An intelligent content generation system based on open source data analysis

Abstract

The present application discloses an intelligent content generation system based on open source data analysis, which relates to the field of monitoring and analysis technology. The system includes a data preprocessing module for performing multimodal content collection on heterogeneous open source data platforms to obtain a structured semantic feature matrix; a graph construction and extraction module for constructing a domain knowledge graph based on the structured semantic feature matrix to obtain a domain evolution feature vector; an intelligent content generation module for inputting the domain evolution feature vector into a multimodal generative adversarial network for semantic-visual cross-modal mapping to obtain style-enhanced intelligent generated content; an optimization module for simulating and predicting the dissemination effect of the style-enhanced intelligent generated content to generate a content dissemination influence matrix, and iterating and optimizing the intelligent generated content according to the content dissemination influence matrix to generate the final optimized content and push it to the target terminal.

Description

Technical Field

[0001] The present application relates to the field of monitoring and analysis technology, and in particular to an intelligent content generation system based on open source data analysis. Background Art

[0002] Currently, with the rapid development of open source data platforms, using public network data for intelligent content generation has become an important direction of digital content production.

[0003] In related technologies, traditional intelligent content generation systems find it difficult to cope with the dynamic evolution of open source data. The delayed data update leads to insufficient timeliness of the generated content. In addition, cross-modal generation technology generally has semantic distortion problems. Key information is seriously lost during the conversion process from text to images and videos, which in turn leads to reduced efficiency in content generation. There is room for improvement. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, this application provides an intelligent content generation system based on open source data analysis.

[0005] This application provides an intelligent content generation system based on open source data analysis, including:

[0006] A data preprocessing module is used to collect multimodal content from heterogeneous open source data platforms to obtain original multi-source heterogeneous data sets, and perform cross-platform semantic consistency analysis on the original multi-source heterogeneous data sets to obtain a structured semantic feature matrix;

[0007] The graph construction and extraction module is used to construct a domain knowledge graph based on a structured semantic feature matrix, generate a dynamically evolving domain knowledge graph, and extract trend-related features from the dynamically evolving domain knowledge graph through a temporal convolutional network to obtain a domain evolution feature vector.

[0008] An intelligent content generation module is used to input the domain evolution feature vector into a multimodal generative adversarial network for semantic-visual cross-modal mapping to generate primary intelligent generated content, and then perform style transfer reinforcement training on the primary intelligent generated content to obtain style-enhanced intelligent generated content;

[0009] The optimization module is used to simulate and predict the dissemination effect of style-enhanced intelligently generated content, generate a content dissemination influence matrix, optimize and iterate the intelligently generated content based on the content dissemination influence matrix, generate the final optimized content and push it to the target terminal.

[0010] Preferably, the data preprocessing module specifically includes:

[0011] Construct a distributed crawler cluster that is compatible with multiple protocols, and then perform parallel data crawling on heterogeneous open source data platforms based on the distributed crawler cluster to obtain original multi-source heterogeneous data sets, which include original text data, original image data, and original video data;

[0012] The adaptive threshold algorithm is used to preprocess the original multi-source heterogeneous data corresponding to the original multi-source heterogeneous data set, and then a benchmark data set is constructed based on the preprocessed multi-source heterogeneous data;

[0013] Perform cross-modal feature alignment on the benchmark dataset to extract a semantic association matrix consisting of text entity relationships, image semantic labels, and video keyframe features;

[0014] Cross-platform feature fusion is performed on the semantic association matrix to eliminate the semantic gap between different data sources and generate a structured semantic feature matrix.

[0015] Preferably, the process of using the adaptive threshold algorithm specifically includes:

[0016] Constructing a quality assessment index system, wherein the quality assessment index system includes a text readability score, an image clarity index, and a video stability parameter;

[0017] Perform filter threshold comparison analysis on text readability scores, image clarity indicators, and video stability parameters, and identify normal data distribution ranges and outlier boundaries based on the results of the threshold comparison analysis;

[0018] A dynamic threshold adjustment unit is designed to automatically update the filtering threshold according to the real-time quality fluctuation of the original multi-source heterogeneous data in the original multi-source heterogeneous dataset.

[0019] Preferably, the map construction and extraction module specifically includes:

[0020] Perform entity relationship analysis on the structured semantic feature matrix to extract the hierarchical association structure and time evolution path data between entities;

[0021] Based on the hierarchical association structure and time evolution path data between entities, the initial domain knowledge graph is constructed. The initial domain knowledge graph is dynamically updated through a sliding time window mechanism. Based on the results of the dynamic update, the relationship evolution data is captured in real time to construct a dynamically evolving domain knowledge graph.

[0022] Trend features are extracted from the dynamically evolving domain knowledge graph, and then evolution pattern features are extracted. The domain evolution feature vector is generated based on the fusion of evolution pattern features.

[0023] Preferably, the intelligent content generation module specifically includes:

[0024] Constructing a multimodal generative adversarial network comprising a generator and a discriminator, wherein the generator comprises a text-image cross-modal conversion unit and a semantic preservation constraint unit, and the discriminator comprises a multi-scale feature identification unit;

[0025] Decomposing the domain evolution feature vector into a semantic content vector and a style feature vector, and then inputting the semantic content vector and the style feature vector into the multimodal generative adversarial network for feature mapping;

[0026] Through the text-image cross-modal conversion unit and the semantic preservation constraint unit, primary intelligent generated content is generated, and a style transfer enhancement model is constructed. Based on the style transfer enhancement model, style features are extracted and injected into the generation process. The generation effect is optimized through style-content decoupling, and style-enhanced intelligent generated content is output.

[0027] Preferably, the optimization module specifically includes:

[0028] Construct a virtual communication environment simulation network, input style-enhanced intelligently generated content into the virtual communication environment simulation network to predict the communication trajectory, and then generate a content communication influence matrix based on the prediction results;

[0029] Identify key optimization parameters based on the content dissemination influence matrix, construct a multi-objective optimization function based on the key optimization parameters, and analyze it to determine the optimal improvement direction;

[0030] Update the optimization strategy based on the optimal improvement direction, achieve iterative optimization through dynamic adjustment of the optimization strategy, output the final optimized content and push it to the target terminal.

[0031] Preferably, the process of predicting the propagation trajectory specifically includes:

[0032] Build a user behavior simulation agent and generate virtual user groups based on real user portraits, and then design a multi-channel communication model. According to the multi-channel communication model, simulate the differences in the corresponding communication paths of virtual user groups on heterogeneous open source data platforms, and introduce influencing factors to simulate the impact on the differences in communication paths; then optimize the selection of communication strategies through deep reinforcement learning and generate a content communication influence matrix.

[0033] In summary, this application has the following beneficial technical effects:

[0034] The present application provides an intelligent content generation system based on open source data analysis, which performs cross-platform semantic consistency analysis on original multi-source heterogeneous data sets to obtain a structured semantic feature matrix, constructs a domain knowledge graph and extracts trend-related features based on the structured semantic feature matrix, and obtains a domain evolution feature vector, thereby effectively responding to the dynamic evolution characteristics of open source data and improving the timeliness of generated content. It also generates primary intelligent generated content based on the domain evolution feature vector, performs style transfer reinforcement training and communication effect simulation prediction on the primary intelligent generated content, generates a content communication influence matrix, optimizes and iterates the intelligent generated content based on the content communication influence matrix, generates the final optimized content and pushes it to the target terminal, thereby effectively reducing the occurrence of semantic distortion problems in cross-modal generation technology, thereby effectively improving the efficiency of content generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0036] Figure 1 It is a schematic diagram of the system module of the present invention. DETAILED DESCRIPTION

[0037] Example

[0038] Reference Figure 1 , an intelligent content generation system based on open source data analysis, including:

[0039] A data preprocessing module is used to collect multimodal content from heterogeneous open source data platforms to obtain original multi-source heterogeneous data sets, and perform cross-platform semantic consistency analysis on the original multi-source heterogeneous data sets to obtain a structured semantic feature matrix;

[0040] The graph construction and extraction module is used to construct a domain knowledge graph based on a structured semantic feature matrix, generate a dynamically evolving domain knowledge graph, and extract trend-related features from the dynamically evolving domain knowledge graph through a temporal convolutional network to obtain a domain evolution feature vector.

[0041] An intelligent content generation module is used to input the domain evolution feature vector into a multimodal generative adversarial network for semantic-visual cross-modal mapping to generate primary intelligent generated content, and then perform style transfer reinforcement training on the primary intelligent generated content to obtain style-enhanced intelligent generated content;

[0042] The optimization module is used to simulate and predict the dissemination effect of style-enhanced intelligently generated content, generate a content dissemination influence matrix, optimize and iterate the intelligently generated content based on the content dissemination influence matrix, generate the final optimized content and push it to the target terminal.

[0043] The data preprocessing module specifically includes:

[0044] Construct a distributed crawler cluster that is compatible with multiple protocols, and then perform parallel data crawling on heterogeneous open source data platforms based on the distributed crawler cluster to obtain original multi-source heterogeneous data sets, which include original text data, original image data, and original video data;

[0045] The adaptive threshold algorithm is used to preprocess the original multi-source heterogeneous data corresponding to the original multi-source heterogeneous data set, and then a benchmark data set is constructed based on the preprocessed multi-source heterogeneous data;

[0046] Perform cross-modal feature alignment on the benchmark dataset to extract a semantic association matrix consisting of text entity relationships, image semantic labels, and video keyframe features;

[0047] Cross-platform feature fusion is performed on the semantic association matrix to eliminate the semantic gap between different data sources and generate a structured semantic feature matrix.

[0048] In the embodiment of the present invention, the implementation process of the data preprocessing module is as follows: first, a distributed crawler cluster consisting of a multi-protocol parsing engine and a load balancer is built, and differentiated collection strategies are set according to the data interface characteristics of different open source platforms. The heterogeneous platform is divided into several data collection units through a task scheduling algorithm for parallel crawling, and the original data set covering text, images and videos is obtained in real time; an adaptive threshold algorithm is used to perform multi-dimensional purification processing on the collected data, and for text data, low-frequency noise words are filtered by dynamically adjusting the word frequency statistical threshold for dynamic adjustment of the word frequency statistical threshold. For image data, the contrast adjustment threshold is dynamically set according to the brightness histogram distribution to enhance the effective features. For video data, the key frame extraction threshold is set according to the difference between adjacent frames to achieve standardized cleaning of the full amount of data; in the cross-modal feature alignment stage The cleaned data is input into a multi-channel feature extraction network, the text data is used to extract subject-verb-object triples through the entity recognition model, the image data uses a deep convolutional network to identify object semantic labels, and the video data uses a temporal action detection algorithm to extract key frame feature sequences. Subsequently, a mapping relationship between the three modal features is established through a vector projection algorithm in a shared semantic space to form a three-dimensional feature matrix reflecting cross-modal semantic correlation; finally, a cross-platform feature fusion mechanism is adopted to identify common semantic items by calculating the cosine similarity between feature vectors from different data sources, fuse the common items and eliminate platform-specific interference factors, and combine the feature contribution weight allocation algorithm to generate a structured feature matrix with unified dimensions and semantic coherence, where the weight allocation is dynamically adjusted based on the product of the feature entropy value and the platform data quality assessment coefficient.

[0049] The process of using the adaptive threshold algorithm specifically includes:

[0050] Constructing a quality assessment index system, wherein the quality assessment index system includes a text readability score, an image clarity index, and a video stability parameter;

[0051] Perform filter threshold comparison analysis on text readability scores, image clarity indicators, and video stability parameters, and identify normal data distribution ranges and outlier boundaries based on the results of the threshold comparison analysis;

[0052] A dynamic threshold adjustment unit is designed to automatically update the filtering threshold according to the real-time quality fluctuation of the original multi-source heterogeneous data in the original multi-source heterogeneous dataset.

[0053] In an embodiment of the present invention, the implementation process of the adaptive threshold algorithm is as follows: first, a multimodal quality assessment system is established, in which the text readability score is obtained by calculating the weighted sum of vocabulary density and the proportion of long and difficult sentences, the image clarity index generates a quantized value based on image edge sharpness analysis using the Laplace gradient operator, and the video stability parameter is comprehensively evaluated through the statistical difference of motion vectors between adjacent frames and the frequency of lens jitter; the normal distribution characteristics of each quality indicator are statistically analyzed using a historical data training set, and the range of three standard deviations from the mean is determined as the normal data distribution range. A dynamic boundary determination mechanism based on an outlier factor is established for abnormal data points that exceed the threshold of this range; for heterogeneous data streams collected in real time, the fluctuation trends of text readability, image clarity, and video stability are continuously monitored within a sliding time window. When a persistent shift in the indicator distribution is detected, a threshold adaptive adjustment mechanism is triggered: first, the mean and variance of each quality indicator in the current window are calculated, and the percentage of the mean deviation from the initial training set baseline value is used as the threshold correction. At the same time, an additional amplification factor is applied to indicators whose fluctuation variance exceeds the warning value, forming a filtering threshold update rule for the dynamic evolution of data quality.

[0054] The graph construction and extraction module specifically includes:

[0055] Perform entity relationship analysis on the structured semantic feature matrix to extract the hierarchical association structure and time evolution path data between entities;

[0056] Based on the hierarchical association structure and time evolution path data between entities, the initial domain knowledge graph is constructed. The initial domain knowledge graph is dynamically updated through a sliding time window mechanism. Based on the results of the dynamic update, the relationship evolution data is captured in real time to construct a dynamically evolving domain knowledge graph.

[0057] Trend features are extracted from the dynamically evolving domain knowledge graph, and then evolution pattern features are extracted. The domain evolution feature vector is generated based on the fusion of evolution pattern features.

[0058] In an embodiment of the present invention, the implementation process of the graph construction and extraction module is as follows: multi-hop relationship reasoning is performed on the structured semantic feature matrix through a deep graph convolutional network to identify the implicit affiliation, causal association and temporal dependency between entities, and generate a three-dimensional association network containing parent-child hierarchy, attribute inheritance relationship and event evolution path; when constructing the initial domain knowledge graph based on the network topology structure, the timestamp annotation technology is used to encode the creation time and duration period information of the entity relationship as the attribute value of the graph edge, forming a spatiotemporal knowledge framework with time dimension traceability; when introducing a sliding time window mechanism to dynamically maintain the graph, the window length is set as the domain evolution cycle parameter, and the time slice is incremented by 0. The quantitative learning algorithm compares the difference between the newly added entity relationships and the historical graph. The difference calculation adopts the weighted value of the graph structure similarity measurement and the cosine distance of the relationship vector space. When the difference exceeds the evolution threshold, the graph update process is triggered, the new relationship is embedded in the existing graph and the invalid associations that exceed the window period are eliminated; in the trend feature extraction stage, the three indicators of the fluctuation trajectory of the relationship strength, the convergence trend of the association density and the evolution rate of the topological structure are statistically analyzed through the time series graph pattern mining algorithm. The self-attention mechanism is used to dynamically assign weights to the multi-dimensional trend features. Finally, a composite feature vector representing the evolution law of the domain is generated through feature splicing and standardization. This vector contains both the long-term evolution principal component and the short-term fluctuation characteristics.

[0059] The intelligent content generation module specifically includes:

[0060] Constructing a multimodal generative adversarial network comprising a generator and a discriminator, wherein the generator comprises a text-image cross-modal conversion unit and a semantic preservation constraint unit, and the discriminator comprises a multi-scale feature identification unit;

[0061] Decomposing the domain evolution feature vector into a semantic content vector and a style feature vector, and then inputting the semantic content vector and the style feature vector into the multimodal generative adversarial network for feature mapping;

[0062] Through the text-image cross-modal conversion unit and the semantic preservation constraint unit, primary intelligent generated content is generated, and a style transfer enhancement model is constructed. Based on the style transfer enhancement model, style features are extracted and injected into the generation process. The generation effect is optimized through style-content decoupling, and style-enhanced intelligent generated content is output.

[0063] In an embodiment of the present invention, the implementation process of the intelligent content generation module is as follows: first, a generative adversarial network framework consisting of a cross-modal conversion engine and a multi-scale discriminator is constructed, and the generator uses a dual-channel encoding structure to decouple the input domain evolution feature vector into a semantic content vector and a style feature vector, wherein the semantic vector extracts core concept elements through a self-attention mechanism, and the style vector uses a Gram matrix to capture statistical distribution characteristics; in the content generation stage, the text-image conversion unit generates a basic visual feature map layer by layer based on the semantic vector through a deconvolution network, and at the same time, the semantic preservation constraint unit dynamically adjusts the generation weight to prevent semantic deviation by calculating the potential space similarity between the generated content and the input vector; at the same time, The style transfer enhancement model constructed in this step extracts the texture spectrum features of the target style through a convolutional network, and uses an adaptive instance normalization algorithm to fuse the style features into the basic visual feature map to achieve gradient optimization with style-content decoupling; a multi-level feature identification mechanism is deployed on the discriminator side, which extracts local image details and global structural features through convolution kernel groups at 64x64, 128x128 and 256x256 pixel scales, and compares the feature distribution differences between the generated images and real samples; finally, the generative adversarial network is optimized through an alternating training mechanism, so that the generated content can achieve multi-style adaptive expression while maintaining the original semantic integrity, and output intelligent generated content with both information accuracy and visual expressiveness.

[0064] The optimization module specifically includes:

[0065] Construct a virtual communication environment simulation network, input style-enhanced intelligently generated content into the virtual communication environment simulation network to predict the communication trajectory, and then generate a content communication influence matrix based on the prediction results;

[0066] Identify key optimization parameters based on the content dissemination influence matrix, construct a multi-objective optimization function based on the key optimization parameters, and analyze it to determine the optimal improvement direction;

[0067] Update the optimization strategy based on the optimal improvement direction, achieve iterative optimization through dynamic adjustment of the optimization strategy, output the final optimized content and push it to the target terminal.

[0068] In an embodiment of the present invention, the implementation process of the optimization module is as follows: a virtual communication environment simulation system is constructed through a graph neural network, and the style-enhanced intelligently generated content is decomposed into multi-dimensional feature vectors such as visual appeal index, semantic association strength and emotional tendency value, and input into a user behavior prediction model trained based on historical communication data to simulate the content's propagation path and hierarchical diffusion effect between social network nodes; a complex network analysis algorithm is used to calculate the time attenuation curve and forwarding probability distribution matrix of the content reaching the user group, and the three-dimensional influence matrix reflecting the potential depth and breadth of communication is generated by combining the similarity of user interest tags and the weight of social relationships; based on this matrix, the content feature dimensions are extracted The key sensitive parameters that affect the forwarding rate are identified. Through a multi-objective optimization framework, the three core indicators of emotional tendency balance, visual impact threshold, and information dissemination rate are used as optimization targets. A genetic algorithm search strategy is used to find the Pareto optimal solution set in the solution space. After eliminating conflicting solutions between objective functions, the improvement direction for maximizing the overall communication efficiency is determined. Finally, a dynamic feedback optimization mechanism is established. By monitoring the deviation between the actual communication effect and the simulated prediction value in real time, the color saturation, keyword density, and emotional expression intensity of the generated content are gradient adjusted. After multiple rounds of iterative optimization, the ultimate optimized content that matches the preferences of the target user group is output and accurately pushed to the user's terminal device. The user behavior prediction model uses a temporal attention mechanism to capture the characteristics of the propagation time window. The multi-objective optimization process eliminates the conflicting constraints between objectives through a non-dominated sorting algorithm. The gradient adjustment coefficient is dynamically updated according to the fitness value of the previous optimization effect.

[0069] The process of predicting the propagation trajectory includes:

[0070] Build a user behavior simulation agent and generate virtual user groups based on real user portraits, and then design a multi-channel communication model. According to the multi-channel communication model, simulate the differences in the corresponding communication paths of virtual user groups on heterogeneous open source data platforms, and introduce influencing factors to simulate the impact on the differences in communication paths; then optimize the selection of communication strategies through deep reinforcement learning and generate a content communication influence matrix.

[0071] In an embodiment of the present invention, the implementation process of the propagation trajectory prediction is as follows: an adaptive generative network is constructed based on the user social relationship graph and behavior log data, and a virtual user group with age, occupation, and interest gradient distribution characteristics is simulated and generated through a conditional generative adversarial model. Each agent is given a social influence weight and an information sensitivity coefficient; in the cross-platform propagation modeling stage, a differentiated propagation rule matrix is ​​designed for different open source platforms, wherein the social media platform sets a two-hop propagation probability attenuation model based on friend relationships, the forum platform adopts a plate heat conduction algorithm driven by topic similarity, and the video platform constructs a composite forwarding decision function of visual attractiveness index and stay time; a dynamic impact factor adjustment mechanism is introduced, and when a dynamic influence factor appears in the simulation, a dynamic influence factor adjustment mechanism is introduced. In the event of a sudden hot topic, a global attention network is used to calculate the match between event relevance and user interests, generating a contagion coefficient that multiplies the basic probability of dissemination. Incorporating a deep reinforcement learning framework, the selection of dissemination strategies is abstracted into a Markov decision process, defining the selection of dissemination channels and the push time window as a set of executable actions for the agent. The reward signal is a weighted composite of dissemination path coverage, user engagement index, and lasting influence. After multiple rounds of iterative optimization of the strategy, the dissemination paths of each virtual user driven by the optimized strategy are calculated. A multidimensional influence matrix is ​​generated by multiplying the node diffusion depth and the activation ratio of users at adjacent levels. This matrix simultaneously reflects the spatial distribution characteristics and temporal evolution of content across different platforms. The social influence weight is dynamically updated based on the user's historical dissemination contribution, and the contagion coefficient is dynamically attenuated and adjusted by real-time monitoring of the information overload index in the virtual environment.

[0072] The above contents are merely examples and explanations of the concept of the present invention. Those skilled in the art may make various modifications or additions to the described specific embodiments or replace them in similar ways. As long as they do not deviate from the concept of the invention, they should all fall within the scope of protection of the present invention.

[0073] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0074] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to specific embodiments. Obviously, many modifications and variations are possible based on the contents of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention.

Claims

1. An intelligent content generation system based on open source data analysis, characterized in that: include: A data preprocessing module is used to collect multimodal content from heterogeneous open source data platforms to obtain original multi-source heterogeneous data sets, and perform cross-platform semantic consistency analysis on the original multi-source heterogeneous data sets to obtain a structured semantic feature matrix; The data preprocessing module specifically includes: Construct a distributed crawler cluster that is compatible with multiple protocols, and then perform parallel data crawling on heterogeneous open source data platforms based on the distributed crawler cluster to obtain original multi-source heterogeneous data sets, which include original text data, original image data, and original video data; The adaptive threshold algorithm is used to preprocess the original multi-source heterogeneous data corresponding to the original multi-source heterogeneous data set, and then a benchmark data set is constructed based on the preprocessed multi-source heterogeneous data; Perform cross-modal feature alignment on the benchmark dataset to extract a semantic association matrix consisting of text entity relationships, image semantic labels, and video keyframe features; Perform cross-platform feature fusion on the semantic correlation matrix to eliminate the semantic gap between different data sources and generate a structured semantic feature matrix; The process of using the adaptive threshold algorithm specifically includes: Constructing a quality assessment index system, wherein the quality assessment index system includes a text readability score, an image clarity index, and a video stability parameter; Perform filter threshold comparison analysis on text readability scores, image clarity indicators, and video stability parameters, and identify normal data distribution ranges and outlier boundaries based on the results of the threshold comparison analysis; A dynamic threshold adjustment unit is designed to automatically update the filtering threshold according to the real-time quality fluctuation of the original multi-source heterogeneous data in the original multi-source heterogeneous dataset; The graph construction and extraction module is used to construct a domain knowledge graph based on a structured semantic feature matrix, generate a dynamically evolving domain knowledge graph, and extract trend-related features from the dynamically evolving domain knowledge graph through a temporal convolutional network to obtain a domain evolution feature vector. An intelligent content generation module is used to input the domain evolution feature vector into a multimodal generative adversarial network for semantic-visual cross-modal mapping to generate primary intelligent generated content, and then perform style transfer reinforcement training on the primary intelligent generated content to obtain style-enhanced intelligent generated content; The optimization module is used to simulate and predict the dissemination effect of style-enhanced intelligently generated content, generate a content dissemination influence matrix, optimize and iterate the intelligently generated content based on the content dissemination influence matrix, generate the final optimized content and push it to the target terminal.

2. The intelligent content generation system based on open source data analysis according to claim 1, characterized in that: The graph construction and extraction module specifically includes: Perform entity relationship analysis on the structured semantic feature matrix to extract the hierarchical association structure and time evolution path data between entities; Based on the hierarchical association structure and time evolution path data between entities, the initial domain knowledge graph is constructed. The initial domain knowledge graph is dynamically updated through a sliding time window mechanism. Based on the results of the dynamic update, the relationship evolution data is captured in real time to construct a dynamically evolving domain knowledge graph. Trend features are extracted from the dynamically evolving domain knowledge graph, and then evolution pattern features are extracted. The domain evolution feature vector is generated based on the fusion of evolution pattern features.

3. The intelligent content generation system based on open source data analysis according to claim 1, characterized in that: The intelligent content generation module specifically includes: Constructing a multimodal generative adversarial network comprising a generator and a discriminator, wherein the generator comprises a text-image cross-modal conversion unit and a semantic preservation constraint unit, and the discriminator comprises a multi-scale feature identification unit; Decomposing the domain evolution feature vector into a semantic content vector and a style feature vector, and then inputting the semantic content vector and the style feature vector into the multimodal generative adversarial network for feature mapping; Through the text-image cross-modal conversion unit and the semantic preservation constraint unit, primary intelligent generated content is generated, and a style transfer enhancement model is constructed. Based on the style transfer enhancement model, style features are extracted and injected into the generation process. The generation effect is optimized through style-content decoupling, and style-enhanced intelligent generated content is output.

4. The intelligent content generation system based on open source data analysis according to claim 1, characterized in that: The optimization module specifically includes: Construct a virtual communication environment simulation network, input style-enhanced intelligently generated content into the virtual communication environment simulation network to predict the communication trajectory, and then generate a content communication influence matrix based on the prediction results; Identify key optimization parameters based on the content dissemination influence matrix, construct a multi-objective optimization function based on the key optimization parameters, and analyze it to determine the optimal improvement direction; Update the optimization strategy based on the optimal improvement direction, achieve iterative optimization through dynamic adjustment of the optimization strategy, output the final optimized content and push it to the target terminal.

5. The intelligent content generation system based on open source data analysis according to claim 4, characterized in that: The process of predicting the propagation trajectory includes: Build a user behavior simulation agent and generate virtual user groups based on real user portraits, and then design a multi-channel communication model. According to the multi-channel communication model, simulate the differences in the corresponding communication paths of virtual user groups on heterogeneous open source data platforms, and introduce influencing factors to simulate the impact on the differences in communication paths; then optimize the selection of communication strategies through deep reinforcement learning and generate a content communication influence matrix.

Images