All-media operation intelligent analysis platform based on big data

Abstract

The invention discloses an all-media operation intelligent analysis platform based on big data, and relates to the technical field of computers, the platform comprises a data acquisition module, the data acquisition module applies a distributed crawler technology, compiles customized scripts, and combines social media official API calling and data interface docking or pushing of a cooperation platform to obtain a data acquisition module; all-media data including texts, images, audios and videos are collected from web pages, social media and news clients in real time; through a distributed crawler technology, social media official API calling and cooperation platform data interface docking, all-media data including texts, images, audios and videos are collected in real time, and cross-modal data are converted into feature vectors in a unified format for fusion by using a dynamic weight fusion algorithm. The problem that in a traditional analysis method, single-modal data information is limited is solved, and a cross-modal dynamic fusion formula is adopted.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to an all-media operation intelligent analysis platform based on big data. Background Art

[0002] With the rapid development of Internet technology, the omnimedia era has arrived. The way information is disseminated has undergone earth-shaking changes. Diversified channels such as web pages, social media, and news clients continue to emerge, generating massive amounts of text, images, audio, and video multimedia data every day. These data contain rich market information, user preferences, and valuable content on social hot spots. For enterprises, governments, and research institutions, how to efficiently collect, process, and analyze these omnimedia data to tap into their potential value has become an urgent problem to be solved. The omnimedia operation intelligent analysis platform has emerged, aiming to use advanced technical means to conduct in-depth mining and analysis of omnimedia data to provide strong support for decision-making.

[0003] However, traditional omni-media data analysis technology has many shortcomings. First, in terms of data collection, traditional methods often rely on a single data source or simple crawler technology, which makes it difficult to cover a wide range of data in the omni-media field, and the real-time and accuracy of data collection are difficult to guarantee. Secondly, in terms of data processing, traditional technologies often lack the ability to deeply integrate multimodal data, and it is difficult to effectively integrate different types of data such as text, images, audio, and video, resulting in one-sided analysis results and an inability to fully reflect the true situation of media content. In addition, in terms of sentiment semantic analysis and trend hotspot prediction, traditional technologies often rely on manual labeling or simple keyword matching, which makes it difficult to accurately capture the sentiment tendencies and semantic relationships in the text, and cannot effectively predict the development trends and hot topics of media content. Finally, in terms of decision support, traditional technologies often lack scientific and reasonable decision-making basis and visualization display means, resulting in a highly subjective and inefficient decision-making process.

[0004] Therefore, developing an intelligent analysis platform for all-media operations based on big data will greatly improve the efficiency and accuracy of all-media data analysis, provide more scientific and reasonable decision-making support for enterprises, governments and research institutions, and promote the intelligent development of the field of all-media operations. Summary of the Invention

[0005] The purpose of the present invention is to make up for the shortcomings of the existing technology and provide an all-media operation intelligent analysis platform based on big data. Through distributed crawlers, API calls and data interface docking, it realizes the real-time collection of all-media data. The platform uses a dynamic weight fusion algorithm to process multimodal data, combines the sentiment tendency judgment model with the semantic role labeling technology, and deeply analyzes the data sentiment and semantic relationship. At the same time, through hot spot screening and trend prediction models, it accurately grasps the development trend of media content. Finally, it integrates the analysis results and provides visual decision-making suggestions, effectively improving the efficiency of all-media operations and the scientific nature of decision-making.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: a big data-based omni-media operation intelligent analysis platform, which includes: Data collection module: Using distributed crawler technology, we write customized scripts, integrate official social media API calls and data interface docking or push from partner platforms to collect all media data including text, images, audio, and video from web pages, social media, and news clients in real time; Data preprocessing module: cleans the collected raw data, converts text, image, audio, and video data into feature vectors in a unified format, and fuses them through a dynamic weight fusion algorithm; Sentiment and semantic analysis module: This module creates a sentiment tendency judgment model to judge the sentiment tendency of the fused data, uses semantic role labeling technology to understand the semantic relationships and implicit information in the text, and uses knowledge graphs to associate entities and relationships in the text; Trend and hotspot analysis module: This module organizes media data into time series to capture temporal features; uses a dynamic hotspot cluster screening algorithm to cluster text and screen potential hotspot clusters; and combines this with a dynamic hotspot trend prediction model to analyze and predict media content development trends and hot topics. Decision support module: Integrates the results of the sentiment semantic analysis module and the trend hotspot analysis module, generates decision recommendations based on the decision recommendation knowledge base, and uses visualization tools to present the analysis results and recommendations in the form of charts and text; Interactive feedback module: Develop a responsive interface compatible with multiple terminals, support users to query data and customize analysis tasks through natural language input or operation instructions, collect user evaluations of analysis results and suggestions through feedback portals, and use them to optimize platform models and strategies after analysis.

[0007] Furthermore, the full media data collected by the data collection module is: Text data includes: web page content, social media text, news client content, and user interaction data; Image data includes: web page images, social media images, and news client images; Audio data includes: social media audio, news client audio, and web page audio; Video data includes: web videos, social media videos, and news client videos.

[0008] Furthermore, the data preprocessing module fuses the text, image, and audio data of the same standard format through a dynamic weight fusion algorithm. The calculation formula of the dynamic weight fusion algorithm is: ,in, is the fused multimodal feature vector, is the modal index, Represent text, image, audio, and video modes respectively. It is The eigenvectors of the modes, is the text modality feature vector, is the image modality feature vector, is the audio modal eigenvector, is the video modality feature vector, It is Moment The weight of the modality, It is Modality attention weights.

[0009] Furthermore, the calculation of parameters in the dynamic weight fusion algorithm is Moment Modal weight The calculation formula is: ,in represents the natural exponential function, which is used to convert the importance score of the mode Converted to nonlinear weight values, For the Moment Mode The importance score is calculated as follows: ,in is modal The temporal variance of the feature, is modal Feature correlations with other modalities, is modal The variance correlation balance coefficient is obtained through historical data statistics; Modality attention weight The calculation formula is: ,in is modal The weight matrix, is modal The key feature vector of is the attention dimension, used to scale dot product calculations, and is set according to the model design and task requirements. It is The eigenvectors of the modes, is a symbolic function that converts multiple real numbers into a probability distribution.

[0010] Furthermore, the emotional tendency judgment model in the emotional semantic analysis module is constructed using the following formula: ,in is the fused multimodal feature vector, It is sentiment polarity projection vector, is the projection vector weight coefficient, which is obtained by optimizing the sentiment classification task through training data. is the emotion intensity adjustment factor, which is dynamically adjusted based on the modal credibility. is a symbolic function, output Represent negative, neutral, and positive emotions respectively. Represents the number of sentiment polarity projection vectors.

[0011] Furthermore, the specific steps of understanding the text semantic relationship and implicit information through semantic role labeling technology in the sentiment semantic analysis module are as follows: (1) Extract verbs / adjectives in sentences as predicates and distinguish their semantics through contextual disambiguation; (2) Label the core arguments and peripheral arguments of the predicate and construct a predicate-argument relationship matrix; (3) Identify the modification relationship between arguments through dependency syntax analysis and extract the sentiment polarity of the modifiers; (4) Complete omitted elements, identify metaphorical expressions, and deduce implicit logical relationships; (5) Align entities in images / audio with textual arguments to verify cross-modal semantic consistency; (6) Mapping semantic roles to knowledge graph nodes and completing implicit semantics through graph relationships; (7) Generate results that include predicate-argument structure, implicit information, and multimodal associations.

[0012] Furthermore, the trend hotspot analysis module uses a hotspot cluster dynamic screening algorithm to cluster texts and screen potential hotspot clusters. The formula is: ,in It is a cluster The text density, is the cluster after text clustering, It is a cluster The mean semantic similarity of the inner text, It is a cluster Normalized semantic distance to the nearest high-density cluster, It is a dynamic threshold that is adjusted dynamically based on time and data volume.

[0013] Furthermore, the trend hotspot analysis module predicts the development trend of media content and hot topics through a hotspot trend dynamic prediction model, and the calculation formula is: ,in, is the cluster after text clustering, Hotspot cluster The importance weight of is: , Hotspot cluster The heat growth rate reflects the growth of the heat cluster over time. represent Hotspot cluster collection filtered at all times Medium Other hotspot clusters outside yes The hotspot cluster set filtered out at a moment includes all hotspot clusters that meet the hotspot conditions at that moment. Indicates Hotspot cluster collection filtered at all times Except for specific hotspot clusters Other hotspot clusters The text density, It is predicted Hot trend value at the moment, yes The hotspot cluster collection filtered out at all times, Hotspot cluster The predicted value of the popularity change trend, Hotspot cluster The similarity between the semantics of and historical outbreak topics, is the trend factor weight, which is determined by training and verification of historical hotspot data and has a value range of , It's time.

[0014] Furthermore, the decision suggestion knowledge base in the decision support module includes the following contents: Industry case library: success stories and failure stories; Policy template library: general policy templates and industry-customized policy templates; Market environment information database: macroeconomic data, policy and regulatory information, and industry trend analysis reports; Expert experience knowledge base: expert opinions and suggestions and expert case interpretations; Internal enterprise data and experience database: historical decision-making data and internal business processes and specifications.

[0015] Furthermore, the decision support module generates decision suggestions based on the decision suggestion knowledge base through the dual-source fusion decision algorithm, and the formula is: ,in It is the best decision suggestion. is a candidate suggestion in the decision suggestion knowledge base, is the eigenvector of the current analysis result, It is a candidate suggestion The corresponding result feature vector, is the characteristic vector of the current market environment, It is a candidate suggestion The corresponding environmental feature vector, It is the similarity weight of the analysis results, which is set according to the specific application scenario and decision-making objectives. The value range is: , is the cosine similarity function.

[0016] Compared with existing technologies, this big data-based all-media operation intelligent analysis platform has the following beneficial effects: 1. The present invention uses distributed crawler technology, social media official API calls and cooperation platform data interface docking to collect all media data including text, images, audio and video in real time, and uses dynamic weight fusion algorithm to convert these cross-modal data into feature vectors in a unified format for fusion. It not only solves the problem of limited information of single modal data in traditional analysis methods, but also ensures that the importance and relevance of different modal data in the fusion process are fully considered through cross-modal dynamic fusion formula. In addition, the platform also creates a sentiment tendency judgment model, combines semantic role labeling technology and knowledge graph, and performs sentiment tendency judgment and semantic relationship understanding on the fused data to generate structured data, so that the platform can understand the sentiment tendency and semantic relationship in the media content more comprehensively and deeply, and provide a richer and more accurate data foundation for subsequent trend hotspot analysis and decision support.

[0017] 2. The present invention organizes media data into time series and uses a hot spot cluster dynamic screening algorithm and a hot spot trend dynamic prediction model to accurately predict the development trend of media content and hot topics. The hot spot cluster dynamic screening algorithm effectively screens out potential hot spot clusters by calculating text density, semantic similarity mean and semantic distance normalization value indicators; and the hot spot trend dynamic prediction model further combines the importance weight, heat growth rate and semantic similarity factors of the hot spot cluster to accurately predict future hot spot trends. In addition, the platform also integrates the results of the sentiment semantic analysis module and the trend hot spot analysis module, generates decision recommendations based on the decision recommendation knowledge base, and displays the analysis results and recommendations in the form of charts and text through visualization tools. This intelligent trend hot spot prediction and accurate decision support function can not only help users grasp the dynamic changes of media content in a timely manner, but also provide users with scientific and reasonable decision-making basis, thereby improving users' operational efficiency and decision-making quality.

[0018] Other advantages, objects and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art based on an examination of the following or may be learned from the practice of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive effort.

[0020] Figure 1 This is a workflow diagram for an all-media operation intelligent analysis platform based on big data; Figure 2 This is a module framework diagram of an all-media operation intelligent analysis platform based on big data. DETAILED DESCRIPTION

[0021] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the specific implementation methods, structures, features and effects of the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments.

[0022] Example 1: In-depth implementation of the big data-based all-media operation intelligent analysis platform in the K12 online education field.

[0023] To improve course conversion rates and user retention rates, a leading K12 online education platform uses omnimedia data analysis to optimize its operational strategy and achieve precise user growth. By integrating dynamic weight fusion algorithms, sentiment tendency judgment models, and hotspot cluster dynamic screening algorithm technology modules, the platform builds a full-link intelligent analysis system from data collection to decision support.

[0024] Data collection module: real-time capture of omni-channel user behavior and public opinion.

[0025] Use distributed crawler technology to connect with the official API to build a multi-dimensional data collection network: Internal behavioral data: SDK tracking is used to collect user click traces, course playback duration, conversion paths from trial classes to full-price classes, and homework submission rates within the app. This covers text (user messages), video (lecture clips), and audio (live class recordings). External public opinion data: Use customized scripts to capture all media content from the WeChat ecosystem (friend circle advertising comments, parent community discussions), short video platforms (TikTok "XX course evaluation" videos and comments, Kuaishou teaching clip interactions), vertical communities (Mama.com "preschool transition course recommendation" posts, Zhihu "how to choose a math online course" Q&A), and simultaneously obtain images (course promotional posters, user homework screenshots) and video data, such as Figure 1 shown.

[0026] Data collection follows the principle of dynamic adaptation. For example, for the algorithm recommendation feature of the Douyin platform, real-time user interaction data on popular education topics is obtained through the official social media API to ensure data integrity of high-conversion channels.

[0027] Data preprocessing module: standardization and dynamic fusion of multimodal features.

[0028] Perform a cleaning process on the raw data: improve data quality by removing duplicate text, filtering outliers, and standardizing formats; The dynamic weight fusion algorithm is used to convert different modal data into a unified feature vector. The calculation formula of the dynamic weight fusion algorithm is: ,in, is the fused multimodal feature vector, is the modal index, Represent text, image, audio, and video modes respectively. It is The eigenvectors of the modes, is the text modality feature vector, is the image modality feature vector, is the audio modal eigenvector, is the video modality feature vector, It is Time the weight of the modal, is the attention weight of the modal, for example, when processing data related to "winter training class", the algorithm automatically calculates the weight of each modal in the current scene according to the dimensions of Douyin advertisement click volume (image / video modal), WeChat community consultation volume (text modal), and trial class playback completion rate (video modal) - if the conversion rate of Douyin channel is significantly higher than that of other channels, the algorithm will increase the attention weight of this modal to highlight the propagation effect characteristics of the advertisement material, such as Figure 2 .

[0029] Sentiment semantic analysis module: deep understanding of user needs and emotional insight.

[0030] Construct an emotional tendency judgment model, the formula is: , where is the fused multi-modal feature vector, is the emotional polarity projection vector, is the projection vector weight coefficient, is the emotional intensity adjustment factor, which is dynamically adjusted based on modal credibility, is the sign function, and represent negative, neutral, and positive emotions, respectively. , indicates the number of emotional polarity projection vectors, which are classified into three categories (positive / neutral / negative) according to the core dimensions of the education scene: Course experience: Analyze the predicate argument structure in the comments through semantic role labeling technology, such as extracting the semantic relationship of "teacher's explanation of geometric auxiliary line skills is very useful" from "teacher's explanation of skills", and judging positive emotion combined with emotional polarity projection vector; Service quality: Identify negative emotions in "class teacher does not timely feedback homework problems", and locate specific service link vulnerabilities through knowledge graph association "after-sales service" entity.

[0031] Use semantic role labeling technology to mine implicit information: such as parent comment "child likes math teacher, but always daydreams in English class", the model can identify the potential needs of "math course interest positive" and "English course attention negative", and confirm the problem authenticity through cross-modal semantic verification (such as associating the playback progress data of the English class video).

[0032] Integrate "grade subject knowledge point difficulty" entity relationship through knowledge graph technology, for example, associate the high-frequency consultation of "grade 2 physics buoyancy calculation" with the corresponding course chapter to form a mapping network of "user question course content".

[0033] Trend hotspot analysis module: dynamic capture and prediction of education demand.

[0034] Use the hotspot cluster dynamic screening algorithm to identify industry cyclical hotspots. The formula is: ,in It is a cluster The text density, is the cluster after text clustering, It is a cluster The mean semantic similarity of the inner text, It is a cluster Normalized semantic distance to the nearest high-density cluster, It is a dynamic threshold that is adjusted dynamically based on time and data volume. Before the midterm exam, the platform clustered text related to "pre-exam sprint" and screened out a potential hotspot cluster of "junior high school math final question breakthrough" by calculating the text density, semantic similarity, and distance to historical hotspots within the cluster. In response to policy trends (such as the release of the "New Curriculum Standards"), we captured emerging topics such as "Core Literacy Cultivation" and "Project-Based Learning." We determined their potential for dissemination through the semantic features of hotspot clusters and analyzed their development trajectories using a dynamic prediction model for hotspot trends. The calculation formula is: ,in, is the cluster after text clustering, Hotspot cluster The importance weight of is: , Hotspot cluster The heat growth rate reflects the growth of the heat cluster over time. represent Hotspot cluster collection filtered at all times Medium Other hotspot clusters outside yes The hotspot cluster set filtered out at a moment includes all hotspot clusters that meet the hotspot conditions at that moment. Indicates Hotspot cluster collection filtered at all times Except for specific hotspot clusters Other hotspot clusters The text density, It is predicted Hot trend value at the moment, yes The hotspot cluster collection filtered out at all times, Hotspot cluster The predicted value of the popularity change trend, Hotspot cluster The similarity between the semantics of and historical outbreak topics, is the trend factor weight, ranging from , At this moment, when it was found that the search volume for "primary school English grammar special course" increased by 150% within a week, the model calculated the popularity growth rate of the hot cluster and matched the outbreak pattern of the "grammar learning" topic in the same period in history. It predicted that it would reach a peak in demand within 3 weeks and recommended preparing the course launch and promotional materials in advance.

[0035] Decision support module: intelligent strategy generation and visual decision assistance.

[0036] Based on the dual-source fusion decision algorithm, decision recommendations are generated according to the decision recommendation knowledge base. The formula is: ,in It is the best decision suggestion. is a candidate suggestion in the decision suggestion knowledge base, is the characteristic vector of the current analysis result, It is a candidate suggestion The corresponding result feature vector, is the characteristic vector of the current market environment, It is a candidate suggestion The corresponding environmental feature vector, It is the similarity weight of the analysis results, which is set according to the specific application scenario and decision-making objectives. The value range is: , The cosine similarity function integrates sentiment analysis results (e.g., 30% of fifth-grade parents’ negative feedback on “the course is too difficult”) with trend predictions (increased search volume for the keyword “differentiated instruction”), and calls upon the decision-making recommendation knowledge base: Industry Case Library: Refer to the promotion strategy of the competitor's "difficulty adaptive course" and retrieve its implementation case that increased the conversion rate by 12%; Strategy Template Library: Enables the "tiered trial class + personalized recommendation" operation template, and combines it with a dynamic weight fusion algorithm to generate course recommendation plans at different difficulty levels; Expert experience database: Integrate the suggestions of educational psychology experts on "age-appropriate content design" to optimize the course outline.

[0037] Generate visual decision suggestions: Dynamic charts display negative review keywords for each subject (e.g., "fast explanation of math word problems") and trend predictions (e.g., the conversion probability of "winter vacation preview class" over the next two weeks). Strategic proposals are then generated using natural language. For example, "Run the 'Fifth Grade Math Level Trial Class' short video on Douyin, paired with a 'Difficulty Test + Smart Recommendation' landing page, and optimize ad creatives based on A / B testing strategies used in competitor products." Interactive feedback module: human-computer collaborative optimization and model iteration.

[0038] It provides a multi-terminal responsive interface, allowing operators to customize analysis tasks through natural language commands. For example, "Analyze the reasons for user churn in junior high school Chinese courses in Shanghai." The platform automatically calls the sentiment and semantic analysis module to parse comment data and locate the specific problem of "lack of depth in classical Chinese explanations" through semantic role annotation. Collect user feedback on recommended courses (such as "the recommended Olympiad course is beyond the child's ability"), input the feedback data into the hot trend prediction model and dynamic weight fusion algorithm, adjust the "difficulty matching" parameters and modal weights, for example, increase the weight of user behavior data (such as trial class completion rate) in the recommendation model, optimize the accuracy of subsequent course recommendations, and form a closed-loop mechanism of "data collection, analysis, decision-making, feedback and optimization."

[0039] To summarize, when the platform is applied in the field of K12 online education, it collects user behavior and public opinion data through all channels, uses the dynamic weight fusion algorithm to realize the standardized processing of multimodal data, uses the sentiment semantic analysis module to explore user needs and emotional tendencies, uses the hot spot cluster dynamic screening algorithm and trend prediction model to grasp the dynamics of educational needs, generates operation strategies based on the dual-source fusion decision algorithm combined with the knowledge base, realizes model iteration through interactive feedback, forms a precise user growth solution, effectively improves the course conversion rate and user retention rate, and provides strong intelligent analysis support for the operation of the online education platform.

[0040] Example 2: In-depth implementation of the big data-based all-media operation intelligent analysis platform in the field of education.

[0041] Holographic monitoring of university brand image and intelligent management platform for public opinion.

[0042] In order to enhance the effectiveness of brand communication, colleges and universities have built a public opinion monitoring and response system covering all media touchpoints to achieve dynamic maintenance and strategic optimization of brand reputation.

[0043] Data collection module: omni-channel dynamic data capture.

[0044] Using distributed crawler technology and connecting to the official API, we can collect three types of data in real time: Proprietary media matrix: official website news, WeChat official account tweets, Weibo topics (such as campus open days), Douyin campus account videos and user comments, and simultaneous capture of user interaction data (likes, reposts, comments); Social media: In-depth discussions on Zhihu's "How to Evaluate XX University," graphic content on the Douban group "XX University Study Experience," and campus Vlogs and commentary posted by UP hosts on Bilibili. Third-party media: reports on school research results from education-related vertical platforms (such as "Ji Mian Dui"), information on school-enterprise cooperation from local news clients, and campus-related content posted by unofficial accounts on short video platforms (such as classroom clips secretly filmed by students).

[0045] Data types cover text (press releases, comments), images (event posters, campus scenery pictures), audio (recordings of admissions briefings), and video (live broadcast and replay of graduation ceremonies), ensuring that all media data is captured without blind spots.

[0046] Data preprocessing module: multimodal feature fusion and standardization.

[0047] Perform a cleansing process on the raw data: remove duplicate information, filter spam comments, correct format errors, and ensure data quality; Using the dynamic weight fusion algorithm, text, images, audio, and video are converted into a unified feature vector. The formula is: For example, when processing data related to "school anniversary live broadcast", the algorithm automatically assigns dynamic weights to different modalities based on the dimensions of live broadcast viewing volume, barrage word frequency (text), on-site photo dissemination volume (image), and theme song playback times (audio), highlighting data with high dissemination value (such as school anniversary video clips that broke the circle).

[0048] Sentimental semantic analysis module: deep semantic understanding and emotional insight.

[0049] Construct a sentiment tendency judgment model, and the construction formula is: , performs ternary classification on the data (positive / neutral / negative). For example, when analyzing comments related to "dormitory conditions", the model extracts the predicate argument structure of "dormitory facilities are outdated" through semantic role labeling technology, combines it with the knowledge graph to associate it with the "logistics management" entity, and determines the specific direction of negative sentiment; Using semantic role annotation to parse implicit information: For example, if a Weibo post mentions "the library reservation platform crashed again," the model can identify "platform crash" as the core event and "impact on learning" as the implicit consequence. It then links this to the "informatization construction" node through the knowledge graph to locate management loopholes. Cross-modal semantic verification: Align scenes in campus activity photos (such as library study rooms) with descriptions in text reviews to verify the consistency of positive comments such as "strong learning atmosphere".

[0050] Trend hotspot analysis module: dynamic hotspot capture and trend prediction.

[0051] Using the hotspot cluster dynamic screening algorithm, cluster analysis is performed on time series data to select potential hotspot clusters. The formula is: For example, during the enrollment season, the platform automatically identifies the topic clusters of "admission score line", "top majors", and "employment quality", and screens potential hotspots based on text density and semantic similarity; Combined with the hot topic trend dynamic prediction model, the development trajectory of the hot topic is analyzed, and the calculation formula is: If it is found that the discussion volume of the "admission score line" topic has increased by 300% within 24 hours, the model predicts that the topic may evolve into a regional hotspot through the growth rate of heat and semantic similarity with historical public opinion events, triggering an early warning mechanism.

[0052] Decision support module: intelligent strategy generation and visual presentation Integrate the results of sentiment analysis (e.g., the proportion of negative public opinion in a certain college reaches 45%) and trend prediction (the public opinion may spread within 48 hours), call resources in the decision-making suggestion knowledge base, and generate decision-making suggestions through a double-source fusion decision-making algorithm, the formula is: The resources in the knowledge base are: Industry case library: refer to the public opinion handling process of "laboratory safety accidents" in other colleges and universities; Strategy template library: enable the "official statement + live rectification" crisis response template; Expert experience library: call the public relations expert's suggestion on "how to rebuild trust through data visualization".

[0053] Generate specific decision-making suggestions: such as "publish a laboratory rectification timeline chart on Weibo, and conduct joint interviews with local media on the spot", and display the public opinion heat curve and sentiment distribution cloud through dynamic charts to assist the school propaganda department in formulating the communication rhythm.

[0054] Interactive feedback module: human-machine collaborative optimization mechanism.

[0055] Provide a multi-terminal responsive interface, and the propaganda personnel can query "international cooperation related public opinion in the past week" through natural language, and the platform automatically retrieves the corresponding data and generates an analysis report; Collect the evaluation of the staff on the suggestions (such as "the live rectification viewing volume exceeds expectations"), input the feedback data into the model, optimize the modal weight of sentiment analysis and the trend factor of hot spot prediction, and form a closed-loop optimization.

[0056] In summary, in the brand image monitoring and public opinion management of universities, the big data-based omni-media operation intelligent analysis platform realizes omni-channel data collection through distributed crawlers and API docking, uses dynamic weight fusion algorithm to integrate multimodal information, combines sentiment tendency judgment model with semantic role labeling technology to analyze public opinion emotions and implicit information, and uses hot spot cluster dynamic screening algorithm and trend prediction model to capture public opinion hot spots and development trends. Finally, it relies on the decision-making recommendation knowledge base to generate strategies, and optimizes the model through interactive feedback, constructing a closed-loop system from data collection to decision support, providing intelligent solutions for university brand maintenance and public opinion response.

[0057] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as above in terms of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can, without departing from the scope of the technical solution of the present invention, make some changes or modifications to equivalent embodiments using the technical contents disclosed above. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A big data-based all-media operation intelligent analysis platform, characterized by: The platform includes: Data collection module: Using distributed crawler technology, we write customized scripts, integrate official social media API calls and data interface docking or push from partner platforms to collect all media data including text, images, audio, and video from web pages, social media, and news clients in real time; Data preprocessing module: cleans the collected raw data, converts text, image, audio, and video data into feature vectors in a unified format, and fuses them through a dynamic weight fusion algorithm; Sentiment and semantic analysis module: This module creates a sentiment tendency judgment model to judge the sentiment tendency of the fused data, uses semantic role labeling technology to understand the semantic relationships and implicit information in the text, and uses knowledge graphs to associate entities and relationships in the text; Trend hotspot analysis module: organizes media data into time series and captures temporal features; uses hotspot cluster dynamic screening algorithm to cluster texts and screen potential hotspot clusters; combines hotspot trend dynamic prediction model to analyze and predict media content development trends and hot topics. Decision support module: Integrates the results of the sentiment semantic analysis module and the trend hotspot analysis module, generates decision recommendations based on the decision recommendation knowledge base, and uses visualization tools to present the analysis results and recommendations in the form of charts and text; Interactive feedback module: Develop a responsive interface compatible with multiple terminals, support users to query data and customize analysis tasks through natural language input or operation instructions, collect user evaluations of analysis results and suggestions through feedback portals, and use them to optimize platform models and strategies after analysis.

2. The big data-based all-media operation intelligent analysis platform according to claim 1, characterized in that: The full media data collected in the data collection module is: Text data includes: web page content, social media text, news client content, and user interaction data; Image data includes: web page images, social media images, and news client images; Audio data includes: social media audio, news client audio, and web page audio; Video data includes: web videos, social media videos, and news client videos.

3. The big data-based all-media operation intelligent analysis platform according to claim 1, characterized in that: The data preprocessing module fuses text, image, and audio data in the same standard format using a dynamic weight fusion algorithm. The calculation formula of the dynamic weight fusion algorithm is: ,in, is the fused multimodal feature vector, is the modal index, Represent text, image, audio, and video modes respectively. It is The eigenvectors of the modes, is the text modality feature vector, is the image modality feature vector, is the audio modal eigenvector, is the video modality feature vector, It is Moment The weight of the modality, It is Modality attention weights.

4. The big data-based all-media operation intelligent analysis platform according to claim 3, characterized in that: The calculation of parameters in the dynamic weight fusion algorithm, Moment Modal weight The calculation formula is: ,in represents the natural exponential function, which is used to convert the importance score of the mode Converted to nonlinear weight values, For the Moment Mode The importance score is calculated as follows: ,in is modal The temporal variance of the feature, is modal Feature correlations with other modalities, is modal The variance correlation balance coefficient is obtained through historical data statistics; Modality attention weight The calculation formula is: ,in is modal The weight matrix, is modal The key feature vector of is the attention dimension, used to scale dot product calculations, and is set according to the model design and task requirements. It is The eigenvectors of the modes, is a symbolic function that converts multiple real numbers into a probability distribution.

5. The big data-based all-media operation intelligent analysis platform according to claim 1, characterized in that: The construction formula of the emotional tendency judgment model in the emotional semantic analysis module is: ,in is the fused multimodal feature vector, It is sentiment polarity projection vector, is the projection vector weight coefficient, which is obtained by optimizing the sentiment classification task through training data. is the emotion intensity adjustment factor, which is dynamically adjusted based on the modal credibility. is a symbolic function, output Represent negative, neutral, and positive emotions respectively. Represents the number of sentiment polarity projection vectors.

6. The big data-based all-media operation intelligent analysis platform according to claim 1, characterized in that: The specific steps of understanding text semantic relationships and implicit information through semantic role labeling technology in the sentiment semantic analysis module are as follows: (1) Extract verbs / adjectives in sentences as predicates and distinguish their semantics through contextual disambiguation; (2) Label the core arguments and peripheral arguments of the predicate and construct a predicate-argument relationship matrix; (3) Identify the modification relationship between arguments through dependency syntax analysis and extract the sentiment polarity of the modifiers; (4) Complete omitted elements, identify metaphorical expressions, and deduce implicit logical relationships; (5) Align entities in images / audio with textual arguments to verify cross-modal semantic consistency; (6) Mapping semantic roles to knowledge graph nodes and completing implicit semantics through graph relationships; (7) Generate results that include predicate-argument structure, implicit information, and multimodal associations.

7. The big data-based all-media operation intelligent analysis platform according to claim 1, characterized in that: The trend hotspot analysis module uses a hotspot cluster dynamic screening algorithm to cluster texts and screen potential hotspot clusters. The formula is: ,in It is a cluster The text density, is the cluster after text clustering, It is a cluster Mean semantic similarity of inner text It is a cluster Normalized semantic distance to the nearest high-density cluster, It is a dynamic threshold that is adjusted dynamically based on time and data volume.

8. The big data-based all-media operation intelligent analysis platform according to claim 1, characterized in that: The trend hotspot analysis module predicts the development trend of media content and hot topics through the hotspot trend dynamic prediction model. The calculation formula is: ,in, is the cluster after text clustering, Hotspot cluster The importance weight of is: , Hotspot cluster The heat growth rate reflects the growth of the heat cluster over time. represent Hotspot cluster collection filtered at all times Medium Other hotspot clusters outside yes The hotspot cluster set filtered out at a moment includes all hotspot clusters that meet the hotspot conditions at that moment. Indicates Hotspot cluster collection filtered at all times Except for specific hotspot clusters Other hotspot clusters The text density, It is predicted Hot trend value at the moment, yes The hotspot cluster collection filtered out at all times, Hotspot cluster The predicted value of the popularity change trend, Hotspot cluster The similarity between the semantics of and historical outbreak topics, is the trend factor weight, which is determined by training and verification of historical hotspot data and has a value range of , It's time.

9. The big data-based all-media operation intelligent analysis platform according to claim 1, characterized in that: The decision support module contains the following content: Industry case library: success stories and failure stories; Policy template library: general policy templates and industry-customized policy templates; Market environment information database: macroeconomic data, policy and regulatory information, and industry trend analysis reports; Expert experience knowledge base: expert opinions and suggestions and expert case interpretations; Internal enterprise data and experience database: historical decision-making data and internal business processes and specifications.

10. The big data-based all-media operation intelligent analysis platform according to claim 1, characterized in that: The decision support module generates decision suggestions based on the decision suggestion knowledge base through the dual-source fusion decision algorithm. The formula is: ,in It is the best decision suggestion. is a candidate suggestion in the decision suggestion knowledge base, is the characteristic vector of the current analysis result, It is a candidate suggestion The corresponding result feature vector, is the characteristic vector of the current market environment, It is a candidate suggestion The corresponding environmental feature vector, It is the similarity weight of the analysis results, which is set according to the specific application scenario and decision-making objectives. The value range is: , is the cosine similarity function.

Images