Integrated media topic selection planning method based on AI technology and media industry data elements

Abstract

The invention discloses an AI technology and media industry data element-based fusion media topic selection planning method, which comprises the following steps of: acquiring fusion media data from a fusion media platform in real time through an API (Application Program Interface) and a distributed crawler, performing data cleaning, alignment and multi-modal data standardization, then extracting key features and constructing a dynamic knowledge graph; a user inputs keywords and selects question types, semantic features of the keywords are extracted, and fusion media data are retrieved according to the semantic features and the question types to generate a selected question recommendation scheme; constructing a propagation value prediction model to analyze and predict the generated topic recommendation scheme, constructing a topic recommendation scheme auditing model, and auditing the generated topic recommendation scheme by adopting an embedded security mechanism; and sending the selected topic recommendation scheme passing the selected topic recommendation scheme auditing and the corresponding propagation value prediction result to the user. According to the invention, the efficiency and accuracy of topic selection planning are improved, and the method has high practicability and popularization value.

Description

Technical Field

[0001] The present invention belongs to the field of converged media technology, and specifically relates to a converged media topic planning method based on AI technology and media industry data elements. Background Art

[0002] Existing converged media topic selection methods commonly suffer from problems such as a single data dimension, insufficient cross-platform relevance, and limited content adaptation capabilities. Traditional systems rely on manual rules or single data sources (such as text keyword extraction), lacking the ability to integrate and analyze multimodal data such as videos and live broadcasts, and are unable to dynamically capture the evolving patterns of public opinion events. Furthermore, the separation of topic generation and security review processes leads to inefficient planning and the risk of lags. For example, currently common technical solutions include using crawlers to collect news data and extract keywords based on TF-IDF to generate topic suggestions, or using user click-through rate data to train recommendation models and generate similar topics. However, these solutions suffer from numerous issues, including a single data dimension and a failure to integrate multimodal data (such as video comments and live broadcast interaction data); a lack of dynamic knowledge graph support, making it difficult to identify potential related hotspots; a lack of multimodal content adaptation capabilities in topic generation (such as automatically generating short video script frameworks); and a lack of a coordinated mechanism between topic planning and content security review. Summary of the Invention

[0003] In response to the above-mentioned shortcomings, the present invention discloses a method for integrated media topic planning based on AI technology and media industry data elements. Through multi-level data integration, dynamic knowledge graph construction and embedded security mechanism, it provides users with topic planning reports from multiple dimensions such as historical big data, current affairs, content value, interaction, dissemination timeliness, and audience groups, thereby improving the efficiency and accuracy of topic planning, having strong practicality and promotion value, and being able to continuously improve its own topic planning scheme through AI incremental learning.

[0004] The present invention is achieved by adopting the following technical solutions: A method for selecting and planning topics for integrated media based on AI technology and media industry data elements, comprising the following steps: (1) Through API interfaces and distributed crawlers, real-time acquisition of converged media data from the converged media platform. The converged media data includes structured data and unstructured data; the structured data includes metadata of manuscripts and user interaction records; the unstructured data includes videos, live streams, bullet comments, and comments. The federated learning framework (FATE) is used to integrate user behavior data of the converged media platform through horizontal federated learning to build a global user portrait model. (2) The integrated media data obtained in step (1) is cleaned, aligned, and multimodal data is standardized. For video or live broadcast data, a null feature extraction model (3D-CNN) is used to extract key frame features. For text data, BERT+CRF is used for entity recognition and semantic annotation. The heterogeneous data is mapped to the same semantic space through a unified knowledge representation framework (RDF triples). Based on the graph attention network (GAT), the cross-platform entity similarity is calculated, and the homonymous or synonymous entities in data at different levels are aligned. A dynamic knowledge graph is constructed, and the temporal graph neural network (TGCN) is used to dynamically update the node relationship according to the time evolution law of the public opinion event. A new subgraph is generated for the emergency in real time, and is associated with the historical event graph through the cross-event propagation path prediction model. (3) The user inputs keywords and selects a report type, extracts the semantic features of the keywords, retrieves the converged media data obtained in step (1) based on the semantic features and report type, arranges the retrieved converged media data based on similarity and time sequence, selects the top 3 to 5 converged media data, and generates a topic recommendation scheme, which includes an overview of the topic content, the reporting angle, and the media formats that can be used; (4) Construct a communication value prediction model to analyze and predict the topic recommendation scheme generated in step (3). GPT-4+TextRank is used to extract keywords from the topic content overview and calculate information entropy to measure the novelty of the content. The historical communication path is analyzed based on heterogeneous graph neural network (HGNN). The interest matching probability of the target user is predicted through collaborative filtering and dynamic knowledge graph embedding (KGE). The Transformer+Prophet model is used to predict the communication peak value by combining the factors of topic release time and holidays. The machine learning method based on multi-task learning is used to optimize the three prediction results of communication breadth, depth and continuity. (5) Construct a topic recommendation review model and use an embedded security mechanism to review the topic recommendation generated in step (3). Call the multimodal sensitive content detection model (ViLBERT) to process the topic recommendation, identify sensitive elements in the video and risk entities in the text, and push high-risk topics to manual review when they are detected; (6) The topic recommendation plan that has passed the topic recommendation plan review and the corresponding communication value prediction results will be sent to the user. The user will create a manuscript based on the topic recommendation plan and publish it on the integrated media platform.

[0005] Furthermore, the integrated media platform includes provincial, municipal, district and county integrated media platforms.

[0006] Furthermore, in the process of constructing the dynamic knowledge graph described in step (2), for video data, the CLIP model is used to extract visual-textual cross-modal features, and the spatiotemporal hotspot labels are generated in combination with the timestamps of the bullet comments; for live broadcast data, the emotional fluctuations of real-time comments are analyzed through the LSTM+attention mechanism, and high-interaction segments are identified and marked as outbreak nodes of the knowledge graph; in the graph structure, a hierarchical topology is adopted, with the provincial node as the root node, and the sub-graphs of prefecture-level cities and / or districts are linked downward, and the multi-dimensional entities of public opinion events, media resources, and communication channels are linked horizontally. Based on the speed of public opinion propagation and the frequency of user interaction, the reinforcement learning model (DQN) is used to dynamically adjust the relationship weights between nodes.

[0007] Furthermore, real-time updates of dynamic knowledge graphs are achieved through an incremental graph embedding algorithm (Node2Vec+Online Learning); high-risk nodes are automatically associated with the audit rule base to trigger embedded security audits.

[0008] Furthermore, the report topic types in step (3) are livelihood, current affairs, sports, and entertainment; and the media forms include text, video, and audio.

[0009] Furthermore, in step (5), an adversarial training mechanism is adopted to generate adversarial samples using the GAN network, and the topic recommendation review model is regularly optimized to improve its robustness in identifying new sensitive content.

[0010] Furthermore, in step (6), the entire life cycle data of topic recommendation schemes, manuscript creation and manuscript dissemination is collected to update the dynamic knowledge graph, and the incremental learning method is used to use the entire life cycle data of topic recommendation schemes, manuscript creation and manuscript dissemination as samples for model optimization.

[0011] Compared with the existing technology, this technical solution has the following beneficial effects: 1. This invention combines artificial intelligence, knowledge bases, public opinion monitoring systems, and communication value prediction systems for multi-dimensional considerations, providing users with topic planning recommendations, continuously proposing reporting corrections and suggestions during the activity, and ultimately enabling incremental learning. It uses a federated learning framework to integrate user behavior data from three platforms through horizontal federated learning, without leaving the data locally. This constructs a global user profile model, resolves data silos, and enables intelligent fusion analysis of multi-source heterogeneous media data.

[0012] 2. This invention not only standardizes multimodal data and performs multimodal semantic fusion, but also leverages a temporal graph neural network (TGCN) to dynamically update node relationships based on the temporal evolution of public opinion events. For example, a breaking news event generates a new subgraph in real time, which is then linked to the historical event graph using a cross-event propagation path prediction model.

[0013] 3. In the process of constructing the dynamic knowledge graph, the present invention supports the joint representation of multi-source data such as text, video, live broadcast, and barrage, and realizes real-time updating of the graph through an incremental graph embedding algorithm (Node2Vec+Online Learning) without the need for full reconstruction.

[0014] 4. The present invention adopts an embedded security mechanism to review the topic planning scheme, calls the multimodal sensitive content detection model, identifies sensitive elements in the video screen and risk entities in the text, and realizes the security review of multimodal content. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flowchart of the integrated media topic planning method based on AI technology and media industry data elements described in Example 1. DETAILED DESCRIPTION

[0016] The present invention is further illustrated by the following examples, which are not intended to limit the present invention. Specific experimental conditions and methods not specified in the following examples are conventional methods well known to those skilled in the art.

[0017] The scheme described in the following embodiments is based on the three-level integrated knowledge base of media integration in Guangxi Zhuang Autonomous Region and the cloud public opinion guidance platform of Guangxi Zhuang Autonomous Region established by the applicant of the present invention. The three-level integrated knowledge base of media integration in Guangxi Zhuang Autonomous Region contains the media history and new manuscript resources of the entire region of Guangxi Zhuang Autonomous Region. The cloud public opinion guidance platform of Guangxi Zhuang Autonomous Region has real-time monitoring of public opinion data on the entire network, multi-dimensional intelligent analysis (including emotional tendencies, communication trends, topic focus and media channel portraits) and automatic early warning functions, supports users to customize monitoring plans and generate visual reports.

[0018] Example 1: A method for selecting and planning topics for integrated media based on AI technology and media industry data elements, comprising the following steps: (1) Through API interfaces and distributed crawlers, real-time acquisition of converged media data from the three-level integrated knowledge base of media convergence in Guangxi Zhuang Autonomous Region is carried out. The converged media data includes structured data and unstructured data; the structured data includes metadata of manuscripts and user interaction records; the unstructured data includes videos, live streams, bullet screens, and comments; the federated learning framework (FATE) is used to integrate user behavior data of the converged media platform through horizontal federated learning to build a global user portrait model; (2) The integrated media data obtained in step (1) is cleaned, aligned, and multimodal data is standardized. For video or live broadcast data, a null feature extraction model (3D-CNN) is used to extract key frame features. For text data, BERT+CRF is used for entity recognition and semantic annotation. The heterogeneous data is mapped to the same semantic space through a unified knowledge representation framework (RDF triples). Based on the graph attention network (GAT), the cross-platform entity similarity is calculated, and the homonymous or synonymous entities in data at different levels are aligned. A dynamic knowledge graph is constructed, and the temporal graph neural network (TGCN) is used to dynamically update the node relationship according to the time evolution law of the public opinion event. A new subgraph is generated for the emergency in real time, and is associated with the historical event graph through the cross-event propagation path prediction model. During the construction of the dynamic knowledge graph, for video data, the CLIP model is used to extract visual-textual cross-modal features, and the spatiotemporal hotspot labels are generated in combination with the timestamps of the bullet comments. For live broadcast data, the LSTM+attention mechanism is used to analyze the emotional fluctuations of real-time comments, identify highly interactive segments, and mark them as outbreak nodes in the knowledge graph. In the graph structure, a hierarchical topology is adopted, with provincial nodes as the root nodes, linking to sub-graphs of prefecture-level cities and / or counties downward, and horizontally linking multi-dimensional entities of public opinion events, media resources, and communication channels. The nodes contain timestamps and geographic tags, supporting the spatiotemporal communication analysis of regional events such as the "March 3 Folk Customs Activities in Guangxi Zhuang Autonomous Region"; Based on the speed of public opinion dissemination and the frequency of user interaction, a reinforcement learning model (DQN) is used to dynamically adjust the weights of relationships between nodes. For example, the weight of the spread of emergencies is significantly increased in the short term. An incremental graph embedding algorithm (Node2Vec + Online Learning) is used to achieve real-time updates of dynamic knowledge graphs. (3) The user inputs a keyword and selects a report type, extracts the semantic features of the keyword, retrieves the integrated media data obtained in step (1) based on the semantic features and the report type, arranges the retrieved integrated media data based on similarity and time sequence, selects the top 3 to 5 integrated media data to generate a topic recommendation scheme, and the topic recommendation scheme includes an overview of the topic content, the reporting angle, and the media formats that can be used; the report types include people's livelihood, current affairs, sports, and entertainment; and the media formats include text, video, and audio; (4) Construct a communication value prediction model to analyze and predict the topic recommendation scheme generated in step (3), using GPT-4+TextRank to extract keywords from the topic content overview and calculate information entropy to measure the novelty of the content. Based on the heterogeneous graph neural network (HGNN), the historical communication path (such as provincial media → prefecture-level big V → district and county communities) is analyzed, and the interest matching probability of the target user is predicted through collaborative filtering and dynamic knowledge graph embedding (KGE); the Transformer+Prophet model is used to combine the factors of topic release time and holidays to predict the communication peak; the machine learning method based on multi-task learning is used to optimize the three prediction results of communication breadth (click volume), depth (forwarding level), and persistence (heat decay cycle); (5) Construct a topic recommendation review model and use an embedded security mechanism to review the topic recommendation generated in step (3), call the multimodal sensitive content detection model (ViLBERT) to process the topic recommendation, identify sensitive elements in the video (such as flags, logos) and risk entities in the text, and automatically associate high-risk nodes (such as sensitive public opinion) with the review rule library to trigger an embedded security review; when a high-risk topic is detected, it is automatically associated with the "1+14+111+N" handling mechanism of the Guangxi Zhuang Autonomous Region Cloud Public Opinion System and pushed to the Cyberspace Administration of China and the corresponding media level for manual review; (6) The topic recommendation plan that has passed the topic recommendation plan review and the corresponding communication value prediction results will be sent to the user. The user will create a manuscript based on the topic recommendation plan and publish it on the integrated media platform.

[0019] Example 2: The method for integrated media topic planning based on AI technology and media industry data elements described in this example differs from the method described in Example 1 only in that, in step (5), an adversarial training mechanism is adopted, and adversarial samples are generated using a GAN network, and the topic recommendation scheme review model is regularly optimized to improve its robustness in recognizing new sensitive content; in step (6), the entire life cycle data of topic recommendation schemes, creative manuscripts, and manuscript dissemination are collected to update the dynamic knowledge graph, and at the same time, the entire life cycle data of topic recommendation schemes, creative manuscripts, and manuscript dissemination are used as samples for model optimization using an incremental learning method.

[0020] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A method for selecting and planning topics for integrated media based on AI technology and media industry data elements, characterized by: The following steps are involved: (1) Through API interfaces and distributed crawlers, real-time acquisition of converged media data from the converged media platform. The converged media data includes structured data and unstructured data; the structured data includes metadata of manuscripts and user interaction records; the unstructured data includes videos, live streams, bullet comments, and comments; using a federated learning framework, through horizontal federated learning, the user behavior data of the converged media platform is integrated to build a global user portrait model; (2) The integrated media data obtained in step (1) is cleaned, aligned, and multimodal data is standardized. For video or live broadcast data, a null feature extraction model is used to extract key frame features, and BERT+CRF is used for entity recognition and semantic annotation of text data. The heterogeneous data is mapped to the same semantic space through a unified knowledge representation framework; cross-platform entity similarity is calculated based on the graph attention network, and homonymous or synonymous entities in data at different levels are aligned; a dynamic knowledge graph is constructed, and a time-series graph neural network is used to dynamically update the node relationship according to the time evolution law of public opinion events, and a new subgraph is generated for emergencies in real time, and is associated with the historical event graph through a cross-event propagation path prediction model; (3) The user inputs keywords and selects a report type, extracts the semantic features of the keywords, retrieves the converged media data obtained in step (1) based on the semantic features and report type, arranges the retrieved converged media data based on similarity and time sequence, selects the top 3 to 5 converged media data, and generates a topic recommendation scheme, which includes an overview of the topic content, the reporting angle, and the media formats that can be used; (4) Construct a communication value prediction model to analyze and predict the topic recommendation scheme generated in step (3), using GPT-4+TextRank to extract keywords from the topic content overview and calculate information entropy to measure the content novelty. The historical communication path is analyzed based on heterogeneous graph neural network, and the interest matching probability of the target user is predicted through collaborative filtering and dynamic knowledge graph embedding. The Transformer+Prophet model is used to predict the communication peak value by combining the factors of topic release time and holidays. The machine learning method based on multi-task learning is used to optimize the three prediction results of communication breadth, depth and continuity. (5) Construct a topic recommendation review model and use an embedded security mechanism to review the topic recommendation generated in step (3), call a multimodal sensitive content detection model to process the topic recommendation, identify sensitive elements in the video and risk entities in the text, and push high-risk topics to manual review when they are detected; (6) The topic recommendation plan that has passed the topic recommendation plan review and the corresponding communication value prediction results will be sent to the user. The user will create a manuscript based on the topic recommendation plan and publish it on the integrated media platform.

2. The method for integrated media topic planning based on AI technology and media industry data elements according to claim 1 is characterized by: The aforementioned integrated media platforms include provincial, municipal, district and county integrated media platforms.

3. The method for integrated media topic planning based on AI technology and media industry data elements according to claim 1 is characterized by: In the process of constructing the dynamic knowledge graph described in step (2), for video data, the CLIP model is used to extract visual-textual cross-modal features, and the spatiotemporal hotspot labels are generated by combining the timestamps of the bullet comments. For live broadcast data, the emotional fluctuations of real-time comments are analyzed through the LSTM+attention mechanism, and high-interaction segments are identified and marked as outbreak nodes of the knowledge graph. In the graph structure, a hierarchical topology is adopted, with provincial nodes as the root nodes, and sub-graphs of cities and / or districts and counties as the downward links, and multi-dimensional entities of public opinion events, media resources, and communication channels are horizontally linked. Based on the speed of public opinion propagation and the frequency of user interaction, a reinforcement learning model is used to dynamically adjust the relationship weights between nodes.

4. The method for integrated media topic planning based on AI technology and media industry data elements according to claim 1 is characterized by: Real-time updates of dynamic knowledge graphs are achieved through incremental graph embedding algorithms; high-risk nodes are automatically associated with the audit rule base to trigger embedded security audits.

5. The method for integrated media topic planning based on AI technology and media industry data elements according to claim 1 is characterized by: The report topic types in step (3) are livelihood, current affairs, sports, and entertainment; the media forms include text, video, and audio.

6. The method for integrated media topic planning based on AI technology and media industry data elements according to claim 1 is characterized by: In step (5), an adversarial training mechanism is adopted to generate adversarial samples using the GAN network, and the topic recommendation review model is regularly optimized.

7. The method for integrated media topic planning based on AI technology and media industry data elements according to claim 1 is characterized by: In step (6), the data of the entire life cycle of topic recommendation schemes, manuscript creation and manuscript dissemination are collected to update the dynamic knowledge graph, and the incremental learning method is used to use the data of the entire life cycle of topic recommendation schemes, manuscript creation and manuscript dissemination as samples for model optimization.

Images