Internet content intelligent creation method and device based on generative adversarial network, equipment and medium

Abstract

The invention relates to an Internet content intelligent creation method and device based on a generative adversarial network, equipment and a medium. The method comprises the following steps: acquiring network real-time hotspot data streams, grouping the data streams by adopting an unsupervised clustering algorithm, and identifying a high-dynamic trend content set; performing multi-dimensional feature analysis on the high-dynamic trend content set by adopting a pre-trained deep semantic analysis model to generate a trend associated content theme; performing feature extraction on the multi-form content under the trend correlation content theme through a feature extractor to obtain a multi-form content feature set; carrying out nonlinear standardization mapping on the multi-form content feature set by adopting a feature standardization algorithm to generate a unified feature expression set; and based on the unified feature expression set, generating a framework by using a pre-constructed generative adversarial network content, and outputting creation content. By adopting the method, the efficiency and the automation level of content generation can be improved, and the consistency of the multi-modal content in semantic and emotional levels is ensured.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of generative artificial intelligence, and particularly relates to an Internet content intelligent creation method and device based on a generative adversarial network, equipment and a medium. BACKGROUND

[0002] With the rapid development of artificial intelligence (AI) technology in the field of content creation, especially the breakthroughs in generative models (such as generative adversarial networks, GANs) and multi-modal learning technology, there has emerged technology that uses automated means to assist or replace human labor in Internet content creation. This kind of technology aims to improve the efficiency and scale of content production by analyzing network data streams, identifying trend themes, and automatically generating related text, images, videos, and other multi-form content. In existing technology, traditional Internet content creation methods or early automated content generation methods generally follow the following process: human labor or simple crawler tools collect hot information from the network; editors or rule / template-based systems create content based on this information, or use single-modal models (such as text generation models or image generation models) to generate text or pictures respectively; and different forms of content are manually combined and published. The core feature of this approach is that content generation is often carried out independently in a single mode (text, image, video are processed separately), the relevance and timeliness of content themes to dynamic network hotspots are highly dependent on human judgment, and the ability to integrate multi-dimensional features such as semantics and emotion for cross-modal collaborative creation is insufficient. The current content creation methods or traditional methods have the following significant problems: low efficiency and strong dependence on human labor, the entire process from hotspot identification, theme analysis to content generation requires human intervention, making it difficult to meet the demand for rapid creation of real-time hotspot content under the massive flow of Internet information; fragmented content forms and poor collaboration, text, images, videos, and other different forms of content are usually generated independently by different tools or processes, lacking internal semantic and emotional consistency, and making it difficult to form organically integrated multi-modal content; hotspot response lag and insufficient dynamic capture, relying on human labor or simple rules to identify hotspots, making it difficult to capture a set of hotspot content with high dynamic change trends in real time, resulting in generated content that may deviate from the latest network focus; limited theme depth and emotional fit, analysis of the creation theme often stops at the level of shallow keywords, lacking the integration of deep semantic understanding and emotional analysis (such as emotional intensity, distribution characteristics), resulting in generated content that deviates from user expectations or the network atmosphere in terms of theme depth and emotional expression; and uneven multi-form content coverage, generated content may be overly focused on a single form (such as text), ignoring the natural multi-form content distribution characteristics (such as the absence of images or videos) under the hotspot theme, resulting in single-form content presentation and incomplete coverage. SUMMARY

[0003] Based on this, it is necessary to provide a method, device, equipment and medium for creating internet content based on a generative adversarial network.

[0004] In a first aspect, the present application provides a method for creating internet content based on a generative adversarial network, comprising:

[0005] Obtaining a network real-time hot data stream, and grouping the data stream using an unsupervised clustering algorithm to identify a high-dynamic trend content set;

[0006] Using a pre-trained deep semantic parsing model to analyze the high-dynamic trend content set in multiple dimensions to generate a trend-related content theme;

[0007] Extracting features of multi-form content under the trend-related content theme through a feature extractor to obtain a multi-form content feature set, the multi-form content including text, image, and video content;

[0008] Using a feature standardization algorithm to perform nonlinear standardization mapping on the multi-form content feature set to generate a unified feature expression set;

[0009] Based on the unified feature expression set, using a pre-constructed generative adversarial network content generation framework to output the created content.

[0010] In one embodiment, obtaining a network real-time hot data stream, and grouping the data stream using an unsupervised clustering algorithm to identify a high-dynamic trend content set, comprises:

[0011] Real-time collection of hot data streams from multiple internet data sources;

[0012] Grouping the hot data streams using an unsupervised clustering algorithm to generate multiple content classification groups;

[0013] Calculating the time series change characteristic value of each content classification group within a preset time window using a trend analysis algorithm;

[0014] In response to the time series change characteristic value of a certain content classification group exceeding a preset threshold, determining that the content of the corresponding content classification group is high-dynamic trend content;

[0015] Integrating all high-dynamic trend content to output a high-dynamic trend content set.

[0016] In one embodiment, using a pre-trained deep semantic parsing model to analyze the high-dynamic trend content set in multiple dimensions to generate a trend-related content theme, comprises:

[0017] Extracting semantic structure feature vectors of the high-dynamic trend content set through the pre-trained deep semantic parsing model to generate a theme clustering group;

[0018] The incremental sentiment dictionary matching algorithm is used to analyze the sentiment distribution characteristics of the topic clustering group and calculate the sentiment intensity value, and the sentiment-enhanced topic group is output;

[0019] Based on the sentiment-enhanced topic group, a semantic-sentiment multi-modal feature space is constructed, and semantic and sentiment features are fused through a cross-modal attention mechanism to generate trend-related content topics.

[0020] In one embodiment, the trend-related content topics are generated by fusing semantic and sentiment features through a cross-modal attention mechanism, including:

[0021] Through the cross-modal attention mechanism, the following formula is used to generate a semantic and sentiment fusion feature matrix:

[0022]

[0023] wherein, is a semantic feature matrix extracted from the topic clustering group, is a sentiment feature matrix extracted from the sentiment-enhanced topic group, is a feature value matrix corresponding to S, is a trainable semantic-sentiment correlation matrix, d is the dimension of the feature vector, K is the number of topics, φ is a nonlinear activation function, and Z output is a fusion feature matrix;

[0024] The fusion feature matrix is mapped to the semantic-sentiment multi-modal feature space by the following formula:

[0025]

[0026] wherein, is a feature set in the semantic-sentiment multi-modal feature space, MLP is a multi-layer perceptron, and z k is the fusion feature matrix of the kth topic, and K is the number of topics;

[0027] Based on the feature set in the semantic-sentiment multi-modal feature space The following formula is used to generate trend-related content topics T:

[0028]

[0029] wherein, θ is the model parameter of the generator, t k is the vector representation of the kth topic, z k is the fusion feature matrix of the kth topic, and P(·) is the conditional probability distribution function of the generator.

[0030] In one of the embodiments, based on the unified feature expression set, an output creative content is generated by using a pre-constructed generative adversarial network content generation framework, including:

[0031] The unified feature expression set is input to a generator of the generative adversarial network;

[0032] The text features, image features and video features are collaboratively encoded by a cross-modal feature fusion layer in the generator to generate a multi-modal latent vector;

[0033] The multi-modal latent vector is input to a discriminator of the generative adversarial network;

[0034] Based on an adversarial loss value output by the discriminator, weight parameters of the generator module are iteratively optimized;

[0035] Based on the multi-modal latent vector, the creative content is generated by the optimized generator.

[0036] In one of the embodiments, the method further includes:

[0037] Based on the trend-related content theme, content form distribution features of the text, image and video under the theme are parsed;

[0038] The coverage of the creative content in the content form distribution features is calculated, and the coverage is obtained by calculating the proportion of the number of the generated content in the text, image and video forms;

[0039] When the coverage is lower than a preset form threshold, the content in the missing form is generated by the generative adversarial network content generation framework;

[0040] The content in the missing form is fused with the creative content to form multi-form generated content.

[0041] In one of the embodiments, the method further includes:

[0042] User feedback data of the creative content is obtained, and the user feedback data includes emotional evaluation and form integrity evaluation;

[0043] Emotional matching degree indicators and form integrity degree indicators are parsed from the user feedback data;

[0044] If the emotional matching degree indicators are lower than a first threshold, feature weight parameters in a discriminator of the generative adversarial network content generation framework are adjusted based on a deviation value of the emotional matching degree indicators;

[0045] If the form integrity degree indicators are lower than a second threshold, a weight correction amount of a cross-modal feature fusion layer in a generator of the generative adversarial network content generation framework is calculated;

[0046] The feature weight parameters and / or the weight correction amount are applied to update configuration parameters of the discriminator and / or the generator.

[0047] In a second aspect, the present application also provides an internet content intelligent creation device based on a generative adversarial network, comprising:

[0048] a dynamic clustering module, configured to acquire a network real-time hotspot data stream, and perform grouping on the data stream by using an unsupervised clustering algorithm to identify a high-dynamic trend content set;

[0049] a semantic sentiment fusion module, configured to perform multi-dimensional feature analysis on the high-dynamic trend content set by using a pre-trained deep semantic analysis model to generate a trend-related content theme;

[0050] a multi-modal extraction module, configured to perform feature extraction on multi-form content under the trend-related content theme by using a feature extractor to obtain a multi-form content feature set, the multi-form content including text, image, and video form content;

[0051] a feature standardization module, configured to perform nonlinear standardization mapping on the multi-form content feature set by using a feature standardization algorithm to generate a unified feature expression set;

[0052] a content creation generation module, configured to output a created content based on the unified feature expression set by using a pre-constructed generative adversarial network content generation framework.

[0053] In a third aspect, the present application also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-described internet content intelligent creation method based on a generative adversarial network when executing the computer program.

[0054] In a fourth aspect, the present application also provides a computer readable storage medium, which stores a computer program, and the computer program implements the steps of the above-described internet content intelligent creation method based on a generative adversarial network when executed by a processor.

[0055] The above-described internet content intelligent creation method based on a generative adversarial network, device, computer device, and storage medium solve the problem of lag in dynamic hotspot response of traditional methods by acquiring a network real-time hotspot data stream and identifying a high-dynamic trend content set by using an unsupervised clustering algorithm; overcome the defects of insufficient theme depth and sentiment fit by performing multi-dimensional feature analysis on the trend-related content theme by using a pre-trained deep semantic analysis model; eliminate the disadvantages of poor multi-modal content synergy and form fragmentation by performing feature extraction and nonlinear standardization mapping on multi-form content such as text, image, and video to generate a unified feature expression set; and significantly improve the efficiency and automation degree of hotspot content generation by using a generative adversarial network content generation framework to output created content, and realize the internal consistency of multi-modal content in the semantic and sentiment levels. BRIEF DESCRIPTION OF DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the embodiments or related art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort on the basis of these drawings.

[0057] Figure 1 The flow chart of the internet content intelligent creation method based on the generative adversarial network of the present application;

[0058] Figure 2 The structure diagram of the internet content intelligent creation device based on the generative adversarial network of the present application. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solutions and advantages of the present application more clear, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0060] In one embodiment, as shown in Figure 1 A method for internet content intelligent creation based on a generative adversarial network is provided. In this embodiment, the method is applied to a terminal for illustration. It should be understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and can be realized through the interaction of the terminal and the server. In the implementation environment, a distributed computing framework can be used, in which the terminal (such as a general intelligent device) is responsible for data acquisition and preliminary processing, the server (such as a cloud computing platform) carries a core algorithm module, and data transmission and result feedback are realized through a standardized network protocol, realizing end-to-end collaboration. The application scenario is oriented to the real-time demand of an internet content platform, such as a social media event outbreak or a news hotspot tracking. The terminal acquires data streams from multiple sources and uploads them, the server performs abstract content clustering, semantic sentiment fusion and generation tasks, and finally outputs the created content to the terminal for distribution, ensuring efficient response and consistent output. In this embodiment, the method includes the following steps:

[0061] S01, acquiring network real-time hotspot data streams, and grouping the data streams by using an unsupervised clustering algorithm to identify a high-dynamic trend content set.

[0062] Among them, network real-time hot spot data stream (a set of data collected in real time from multiple Internet data sources (such as social media platforms, news websites or forums), containing popular topics, events or user focus on the current network); unsupervised clustering algorithm (a data grouping technique without pre-labeled, used to automatically divide hot spot data stream into multiple content classification groups according to data similarity, such as groups based on theme or semantic similarity); high dynamic trend content set (by analyzing the time series change characteristic value of the content classification group, the subset with significant dynamic change can be identified, where the time series change characteristic value exceeds the preset threshold, such as the change rate threshold, indicating that the content group is experiencing rapid evolution or high attention.); In implementation, hot spot data stream can be collected in real time from multiple Internet data sources (such as API interface or web crawler); apply unsupervised clustering algorithm (such as K-means or DBSCAN) to automatically group data stream (based on data feature similarity, such as keyword frequency or semantic vector), generate multiple content classification groups; calculate the time series change characteristic value of each content classification group in the preset time window (such as within 1 hour), if the time series change characteristic value of a content classification group exceeds the preset threshold (such as change rate > 10%), it is determined that the content of this group is high dynamic trend content, integrate all identified high dynamic trend content, output as high dynamic trend content set.

[0063] S02, using a pre-trained deep semantic analysis model to analyze the high dynamic trend content set in multiple dimensions to generate trend-related content themes.

[0064] Wherein, the pre-trained deep semantic parsing model (a neural network model pre-trained on a large corpus (such as based on the Transformer architecture), capable of parsing the deep semantic structure of the text content, and outputting a semantic feature vector); multi-dimensional feature analysis (joint analysis of semantic dimensions such as topic clustering, keyword association, and sentiment dimensions such as sentiment orientation, intensity distribution); trend-related content theme (high-order theme representation closely related to real-time hotspots generated by fusing semantic and sentiment features, with semantic consistency and emotional expression); when implemented, the pre-trained deep semantic parsing model (such as the neural network model of BERT or GPT variants) can be used to extract the semantic structure feature vector of the high-dynamic trend content set, and the initial topic clustering group is generated based on vector similarity clustering; the incremental sentiment dictionary matching algorithm can be used to dynamically analyze the sentiment distribution features of the topic clustering group: through sentiment dictionary library matching, the sentiment intensity value (such as positive / negative intensity score) is calculated in combination with the context, and the sentiment-enhanced theme group is output; based on the sentiment-enhanced theme group, a semantic-sentiment multi-modal feature space is constructed, the semantic matrix and the sentiment matrix are fused through the cross-modal attention mechanism, and the fusion feature matrix is generated; the fusion feature matrix is mapped to the semantic-sentiment multi-modal feature space through a multi-layer perceptron (MLP), and the trend-related content theme is generated according to the probability distribution function.

[0065] S03, feature extraction is performed on the multi-form content under the trend-related content theme by the feature extractor, and a multi-form content feature set is obtained, the multi-form content including text, image, and video form content.

[0066] Wherein, the feature extractor (a pre-constructed algorithm module for extracting deep feature representation from heterogeneous data); multi-form content (including text, image, and video form content, covering structured and unstructured data under the trend-related content theme); multi-form content feature set (standardized feature set formed after extraction, representing the vectorized expression of the content in the feature space) when implemented, the feature extractor can be used to extract semantic feature vectors for text content using natural language processing models (such as BERT or word embedding models), extract visual features (such as edge, texture, and object recognition features) for image content using convolutional neural networks (CNN), and extract motion sequence and scene features for video content using spatio-temporal models (such as 3D CNN). The original data of text, image, and video are integrated through a parallel processing mechanism, and the output is the multi-form content feature set.

[0067] S04, a feature standardization algorithm is used to perform nonlinear standardization mapping on the multi-form content feature set to generate a unified feature expression set.

[0068] Among them, the characteristic standardization algorithm (a mathematical method of scaling and mapping feature values through a nonlinear function such as Sigmoid or Tanh); the unified feature expression set (a same-scale feature matrix formed after standardization, and each modal feature is embedded in the same vector space); when implemented, it can perform feature alignment operations on multiple form content feature sets (such as semantic vectors of text, CNN features of images, and 3D-CNN features of videos), and unify all feature vector dimensions to a preset value d through dimension reduction or padding; a nonlinear normalization mapping function (such as an improved Min-Max scaling: Where μ is the mean, σ is the standard deviation, and α and β are learnable parameters) is applied to each feature dimension to eliminate the numerical scale difference of different modal features; feature weighting based on attention mechanism can be introduced to dynamically adjust the contribution weight of each modal feature, and a unified feature expression set is output to eliminate form differences through mathematical transformation and to strengthen key features through attention mechanism.

[0069] S05, based on the unified feature expression set, using a pre-constructed generative adversarial network content generation framework, output creative content.

[0070] Among them, the pre-constructed generative adversarial network content generation framework (a modular system based on a generative adversarial network (GAN), including generator and discriminator components, the generator is responsible for content synthesis, and the discriminator is used for quality evaluation); creative content (the final output of the multi-form product, including text, image and video combination); when implemented, the unified feature expression set can be input into the generator module of the generative adversarial network, and the text features, image features and video features are cooperatively encoded through the cross-modal feature fusion layer (multi-head attention mechanism) in the generator, to generate a multi-modal latent vector; the multi-modal latent vector is input into the discriminator module, and the weight parameters of the generator are iteratively optimized based on the adversarial loss value (such as Wasserstein distance calculation) output by the discriminator, and the loss is minimized using the gradient descent algorithm; the creative content is generated based on the multi-modal latent vector through the optimized generator, such as using a decoder network to map the multi-modal latent vector to a text sequence, an image pixel matrix or a video frame sequence.

[0071] The above internet content intelligent creation method based on the generative adversarial network realizes real-time capture of dynamic network hotspots by obtaining network real-time hotspot data streams and grouping the data streams using an unsupervised clustering algorithm to identify a high-dynamic trend content set, overcoming the problem of content deviating from the latest focus due to response lag; a pre-trained deep semantic analysis model is used to analyze the content set in multiple dimensions to generate trend-related content themes, and deep semantic understanding and emotional feature fusion are used to enhance the depth of the theme and the emotional fit; the feature extractor extracts the text, image and video content under the theme to obtain a multi-form content feature set, and a feature standardization algorithm is used for nonlinear standardization mapping to generate a unified feature expression set, eliminating the defect of fragmentation of multi-modal content forms; based on the unified feature expression set, a pre-constructed generative adversarial network content generation framework is used to output the created content, and a cross-modal feature fusion layer is used to realize collaborative coding of text, image and video features, improving the content generation efficiency while ensuring the consistency of multi-modal content in the semantic and emotional aspects, forming a closed loop from dynamic clustering to content generation, and improving the automation level and quality of hotspot content creation.

[0072] In one of the embodiments, the network real-time hotspot data stream is obtained, and an unsupervised clustering algorithm is used to group the data stream to identify a high-dynamic trend content set, including:

[0073] S11, real-time collection of hotspot data streams from multiple internet data sources;

[0074] S12, application of an unsupervised clustering algorithm to group the hotspot data streams to generate a plurality of content classification groups;

[0075] S13, calculation of the time sequence change characteristic value of each content classification group within a preset time window using a trend analysis algorithm;

[0076] S14, in response to the time sequence change characteristic value of a certain content classification group exceeding a preset threshold, determining that the content of the corresponding content classification group is high-dynamic trend content;

[0077] S15, integration of all high-dynamic trend content, output as a high-dynamic trend content set.

[0078] Specifically, a set of data collected in real time from multiple internet data sources (such as social media platforms, news websites, or forums) through API interfaces or web crawlers contains current network hot topics or events; an unsupervised clustering algorithm uses an automatic grouping technique based on data similarity (such as K-means or DBSCAN) to group the collected hot data stream, generating multiple content classification groups, where the content classification group refers to a group automatically divided according to theme or semantic similarity; a trend analysis algorithm is used to calculate the time series change characteristic value of each content classification group in a preset time window (such as within 1 hour), which can be content activity or growth rate, used to quantify dynamic changes; in response to the time series change characteristic value of a certain content classification group exceeding a preset threshold (such as a change rate > 10%, which can be calibrated according to the change rate of the content classification group characteristic value of the historical hot data stream), it is determined that the content of this group is high-dynamic trend content, indicating that it is experiencing rapid evolution or high attention; all identified high-dynamic trend contents are integrated and output as a high-dynamic trend content set as input for subsequent processing. Through a distributed computing framework (terminals are responsible for preliminary collection, servers perform clustering), terminals upload data streams obtained from multiple sources in real time, servers apply clustering algorithms to automatically group based on keyword frequency or semantic vectors, calculate time series characteristic values (use sliding window algorithm to analyze time series data), trigger dynamic content labeling when threshold is exceeded, and integrate output set, eliminating the response lag problem caused by relying on manual in traditional methods, improving the automation efficiency and accuracy of hot spot recognition.

[0079] In one embodiment, a pre-trained deep semantic parsing model is used to analyze the high-dynamic trend content set in multiple dimensions to generate trend-related content themes, including:

[0080] S21, extract the semantic structure feature vector of the high-dynamic trend content set through the pre-trained deep semantic parsing model, and generate a theme clustering group;

[0081] S22, analyze the sentiment distribution characteristics of the theme clustering group and calculate the sentiment intensity value using an incremental sentiment dictionary matching algorithm, and output a sentiment-enhanced theme group;

[0082] S23, based on the sentiment-enhanced theme group, construct a semantic-sentiment multimodal feature space, and fuse semantic and sentiment features through cross-modal attention mechanism to generate trend-related content themes.

[0083] Exemplarily, in a distributed computing framework (e.g., terminals are responsible for data collection, and servers execute core algorithms), a semantic structure feature vector of a high dynamic trend content set is extracted by a pre-trained deep semantic parsing model, a topic clustering group (a group automatically divided based on semantic similarity) is generated based on vector similarity clustering; an incremental sentiment dictionary matching algorithm (the update rule is: when the frequency of a new word is >1000 times / day, manually annotate the sentiment polarity and update the dictionary) is used to dynamically analyze the sentiment distribution characteristics of the topic clustering group (the sentiment intensity value is calculated by matching the sentiment keywords in the sentiment dictionary library and combining the context, such as positive or negative scores), and the sentiment-enhanced topic group (the topic representation after integrating sentiment intensity analysis) is output; a semantic-sentiment multi-modal feature space (a vector space that integrates text semantics and sentiment features) is constructed based on the sentiment-enhanced topic group, and the semantic and sentiment features are fused through a cross-modal attention mechanism to generate a fusion feature matrix. The fusion feature matrix is mapped to the semantic-sentiment multi-modal feature space to generate a trend-related content topic.

[0084] In one of the embodiments, the trend-related content topic is generated by fusing the semantic and sentiment features through the cross-modal attention mechanism, including:

[0085] S31, the semantic and sentiment fusion feature matrix is generated by using the following formula through the cross-modal attention mechanism:

[0086]

[0087] wherein, is a semantic feature matrix extracted from the topic clustering group, is a sentiment feature matrix extracted from the sentiment-enhanced topic group, is a feature value matrix corresponding to S, is a trainable semantic-sentiment correlation matrix, d is the dimension of the feature vector, K is the number of topics, φ is a nonlinear activation function, and Z output is a fusion feature matrix;

[0088] S32, the fusion feature matrix is mapped to the semantic-sentiment multi-modal feature space by using the following formula:

[0089]

[0090] wherein, is a feature set in the semantic-sentiment multi-modal feature space, MLP is a multi-layer perception, and z k is a fusion feature matrix of the kth topic, and K is the number of topics;

[0091] S33, based on the feature set in the semantic-sentiment multi-modal feature space the following formula is used to generate a trend-related content topic T:

[0092]

[0093] where θ is the model parameter of the generator, t k is the vector representation of the kth topic, z k is the fusion feature matrix of the kth topic, P(·) is the conditional probability distribution function of the generator.

[0094] Specifically, the semantic feature matrix S (a feature matrix extracted from the topic clustering group, based on semantic similarity division, with dimensions Kxd, representing semantic structure features); the sentiment feature matrix E (a feature matrix extracted from the sentiment-enhanced topic group, with dimensions Kxd, representing sentiment distribution features); the eigenvalue matrix V (the eigenvalue matrix corresponding to S, with dimensions Kxd); the trainable semantic-sentiment correlation matrix W s (dxd dimension of the optimizable parameter matrix, used to model the correlation weight between semantics and sentiment); d is the dimension of the feature vector, for scaling to prevent gradient disappearance, K is the number of topics; the nonlinear activation function φ (ReLU function) is used to introduce nonlinear transformation; when the fusion feature matrix Z (the fusion output of semantic and sentiment features) is implemented, it can be implemented in a distributed computing framework (such as the terminal responsible for data collection, and the server executing the core algorithm), through the cross-modal attention mechanism, using the formula to generate the fusion feature matrix Z, where the softmax function is used to normalize the attention weight, for scaling to avoid gradient disappearance, the semantic feature S and the sentiment feature E are passed through the correlation matrix W s dynamic fusion; the feature set F (the feature set in the semantic-sentiment multi-modal feature space) fuses the feature matrix z k of the kth topic through the multi-layer perception MLP (multi-layer neural network structure), using the formula maps the fusion feature matrix Z to the semantic-sentiment multi-modal feature space, where the MLP performs nonlinear transformation on the vector of each topic to generate the feature set based on the feature set using the formula to generate the trend-related content topic T, where t k |z k ; θ is substituted into the conditional probability distribution function P(·) of the generator to obtain the multi-class probability distribution P(t k |z k ; θ) based on the Softmax function; the argmax function maximizes the likelihood of topic generation based on the probability distribution, which can ensure the consistency and depth of the topic in semantics and sentiment.

[0095] In one of the embodiments, based on the unified feature expression set, an output creative content is generated by using a pre-constructed generative adversarial network content generation framework, including:

[0096] S41, inputting the unified feature expression set to a generator of the generative adversarial network;

[0097] S42, performing collaborative coding on the text feature, the image feature and the video feature by a cross-modal feature fusion layer in the generator to generate a multi-modal latent vector;

[0098] S43, inputting the multi-modal latent vector to a discriminator of the generative adversarial network;

[0099] S44, iteratively optimizing weight parameters of the generator module based on an adversarial loss value output by the discriminator;

[0100] S45, generating the creative content by the generator based on the multi-modal latent vector.

[0101] Exemplarily, the unified feature expression set can be input to the generator of the generative adversarial network, the text feature, the image feature and the video feature are collaboratively coded by a cross-modal feature fusion layer (multi-head attention mechanism) in the generator to generate a multi-modal latent vector (representing a low-dimensional representation of fused semantics and structure), the multi-modal latent vector is input to the discriminator module, weight parameters of the generator module are iteratively optimized based on an adversarial loss value (such as a loss function calculated by a Wasserstein distance) output by the discriminator, and a gradient descent algorithm is used to minimize the loss; the creative content is generated by the generator based on the multi-modal latent vector, for example, the latent vector is mapped to a text sequence (such as LSTM decoding), an image pixel matrix (such as a CNN generator) or a video frame sequence (such as a 3D-GAN) using a decoder network, so that the multi-modal content has internal consistency in the semantic and emotional levels, and the defect of form fragmentation in traditional methods is eliminated.

[0102] In one of the embodiments, the method further includes:

[0103] S51, analyzing content form distribution characteristics of the text, the image and the video under the theme based on the trend-related content theme;

[0104] S52, calculating a coverage rate of the creative content in the content form distribution characteristics, the coverage rate being obtained by calculating a proportion of the number of the text, the image and the video forms of the generated content;

[0105] S53, when the coverage rate is lower than a preset form threshold, generating the content of the missing form by using the generative adversarial network content generation framework;

[0106] S54, fusing the content of the missing form with the creative content to form multi-form generated content.

[0107] Specifically, content form distribution feature (the proportion distribution of text, image and video form content parsed based on the theme in historical or real-time data, for example, by analyzing multi-source data under the theme, calculating the quantitative features of text proportion 40%, image proportion 35%, and video proportion 25%); coverage rate (the proportion of the number of each form in the generated content in the content form distribution feature (such as if the text form accounts for 30% in the generated content, the text coverage rate = 30% / 40% = 75%), which is obtained by counting the number of text, image and video instances in the generated content and dividing by the benchmark value of the corresponding form in the distribution feature); preset form threshold (a predefined lower limit of coverage rate (such as 70%), used to trigger the generation of missing content); missing form content (the form to be supplemented (such as image missing) identified when the coverage rate is lower than the threshold); in implementation, the content theme can be associated based on the trend, the content form distribution feature of text, image and video under the theme is parsed, the proportion of the number of each form is calculated through statistical analysis, and a distribution vector is generated; the coverage rate of the generated content in the content form distribution feature is calculated, the number of instances of each form in the generated content is counted by using a counter, and the proportion value is calculated in combination with the distribution vector; in response to the coverage rate being lower than the preset form threshold (such as image coverage rate < 70%), the missing form content is generated by using the generative adversarial network content generation framework, the cross-modal feature fusion layer (multi-head attention mechanism) of the generator is called to synthesize new content for the missing form (such as image), for example, a CNN decoder is used to generate a pixel matrix; the missing form content is fused with the generated content to form multi-form generated content through content integration (such as feature splicing or weighted average algorithm), so as to ensure that the output covers all forms and the proportions are balanced.

[0108] In one of the embodiments, the method further comprises:

[0109] S61, obtaining user feedback data of the generated content, the user feedback data including emotional evaluation and form integrity evaluation;

[0110] S62, parsing an emotional matching degree index and a form integrity index from the user feedback data;

[0111] S63, if the emotional matching degree index is lower than a first threshold, adjusting a feature weight parameter in a discriminator of the content generation framework based on a deviation value of the emotional matching degree index;

[0112] S64, if the form integrity index is lower than a second threshold, calculating a weight correction amount of a cross-modal feature fusion layer in a generator of the generative adversarial network content generation framework;

[0113] S65, applying the feature weight parameter and / or the weight correction amount to update the configuration parameters of the discriminator and / or the generator.

[0114] Exemplarily, the user feedback data is a set of data collected from an Internet platform such as a social media or a content distribution channel in real time after the creation content is published, containing sentiment evaluation such as subjective score or comment of the user on the emotional expression of the content and form integrity evaluation such as feedback of the user on the form coverage of the text, image and video; the sentiment matching degree index is a sentiment consistency score quantified after the user feedback data is analyzed by a natural language processing algorithm (such as a sentiment classifier based on a pre-trained model) (such as calculating the cosine similarity of the sentiment tendency of the generated content and the sentiment of the user feedback); the form integrity index is a form coverage integrity score calculated after the feedback data is analyzed by a statistical analysis module (such as calculating the proportion of missing forms pointed out by the user); the first threshold and the second threshold are respectively a preset lower limit of the sentiment matching degree (such as 0.7) and a preset lower limit of the form integrity (such as 0.75), used to trigger the parameter adjustment mechanism. In implementation, the user feedback data of the creation content is obtained, the feedback data is analyzed to generate the sentiment matching degree index and the form integrity index; if the sentiment matching degree index is lower than the first threshold, the feature weight parameter in the discriminator of the content generation framework of the generative adversarial network is adjusted based on the sentiment matching degree index bias value (the sentiment matching error is calculated by a loss function such as cross-entropy loss, for example, the weight matrix is updated by gradient backpropagation, and the formula is wherein η is a learning rate, L 情感 is a sentiment loss function based on cross-entropy loss); if the form integrity index is lower than the second threshold, the weight correction amount of the cross-modal feature fusion layer in the generator is calculated (the weight increment is calculated using an optimization algorithm such as Adam, and the formula is ΔW g = α · mean(|missing form feature vector|), wherein α is a scaling coefficient); the feature weight parameter and / or the weight correction amount are applied, and the configuration parameters of the discriminator and the generator are dynamically updated by a parameter server (such as recompiling the model after updating the weight), forming a closed-loop optimization mechanism to continuously improve the content quality.

[0115] The above-mentioned internet content intelligent creation method based on a generative adversarial network acquires real-time hotspot data streams and uses an unsupervised clustering algorithm to group the data streams, combines real-time collection of multi-source data and application of a trend analysis algorithm to calculate time series change characteristic values, identifies a high dynamic trend content set when the threshold is exceeded, solves the problems of hotspot response lag and insufficient dynamic capture in traditional methods, and ensures real-time synchronization of content generation with the latest network hotspots. A pre-trained deep semantic analysis model is used to analyze the multi-dimensional features of the high dynamic trend content set, including extracting semantic structure feature vectors, applying an incremental sentiment dictionary matching algorithm to analyze sentiment distribution features and calculate sentiment intensity values, and generating trend-related content themes by fusing semantic and sentiment features through a cross-modal attention mechanism, overcoming the limitations of theme depth and sentiment compatibility, and realizing the internal fusion of deep semantic understanding and sentiment analysis. The feature extractor extracts features from the multi-form content (text, image, video) under the trend-related content theme, and uses a feature standardization algorithm for nonlinear standardization mapping to generate a unified feature expression set, eliminating the disadvantages of fragmented content forms and poor collaboration, and ensuring consistent expression of multi-modal features in a unified vector space. Based on the unified feature expression set, the pre-constructed generative adversarial network content generation framework outputs the created content, the cross-modal feature fusion layer in the generator collaboratively encodes text, image and video features to generate multi-modal latent vectors, and the parameters are iteratively optimized based on the adversarial loss value of the discriminator to improve the efficiency and automation of content generation, solving the problems of strong artificial dependence and low efficiency. By analyzing the content form distribution characteristics under the theme, calculating the coverage rate, and generating missing form content when the threshold is below, and adjusting the sentiment matching degree index and form completeness index based on user feedback data (such as correcting feature weight parameters and weight correction amount through bias value), the multi-form content coverage is balanced, avoiding the defects of single form content generation or incomplete coverage. Through unsupervised clustering, deep semantic sentiment fusion, multi-modal feature standardization and collaborative optimization of the generative adversarial network, the consistency of internet content generation is realized in the aspects of semantics, sentiment and multi-form, significantly improving the real-time performance, accuracy and user experience of hotspot content creation.

[0116] It should be understood that although each step in the flowchart involved in each embodiment as described above is shown in sequence according to the direction of the arrow, these steps are not necessarily executed in the order indicated by the arrow. Unless otherwise explicitly stated herein, there is no strict order limitation for the execution of these steps, and these steps can be executed in other orders. Moreover, at least some of the steps in the flowchart involved in each embodiment as described above can include multiple steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but can be alternately or alternately executed with at least some of the other steps or steps or stages in other steps.

[0117] Based on the same inventive concept, the embodiments of the present application also provide a generative adversarial network-based internet content intelligent creation device for implementing the above-mentioned generative adversarial network-based internet content intelligent creation method. The problem-solving implementation scheme provided by the device is similar to the implementation scheme described in the above method, so the specific limitations in one or more generative adversarial network-based internet content intelligent creation device embodiments provided below can be referred to the limitations of the generative adversarial network-based internet content intelligent creation method described above, which will not be repeated here.

[0118] In one exemplary embodiment, as shown in Figure 2 a generative adversarial network-based internet content intelligent creation device is provided, comprising:

[0119] The dynamic clustering module 101 is configured to obtain a network real-time hot spot data stream, and group the data stream by using an unsupervised clustering algorithm to identify a high-dynamic trend content set;

[0120] The semantic sentiment fusion module 102 is configured to perform multi-dimensional feature analysis on the high-dynamic trend content set by using a pre-trained deep semantic analysis model to generate a trend-related content theme;

[0121] The multi-modal extraction module 103 is configured to perform feature extraction on multi-form content under the trend-related content theme by using a feature extractor to obtain a multi-form content feature set, the multi-form content including text, image, and video form content;

[0122] The feature standardization module 104 is configured to perform nonlinear standardization mapping on the multi-form content feature set by using a feature standardization algorithm to generate a unified feature expression set;

[0123] The content creation generation module 105 is configured to output a created content based on the unified feature expression set by using a pre-built generative adversarial network content generation framework.

[0124] In one of the embodiments, the dynamic clustering module 101 is further configured to:

[0125] collecting real-time hotspot data streams from multiple internet data sources;

[0126] grouping the hotspot data streams by applying unsupervised clustering algorithm to generate multiple content classification groups;

[0127] calculating time-series change feature values of each content classification group within a preset time window by using a trend analysis algorithm;

[0128] determining that the content of a corresponding content classification group is high-dynamic trend content in response to the time-series change feature value of the content classification group exceeding a preset threshold;

[0129] integrating all high-dynamic trend contents and outputting a high-dynamic trend content set.

[0130] In one of the embodiments, the semantic sentiment fusion module 102 is further configured to:

[0131] extracting semantic structure feature vectors of the high-dynamic trend content set by using a pre-trained deep semantic parsing model to generate a topic clustering group;

[0132] analyzing sentiment distribution features of the topic clustering group and calculating sentiment intensity values by using an incremental sentiment dictionary matching algorithm to output a sentiment-enhanced topic group;

[0133] constructing a semantic-sentiment multimodal feature space based on the sentiment-enhanced topic group and fusing semantic and sentiment features by using a cross-modal attention mechanism to generate a trend-related content topic.

[0134] In one of the embodiments, the semantic sentiment fusion module 102 is further configured to:

[0135] using the following formula to generate a semantic-sentiment fusion feature matrix by using a cross-modal attention mechanism:

[0136]

[0137] wherein, S is a semantic feature matrix extracted from the topic clustering group, E is a sentiment feature matrix extracted from the sentiment-enhanced topic group, is a feature value matrix corresponding to S, is a trainable semantic-sentiment correlation matrix, d is a feature vector dimension, K is the number of topics, φ is a nonlinear activation function, and Z output is a fusion feature matrix;

[0138] mapping the fusion feature matrix to a semantic-sentiment multimodal feature space by using the following formula:

[0139]

[0140] wherein, is a feature set in the semantic-affective multi-modal feature space, MLP is a multi-layer perception, z k is a fusion feature matrix of the kth topic, K is the number of topics;

[0141] based on the feature set in the semantic-affective multi-modal feature space The trend-related content theme T is generated using the following formula:

[0142]

[0143] wherein, θ is a model parameter of the generator, t k is a vector representation of the kth topic, z k is a fusion feature matrix of the kth topic, P(·) is a conditional probability distribution function of the generator.

[0144] In one of the embodiments, the content creation generation module 105 is further configured to:

[0145] input the unified feature expression set into the generator of the generative adversarial network;

[0146] collaboratively encode the text features, image features and video features through the cross-modal feature fusion layer in the generator to generate a multi-modal latent vector;

[0147] input the multi-modal latent vector into the discriminator of the generative adversarial network;

[0148] based on the adversarial loss value output by the discriminator, iteratively optimize the weight parameters of the generator module;

[0149] based on the multi-modal latent vector, generate the created content through the optimized generator.

[0150] In one of the embodiments, the content creation generation module 105 is further configured to:

[0151] based on the trend-related content theme, analyze the content form distribution features of the text, image and video under the theme;

[0152] calculate the coverage rate of the created content in the content form distribution features, the coverage rate is obtained by calculating the proportion of the number of the created content in the form of text, image and video;

[0153] when the coverage rate is lower than a preset form threshold, generate the content in the missing form through the generative adversarial network content generation framework;

[0154] fuse the content in the missing form with the created content to form the multi-form generated content.

[0155] In one of the embodiments, the content creation generation module 105 is further configured to:

[0156] obtain user feedback data of the created content, the user feedback data comprising sentiment evaluation and form completeness evaluation;

[0157] parse a sentiment matching degree index and a form completeness index from the user feedback data;

[0158] if the sentiment matching degree index is lower than a first threshold, adjust a feature weight parameter in a discriminator of the generative adversarial network content generation framework based on a deviation value of the sentiment matching degree index;

[0159] if the form completeness index is lower than a second threshold, calculate a weight correction amount of a cross-modal feature fusion layer in a generator of the generative adversarial network content generation framework;

[0160] apply the feature weight parameter and / or the weight correction amount to update configuration parameters of the discriminator and / or the generator.

[0161] In one of the embodiments, a computer device is provided, comprising a memory and a processor, the memory storing a computer program, and the processor implementing the steps of the generative adversarial network based internet content intelligent creation method as described above when executing the computer program.

[0162] In one of the embodiments, a computer readable storage medium is provided, storing a computer program, and the computer program is executed by a processor to implement the steps of the above method embodiments.

[0163] For the device embodiments, since they basically correspond to the method embodiments, the related parts can be referred to the part of the method embodiments. The device embodiments described above are merely illustrative, and the components described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e. they can be located in one place or distributed on multiple network units. Some or all of the modules can be selected to achieve the purpose of the present disclosure according to the actual needs. Those skilled in the art can understand and implement it without creative labor.

[0164] The above described embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the patent scope of the application. It should be noted that, for those skilled in the art, without departing from the concept of the present application, some modifications and improvements can be made, which are all within the protection scope of the present application.

Claims

1. An Internet content intelligent creation method based on a generative adversarial network, characterized in that, The method comprises: acquiring a network real-time hot spot data stream, and grouping the data stream by using an unsupervised clustering algorithm to identify a high-dynamic trend content set; performing multi-dimensional feature analysis on the high-dynamic trend content set by using a pre-trained deep semantic analysis model to generate a trend-related content theme; performing feature extraction on multi-form content under the trend-related content theme by using a feature extractor to obtain a multi-form content feature set, wherein the multi-form content includes text, image, and video form content; performing nonlinear standardization mapping on the multi-form content feature set by using a feature standardization algorithm to generate a unified feature expression set; based on the unified feature expression set, using a pre-constructed generative adversarial network content generation framework to output creative content.

2. The method of claim 1, wherein, The acquisition of the network real-time hot spot data stream and the grouping of the data stream by using an unsupervised clustering algorithm to identify a high-dynamic trend content set comprises: real-time collection of the hot spot data stream from multiple Internet data sources; grouping of the hot spot data stream by using an unsupervised clustering algorithm to generate multiple content classification groups; calculation of the time sequence change characteristic value of each content classification group within a preset time window by using a trend analysis algorithm; determination that the content of a corresponding content classification group is high-dynamic trend content in response to the time sequence change characteristic value of the content classification group exceeding a preset threshold value; integration of all high-dynamic trend content to output the high-dynamic trend content set.

3. The method of claim 1, wherein, The multi-dimensional feature analysis on the high-dynamic trend content set by using a pre-trained deep semantic analysis model to generate a trend-related content theme comprises: extraction of a semantic structure feature vector of the high-dynamic trend content set by using the pre-trained deep semantic analysis model to generate a theme clustering group; analysis of the sentiment distribution characteristics of the theme clustering group and calculation of a sentiment intensity value by using an incremental sentiment dictionary matching algorithm to output a sentiment-enhanced theme group; construction of a semantic-sentiment multi-modal feature space based on the sentiment-enhanced theme group, and fusion of semantic and sentiment features by using a cross-modal attention mechanism to generate the trend-related content theme.

4. The method of claim 3, wherein, The fusion of semantic and sentiment features by using a cross-modal attention mechanism to generate the trend-related content theme comprises: the generation of a semantic-sentiment fusion feature matrix by using the following formula by using a cross-modal attention mechanism: wherein, is a semantic feature matrix extracted from the topic cluster group, is a sentiment feature matrix extracted from the sentiment-enhanced topic group, is a feature value matrix corresponding to S, is a trainable semantic-sentiment correlation matrix, d is the dimension of the feature vector, K is the number of topics, φ is a nonlinear activation function, and Z output is a fusion feature matrix. the mapping of the fusion feature matrix to a semantic-sentiment multi-modal feature space by using the following formula: wherein, is a feature set in the semantic-affective multimodal feature space, MLP is a multi-layer perceptron, z k is a fusion feature matrix of the kth topic, and K is the number of topics; based on a feature set in the semantic-sentiment multimodal feature space The trend-related content topic T is generated using the following formula: where θ is the model parameter of the generator, t k is the vector representation of the kth topic, z k is the fusion feature matrix of the kth topic, P(·) is the conditional probability distribution function of the generator.

5. The method of claim 1, wherein, The output of creative content based on the unified feature expression set by using a pre-constructed generative adversarial network content generation framework comprises: input of the unified feature expression set into a generator of the generative adversarial network; collaborative coding of text features, image features, and video features by using a cross-modal feature fusion layer in the generator to generate a multi-modal latent vector; input of the multi-modal latent vector into a discriminator of the generative adversarial network; iterative optimization of the weight parameters of the generator module based on the adversarial loss value output by the discriminator; generation of the creative content by using the optimized generator based on the multi-modal latent vector.

6. The method of claim 1, wherein, The method further comprises: Correlate a content theme based on the trend, analyze content form distribution characteristics of text, image, and video under the theme; Calculate the coverage of the created content in the content form distribution characteristics, which is obtained by calculating the proportion of the number of text, image, and video forms in the generated content; When the coverage is lower than a preset form threshold, generate content in the missing form through the generative adversarial network content generation framework; Fuse the content in the missing form with the created content to form multi-form generated content.

7. The method of claim 6, wherein, The method further comprises: Obtain user feedback data of the created content, which includes emotional evaluation and form integrity evaluation; Analyze emotional matching degree indicators and form integrity indicators from the user feedback data; If the emotional matching degree indicators are lower than a first threshold, adjust feature weight parameters in a discriminator of the generative adversarial network content generation framework based on deviation values of the emotional matching degree indicators; If the form integrity indicators are lower than a second threshold, calculate weight correction amounts of cross-modal feature fusion layers in a generator of the generative adversarial network content generation framework; Apply the feature weight parameters and / or the weight correction amounts to update configuration parameters of the discriminator and / or the generator.

8. An Internet content intelligent creation device based on a generative adversarial network, characterized by, The device comprises: A dynamic clustering module that obtains network real-time hotspot data streams and groups the data streams by using an unsupervised clustering algorithm to identify a high-dynamic trend content set; A semantic sentiment fusion module that uses a pre-trained deep semantic analysis model to analyze multi-dimensional features of the high-dynamic trend content set to generate a trend-related content theme; A multi-modal extraction module that extracts features of multi-form content under the trend-related content theme by using a feature extractor to obtain a multi-form content feature set, wherein the multi-form content includes text, image, and video forms; A feature standardization module that uses a feature standardization algorithm to perform nonlinear standardization mapping on the multi-form content feature set to generate a unified feature expression set; A content creation and generation module that uses a pre-built generative adversarial network content generation framework to output created content based on the unified feature expression set. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to implement the steps of the method in any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method in any one of claims 1 to 7.

Images