Self-evolution model generation and real-time optimization method for intelligent dialogue system and related device

Abstract

The invention provides a self-evolution model generation and real-time optimization method for an intelligent dialogue system and a related device, and relates to the technical field of artificial intelligence. By establishing a collaborative optimization mechanism, dynamically expanding a model architecture and realizing real-time iterative update, the problems of model update lag, learning mechanism isolation, architecture stiffness and iteration delay are solved, and the method has the advantages of improving the dynamic adaptability, collaborative optimization capability and real-time response efficiency of the system.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method and related apparatus for generating and optimizing self-evolving models in real time for intelligent dialogue systems. Background Technology

[0002] With the rapid development of artificial intelligence technology, intelligent dialogue systems are playing an increasingly crucial role in enterprise services. Their core relies on natural language processing and machine learning models to accurately identify user intent and efficiently generate responses. However, in real-world business scenarios, when faced with rapidly changing market environments and user demands, existing intelligent dialogue models reveal deep-seated technical bottlenecks, severely hindering the continuous improvement of service efficiency.

[0003] The problems of lagging model updates and poor adaptability are particularly prominent. Traditional models rely on static training and periodic offline update mechanisms, making them unable to dynamically respond to business evolution. For example, when a company launches a new product line or marketing campaign, relevant new knowledge and user expressions are difficult to incorporate into the model training system in a timely manner. This leads to a significant decrease in the model's confidence in intent recognition when dealing with emerging business topics, resulting in responses that deviate from the user's actual needs and ultimately frequently triggering human customer service transfers. This kind of "model rigidity" not only increases the company's operating costs but also causes a decline in user satisfaction, making it difficult for intelligent dialogue systems to keep up with the pace of business dynamics. During seasonal promotions or emergencies, the rapid changes in user query patterns further amplify this deficiency, causing a precipitous drop in model performance.

[0004] The isolation of incremental learning and reinforcement learning exacerbates the complexity of system optimization. While some technical solutions incorporate incremental learning to continuously absorb new dialogue data and reinforcement learning to optimize dialogue strategies, the two are often implemented in isolation. Incremental learning focuses solely on the linear expansion of the knowledge base, while reinforcement learning is limited to policy-level adjustments, lacking an inherent synergistic mechanism. This separation puts the model in a dilemma: on the one hand, newly learned knowledge cannot be effectively transformed into decision-making ability, resulting in inappropriate responses to identified intentions; on the other hand, policy optimization encounters bottlenecks due to lagging knowledge base updates. For example, after introducing new services, although the model can identify user intentions, it cannot generate effective scripts to guide self-service. This disconnect between knowledge learning and policy optimization hinders the coordinated evolution of the system's overall intelligence level.

[0005] The rigidity of the model architecture further limits generalization ability. The network structure of existing models is fixed in the initial design stage, including key parameters such as the dimension and number of layers of the classification network. When encountering entirely new business domains not covered by training data, such as suddenly expanding from conventional telecommunications services to financial advisory services, the significantly different distribution of dialogue data makes the fixed architecture difficult to adapt. Relying solely on parameter fine-tuning cannot solve the fundamental structural defects, the system performance deteriorates sharply, and cross-domain adaptive migration is impossible. This rigid architecture forces enterprises to invest a lot of resources in model reconstruction when diversifying their businesses, resulting in extended development cycles and wasted resources.

[0006] Feedback delays and excessively long iteration cycles prevent the system from evolving in real time. Current optimization processes are typically performed in batches on a weekly or monthly basis. Real-time feedback data generated during user interactions, such as conversation interruptions, low satisfaction ratings, or repeated questions, cannot be captured and used for model correction in a timely manner. This delay prevents the model from quickly identifying and correcting errors. For example, negative user feedback on new features may not be processed for weeks, allowing problems to fester. The system gradually evolves into a passively responding "static" entity, rather than a "living" system that can evolve in sync with user behavior and environmental changes, ultimately weakening the practical value and market competitiveness of intelligent customer service.

[0007] In summary, existing technologies have systemic deficiencies in terms of model dynamic adaptability, learning mechanism synergy, architectural flexibility, and iteration timeliness. There is an urgent need for an intelligent dialogue model generation method that can deeply integrate knowledge updates and strategy optimization and has the ability to autonomously adjust the architecture, so as to support the continuous evolution of the system in complex and ever-changing business environments.

[0008] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0009] This application provides a method and related apparatus for generating and optimizing self-evolving models in real time for intelligent dialogue systems. It addresses issues such as lagging model updates, isolated learning mechanisms, rigid architectures, and iteration delays, thereby improving the system's dynamic adaptability, collaborative optimization capabilities, and real-time response efficiency.

[0010] Firstly, the self-evolving model generation and real-time optimization method for intelligent dialogue systems provided in this application adopts the following technical solution: A method for generating and optimizing self-evolving models in real time for intelligent dialogue systems includes: Model initialization steps: Build an initial intelligent dialogue model, which includes a feature extraction network, an intent classification network, and a response generation network; The steps for establishing a collaborative optimization loop are as follows: A collaborative optimization mechanism for incremental learning and reinforcement learning is established. The incremental learning process is configured as follows: Newly generated user-customer service dialogue records are periodically collected, cleaned, and labeled. New intent features and question-and-answer knowledge are extracted, and the parameters of the feature extraction network and intent classification network are updated in a fine-tuning manner to expand and correct the model's knowledge base. The reinforcement learning process is configured as follows: Each intelligent dialogue is treated as a decision-making process. Successful completion of the dialogue and user satisfaction evaluation are used as positive feedback rewards, while the interruption of the dialogue by the user or transfer to human customer service is used as negative feedback rewards. Based on these reward signals, the policy function of the response generation network is optimized using a policy gradient algorithm, enabling it to learn the policy for generating the optimal response from the current dialogue state. The incremental learning process and the reinforcement learning process share a collaborative trainer, which balances the model's ability to learn new knowledge and its optimal decision-making ability based on existing knowledge by dynamically adjusting the weights of knowledge distillation loss and policy reward loss. Self-evolution triggering and architecture adjustment steps: The system monitors key performance indicators in real time; when the confidence level of intent recognition for a specific emerging business topic is consistently lower than the first threshold, or the conversion rate of dialogues related to the topic to human agents is consistently higher than the second threshold, it is determined that the current model architecture is insufficient to handle the new topic, and the self-evolution mechanism is triggered; the execution method of the self-evolution mechanism is as follows: freeze the original main structure of the model, connect a dedicated classification sub-network for the new topic in parallel after the intent classification network, and add a policy routing module corresponding to the sub-network before the response generation network, thereby dynamically expanding the architecture of the model to adapt to the new scenario; Online real-time iteration steps: Establish an online learning closed loop; stream user feedback and dialogue logs generated online into the collaborative optimization loop to perform hourly or daily incremental updates to the model parameters; simultaneously, for newly added network modules generated in the self-evolution step, use recent highly relevant dialogue data for centralized training to achieve rapid integration and deployment of new functions.

[0011] Optionally, in the collaborative optimization loop establishment step, the method for dynamically adjusting the weights is as follows: when the amount of new dialogue data increases rapidly within a unit of time, the weight of knowledge distillation loss is increased to enhance new knowledge learning; when the user negative feedback rate increases rapidly within a unit of time, the weight of strategy reward loss is increased to optimize the decision-making strategy.

[0012] Optionally, in the self-evolution triggering and architecture adjustment steps, the key performance indicators also include the rate of repeated questions asked by the user again within a preset time window after the user has had a dialogue with the intelligent system. When this rate exceeds a specific threshold, it serves as an auxiliary judgment condition for triggering the self-evolution mechanism.

[0013] Optionally, in the self-evolution triggering and architecture adjustment steps, the dedicated classification sub-network is a lightweight network whose training data mainly comes from samples obtained by data augmentation of corpora related to emerging business topics in historical dialogues.

[0014] Optionally, in the online real-time iteration step, the streaming input data first passes through a feedback filter, which scores the feedback information based on its clarity and information content, and only sends high-quality feedback data with scores higher than a threshold into the model for iterative updates.

[0015] Optionally, the method further includes a model stability maintenance step: periodically testing the iteratively evolved model with a validation set covering the basic business; when the model's accuracy in recognizing the intent of the basic business drops beyond the tolerance level, triggering a model rollback mechanism and initiating special reinforcement training for catastrophic forgetting.

[0016] Optionally, the policy function optimization of the response generation network adopts a constrained policy optimization algorithm, and an additional reward is set in the reward function for successfully guiding users to use self-service.

[0017] Secondly, this application provides a self-evolving model generation and real-time optimization system for intelligent dialogue systems, comprising: The model initialization module is used to build an initial intelligent dialogue model, which includes a feature extraction network, an intent classification network, and a response generation network. A collaborative optimization loop establishment module is used to establish a collaborative optimization mechanism for incremental learning and reinforcement learning. The incremental learning process is configured to: periodically collect newly generated user-customer service dialogue records, clean and label them, extract new intent features and question-and-answer knowledge, and fine-tune the parameters of the feature extraction network and intent classification network to expand and correct the model's knowledge base. The reinforcement learning process is configured to: treat each intelligent dialogue as a decision-making process, using successful completion of the dialogue and user satisfaction evaluation as positive feedback rewards, and interruption of the dialogue by the user or transfer to human customer service as negative feedback rewards. Based on these reward signals, the policy function of the response generation network is optimized through a policy gradient algorithm, enabling it to learn the policy for generating the optimal response from the current dialogue state. The incremental learning process and the reinforcement learning process share a collaborative trainer, which dynamically adjusts the weights of knowledge distillation loss and policy reward loss to balance the model's ability to learn new knowledge and its optimal decision-making ability based on existing knowledge. The self-evolution triggering and architecture adjustment module is used for real-time monitoring of key performance indicators of the system. When the confidence level of intent recognition for a specific emerging business topic is continuously lower than the first threshold, or the conversion rate of dialogues related to the topic to human agents is continuously higher than the second threshold, it is determined that the current model architecture is insufficient to handle the new topic, and the self-evolution mechanism is triggered. The execution method of the self-evolution mechanism is as follows: freeze the original main structure of the model, connect a dedicated classification sub-network for the new topic in parallel after the intent classification network, and add a policy routing module corresponding to the sub-network before the response generation network, thereby dynamically expanding the architecture of the model to adapt to the new scenario. An online real-time iteration module is used to establish an online learning loop; real-time user feedback and dialogue logs generated online are streamed into the collaborative optimization loop to perform incremental updates of model parameters on an hourly or daily basis; at the same time, newly added network modules generated in the self-evolution step are trained intensively using recent highly relevant dialogue data to achieve rapid integration and deployment of new functions.

[0018] Thirdly, this application provides a computer device, the device comprising: a memory and a processor, wherein the processor, when executing computer instructions stored in the memory, performs the method described above.

[0019] Fourthly, this application provides a computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform the method described above.

[0020] In summary, this application solves the problems of lagging model updates, isolated learning mechanisms, rigid architecture, and iteration delays by establishing a collaborative optimization mechanism, dynamically expanding the model architecture, and realizing real-time iterative updates. It has the advantages of improving the system's dynamic adaptability, collaborative optimization capabilities, and real-time response efficiency. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the computer device structure of the hardware operating environment involved in the embodiments of this application; Figure 2 This is a flowchart illustrating the first embodiment of the self-evolutionary model generation and real-time optimization method for intelligent dialogue systems according to this application; Figure 3 This is a structural block diagram of the first embodiment of the self-evolving model generation and real-time optimization system for intelligent dialogue systems according to this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0023] Reference Figure 1 , Figure 1 This is a schematic diagram of the computer device structure of the hardware operating environment involved in the embodiments of this application.

[0024] like Figure 1 As shown, the computer device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be high-speed random access memory (RAM) or stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0025] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the computer device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0026] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a network communication module, a user interface module, and a self-evolving model generation and real-time optimization program for an intelligent dialogue system.

[0027] exist Figure 1 In the computer device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in this application can be set in the computer device. The computer device calls the self-evolutionary model generation and real-time optimization program for intelligent dialogue system stored in the memory 1005 through the processor 1001, and executes the self-evolutionary model generation and real-time optimization method for intelligent dialogue system provided in the embodiment of this application.

[0028] This application provides a method for generating and optimizing a self-evolving model in real time for an intelligent dialogue system, referring to... Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the self-evolutionary model generation and real-time optimization method for intelligent dialogue systems according to this application.

[0029] In this embodiment, the method for generating and optimizing a self-evolving model in real time for an intelligent dialogue system includes the following steps: Model initialization step 10: Construct an initial intelligent dialogue model, which includes a feature extraction network, an intent classification network, and a response generation network; Step 20 of the collaborative optimization loop establishment: Establish a collaborative optimization mechanism for incremental learning and reinforcement learning; wherein, the incremental learning process is configured as follows: periodically collect newly generated user-customer service dialogue records, clean and label them, extract new intent features and question-and-answer knowledge, and update the parameters of the feature extraction network and intent classification network in a fine-tuning manner to expand and correct the model's knowledge base; the reinforcement learning process is configured as follows: treat each intelligent dialogue as a decision-making process, use the successful completion of the dialogue and user satisfaction evaluation as positive feedback rewards, and use the interruption of the dialogue by the user or transfer to human customer service as negative feedback rewards, and optimize the policy function of the response generation network through the policy gradient algorithm based on these reward signals, so that it learns the policy to generate the optimal response from the current dialogue state; the incremental learning process and the reinforcement learning process share a collaborative trainer, which balances the model's ability to learn new knowledge and its ability to make optimal decisions based on existing knowledge by dynamically adjusting the weights of knowledge distillation loss and policy reward loss; Self-evolution triggering and architecture adjustment step 30: The system monitors key performance indicators in real time; when the confidence level of intent recognition for a specific emerging business topic is continuously lower than the first threshold, or the conversion rate of dialogues related to the topic to human agents is continuously higher than the second threshold, it is determined that the current model architecture is insufficient to handle the new topic, and the self-evolution mechanism is triggered; the execution method of the self-evolution mechanism is as follows: freeze the original main structure of the model, connect a dedicated classification sub-network for the new topic in parallel after the intent classification network, and add a policy routing module corresponding to the sub-network before the response generation network, thereby dynamically expanding the architecture of the model to adapt to the new scenario; Online real-time iteration step 40: Establish an online learning closed loop; input real-time user feedback and dialogue logs generated online into the collaborative optimization loop to perform incremental updates of model parameters on an hourly or daily basis; at the same time, for the newly added network modules generated in the self-evolution step, use recent highly relevant dialogue data for centralized training to achieve rapid integration and deployment of new functions.

[0030] Traditional intelligent dialogue systems suffer from lagging model updates and poor adaptability when facing rapidly changing business environments and user needs, resulting in low confidence in intent recognition and inaccurate responses. Incremental learning and reinforcement learning applications are isolated, making it difficult to balance new knowledge learning with decision strategy optimization. Fixed model architecture leads to insufficient generalization ability, feedback delays, and long iteration cycles, making it difficult for the system to adapt quickly to changes.

[0031] In response, this embodiment proposes a method for generating and optimizing self-evolving models in real time for intelligent dialogue systems. The method aims to achieve self-evolution and real-time optimization of the model by deeply integrating continuous learning and decision optimization and dynamically adjusting its own architecture, thereby improving the intelligence level and service efficiency of the intelligent dialogue system.

[0032] For ease of understanding, the following explains some key terms in this embodiment: Intelligent dialogue models are software systems designed to simulate natural language interactions between human customer service representatives and users. Their core function is to understand user intent and generate appropriate responses. These models typically consist of multiple neural network modules working together to handle each stage of the dialogue.

[0033] The feature extraction network is responsible for extracting semantic features and contextual information from the user's input text. These features form the basis for subsequent intent classification and response generation, and their quality directly affects the model's understanding ability.

[0034] The intent classification network identifies the deep intent of user dialogues based on the features output by the feature extraction network, such as inquiring about orders, processing transactions, or seeking help. The accuracy of this network is crucial for the intelligent dialogue system to respond correctly to user requests.

[0035] The response generation network generates natural, accurate, and context-appropriate text responses based on user intent and the current dialogue state. This network typically employs a sequence-to-sequence architecture or is fine-tuned based on a large language model.

[0036] Incremental learning is a machine learning paradigm that allows models to continuously update and expand their knowledge base by learning from new data without complete retraining. This approach helps models adapt to changing business needs and the emergence of new knowledge.

[0037] Reinforcement learning is a machine learning method that learns optimal decision-making strategies through interaction with the environment. In this method, the model adjusts its behavior based on reward or penalty signals obtained from the environment to maximize long-term rewards.

[0038] The co-trainer is a component in this method used to coordinate the incremental learning process and the reinforcement learning process. Its role is to dynamically adjust the weights between different learning objectives to ensure that the model can optimize its decision-making strategy while learning new knowledge.

[0039] Knowledge distillation loss is a loss function used to measure the extent to which a student model learns knowledge from a teacher model. In this method, it is used to guide the model to effectively absorb new knowledge.

[0040] Policy reward loss is a loss function used in reinforcement learning to measure the quality of the actions generated by the model's current policy. By optimizing this loss, the model can learn a policy that generates the optimal response in a specific dialogue state.

[0041] The self-evolution mechanism refers to the system's ability to automatically trigger and dynamically adjust its network structure when it detects that the existing model architecture cannot effectively handle new business scenarios. This mechanism enables the model to fundamentally adapt to new challenges.

[0042] The dedicated classification subnetwork is a network module dynamically connected in parallel to the intent classification network to handle specific emerging business topics. It focuses on identifying user intents related to the new topic, thereby improving adaptability to new scenarios.

[0043] The policy routing module corresponds to the dedicated classification sub-network and is added before the response generation network. Its function is to route the dialogue flow to the appropriate response generation path based on the intent classification results, ensuring that responses to new topics are generated correctly.

[0044] An online learning closed loop refers to a mechanism that continuously inputs real-time user feedback and dialogue logs generated online into the model for rapid iterative updates. This closed loop enables the model to achieve incremental updates on an hourly or daily basis, thereby quickly responding to changes in user needs.

[0045] The self-evolutionary model generation and real-time optimization method in this embodiment specifically includes the following steps: First, the model initialization step is performed. This step aims to build an initial intelligent dialogue model, configured to include a feature extraction network, an intent classification network, and a response generation network. For example, a pre-trained general language model can be used as a base and fine-tuned to adapt to the intelligent dialogue scenario. The feature extraction network can be a Transformer encoder responsible for extracting semantic features from the user's input text; the intent classification network can be a multilayer perceptron used to recognize user intent; and the response generation network can be a Transformer decoder used to generate text responses. These networks are initially configured to handle dialogue in basic business scenarios.

[0046] Secondly, a collaborative optimization loop is established. This step aims to establish a collaborative optimization mechanism between incremental learning and reinforcement learning. Specifically, the incremental learning process is configured to periodically collect newly generated user-customer service dialogue records, clean and label them, and extract new intent features and question-and-answer knowledge. For example, new dialogue records can be manually filtered and labeled periodically from online logs; the cleaning process may include removing duplicates and filtering low-quality dialogues. Subsequently, the parameters of the feature extraction network and intent classification network are updated in a fine-tuning manner to expand and correct the model's knowledge base. The reinforcement learning process is configured to treat each intelligent dialogue as a decision-making process, with successful completion of the dialogue and user satisfaction evaluation as positive feedback rewards, and the interruption of the dialogue by the user or transfer to human customer service as negative feedback rewards. For example, reward signals can be obtained from the user's proactive evaluation at the end of the dialogue or from indirect indicators such as dialogue duration and user behavior. Based on these reward signals, the policy function of the response generation network is optimized through a policy gradient algorithm, enabling it to learn the strategy for generating the optimal response from the current dialogue state. This incremental learning process shares a co-trainer with the reinforcement learning process. This co-trainer balances the model's ability to learn new knowledge with its ability to make optimal decisions based on existing knowledge by dynamically adjusting the weights of the knowledge distillation loss and the policy reward loss. For example, the weight adjustment logic can be based on preset heuristic rules to adapt to different learning stages.

[0047] Further, the system executes a self-evolution triggering and architecture adjustment step. This step aims to enable the system to dynamically expand the model's architecture to adapt to new scenarios. The system monitors key performance indicators in real time, such as the accuracy of intent recognition and user satisfaction. When the confidence level of intent recognition for a specific emerging business topic consistently falls below a first threshold, or the conversion rate of dialogues related to that topic to human intervention consistently exceeds a second threshold, the system determines that the current model architecture is insufficient to handle the new topic, thus triggering the self-evolution mechanism. This self-evolution mechanism is executed as follows: the original main structure of the model is frozen, a dedicated classification sub-network for the new topic is connected in parallel after the intent classification network, and a policy routing module corresponding to this sub-network is added before the response generation network. For example, the dedicated classification sub-network could be a small feedforward neural network that focuses on recognizing user intents related to the new topic; the policy routing module could be a simple classifier that determines whether to route the request to the newly added response generation path based on the output of the intent classification network.

[0048] Finally, an online real-time iteration step is executed. This step aims to establish an online learning loop, enabling rapid integration and deployment of new features. Real-time user feedback and dialogue logs generated online are streamed into this collaborative optimization loop, providing hourly or daily incremental updates to the model parameters. For example, real-time user feedback can include user behavior data such as clicks, likes, and negative reviews on replies, while the streamed input of dialogue logs can be achieved through a message queue system. Simultaneously, newly added network modules generated in the self-evolution step are trained intensively using recent, highly relevant dialogue data. For example, the intensive training data for new network modules can be obtained by selecting initial dialogues after the new service launch and performing manual or semi-automatic annotation.

[0049] This embodiment effectively balances the model's ability to learn new knowledge with its ability to make optimal decisions based on existing knowledge by establishing a collaborative optimization mechanism of incremental learning and reinforcement learning, thus solving the problems of lagging model updates and isolated learning mechanisms in traditional models. Simultaneously, the introduction of self-evolution triggering and architecture adjustment steps enables the model architecture to dynamically expand to adapt to emerging business scenarios, overcoming the limitations of insufficient generalization ability in fixed architectures. Combined with online real-time iteration, hourly or daily updates of model parameters are achieved, significantly improving the adaptability and response speed of the intelligent dialogue system to business changes.

[0050] In some of the embodiments described above, a collaborative optimization loop is proposed. This loop dynamically adjusts the weights of knowledge distillation loss and policy reward loss to balance the model's ability to learn new knowledge with its optimal decision-making ability based on existing knowledge. However, in the actual operation of intelligent dialogue systems, facing constantly changing business scenarios and user interaction patterns, how to intelligently and effectively adjust these weights according to the real-time state of the system, such as the rate of new data inflow or fluctuations in user satisfaction, to ensure that the model can continuously and efficiently learn new knowledge and optimize decision-making strategies, is a problem that needs to be solved.

[0051] In response, this embodiment further proposes that, in the collaborative optimization loop establishment step, the method for dynamically adjusting the weights is as follows: when the amount of new dialogue data increases rapidly within a unit of time, the weight of knowledge distillation loss is increased to enhance new knowledge learning; when the user negative feedback rate increases rapidly within a unit of time, the weight of strategy reward loss is increased to optimize the decision-making strategy.

[0052] The described dynamic weight adjustment method aims to enable the collaborative trainer to adaptively adjust the emphasis of incremental learning and reinforcement learning processes in model optimization based on the real-time operational status of the intelligent dialogue system. This dynamic adjustment mechanism avoids the problems of low learning efficiency or decreased decision quality that may result from fixed weights, thus enabling the model to better adapt to dynamically changing business environments. Specifically, when the system detects a rapid increase in the amount of new dialogue data per unit time, this usually means that a large number of new user inquiries, new business topics, or new knowledge points are emerging. In this case, to ensure that the model can absorb this new knowledge in a timely and effective manner, this embodiment enhances the learning of new knowledge by increasing the weight of the knowledge distillation loss. Knowledge distillation is a model compression and knowledge transfer technique that reduces model complexity while maintaining performance by having a small model (student model) learn the output of a large model (teacher model). In this context, increasing the weight of the knowledge distillation loss means that the co-trainer will focus more on efficiently integrating knowledge from newly collected and labeled dialogue data into the feature extraction network and intent classification network in a fine-tuned manner. This rapidly expands and corrects the model's knowledge base, preventing lag in the model's understanding and processing of new business due to insufficient new knowledge learning. On the other hand, when the user negative feedback rate increases rapidly within a unit of time, it usually indicates that the response generated by the current response generation network fails to effectively solve the user's problem or leads to user dissatisfaction. Negative feedback may manifest as the user interrupting the conversation, transferring to human customer service, or explicitly expressing dissatisfaction. To quickly correct this poor decision-making behavior, this embodiment optimizes the decision-making strategy by increasing the weight of the policy reward loss. The policy reward loss is a key indicator in reinforcement learning, directly reflecting the quality of the model's decision-making when generating responses. Increasing its weight means that the co-trainer will focus more on utilizing reward signals in the reinforcement learning process (positive feedback such as successful dialogue completion and user satisfaction evaluation, and negative feedback such as dialogue interruption and transfer to human customer service) to optimize the policy function of the response generation network, enabling it to learn the strategy for generating the optimal response from the current dialogue state.

[0053] By employing the aforementioned dynamic weight adjustment method, this embodiment can intelligently adjust the learning focus of the collaborative optimization loop based on the different challenges faced by the intelligent dialogue system in actual operation. Specifically, when the system faces a large influx of new business or knowledge, by increasing the weight of the knowledge distillation loss, the model can more effectively absorb and integrate this new knowledge, avoiding biases in intent recognition or response generation due to untimely knowledge updates. Conversely, when users show high dissatisfaction with customer service responses, by increasing the weight of the strategy reward loss, the system can invest more learning resources in optimizing the decision-making strategy of the response generation network, thereby quickly correcting poor dialogue behavior and improving user satisfaction and dialogue success rate. This adaptive weight adjustment mechanism enables the model to maintain efficient learning capabilities and optimal decision-making abilities when facing dynamically changing business scenarios and user feedback, significantly improving the robustness and user experience of the intelligent dialogue system.

[0054] In some of the above implementations, the intelligent dialogue system determines whether the model needs self-evolution by monitoring whether the confidence level of intent recognition consistently falls below a first threshold, or whether the rate of conversion to human intervention in conversations related to the topic consistently exceeds a second threshold. However, in practical applications, these two indicators alone may not fully reflect user satisfaction with customer service responses or the model's ability to handle specific issues. Sometimes, the model may provide a response with high confidence, but this response may not actually solve the user's problem, forcing the user to initiate another inquiry. This repeated questioning phenomenon is not effectively captured by existing triggering mechanisms, thus delaying the model's discovery and optimization of potential problems.

[0055] In this regard, this embodiment further proposes that, in the self-evolution triggering and architecture adjustment steps, the key performance indicators also include the rate of repeated questions asked by the user again within a preset time window after the user has had a dialogue with the intelligent system. When this rate exceeds a specific threshold, it serves as an auxiliary judgment condition for triggering the self-evolution mechanism.

[0056] Specifically, the rate at which a user re-inquires about the same issue within a preset time window after an initial intelligent dialogue is a crucial indicator of user satisfaction with the customer service system's response or the degree of problem resolution. It reflects whether a user, after a dialogue ends, re-inquires about the same issue within a short period because the problem was not effectively resolved. To accurately calculate this re-inquiry rate, the system needs to be able to identify whether two inquiries pertain to the "same issue." This can be achieved in several ways. For example, semantic similarity analysis can be used to determine if the semantic content of two inquiries is highly consistent by calculating the cosine similarity of word vectors or sentence embedding models; keyword matching technology can be used to identify whether there are identical or highly related core keywords in two inquiries; or, by combining intent recognition results, if the intent classification results of two dialogues are the same or highly similar, and the core entity information is consistent, then it can be determined to be the same issue. Furthermore, the preset time window is a configurable parameter, such as 1 hour, 24 hours, or 72 hours. Only re-inquiries occurring within this time window are counted in the statistics to distinguish between genuine re-inquiries and different user needs for the same service at different times. The settings for this window should be optimized based on business experience and user behavior patterns. Within each statistical period (e.g., daily or weekly), the system will count the number of times a user repeatedly asks a question within a preset time window in all conversations, and divide this number by the total number of conversations or the total number of users to obtain the repeat question rate.

[0057] The specific threshold is a pre-set value used to determine whether the repeat question rate has reached a level that warrants triggering the self-evolution mechanism. This threshold is typically determined through analysis of historical data, accumulation of expert experience, or A / B testing. For example, if historical data shows that user satisfaction significantly decreases when the repeat question rate exceeds 5%, then 5% can be set as this threshold. The system continuously monitors the calculated repeat question rate and compares it to this threshold. When the repeat question rate exceeds this specific threshold, it indicates that the model may have potential shortcomings in handling this type of question.

[0058] Using the repeated question rate as an auxiliary criterion for triggering the self-evolution mechanism means that it is not the only triggering condition, but rather works in conjunction with other key performance indicators such as intent recognition confidence and conversion rate to human intervention. This auxiliary criterion mechanism can employ a logical "OR" relationship, meaning that self-evolution is triggered as long as any one of the triggering conditions is met; or it can use a more complex weighted or multi-condition combination, for example, triggering self-evolution only when the repeated question rate exceeds a threshold and the intent recognition confidence, while not low, is not exceptionally high. This mechanism allows the model to perceive its own performance deficiencies more comprehensively and precisely.

[0059] By introducing the rate of users repeatedly asking the same question again within a preset time window after a dialogue with the intelligent system as a key performance indicator, and using it as an auxiliary judgment condition to trigger the self-evolution mechanism, this embodiment can more sensitively capture the model's shortcomings in solving users' actual problems. Even if the confidence level of intent recognition is high or the rate of transferring to human agents has not increased significantly, if users frequently ask the same question repeatedly, it indicates that the responses provided by the model may not truly solve the user's pain points or the information is not clear enough. By incorporating this indicator into the self-evolution trigger condition, the system can promptly identify scenarios that appear to be working well but are actually ineffective, thereby more comprehensively and accurately evaluating model performance and promptly initiating the self-evolution mechanism for targeted optimization. This avoids users repeatedly consulting due to unresolved problems, significantly improving user experience and problem-solving efficiency.

[0060] In some of the embodiments described above, when the system detects that the confidence level of intent recognition for a specific emerging business topic is consistently lower than a first threshold, or the conversion rate of dialogues related to that topic to human intervention is consistently higher than a second threshold, a self-evolution mechanism is triggered. This mechanism dynamically expands the model's architecture to adapt to new scenarios by connecting a dedicated classification sub-network for the new topic in parallel after the intent classification network and adding a policy routing module corresponding to the sub-network before the response generation network. However, in practical applications, initial training data for emerging business topics is often scarce, and training a complex dedicated classification sub-network from scratch may require a long period and a large amount of computing resources, which can affect the efficiency of rapid integration and deployment of new functions.

[0061] In this regard, this embodiment further proposes that the dedicated classification sub-network is a lightweight network, and its training data mainly comes from samples obtained by data augmentation of corpora related to emerging business topics in historical dialogues.

[0062] Specifically, lightweight networks refer to neural network structures that, while maintaining certain performance, reduce the number of network layers, parameters, and computational complexity, resulting in smaller model size, faster inference speed, and lower computational resource requirements. For example, techniques such as depthwise separable convolution and grouped convolution can be used to construct this network, or network pruning and quantization methods can be used to optimize the network and reduce its computational and storage overhead. Furthermore, small pre-trained models such as MobileNet and ShuffleNet can be chosen as the backbone network; these models are themselves lightweight architectures designed for resource-constrained environments. By adopting lightweight networks, the training and inference resource consumption of dedicated classification sub-networks can be significantly reduced, thereby accelerating the deployment of new features.

[0063] Meanwhile, to address the issue of scarce data in the initial stages of emerging business topics, the training data for the dedicated classification sub-network is primarily obtained by augmenting historical dialogues related to these topics. Data augmentation is a method of generating new data by transforming existing data to expand the training dataset and improve the model's generalization ability and robustness. For example, text data augmentation techniques such as synonym replacement, random insertion / deletion / swapping of words, and back-translation can be used to generate new training samples. Furthermore, methods based on pre-trained language models, such as BERT and GPT, can be used to generate sentences that are semantically related to the original corpus but express different ideas. When selecting corpora, priority is given to selecting corpora from historical dialogues that are highly relevant to the current emerging business topics in terms of keywords, semantics, and context to ensure the effectiveness of the augmented data.

[0064] By employing a lightweight network as a dedicated classification subnetwork, the above technical solution significantly reduces the computational resources and time costs required for model training, enabling rapid deployment of the intent recognition module for emerging business topics. Simultaneously, data augmentation using historical dialogue corpora related to emerging business topics effectively addresses the challenge of data scarcity in the early stages of emerging business operations, providing ample and diverse training samples for the lightweight network. This ensures that, with limited data and resources, the dedicated classification subnetwork can efficiently learn and accurately identify new intents, thereby accelerating the intelligent dialogue system's response speed and adaptability to new business needs and enhancing the user experience.

[0065] In some of the embodiments described above, it is proposed to stream user feedback and dialogue logs generated online into a collaborative optimization loop to incrementally update model parameters on an hourly or daily basis, thereby achieving online real-time model iteration. However, in practical applications, the user feedback data generated online may contain a large amount of low-quality, unclear, or insufficiently informative content. If all data is used for model iteration updates without differentiation, noise may be introduced, leading to reduced model learning efficiency and even affecting the model's stability and accuracy.

[0066] In this regard, this embodiment further proposes that in the above-mentioned online real-time iteration step, the streaming input data will first pass through a feedback filter. The filter scores the feedback information based on the clarity and information content, and only sends high-quality feedback data with scores higher than the threshold into the model for iterative updates.

[0067] Specifically, the feedback filter is a module specifically designed to evaluate and filter streaming input data in online real-time iterative steps. Its core function is to ensure that only high-quality, valuable feedback data is used for incremental model updates, effectively preventing low-quality data from negatively impacting model performance. This feedback filter can be configured as a rule-based system, pre-setting a series of rules to identify low-quality feedback, such as determining if the text is too short, contains a large number of repetitive words, has an ambiguous sentiment, or is irrelevant to the current conversation topic. Alternatively, the feedback filter can be a machine learning-based model, pre-training a classifier to judge the quality of the feedback data, for example, using a text classification model to categorize feedback as "high-quality," "medium-quality," or "low-quality." Furthermore, the feedback filter can also incorporate natural language processing techniques to perform semantic analysis on the feedback text to assess its clarity, completeness, and information density.

[0068] The feedback filter assesses data quality by scoring feedback information based on clarity and informativeness. Clarity scoring ensures the feedback is easy to understand and can be quantified by evaluating grammatical correctness, lexical diversity, syntactic complexity, and the presence of ambiguous or vague expressions. For example, language models can be used to calculate text perplexity; lower perplexity generally indicates clearer text. Keyword matching and entity recognition can also be used to determine if the feedback explicitly addresses a specific problem or intent. Informativeness scoring ensures the feedback contains sufficient useful information and can be quantified by calculating the number of novel words, their relevance to the model's current knowledge base, whether they contain specific user needs or problem descriptions, and whether they provide actionable suggestions or error reports. For example, TF-IDF (Term Frequency-Inverse Document Frequency) or BERT embedding models can be used to measure the relevance of the feedback text to a specific topic or intent; higher relevance indicates greater informativeness. Furthermore, informativeness can be assessed by analyzing whether user feedback contains structured or semi-structured information such as specific error codes, operational steps, and expected results. Finally, a comprehensive score is obtained by weighting the clarity score and the information content score through a weighted average or other fusion method.

[0069] Based on this, the feedback filter only sends high-quality feedback data with scores above a threshold to the model for iterative updates. This means the system sets a predefined scoring threshold. When the feedback filter calculates a comprehensive score for a piece of feedback data that is higher than this threshold, the data is marked as "high-quality feedback" and allowed to enter the collaborative optimization loop for incremental learning or reinforcement learning. Data below the threshold is discarded or stored for manual review and does not participate in real-time iteration. The threshold can be determined through offline experiments, expert experience, or based on historical data analysis to balance the needs of data volume and data quality.

[0070] Through the aforementioned technical solutions, the system can effectively identify and filter out low-quality, unclear, or insufficiently informative user feedback data during online real-time iteration. This avoids introducing noise into the model training process, significantly improving the efficiency and accuracy of incremental model updates. The model learns only based on filtered, high-quality data, thus evolving more stably and reducing the risk of performance degradation or knowledge contamination caused by poor data. Ultimately, this enables the intelligent dialogue system to adapt to new business scenarios and user needs more quickly and reliably, continuously improving dialogue quality and user satisfaction.

[0071] In some of the embodiments described above, a method for generating and optimizing self-evolving models in real time for intelligent dialogue systems is proposed. This method enables the intelligent dialogue system to continuously learn new knowledge, adapt to new scenarios, and optimize decision-making strategies through steps such as model initialization, establishment of a collaborative optimization loop, self-evolution triggering and architecture adjustment, and online real-time iteration. However, in the process of the model continuously performing incremental learning and architecture adjustment to adapt to emerging business and user needs, there is a potential technical challenge: while learning new knowledge, the model may suffer from "catastrophic forgetting" of existing, basic business knowledge, leading to a decline in performance when handling core and common user problems, thereby affecting the overall service quality and user experience.

[0072] To address these challenges, this embodiment further proposes a model stability maintenance step. This step includes: periodically testing the iteratively evolved model using a validation set covering the basic business logic; triggering a model rollback mechanism when the model's accuracy in recognizing intent related to the basic business logic drops beyond the tolerance threshold; and initiating specialized reinforcement training to address catastrophic forgetting.

[0073] Specifically, the model stability maintenance steps aim to ensure that the intelligent dialogue system's ability to handle core business processes does not degrade as it continuously evolves and adapts to new scenarios. "Regularly testing the iteratively evolved model with a validation set covering core business processes" means that the system periodically (e.g., daily, weekly, or after each major model update) uses a pre-prepared set of dialogue data representing the system's core functions and common user queries to evaluate the current model's performance. These validation sets typically contain high-quality annotated intent recognition, entity extraction, and response generation test cases, accurately reflecting the model's baseline performance in handling core business processes. This regular testing allows for continuous monitoring of the model's performance on core business processes, enabling the timely detection of any signs of performance degradation.

[0074] "Triggering the model rollback mechanism when the model's accuracy in recognizing intent for basic business operations drops beyond the tolerance threshold" means that the system sets an acceptable performance degradation threshold, or "tolerance threshold." Once the model's accuracy in recognizing intent for basic business operations falls below this threshold, the system will determine that the current model may have a serious defect or has suffered a catastrophic failure, and immediately initiate the "model rollback mechanism." This mechanism will replace the currently deployed model version with a previously stable and validated older version, thereby quickly restoring the system's normal operation and service quality, and avoiding a continuous negative impact on user experience due to model defects.

[0075] "Initiating specialized reinforcement training for catastrophic forgetting" refers to the system launching a dedicated training process to address catastrophic forgetting issues after model rollback or performance degradation is detected. This specialized reinforcement training can employ various techniques. For example, through experience replay mechanisms, old data from basic business processes can be mixed with new data for training, or old data can be periodically used for small-scale training iterations to solidify the model's memory of fundamental knowledge. Additionally, regularization methods such as Elastic Weight Consolidation (EWC) or Learning Without Forgetting (LwF) can be used, penalizing changes to important parameters of old tasks by adding a regularization term to the loss function. Another approach is to utilize knowledge distillation techniques, using a stable "teacher model" to guide the current "student model's" learning, ensuring it maintains accuracy in learning old knowledge while acquiring new knowledge. These training methods aim to specifically consolidate the model's mastery of fundamental business knowledge, ensuring that the model can reliably and accurately handle core business processes while continuously adapting to new scenarios.

[0076] Through the above technical solution, this embodiment effectively solves the problem that during continuous self-evolution and iteration, the model may forget or weaken its ability to process basic business logic as it learns new knowledge. Regular testing with a validation set covering basic business logic can promptly detect performance degradation in core functions. Once an intention recognition accuracy drop exceeds a preset tolerance level, a model rollback mechanism is immediately triggered, quickly restoring system stability and service quality, and preventing long-term negative impacts on user experience due to model defects. Simultaneously, initiating specialized reinforcement training for catastrophic forgetting can specifically consolidate the model's mastery of basic business knowledge, ensuring that the model can reliably and accurately handle core business logic while continuously adapting to new scenarios, thereby improving the overall robustness and user satisfaction of the intelligent dialogue system.

[0077] In some of the above implementations, the intelligent dialogue system optimizes the policy function of the response generation network through reinforcement learning to learn a strategy for generating the optimal response from the current dialogue state. However, in practical applications, simply pursuing dialogue success or user satisfaction may cause the model to tend to provide direct answers, neglecting to guide users to use more efficient and convenient self-service channels. This may increase the pressure on human customer service and reduce the overall operational efficiency of the system.

[0078] To address this, this embodiment further proposes using a constrained policy optimization algorithm to optimize the policy function of the response generation network, and setting an additional reward in the reward function for successfully guiding users to use self-service.

[0079] Specifically, the policy function optimization of the response generation network employs a constrained policy optimization (CPO) algorithm. CPO is a reinforcement learning algorithm that, in optimizing the policy function, considers satisfying certain constraints in addition to maximizing the expected reward. These constraints can be related to safety, resource consumption, or specific behavioral norms, such as limiting the probability of the model generating inappropriate responses, controlling the number of dialogue turns, or ensuring the accurate delivery of key information. This algorithm restricts policy updates to a policy space that satisfies the constraints by introducing Lagrange multipliers or projection operations during the policy update process, thereby optimizing performance while ensuring policy stability and compliance. Simultaneously, an additional reward is set in the reward function for successfully guiding users to use self-service. The reward function is a key component in reinforcement learning that guides model learning; it defines the degree of "goodness" or "badness" of different actions taken by the model in different states. Adding an extra reward to successfully guide users to self-service means that when the response generation network generates a response that effectively directs users to self-service (e.g., providing self-service query links, guiding operation paths, recommending FAQ pages, etc.) and ultimately leads to the user's successful use of self-service, the model receives a higher positive reward than simply solving the problem. This extra reward can be a fixed positive value or dynamically adjusted based on factors such as the type of self-service and the user's efficiency in using it. For example, the reward for guiding a user to successfully complete a complex self-service operation may be higher than the reward for guiding a user to view a simple FAQ.

[0080] Through the aforementioned technical solutions, the intelligent dialogue system, when optimizing the policy function of the response generation network, not only considers the successful completion of the dialogue and user satisfaction, but also ensures the stability and compliance of policy updates by introducing a constrained policy optimization algorithm, avoiding unintended behaviors that might arise when the model pursues high rewards. Simultaneously, by setting an additional reward in the reward function for successfully guiding users to use self-service, the model can proactively explore and tend to generate responses that guide users to use self-service when learning the optimal response strategy. This effectively solves the problem that simply pursuing dialogue success might lead to the model over-reliance on direct answers and neglect of self-service guidance. In this way, the model can more intelligently balance users' immediate needs with system operational efficiency, improving user experience while effectively diverting pressure from human customer service, reducing operating costs, and promoting users' habit of solving problems themselves, thereby improving the overall service capability and efficiency of the intelligent dialogue system.

[0081] The following example will provide a more detailed explanation of the above technical solution: Suppose a large telecommunications operator's intelligent dialogue system initially handles traditional services such as broadband fault diagnosis, package inquiries, and phone bill top-ups. During the model initialization step, the system has built an initial intelligent dialogue model, which includes a feature extraction network, an intent classification network, and a response generation network. This model can effectively handle the aforementioned traditional services.

[0082] However, with the expansion of its operator business, the company launched a brand-new "Smart Home" service, covering emerging businesses such as smart home device installation, remote control, and home network optimization. Faced with these new services, the original intelligent dialogue model began to experience problems: when users inquired about "how to remotely unlock a smart door lock" or "what to do if home network coverage is poor," the model's intent recognition confidence consistently fell below the first threshold, and a large number of related conversations were transferred to human customer service, with the transfer rate consistently exceeding the second threshold. This indicates that the model lacks sufficient knowledge of the new services, and the fixed architecture struggles to effectively handle these new intents, highlighting the problems of lagging model updates, poor adaptability, and fixed architecture in existing technologies.

[0083] To address the aforementioned issues, this method establishes a collaborative optimization loop. In the incremental learning process, the system periodically collects new dialogue records between users and customer service regarding "smart home" services. For example, user A inquires about "smart camera installation steps," and customer service provides detailed instructions. These dialogue records are cleaned and labeled, extracting new intent features such as "smart camera installation" and "home network optimization," along with corresponding question-and-answer knowledge. Subsequently, the system updates the parameters of the feature extraction network and intent classification network through fine-tuning, thereby expanding and refining the model's knowledge base regarding "smart home" services.

[0084] Meanwhile, the reinforcement learning process treats each intelligent dialogue as a decision-making process. When user B successfully resolves the issue of "smart speaker unable to connect to the network" through intelligent customer service and gives a "satisfactory" rating, the system considers this a positive feedback reward. However, when user C inquires about "troubleshooting smart door locks," and the intelligent customer service fails to provide effective assistance, causing the dialogue to be interrupted or transferred to human customer service, the system considers this a negative feedback reward. Based on these reward signals, the system optimizes the policy function of the response generation network through a policy gradient algorithm, enabling it to learn the strategy for generating the optimal response from the current dialogue state. For example, when a user inquires about a smart device malfunction, the system prioritizes guiding the user to conduct self-troubleshooting and sets additional rewards for successfully guiding the user to use self-service.

[0085] The incremental learning and reinforcement learning processes share a co-trainer. When the system detects a rapid increase in the amount of new "smart home" dialogue data within a unit of time, the co-trainer dynamically increases the weight of the knowledge distillation loss to enhance the model's ability to learn new "smart home" knowledge. When the negative feedback rate of users regarding "smart home" services increases rapidly within a unit of time, the co-trainer increases the weight of the policy reward loss to optimize the response generation network's decision-making strategy in the "smart home" scenario. This collaborative mechanism avoids the problem of isolated applications of incremental learning and reinforcement learning in existing technologies, achieving a balance between new knowledge learning and decision-making strategy optimization.

[0086] When the system continuously monitors and finds that the confidence level of intent recognition for the topic of "smart home" is consistently below the first threshold, or that the rate of transferring conversations related to this topic to human agents is consistently above the second threshold, the system determines that the current model architecture is insufficient to handle the new topic and triggers the self-evolution mechanism. Furthermore, if the rate of repeated inquiries about the same "smart home" question from a user within a preset time window after a conversation with the intelligent system exceeds a specific threshold, this will also serve as an auxiliary condition for triggering the self-evolution mechanism.

[0087] The self-evolution mechanism works as follows: The original main structure of the model is frozen, and a dedicated classification sub-network for the emerging topic of "smart home" is connected in parallel after the intent classification network. This dedicated classification sub-network is a lightweight network, and its training data mainly comes from data augmentation samples obtained from historical dialogues related to the emerging business topic of "smart home." Simultaneously, a policy routing module corresponding to this sub-network is added before the response generation network. When the system identifies that the user's intent belongs to the "smart home" domain, the policy routing module directs the request to the newly connected dedicated classification sub-network for more refined intent recognition and generates a response based on the results. This dynamically expanded model architecture effectively solves the problem of insufficient generalization ability caused by the fixed architecture of existing models, enabling the model to adapt to entirely new business scenarios.

[0088] Subsequently, the system establishes an online learning loop and enters the online real-time iteration step. Real-time user feedback generated online (such as user satisfaction ratings for "Smart Home" responses and click behaviors) and dialogue logs are streamed into the collaborative optimization loop. This streamed input data first passes through a feedback filter, which scores the feedback information based on clarity and information content, sending only high-quality feedback data with scores above a threshold to the model for iterative updates. Model parameters are incrementally updated hourly or daily to ensure the model can quickly correct errors and capture changes in user preferences, solving the problems of feedback latency and long iteration cycles in existing technologies. Simultaneously, for the newly added "Smart Home" dedicated classification subnetwork and policy routing module generated in the self-evolution step, the system uses recent highly relevant "Smart Home" dialogue data for centralized training to achieve rapid integration and deployment of new functions.

[0089] To maintain model stability, the system periodically tests the iteratively evolved model using a validation set covering basic telecommunications services. When the model's accuracy in recognizing intent related to basic telecommunications services drops beyond the tolerance threshold, the system triggers a model rollback mechanism and initiates specialized reinforcement training to address catastrophic forgetting, ensuring that the model learns new knowledge without forgetting its original core business capabilities.

[0090] Furthermore, embodiments of this application also propose a computer-readable storage medium storing a program for generating and optimizing a self-evolving model for an intelligent dialogue system in real time. When the program for generating and optimizing a self-evolving model for an intelligent dialogue system is executed by a processor, it implements the steps of the method for generating and optimizing a self-evolving model for an intelligent dialogue system as described above.

[0091] Reference Figure 3 , Figure 3 This is a structural block diagram of the first embodiment of the self-evolving model generation and real-time optimization system for intelligent dialogue systems according to this application.

[0092] like Figure 3 As shown in the embodiments of this application, the self-evolving model generation and real-time optimization system for intelligent dialogue systems includes: The model initialization module 10 is used to build an initial intelligent dialogue model, which includes a feature extraction network, an intent classification network, and a response generation network. The collaborative optimization loop establishment module 20 is used to establish a collaborative optimization mechanism for incremental learning and reinforcement learning. The incremental learning process is configured to: periodically collect newly generated user-customer service dialogue records, clean and label them, extract new intent features and question-and-answer knowledge, and fine-tune the parameters of the feature extraction network and intent classification network to expand and correct the model's knowledge base. The reinforcement learning process is configured to: treat each intelligent dialogue as a decision-making process, using successful completion of the dialogue and user satisfaction evaluation as positive feedback rewards, and interruption of the dialogue by the user or transfer to human customer service as negative feedback rewards. Based on these reward signals, the policy function of the response generation network is optimized through a policy gradient algorithm, enabling it to learn the policy for generating the optimal response from the current dialogue state. The incremental learning process and the reinforcement learning process share a collaborative trainer, which dynamically adjusts the weights of knowledge distillation loss and policy reward loss to balance the model's ability to learn new knowledge and its optimal decision-making ability based on existing knowledge. The self-evolution triggering and architecture adjustment module 30 is used for real-time monitoring of key performance indicators of the system. When the confidence level of intent recognition for a specific emerging business topic is continuously lower than the first threshold, or the conversion rate of dialogues related to the topic to human agents is continuously higher than the second threshold, it is determined that the current model architecture is insufficient to handle the new topic and triggers the self-evolution mechanism. The execution method of the self-evolution mechanism is as follows: freeze the original main structure of the model, connect a dedicated classification sub-network for the new topic in parallel after the intent classification network, and add a policy routing module corresponding to the sub-network before the response generation network, thereby dynamically expanding the architecture of the model to adapt to the new scenario. The online real-time iteration module 40 is used to establish an online learning closed loop; it streams user feedback and dialogue logs generated online into the collaborative optimization loop to perform incremental updates of model parameters on an hourly or daily basis; at the same time, for newly added network modules generated in the self-evolution step, it uses recent highly relevant dialogue data for centralized training to achieve rapid integration and deployment of new functions.

[0093] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solution of this application. In specific applications, those skilled in the art can make settings as needed, and this application does not impose any restrictions on this.

[0094] This embodiment solves the problems of lagging model updates, isolated learning mechanisms, rigid architecture, and iteration delays by establishing a collaborative optimization mechanism, dynamically expanding the model architecture, and realizing real-time iterative updates. It has the advantages of improving the system's dynamic adaptability, collaborative optimization capabilities, and real-time response efficiency.

[0095] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this application. In practical applications, those skilled in the art can select some or all of it to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0096] In addition, for technical details not described in detail in this embodiment, please refer to the method for generating and optimizing a self-evolving model for an intelligent dialogue system provided in any embodiment of this application, which will not be repeated here.

[0097] Furthermore, it should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0098] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0099] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this application. The above are only preferred embodiments of this application and do not limit the patent scope of this application. All equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for generating and optimizing a self-evolving model in real time for an intelligent dialogue system, characterized in that, include: Model initialization steps: Build an initial intelligent dialogue model, which includes a feature extraction network, an intent classification network, and a response generation network; The steps for establishing a collaborative optimization loop are as follows: A collaborative optimization mechanism for incremental learning and reinforcement learning is established. The incremental learning process is configured as follows: Newly generated user-customer service dialogue records are periodically collected, cleaned, and labeled. New intent features and question-and-answer knowledge are extracted, and the parameters of the feature extraction network and intent classification network are updated in a fine-tuning manner to expand and correct the model's knowledge base. The reinforcement learning process is configured as follows: Each intelligent dialogue is treated as a decision-making process. Successful completion of the dialogue and user satisfaction evaluation are used as positive feedback rewards, while the interruption of the dialogue by the user or transfer to human customer service is used as negative feedback rewards. Based on these reward signals, the policy function of the response generation network is optimized using a policy gradient algorithm, enabling it to learn the policy for generating the optimal response from the current dialogue state. The incremental learning process and the reinforcement learning process share a collaborative trainer, which balances the model's ability to learn new knowledge and its optimal decision-making ability based on existing knowledge by dynamically adjusting the weights of knowledge distillation loss and policy reward loss. Self-evolution triggering and architecture adjustment steps: The system monitors key performance indicators in real time; when the confidence level of intent recognition for a specific emerging business topic is consistently lower than the first threshold, or the conversion rate of dialogues related to the topic to human agents is consistently higher than the second threshold, it is determined that the current model architecture is insufficient to handle the new topic, and the self-evolution mechanism is triggered; the execution method of the self-evolution mechanism is as follows: freeze the original main structure of the model, connect a dedicated classification sub-network for the new topic in parallel after the intent classification network, and add a policy routing module corresponding to the sub-network before the response generation network, thereby dynamically expanding the architecture of the model to adapt to the new scenario; Online real-time iteration steps: Establish an online learning closed loop; stream user feedback and dialogue logs generated online into the collaborative optimization loop to perform hourly or daily incremental updates to the model parameters; simultaneously, for newly added network modules generated in the self-evolution step, use recent highly relevant dialogue data for centralized training to achieve rapid integration and deployment of new functions.

2. The method according to claim 1, characterized in that, In the collaborative optimization loop establishment step, the method for dynamically adjusting the weights is as follows: when the amount of new dialogue data increases rapidly within a unit of time, the weight of knowledge distillation loss is increased to enhance new knowledge learning; when the user negative feedback rate increases rapidly within a unit of time, the weight of strategy reward loss is increased to optimize the decision-making strategy.

3. The method according to claim 1, characterized in that, In the self-evolution triggering and architecture adjustment steps, the key performance indicators also include the rate of repeated questions asked by users within a preset time window after a dialogue with the intelligent system. When this rate exceeds a specific threshold, it serves as an auxiliary judgment condition for triggering the self-evolution mechanism.

4. The method according to claim 1, characterized in that, In the self-evolution triggering and architecture adjustment steps, the dedicated classification sub-network is a lightweight network whose training data mainly comes from samples obtained by data augmentation of corpora related to emerging business topics in historical dialogues.

5. The method according to claim 1, characterized in that, In the online real-time iteration step, the streaming input data first passes through a feedback filter. This filter scores the feedback information based on its clarity and information content, and only sends high-quality feedback data with scores higher than a threshold into the model for iterative updates.

6. The method according to claim 1, characterized in that, The method also includes a model stability maintenance step: periodically testing the iteratively evolved model with a validation set covering the basic business; when the model’s accuracy in recognizing the intent of the basic business drops beyond the tolerance level, triggering a model rollback mechanism and initiating special reinforcement training for catastrophic forgetting.

7. The method according to claim 1, characterized in that, The policy function optimization of the response generation network adopts a constrained policy optimization algorithm, and an additional reward is set in the reward function for successfully guiding users to use self-service.

8. A self-evolving model generation and real-time optimization system for intelligent dialogue systems, characterized in that, include: The model initialization module is used to build an initial intelligent dialogue model, which includes a feature extraction network, an intent classification network, and a response generation network. A collaborative optimization loop establishment module is used to establish a collaborative optimization mechanism for incremental learning and reinforcement learning. The incremental learning process is configured to: periodically collect newly generated user-customer service dialogue records, clean and label them, extract new intent features and question-and-answer knowledge, and fine-tune the parameters of the feature extraction network and intent classification network to expand and correct the model's knowledge base. The reinforcement learning process is configured to: treat each intelligent dialogue as a decision-making process, using successful completion of the dialogue and user satisfaction evaluation as positive feedback rewards, and interruption of the dialogue by the user or transfer to human customer service as negative feedback rewards. Based on these reward signals, the policy function of the response generation network is optimized through a policy gradient algorithm, enabling it to learn the policy for generating the optimal response from the current dialogue state. The incremental learning process and the reinforcement learning process share a collaborative trainer, which dynamically adjusts the weights of knowledge distillation loss and policy reward loss to balance the model's ability to learn new knowledge and its optimal decision-making ability based on existing knowledge. The self-evolution triggering and architecture adjustment module is used for real-time monitoring of key performance indicators of the system. When the confidence level of intent recognition for a specific emerging business topic is continuously lower than the first threshold, or the conversion rate of dialogues related to the topic to human agents is continuously higher than the second threshold, it is determined that the current model architecture is insufficient to handle the new topic, and the self-evolution mechanism is triggered. The execution method of the self-evolution mechanism is as follows: freeze the original main structure of the model, connect a dedicated classification sub-network for the new topic in parallel after the intent classification network, and add a policy routing module corresponding to the sub-network before the response generation network, thereby dynamically expanding the architecture of the model to adapt to the new scenario. An online real-time iteration module is used to establish an online learning loop; real-time user feedback and dialogue logs generated online are streamed into the collaborative optimization loop to perform incremental updates of model parameters on an hourly or daily basis; at the same time, newly added network modules generated in the self-evolution step are trained intensively using recent highly relevant dialogue data to achieve rapid integration and deployment of new functions.

9. A computer device, characterized in that, The device includes a memory and a processor, wherein the processor, when executing computer instructions stored in the memory, performs the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Includes instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1 to 7.

Images