findings, and preliminary strategy optimization suggestions; The knowledge base update unit stores effective strategy patterns and experiences derived from the review into the strategy knowledge base to improve the quality of future strategy recommendations.

[0028] The knowledge base update unit employs a case-based incremental learning mechanism to dynamically optimize the strategy knowledge base. Specific implementation methods include: Strategy Pattern Extraction and Quantification Phase: Based on the comparative analysis report of dissemination effects derived from the post-mortem analysis, the system extracts the validated and effective combinations of differentiated strategy parameters and transforms them into quantifiable strategy feature vectors. This is implemented as follows: First, the system receives the post-mortem analysis report output by the comparative analysis unit and extracts validated and effective strategy patterns from it. A pattern is defined as a specific combination of strategies employed under specific conditions (content characteristics, platform) that results in a significant positive effect improvement E. The system encodes the conditions (such as news category, sentiment), platform, and strategies (such as headline template ID, cover type, and publication time) into a unified feature vector. The effect improvement E is then quantified into a scalar value, for example... ; Similar Case Retrieval and Fusion Stage: Based on the core features of the current news, such as its theme and sentiment, strategy vectors of historical similar cases are retrieved from the strategy knowledge base. New effective strategy patterns are then fused with historical patterns using weighted averaging or clustering methods. This includes the following steps: Case retrieval based on vector similarity: When a new effective strategy pattern vector is obtained... Then, the system needs to integrate it into the strategy knowledge base; first, the system calculates... Compared with all existing case vectors in the knowledge base Cosine similarity between them: This similarity measure measures the overall similarity between new and old cases in terms of conditional context and policy actions; the system sets a similarity threshold. All The old cases were retrieved to form a set of similar cases. ; Weighted fusion and extraction of strategic knowledge: If The fact that the value is not empty indicates the existence of similar policy experiences in the past; the system does not simply store these repeatedly, but rather performs knowledge fusion, the core of which is the extraction of the policy action part; the system uses a weighted average method to generate a fused, more generalizable policy vector. The weights are determined by the confidence level of the effects in previous cases. and similarity Joint decision: ; in The fused policy vector For the first The weights of each policy vector. For the first The old policy vector ensures that older cases with better performance, higher confidence, and greater similarity have a greater influence on the formation of the new policy vector; Generation of fusion cases and inheritance of effect labels: While fusing strategy vectors, the system also performs similar operations on context vectors, or directly adopts... The context section. This ultimately generates a fusion case. Its vector representation is The effect tag of this integration case Then through the The effect labels of the cases were weighted and averaged to calculate the initial confidence level. Set as The sum of the average confidence of the cases and a reward factor based on the number of fused cases is expressed as: ,in It is an adjustment parameter that encourages integration based on more historical experience; Knowledge base iteration and update phase: The integrated optimization strategy pattern is stored as a new case in the strategy knowledge base, and effect tags and confidence scores are added to it to improve the recommendation quality and adaptability of the subsequent strategy generation module. This includes the following steps: Case entry and metadata annotation: Whether it's a brand new case (when (When empty) or the merged case At this stage, each case vector will be stored as a new entry in the strategy knowledge base. During storage, in addition to the case vector itself, the system will also attach rich metadata, including: effect tags. Initial confidence level Generate timestamps, traceability information, and a summary of applicable conditions; Dynamic decay and reinforcement update of confidence: The system introduces a time decay factor. Periodically decay the confidence level of all cases in the database: Furthermore, if the context and strategy of an old case are highly similar to those of a newly occurring successful case (i.e., it was retrieved and weighted during the fusion phase of the new cases), the confidence of the old case will be strengthened, and its update can be expressed as follows: ,in It's the learning rate. It is the weight of new cases when they are integrated; this mechanism allows the knowledge base to discard outdated or unverified knowledge while reinforcing valid knowledge that has been repeatedly verified.

[0029] The user interaction and decision-making module includes an editing workbench unit and a publishing approval unit; The editing workbench unit is used to centrally display and edit video drafts, strategy suggestions, and simulation reports generated by the system; The publication approval unit provides a review interface before final publication and a one-click multi-platform submission function, and all publication operations must be completed through an authorized user account.

[0030] The data storage and management module includes a real-time database unit, a historical archive unit, and a permission management unit; The real-time database unit is used to support real-time access during the editing process; The historical archive unit is used to archive news materials, approved strategy versions, and dissemination records, forming a case library; The permission management unit implements role-based content access, editing, and publishing permission control, and implements a fine-grained role-based access control model. First, core roles (such as intern, editor, chief editor, and administrator) and their corresponding permission sets are defined. Permissions are defined in the form of (resource, operation, condition) triples, such as (policy package, modify, project.status=editing). The mapping relationship between roles and permissions is stored in a dedicated permission policy table. In addition, the system supports attribute-based access control, and permissions can dynamically depend on resource attributes (such as the department to which the project belongs) or user attributes (such as the team to which the user belongs).

[0031] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0032] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A system for automatic short video cutting and multi-platform differentiated dissemination strategy generation, characterized in that, The system includes: The intelligent analysis module for news materials extracts keywords from news texts to generate short video tags, identifies entities to match portrait materials, generates summaries as the basis for generating subtitles, identifies targets, scenes and focus in video content, and performs multimodal sentiment judgment and public opinion sensitivity scanning and labeling. The short video automatic editing module automatically recommends candidate materials based on the importance of the news, automatically generates subtitle drafts and supports manual proofreading, and recommends matching background music and sound effects based on sentiment analysis results to help users efficiently complete the initial production of videos to adapt to different titles, covers and publishing rules; Multi-platform differentiation strategy generation module: used to adapt to the publishing specifications of different platforms, automatically generate multiple style title suggestions, and recommend cover images that are adapted to the size and style of each platform, so as to generate a draft of a differentiated communication strategy that conforms to the characteristics of multiple platforms; The dissemination effect simulation and prediction module uses an ensemble learning model trained on historical data to perform multi-dimensional feature fusion and multi-indicator correlation analysis on the input news features and strategy drafts. It outputs the potential dissemination indicator ranges and risk indicators for each platform and generates a visual simulation analysis report to predict the dissemination effect and provide risk warnings. The news features include summary, people, scenes, sentiment, and public opinion sensitivity. The strategy dynamic review and optimization module collects actual dissemination data after content release, compares and analyzes it with the initial goals and historical cases, and extracts effective strategy patterns and updates the strategy knowledge base through an incremental learning mechanism based on case reasoning, so as to continuously improve the accuracy and adaptability of strategy generation. User interaction and decision-making module: It provides a centralized editing workbench to display and modify video drafts, strategy suggestions and simulation reports, and provides an approval interface for authorized users to complete review and one-click multi-platform publishing; Data storage and management module: It supports data access during the editing process through a real-time database, forms a traceable case library through a historical archiving unit, and implements role-based access control through a permission management unit. 2.The system of claim 1, wherein, The intelligent analysis module for news materials includes a text analysis unit, a visual analysis unit, and a sentiment recognition unit; The text analysis unit extracts news keywords using the TF-IDF algorithm, performs entity recognition using a model based on the BERT architecture, and generates summaries using the TextRank algorithm. The visual analysis unit identifies people, scenes, actions, and visual focus areas in the video through object detection and scene classification models. The emotion recognition unit determines the emotional tendency and public opinion sensitivity of news content through multimodal emotion extraction and fusion and sensitive word scanning, and provides annotation prompts. 3.The news short video automatic clipping and multi-platform differentiated propagation strategy generation system according to claim 1, characterized in that: The short video automatic editing module includes a material recommendation unit, a subtitle assistance unit, and a sound effect recommendation unit; The material recommendation unit is used to automatically recommend candidate video clips and image sequences based on the importance of the news. The subtitle assistance unit is used to automatically generate subtitle text drafts and supports manual proofreading and synchronization. The sound effect recommendation unit recommends background music and sound effect library options that match the sentiment of the news. 4.The system of claim 1, wherein: The multi-platform differentiation strategy generation module includes a platform rule adaptation unit, a title suggestion generation unit, and a cover suggestion optimization unit; The platform rule adaptation unit stores the publishing specifications of each platform and verifies the compliance of the generated content; The title suggestion generation unit is used to generate multiple title suggestions with different style preferences for editors to choose from or modify; The cover suggestion optimization unit is used to automatically generate or recommend multiple cover images from the materials that are suitable for the size and style of each platform for the editor to select. 5.The news short video automatic clipping and multi-platform differentiated propagation strategy generation system according to claim 1, characterized in that: The propagation effect simulation and prediction module includes a data training unit, a model simulation unit, and a report generation unit; The data training unit trains a multi-platform propagation effect simulation and analysis model based on historical propagation data; When the model simulation unit is input with the current news characteristics and strategy package draft, it outputs the potential dissemination indicator range and risk indicators for each platform. The report generation unit is used to generate visual simulation analysis reports and risk warnings for personnel to edit and reference. 6.The system of claim 5, wherein: The model simulation unit employs an ensemble learning model to perform multi-indicator correlation analysis, specifically implemented as follows: Feature Fusion Layer: News features, platform features, and strategy features are fused in multiple dimensions using a multi-head self-attention network; Multi-indicator correlation analysis layer: Outputs the correlation trends and confidence intervals of key indicators; Risk scanning step: Based on the sensitive word database and historical public opinion pattern matching, identify risk points in the content that may cause controversy.

7. The news short video automatic clipping and multi-platform differentiated propagation strategy generation system according to claim 1, characterized in that: The strategy dynamic review and optimization module includes a data collection unit, a comparative analysis unit, and a knowledge base update unit. The data collection unit collects actual dissemination data at a set period after the content is published. The comparative analysis unit compares the actual data with the project's initial goals and the performance of similar content in the past to analyze the gains and losses of the strategy; The knowledge base update unit stores the effective strategy patterns and experiences derived from the review into the strategy knowledge base to improve the quality of future strategy recommendations.

8. The news short video automatic clipping and multi-platform differentiated propagation strategy generation system according to claim 7, characterized in that: The knowledge base update unit employs a case-based incremental learning mechanism to dynamically optimize the strategy knowledge base. Specific implementation methods include: Strategy pattern extraction and quantification stage: Based on the comparative analysis report of the dissemination effect obtained from the review, extract the effective combination of differentiated strategy parameters and transform them into quantifiable strategy feature vectors; Similar case retrieval and fusion stage: Based on the core features of the current news theme and sentiment, retrieve the strategy vectors of historical similar cases from the strategy knowledge base, and fuse the new effective strategy patterns with historical patterns through weighted averaging or clustering methods; Knowledge base iteration and update phase: The integrated optimization strategy pattern is stored in the strategy knowledge base as a new case, and effect tags and confidence scores are added to it to improve the recommendation quality and adaptability of the subsequent strategy generation module. 9.The system of claim 1, wherein: The user interaction and decision-making module includes an editing workbench unit and a publishing approval unit; The editing workbench unit is used to centrally display and edit video drafts, strategy suggestions, and simulation reports generated by the system; The publication approval unit provides a review interface before final publication and a one-click multi-platform submission function, and all publication operations must be completed through an authorized user account. 10.The news short video automatic clipping and multi-platform differentiated propagation strategy generation system according to claim 1, characterized in that: The data storage and management module includes a real-time database unit, a historical archive unit, and a permission management unit; The real-time database unit is used to support real-time access during the editing process; The historical archiving unit is used to archive news materials, approved strategy versions, and dissemination records to form a case library; The permission management unit implements role-based content access, editing, and publishing permission control.

Images