Man-machine interaction method and system based on AI large model

Abstract

The embodiment of the invention provides a man-machine interaction method and system based on an AI large model. The man-machine interaction method comprises the steps that the AI large model obtained through training based on a preset algorithm learns interaction behaviors of a user to obtain a personalized response rule; on the premise of ensuring data desensitization processing, obtaining user interaction data for optimization and iteration of the model; differentiated feedback information for different users is dynamically generated based on the personalized response rule so as to enhance the user experience; and the strategy weight of the AI large model is adjusted according to the feedback, so that the personalized demand of user experience is balanced, and privacy protection is realized. Through the scheme of the embodiment of the invention, the problem of how to solve the contradiction between user experience individuation and privacy protection in man-machine interaction can be solved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and specifically to a human-computer interaction method and system based on an AI big model. Background Art

[0002] The human-computer interaction method and system based on AI big model is a technical solution that uses advanced artificial intelligence technology and large-scale data sets to enhance the interactive experience between humans and computers. This method usually involves deep learning algorithms, which enable computers to adaptively improve their response methods by learning user preferences, habits and behavior patterns, providing a more personalized and natural interactive interface.

[0003] However, in the process of improving the personalization of the user experience of human-computer interaction, there is an important challenge, that is, how to ensure personalization without infringing on the privacy of users. This contradiction is reflected in the fact that in order to provide more accurate and personalized services, the system often needs to collect and process a large amount of user personal information and usage behavior data; this not only involves data security issues, but also raises concerns about privacy infringement and social ethics. In order to resolve this contradiction, research and practice often need to adopt a combination of multiple technologies and strategies, such as using data encryption technology, anonymization processing, and strengthening user control over their own data, so as to optimize the user experience while protecting the user's personal privacy rights from infringement. Summary of the invention

[0004] In view of this, the embodiments of the present disclosure provide a human-computer interaction method and system based on an AI big model, which at least partially solves the problems existing in the prior art.

[0005] Human-computer interaction method and system based on AI big model, including:

[0006] The AI ​​big model trained based on the preset algorithm learns the user's interactive behavior to obtain personalized response rules;

[0007] Obtain user interaction data for model optimization and iteration while ensuring data anonymization;

[0008] Dynamically generate differentiated feedback information for different users based on personalized response rules to enhance user experience;

[0009] Adjust the strategy weight of the AI ​​model based on feedback to balance the personalized needs of user experience and privacy protection.

[0010] In a specific embodiment, the steps specifically include:

[0011] Collect user interaction logs as original training samples;

[0012] Performing feature engineering processing on the original training samples to extract effective features;

[0013] Use preset algorithms and processed samples to conduct initial training of the AI ​​model;

[0014] Analyze the model's performance indicators based on preliminary training results and make necessary fine-tuning;

[0015] The deviation between the model-predicted response and the actual user feedback is evaluated to form personalized response rules.

[0016] In a specific embodiment, the steps specifically include:

[0017] Use de-identification technology to remove or replace personally identifiable information;

[0018] Apply data desensitization algorithms to blur sensitive fields by adding noise, etc.

[0019] Regularly update desensitization strategies to ensure data security;

[0020] Establish access control and monitoring mechanisms to protect the security of processed user data;

[0021] The processed user interaction data is input into the big AI model as a data source for retraining.

[0022] In a specific embodiment, the steps specifically include:

[0023] Capture the contextual information of the current user interaction state in real time;

[0024] Call personalized response rules for intelligent matching and generate basic response message templates;

[0025] Compare multiple possible response messages and select the optimal solution based on the user's historical preferences;

[0026] Dynamically adjust the response speed to keep the interaction rhythm in sync with user expectations;

[0027] Combine user feedback to further improve the content and presentation of response information.

[0028] In a specific embodiment, the steps specifically include:

[0029] Establish a strategy weight update mechanism to adapt to different feedbacks;

[0030] Quantitatively evaluate user satisfaction to determine positive and negative reward values;

[0031] Use reinforcement learning techniques to adjust model parameters to achieve a better balance;

[0032] Introduce expert rules to avoid excessive bias towards personal preferences and ignoring privacy and security risks;

[0033] Monitor the impact of the strategy update process on the overall model effect to ensure stable operation.

[0034] In a specific implementation, the AI ​​big model obtained by training based on the preset algorithm learns the user's interactive behavior to obtain personalized response rules, including:

[0035] Build training datasets to cover multiple types of user interaction histories;

[0036] Set the algorithm initialization parameters and model architecture and train the large model based on the above training data set;

[0037] Adjust the number of model iterations according to the interaction effect until it converges to the predetermined accuracy;

[0038] An offline test set is used to verify the effectiveness of the personalized response rules obtained during the learning process.

[0039] In a specific embodiment, setting algorithm initialization parameters and model architecture and training a large model based on the above training data set includes:

[0040] Determine algorithm hyperparameters such as learning rate and batch size as initialization configuration;

[0041] Use backpropagation to calculate the derivative of each parameter and adjust its value to optimize the cost function (J);

[0042] Once the change in the cost function (J) is less than a predefined small value, it is considered converged and the training is stopped;

[0043] Choose the appropriate number of hidden layers and neuron nodes to meet the training requirements.

[0044] In a specific embodiment, using back propagation to calculate the derivative of each parameter and then adjust its value to optimize the cost function (J) includes:

[0045] Calculate the cost difference (C) of the output layer relative to the true value of the label;

[0046] Back-step through each layer of network nodes to find partial derivatives to quantify the direction and magnitude of each update;

[0047] If (C < threshold C_min), the cost is considered to have been sufficiently reduced, and the next round of training will continue until it no longer decreases or meets the preset stopping point;

[0048] The weights of nodes at each level are updated through gradient (where (α) is a learning rate fixed in advance).

[0049] In a specific embodiment, if (C<threshold C_min), it is considered that the cost has been sufficiently reduced, and continuing to the next round of training until it no longer decreases or meets the preset stopping point includes:

[0050] When the difference between the current cost (C) and the previous iteration (△(C)) is less than the predefined ratio (r%, rge0);

[0051] Check whether the specified number of training rounds has been reached or the model performance on the validation set is no longer significantly improved;

[0052] If any of the conditions are met, this round of learning cycle ends and the testing and deployment preparation phase begins;

[0053] If any condition is not met, the current learning state is continued for further optimization (C).

[0054] In a specific implementation, when the difference of the current cost (C) drop compared to the previous iteration (△(C)) is less than a predefined ratio (r%, rge0) includes:

[0055] Obtain the difference between two adjacent sets of (C) data (|C^{t}C^{t1}|);

[0056] Compare this difference with (|C^{t}|cdot(r / 100));

[0057] If the former is less than the latter, it indicates that the model is approaching a stable state or a minimum point, thus deciding to terminate the iteration operation; otherwise, the iteration is continued according to the existing trend;

[0058] During the model training process, the (|C^{t}|cdot(r / 100)) value is updated in real time to track the changes and ensure that the convergence point is detected in time.

[0059] The disclosed embodiment provides a human-computer interaction method and system based on an AI big model, including: learning the user's interactive behavior based on the AI ​​big model trained by a preset algorithm to obtain personalized response rules; obtaining user interaction data for model optimization and iteration under the premise of ensuring data desensitization; dynamically generating differentiated feedback information for different users based on personalized response rules to enhance user experience; adjusting the strategy weight of the AI ​​big model based on feedback to achieve a balance between personalized user experience needs and privacy protection. Through the solution of the disclosed embodiment, it is possible to solve the contradiction between personalized user experience and privacy protection in human-computer interaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In the accompanying drawings, unless otherwise specified, the same reference numerals throughout the multiple drawings represent the same or similar parts or elements. These drawings are not necessarily drawn to scale. It should be understood that these drawings only depict some embodiments disclosed in the present application and should not be regarded as limiting the scope of the present application.

[0061] Figure 1 It is a flow chart of the human-computer interaction method and system based on the AI ​​big model;

[0062] Figure 2 Yes, a detailed flowchart of the steps;

[0063] Figure 3 Yes, a detailed flowchart of the steps;

[0064] Figure 4 Yes, a detailed flowchart of the steps;

[0065] Figure 5 Yes, a detailed flowchart of the steps;

[0066] Figure 6 It is a flowchart of how the AI ​​big model trained by the preset algorithm learns the user's interactive behavior to obtain personalized response rules;

[0067] Figure 7 It is a flowchart for setting the algorithm initialization parameters and model architecture and training the large model based on the above training data set;

[0068] Figure 8 It is a flowchart of using back propagation to calculate the derivative of each parameter and then adjust its value to optimize the cost function (J);

[0069] Fig. 9 If (C < threshold C_min), the cost is considered to have been sufficiently reduced, and the next round of training is continued until it no longer decreases or meets the preset stopping point flowchart;

[0070] Fig.10 It is a flow chart in which the difference of the current cost (C) drop compared with the previous iteration (△(C)) is less than a predefined ratio (r%, rge0). DETAILED DESCRIPTION

[0071] In order to make the purpose, technical solutions and advantages of the embodiments of the present disclosure more clear, the embodiments of the present disclosure are further described in detail in combination with the embodiments and drawings. The schematic implementation methods of the embodiments of the present disclosure and their descriptions are only used to explain the embodiments of the present disclosure and are not intended to limit the embodiments of the present disclosure.

[0072] In order to explore this technical system in depth, we first need to start from the overall architecture and understand the four key steps of the system and its working principles. These steps include using the AI ​​big model trained by the preset algorithm to learn the user's interactive behavior to obtain personalized response rules; obtaining the user's interaction data to optimize and upgrade the big model without exposing sensitive information; using the above-obtained rules to generate customized feedback information to improve the user's experience level; and adjusting the decision-making trade-off parameters of the AI ​​big model according to the recipient's response to ensure that the specific preferences of individual users are met and the dual requirements of protecting their personal privacy are protected.

[0073] For example, suppose there is an online shopping app that wants to improve the quality of its interactions with users - to increase user satisfaction by better adapting to each user's unique shopping preferences. The big model in the system is initially a pre-trained basic model with basic functions; it will then be given more complex features - learning user behavior patterns and generating customized response rule sets to respond to user queries faster and more usefully in the future. In this example, we can assume that the big model first gained some initial understanding and skills from the learning task of recording tens of millions of historical purchase records and related evaluation information.

[0074] Next, we need to enable this trained model to obtain new user interaction data to further optimize itself while ensuring data anonymization. The key point here is to utilize the value of data while maintaining personal identity from any risk of leakage. Back to the example of the shopping app, we collect behavioral data such as the length of time users browse product pages, click behavior preferences such as viewing product images or adding directly to the shopping cart, and search terms entered by users. These data will first be removed from specific information that can be directly linked to the user before being used for model analysis.

[0075] With the personalized response criteria trained on the above user behavior patterns, the next step is to generate specific feedback messages for each specific and differentiated user based on this rule. For example, if a consumer has recently frequently browsed products related to baby clothing, then according to the personalized response mechanism, this APP may display recommended products or discount information related to baby clothing, or send a questionnaire request to the user to collect more information to further improve the accuracy of the recommendation results, thereby creating a more intimate service environment for this special user and significantly improving the user experience. Another user may only be interested in new technology products, and will receive the latest technology product recommendations or discount codes automatically provided from within the application, which also achieves accurate service push effects.

[0076] As the AI ​​system continues to extract value from user interactions, the final stage is to adjust decision-making trade-offs based on user responses. In the application scenario described above, suppose a user responds positively to the content pushed to him, such as a high click-through rate on the recommended baby clothing or even a successful purchase. At this time, the application will further increase the proportion of the baby clothing category in the list of recommended content for this user in the future based on such positive user feedback to better serve this type of target customers. On the contrary, when encountering situations where there is no interest in a certain topic (for example, a customer has deleted the information flow tweets about maternal and child products many times), the AI ​​will reduce the chances of such products appearing to avoid boredom. Such a dynamic policy adjustment mechanism makes the application more accurate and humane, while also avoiding infringing on the personal privacy rights of customers in the process of obtaining and using their sensitive personal information.

[0077] In general, the "human-computer interaction method and system based on AI large model" effectively improves user satisfaction and ensures data security by continuously collecting learning, iteratively updating, and providing personalized feedback content. Through strict management and implementation of each link, the best balance between meeting personalized needs and protecting privacy goals is finally achieved.

[0078] Next, when describing “the specific steps”, we must first understand that these five steps are the key components of the entire human-computer interaction method based on the AI ​​big model.

[0079] The step of collecting user interaction logs as raw training samples refers to the collection of data records generated in past human-computer dialogues on the platform. For example, suppose a voice assistant interacts with users, asks questions, and responds to questions during daily operations. All these interactions will leave corresponding electronic logs, such as questions or instructions raised by users and the content of the system's response. These contents constitute a large amount of user interaction log data; the next step is to capture useful elements from the log information as raw materials for training large models. These data lay the foundation for later machine learning.

[0080] The next step is data preprocessing, which involves so-called feature engineering. The goal is to extract data features that reflect actual business conditions and are highly consistent with specific needs, and thereby enhance the model's learning ability. For example, the main keywords of the user's questions and context-related information are extracted from the log file to form a richer semantic expression. After a series of operations like this, the original simple conversation content will be transformed into a new sample containing more valuable feature information for the model to further learn.

[0081] Next, the initial machine learning work is performed on the preliminarily constructed samples using the preset algorithm. This step gradually establishes prediction rules or optimizes the original parameter configuration by allowing the AI ​​model to absorb and digest the information that has been sorted out. Here is an example to help illustrate: if the word vector technology is selected to calculate text similarity in the above work, then this step may introduce a deep neural network structure, input the key feature points obtained in each round into the model to try to simulate the human brain workflow, and gradually approach the ideal result by repeatedly adjusting the weight parameters.

[0082] After completing the above training activities, it is necessary to continue to quantify the generated model effects to clarify its performance. This means that the existing model effects need to be fully tested to see whether it accurately, timely and reliably achieves the established design indicators and user expectations. Taking the previous voice assistant project as an example, we can compare the differences in answers between the two situations when the user asks the same question (once directly giving correct feedback, and once indirectly guiding the answer with the help of a new model), and judge where there is room for improvement based on the results; then make appropriate parameter fine-tuning and other measures to ensure that the machine can be closer to the user's real thoughts.

[0083] Finally, when the model is initially stable, it is necessary to compare the consistency of each output result with the actual user expectations. That is, to evaluate whether the answer or action given by the system each time is in line with human habits, and update the corresponding response strategy accordingly; for example, after the above project runs for a certain period of time, if enough feedback evidence is collected to point out that in some cases, the assistant provides guidance information too quickly, making the other party unable to keep up with the thinking rhythm, etc., the answer strategy can be adjusted by adding appropriate pauses or tone assistance to make it more suitable for the majority of people to accept and further enhance the user experience value.

[0084] Next, describe the "specific steps":

[0085] Regarding the specific implementation steps in the claims, de-identification technology should first be used to remove or replace personal information to achieve basic privacy protection of data. For example, for a set of collected user voice records, before these voice samples are input into the big model, the system will automatically identify the real name, address and other information that may clearly point to the user's identity, and erase or replace them with other non-identifiable data content, ensuring that even direct observers of the data cannot trace the identity of the data source.

[0086] After the data is initially cleaned, data desensitization technology should be used to further process sensitive fields. For example, by adding noise or blurring fields such as the amount and mobile phone number in user feedback data. In this way, even malicious attempts to parse it cannot accurately extract the true value of this information. This measure is particularly applicable to chatbot systems in the field of financial consulting - while providing services to customers, their account details are kept safe.

[0087] In order to ensure that information security is permanently and reliably protected throughout the process, the desensitization rules must be regularly re-evaluated and updated to prevent potential new threats or associated risk points in the data over time. This means that changes in the external environment need to be constantly monitored, such as new regulations that require the strengthening of certain types of personal information protection, or the emergence of new technologies that may challenge existing desensitization technologies. For example, adding biometric data desensitization steps according to regulatory requirements, or updating denoising algorithms to improve data quality due to technology upgrades.

[0088] In addition to internal data preprocessing, it is also necessary to establish a comprehensive access control and monitoring strategy to ensure that the processed interactive information is not used without authorization and to avoid leaks from any link. For example, a multi-authority verification process can be established, and a multi-layer identity verification threshold can be set for the system that stores the pre-processed interactive information library; non-compliant behaviors can be identified through traffic monitoring and analysis, and channels can be cut off immediately when suspicious actions are found, and emergency plans can be initiated when necessary to reduce losses; third-party independent review services can also be introduced to ensure that there are no blind spots in the overall security architecture.

[0089] After completing all the above preparations, the last step is to input the desensitized, safer and more reliable user interaction data into the artificial intelligence model for training. For example, the interactive system integrated with the navigation and multimedia center in the smart car can obtain richer and more diverse driving situation feedback data through such a series of operations without compromising the confidentiality of passenger information, so as to continuously iterate and improve its own intelligence level and better assist car owners in dealing with various challenges during the journey.

[0090] Through such a progressive process from removing individual information identifiers, blurring high-risk details to establishing a comprehensive management and control framework, we can effectively balance the relationship between innovative exploration and legitimate rights and interests when building new human-computer interaction solutions based on big models, and ultimately promote the implementation of more humane and safer products.

[0091] Next, "the steps described are specific" are described.

[0092] Real-time capture of contextual information in the current user interaction state means that the system captures every interaction detail with the user through voice recognition, text analysis, etc., to ensure that the system can generate the most appropriate and coherent answer or behavior based on the current situation. For example, in the application scenario of a chatbot, if a user asks "What will the weather be like tomorrow?" the system needs to capture the time and place of the inquiry as contextual information for providing an answer.

[0093] Calling personalized response rules for intelligent matching and generating basic response message templates means comparing the currently captured information with a series of predefined rules to find the best answer for the current situation. In the weather question and answer example, the system will obtain the future weather forecast information for the corresponding location from the database or online API and then build an initial response draft. For example, it may build a basic template for a message like "Tomorrow in Beijing..." to prepare a response.

[0094] Compare multiple possible response messages and select the best solution based on the user's historical preferences. This item refers to comparing and evaluating the above basic template and other alternative solutions based on the user's past preferences. Continuing with the example of asking about the weather in Beijing, if the chat record shows that the user is more concerned about wind speed than the possibility of rain, the answer content that focuses more on introducing future wind speed forecasts will be selected.

[0095] Dynamically adjusting the response speed and keeping the interaction rhythm in sync with user expectations means that the system must adjust its own request processing speed to ensure that there are no dull moments or breakpoints in the communication process, thereby improving the user experience and engagement. If the chat history shows that the user tends to get a response faster, the system will try to speed up the response process for the next inquiry.

[0096] Combine user feedback to further improve the content and presentation of response information. This step emphasizes using the collected feedback to adjust the prepared replies or services in order to achieve better interaction results. For example, if the user has expressed through text comments that they prefer a direct answer to the question, the chatbot will simplify the additional content in the reply based on this information to convey only the core information points.

[0097] The above five steps, through continuous iteration and improvement, form an overall mechanism for more efficient and smoother human-computer interaction.

[0098] Next, describe the "specific steps":

[0099] Establishing a strategy weight update mechanism to adapt to different feedback means designing a method to enable AI in human-computer interaction to understand and adjust its behavior in a timely manner based on user feedback. For example, when this technology is applied to a recommendation engine app, when a user chooses to like or dislike a displayed product, the algorithm can adjust the weight based on this information and make recommendations that are more suitable for user needs in the future.

[0100] The purpose of quantitatively evaluating user satisfaction is to provide a numerical measure for user interaction, so that we can accurately know which decisions are affirmed by users and which decisions should be avoided. In the use scenario of an online customer service robot, by analyzing the emojis used by users after the conversation with the robot or the actions of clicking the rating button, the user's satisfaction with the experience can be scored. For example, a "thumbs up" is considered to have a high feedback value and represents a positive reward, while an "angry face" expression indicates a negative reward, which will affect the machine's learning path.

[0101] Using reinforcement learning technology to adjust model parameters means that the AI ​​model is continuously learning and optimizing through a series of trial and error and feedback loops. In the context of a voice-controlled home smart system, if the model is initially unable to accurately recognize the voice command "please close the window", it may try different processing logic such as changing the tone and speech speed recognition, and gradually improve its response until the success rate increases. This adjustment makes the control system of the smart home more adaptable to inputs of various accents, enables more efficient operation of functions, and improves user satisfaction.

[0102] The purpose of introducing expert rules is to prevent the learning process from making inappropriate decisions based entirely on the preferences of individual users, which may lead to ignoring other factors such as privacy leakage risks. Consider an algorithm for recommending new friends in a social software. If reasonable rules are not set, the system may pay too much attention to personal interests and ignore the potential spread of inappropriate content. For this reason, the development team will set up a filtering mechanism to push dating options to users only after a security assessment, reducing risks while taking into account personal preferences.

[0103] Monitoring the strategy update process to ensure the effect means that it is necessary to regularly evaluate whether the model adjustment has a negative impact on the overall performance, and make timely changes or corrections. For example, when the above recommendation software adds a push function based on popular tags, there may be a short period of time when the original active user volume declines due to the adjustment of the new strategy. At this time, the team will monitor this anomaly through background data analysis and quickly roll back to the previous version to stabilize the service status and prevent potential revenue losses. Through this process, we ensure that the strategy adjustment will bring positive changes.

[0104] Next, the steps of "learning the user's interactive behavior based on the AI ​​big model trained by the preset algorithm to obtain personalized response rules" are described as follows:

[0105] First, build a rich training data set to cover the user's interaction history with the system. These historical records may contain a large amount of user usage data, feedback, and responses in specific situations. Data collection should be extensive, comprehensive, and representative of the interaction habits of different types of users. For example, in smart home applications, the training data set can start with the specific requests when users communicate with smart assistants - such as setting a timer alarm, checking a schedule, or asking about the current weather.

[0106] Then set the structure and parameters of the initial model and start training the large model using these historical records. This means choosing a suitable machine learning algorithm and the corresponding model configuration, such as a deep neural network, and determining the layer structure and its number of nodes, activation function, loss function, and other hyperparameters. For example, for the above-mentioned intelligent assistant scenario, you can choose a model with a Transformer structure and set its embedding layer size to 768 to adapt to the understanding of complex conversations.

[0107] The training process also requires continuous analysis of the model's performance results to adjust optimization strategies, such as the number of iterations and related convergence conditions, to improve prediction accuracy. Specifically, if it is found that the intelligent assistant still cannot accurately interpret certain user instructions, the number of training iterations can be appropriately increased so that the model can better understand the user's intentions.

[0108] The last step is to select a portion of the non-training set as an offline test set to verify whether the learned rules can really provide accurate responses to individual users with high accuracy. After completing a series of internal tests, assuming that the intelligent assistant can respond quickly and accurately to new user instructions and maintain a high success rate, the personalized response mechanism can be considered to have been established. For example, the test set can include new question types or scenarios that have never appeared in the training data to examine the performance of the model in a real environment.

[0109] Through these steps, we can gradually build a large-scale artificial intelligence system that not only understands general language communication norms, but also can give personalized responses based on the characteristics of each user.

[0110] Next, the various steps and specific examples of "setting algorithm initialization parameters and model architecture and training a large model based on the above training data set" are described.

[0111] First, before training a machine learning model, appropriate initialization parameters and network structures must be selected. The key parts include determining the learning rate (controlling the speed of parameter updates), batch size (the number of samples required for each step of gradient descent), and other hyperparameters, such as weight decay terms. For example, in the training of deep learning models for natural language processing tasks, the initial learning rate may be set to 0.001, the batch size to 32 or 64, and L2 regularization may be enabled to avoid overfitting.

[0112] Then, after the model initialization configuration is completed, the model is trained using the data set. In this process, the back propagation algorithm is involved to calculate the derivative of each parameter layer by layer. Taking the neural network in supervised learning as an example, after forward propagation, the gap between the network output and the true label is calculated to form the so-called cost function (J). Subsequently, back propagation is used to calculate the partial derivative of each layer of weights to the cost function, which is used as the basis for updating the weights, and the parameters are adjusted by the gradient descent method. In a specific embodiment, the cross entropy loss function may be used as a method for calculating the cost function.

[0113] The training process is to repeat forward propagation and back propagation operations. At the end of each round (i.e., an Epoch), by checking whether the new (J) has changed compared with the previous one, if the change is lower than the set tolerance limit, it means that the training has reached a convergence state, and the update process can be stopped and no further iterations can be made. For example, when training a model for a classification task, if the loss reduction between two rounds is less than 0.001, the training is considered to be completed and can be terminated.

[0114] Finally, the choice of model architecture usually includes the number of hidden layers and nodes. For complex natural language processing tasks, a deeper and wider network model may be used, such as a five-layer hidden layer, and each layer contains 256 to 1024 neurons to process semantic information; for simpler problems, such as identifying small-scale targets in images, a small neural network with only one or two layers can be considered to achieve sufficiently high accuracy.

[0115] In summary, setting appropriate initialization parameters and network structure, and iterating on the training set according to these parameters can effectively optimize the model performance and obtain better results. In this way, an efficient AI model for specific problems can be obtained.

[0116] Next, the process of "using backpropagation to calculate the derivative of each parameter and then adjust its value to optimize the cost function (J)" is described. In machine learning, especially in deep learning neural networks, the backpropagation algorithm is a widely used gradient descent optimization technique that aims to effectively calculate the gradient of the loss function with respect to the parameters of each layer of the network and update these parameters accordingly to minimize the difference between the network output and the expected value.

[0117] The first step of backpropagation is to calculate the cost difference (C) of the output layer relative to the true label. Here, we choose an appropriate loss function, which measures the mismatch between the network's predictions and the actual target values. For example, cross entropy is often chosen as the loss function in classification tasks. The goal of this step is to provide a quantitative representation of the inaccuracy between the current network output and the true label.

[0118] Once the cost difference (C) is calculated, the next step is to use backpropagation to calculate the derivatives (i.e. gradients) of the parameters of each layer to determine which updates the network needs to make. Working backward from the output layer to the input layer, we use the chain rule to calculate the gradient of the loss with respect to each weight, which helps us understand how much adjustment to a particular weight is effective in reducing (C). For example, suppose our last layer has two units ((a^{[L]}) and (a^{[L-1]}), the activations of the current and previous layers respectively), and one output node. We will first determine the gradient of the current cost with respect to the activations of the last layer, and then work backwards, by calculating the partial derivatives of the loss function with respect to each level of weights, to identify the direction in which adjusting these parameters will contribute the most (or least, in absolute terms) to reducing the cost.

[0119] Then, if (C < threshold C_{min}) is found, it means that the current network is performing well to a certain extent, and the loss function has been reduced to the set threshold lower limit. It can be considered that the network has been well trained to some extent. According to the training strategy, it may enter the next iteration cycle at this time, or decide to terminate the training process according to other pre-set criteria.

[0120] Finally, based on the partial derivatives obtained above and the set learning rate (alpha), the weights of each layer in the network are updated. The specific mathematical update step is to add the negative learning rate to the weight and multiply it by the gradient calculated last time. The expression is written as the new weight

[0121]

[0122] To illustrate, assume that we are in a binary classification problem, and the weights of the last layer of the network ((w=(0.2))) correspond to a single neuron in the output layer. For a given set of data, we calculate the loss between the output result and the true label and find If (alpha = 0.05) at this time, if we calculate that (left|frac{dJ}{dw}right| = 2.5), the updated weight is (w = w-α×2.5 = 0.2-0.05times2.5 = 0.175). The whole process is executed repeatedly until the cost converges or the preset maximum number of training times is reached.

[0123] In short, the above process is a typical back-propagation mechanism and parameter adjustment method, which is one of the core algorithms for neural network optimization. This mechanism can help adjust complex weights and bias combinations, and then approach the ideal weight set to reduce the gap between the final prediction result and the correct label.

[0124] Next, it is described that “if (C<threshold C_min), the cost is considered to have been sufficiently reduced, and the next round of training is continued until it no longer decreases or meets the preset stopping point”.

[0125] This section describes the criteria for deciding when to proceed to the next iteration during training. Here, we will focus on two core stopping conditions: the first is to determine whether the model cost is low enough; the second is to determine whether the cost is still decreasing or stable by comparing the difference between the current model cost and the most recent model cost. If either condition is met, the model training can be terminated. The following first explains the significance of these criteria, and then gives specific examples to illustrate.

[0126] Under a given cost threshold, the learning progress of the model needs to be evaluated during each iteration of the model. After a certain number of learning steps, the "current cost (C)" will be compared with the "previous cost (C')". The difference between the two "(Δ(C) = |C-C'|)" is less than a set proportional coefficient "(r%)" relative to the difference in the previous round. If the difference becomes small enough or below the preset threshold, it means that the cost is difficult to reduce further. In addition, the system will also check how many rounds of training have been carried out or whether the performance of the model still shows an improving trend.

[0127] Another criterion is whether the maximum number of training times has been reached or the effect on the validation set has stagnated. This is to ensure that no unnecessary time is spent in a situation where the limit is almost reached. If any of these conditions are met, the training phase is considered to have achieved its goal, and the system will be ready to enter the final model test and subsequent application deployment.

[0128] - Assume that the model has an initial cost (C_{init}=10.0) and each round of model training is sufficient so that (r%=0.01(1%)) is the preset proportionality coefficient. "(C_min)" is set as a threshold value of 3.0.

[0129] -After several rounds, the "current cost (C)" is assumed to drop to 7.8, and the "last round cost" is recorded as "7.5" (assuming (△(C)=|7.8-7.5|=0.3), relative to the previous (0.3>r%*C), where "(C=7.5)" in the last training, thus satisfying the condition that the cost is still decreasing).

[0130] -But suppose that at a certain point in time, the cost drops to 3.0, and at this time (C_{t+1}-C_t=0.02<0.01*3); therefore, the stopping condition that (C) no longer drops to a significant degree and is less than the threshold is met.

[0131] Under this condition, "(△C=|3.0-3.02|)", and because it is close to (1%) of the (C) value, and (0.01cdot3approx0.03), that is, it meets the standard of "the difference becomes quite small", indicating that the training process can be ended.

[0132] At the same time, in another assumption, although (C = 2.99 < 3) (which means it is very close to the minimum limit), the model may not make significant progress in several consecutive rounds (for example, the accuracy on the validation set is less than 1%) or the maximum allowed number of iterations, 200, has been reached. In this case, the condition of "the cost reduction rate is not significant enough" is considered to have been met. Therefore, according to the description logic, even if (C) is lower than (C_min), as long as any of the other termination conditions is met, the training will stop and enter the final testing and deployment process.

[0133] This represents a flexible learning strategy based on conditions, which ensures efficient use of resources and high-quality machine learning results. This training mechanism ensures that not only the cost is sufficiently reduced, but also the model itself is continuously iterated and optimized until it reaches the established standard or the performance is no longer significantly improved, thus preparing a more accurate and reliable large-scale AI system for practical applications.

[0134] Next, describe "when the difference between the current cost (C) and the previous iteration (△(C)) is less than a predefined ratio (r%, rge0)":

[0135] This description involves the judgment of the stability of the iterative algorithm and the minimum cost solution in the optimization process of the human-computer interaction method and system using artificial intelligence large models. According to the given requirements, when the cost reduction between two steps is not enough to meet the predefined criteria relative to the current cost value, it is judged that the stability point has been reached, and further iterations are stopped.

[0136] The difference between two adjacent sets of (C) data is the difference in cost between two consecutive iterations (|C^{t}-C^{t-1}|); (C^{t}) represents the cost value of the current iteration, while (C^{t-1}) refers to the cost obtained in the previous iteration. This difference value reflects the degree of decrease in the current cost compared to the previous one, and can be used as a standard reference for whether the iteration needs to be terminated.

[0137] This gap value is then compared with a fraction of the current cost adjusted according to a preset threshold: (|C^{t}|cdot(frac{r}{100})). The comparison here is to determine whether the gap value is small enough to be considered to have reached the expected convergence level, or whether more iterations are needed to further reduce the difference so that the iterations can move closer to the global minimum. This condition is actually a termination judgment condition for the training process - if (|C^{t}-C^{t-1}|<|C^{t}|cdotfrac{r}{100}), then the model is considered to be close to the optimal result or converged, and the training process can be stopped. Otherwise, it means that more calculations are needed to reach the convergence point required for the expected effect.

[0138] As a specific example, assume that the current cost (C^{t}=50) and the last cost (C^{t-1}=70), and use (r=0.02) or (2%) as the convergence criterion:

[0139] - The cost difference between the two steps is: (|C^{t}-C^{t-1}||=|50-70|=20)

[0140] - Let (|C^{t}cdotfrac{0.02}{1})|=|50*0.02|=1),

[0141] Due to the difference between the current two steps (20>1), the convergence condition is not met at this time.

[0142] After continuing the iterative training for a period of time, assuming that in a certain update (C^{*t}=50.25) is obtained, and the cost at the previous moment (C^{*t-1}=50.32),

[0143] - Re-evaluation: (|50.25-50.32|=0.07)

[0144] - Check condition: (50.25*frac{0.02}{1}=1)

[0145] Now we find that 0.07 is less than 1, which means that the condition is met. Therefore, we can stop the iteration operation and the model has reached the preset stable point.

[0146] In addition, (|C^{t}|cdot(r / 100)) is continuously updated during the entire iteration process. This is because as time goes by, the algorithm runs, and the model learning changes, this value also needs to be updated accordingly to accurately reflect the iteration termination conditions as the cost (C^{t}) changes.

[0147] The "human-computer interaction method and system based on AI big model" of the present invention includes: an innovative way to overcome the common conflict between personalized experience and user privacy protection. This technical solution mainly covers four steps:

[0148] 1. Personalized response rule learning: The present invention uses a large-scale artificial intelligence (AI) model obtained through pre-set algorithm training to comprehensively learn various types of user interaction behaviors, aiming to extract personalized characteristics from these behavior patterns, and build a response mechanism based on personal preferences. This mechanism can generate more accurate and personalized interaction responses based on the user's specific behavioral habits, thereby greatly improving the user experience level. This step is the core of the entire system. Through intelligent learning, it realizes the precise grasp of user preferences and enables the accurate generation of personalized feedback information.

[0149] 2. Safe processing of user interaction data: Considering the privacy leakage issues that may arise in the process of optimizing user experience, this method takes necessary measures to ensure that all collected user data is fully desensitized before being used as the basis for iteration of the AI ​​large model. This means that even the most basic level of information will not be exposed without authorization, effectively protecting the privacy of the data and complying with the existing legal requirements for data security.

[0150] 3. Dynamic generation of differentiated feedback information: Based on the hard work in the above two stages, a complete set of personalized feedback logic has been generated. This logic can generate dynamic interactive instructions or suggestions with high individual matching and strong pertinence in real time according to the changes in the usage scenarios of specific users, which significantly increases the flexibility and friendliness of the interaction process, making users feel more cared for, thereby promoting an increase in long-term willingness to use and an improvement in word-of-mouth effect, enhancing the overall experience while maintaining user privacy.

[0151] 4. Policy weight adjustment: The system fine-tunes the feedback mechanism and optimizes the whole system by continuously collecting new information from users and analyzing past behavior results. This method not only allows AI to gradually understand the unique characteristics of each user and continuously improve its response accuracy, but also ensures that even the smallest adjustments will respect and protect the privacy rights of each individual. Ultimately, by dynamically balancing the differences between user experience needs and privacy boundaries, the maximum compatibility of the interests of both parties is achieved, and the core challenges existing in the current human-computer interface are effectively addressed.

[0152] In summary, the present invention uses advanced AI technology and rigorous data management methods to successfully build a powerful human-computer communication bridge that is highly personalized while taking into account data privacy protection functions, solving the difficult problem of how to ensure the quality of user experience while avoiding infringement of user privacy.

[0153] The methods, programs, systems, devices, etc. of the embodiments of the present invention may be executed or implemented in a single or multiple networked computers, or may be practiced in a distributed computing environment. In the embodiments of this specification, in these distributed computing environments, tasks may be performed by remote processing devices connected via a communication network.

[0154] Those skilled in the art should understand that the embodiments of this specification may provide methods, systems or computer program products. Therefore, those skilled in the art may imagine that the implementation of the functional modules / units or controllers and related method steps described in the above embodiments may be implemented in software, hardware or a combination of software / hardware.

[0155] Unless explicitly stated, the actions or steps of the methods, programs, and embodiments of the present invention do not have to be performed in a specific order and can still achieve the desired results. In some implementations, multitasking and parallel processing are also possible or may be advantageous.

[0156] In this article, multiple embodiments of the present invention are described, but for the sake of brevity, the description of each embodiment is not exhaustive, and the same or similar features or parts between the embodiments may be omitted. In this article, "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" are meant to be applicable to at least one embodiment or example according to the present invention, but not all embodiments. The above terms do not necessarily mean to refer to the same embodiment or example. In the absence of contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples.

[0157] The exemplary systems and methods of the present invention have been specifically shown and described with reference to the above embodiments, which are merely examples of the best modes for implementing the present systems and methods. It will be appreciated by those skilled in the art that various changes may be made to the embodiments of the systems and methods described herein when implementing the present systems and / or methods without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims

1. A human-computer interaction method and system based on an AI big model, characterized in that: include: The AI ​​big model trained based on the preset algorithm learns the user's interactive behavior to obtain personalized response rules; Obtain user interaction data for model optimization and iteration while ensuring data anonymization; Dynamically generate differentiated feedback information for different users based on personalized response rules to enhance user experience; Adjust the strategy weight of the AI ​​model based on feedback to balance the personalized needs of user experience and privacy protection.

2. The human-computer interaction method and system based on AI big model according to claim 1 is characterized in that: The steps specifically include: Collect user interaction logs as original training samples; Performing feature engineering processing on the original training samples to extract effective features; Use preset algorithms and processed samples to conduct initial training of the AI ​​model; Analyze the model's performance indicators based on preliminary training results and make necessary fine-tuning; The deviation between the model-predicted response and the actual user feedback is evaluated to form personalized response rules.

3. The human-computer interaction method and system based on AI big model according to claim 1 is characterized in that: The steps specifically include: Use de-identification technology to remove or replace personally identifiable information; Apply data desensitization algorithms to blur sensitive fields by adding noise, etc. Regularly update desensitization strategies to ensure data security; Establish access control and monitoring mechanisms to protect the security of processed user data; The processed user interaction data is input into the big AI model as a data source for retraining.

4. The human-computer interaction method and system based on AI big model according to claim 1 is characterized in that: The steps specifically include: Capture the contextual information of the current user interaction state in real time; Call personalized response rules for intelligent matching and generate basic response message templates; Compare multiple possible response messages and select the optimal solution based on the user's historical preferences; Dynamically adjust the response speed to keep the interaction rhythm in sync with user expectations; Combine user feedback to further improve the content and presentation of response information.

5. The human-computer interaction method and system based on AI big model according to claim 1 is characterized in that: The steps specifically include: Establish a strategy weight update mechanism to adapt to different feedbacks; Quantitatively evaluate user satisfaction to determine positive and negative reward values; Use reinforcement learning techniques to adjust model parameters to achieve a better balance; Introduce expert rules to avoid excessive bias towards personal preferences and ignoring privacy and security risks; Monitor the impact of the strategy update process on the overall model effect to ensure stable operation.

6. The human-computer interaction method and system based on AI big model according to claim 1 is characterized in that: The AI ​​big model trained based on the preset algorithm learns the user's interactive behavior to obtain personalized response rules including: Build training datasets to cover multiple types of user interaction histories; Set the algorithm initialization parameters and model architecture and train the large model based on the above training data set; Adjust the number of model iterations according to the interaction effect until it converges to the predetermined accuracy; An offline test set is used to verify the effectiveness of the personalized response rules obtained during the learning process.

7. The human-computer interaction method and system based on AI big model according to claim 2 is characterized in that: Setting the algorithm initialization parameters and model architecture and training the large model based on the above training data set includes: Determine algorithm hyperparameters such as learning rate and batch size as initialization configuration; Use backpropagation to calculate the derivative of each parameter and adjust its value to optimize the cost function (J); Once the change in the cost function (J) is less than a predefined small value, it is considered converged and the training is stopped; Choose the appropriate number of hidden layers and neuron nodes to meet the training requirements.

8. The human-computer interaction method and system based on AI big model according to claim 3 is characterized in that: Using backpropagation to calculate the derivative of each parameter and then adjust its value to optimize the cost function (J) includes: Calculate the cost difference (C) of the output layer relative to the true value of the label; Back-step through each layer of network nodes to find partial derivatives to quantify the direction and magnitude of each update; If (C < threshold C_min), the cost is considered to have been sufficiently reduced, and the next round of training will continue until it no longer decreases or meets the preset stopping point; The weights of nodes at each level are updated through gradient (where (α) is a learning rate fixed in advance).

9. The human-computer interaction method and system based on AI big model according to claim 4 is characterized in that: If (C < threshold C_min), the cost is considered to have been sufficiently reduced, and the next round of training is continued until it no longer decreases or meets the preset stopping point, including: When the difference between the current cost (C) and the previous iteration (△(C)) is less than the predefined ratio (r%, rge0); Check whether the specified number of training rounds has been reached or the model performance on the validation set is no longer significantly improved; If any of the conditions are met, this round of learning cycle ends and the testing and deployment preparation phase begins; If any condition is not met, the current learning state is continued for further optimization (C).

10. The human-computer interaction method and system based on AI big model according to claim 5, characterized in that: When the difference between the current cost (C) and the previous iteration (△(C)) is less than a predefined ratio (r%, rge0), it includes: Obtain the difference between two adjacent sets of (C) data (|C^{t}C^{t1}|); Compare this difference with (|C^{t}|cdot(r / 100)); If the former is less than the latter, it indicates that the model is approaching a stable state or a minimum point, thus deciding to terminate the iteration operation; otherwise, the iteration is continued according to the existing trend; During the model training process, the (|C^{t}|cdot(r / 100)) value is updated in real time to track the changes and ensure that the convergence point is detected in time.

Images