AIGC-based Information Push Method

Abstract

The present invention belongs to the technical field of information push, and discloses an information push method based on AIGC. The method includes the following steps: constructing a knowledge graph generation model, a user portrait generation model, an information push content generation model, and an information push strategy generation model; using the knowledge graph generation model to generate a knowledge graph according to a number of real-time knowledge data; using the user portrait generation model to generate a user portrait and perform knowledge mapping according to real-time user behavior data; using the information push strategy generation model to generate an information push strategy according to the real-time user portrait after knowledge mapping; and using the information push content generation model to generate information push content according to the real-time information push strategy and the real-time user portrait after knowledge mapping. The present invention solves the problems of insufficient personalization, single content generation, weak push timeliness, and poor user experience existing in the prior art.

Description

Technical Field

[0001] The present invention belongs to the technical field of information push, and particularly relates to an information push method based on AIGC. Background Art

[0002] With the development of Internet technology, the information push technology has entered a rapid development era. The so-called information push technology is that information providers send information, data or content to be pushed to users on a public platform according to certain rules and strategies, so that users can understand and access the information. Especially in the protection and inheritance of intangible cultural heritage, it is necessary to actively push information and content related to intangible cultural heritage to users to achieve the dissemination of intangible cultural heritage and attract more users' attention and protection of intangible cultural heritage.

[0003] The existing information push technology has the following defects:

[0004] (1) Insufficient personalization: The existing information push technology often adopts a fixed push strategy, which cannot fully capture the dynamic interests and real-time behavior changes of users, resulting in a deviation between the pushed content and the actual needs of users;

[0005] (2) Single content generation: The content generation in the existing technology mostly relies on templates or predefined rules, and the generated information lacks diversity and innovation, making it difficult to meet users' needs for high-quality and personalized content;

[0006] (3) Weak push timeliness: The existing push relies on manual intervention, with a large cost investment and a lack of rapid adaptability to environmental changes, resulting in weak push timeliness;

[0007] (4) Poor user experience: Due to the low matching degree between the pushed content and users' needs, users often receive irrelevant or repetitive information, which affects the user experience and may lead to users' dissatisfaction with the push service or even abandonment of use. Summary of the Invention

[0008] In order to solve the problems of insufficient personalization, single content generation, weak push timeliness and poor user experience existing in the prior art, the purpose of the present invention is to provide an information push method based on AIGC.

[0009] The technical solution adopted by the present invention is as follows:

[0010] An information push method based on AIGC, comprising the following steps:

[0011] Using natural language processing algorithms, construct a knowledge graph generation model, use AIGC algorithms to construct a user profile generation model and an information push content generation model, and use reinforcement learning algorithms to construct an information push strategy generation model;

[0012] Collect a number of real-time knowledge data in the target domain, and based on the number of real-time knowledge data, use the knowledge graph generation model to generate a knowledge graph and obtain the real-time knowledge graph of the target domain;

[0013] Collect the real-time user behavior data of users on the public platform. Based on the real-time user behavior data, use the user profile generation model to generate a user profile to obtain a real-time user profile, and use the real-time knowledge graph to perform knowledge mapping on the real-time user profile of the user to obtain a real-time user profile after knowledge mapping;

[0014] Based on the real-time user profile after knowledge mapping, use the information push strategy generation model to generate an information push strategy to obtain a real-time information push strategy;

[0015] Based on the real-time information push strategy and the real-time user profile after knowledge mapping, use the information push content generation model to generate information push content to obtain real-time information push content, and push the real-time information push content to the corresponding users on the public platform.

[0016] Furthermore, the user profile generation model is constructed based on the RF-Attention-DBN algorithm, and the user profile generation model includes a key feature extraction module constructed based on the RF algorithm, an attention weight module constructed based on the Attention mechanism, and a user profile generation module constructed based on the DBN algorithm, which are connected in sequence.

[0017] Furthermore, the information push content generation model is constructed based on the MLP-AIGC algorithm, and the information push content generation model includes a feature processing module constructed based on the MLP algorithm and a group of information push content generation modules constructed based on the AIGC algorithm, which are connected in sequence.

[0018] Furthermore, the information push content generation module group includes a text generation module constructed based on the Transformer algorithm, an image generation module constructed based on GAN, an audio generation module constructed based on the WaveNet algorithm, and a video generation module constructed based on VGAN. The feature processing module is respectively connected to the text generation module, the image generation module, the audio generation module, and the video generation module.

[0019] Further, the information push strategy generation model is constructed based on the MPO-PPO algorithm, and the information push strategy generation model includes a meta-policy optimization module constructed based on the MPO algorithm and a reinforcement learning module constructed based on the PPO algorithm. The reinforcement learning module is provided with an agent, a policy network, and an experience replay pool. The agent is respectively connected to the policy network, the experience replay pool, and the meta-policy optimization module, and the meta-policy optimization module is connected to the experience replay pool.

[0020] Further, the knowledge graph generation model is constructed based on the BERT-Double CRF algorithm, and the knowledge graph generation model includes a text feature extraction module constructed based on the BERT algorithm, a named entity extraction module constructed based on the CRF algorithm, and an entity relationship extraction module constructed based on the CRF algorithm. The text feature extraction module is respectively connected to the named entity extraction module and the entity relationship extraction module.

[0021] Further, using natural language processing algorithms, a knowledge graph generation model is constructed, using AIGC algorithms, a user portrait generation model and an information push content generation model are constructed, and using reinforcement learning algorithms, an information push strategy generation model is constructed, including the following steps:

[0022] Collect a number of historical knowledge data, a number of multi-modal content data sets, and historical user behavior data of a number of users in a preset domain, and perform preprocessing to obtain a number of preprocessed historical knowledge data, a number of preprocessed multi-modal content data sets, and a number of preprocessed historical user behavior data;

[0023] According to a number of preprocessed historical knowledge data, using natural language processing algorithms, a knowledge graph generation model is constructed, and a number of historical knowledge graphs in a preset domain are generated;

[0024] According to a number of preprocessed historical user behavior data, using AIGC algorithms, a user portrait generation model is constructed, and a number of historical user portraits are generated;

[0025] According to a number of historical knowledge graphs, perform knowledge mapping on a number of historical user portraits to obtain a number of knowledge-mapped historical user portraits;

[0026] According to a number of knowledge-mapped historical user portraits, using reinforcement learning algorithms, an information push strategy generation model is constructed, and a number of historical information push strategies are generated;

[0027] According to a number of historical information push strategies, a number of multi-modal content data sets, and a number of knowledge-mapped historical user portraits, using AIGC algorithms, an information push content generation model is constructed.

[0028] Further, according to the historical user portraits after several knowledge mappings, use the reinforcement learning algorithm to construct an information push policy generation model and generate several historical information push policies, including the following steps:

[0029] Analyze the historical user portraits after knowledge mapping to obtain several historical user portrait states. According to the several historical user portrait states, define the state space of the intelligent agent, set several preset information push actions, and define the action space of the intelligent agent according to the several preset information push actions;

[0030] Define the reward function according to the influence of the preset information push actions on the historical user portrait states. Based on the state space, action space, and reward function, construct the intelligent agent and policy network of the reinforcement learning module, and use the experience replay mechanism to initialize the experience replay pool;

[0031] Regard the information push policy generation problem as a simulation environment, use the PPO algorithm to construct a policy network, and combine the policy network, experience replay pool, and intelligent agent to obtain an initial reinforcement learning module;

[0032] Set the information push policy generation tasks in several preset domains as several sub-scenarios for meta-policy optimization, and use the MPO algorithm to construct an initial meta-policy optimization module;

[0033] According to the several historical user portraits after knowledge mapping, train and optimize the initial reinforcement learning module, extract the historical policy network parameters of the policy network of the reinforcement learning module during the training and optimization, obtain the final reinforcement learning module, and generate several historical information push policies and corresponding historical reinforcement learning experiences;

[0034] Based on the several sub-scenarios, train and optimize the initial meta-policy optimization module according to the several historical policy network parameters, obtain the final meta-policy optimization module, and generate several historical meta-policy optimization experiences;

[0035] Store the several historical meta-policy optimization experiences and several historical reinforcement learning experiences into the experience replay pool of the final reinforcement learning module, and integrate the final meta-policy optimization module and the final reinforcement learning module to obtain the information push policy generation model.

[0036] Further, according to several historical information push policies, several multi-modal content data sets, and several historical user portraits after knowledge mapping, use the AIGC algorithm to construct an information push content generation model, including the following steps:

[0037] Using the AIGC algorithm, construct an initial information push content generation model; the initial information push content generation model includes an initial feature processing module, an initial text generation module, an initial image generation module, an initial audio generation module, and an initial video generation module;

[0038] Parse several multimodal content datasets in different preset fields to obtain several text content data, several image content data, several audio content data, and several video content data in different preset fields;

[0039] According to several text content data, several image content data, several audio content data, and several video content data, optimize and train the corresponding initial text generation module, initial image generation module, initial audio generation module, and initial video generation module to obtain an optimized text generation module, an optimized image generation module, an optimized audio generation module, and an optimized video generation module;

[0040] Extract the historical information push strategy features of the historical information push strategy and the knowledge-mapped historical user portrait features of the knowledge-mapped historical user portrait. According to several historical information push strategy features and several knowledge-mapped historical user portrait features, optimize and train the initial feature processing module to obtain a final feature processing module and generate several processed historical features;

[0041] According to several processed historical features, debug the optimized text generation module, optimized image generation module, optimized audio generation module, and optimized video generation module to obtain a final text generation module, a final image generation module, a final audio generation module, and a final video generation module;

[0042] Integrate the final feature processing module, final text generation module, final image generation module, final audio generation module, and final video generation module to obtain a final information push content generation model.

[0043] An information push system based on AIGC for implementing an information push method. The system includes an AIGC model construction unit, a knowledge graph generation unit, a user portrait generation unit, an information push strategy generation unit, and an information push content generation unit connected in sequence.

[0044] The beneficial effects of the present invention are:

[0045] An information push method based on AIGC provided by the present invention can construct a dynamic user portrait and analyze user behavior in real time through a user portrait generation model, capture users' interests and needs more accurately, achieve highly personalized information push, and improve user satisfaction; it can automatically and intelligently generate information push content through an information push content generation model, generate rich, diverse, and innovative push content, including various modalities such as text, image, audio, and video, to meet users' needs for high-quality content; use a knowledge graph generation model to construct a more complete knowledge graph, which can identify complex domain relationships and deep semantics, thereby improving the relevance and accuracy of push content, and generate an information push strategy according to users' habits through an information push strategy generation model, avoiding manual intervention, reducing cost investment, and being able to quickly adapt to environmental changes to ensure the best match between push content and user needs; through precise content push, the push of irrelevant and repetitive information is reduced, significantly improving the user experience and enhancing users' trust and dependence on the push service.

[0046] Other beneficial effects of the present invention will be further described in the specific implementation manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flowchart of the information push method based on AIGC in the present invention.

[0048] Figure 2 is a structural block diagram of the information push system based on AIGC in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0049] The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments.

[0050] Embodiment 1:

[0051] As Figure 1 shown, this embodiment provides an information push method based on AIGC, including the following steps:

[0052] S1: Use natural language processing algorithms to construct a knowledge graph generation model, use Artificial Intelligence Generated Content (AIGC) algorithms to construct a user portrait generation model and an information push content generation model, and use reinforcement learning algorithms to construct an information push strategy generation model, including the following steps:

[0053] S1-1: Collect a number of historical knowledge data, a number of multimodal content data sets, and historical user behavior data of a number of users in a preset field, and perform preprocessing to obtain a number of preprocessed historical knowledge data, a number of preprocessed multimodal content data sets, and a number of preprocessed historical user behavior data;

[0054] In this embodiment, the protection and inheritance of intangible cultural heritage are used as the application scenario. Through media, film and television, and new media public platforms as data sources, information is pushed in different intangible cultural heritage fields, and the generated intangible cultural heritage content is pushed to public platform programs such as users' social media, news clients, and video programs; Social media platform: According to the user's interest preferences, push relevant intangible cultural heritage content generated using the AIGC model, such as intangible cultural heritage project introductions, inheritor stories, intangible cultural heritage activity information, etc.; News client: According to the user's reading habits, push relevant intangible cultural heritage news and information generated using the AIGC model; Video program: According to the user's viewing preferences, push relevant intangible cultural heritage videos generated using the AIGC model, such as intangible cultural heritage skill demonstrations, intangible cultural heritage documentary films, etc.;

[0055] S1-2: According to a number of preprocessed historical knowledge data, use natural language processing algorithms to construct a knowledge graph generation model and generate a number of historical knowledge graphs in a preset field;

[0056] The knowledge graph generation model is constructed based on the Bidirectional Encoder Representations from Transformers (BERT)-Double Conditional Random Fields (CRF) algorithm, and the knowledge graph generation model includes a text feature extraction module constructed based on the BERT algorithm, a named entity extraction module constructed based on the CRF algorithm, and an entity relationship extraction module constructed based on the CRF algorithm. The text feature extraction module is respectively connected to the named entity extraction module and the entity relationship extraction module;

[0057] The BERT module can capture the deep semantic information in the knowledge data, which is very useful for identifying different types of entities and extracting the relationships between entities. The CRF module can consider the dependencies between adjacent labels, which can help the model learn the sequence dependencies of entity labels, thereby improving the accuracy of named entity annotation. And in the entity relationship annotation task, the CRF module processes the relationship classification between entity pairs, especially when dealing with multiple entities and relationships, to achieve entity relationship annotation;

[0058] S1-3: According to a number of preprocessed historical user behavior data, use the AIGC algorithm to construct a user portrait generation model and generate a number of historical user portraits;

[0059] The user portrait generation model is constructed based on the Random Forest (RF)-Attention-Deep Belief Network (DBN) algorithm, and the user portrait generation model includes a key feature extraction module constructed based on the RF algorithm, an attention weight module constructed based on the Attention mechanism, and a user portrait generation module constructed based on the DBN algorithm, which are connected in sequence;

[0060] The key feature extraction module screens key features from the input user behavior data through the internal Classification And Regression Tree (CART), and extracts several key features related to the user portrait tags, including platform type, access frequency, access time, followed entries, comment methods, etc.; the attention weight module realizes weighted fusion of several key features through preset attention weights, enhances the influence degree of the attention features on the label prediction, and improves the label prediction accuracy; the user portrait generation module generates labels according to the fused features, pre-trains the network in an unsupervised manner, and then performs supervised fine-tuning, which helps to improve the accuracy of label prediction;

[0061] S1-4: Perform knowledge mapping on several historical user portraits according to several historical knowledge graphs to obtain several knowledge-mapped historical user portraits;

[0062] S1-5: According to several knowledge-mapped historical user portraits, use the reinforcement learning algorithm to construct an information push policy generation model and generate several historical information push policies;

[0063] The information push policy generation model is constructed based on the Meta-Policy Optimization (MPO)-Proximal Policy Optimization (PPO) algorithm, and the information push policy generation model includes a meta-policy optimization module constructed based on the MPO algorithm and a reinforcement learning module constructed based on the PPO algorithm. The reinforcement learning module is provided with an agent, a policy network, and an experience replay pool. The agent is respectively connected to the policy network, the experience replay pool, and the meta-policy optimization module, and the meta-policy optimization module is connected to the experience replay pool;

[0064] The meta - policy optimization module can quickly adapt to the changing information push environment, initialize the policy network based on historical experience, and utilize the knowledge learned from other tasks when dealing with new or unseen tasks, improving the generalization ability of the reinforcement learning module, reducing the need to train from scratch on new tasks, and saving computational resources and time. The PPO algorithm of the reinforcement learning module improves the learning stability by restricting the policy update amplitude, reducing the risk of policy collapse. By using the experience replay pool, data samples can be reused, improving the sample utilization rate and learning efficiency. Through continuous iterative updates, the policy network can learn a better information push policy. The experience replay pool is used to store meta - policy optimization experiences and reinforcement learning experiences. Reinforcement learning experiences include states, actions, rewards, and next states. Meta - policy optimization experiences include the network parameters of the policy network in the reinforcement learning module and the corresponding policy adjustment categories.

[0065] According to the historical user portraits after several knowledge mappings, use the reinforcement learning algorithm to construct an information push policy generation model and generate several historical information push policies, including the following steps:

[0066] S1 - 5 - 1: Analyze the historical user portraits after knowledge mapping to obtain several historical user portrait states. According to the several historical user portrait states, define the state space of the agent, set several preset information push actions, and define the action space of the agent according to the several preset information push actions.

[0067] S1 - 5 - 2: Define the reward function according to the influence of the preset information push actions on the historical user portrait states. Based on the state space, action space, and reward function, construct the agent and policy network of the reinforcement learning module, and use the experience replay mechanism to initialize the experience replay pool. The reward function is a key component in reinforcement learning, which guides the agent to learn how to select actions from the state space to obtain the maximum long - term reward, thereby optimizing the push policy.

[0068] S1 - 5 - 3: Regard the information push policy generation problem as a simulation environment, use the PPO algorithm to construct the policy network, and combine the policy network, experience replay pool, and agent to obtain the initial reinforcement learning module.

[0069] S1 - 5 - 4: Set the information push policy generation tasks in several preset domains as several sub - scenarios of meta - policy optimization, and use the MPO algorithm to construct the initial meta - policy optimization module.

[0070] S1-5-5: Train and optimize the initial reinforcement learning module based on the historical user portraits after several knowledge mappings, extract the historical policy network parameters of the policy network in the reinforcement learning module during the training and optimization process to obtain the final reinforcement learning module, and generate several historical information push policies and corresponding historical reinforcement learning experiences; Through training, the policy network can learn better information push policies, thereby improving the relevance and effectiveness of the push;

[0071] S1-5-6: Based on several sub-scenarios, train and optimize the initial meta-policy optimization module according to several historical policy network parameters to obtain the final meta-policy optimization module, and generate several historical meta-policy optimization experiences;

[0072] S1-5-7: Store several historical meta-policy optimization experiences and several historical reinforcement learning experiences in the experience replay pool of the final reinforcement learning module, and integrate the final meta-policy optimization module and the final reinforcement learning module to obtain an information push policy generation model; Storing experiences allows the model to continue to benefit from past learning, while the integration of the modules ensures the consistency and collaborative work of the entire system, thereby generating more effective push policies;

[0073] S1-6: Construct an information push content generation model using the AIGC algorithm based on several historical information push policies, several multi-modal content datasets, and several historical user portraits after knowledge mapping;

[0074] The information push content generation model is constructed based on the Multi-Layer Perceptron (MLP)-AIGC algorithm, and the information push content generation model includes a feature processing module constructed based on the MLP algorithm and an information push content generation module group constructed based on the AIGC algorithm;

[0075] The information push content generation module group includes a text generation module constructed based on the Transformer algorithm, an image generation module constructed based on the Generative Adversarial Networks (GAN), an audio generation module constructed based on the WaveNet algorithm, and a video generation module constructed based on the VidioGenerative Adversarial Networks (VGAN). The feature processing module is respectively connected to the text generation module, the image generation module, the audio generation module, and the video generation module;

[0076] The AIGC algorithm refers to an algorithm that uses artificial intelligence technology to generate content. Such algorithms usually include natural language processing, deep learning, generative adversarial networks, variational autoencoders, etc., and are used to create text, images, audio, video, and other forms of data or media content; the feature processing module processes the information push strategy features and the user portrait features after knowledge mapping to obtain processed features. The processed features include the push time selection action feature, content customization action feature, content type selection action feature, etc. of the information push strategy. The user portrait features include user features such as interest tag features, behavior pattern features, knowledge preference features, etc., and are feature vectors of a fixed length to ensure that the feature vectors can be used by the subsequent content generation module;

[0077] According to a number of historical information push strategies, a number of multi-modal content data sets, and a number of historical user portraits after knowledge mapping, use the AIGC algorithm to construct an information push content generation model, including the following steps:

[0078] S1-6-1: Use the AIGC algorithm to construct an initial information push content generation model; the initial information push content generation model includes an initial feature processing module, an initial text generation module, an initial image generation module, an initial audio generation module, and an initial video generation module; use the AIGC algorithm to construct an initial model that includes a feature processing module and multiple content generation modules (text, image, audio, video), and this model will be used to generate multi-modal content;

[0079] S1-6-2: Parse the data of a number of multi-modal content data sets in different preset fields to obtain a number of text content data, a number of image content data, a number of audio content data, and a number of video content data in different preset fields; data parsing ensures that the model can receive high-quality and structured input data, which is crucial for training an efficient content generation model;

[0080] S1-6-3: According to a number of text content data, a number of image content data, a number of audio content data, and a number of video content data, optimize and train the corresponding initial text generation module, initial image generation module, initial audio generation module, and initial video generation module to obtain an optimized text generation module, an optimized image generation module, an optimized audio generation module, and an optimized video generation module; optimization training improves the accuracy and diversity of each module when generating corresponding content, making the generated content more in line with the characteristics of the intangible cultural heritage field and user preferences;

[0081] S1-6-4: Extract the historical information push policy features of the historical information push policy and the features of the historical user profile after knowledge mapping of the historical user profile. Optimize and train the initial feature processing module based on a number of historical information push policy features and a number of features of the historical user profile after knowledge mapping to obtain the final feature processing module, and generate a number of features after historical processing. Optimizing the feature processing module helps to better convert the user profile and push policy into a form that the content generation module can understand, improving the relevance and personalization of content generation.

[0082] S1-6-5: Debug the optimized text generation module, optimized image generation module, optimized audio generation module, and optimized video generation module based on a number of features after historical processing to obtain the final text generation module, final image generation module, final audio generation module, and final video generation module. Debugging ensures that the content generation module can generate high-quality content based on the optimized features, improving the attractiveness of the pushed content and user satisfaction.

[0083] S1-6-6: Integrate the final feature processing module, final text generation module, final image generation module, final audio generation module, and final video generation module to obtain the final information push content generation model. The integrated model can work collaboratively to generate multi-modal content based on the user profile and push policy, improving the efficiency and effectiveness of the entire push system.

[0084] S2: Collect a number of real-time knowledge data in the target domain, and use the knowledge graph generation model to generate a knowledge graph for the target domain based on the number of real-time knowledge data, including the following steps:

[0085] S2-1: Collect a number of real-time knowledge data in the target domain and perform preprocessing to obtain a number of preprocessed real-time knowledge data. Collect the latest knowledge data from the target domain, such as news, research reports, social media posts, etc., and then perform preprocessing operations such as data cleaning, deduplication, and formatting to ensure the quality and consistency of the data. The preprocessed data is more suitable for subsequent knowledge graph generation, reducing the impact of noise and incorrect data, and improving the accuracy and reliability of the knowledge graph.

[0086] S2-2: Use the knowledge graph generation model to extract named entities and entity relationships from a number of preprocessed real-time knowledge data to obtain a number of knowledge named entities and a number of knowledge entity relationships. Extracting named entities and entity relationships is a key step in constructing a knowledge graph, which provides a structured data foundation for the knowledge graph, enabling the knowledge graph to accurately reflect the relationships between entities.

[0087] S2-3: Construct a knowledge graph based on a number of knowledge named entities and a number of knowledge entity relationships to obtain a real-time knowledge graph of the target domain; the knowledge graph represents knowledge in a graphical way, which is convenient for understanding and analysis, integrates scattered knowledge points into an organic whole, and provides support for subsequent knowledge mapping;

[0088] S3: Collect the real-time user behavior data of users on the public platform. According to the real-time user behavior data, use the user portrait generation model to generate a user portrait to obtain a real-time user portrait, and use the real-time knowledge graph to perform knowledge mapping on the real-time user portrait of the user to obtain a real-time user portrait after knowledge mapping, including the following steps:

[0089] S3-1: Collect the real-time user behavior data of users on the public platform and perform preprocessing to obtain preprocessed real-time user behavior data; collect the real-time behavior data of users from public platforms (such as social media, news websites, forums, etc.), and then perform preprocessing operations such as data cleaning, denoising, and normalization;

[0090] S3-2: Use the user portrait generation model to extract a number of real-time key features from the preprocessed real-time user behavior data, perform feature fusion on the number of real-time key features according to the preset attention weight value to obtain a real-time fusion feature, and generate a user portrait according to the real-time fusion feature to obtain a real-time user portrait; by focusing on key features, the model can more accurately capture the personalized information of users, and the real-time user portrait can reflect the latest behavior and interest changes of users, improving the accuracy and timeliness of information push;

[0091] S3-3: Obtain a number of real-time user portrait tags of the real-time user portrait, and obtain the similarity between each real-time user portrait tag and a number of knowledge named entities in the real-time knowledge graph; extract tags from the real-time user portrait and calculate the similarity between these tags and the knowledge named entities in the real-time knowledge graph to find the association between user interests and knowledge entities. Similarity matching helps to connect the user's interests with the professional knowledge in the intangible cultural heritage field and provide more relevant content for users;

[0092] S3-4: Map the knowledge named entities in the real-time knowledge graph to the real-time user portrait tag with the highest similarity, and associate the corresponding knowledge entity relationship with the real-time user portrait tag with the highest similarity to obtain a real-time user portrait after knowledge mapping; through knowledge mapping, the user portrait contains richer knowledge information, which helps to understand the user's needs more deeply. The mapped user portrait can better guide the information push strategy, ensure that the pushed content is highly relevant to the user's interests and knowledge background, provide a more accurate basis for personalized information push, and improve the accuracy of push and user satisfaction;

[0093] S4: According to the real-time user portrait after knowledge mapping, use the information push strategy generation model to generate the information push strategy, including the following steps:

[0094] S4-1: According to the target domain, extract the corresponding target historical meta-policy optimization experience from several historical meta-policy optimization experiences of different sub-scenarios in the experience replay pool of the information push strategy generation model. According to the target historical meta-policy optimization experience, use the meta-policy optimization module to initialize the real-time policy network parameters of the policy network of the reinforcement learning module to obtain an updated policy network; by reusing historical experience, the training process of the policy network can be accelerated, and the historical experience for a specific domain can improve the domain adaptability of the policy network;

[0095] S4-2: Randomly extract several historical reinforcement learning experiences from the experience replay pool, and update the action space of the agent according to the preset information push actions of the several historical reinforcement learning experiences to obtain an updated action space; due to the diversity of historical experience, the action space is more abundant and can cover more possible push strategies. The update of the action space enables the agent to better adapt to different user needs and scenarios;

[0096] S4-3: Analyze the real-time user portrait after knowledge mapping to obtain several real-time user portrait states. According to the several real-time user portrait states, update the state space of the agent to obtain an updated state space; the update of the state space enables the agent to formulate push strategies based on the personalized characteristics of users, and a more accurate state representation helps to generate more effective push strategies;

[0097] S4-4: Connect the updated state space to the input end of the updated policy network, connect the updated action space to the output end of the updated policy network, and use the agent to control the updated policy network to generate the probability distribution of all possible information push actions in the updated action space corresponding to each real-time user portrait state in the updated state space;

[0098] S4-5: Take the possible information push action with the highest probability distribution in the updated action space as the execution information push action for the real-time user portrait state, traverse all real-time user portrait states in the updated state space to obtain several execution information push actions, and obtain the real-time information push strategy according to the several execution information push actions;

[0099] The information push strategy includes:

[0100] Content selection action: Select specific types of content related to user interests or historical behaviors (such as specific content in the intangible cultural heritage field, etc.);

[0101] Content customization actions: Customize the content style, language, or format according to user preferences; Personalize and adjust the difficulty or depth of the content;

[0102] Push timing actions: Determine the optimal push time, such as during user active periods or when specific events occur; Adjust the push frequency according to the user's living habits;

[0103] Push channel actions: Push content on different platforms according to the user's usage habits;

[0104] S5: According to the real-time information push strategy and the real-time user portrait after knowledge mapping, use the information push content generation model to generate the information push content, obtain the real-time information push content, and push the real-time information push content to the corresponding users on the public platform, including the following steps:

[0105] S5-1: Extract the real-time information push strategy features of the real-time information push strategy and the knowledge mapping real-time user portrait features of the real-time user portrait after knowledge mapping; Extract key features from the real-time information push strategy, such as push time selection action features, content customization action features, content type selection action features, etc., and at the same time extract user features from the real-time user portrait after knowledge mapping, such as interest tag features, behavior pattern features, knowledge preference features, etc.;

[0106] S5-2: Use the information push content generation model to process the real-time information push strategy features and the knowledge mapping real-time user portrait features to obtain the real-time processed features; Feature processing improves the generalization ability of the model, enabling it to better adapt to different push scenarios and providing more accurate guidance for content generation;

[0107] S5-3: According to the real-time processed features, generate the information push content to obtain the real-time information push content, such as text, images, videos, etc., and push the real-time information push content to the corresponding users on the public platform according to the real-time information push strategy.

[0108] Embodiment 2:

[0109] As Figure 2 shown, this embodiment provides an information push system based on AIGC for implementing the information push method. The system includes an AIGC model construction unit, a knowledge graph generation unit, a user portrait generation unit, an information push strategy generation unit, and an information push content generation unit connected in sequence;

[0110] The AIGC model construction unit is used to construct a knowledge graph generation model using natural language processing algorithms, construct a user portrait generation model and an information push content generation model using AIGC algorithms, and construct an information push strategy generation model using reinforcement learning algorithms;

[0111] A knowledge graph generation unit, configured to collect a number of real-time knowledge data in a target domain, and based on the number of real-time knowledge data, use a knowledge graph generation model to perform knowledge graph generation to obtain a real-time knowledge graph of the target domain;

[0112] A user portrait generation unit, configured to collect real-time user behavior data of a user on a public platform, and based on the real-time user behavior data, use a user portrait generation model to perform user portrait generation to obtain a real-time user portrait, and use the real-time knowledge graph to perform knowledge mapping on the real-time user portrait of the user to obtain a real-time user portrait after knowledge mapping;

[0113] An information push strategy generation unit, configured to generate an information push strategy based on the real-time user portrait after knowledge mapping, using an information push strategy generation model to obtain a real-time information push strategy;

[0114] An information push content generation unit, configured to generate real-time information push content based on the real-time information push strategy and the real-time user portrait after knowledge mapping, using an information push content generation model, and push the real-time information push content to the corresponding user on the public platform.

[0115] An information push method based on AIGC provided by the present invention can construct a dynamic user portrait and analyze user behavior in real time through a user portrait generation model, can capture the interests and needs of users more accurately, realize highly personalized information push, and improve user satisfaction; through an information push content generation model, it can generate information push content automatically and intelligently, and can generate rich and diverse, innovative push content, including various modalities such as text, image, audio and video, to meet the needs of users for high-quality content; using a knowledge graph generation model to construct a more perfect knowledge graph can identify complex domain relationships and deep semantics, thereby improving the relevance and accuracy of push content. By using an information push strategy generation model to generate an information push strategy according to the user's habits, it avoids manual intervention, reduces cost investment, and can quickly adapt to environmental changes to ensure the best match between push content and user needs; through accurate content push, it reduces the push of irrelevant and duplicate information, significantly improves the user experience, and enhances the user's trust and dependence on the push service.

[0116] The present invention is not limited to the above optional embodiments, and anyone can obtain other various forms of products under the inspiration of the present invention. The above specific embodiments should not be construed as limiting the protection scope of the present invention, and the protection scope of the present invention should be defined by the claims, and the description can be used to interpret the claims.

Claims

1. An information push method based on AIGC, characterized in that: The steps include: Use natural language processing algorithms to build a knowledge graph generation model, use artificial intelligence content creation AIGC algorithms to build user portrait generation models and information push content generation models, and use reinforcement learning algorithms to build information push strategy generation models; The user portrait generation model is constructed based on the random forest RF-attention mechanism Attention-deep belief network DBN algorithm, and the user portrait generation model includes a key feature extraction module constructed based on the RF algorithm, an attention weight module constructed based on the Attention mechanism, and a user portrait generation module constructed based on the DBN algorithm, which are sequentially connected; The information push content generation model is constructed based on a multi-layer perceptron MLP-AIGC algorithm, and the information push content generation model includes a feature processing module constructed based on the MLP algorithm and an information push content generation module group constructed based on the AIGC algorithm, which are connected in sequence; The information push strategy generation model is constructed based on the meta-strategy optimization MPO-similarity and difference strategy optimization PPO algorithm, and the information push strategy generation model includes a meta-strategy optimization module constructed based on the MPO algorithm and a reinforcement learning module constructed based on the PPO algorithm, the reinforcement learning module is provided with an intelligent agent, a strategy network and an experience replay pool, the intelligent agent is respectively connected to the strategy network, the experience replay pool and the meta-strategy optimization module, and the meta-strategy optimization module is connected to the experience replay pool; Collect some real-time knowledge data in the target field, and use the knowledge graph generation model to generate a knowledge graph based on the real-time knowledge data to obtain a real-time knowledge graph in the target field; Collect the real-time user behavior data of users on the public platform, use the user portrait generation model to generate user portraits based on the real-time user behavior data, obtain real-time user portraits, and use the real-time knowledge graph to perform knowledge mapping on the real-time user portraits of users to obtain real-time user portraits after knowledge mapping, including the following steps: Collecting some real-time knowledge data in the target field and preprocessing them to obtain some preprocessed real-time knowledge data; Use the knowledge graph generation model to extract named entities and entity relationships from some pre-processed real-time knowledge data to obtain some knowledge named entities and some knowledge entity relationships; Based on several knowledge named entities and several knowledge entity relationships, a knowledge graph is constructed to obtain a real-time knowledge graph of the target field, including the following steps: Collecting real-time user behavior data of users on the public platform and preprocessing it to obtain real-time user behavior data after preprocessing; Use the user portrait generation model to extract several real-time key features of the pre-processed real-time user behavior data, perform feature fusion on several real-time key features according to preset attention weight values ​​to obtain real-time fusion features, and generate user portraits based on the real-time fusion features to obtain real-time user portraits; Obtain several real-time user portrait tags of the real-time user portrait, and obtain the similarity between each real-time user portrait tag and several knowledge named entities of the real-time knowledge graph; Map the knowledge named entities of the real-time knowledge graph to the real-time user portrait tag with the highest similarity, and associate the corresponding knowledge entity relationship with the real-time user portrait tag with the highest similarity to obtain the real-time user portrait after knowledge mapping; According to the real-time user portrait after knowledge mapping, the information push strategy generation model is used to generate the information push strategy to obtain the real-time information push strategy, including the following steps: According to the target domain, the corresponding target historical meta-strategy optimization experience is extracted from several historical meta-strategy optimization experiences of different sub-scenarios in the experience playback pool of the information push strategy generation model, and the real-time policy network parameters of the policy network of the reinforcement learning module are initialized using the meta-strategy optimization module based on the target historical meta-strategy optimization experience to obtain an updated policy network. Randomly extract a number of historical reinforcement learning experiences from the experience replay pool, and push actions according to the preset information of the historical reinforcement learning experiences to update the action space of the agent and obtain an updated action space; The real-time user portrait after knowledge mapping is parsed to obtain a number of real-time user portrait states, and the state space of the intelligent agent is updated according to the real-time user portrait states to obtain an updated state space; Connect the updated state space to the input of the updated policy network, connect the updated action space to the output of the updated policy network, use the agent to control the updated policy network, and generate a probability distribution of all possible information push actions in the updated action space corresponding to each real-time user profile state in the updated state space; The possible information push action with the highest probability distribution in the updated action space is used as the execution information push action of the real-time user portrait state, all real-time user portrait states in the updated state space are traversed to obtain a number of execution information push actions, and a real-time information push strategy is obtained according to the number of execution information push actions; Information push strategies include: Content selection actions, content customization actions, push timing actions, and push channel actions; According to the real-time information push strategy and the real-time user portrait after knowledge mapping, the information push content generation model is used to generate information push content, obtain real-time information push content, and push the real-time information push content to the corresponding users on the public platform, including the following steps: Extracting the real-time information push strategy features of the real-time information push strategy and the real-time user portrait features after knowledge mapping of the real-time user portrait; Use the information push content generation model to process the real-time information push strategy features and the real-time user portrait features after knowledge mapping to obtain real-time processed features; According to the real-time processed features, information push content is generated to obtain real-time information push content, and according to the real-time information push strategy, the real-time information push content is pushed to the corresponding users on the public platform.

2. The information push method based on AIGC according to claim 1, characterized in that: The information push content generation module group includes a text generation module constructed based on the Transformer algorithm, an image generation module constructed based on the generative adversarial network GAN, an audio generation module constructed based on the wavelet transform network WaveNet algorithm, and a video generation module constructed based on the video generative adversarial network VGAN. The feature processing module is respectively connected to the text generation module, the image generation module, the audio generation module and the video generation module.

3. The information push method based on AIGC according to claim 1, characterized in that: The knowledge graph generation model is constructed based on the BERT-Double Conditional Random Field (CRF) algorithm, a bidirectional encoder representation from Transformers, and the knowledge graph generation model includes a text feature extraction module constructed based on the BERT algorithm, a named entity extraction module constructed based on the CRF algorithm, and an entity relationship extraction module constructed based on the CRF algorithm. The text feature extraction module is connected to the named entity extraction module and the entity relationship extraction module respectively.

4. The information push method based on AIGC according to claim 3, characterized in that: Use the natural language processing algorithm to build a knowledge graph generation model, use the AIGC algorithm to build a user portrait generation model and an information push content generation model, and use the reinforcement learning algorithm to build an information push strategy generation model, including the following steps: Collecting a number of historical knowledge data in a preset field, a number of multimodal content data sets, and a number of historical user behavior data of users, and preprocessing them to obtain a number of preprocessed historical knowledge data, a number of preprocessed multimodal content data sets, and a number of preprocessed historical user behavior data; Based on some pre-processed historical knowledge data, a natural language processing algorithm is used to build a knowledge graph generation model and generate historical knowledge graphs in several preset fields; Based on some pre-processed historical user behavior data, use the AIGC algorithm to build a user portrait generation model and generate some historical user portraits; According to several historical knowledge graphs, knowledge mapping is performed on several historical user portraits to obtain several historical user portraits after knowledge mapping; Based on the historical user portraits after knowledge mapping, a reinforcement learning algorithm is used to build an information push strategy generation model and generate several historical information push strategies; Based on several historical information push strategies, several multimodal content data sets, and several historical user portraits after knowledge mapping, the AIGC algorithm is used to build an information push content generation model.

5. The information push method based on AIGC according to claim 4, characterized in that: Based on the historical user portraits after several knowledge mappings, a reinforcement learning algorithm is used to build an information push strategy generation model and generate several historical information push strategies, including the following steps: Analyze the historical user portraits after knowledge mapping to obtain several historical user portrait states, define the state space of the intelligent agent based on the several historical user portrait states, set several preset information push actions, and define the action space of the intelligent agent based on the several preset information push actions; According to the impact of the preset information push action on the historical user portrait status, the reward function is defined. Based on the state space, action space and reward function, the intelligent agent and policy network of the reinforcement learning module are constructed, and the experience replay mechanism is used to initialize the experience replay pool. The information push strategy generation problem is used as a simulation environment, and the PPO algorithm is used to build a policy network. The policy network, experience replay pool, and agent are combined to obtain the initial reinforcement learning module. The information push strategy generation tasks in several preset fields are set as several sub-scenarios of meta-strategy optimization, and the MPO algorithm is used to build the initial meta-strategy optimization module; According to the historical user portraits after several knowledge mappings, the initial reinforcement learning module is trained and optimized, and the historical policy network parameters of the policy network of the reinforcement learning module in the training optimization are extracted to obtain the final reinforcement learning module, and several historical information push strategies and corresponding historical reinforcement learning experiences are generated; Based on several sub-scenarios and several historical policy network parameters, the initial meta-policy optimization module is trained and optimized to obtain the final meta-policy optimization module, and several historical meta-policy optimization experiences are generated; Several historical meta-strategy optimization experiences and several historical reinforcement learning experiences are stored in the experience replay pool of the final reinforcement learning module, and the final meta-strategy optimization module and the final reinforcement learning module are integrated to obtain an information push strategy generation model.

6. The information push method based on AIGC according to claim 5, characterized in that: Based on several historical information push strategies, several multimodal content data sets, and several historical user portraits after knowledge mapping, the AIGC algorithm is used to build an information push content generation model, including the following steps: Using the AIGC algorithm, construct an initial information push content generation model; the initial information push content generation model includes an initial feature processing module, an initial text generation module, an initial image generation module, an initial audio generation module and an initial video generation module; Performing data analysis on a plurality of multimodal content data sets in different preset fields to obtain a plurality of text content data, a plurality of image content data, a plurality of audio content data, and a plurality of video content data in different preset fields; According to a plurality of text content data, a plurality of image content data, a plurality of audio content data and a plurality of video content data, an initial text generation module, an initial image generation module, an initial audio generation module and an initial video generation module are optimized and trained to obtain an optimized text generation module, an optimized image generation module, an optimized audio generation module and an optimized video generation module; Extract the historical information push strategy features of the historical information push strategy and the knowledge-mapped historical user portrait features of the historical user portrait after knowledge mapping, optimize and train the initial feature processing module based on the historical information push strategy features and the knowledge-mapped historical user portrait features, obtain the final feature processing module, and generate a number of historical processed features; According to several historical processed features, the optimized text generation module, the optimized image generation module, the optimized audio generation module, and the optimized video generation module are debugged to obtain a final text generation module, a final image generation module, a final audio generation module, and a final video generation module; The final feature processing module, the final text generation module, the final image generation module, the final audio generation module and the final video generation module are integrated to obtain the final information push content generation model.

Images