Block chain and deep learning-based enterprise propaganda virtual digital human system

Abstract

The invention relates to the technical field of digital human systems, and discloses an enterprise propaganda virtual digital human system based on a block chain and deep learning, and the system ensures the safety and non-tampering of enterprise propaganda data and user interaction data through a block chain storage unit; the intelligent level of the virtual digital human is improved by utilizing a deep learning model construction unit, and the functions of language understanding, dialogue generation, image rendering and the like are realized; the enterprise information extraction unit and the user request processing unit are respectively responsible for data updating and request analysis; the dialogue management unit and the image rendering unit generate proper replies and dynamic images according to user requirements and scenes; the block chain verification unit guarantees the authenticity of the data; the feedback optimization unit realizes continuous optimization of the model; and the multi-channel publishing and managing unit ensures multi-platform synchronization and consistency of contents. According to the system, the personalized, interactive and efficient levels of enterprise propaganda are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of digital human systems, and in particular to a corporate publicity virtual digital human system based on blockchain and deep learning. Background Art

[0002] With the rapid development of information technology, the way of corporate publicity is undergoing unprecedented changes. Traditional publicity methods, such as paper advertisements and television advertisements, can no longer meet the needs of modern enterprises for personalized, interactive and efficient publicity. In this context, virtual digital people, as an emerging publicity carrier, are gradually emerging and becoming the new favorite of corporate publicity with their unique charm and unlimited potential.

[0003] However, the existing virtual digital human system still faces many challenges in practical application. On the one hand, the safe storage and effective management of corporate promotional materials is an urgent problem to be solved. Traditional storage methods are vulnerable to attacks, data security is difficult to guarantee, and data management and update efficiency is low. On the other hand, the intelligence level of virtual digital humans still needs to be improved. Current virtual digital humans often lack the ability of deep learning and self-optimization, and it is difficult to respond flexibly according to user needs and feedback, resulting in a poor interactive experience.

[0004] The image rendering and dialogue generation of virtual digital people are also one of the technical difficulties. Traditional rendering technology is difficult to achieve highly realistic dynamic images and facial animations, which affects the user's immersion and interactivity. At the same time, dialogue generation models often lack context understanding and context perception capabilities, resulting in the lack of coherence and pertinence in the generated reply text. Summary of the invention

[0005] The purpose of the present invention is to provide a corporate promotion virtual digital human system based on blockchain and deep learning to solve the problems raised in the above background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a virtual digital human system for enterprise publicity based on blockchain and deep learning, the system comprising: A user interaction interface unit, used to receive enterprise promotion related requests input by users and display interactive responses of virtual digital people; Blockchain storage unit, used to securely store corporate promotional materials, user interaction data, and virtual digital human learning and processing records; A deep learning model building unit, which is used to build and train the deep learning model of virtual digital humans, including language understanding models, dialogue generation models, and image rendering models; An enterprise information extraction unit, connected to the blockchain storage unit, for extracting and updating enterprise promotional materials, including enterprise profiles, product services, and news information; A user request processing unit, connected to the user interaction interface unit, for receiving and parsing user requests and converting the requests into structured data for subsequent model processing; A dialogue management unit, connected to the deep learning model construction unit and the user request processing unit, for invoking the dialogue generation model to generate an appropriate reply text based on the user request and the current dialogue context; An image rendering unit, connected to the deep learning model building unit, is used to call the image rendering model to generate a dynamic image and expression animation of a virtual digital human according to the content and scene of the conversation; A blockchain verification unit, used to perform blockchain verification on user interaction data and record the verified data into a blockchain storage unit; Feedback optimization unit, which is used to collect user feedback and interaction data, analyze model performance, and regularly tune and update deep learning models; The multi-channel publishing and management unit is used to publish the interactive content of virtual digital people to multiple social platforms and corporate websites, while managing the content updates and consistency of different channels.

[0007] Preferably, the language understanding model in the deep learning model construction unit is constructed using the Transformer architecture and incorporates an attention mechanism.

[0008] Preferably, the step of training the language understanding model includes: Step A: Collect text data related to corporate publicity, including corporate official website content, product manuals and press releases, and perform data cleaning and preprocessing; Step B: Perform word segmentation, part-of-speech tagging and named entity recognition on the preprocessed text data to construct a text feature vector; Step C: Construct a Transformer-based language understanding model, including input layer, encoding layer, attention layer and output layer; the input layer receives the text feature vector, the encoding layer processes the text sequence through a multi-layer Transformer encoder, the attention layer introduces a self-attention mechanism to capture the internal relationship of the text, and the output layer is used to output the understanding representation of the text; Step D: Use the data set processed in step A to train the model and use the cross entropy loss function to optimize the model parameters; Step E: Evaluate the model performance on the validation set, and measure the model effect through accuracy and F1 value indicators; Step F: Adjust the model structure or hyperparameters based on the evaluation results, and repeat steps C to E until the model performance meets the requirements.

[0009] Preferably, the data cleaning method in step A includes removing irrelevant information, correcting spelling errors, unifying text formats and filtering low-quality texts.

[0010] Preferably, the preprocessing step in step A also includes text enhancement, which increases the diversity of training data through synonym replacement and sentence transformation.

[0011] Preferably, the dialogue generation model adopts a sequence-to-sequence Seq2Seq architecture and is trained in combination with reinforcement learning; the specific implementation methods include: Step 1: The dialogue generation model consists of an encoder, a decoder, and an attention mechanism. The encoder uses a multi-layer LSTM network to encode the user input sequence into a fixed-length context vector. The decoder also uses a multi-layer LSTM network to generate a reply sequence based on the context vector. The attention mechanism is used to calculate the correlation between each hidden state in the encoder and the current hidden state of the decoder, and dynamically adjust the weight of the context vector. Step 2: During the Seq2Seq model training phase, the cross entropy loss function is used to measure the difference between the generated response and the true response; for a response sequence of length T, the cross entropy loss function is defined as: , in, represents the true label of the tth word in the reply sequence, represents the generated sequence of the first t-1 words, x represents the user input sequence, It means that given the user input and the previous generated sequence, the tth word is probability; Step 3: Based on the Seq2Seq model, a reinforcement learning framework is introduced for further optimization; a reward function R is defined to give rewards based on the naturalness, relevance, diversity, and user satisfaction of the generated reply sequence; and a policy gradient method is used to optimize the model parameters θ to maximize the expected reward.

[0012] Preferably, the step of optimizing the dialogue generation model using a reinforcement learning framework includes: S1: Use supervised learning methods to pre-train the Seq2Seq model and obtain the initial parameters θ; S2: Build a simulated user environment and conduct multiple rounds of dialogue simulation with the user model; in each round of dialogue, generate a response sequence based on the user input and the current model parameters; S3: Design a reward function R that comprehensively considers the naturalness, relevance, diversity, and simulated user feedback of the response; naturalness is measured by the language model score, relevance is calculated by the match with the preset answer, and diversity is evaluated by calculating the entropy index of the generated response; S4: Based on the generated response sequence and the corresponding reward value, the policy gradient method is used to update the model parameter θ; the REINFORCE algorithm is used, and the gradient calculation formula is: , in, represents the expected operator, which means taking the expected value of the internal expression under given model parameters θ; represents the reward value, which is the comprehensive reward calculated based on the generated reply sequence and the simulated user’s feedback; The conditional probability of the model parameters θ on the generated response sequence Y The gradient of Approximate the expectation through Monte Carlo sampling, use the generated response sequence and reward value to calculate the gradient and update the model parameters; S5: Repeat steps S2 to S4 until the model performance converges or reaches the preset number of training rounds.

[0013] Preferably, the reward function R in the reinforcement learning training is designed as follows: Naturalness reward: Use the pre-trained language model to score the generated response sequence. The higher the score, the greater the reward. The inverse of the perplexity is used as the scoring indicator. The lower the perplexity, the more natural the response. Relevance reward: Calculate the matching degree between the generated response sequence and the preset answer or user intention. The higher the matching degree, the greater the reward. The cosine similarity indicator is used to measure the matching degree. Diversity reward: Encourage the model to generate diverse and novel responses, avoid repetitive and templated responses; calculate the entropy index of the generated responses, the greater the entropy, the greater the reward; User satisfaction rewards: In simulated conversations or real user interactions, collect user satisfaction feedback on generated responses. The higher the satisfaction, the greater the reward. User satisfaction is measured using user ratings, number of likes, or interaction duration. Comprehensive reward: Take the weighted sum of the above rewards to get the comprehensive reward value R.

[0014] Preferably, the image rendering model also includes a style transfer function for adjusting the appearance and animation style of the virtual digital human according to the corporate brand style.

[0015] Preferably, the multi-channel publishing and management unit includes an API interface docking function for integration with a third-party social platform and an enterprise's internal system.

[0016] Compared with the prior art, the present invention has the following beneficial effects: The system adopts blockchain storage units and utilizes the decentralized, tamper-proof and encrypted characteristics of blockchain technology to ensure the safe storage of corporate promotional materials, user interaction data, and virtual digital human learning and processing records. It not only effectively prevents data leakage and tampering risks, but also improves data reliability and traceability, providing solid data protection for corporate promotion.

[0017] Through the deep learning model construction unit, the system has built a deep learning model system including language understanding model, dialogue generation model and image rendering model. These models can continuously learn user needs and feedback, achieve self-optimization and intelligent upgrade, thereby significantly improving the intelligence level and interactive ability of virtual digital people, enabling virtual digital people to more accurately understand user intentions, generate coherent and targeted reply texts, and show realistic dynamic images and expression animations, greatly enhancing the user's interactive experience and immersion.

[0018] The image rendering unit uses rendering technology to dynamically generate realistic images and facial expressions of virtual digital people according to the content and scene of the dialogue, making the virtual digital people more vivid and realistic in visual performance. At the same time, the dialogue management unit combines context understanding and context perception capabilities to generate more coherent and natural reply texts, effectively solving the technical difficulties of traditional virtual digital people in image rendering and dialogue generation.

[0019] The multi-channel publishing and management unit enables the interactive content of virtual digital people to be easily published on multiple social platforms and corporate official websites, while ensuring content updates and consistency management on different channels. This not only improves the coverage and influence of corporate publicity, but also reduces management costs and workload, achieving efficient and convenient publicity effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A working principle diagram of the enterprise promotion virtual digital human system based on blockchain and deep learning according to the present invention; Figure 2 Flowchart for building a language understanding model; Figure 3 Flowchart for building a model for conversation generation. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] See also Figure 1-3 The present invention provides a technical solution: a virtual digital human system for enterprise publicity based on blockchain and deep learning, the system comprising: User interaction interface unit: This unit is designed as a graphical user interface (GUI) or web interface, which can receive corporate promotion-related requests entered by users through keyboard, mouse, touch screen or voice, and display the interactive responses of virtual digital people in real time, including text, voice, video and other forms, to ensure smooth communication between users and virtual digital people.

[0023] Blockchain storage unit: It is built with distributed ledger technology and is used to securely store corporate promotional materials (such as corporate profiles, product introductions, news updates, etc.), user interaction data (such as user requests, conversation records, etc.), and virtual digital human learning and processing records. Encryption algorithms and consensus mechanisms are used to ensure data security and immutability, while enabling efficient access and management of data.

[0024] Deep learning model construction unit: Use the deep learning framework to build and train the core models of virtual digital humans, including language understanding model, dialogue generation model and image rendering model. The language understanding model is used to parse user input, extract intent and key information; the dialogue generation model is based on the Seq2Seq architecture and reinforcement learning technology to generate natural, coherent reply text that meets the needs of corporate publicity; the image rendering model uses computer graphics technology to generate realistic virtual digital human images and expression animations.

[0025] Enterprise information extraction unit: closely connected with the blockchain storage unit, extracting and updating enterprise promotional materials regularly or according to demand. Interacting with the blockchain storage unit through smart contracts or API interfaces to ensure the accuracy and timeliness of information.

[0026] User request processing unit: receives user requests from the user interaction interface unit, performs preprocessing and parsing, and converts unstructured user input into a structured data format to facilitate subsequent model processing and analysis.

[0027] Dialogue management unit: As the core of the system, it connects the deep learning model building unit and the user request processing unit. According to the user request and the current dialogue context, it calls the dialogue generation model to generate appropriate reply text and controls the dialogue process to ensure the continuity and pertinence of the dialogue.

[0028] Image rendering unit: Based on the content of the conversation and the scene requirements, the image rendering model is called to generate the dynamic image and expression animation of the virtual digital human. Real-time rendering technology is used to ensure that the image of the virtual digital human is updated synchronously with the content of the conversation, enhancing the user's immersion and interactivity.

[0029] Blockchain verification unit: Blockchain verification of user interaction data to ensure the authenticity and integrity of the data. Through digital signatures, hash algorithms and other technical means, user requests, conversation records and other data are verified, and the verified data is recorded in the blockchain storage unit.

[0030] Feedback optimization unit: collects user feedback and interaction data, and uses data analysis technology to evaluate and analyze model performance. Based on the analysis results, the deep learning model is regularly tuned and updated to improve the model's intelligence level and user satisfaction.

[0031] Multi-channel publishing and management unit: publish the interactive content of virtual digital people to multiple social platforms and corporate official websites to achieve multi-channel publicity. At the same time, manage the content updates and consistency of different channels to ensure the uniformity and accuracy of corporate publicity information.

[0032] This invention integrates blockchain technology and deep learning technology to build a safe, efficient and intelligent virtual digital human system for corporate publicity. This system can not only effectively solve the problem of safe storage and effective management of corporate publicity materials, but also significantly improve the intelligence level and interactive experience of virtual digital humans, bringing a new solution to corporate publicity.

[0033] The present invention will be further described below in conjunction with Examples 1 to 4: Embodiment 1:

[0034] In the deep learning model building unit, the language understanding model is built based on the Transformer architecture and incorporates the attention mechanism to improve the model's ability to understand corporate promotional text data. The following are the detailed steps and implementation methods for training the language understanding model: Step A: Data collection and preprocessing Data collection: Collect text data related to corporate publicity from multiple sources, including but not limited to corporate official website content, product manuals, press releases, etc. These data should cover all aspects of the company to ensure the comprehensiveness and accuracy of the model.

[0035] Data cleaning: Clean the collected data, remove irrelevant information (such as advertisements, navigation bar content, etc.), correct spelling errors, unify text formats (such as converting full-width characters to half-width characters, unifying date formats, etc.), and filter out low-quality texts (such as texts that are too short or repetitive).

[0036] Text enhancement: In order to increase the diversity of training data, the cleaned text is subjected to synonym replacement, sentence transformation and other operations. For example, "the enterprise was established in 2000" is transformed into "the enterprise was founded in 2000", or "the product performance is excellent" is replaced into "the product performs well".

[0037] Step B: Text feature extraction Word segmentation: Use word segmentation tools (such as Jieba) to segment the preprocessed text and divide the continuous text into independent words or phrases.

[0038] Part-of-speech tagging: Label each word with its part of speech (such as noun, verb, adjective, etc.) so that the model can better understand the structure and semantics of the text.

[0039] Named Entity Recognition: Identify named entities in text (such as company names, product names, personal names, etc.) and assign specific labels to them so that the model can accurately recognize and process this key information.

[0040] Constructing text feature vector: Convert the results of word segmentation, part-of-speech tagging and named entity recognition into numerical feature vectors as the input of the model. Text feature vector is the process of converting text into numerical representation after word segmentation, part-of-speech tagging and named entity recognition. Specifically, for each word or phrase, the word embedding technology Word2Vec can be used to convert it into a vector of fixed dimension. Let the text feature vector be ,in represents the word embedding vector of the i-th word, and n is the total number of words in the text.

[0041] Step C: Model construction Input layer: Receives the text feature vector as input and passes it to the encoding layer.

[0042] Encoding layer: It consists of multiple layers of Transformer encoders, each of which contains a self-attention mechanism and a feedforward neural network. The self-attention mechanism can capture the relationship between words in the text sequence, and the feedforward neural network performs further nonlinear transformation on the output of the self-attention. The formula of the feedforward neural network is as follows: , in, , is the weight matrix, , is the bias vector, Represents the ReLU activation function.

[0043] Attention layer: Additional attention mechanisms, such as multi-head attention mechanisms, are introduced based on the encoding layer to capture more complex relationships and dependencies within the text.

[0044] Output layer: Based on the processing results of the encoding layer and the attention layer, the output is the understanding representation of the text. This can be the classification label, sentiment tendency, key information extraction results, etc. of the text.

[0045] Step D: Model training Use the data set processed in step A to train the model. Use the cross entropy loss function as the optimization target and update the model parameters through the back propagation algorithm. During the training process, you can use optimization algorithms such as stochastic gradient descent (SGD) and Adam to accelerate convergence and improve training efficiency.

[0046] Step E: Model Evaluation Evaluate the performance of the model on the validation set. Select accuracy and F1 value as evaluation indicators to measure the model's ability to understand corporate promotional text data. Accuracy reflects the proportion of correct predictions made by the model, while F1 value is the harmonic mean of precision and recall, which can more comprehensively evaluate the effectiveness of the model.

[0047] Step F: Model tuning and iteration According to the evaluation results, adjust the structure or hyperparameters of the model. For example, you can increase or decrease the number of layers of the Transformer encoder, adjust the learning rate, change the type of attention mechanism, etc. Then repeat steps C to E to iteratively train and optimize the model until the model performance meets the requirements. Embodiment 2:

[0048] This embodiment is used to implement a dialogue generation model based on a sequence-to-sequence (Seq2Seq) architecture combined with reinforcement learning. The specific implementation method is as follows: Step 1: Model construction: The dialogue generation model mainly consists of three parts: encoder, decoder and attention mechanism.

[0049] Encoder: Use a multi-layer long short-term memory network (LSTM) to encode the user input sequence. Suppose the user input sequence is ,in Represents the i-th word in the input sequence. The encoder processes the input sequence word by word through a multi-layer LSTM network, and uses the hidden state of the last LSTM unit as a fixed-length context vector C, which contains all the information of the user input sequence.

[0050] Decoder: It also uses a multi-layer LSTM network to generate a response sequence based on the context vector C. During the decoding process, the initial hidden state of the decoder LSTM is set to the context vector C. Subsequently, the decoder generates a response sequence word by word ,in Represents the jth word in the reply sequence. The generation of each word depends on the generation result of the previous word and the context vector C.

[0051] Attention mechanism: In order to dynamically adjust the weight of the context vector, an attention mechanism is introduced. When , the attention mechanism evaluates the relevance of each hidden state in the encoder to the current hidden state of the decoder, generating a weighted context vector In this way, the generation of each reply word can more accurately focus on the relevant information in the user input sequence.

[0052] Step 2: Seq2Seq model training: During the training phase of the Seq2Seq model, the cross entropy loss function is used to measure the difference between the generated response and the real response. For a response sequence of length T, the cross entropy loss function is defined as follows: , in, represents the true label of the tth word in the reply sequence, represents the generated sequence of the first t-1 words, x represents the user input sequence, It means that given the user input and the previous generated sequence, the tth word is By minimizing the cross entropy loss function, a Seq2Seq model that can generate reasonable responses can be initially trained.

[0053] Step 3: Reinforcement learning optimization: In order to further improve the performance of the dialogue generation model, a reinforcement learning framework is introduced for further optimization.

[0054] A comprehensive reward function R is defined based on the multi-dimensional indicators of the generated response sequence, such as naturalness, relevance, diversity, and user satisfaction. Naturalness measures the language fluency and grammatical correctness of the response; relevance evaluates the degree of association between the response and the user input; diversity reflects the richness and innovation of the response; and user satisfaction is obtained through user feedback or online evaluation.

[0055] The policy gradient method is used to treat the model parameters θ as policy parameters, and these parameters are optimized to maximize the expected reward. Specifically, for each generated response sequence, its reward value R is calculated, and the model parameters θ are updated according to the reward value. The update rule is usually based on the gradient ascent method, that is, the parameters are adjusted in the direction of increasing rewards. Embodiment 3:

[0056] This embodiment is used to implement a dialogue generation model based on reinforcement learning optimization. The specific implementation method is described in detail as follows: S1 supervised learning pre-training: First, the Seq2Seq model is pre-trained using supervised learning methods to obtain the initial parameters θ. In this stage, a large-scale conversation dataset is used to train the model by minimizing the cross entropy loss function, so that it can initially understand user input and generate reasonable responses. After pre-training, the model has basic conversation generation capabilities.

[0057] S2 builds a simulated user environment: In order to simulate real conversation scenarios, a simulated user environment is constructed. This environment can simulate the user's input behavior and conduct multiple rounds of conversation simulation with the conversation generation model. In each round of conversation, the simulated user generates user input according to the preset conversation process or random strategy, and the model generates a reply sequence according to the current parameter θ.

[0058] S3 Design reward function: The reward function R is the key to reinforcement learning optimization. It takes into account the naturalness, relevance, diversity and simulated user feedback of the response. The specific design method is as follows: Naturalness reward: Use pre-trained language models (such as GPT, BERT, etc.) to score the generated response sequence. The higher the score, the more natural the response, and the greater the reward. Specifically, the inverse of the perplexity is used as the scoring indicator, and the lower the perplexity, the higher the score.

[0059] Relevance reward: Calculate the matching degree of the generated reply sequence with the preset answer or user intent. The cosine similarity metric is used to measure, and the higher the matching degree, the greater the reward. This can be achieved by calculating the cosine value between the reply vector and the answer vector or user intent vector.

[0060] Diversity Reward: To encourage the model to generate diverse and novel responses, we calculate the entropy of the responses. The greater the entropy, the more diverse the responses, and the greater the reward. This can be calculated by counting the distribution of different words or phrases in the responses.

[0061] User satisfaction reward: In simulated conversations or real user interactions, collect user satisfaction feedback on generated responses. The higher the satisfaction, the greater the reward. User satisfaction can be measured using indicators such as user ratings, number of likes, and interaction time.

[0062] Comprehensive reward: The above rewards are weighted and summed to obtain the comprehensive reward value R. The weight can be adjusted according to actual needs and experimental results to balance the importance of each reward.

[0063] S4 policy gradient updates model parameters: Use the policy gradient method to update the model parameters θ to maximize the expected reward. The REINFORCE algorithm is used, and its gradient calculation formula is: , in, represents the expected operator, which means taking the expected value of the internal expression under given model parameters θ; represents the reward value, which is the comprehensive reward calculated based on the generated reply sequence and the simulated user’s feedback; The conditional probability of the model parameters θ on the generated response sequence Y The gradient of . Monte Carlo sampling approximates the expectation, uses the generated reply sequence and reward value to calculate the gradient, and updates the model parameters. Specifically, after each round of dialogue simulation, the gradient is calculated and the model parameters are updated based on the generated reply sequence and the corresponding reward value.

[0064] S5 Iteration Training: Repeat steps 2 to 4 until the model performance converges or the preset number of training rounds is reached. During the iterative training process, the model gradually learns to adjust the strategy of generating responses according to the reward function to generate more natural, relevant, diverse and user-friendly responses. Embodiment 4:

[0065] This embodiment is used to implement an image rendering model with a style transfer function and a multi-channel publishing and management unit that supports API interface docking. The specific implementation method is as follows: 1. Style transfer function in image rendering model Style feature extraction: Build or select a pre-trained style feature extraction network that can automatically extract key style features from a given corporate brand style image, such as color matching, texture pattern, shape design, etc. These style features will be encoded into vector form to facilitate subsequent integration with the basic appearance and animation style of the virtual digital human.

[0066] Design a parameterizable basic model of virtual digital human, which includes basic elements such as bone structure, muscle form, facial expression, etc., and has basic animation generation capabilities.

[0067] Develop a style fusion module that receives the style feature vector extracted from the brand style image and the basic model parameters of the virtual digital human, and fuses the style features into the basic model through a deep neural network or other machine learning algorithm to generate a virtual digital human image with a specific brand style. Integrate efficient real-time rendering engines such as Unity and Unreal Engine to support the real-time display of virtual digital humans in different scenarios.

[0068] Develop a user-friendly interface that allows users to adjust the appearance and animation style of virtual people according to actual needs, such as changing clothes, adjusting skin color, changing movement rhythm, etc., while maintaining consistency with the corporate brand style.

[0069] Implement a style consistency verification mechanism to ensure the accuracy and consistency of style transfer by comparing the similarity between the rendered virtual digital human image and the original brand style image.

[0070] If the similarity is lower than the preset threshold, adjustments will be made automatically or the user will be prompted to make adjustments until a satisfactory effect is achieved.

[0071] 2. API interface connection function in multi-channel publishing and management unit Design a set of standardized and easily extensible API interfaces for integration with third-party social platforms (such as WeChat, Weibo, Douyin, etc.) and internal enterprise systems (such as CRM, ERP, etc.).

[0072] The API interface should support multiple functions such as data query, data push, user authentication, and permission management to meet the integration requirements of different platforms. Develop a data exchange module that can automatically exchange and synchronize virtual digital human rendering results, animation data, user interaction data, etc. with third-party platforms through the API interface. Ensure the security, integrity, and real-time performance of the data exchange process, and use technical means such as encrypted transmission, data verification, and asynchronous processing to improve the reliability of data exchange. Develop corresponding adaptation modules for different third-party social platforms to support the display and interaction of virtual digital humans on different platforms.

[0073] The one-click publishing function is realized. Users only need to select the target platform in the management system to quickly publish the virtual digital human and its related content to the designated platform.

[0074] Permission management and user authentication: An integrated rights management system is used to assign different access rights and operation permissions to different users to ensure data security and privacy.

[0075] Implement user authentication mechanisms such as OAuth, JWT, etc. to support seamless login and authorization of third-party platform users.

[0076] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0077] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. The enterprise publicity virtual digital human system based on blockchain and deep learning is characterized by: The system comprises: A user interaction interface unit, used to receive enterprise promotion related requests input by users and display interactive responses of virtual digital people; Blockchain storage unit, used to securely store corporate promotional materials, user interaction data, and virtual digital human learning and processing records; A deep learning model building unit, which is used to build and train the deep learning model of virtual digital humans, including language understanding models, dialogue generation models, and image rendering models; An enterprise information extraction unit, connected to the blockchain storage unit, for extracting and updating enterprise promotional materials, including enterprise profiles, product services, and news information; A user request processing unit, connected to the user interaction interface unit, for receiving and parsing user requests and converting the requests into structured data for subsequent model processing; A dialogue management unit, connected to the deep learning model construction unit and the user request processing unit, for invoking the dialogue generation model to generate an appropriate reply text based on the user request and the current dialogue context; An image rendering unit, connected to the deep learning model building unit, is used to call the image rendering model to generate a dynamic image and expression animation of a virtual digital human according to the content and scene of the conversation; A blockchain verification unit, used to perform blockchain verification on user interaction data and record the verified data into a blockchain storage unit; Feedback optimization unit, which is used to collect user feedback and interaction data, analyze model performance, and regularly tune and update deep learning models; The multi-channel publishing and management unit is used to publish the interactive content of virtual digital people to multiple social platforms and corporate websites, while managing the content updates and consistency of different channels.

2. The enterprise promotion virtual digital human system based on blockchain and deep learning according to claim 1 is characterized by: The language understanding model in the deep learning model building unit is built using the Transformer architecture and incorporates an attention mechanism.

3. The enterprise promotion virtual digital human system based on blockchain and deep learning according to claim 2 is characterized in that: The steps to train a language understanding model include: Step A: Collect text data related to corporate publicity, including corporate official website content, product manuals and press releases, and perform data cleaning and preprocessing; Step B: Perform word segmentation, part-of-speech tagging and named entity recognition on the preprocessed text data to construct a text feature vector; Step C: Construct a Transformer-based language understanding model, including input layer, encoding layer, attention layer and output layer; the input layer receives the text feature vector, the encoding layer processes the text sequence through a multi-layer Transformer encoder, the attention layer introduces a self-attention mechanism to capture the internal relationship of the text, and the output layer is used to output the understanding representation of the text; Step D: Use the data set processed in step A to train the model and use the cross entropy loss function to optimize the model parameters; Step E: Evaluate the model performance on the validation set, and measure the model effect through accuracy and F1 value indicators; Step F: Adjust the model structure or hyperparameters based on the evaluation results, and repeat steps C to E until the model performance meets the requirements.

4. The enterprise promotion virtual digital human system based on blockchain and deep learning according to claim 3 is characterized by: The data cleaning method in step A includes removing irrelevant information, correcting spelling errors, unifying text formats, and filtering low-quality texts.

5. The enterprise promotion virtual digital human system based on blockchain and deep learning according to claim 4 is characterized by: The preprocessing step in step A also includes text enhancement, which increases the diversity of training data through synonym replacement and sentence transformation.

6. The enterprise promotion virtual digital human system based on blockchain and deep learning according to claim 5 is characterized in that: The dialogue generation model adopts a sequence-to-sequence Seq2Seq architecture and is trained in combination with reinforcement learning. The specific implementation methods include: Step 1: The dialogue generation model consists of an encoder, a decoder, and an attention mechanism. The encoder uses a multi-layer LSTM network to encode the user input sequence into a fixed-length context vector. The decoder also uses a multi-layer LSTM network to generate a reply sequence based on the context vector. The attention mechanism is used to calculate the correlation between each hidden state in the encoder and the current hidden state of the decoder, and dynamically adjust the weight of the context vector. Step 2: During the Seq2Seq model training phase, the cross entropy loss function is used to measure the difference between the generated response and the true response; for a response sequence of length T, the cross entropy loss function is defined as: ， in, represents the true label of the tth word in the reply sequence, represents the generated sequence of the first t-1 words, x represents the user input sequence, It means that given the user input and the previous generated sequence, the tth word is probability; Step 3: Based on the Seq2Seq model, a reinforcement learning framework is used for further optimization. A reward function R is defined to give rewards based on the naturalness, relevance, diversity, and user satisfaction of the generated reply sequence. The policy gradient method is used to optimize the model parameters θ to maximize the expected reward.

7. The enterprise promotion virtual digital human system based on blockchain and deep learning according to claim 6 is characterized in that: The steps of optimizing the dialogue generation model using the reinforcement learning framework include: S1: Use supervised learning methods to pre-train the Seq2Seq model and obtain the initial parameters θ; S2: Build a simulated user environment and conduct multiple rounds of dialogue simulation with the user model; in each round of dialogue, generate a response sequence based on the user input and the current model parameters; S3: Design a reward function R that comprehensively considers the naturalness, relevance, diversity, and simulated user feedback of the response; naturalness is measured by the language model score, relevance is calculated by the match with the preset answer, and diversity is evaluated by calculating the entropy index of the generated response; S4: Based on the generated response sequence and the corresponding reward value, the policy gradient method is used to update the model parameter θ; the REINFORCE algorithm is used, and the gradient calculation formula is: ， in, represents the expected operator, which means taking the expected value of the internal expression under given model parameters θ; represents the reward value, which is the comprehensive reward calculated based on the generated reply sequence and the simulated user’s feedback; The conditional probability of the model parameters θ on the generated response sequence Y The gradient of Approximate the expectation through Monte Carlo sampling, use the generated response sequence and reward value to calculate the gradient and update the model parameters; S5: Repeat steps S2 to S4 until the model performance converges or reaches the preset number of training rounds.

8. The enterprise promotion virtual digital human system based on blockchain and deep learning according to claim 7 is characterized in that: The design method of the reward function R in the reinforcement learning training includes: Naturalness reward: Use the pre-trained language model to score the generated response sequence. The higher the score, the greater the reward. The inverse of the perplexity is used as the scoring indicator. The lower the perplexity, the more natural the response. Relevance reward: Calculate the matching degree between the generated response sequence and the preset answer or user intention. The higher the matching degree, the greater the reward. The cosine similarity indicator is used to measure the matching degree. Diversity reward: Encourage the model to generate diverse and novel responses, avoid repetitive and templated responses; calculate the entropy index of the generated responses, the greater the entropy, the greater the reward; User satisfaction rewards: In simulated conversations or real user interactions, collect user satisfaction feedback on generated responses. The higher the satisfaction, the greater the reward. User satisfaction is measured using user ratings, number of likes, or interaction duration. Comprehensive reward: Take the weighted sum of the above rewards to get the comprehensive reward value R.

9. The enterprise promotion virtual digital human system based on blockchain and deep learning according to claim 1 is characterized by: The image rendering model also includes a style transfer function for adjusting the appearance and animation style of the virtual digital human according to the corporate brand style.

10. The enterprise promotion virtual digital human system based on blockchain and deep learning according to claim 1 is characterized by: The multi-channel publishing and management unit includes an API interface docking function for integration with third-party social platforms and internal enterprise systems.

Images