Intelligent news broadcasting system and method based on deep learning

Abstract

The invention discloses an intelligent news broadcasting system and method based on deep learning, and the method comprises the following steps: S1, constructing a multi-modal news data set, carrying out the semantic coding, and generating a preliminary semantic feature; s2, extracting a time feature by using a dynamic time convolutional network, and generating the time feature and a hierarchical semantic feature through a hierarchical Transform; s3, performing cross-modal alignment on the time features and the hierarchical semantic features to generate cross-modal fusion features; s4, performing multi-scale semantic modeling, constructing a dynamic event tree, and generating a multi-scale dynamic semantic structure; s5, inputting the multi-scale dynamic semantic structure into the improved three-stage generative network to generate summary, details and predictive broadcast content, and pushing the summary, the details and the predictive broadcast content; s6, optimizing a dynamic event tree and a semantic generation rule in combination with user feedback; and S7, synchronously outputting the optimized broadcast content in the forms of characters, voices and videos. According to the method, deep learning, multi-modal alignment, a dynamic modeling method and the like are combined, so that intelligent news broadcasting is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning and news broadcasting, and in particular to a news intelligent broadcasting system and method based on deep learning. Background Art

[0002] With the rapid development of the Internet and artificial intelligence technology, the way of news dissemination has undergone multiple innovations from traditional media to digital media and then to intelligent media. At this stage, news broadcasting has gradually evolved from a single text or voice form to a comprehensive dissemination form that integrates multimodal information, including a combination of text, images, videos and other forms to meet users' needs for real-time, diverse and accurate information acquisition. However, the existing news broadcasting system still has obvious technical limitations when dealing with multimodal data processing, complex event modeling and user personalized needs.

[0003] At present, many news broadcast systems rely on template-based or rule-driven methods for content generation. Although this method can meet basic automation needs, it shows significant limitations when dealing with complex news events or multimodal data. Specifically, template-based methods are usually unable to flexibly adapt to the dynamic development of events, and the generated broadcast content is often rigid and lacks real-time adjustment capabilities. In addition, when faced with multimodal data such as images and videos, existing technologies find it difficult to achieve semantic alignment and deep fusion of data, resulting in a low semantic correlation between the generated content and actual news events, making it difficult to meet users' requirements for the authenticity and diversity of news broadcasts.

[0004] Complex news events usually have the characteristics of multi-stage dynamic evolution, including the occurrence of events, development of events, public reactions and subsequent processing. This dynamic change in the time dimension puts higher modeling requirements on news broadcasting. However, most existing technologies use static modeling methods, which cannot fully capture the temporal dynamic characteristics of news events. For example, for the broadcast of breaking news, existing methods are often limited to quickly extracting basic information about the event, but lack the ability to dynamically track the development of the event and update semantics in subsequent reports. This defect makes the broadcast content unable to reflect the latest developments of the event, reducing the timeliness and coherence of the broadcast.

[0005] The ability to process multimodal data is also an important technical bottleneck faced by existing news broadcasting systems. News content is usually presented in the form of text, images, and videos, and each modal data contains specific semantic information. However, when processing multimodal data, existing technologies often use independent analysis methods, and lack a unified semantic modeling and alignment mechanism for different modal data. Due to the different characteristics of each modal data, there may be differences or conflicts in semantic information between modalities. It is difficult for existing systems to solve these problems through effective fusion methods, resulting in the lack of correlation and consistency between multimodal generated news content, which further affects the accuracy and expressiveness of the broadcast.

[0006] User interaction feedback is an important part of the news intelligent broadcasting system, which can directly reflect the user's attention and preference for news content. However, existing technologies are insufficient in optimizing the generation of content using user feedback. Although some systems can sort popular content through simple user behavior analysis (such as number of clicks and browsing time), they cannot effectively integrate users' real-time interaction data into the generation model for dynamic adjustment of broadcast content. The lack of in-depth use of user feedback makes it difficult for news broadcasts to meet users' personalized needs and improve the accuracy of content generation.

[0007] In addition, the existing news broadcasting system also has deficiencies in the hierarchical and predictive capabilities of content generation. News events usually contain core information (such as the event subject, time clues) and background information (such as related historical events or peripheral details), but the existing system lacks effective distinction between semantic levels when generating broadcast content, resulting in lengthy and lack of focus in the generated content. At the same time, the predictive broadcasting ability for future development trends of events is weak. Existing technologies often only provide a description of the current state of the event, but lack reasonable speculation on the possible future evolution direction, which further limits the practicality of the broadcasting system.

[0008] Therefore, how to provide a news intelligent broadcasting system and method based on deep learning is an urgent problem that technicians in this field need to solve. Summary of the invention

[0009] One purpose of the present invention is to propose a news intelligent broadcasting system and method based on deep learning. The present invention combines deep learning technology, multimodal data alignment mechanism and dynamic event modeling method to fully realize the intelligence and efficiency of news broadcasting. By introducing the dynamic time convolutional network and hierarchical Transformer architecture, the temporal features and semantic features of the news content are accurately extracted, the semantic alignment of multimodal data is achieved by cross-modal comparative learning, and the modeling ability of the multi-stage evolution of news events is improved by the construction of a dynamic event tree. The system has strong real-time performance, accurate content, rich levels and personalized adaptability, and can be widely used in the fields of intelligent media and online news services, greatly optimizing the user experience.

[0010] A news intelligent broadcasting method based on deep learning according to an embodiment of the present invention comprises the following steps: S1. Collect text data, image data and video data of news events, build a multimodal news dataset and perform semantic encoding to generate preliminary semantic features; S2, extracting the time dimension of the preliminary semantic features based on a dynamic time convolutional network, and semantically grading the news content through a hierarchical Transformer architecture to generate time features and hierarchical semantic features; S3, through cross-modal contrastive learning driven by event characteristics, aligning the temporal features and hierarchical semantic features with multimodal news data to generate cross-modal fusion features; S4. Using a multi-head self-attention mechanism to perform multi-scale semantic modeling on the cross-modal fusion features, and constructing a dynamic event tree based on time nodes and semantic relevance to generate a multi-scale dynamic semantic structure; S5, inputting the multi-scale dynamic semantic structure into the improved three-stage generation network, sequentially generating summary broadcast content, detailed broadcast content, and predictive broadcast content based on event evolution, and pushing them to the user terminal in real time; S6. Based on the user interaction feedback information, the event feedback optimization mechanism is used to dynamically adjust the node weights and semantic generation rules of the dynamic event tree to generate optimized broadcast content; S7. Synchronously output the optimized broadcast content in the form of text, voice and video.

[0011] Optionally, the S2 specifically includes: S21. Embed the preliminary semantic features and embed the feature vector using the improved position encoding method: ; in, Represents the partial code value of the posth eigenvector at the even dimension index 2q, represents the partial code value of the posth feature vector at the odd dimension index 2q+1, pos represents the position index of the feature vector, exp represents the exponential function, t represents the time node, and d represents the dimension of the embedded feature vector. represents the time decay factor; S22. Use dynamic temporal convolutional networks to extract the time dimension of embedded coding features, and capture the time features in irregular time series through dynamically adjusted convolution kernels: ; in, represents the output feature value of time step t, k represents the size of the convolution kernel, represents the weight parameter of the convolution kernel, represents the input feature value at time step tr, represents the time node difference, b represents the bias term; S23. Construct the first layer of hierarchical Transformer, use the multi-head self-attention mechanism to semantically encode the time features extracted by the dynamic time convolutional network, and extract the core semantic features in the news content. The core semantics includes the subject, keywords and time node features of the news event: ; in, represents the core semantic features, softmax represents the normalization function, The query vector representing the core semantic features, The key vector representing the core semantic features, The transpose of the key vector representing the core semantic features, A value vector representing the core semantic features, The weight matrix representing the time feature, represents the normalization factor; S24. Construct the second layer of hierarchical Transformer, perform background semantic decoding on the core semantic features, and extract secondary semantic features: ; in, Represents secondary semantic features, The query vector representing the secondary semantic features, The key vector representing the secondary semantic features, The transpose of the key vector representing the secondary semantic features, A value vector representing a secondary semantic feature, and represents the weight parameter, Positional encoding to represent background semantics; S25. Based on the core semantic features and secondary semantic features, the temporal features and hierarchical semantic features are generated through a multiple dynamic weighting mechanism of the temporal dimension and the semantic level: ; in, Represents temporal features and hierarchical semantic features, Represents the weight matrix of core semantic features, Represents the weight matrix of secondary semantic features, Represents time characteristics.

[0012] Optionally, the event feature-driven cross-modal contrastive learning in S3 specifically includes: obtaining a set of event feature information, including time clues, semantic subjects and background features, as a guiding signal for feature alignment; based on cross-modal contrastive learning, aligning time features and hierarchical semantic features with multimodal news data.

[0013] Time clues: refers to the time node or time period when a news event occurs, such as timestamps, time sequence, and event development stages. This information is used to mark and organize the time dimension of news data and provide a time reference for subsequent alignment. Time clues help the feature alignment model to concentrate data with similar time characteristics (such as images and text at different time points of the same event) into a unified time semantic space.

[0014] Semantic subject: refers to the core object or subject of a news event, such as the relevant people, places or topics of a news event. These semantic subjects are the core information of the alignment process. By extracting semantic subjects, the focus on the core content in the alignment process can be enhanced, and irrelevant or noise information can be ignored.

[0015] Background features: refers to the contextual information surrounding the news event, such as the background environment mentioned in the news, related historical events, or supplementary description information. The introduction of background features can provide broader semantic support for alignment, especially when multimodal data (such as video or image background) is involved, which helps to improve the accuracy and completeness of alignment.

[0016] As guidance signals for feature alignment: After temporal clues, semantic subjects, and background features are extracted, a set of multi-dimensional guidance signals (usually called event feature vectors) are formed. These vectors are used to guide cross-modal feature alignment, so that data from different modalities (text, images, and videos) can find common reference points in the semantic space.

[0017] Cross-modal contrastive learning is a machine learning method that aims to compare and match features of different modalities (such as text, images, videos, etc.).

[0018] Core idea: By calculating the similarity of feature vectors from different modalities, data with high semantic associations are screened out and data with low associations are ignored.

[0019] In this scheme, temporal features and hierarchical semantic features come from dynamic event trees, respectively, while multimodal news data comes from different modalities (e.g., video, image, and text).

[0020] Comparison method: By comparing temporal features with hierarchical semantic features, and comparing these features with text, image, and video modal data, the model can learn the semantic alignment relationship between modalities.

[0021] Temporal features are aligned with hierarchical semantic features: Temporal features: derived from the time nodes of the dynamic event tree, such as the time sequence of news and the development stage of events. Through alignment, cross-modal data is ensured to be consistent in the time dimension, for example, images, texts, and video content in the same time period can be matched to the corresponding temporal features.

[0022] Hierarchical semantic features: derived from the semantic hierarchy of the dynamic event tree (core layer, sub-core layer, and background layer). Through alignment, the model can associate data of different modalities to the corresponding semantic hierarchy, for example, key semantic information corresponds to the main scene in the image, the core description in the text, etc.

[0023] Multimodal news data alignment: The alignment process includes: projecting temporal features, hierarchical semantic features and multimodal news data into a unified semantic space; calculating the semantic similarity between temporal features, hierarchical semantic features and multimodal data; and screening out the most relevant feature combinations to form cross-modal fusion features.

[0024] Optionally, the S4 specifically includes: S41. Divide the cross-modal fusion features by time nodes to form feature sets of several time segments ,in Indicates time node The corresponding feature subset; S42, use the multi-head self-attention mechanism to perform semantic modeling on the feature subset within each time node, extract local context associations, and obtain preliminary time node features; S43, comparing the features of different preliminary time nodes in pairs, defining the semantic relevance between the time nodes, and calculating the semantic relevance by weighted combination of feature similarity and time difference: ; Among them, Indicates time node and time nodes The semantic association between them, A represents the dimension of the feature vector, Indicates time node The eigenvalue in the kth dimension, Indicates time node In the k-th dimension, exp represents the exponential function. Indicates time node and time nodes The time difference between represents the time decay parameter; S44. Construct a dynamic event tree T=(N,E) based on each time node and semantic relevance, where Represents a set of time nodes and a set of edges , represents the semantic relevance threshold; S45. Divide the time nodes in the dynamic event tree into layers: ; in, Indicates time node The level to which it belongs, Indicates time node The semantic contribution of represents the semantic importance threshold, represents the correlation stratification threshold, represents the time interval threshold, and n represents the total number of time nodes; S46. Iteratively update the dynamic event tree, calculate and correct the edge weights and hierarchical divisions, and finally generate a multi-scale dynamic semantic structure.

[0025] Optionally, the S5 specifically includes: S51, initializing an improved three-stage generation network, including an overview generation module, a detail generation module and a prediction generation module, and loading core layer features, sub-core layer features and background layer features in a multi-scale dynamic semantic structure; S52, in the summary generation module, extracting elements from the core layer features to generate summary broadcast content of emergency and breaking news, the summary generation module including a sequence encoder and an attention decoder, outputting a summary broadcast sequence of the first stage; S53, in the detail generation module, the summary broadcast sequence output by the summary generation module and the sub-core layer features and background layer features in the multi-scale dynamic semantic structure are received, and are integrated into the multimodal decoder, and the event subject, background and context details are supplemented through the fine-grained attention mechanism, and the detail broadcast sequence of the second stage is output; S54, in the prediction generation module, combining the temporal feature information of the summary broadcast sequence, the detailed broadcast sequence and the multi-scale dynamic semantic structure, capturing the future trend of the event evolution trend through the recursive time network, and outputting the predictive broadcast content based on the event evolution; S55, integrating summary broadcast content, detailed broadcast content and predictive broadcast content based on event evolution into a unified multi-modal output buffer area, and cross-validating them to screen out redundant and contradictory content; S56: Output the cross-verified broadcast content to the user terminal through a real-time push channel.

[0026] Optionally, the S6 specifically includes: S61. Collect user interaction feedback information, including user click times, stay time, comment sentiment analysis results and sharing behavior, to form a user feedback matrix ,in Indicates user a's response to a time node Overall feedback score: ; in, Indicates user a's response to a time node MaxClicks represents the maximum number of clicks by all users. Indicates user a's response to a time node The duration of stay, MaxTime indicates the maximum duration of stay of all users, Indicates user a's response to a time node The sentiment score of the review, , and Represents the weight parameter of the feedback indicator; S62. Dynamically update the weight of each time node of the dynamic event tree based on the user feedback matrix: ; in, Indicates time node The update weight of represents the equilibrium parameter, Indicates time node The original weight of Indicates time node The average feedback value of S63. Adjust the semantic generation rules of the dynamic event tree according to the updated weight, and recalculate the semantic contribution of each node: ; in, Indicates time node The semantic contribution of Indicates time node The update weight of n represents the total number of time nodes, tanh represents the hyperbolic tangent function, and ρ represents the amplification factor of the feedback change. Indicates time node The baseline feedback value of S64, based on the updated weights and semantic contributions, dynamically optimize the hierarchical structure of the dynamic event tree, perform pruning operations, and re-divide the core layer and the background layer; S65, using the updated time node weights and semantic generation rules, regenerate summary broadcast content, detailed broadcast content, and predictive broadcast content based on event evolution, so that the semantics of the generated content is consistent with user feedback; S66: Push the optimized broadcast content to the user terminal, and record a new round of user feedback information in real time, forming a closed-loop optimization mechanism of user feedback and dynamic event tree adjustment.

[0027] A news intelligent broadcasting system based on deep learning according to an embodiment of the present invention includes the following modules: The data collection and coding module is used to collect text data, image data and video data of news events, build a multimodal news data set, and perform semantic coding on the multimodal news data to generate preliminary semantic features; The temporal and semantic feature extraction module is used to extract the temporal dimension of the preliminary semantic features based on the dynamic temporal convolutional network, and to semantically classify the news content through the hierarchical Transformer architecture to generate temporal features and hierarchical semantic features; The cross-modal alignment module is used to align temporal features and hierarchical semantic features with multimodal news data through cross-modal contrastive learning driven by event characteristics to generate cross-modal fusion features; The dynamic event tree generation module is used to use the multi-head self-attention mechanism to perform multi-scale semantic modeling on cross-modal fusion features, build a dynamic event tree based on time nodes and semantic relevance, and generate a multi-scale dynamic semantic structure; A three-stage content generation module, which is used to input the multi-scale dynamic semantic structure into the improved three-stage generation network, and sequentially generate summary broadcast content, detailed broadcast content, and predictive broadcast content based on event evolution; The event feedback optimization module is used to dynamically adjust the node weights and semantic generation rules of the dynamic event tree based on user interaction feedback information using the event feedback optimization mechanism to generate optimized broadcast content; The multimodal output module is used to synchronously output the optimized broadcast content in the form of text, voice and video, and transmit it to the user terminal through a real-time push channel.

[0028] The beneficial effects of the present invention are: First, the present invention improves the system's ability to extract the time dimension features of news content by introducing a dynamic time convolutional network and a hierarchical Transformer architecture. The dynamic time convolutional network can accurately capture time features in irregular time series and extract the multi-stage dynamic changes of news events. The hierarchical Transformer architecture semantically classes news content, distinguishes core semantic features (such as news subjects, keywords, and time clues) from secondary semantic features (such as background and auxiliary information), and realizes the hierarchical expression of news content. This hierarchical processing method enables the generated broadcast content to highlight the key points while taking into account the background details, which can meet the needs of users to quickly obtain core information and have an in-depth understanding of the news background.

[0029] Secondly, the system achieves semantic alignment and multi-scale modeling of multimodal news data through cross-modal contrastive learning and multi-head self-attention mechanism driven by event characteristics. Cross-modal contrastive learning uses time clues, semantic subjects and background features as guiding signals to project the features of multimodal data such as text, images and videos into a unified semantic space, significantly improving the semantic consistency between different modal data. The multi-head self-attention mechanism further performs multi-scale semantic modeling on cross-modal fusion features, effectively capturing the dynamic changes of news content at time nodes and the correlation between semantic levels. The dynamic event tree constructed based on time nodes and semantic correlation can not only clearly present the evolution path of news events, but also provide high-quality semantic input for subsequent generation tasks.

[0030] In addition, the improved three-stage generation network designed by the present invention significantly improves the hierarchy and diversity of news broadcast content. Through the progressive processing of the summary generation module, the detail generation module and the prediction generation module, the system can respectively generate a quick summary broadcast of emergency events, complete news details and predictive broadcasts based on event evolution trends. This generation mode ensures the comprehensiveness and foresight of the broadcast content, while meeting the user's needs for rapid response to breaking news and insight into future trends. Especially in the prediction generation stage, the system combines the time node information of the dynamic event tree and the recursive time network to achieve a reasonable speculation on the future development of the event, filling the gap in the existing news broadcast technology in predictive content generation.

[0031] The introduction of user interaction feedback further enhances the dynamic adaptability and personalization capabilities of the system. The present invention constructs a user feedback matrix by collecting interactive data such as the number of clicks, stay time, emotional comments, and sharing behaviors of users, and dynamically adjusts the node weights and semantic generation rules of the dynamic event tree using an event feedback optimization mechanism. This closed-loop optimization mechanism enables the generated broadcast content to adapt to the user's interest changes and personalized needs in real time, significantly improving the accuracy of the content and user satisfaction.

[0032] Finally, the present invention realizes the synchronous output of multimodal content and supports multiple forms of broadcasting, such as text, voice and video. This multimodal output mechanism can not only improve the expressiveness of news broadcasting, but also provide users with a variety of choices, further optimizing the user experience. At the same time, the real-time push function ensures that users can obtain news content in the first place, meeting the high requirements for the timeliness of news broadcasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 This is an overall flow chart of a news intelligent broadcasting method based on deep learning proposed by the present invention; Figure 2 This is a structural diagram of a news intelligent broadcasting system based on deep learning proposed by the present invention. DETAILED DESCRIPTION

[0034] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0035] refer to Figure 1 , a news intelligent broadcasting method based on deep learning, comprising the following steps: S1. Collect text data, image data and video data of news events, build a multimodal news dataset and perform semantic encoding to generate preliminary semantic features; S2, extracting the time dimension of the preliminary semantic features based on a dynamic time convolutional network, and semantically grading the news content through a hierarchical Transformer architecture to generate time features and hierarchical semantic features; S3, through cross-modal contrastive learning driven by event characteristics, aligning the temporal features and hierarchical semantic features with multimodal news data to generate cross-modal fusion features; S4. Using a multi-head self-attention mechanism to perform multi-scale semantic modeling on the cross-modal fusion features, and constructing a dynamic event tree based on time nodes and semantic relevance to generate a multi-scale dynamic semantic structure; S5, inputting the multi-scale dynamic semantic structure into the improved three-stage generation network, sequentially generating summary broadcast content, detailed broadcast content, and predictive broadcast content based on event evolution, and pushing them to the user terminal in real time; S6. Based on the user interaction feedback information, the event feedback optimization mechanism is used to dynamically adjust the node weights and semantic generation rules of the dynamic event tree to generate optimized broadcast content; S7. Synchronously output the optimized broadcast content in the form of text, voice and video.

[0036] In this implementation, S2 specifically includes: S21. Embed the preliminary semantic features and embed the feature vector using the improved position encoding method: ; in, Represents the partial code value of the posth eigenvector at the even dimension index 2q, represents the partial code value of the posth feature vector at the odd dimension index 2q+1, pos represents the position index of the feature vector, exp represents the exponential function, t represents the time node, and d represents the dimension of the embedded feature vector. represents the time decay factor; S22. Use dynamic temporal convolutional networks to extract the time dimension of embedded coding features, and capture the time features in irregular time series through dynamically adjusted convolution kernels: ; in, represents the output feature value of time step t, k represents the size of the convolution kernel, represents the weight parameter of the convolution kernel, represents the input feature value at time step tr, represents the time node difference, b represents the bias term; S23. Construct the first layer of hierarchical Transformer, use the multi-head self-attention mechanism to semantically encode the time features extracted by the dynamic time convolutional network, and extract the core semantic features in the news content. The core semantics includes the subject, keywords and time node features of the news event: ; in, represents the core semantic features, softmax represents the normalization function, The query vector representing the core semantic features, The key vector representing the core semantic features, The transpose of the key vector representing the core semantic features, A value vector representing the core semantic features, The weight matrix representing the time feature, represents the normalization factor; S24. Construct the second layer of hierarchical Transformer, perform background semantic decoding on the core semantic features, and extract secondary semantic features: ; in, Represents secondary semantic features, The query vector representing the secondary semantic features, The key vector representing the secondary semantic features, The transpose of the key vector representing the secondary semantic features, A value vector representing a secondary semantic feature, and represents the weight parameter, Positional encoding to represent background semantics; S25. Based on the core semantic features and secondary semantic features, the temporal features and hierarchical semantic features are generated through a multiple dynamic weighting mechanism of the temporal dimension and the semantic level: ; in, Represents temporal features and hierarchical semantic features, Represents the weight matrix of core semantic features, Represents the weight matrix of secondary semantic features, Represents time characteristics.

[0037] In this embodiment, the event feature-driven cross-modal contrastive learning in S3 specifically includes: obtaining a set of event feature information, including time clues, semantic subjects and background features, as a guiding signal for feature alignment; based on cross-modal contrastive learning, aligning the time features and hierarchical semantic features with the multimodal news data.

[0038] In this implementation, S4 specifically includes: S41. Divide the cross-modal fusion features by time nodes to form feature sets of several time segments ,in Indicates time node The corresponding feature subset; S42, use the multi-head self-attention mechanism to perform semantic modeling on the feature subset within each time node, extract local context associations, and obtain preliminary time node features; S43, comparing the features of different preliminary time nodes in pairs, defining the semantic relevance between the time nodes, and calculating the semantic relevance by weighted combination of feature similarity and time difference: ; in, Indicates time node and time nodes The semantic association between them, A represents the dimension of the feature vector, Indicates time node The eigenvalue in the kth dimension, Indicates time node In the k-th dimension, exp represents the exponential function. Indicates time node and time nodes The time difference between represents the time decay parameter; S44. Construct a dynamic event tree T=(N,E) based on each time node and semantic relevance, where Represents a set of time nodes and a set of edges , represents the semantic relevance threshold; S45. Divide the time nodes in the dynamic event tree into layers: ; in, Indicates time node The level to which it belongs, Indicates time node The semantic contribution of represents the semantic importance threshold, represents the correlation stratification threshold, represents the time interval threshold, and n represents the total number of time nodes; S46. Iteratively update the dynamic event tree, calculate and correct the edge weights and hierarchical divisions, and finally generate a multi-scale dynamic semantic structure.

[0039] In this implementation manner, S5 specifically includes: S51, initializing an improved three-stage generation network, including an overview generation module, a detail generation module and a prediction generation module, and loading core layer features, sub-core layer features and background layer features in a multi-scale dynamic semantic structure; S52, in the summary generation module, extracting elements from the core layer features to generate summary broadcast content of emergency and breaking news, the summary generation module including a sequence encoder and an attention decoder, outputting a summary broadcast sequence of the first stage; S53, in the detail generation module, the summary broadcast sequence output by the summary generation module and the sub-core layer features and background layer features in the multi-scale dynamic semantic structure are received, and are integrated into the multimodal decoder, and the event subject, background and context details are supplemented through the fine-grained attention mechanism, and the detail broadcast sequence of the second stage is output; S54, in the prediction generation module, combining the temporal feature information of the summary broadcast sequence, the detailed broadcast sequence and the multi-scale dynamic semantic structure, capturing the future trend of the event evolution trend through the recursive time network, and outputting the predictive broadcast content based on the event evolution; S55, integrating summary broadcast content, detailed broadcast content and predictive broadcast content based on event evolution into a unified multi-modal output buffer area, and cross-validating them to screen out redundant and contradictory content; S56: Output the cross-verified broadcast content to the user terminal through a real-time push channel.

[0040] In this implementation manner, S6 specifically includes: S61. Collect user interaction feedback information, including user click times, stay time, comment sentiment analysis results and sharing behavior, to form a user feedback matrix ,in Indicates user a's response to a time node Overall feedback score: ; in, Indicates user a's response to a time node MaxClicks represents the maximum number of clicks by all users. Indicates user a's response to a time node The duration of stay, MaxTime indicates the maximum duration of stay of all users, Indicates user a's response to a time node The sentiment score of the review, , and Represents the weight parameter of the feedback indicator; S62. Dynamically update the weight of each time node of the dynamic event tree based on the user feedback matrix: ; in, Indicates time node The update weight of represents the equilibrium parameter, Indicates time node The original weight of Indicates time node The average feedback value of S63. Adjust the semantic generation rules of the dynamic event tree according to the updated weight, and recalculate the semantic contribution of each node: ; in, Indicates time node The semantic contribution of Indicates time node The update weight of n represents the total number of time nodes, tanh represents the hyperbolic tangent function, and ρ represents the amplification factor of the feedback change. Indicates time node The baseline feedback value of S64, based on the updated weights and semantic contributions, dynamically optimize the hierarchical structure of the dynamic event tree, perform pruning operations, and re-divide the core layer and the background layer; S65, using the updated time node weights and semantic generation rules, regenerate summary broadcast content, detailed broadcast content, and predictive broadcast content based on event evolution, so that the semantics of the generated content is consistent with user feedback; S66: Push the optimized broadcast content to the user terminal, and record a new round of user feedback information in real time, forming a closed-loop optimization mechanism of user feedback and dynamic event tree adjustment.

[0041] refer to Figure 2 , a news intelligent broadcasting system based on deep learning, including the following modules: The data collection and coding module is used to collect text data, image data and video data of news events, build a multimodal news data set, and perform semantic coding on the multimodal news data to generate preliminary semantic features; The temporal and semantic feature extraction module is used to extract the temporal dimension of the preliminary semantic features based on the dynamic temporal convolutional network, and to semantically classify the news content through the hierarchical Transformer architecture to generate temporal features and hierarchical semantic features; The cross-modal alignment module is used to align temporal features and hierarchical semantic features with multimodal news data through cross-modal contrastive learning driven by event characteristics to generate cross-modal fusion features; The dynamic event tree generation module is used to use the multi-head self-attention mechanism to perform multi-scale semantic modeling on cross-modal fusion features, build a dynamic event tree based on time nodes and semantic relevance, and generate a multi-scale dynamic semantic structure; A three-stage content generation module, which is used to input the multi-scale dynamic semantic structure into the improved three-stage generation network, and sequentially generate summary broadcast content, detailed broadcast content, and predictive broadcast content based on event evolution; The event feedback optimization module is used to dynamically adjust the node weights and semantic generation rules of the dynamic event tree based on user interaction feedback information using the event feedback optimization mechanism to generate optimized broadcast content; The multimodal output module is used to synchronously output the optimized broadcast content in the form of text, voice and video, and transmit it to the user terminal through a real-time push channel.

[0042] Embodiment 1: In order to verify the feasibility of the present invention in implementation, the present invention is applied to the intelligent news broadcasting system of an online news media platform, which has about 5 million daily active users, covering news events in multiple fields such as politics, technology, sports, and entertainment, with an average daily update volume of more than 3,000 news articles. Users have extremely high requirements for the real-time, diversity, and personalization of news acquisition. However, the traditional news broadcasting system has obvious deficiencies in data processing efficiency, broadcast content accuracy, and user satisfaction. This embodiment illustrates the actual effect of the present invention through specific scenario description and data verification.

[0043] During a certain science and technology press conference, the platform needs to broadcast the press conference related events in real time, including multimodal content such as text summaries, image displays and video interpretation. Traditional systems can often only manually edit news headlines and short articles, which is inefficient and cannot effectively handle the extraction of key content in press conference videos. The present invention uses a data acquisition and encoding module to semantically encode press conference videos, pictures and releases, generates preliminary semantic features, uses a dynamic time convolutional network to extract time dimension features, and distinguishes core semantics from background semantics through a hierarchical Transformer architecture, thereby quickly constructing a dynamic event tree that fully covers the multi-stage evolution of events.

[0044] During the application process, the three-stage generation network of the present invention first generates summary broadcast content based on core semantic features. For example, when the press conference just started, the system generated a breaking news summary including "XX brand releases new AI chip, focusing on high performance and low power consumption" within 30 seconds. As the content of the press conference deepened, the system further extracted secondary semantic features and generated detailed broadcast content, including the chip's technical parameters, application scenarios, market forecasts, etc. In addition, through the time node modeling of the dynamic event tree, the system predicted the media and market response after the press conference, such as "It is expected that the chip will lead the new industry trend in the next six months", and pushed it to the user terminal in real time.

[0045] In order to verify the beneficial effects of the present invention, the operation data of the platform during the press conference was analyzed in detail. In the comparative experiment, the traditional news broadcast system and the method of the present invention were used to process the same press conference news event, and key indicators such as news generation efficiency, user click rate, content matching and user satisfaction were recorded. Table 1 below is the statistical results of the comparative experimental data.

[0046] Table 1 Comparative effect analysis table of the traditional system and the system of the present invention ; It can be seen from the data in Table 1 above that the present invention has shown significant advantages in news generation time, user click-through rate and breaking news response time. In terms of news generation time, the traditional system takes an average of 180 seconds, while the present invention only takes 30 seconds, an improvement of 500%. In terms of user click-through rate, the click-through rate of the content generated by the present invention reached 25.8%, an increase of 106.4% compared to 12.5% ​​of the traditional system. In addition, in the content matching and user satisfaction scores, the present invention reached 9.3 and 9.5 respectively, an increase of 50% and 33.8% respectively compared to the traditional system. Especially in terms of breaking news response time, the performance of the present invention is particularly outstanding, and it can complete the generation and push of breaking news within 15 seconds, while the traditional system takes an average of 300 seconds.

[0047] The present invention dynamically adjusts the news generation rules through the event feedback optimization module, so that the system can adapt to the user's feedback needs in real time. For example, when the user shows a high interest in technical details, the system strengthens the generation of technical parameter content through the weight adjustment mechanism and reduces the proportion of background information, thereby further improving the user's reading experience and satisfaction.

[0048] In summary, the present invention significantly improves the efficiency, accuracy and user satisfaction of news broadcasting in practical applications, effectively solves the shortcomings of traditional systems in multimodal data processing, event dynamic modeling and personalized content generation, and provides an efficient and intelligent solution for the news media industry.

[0049] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A news intelligent broadcasting method based on deep learning, characterized in that: The steps include: S1. Collect text data, image data and video data of news events, build a multimodal news dataset and perform semantic encoding to generate preliminary semantic features; S2, extracting the time dimension of the preliminary semantic features based on a dynamic time convolutional network, and semantically grading the news content through a hierarchical Transformer architecture to generate time features and hierarchical semantic features; S3, through cross-modal contrastive learning driven by event characteristics, aligning the temporal features and hierarchical semantic features with multimodal news data to generate cross-modal fusion features; S4. Using a multi-head self-attention mechanism to perform multi-scale semantic modeling on the cross-modal fusion features, and constructing a dynamic event tree based on time nodes and semantic relevance to generate a multi-scale dynamic semantic structure; S5, inputting the multi-scale dynamic semantic structure into the improved three-stage generation network, sequentially generating summary broadcast content, detailed broadcast content, and predictive broadcast content based on event evolution, and pushing them to the user terminal in real time; S6. Based on the user interaction feedback information, the event feedback optimization mechanism is used to dynamically adjust the node weights and semantic generation rules of the dynamic event tree to generate optimized broadcast content; S7. Synchronously output the optimized broadcast content in the form of text, voice and video.

2. According to the deep learning-based intelligent news broadcasting method of claim 1, it is characterized in that: The S2 specifically includes: S21. Embed the preliminary semantic features and embed the feature vector using the improved position encoding method: ； in, Represents the partial code value of the posth eigenvector at the even dimension index 2q, represents the partial code value of the posth feature vector at the odd dimension index 2q+1, pos represents the position index of the feature vector, exp represents the exponential function, t represents the time node, and d represents the dimension of the embedded feature vector. represents the time decay factor; S22. Use dynamic temporal convolutional networks to extract the time dimension of embedded coding features, and capture the time features in irregular time series through dynamically adjusted convolution kernels: ； in, represents the output feature value of time step t, k represents the size of the convolution kernel, represents the weight parameter of the convolution kernel, represents the input feature value at time step tr, represents the time node difference, b represents the bias term; S23. Construct the first layer of hierarchical Transformer, use the multi-head self-attention mechanism to semantically encode the time features extracted by the dynamic time convolutional network, and extract the core semantic features in the news content. The core semantics includes the subject, keywords and time node features of the news event: ； in, represents the core semantic features, softmax represents the normalization function, The query vector representing the core semantic features, The key vector representing the core semantic features, The transpose of the key vector representing the core semantic features, A value vector representing the core semantic features, The weight matrix representing the time feature, represents the normalization factor; S24. Construct the second layer of hierarchical Transformer, perform background semantic decoding on the core semantic features, and extract secondary semantic features: ； in, Represents secondary semantic features, The query vector representing the secondary semantic features, The key vector representing the secondary semantic features, The transpose of the key vector representing the secondary semantic features, A value vector representing a secondary semantic feature, and represents the weight parameter, Positional encoding to represent background semantics; S25. Based on the core semantic features and secondary semantic features, the temporal features and hierarchical semantic features are generated through a multiple dynamic weighting mechanism of the temporal dimension and the semantic level: ； in, Represents temporal features and hierarchical semantic features, Represents the weight matrix of core semantic features, Represents the weight matrix of secondary semantic features, Represents time characteristics.

3. The method for intelligent news broadcasting based on deep learning according to claim 1, characterized in that: The event feature-driven cross-modal contrastive learning in S3 specifically includes: obtaining a set of event feature information, including time clues, semantic subjects and background features, as a guiding signal for feature alignment; based on cross-modal contrastive learning, aligning time features and hierarchical semantic features with multimodal news data.

4. According to claim 1, a method for intelligent news broadcasting based on deep learning is characterized in that: The S4 specifically includes: S41. Divide the cross-modal fusion features by time nodes to form feature sets of several time segments ,in Indicates time node The corresponding feature subset; S42, use the multi-head self-attention mechanism to perform semantic modeling on the feature subset within each time node, extract local context associations, and obtain preliminary time node features; S43, comparing the features of different preliminary time nodes in pairs, defining the semantic relevance between the time nodes, and calculating the semantic relevance by weighted combination of feature similarity and time difference: ； in, Indicates time node and time nodes The semantic association between them, A represents the dimension of the feature vector, Indicates time node The eigenvalue in the kth dimension, Indicates time node In the k-th dimension, exp represents the exponential function. Indicates time node and time nodes The time difference between represents the time decay parameter; S44. Construct a dynamic event tree T=(N,E) based on each time node and semantic relevance, where Represents a set of time nodes and edge sets , represents the semantic relevance threshold; S45. Divide the time nodes in the dynamic event tree into layers: ; in, Indicates time node The level to which it belongs, Indicates time node The semantic contribution of represents the semantic importance threshold, represents the correlation stratification threshold, represents the time interval threshold, and n represents the total number of time nodes; S46. Iteratively update the dynamic event tree, calculate and correct the edge weights and hierarchical divisions, and finally generate a multi-scale dynamic semantic structure.

5. According to claim 1, a method for intelligent news broadcasting based on deep learning is characterized in that: The S5 specifically includes: S51, initializing an improved three-stage generation network, including an overview generation module, a detail generation module and a prediction generation module, and loading core layer features, sub-core layer features and background layer features in a multi-scale dynamic semantic structure; S52, in the summary generation module, extracting elements from the core layer features to generate summary broadcast content of emergency and breaking news, the summary generation module including a sequence encoder and an attention decoder, outputting a summary broadcast sequence of the first stage; S53, in the detail generation module, the summary broadcast sequence output by the summary generation module and the sub-core layer features and background layer features in the multi-scale dynamic semantic structure are received, and are integrated into the multimodal decoder, and the event subject, background and context details are supplemented through the fine-grained attention mechanism, and the detail broadcast sequence of the second stage is output; S54, in the prediction generation module, combining the temporal feature information of the summary broadcast sequence, the detailed broadcast sequence and the multi-scale dynamic semantic structure, capturing the future trend of the event evolution trend through the recursive time network, and outputting the predictive broadcast content based on the event evolution; S55, integrating summary broadcast content, detailed broadcast content and predictive broadcast content based on event evolution into a unified multimodal output module, and cross-validating them to screen out redundant and contradictory content; S56: Output the cross-verified broadcast content to the user terminal through a real-time push channel.

6. The method for intelligent news broadcasting based on deep learning according to claim 1, characterized in that: The S6 specifically includes: S61. Collect user interaction feedback information, including user click times, stay time, comment sentiment analysis results and sharing behavior, to form a user feedback matrix ,in Indicates user a's response to a time node Overall feedback score: ; in, Indicates user a's response to a time node MaxClicks represents the maximum number of clicks by all users. Indicates user a's response to a time node The duration of stay, MaxTime indicates the maximum duration of stay of all users, Indicates user a's response to a time node The sentiment score of the review, , and Represents the weight parameter of the feedback indicator; S62. Dynamically update the weight of each time node of the dynamic event tree based on the user feedback matrix: ; in, Indicates time node The update weight of represents the equilibrium parameter, Indicates time node The original weight of Indicates time node The average feedback value of S63. Adjust the semantic generation rules of the dynamic event tree according to the updated weight, and recalculate the semantic contribution of each node: ; in, Indicates time node The semantic contribution of Indicates time node The update weight of n represents the total number of time nodes, tanh represents the hyperbolic tangent function, and ρ represents the amplification factor of the feedback change. Indicates time node The baseline feedback value of S64, based on the updated weights and semantic contributions, dynamically optimize the hierarchical structure of the dynamic event tree, perform pruning operations, and re-divide the core layer and the background layer; S65, using the updated time node weights and semantic generation rules, regenerate summary broadcast content, detailed broadcast content, and predictive broadcast content based on event evolution, so that the semantics of the generated content is consistent with user feedback; S66: Push the optimized broadcast content to the user terminal, and record a new round of user feedback information in real time, forming a closed-loop optimization mechanism of user feedback and dynamic event tree adjustment.

7. A news intelligent broadcasting system based on deep learning, executing the news intelligent broadcasting method based on deep learning according to any one of claims 1 to 6, characterized in that: Includes the following modules: The data collection and coding module is used to collect text data, image data and video data of news events, build a multimodal news data set, and perform semantic coding on the multimodal news data to generate preliminary semantic features; The temporal and semantic feature extraction module is used to extract the temporal dimension of the preliminary semantic features based on the dynamic temporal convolutional network, and to semantically classify the news content through the hierarchical Transformer architecture to generate temporal features and hierarchical semantic features; The cross-modal alignment module is used to align temporal features and hierarchical semantic features with multimodal news data through cross-modal contrastive learning driven by event characteristics to generate cross-modal fusion features; The dynamic event tree generation module is used to use the multi-head self-attention mechanism to perform multi-scale semantic modeling on cross-modal fusion features, build a dynamic event tree based on time nodes and semantic relevance, and generate a multi-scale dynamic semantic structure; A three-stage content generation module, which is used to input the multi-scale dynamic semantic structure into the improved three-stage generation network, and sequentially generate summary broadcast content, detailed broadcast content, and predictive broadcast content based on event evolution; The event feedback optimization module is used to dynamically adjust the node weights and semantic generation rules of the dynamic event tree based on user interaction feedback information using the event feedback optimization mechanism to generate optimized broadcast content; The multimodal output module is used to synchronously output the optimized broadcast content in the form of text, voice and video, and transmit it to the user terminal through a real-time push channel.

Images