Model optimization method, device and equipment for intelligent comprehensive adjustment customer service and storage medium

Abstract

A model optimization method, device, equipment and storage medium for an intelligent comprehensive adjustment customer service are applied to the technical field of artificial intelligence, and the method comprises the steps of obtaining key index data corresponding to a plurality of key indexes in session data under the condition of determining that a session ends, and archiving the session data and the key index data to an optimization data pool; according to a preset optimization period and the data in the optimization data pool, obtaining a judgment result of whether to execute model optimization; and if the judgment result is that model optimization needs to be executed, performing model optimization according to the optimization data pool and LoRA fine tuning driven by Bayesian optimization. Automatic and continuous optimization of the knowledge base is realized, and subsequent rapid processing of related problems is ensured. The accuracy and operability of evaluation are comprehensively improved; and the time cost and the computing resource demand of model optimization are effectively reduced. And the self-adaptive capability and the optimization capability of the intelligent comprehensive adjustment customer service in a dynamic environment are improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to a model optimization method, apparatus, device, and storage medium for intelligent integrated customer service. Background Technology

[0002] Currently, the application of intelligent customer service systems in the field of artificial intelligence (e.g., home broadband intelligent agents) in the telecommunications industry mainly relies on pre-trained language models combined with Natural Language Processing (NLP) technology to complete user intent recognition and response generation. Common methods include keyword matching, semantic parsing, and rule-guided fixed process orchestration. In traditional solutions, intelligent customer service systems categorize user questions based on static intent classification models and respond using fixed task instructions or knowledge base information, lacking dynamic adaptability. Some systems integrate simple task scheduling mechanisms or rule-based priority adjustment methods to improve interaction efficiency. However, in practical applications, facing the diversity of user questions and the complexity of network environments, existing technologies generally suffer from high intent recognition deviation rates, delayed interaction responses, frequent misleading answers, and poor system stability. Furthermore, the model optimization process for intelligent customer service systems often uses fixed fine-tuning parameters and manually specified evaluation index weights, making it difficult to balance multi-dimensional performance. Although existing solutions have attempted to introduce reinforcement learning optimization strategies, they typically lack a dynamic optimization mechanism based on periodic evaluation metrics and have not built a unified evaluation-optimization closed-loop system. This results in untimely iteration of large models and a lack of systematic evaluation feedback, which limits the improvement of service quality of intelligent integrated customer service in complex scenarios.

[0003] The main problems and shortcomings of the existing intelligent customer service model optimization are as follows:

[0004] 1. Insufficient dynamic adaptability: Most existing intelligent integrated customer service model optimizations and evaluations are based on static parameter configurations, lacking the ability to adapt to different network environments, user intentions and service scenarios in real time, resulting in performance degradation under complex or sudden situations.

[0005] 2. Limited Evaluation System: Existing evaluations of intelligent customer service systems often focus on single indicators (such as response time or task completion rate), lacking a multi-dimensional and comprehensive evaluation indicator system, making it difficult to fully reflect the service quality of intelligent customer service systems.

[0006] 3. Low optimization efficiency and reliance on manual intervention: Current optimization mechanisms for intelligent integrated customer service (such as model updates and knowledge base maintenance) often rely on manual triggering and intervention, lacking the ability to automatically trigger and optimize based on performance data, and unable to dynamically adjust weights to adapt to real-time demand changes. Summary of the Invention

[0007] At least one embodiment of this application provides a model optimization method, apparatus, device, and storage medium for intelligent integrated customer service, which addresses the problems of insufficient dynamic adaptability, single evaluation system, and low optimization efficiency in the prior art.

[0008] To solve the above-mentioned technical problems, this application is implemented as follows:

[0009] In a first aspect, embodiments of this application provide a model optimization method for intelligent integrated customer service, including:

[0010] When the session is determined to be over, the key indicator data corresponding to multiple key indicators in the session data is obtained, and the session data and the key indicator data are archived to the optimization data pool. The key indicators include at least one of the following: task completion rate, intent deviation rate, misleading answer rate, crash rate, user intervention rate, and user satisfaction.

[0011] Based on the preset optimization cycle and the data in the optimization data pool, obtain the judgment result of whether to perform model optimization;

[0012] If the determination result indicates that model optimization is required, then model optimization is performed based on the optimized data pool and Bayesian optimization-driven low-rank adaptation (LoRA) fine-tuning.

[0013] Specifically, the method described above determines the end of the session if at least one of the following conditions is met:

[0014] The user input was detected as a session end message;

[0015] No user input was detected within the preset time period.

[0016] Preferably, in the method described above, obtaining the judgment result on whether to perform model optimization based on the preset optimization period and the data in the optimization data pool includes:

[0017] If any of the aforementioned optimization cycles are reached within a given timeframe, obtain the comprehensive service evaluation score for the current optimization cycle based on the aforementioned key indicators.

[0018] Based on the multi-dimensional operating indicators obtained by the dynamic perception layer, the optimization trigger threshold for the current optimization cycle is obtained;

[0019] If the overall service evaluation score is less than or equal to the optimization trigger threshold, then the judgment result is determined to be that model optimization needs to be performed; otherwise, the judgment result is determined to be that model optimization does not need to be performed.

[0020] Preferably, in the method described above, obtaining the comprehensive service evaluation score for the current optimization cycle based on the key indicators includes:

[0021] Based on the data type of the key indicators, post-processing is performed to obtain the indicator values ​​corresponding to each key indicator. The post-processing includes at least one of the following: inverting the inverse meaning of indicators, outlier detection and correction, and unified dimension normalization.

[0022] The comprehensive service evaluation score is obtained based on the index values ​​corresponding to each of the key indicators.

[0023] Preferably, in the method described above, obtaining the comprehensive service evaluation score based on the indicator values ​​corresponding to each of the key indicators includes:

[0024] Construct a comprehensive service evaluation decision vector and a multi-objective function for the key indicators;

[0025] The multi-objective function is sorted using non-dominated methods to obtain multiple non-dominated solution sets;

[0026] The comprehensive service evaluation score is calculated based on the non-dominated solution set, the first preset algorithm, and the initialized reference point.

[0027] Preferably, in the method described above, obtaining the optimization trigger threshold for the current optimization cycle based on the multi-dimensional operating indicators obtained by the dynamic perception layer includes:

[0028] Construct the state space of the current optimization cycle based on the multidimensional operating indicators;

[0029] The optimization trigger threshold of the previous optimization cycle is adjusted according to the state space to obtain the optimization trigger threshold of the current optimization cycle.

[0030] Specifically, as described above, the model optimization based on the optimized data pool and Bayesian optimization-driven LoRA fine-tuning includes:

[0031] The session data in the optimized data pool is preprocessed, and the training set and validation set are determined;

[0032] The objective function of the Bayesian optimization is modeled based on the Gaussian process, and the objective function is the fine-tuned performance gain.

[0033] Construct a joint hyperparameter space that includes LoRA structural parameters and hyperparameters related to fine-tuning;

[0034] Based on the Bayesian optimization, an iterative search is performed in the hyperparameter space to obtain the target hyperparameter combination;

[0035] The model is optimized based on the target hyperparameter combination, and the model is fine-tuned, deployed, and replaced.

[0036] Preferably, in the method described above, the step of iteratively searching within the hyperparameter space according to the Bayesian optimization to obtain the target hyperparameter combination includes:

[0037] Initialize the objective function of Gaussian regression, and initialize the acquisition function of the Bayesian optimization;

[0038] Hierarchical sampling is performed in the hyperparameter space to obtain the optimized sample set for the current iteration, where each sample in the optimized sample set is a hyperparameter combination;

[0039] Perform LoRA fine-tuning training once for each sample in the optimized sample set to obtain the corresponding target performance gain;

[0040] Gaussian regression probability modeling is performed based on the optimized sample set;

[0041] Obtain the function value of the acquisition function corresponding to each sample in the optimized sample set, and determine the target sample corresponding to the maximum value among the function values;

[0042] If the number of iterations corresponding to the current iteration is greater than the preset number of iterations threshold, then the target sample is determined to be the target hyperparameter combination; otherwise, the number of iterations is incremented by one, and the process returns to the step of performing stratified sampling in the hyperparameter space to obtain the optimized sample set for the current iteration.

[0043] Secondly, embodiments of this application provide a control device, including:

[0044] The first processing module is used to, upon determining that the session has ended, acquire key indicator data corresponding to multiple key indicators from the session data, and archive the session data and the key indicator data into an optimized data pool. The key indicators include at least one of the following: task completion rate, intent deviation rate, misleading answer rate, crash rate, user intervention rate, and user satisfaction.

[0045] The third processing module is used to obtain a judgment result on whether to perform model optimization based on the preset optimization cycle and the data in the optimization data pool;

[0046] The fourth processing module is used to perform model optimization based on LoRA fine-tuning driven by Bayesian optimization if the judgment result indicates that model optimization is required.

[0047] Thirdly, embodiments of this application provide an electronic device, including: a processor, a memory, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, it implements the steps of the intelligent integrated customer service model optimization method as described above.

[0048] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the intelligent integrated customer service model optimization method described above.

[0049] Fifthly, embodiments of this application provide a computer program product, including computer instructions that, when executed by a processor, implement the steps of the intelligent integrated customer service model optimization method as described above.

[0050] Compared with existing technologies, the intelligent integrated customer service model optimization method, apparatus, device, and storage medium provided in this application can achieve automatic and continuous optimization of the knowledge base by archiving new solutions to the knowledge base, ensuring rapid processing of related issues in the future. Through comprehensive evaluation based on multiple key session indicators, the accuracy and operability of the evaluation are comprehensively improved; by using a combined Bayesian optimization and LoRA efficient fine-tuning method to efficiently optimize the model and hyperparameters, the time cost and computational resource requirements for model optimization are effectively reduced. This reduces the cost of system maintenance and upgrades, and enhances the adaptive and optimization capabilities of the intelligent integrated customer service in dynamic environments, significantly reducing tuning costs. Attached Figure Description

[0051] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0052] Figure 1 This is one of the flowcharts illustrating the model optimization method for the intelligent integrated customer service system in this application;

[0053] Figure 2 This is a flowchart illustrating the conversation steps of the intelligent integrated customer service system in this application.

[0054] Figure 3 This is the second flowchart illustrating the model optimization method for the intelligent integrated customer service system in this application.

[0055] Figure 4 This is the third flowchart illustrating the model optimization method for the intelligent integrated customer service system in this application;

[0056] Figure 5 This is the fourth flowchart illustrating the model optimization method for the intelligent integrated customer service system in this application.

[0057] Figure 6 This is the fifth flowchart illustrating the model optimization method for the intelligent integrated customer service system in this application.

[0058] Figure 7 This is the sixth flowchart illustrating the model optimization method for the intelligent integrated customer service system in this application;

[0059] Figure 8 This is the seventh flowchart illustrating the model optimization method for the intelligent integrated customer service system in this application.

[0060] Figure 9 This is a schematic diagram of the control device in this application;

[0061] Figure 10 This is a schematic diagram of the structure of the electronic device in this application. Detailed Implementation

[0062] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the scope of the present application to those skilled in the art.

[0063] The terms “first,” “second,” etc., used in the specification and claims of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus. The terms “and / or” in the specification and claims indicate at least one of the connected objects.

[0064] See Figure 1 This application provides a model optimization method for intelligent integrated customer service, including:

[0065] Step S101: If the session is determined to be over, obtain key indicator data corresponding to multiple key indicators from the session data, and archive the session data and the key indicator data to the optimization data pool, and / or archive the new solutions in the session data to the knowledge base. The key indicators include at least one of the following: task completion rate, intent deviation rate, misleading answer rate, crash rate, user intervention rate, and user satisfaction.

[0066] Step S102: Based on the preset optimization period and the data in the optimization data pool, obtain the judgment result of whether to perform model optimization;

[0067] Step S103: If the judgment result indicates that model optimization is required, then model optimization is performed based on the optimized data pool and the LoRA fine-tuning driven by Bayesian optimization.

[0068] In the technical solution provided in this embodiment, the intelligent customer service system monitors the information and number of interactions with users. If the session ends, it acquires key indicator data corresponding to multiple key metrics from the session data and archives it to the optimization data pool. Alternatively, it archives new solutions from the session data to the knowledge base, facilitating performance evaluation from multiple perspectives and model optimization. Specifically, if the session data includes new solutions (e.g., obtained through manual intervention), these are summarized through a prompting process and archived to the knowledge base to ensure rapid handling of related issues. This allows for automatic and continuous optimization of the knowledge base through prompting engineering. Using prompting engineering as its core, it dynamically collects knowledge fragments from user interactions and automatically generates or optimizes prompt content based on dialogue intent and task performance, achieving iterative accumulation of knowledge within the prompt structure. By constructing a knowledge update strategy and prompt template linkage mechanism, the intelligent customer service system can adaptively adjust the knowledge base without manual intervention, maintaining long-term consistency and optimality between knowledge content and prompt behavior. It effectively overcomes the shortcomings of traditional knowledge bases, such as reliance on manual maintenance, delayed updates, and disconnection from models, providing key support for building an intelligent integrated customer service system with continuous evolution capabilities.

[0069] In one specific embodiment, the aforementioned key indicators include at least one of the following: task completion rate, intent deviation rate, misleading response rate, crash rate, user intervention rate, and user satisfaction.

[0070] Specifically, Task Completion Rate (TCR) measures whether the intelligent customer service system has successfully completed the user's assigned task objective. For each task interaction, the intelligent customer service system automatically calculates the total number of tasks. The number of tasks completed in the final closed loop with the intelligent integrated customer service system. The task completion rate is: .

[0071] Intent Drift Rate (IDR): This represents the frequency with which the user's intent recognition changes during multi-turn dialogues, reflecting the stability of the model's contextual understanding. The intelligent customer service system utilizes a dedicated "intent classification sub-model" to categorize user input in each turn. It uses semantic similarity calculations to determine whether the intent of two consecutive turns of user input (e.g., turn i and turn i+1) should remain consistent. If the similarity is high (e.g., cosine similarity > 0.7) but the intent labels are different, the system counts the number of turns in which intent recognition drift occurs. Add 1, the total number of rounds in the dialogue is 1. The intention offset rate is: .

[0072] Misleading Response Rate (MRR): This measures the proportion of responses generated by the model that contain hallucinatory, logically flawed, or factual errors. The intelligent customer service system automatically breaks down responses into "semantic + factual + business" triples and matches them against a pre-built, broadened knowledge base. If no match is found or the similarity is too low, the number of model responses containing hallucinatory or misleading content is counted. Add 1, the total number of responses for the model is The misleading response rate is: .

[0073] Crash Rate (CR): In a period of time Within a preset optimization period, the intelligent integrated customer service system will automatically count the number of system crashes. The crash rate, defined as the percentage of times the intelligent customer service system experiences malfunctions or interruptions during processing, is the system crash rate. The system crash rate is: .

[0074] User Intervention Rate (UIR): This measures the percentage of users who actively interrupt a conversation and request human intervention. Throughout the conversation, if the automated customer service system is unable to provide correct and useful answers, the customer will proactively interrupt the interaction and request human intervention. The number of sessions in which this occurs is counted. Add 1. Intelligent integrated customer service operates within a cycle. The total number of sessions processed internally is User intervention rate: .

[0075] User Satisfaction (US): User satisfaction scores are evaluated using a Likert scale. The rating for the intelligent integrated customer service interaction is: In one cycle Within the system, users' overall rating for the intelligent customer service interaction was [rating missing]. User satisfaction (US) is: , ,in, It is a certain cycle Total number of rounds of internal interaction.

[0076] Subsequently, based on the preset optimization cycle and the data in the optimization data pool, a judgment is made to determine whether model optimization is needed. If the judgment result is that model optimization is not needed, it means that the accuracy and other aspects of the current intelligent customer service model meet the requirements, reducing unnecessary model optimization and helping to avoid resource waste. If the judgment result is that model optimization is needed, model optimization is performed using Bayesian optimization-driven LoRA fine-tuning based on the data in the optimization data pool. Since the data in this optimization data pool includes key indicator data corresponding to multiple key indicators, the model can be evaluated based on a comprehensive evaluation system of multiple key indicators of the conversation, comprehensively improving the accuracy and operability of the evaluation. Furthermore, by combining Bayesian optimization and LoRA's efficient fine-tuning method, the model and hyperparameters are optimized efficiently, effectively reducing the time cost and computational resource requirements of model optimization. Through the automatic optimization process, system iteration can be completed without a lot of manual intervention, avoiding traditional offline training, reducing the cost of system maintenance and upgrades, and supporting rapid adaptation to real-world scenarios, improving the intelligent customer service's adaptability and optimization capabilities in dynamic environments, and significantly reducing tuning costs.

[0077] It should be noted that, for the convenience of those skilled in the art, the conversation steps before determining the end of the conversation are illustrated here. When the intelligent customer service receives a conversation initiated by a user, it determines the scenario and provides feedback based on its own capabilities, and monitors the number of interactions in the conversation. If the number of interactions within the same scenario reaches the interaction threshold corresponding to that scenario, it is determined that the current conventional processing method cannot solve the user's problem. Therefore, the interaction count is initialized, and a more effective processing method is used for feedback. In this embodiment, a chain-like response mechanism based on the linkage of a knowledge base and a Model Context Protocol (MCP) is preferred to regenerate the solution. This chain-like response mechanism first relies on the knowledge base to understand the user's intent and provides preliminary solution suggestions. Then, combined with the real-time acquisition of user context information or triggering of external actions by the MCP platform, multi-step responses are generated through chain-like reasoning to gradually advance the problem-solving process.

[0078] See Figure 2 In one specific embodiment, the above-described session steps can be specifically represented as follows.

[0079] Step S201: After receiving user input, the user input is preprocessed to obtain preprocessed user input. Specifically, based on the form of the user input, a corresponding text conversion model is used to recognize the user input as text input. User input forms include, but are not limited to, voice, images, and / or text. Text conversion models include, but are not limited to, speech-to-text models for speech recognition (with a conversion accuracy exceeding 90%) and optical character recognition (OCR) models for extracting text from images (with a conversion accuracy exceeding 85%). In the home broadband field, user input issues include, but are not limited to, unstable network cards, equipment failures, renewals, and equipment replacements.

[0080] Step S202: Based on the preprocessed user input, user intent is identified, and the corresponding scenario is determined. Specifically, information features are extracted using the text semantic feature extraction module and scene feature encoder of an intent recognition model (such as a Long Short-Term Memory (LSTM) network). Then, the scenario vector generation module of the intent recognition model is called to form a unified scenario vector from the information features. Finally, the scenario is classified using the intent classification module of the intent recognition model to identify the user's intent, and the large model in the integrated customer service system is called to enter the scenario closest to the customer's input, such as: network anomaly feedback, equipment replacement application, renewal issue consultation, fault diagnosis request, business processing consultation, etc.

[0081] Step S203: After entering the matching scenario, provide feedback based on the user's input and guide the user to supplement relevant environmental information. For example, in the broadband diagnostic scenario, it may be necessary to guide the customer to input information such as account, device code, and real-time image of the router.

[0082] Step S204: Based on the scenario and the environmental information provided by the user, comprehensively judge the user's needs and provide the user with corresponding solutions based on its own capabilities and knowledge base;

[0083] If step S205 detects that the number of interactions of a user in the same scenario reaches the corresponding interaction threshold, then initialize the number of interactions and regenerate the solution based on the chain-like response mechanism of knowledge base and MCP linkage.

[0084] Furthermore, the steps for regenerating solutions based on the thought chain response mechanism linked by the knowledge base and the Model Context Protocol (MCP) also include:

[0085] If the number of interactions is detected to reach the interaction threshold again, and it is determined that the user's problem has not been resolved, a manual intervention mode is triggered. The manual intervention mode is a collaborative problem-solving mode involving the user, the comprehensive coordination personnel, and the intelligent comprehensive coordination customer service.

[0086] Alternatively, if a user's request for manual intervention is detected, the manual intervention mode is triggered.

[0087] In other words, when regenerating a solution based on a chain-of-thinking response mechanism, it determines whether the user needs human intervention. This involves switching to a human intervention mode where the user, coordination personnel, and intelligent customer service collaborate to resolve the issue, ensuring timely resolution. The determination prioritizes whether the user has requested human intervention, which can be determined through intent recognition of user input or by detecting virtual key presses related to human intervention. Furthermore, it considers the progress of the problem's resolution. If the user's problem remains unresolved after reaching a certain interaction threshold, human intervention is deemed necessary. It's important to note that when human intervention is required, upon confirmation of the session's end, key performance indicator (KPI) data from the session in this human intervention mode is also acquired and archived into the optimization data pool.

[0088] Specifically, the method described above determines the end of the session if at least one of the following conditions is met:

[0089] The user input was detected as a session end message;

[0090] No user input was detected within the preset time period.

[0091] This embodiment provides an example of how to determine the end of a session, which includes at least one of the following: detecting again that the number of interactions has reached the interaction threshold, detecting that the user input is session end information, and not detecting the user input within a preset time period.

[0092] It should be noted that the session end information is generated by the intelligent dispatching customer recognizing the need to end the session based on user input, and is not limited to user input of "end session" or operation of virtual buttons to end the session.

[0093] See Figure 3 Preferably, in the method described above, obtaining the judgment result on whether to perform model optimization based on the preset optimization period and the data in the optimization data pool includes:

[0094] Step S301: If the time reaches any of the optimization cycles, obtain the comprehensive service evaluation score for the current optimization cycle based on the key indicators.

[0095] Step S302: Obtain the optimization trigger threshold for the current optimization cycle based on the multi-dimensional operating indicators obtained by the dynamic perception layer;

[0096] Step S303: If the comprehensive service evaluation score is less than or equal to the optimization trigger threshold, then the judgment result is determined to be that model optimization needs to be performed; otherwise, the judgment result is determined to be that model optimization does not need to be performed.

[0097] In this embodiment, when determining whether to perform model optimization, the process first checks if the preset optimization period has been reached. If not, model optimization is not performed. If any optimization period has been reached, a comprehensive service evaluation score is obtained based on the aforementioned key indicators. Then, an optimization trigger threshold corresponding to the current optimization period is determined based on the multi-dimensional operational indicators obtained from the dynamic perception layer. If the comprehensive service evaluation score is less than or equal to the optimization trigger threshold, model optimization is determined; otherwise, it is determined that model optimization is not required. This allows for a multi-faceted evaluation of the intelligent customer service system using multiple key indicators of the session, comprehensively improving the accuracy, adaptability, and operability of the evaluation, and significantly enhancing the overall performance evaluation capability of the intelligent customer service system in complex scenarios.

[0098] See Figure 4 Preferably, in the method described above, obtaining the comprehensive service evaluation score for the current optimization cycle based on the key indicators includes:

[0099] Step S401: According to the data type of the key indicators, post-processing is performed respectively to obtain the indicator value corresponding to each key indicator. The post-processing includes at least one of the following: inverting the inverse meaning indicator, outlier detection and correction, and unified dimension normalization.

[0100] Step S402: Obtain the comprehensive service evaluation score based on the indicator values ​​corresponding to each of the key indicators.

[0101] In this embodiment, to ensure the smooth acquisition of the comprehensive service evaluation score, after obtaining the comprehensive service evaluation score for the current optimization cycle, post-processing is first performed based on the data type of the key indicators to obtain the corresponding indicator values ​​for each key indicator. This ensures that all evaluation indicators consistently adhere to the principle of "the higher the value, the better" during subsequent model optimization. This post-processing includes at least one of the following: inverting inverse-meaning indicators, outlier detection and correction, and unified dimensional normalization. Then, based on the indicator values ​​corresponding to each key indicator, the required comprehensive service evaluation score can be obtained.

[0102] Specifically, the reverse meaning indicator inversion is used to invert evaluation indicators with reverse meaning, such as intent shift rate, user intervention rate, crash rate, and misleading response rate. The inversion formula can be expressed as: ,in, These are the original evaluation index values. The values ​​are the inverted values ​​of the indicators. Outlier detection and correction: For indicators whose values ​​may be affected by external interference, leading to occasional outliers that could impact model performance (e.g., user satisfaction rate), the Z-score method and sliding window are used to detect outliers. If the fluctuation of an indicator in the current period exceeds three times the standard deviation, the system will mark it as an outlier and correct it using historical periodic data (such as the median or mean) to ensure data stability.

[0103] Standardized Dimensioning: For different evaluation indicators with differing dimensions (e.g., user satisfaction is generally based on a Likert scale of 1 to 5), normalization is used to unify the indicator (user satisfaction rate) to the same dimension. The normalized quantification formula can be expressed as: ,in, These are the original index values. This is the maximum value in the historical records of this indicator. This is the minimum value in the historical records of this indicator. This is the normalized value of the indicator.

[0104] See Figure 5Preferably, in the method described above, obtaining the comprehensive service evaluation score based on the indicator values ​​corresponding to each of the key indicators includes:

[0105] Step S501: Construct a comprehensive service evaluation decision vector and a multi-objective function for the key indicators;

[0106] Step S502: Perform non-dominated sorting on the multi-objective function to obtain multiple non-dominated solution sets;

[0107] Step S503: Calculate the comprehensive service evaluation score based on the non-dominated solution set, the first preset algorithm, and the initialized reference point.

[0108] This embodiment illustrates the steps for obtaining a comprehensive service evaluation score based on the post-processed values ​​of each key indicator. First, a comprehensive service evaluation decision vector and a multi-objective function relating to the aforementioned key indicators are constructed. Taking the key indicators as an example, this comprehensive service evaluation decision vector... ,in, for , for , for , for , for , for , for , for , for , for , for , for The multi-objective function of the key indicators is: ,in, This represents the standardized index value of TCR. This represents the standardized metric value of IDR. This represents the standardized index value of MRR. This represents the standardized index value of CR. The standardized metric value representing UIR. This represents the standardized index value of the US.

[0109] Then, using optimization algorithms such as the non-dominated sorting method proposed by the Non-dominated Sorting Genetic Algorithm 2 (NSGA2), the n non-dominated solution sets in the t-th period are calculated. Since the objective function needs to be maximized, the initial reference point is set to... Then, based on the first preset algorithm, the non-dominated solution set and reference point are calculated to obtain the required comprehensive service evaluation score. The first preset algorithm is as follows: ,in, This represents the Lebesgue measure (area in two dimensions, volume in three dimensions, and hypervolume in higher dimensions). Indicates from non-dominant point To the reference point The resulting supercubic prism; This represents the union of multiple hypervolumes, avoiding redundant calculations of overlapping regions. The higher the value, the better the performance of the intelligent integrated customer service system in multiple dimensions such as task completion rate, semantic stability, and user satisfaction in the t-th cycle.

[0110] See Figure 6 Preferably, in the method described above, obtaining the optimization trigger threshold for the current optimization cycle based on the multi-dimensional operating indicators obtained by the dynamic perception layer includes:

[0111] Step S601: Construct the state space of the current optimization cycle based on the multi-dimensional operating indicators;

[0112] Step S602: Adjust the optimization trigger threshold of the previous optimization cycle according to the state space to obtain the optimization trigger threshold of the current optimization cycle.

[0113] This embodiment illustrates the steps for obtaining the optimization trigger threshold for the current optimization cycle. When obtaining the optimization trigger threshold, firstly, multi-dimensional operational metrics are acquired in real-time based on the dynamic perception layer. These metrics cover task dimensions (including task completion rate and intent deviation rate), quality dimensions (including misleading response rate), stability dimensions (including crash rate), and interaction dimensions (including user intervention rate and satisfaction rating), thereby constructing a state space. state space This includes the values ​​and trends of current evaluation metrics (task completion rate, intent deviation rate, misleading response rate, crash rate, user intervention rate, user satisfaction), specifically, the state space. Represented as: ,in, These are the standardized indicator values ​​after preprocessing for the two evaluation dimensions. Indicators The rate of change within the most recent adjacent period. The action space is defined as the adjustment operations on the optimized trigger threshold λ, including increasing, decreasing, or keeping the threshold constant.

[0114] Then, the optimization trigger threshold of the previous optimization cycle is obtained, and the initial optimization trigger threshold is adjusted according to the state space described above to obtain the optimization trigger threshold of the current optimization cycle. Specifically, it can be expressed as: ,in, That is, the state space constructed under the current optimization period t is used to represent the adjustment amount of the threshold; That is, the optimization trigger threshold corresponding to the current optimization period t.

[0115] It should be noted that when the current optimization cycle is the first optimization cycle (i.e., there is no previous optimization cycle), statistical analysis is performed based on historical evaluation data accumulated during the operation of the intelligent customer service system, including task completion rate, intent deviation rate, misleading answer rate, crash rate, user intervention rate, and user satisfaction. The distribution characteristics of the evaluation scores are obtained, and their mean and standard deviation are calculated. This serves as the initial reference for determining whether the intelligent customer service system needs to automatically initiate the optimization process; that is, the initial optimization trigger threshold. Specifically, it can be expressed as: ,in, This represents the average of the overall scores across historical evaluation data. denoted as standard deviation, and k is the adjustment factor.

[0116] See Figure 7 Specifically, as described above, the model optimization based on the optimized data pool and Bayesian optimization-driven LoRA fine-tuning includes:

[0117] Step S701: Preprocess the session data in the optimized data pool and determine the training set and validation set;

[0118] Step S702: Model the objective function of the Bayesian optimization based on the Gaussian process, wherein the objective function is the fine-tuned performance gain;

[0119] Step S703: Construct a joint hyperparameter space including LoRA structural parameters and hyperparameters related to fine-tuning;

[0120] Step S704: Based on the Bayesian optimization, perform an iterative search in the hyperparameter space to obtain the target hyperparameter combination;

[0121] Step S705: Optimize the model based on the target hyperparameter combination to complete the model fine-tuning, deployment, and replacement.

[0122] In this embodiment, the steps of optimizing the model based on the optimized data pool and LoRA fine-tuning driven by Bayesian optimization are illustrated. First, the session data in the optimized data pool is preprocessed, and the training set and validation set are determined. The session data includes, but is not limited to, user input text and corresponding intent tags (such as network anomaly feedback, device replacement request, etc.). The preprocessing includes, but is not limited to, data cleaning.

[0123] Then, the performance gain after fine-tuning (e.g., an improvement in accuracy on the training set) is used as the objective function, denoted as... The objective function is modeled using a Gaussian process, and the Bayesian optimization objective function is set to maximize performance gain. The objective function can be specifically expressed as: ,in, This represents the accuracy of the model after fine-tuning using a specific combination of parameters. This indicates the current model accuracy.

[0124] Furthermore, a joint hyperparameter space is constructed, including the LoRA structural parameters and the hyperparameters related to fine-tuning. This hyperparameter space can be represented as: , among which, among which, Its general value ranges from 1 to 64, representing the rank of a low-rank matrix; , representing the learning rate; , indicating the regularization strength; This represents the set of layer indices in the network to which the LoRA module is to be inserted, such as φ∈{3rd layer, 6th layer, 9th layer}, etc.

[0125] Furthermore, based on the Bayesian optimization of the objective function obtained above, an iterative search is performed within the hyperparameter space to obtain the optimized optimal solution, i.e., the target hyperparameter combination. Then, based on this target hyperparameter combination, the model in the intelligent integrated customer service can be fine-tuned, and the original model in the intelligent integrated customer service (including but not limited to the intent recognition model) can be deployed and replaced, thereby achieving model optimization for the intelligent integrated customer service.

[0126] See Figure 8 Preferably, in the method described above, the iterative search within the hyperparameter space based on the Bayesian optimization to obtain the target hyperparameter combination includes:

[0127] Step S801: Initialize the objective function of Gaussian regression and initialize the Bayesian optimized acquisition function;

[0128] Step S802: Perform hierarchical sampling in the hyperparameter space to obtain the optimized sample set for the current iteration, where each sample in the optimized sample set is a hyperparameter combination;

[0129] Step S803: Perform LoRA fine-tuning training once for each sample in the optimized sample set to obtain the corresponding target performance gain;

[0130] Step S804: Perform Gaussian regression probability modeling based on the optimized sample set;

[0131] Step S805: Obtain the function value of the acquisition function corresponding to each sample in the optimized sample set, and determine the target sample corresponding to the maximum value among the function values;

[0132] Step S806: If the number of iterations corresponding to the current iteration is greater than the preset number of iterations threshold, then the target sample is determined to be the target hyperparameter combination; otherwise, the number of iterations is incremented by one, and the process returns to the step of performing stratified sampling in the hyperparameter space to obtain the optimized sample set of the current iteration.

[0133] In one specific embodiment, an example is given of how to obtain the target hyperparameter combination by iteratively searching in the hyperparameter space based on Bayesian optimization. First, the objective function of Gaussian regression in Bayesian optimization is initialized, i.e. And the initialization of the Bayesian optimization acquisition function, which in this embodiment is represented as the expected improvement function. Specifically, it is expressed as: ,in, Represents the historical best function value. .in, Indicates a normal distribution. This represents the mean. It represents the standard deviation.

[0134] Then, stratified sampling (preferably using Latin cube sampling) is performed in the hyperparameter space to obtain the optimized sample set for the current iteration (the G-th iteration). Each sample in the optimized sample set For a combination of hyperparameters, ,in, This indicates the total number of samples collected.

[0135] Then, LoRA fine-tuning training is performed once for each sample in the optimized sample set to obtain the performance gain corresponding to each sample. , This represents the performance gain of the i-th sample obtained from the G-th iteration sampling.

[0136] Furthermore, Gaussian regression probability modeling is performed based on the samples in the optimized sample set, and the global Gaussian process is trained to ensure the accuracy of Bayesian optimization. Then, based on the acquisition function initialized above, the function value of the acquisition function corresponding to each sample in the optimized sample set is obtained, and the target sample corresponding to the maximum value among the function values ​​is determined. This target sample is preferably the optimal new point in the search space.

[0137] Finally, a judgment is made based on the iteration number corresponding to the current iteration and the preset iteration number threshold. If the iteration number is greater than the preset iteration number threshold, it means that the expected effect has been achieved. Therefore, the target sample obtained in the last iteration is determined to be the required target hyperparameter combination. If the preset iteration number threshold is not reached, the iteration number corresponding to the current iteration is incremented by one, and the process returns to the step of performing stratified sampling in the hyperparameter space to obtain the optimized sample set of the current iteration. The process is repeated to ensure that the final target hyperparameter combination is the optimal solution.

[0138] To facilitate understanding of the beneficial effects of the technical solutions in the embodiments of this application by those skilled in the art, the performance data of the technical solutions in this application are also compared with those of traditional solutions using traditional evaluation methods and static optimization mechanisms based on actual data. In this embodiment, 1286 sets of user service requests were extracted from actual business scenarios for comparison, resulting in the comparison table shown below:

[0139] Evaluation indicators control group average average value of experimental group Increase TCR 82.5% 91.3% +10.6% IDR 11.3% 5.5% -6.5% US 4.1 4.8 +17.1% MRT 2.7% 1.2% -1.5% CR 3.5% 1.8% -2.7% UIR 6.4% 3.1% -3.3% SE 76.3 89.7 +17.6%

[0140] The control group is the traditional scheme, the experimental group is the technical scheme in this application, and the improvement is the value of the technical scheme in this application minus the traditional scheme.

[0141] The experimental results in the table show that the technical solution in this application significantly improves the overall performance of the intelligent customer service coordination system: the task completion rate of the experimental group increased by 10.6% compared with the control group, indicating that the adaptive optimization mechanism enhances the service capability; user satisfaction increased from Likert score of 4.1 to 4.8, showing a significant improvement in user experience; the crash rate decreased to 1.8%, significantly enhancing system stability; the user intervention rate decreased to 3.1%, indicating that the intelligent customer service coordination system's processing capability is enhanced, reducing the need for manual intervention; and the overall evaluation score (SE) increased by 17.6% compared with the control group, verifying the global optimization effect of the technical solution in this application under a multi-dimensional evaluation system.

[0142] The various methods of the embodiments of this application have been described above. Apparatus for implementing the above methods will now be provided.

[0143] See Figure 9 This application also provides a control device, including:

[0144] The first processing module 901 is configured to, if it is determined that the session has ended, acquire key indicator data corresponding to multiple key indicators from the session data, and archive the session data and the key indicator data to the optimization data pool, and / or archive new solutions from the session data to the knowledge base. The key indicators include at least one of the following: task completion rate, intent deviation rate, misleading answer rate, crash rate, user intervention rate, and user satisfaction.

[0145] The second processing module 902 is used to obtain a judgment result on whether to perform model optimization based on the preset optimization cycle and the data in the optimization data pool;

[0146] The third processing module 903 is used to perform model optimization based on LoRA fine-tuning driven by Bayesian optimization if the judgment result indicates that model optimization needs to be performed.

[0147] Specifically, the apparatus described above determines the end of the session if at least one of the following conditions is met:

[0148] The user input was detected as a session end message;

[0149] No user input was detected within the preset time period.

[0150] Preferably, in the apparatus described above, the second processing module includes:

[0151] The first processing submodule is used to obtain the comprehensive service evaluation score of the current optimization period based on the key indicators when the time reaches any of the optimization periods.

[0152] The second processing submodule is used to obtain the optimization trigger threshold of the current optimization cycle based on the multi-dimensional operation indicators obtained by the dynamic perception layer.

[0153] The third processing submodule is used to determine that if the comprehensive service evaluation score is less than or equal to the optimization trigger threshold, the judgment result is that model optimization needs to be performed; otherwise, the judgment result is that model optimization does not need to be performed.

[0154] Preferably, in the apparatus described above, the first processing submodule includes:

[0155] The first processing unit is used to perform post-processing according to the data type of the key indicators to obtain the indicator value corresponding to each key indicator. The post-processing includes at least one of the following: inverting the inverse meaning indicator, outlier detection and correction, and unified dimension normalization processing.

[0156] The second processing unit is used to obtain the comprehensive service evaluation score based on the indicator values ​​corresponding to each of the key indicators.

[0157] Preferably, in the apparatus described above, the second processing unit includes:

[0158] The first processing subunit is used to construct a service comprehensive evaluation decision vector and a multi-objective function about the key indicators;

[0159] The second processing subunit is used to perform non-dominated sorting on the multi-objective function to obtain multiple non-dominated solution sets;

[0160] The third processing subunit is used to calculate the comprehensive service evaluation score based on the non-dominated solution set, the first preset algorithm, and the initialized reference point.

[0161] Preferably, in the apparatus described above, the second processing submodule includes:

[0162] The third processing unit is used to construct the state space of the current optimization cycle based on the multi-dimensional operating indicators.

[0163] The fourth processing unit is used to adjust the optimization trigger threshold of the previous optimization cycle according to the state space to obtain the optimization trigger threshold of the current optimization cycle.

[0164] Specifically, in the apparatus described above, the third processing module includes:

[0165] The fourth processing submodule is used to preprocess the session data in the optimized data pool and determine the training set and validation set;

[0166] The fifth processing submodule is used to model the objective function of the Bayesian optimization based on the Gaussian process, wherein the objective function is the fine-tuned performance gain;

[0167] The sixth processing submodule is used to construct a joint hyperparameter space including LoRA structural parameters and hyperparameters related to fine-tuning;

[0168] The seventh processing submodule is used to perform an iterative search in the hyperparameter space based on the Bayesian optimization to obtain the target hyperparameter combination;

[0169] The eighth processing submodule is used to optimize the model based on the target hyperparameter combination, and to complete the fine-tuning, deployment and replacement of the model.

[0170] Preferably, in the apparatus described above, the seventh processing submodule includes:

[0171] The fifth processing unit is used to initialize the objective function of Gaussian regression and the acquisition function optimized by Bayes.

[0172] The sixth processing unit is used to perform hierarchical sampling in the hyperparameter space to obtain the optimized sample set for the current iteration, wherein each sample in the optimized sample set is a hyperparameter combination.

[0173] The seventh processing unit is used to perform LoRA fine-tuning training once for each sample in the optimized sample set to obtain the corresponding target performance gain;

[0174] The eighth processing unit is used to perform Gaussian regression probability modeling based on the optimized sample set;

[0175] The ninth processing unit is used to obtain the function value of the acquisition function corresponding to each sample in the optimized sample set, and to determine the target sample corresponding to the maximum value among the function values;

[0176] The tenth processing unit is configured to determine the target sample as the target hyperparameter combination if the number of iterations corresponding to the current iteration is greater than a preset iteration number threshold; otherwise, increment the iteration number by one and return to the step of performing stratified sampling in the hyperparameter space to obtain the optimized sample set of the current iteration.

[0177] The control device embodiment of this application corresponds to the embodiment of the model optimization method for the above-described intelligent integrated customer service system. All implementation means in the above-described method embodiments are applicable to the embodiment of this device and can achieve the same technical effect. The diagnostic device provided in this application embodiment can implement all the method steps implemented in the above-described method embodiments and can achieve the same technical effect. Therefore, the parts that are the same as those in the method embodiments and their beneficial effects will not be described in detail here.

[0178] See Figure 10 Another embodiment of this application provides an electronic device 100, including: a processor 1001, a memory 1002, and a program stored in the memory 1002 and executable on the processor 1001. When the program is executed by the processor 1001, it implements the steps of the intelligent integrated customer service model optimization method as described above and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0179] Another embodiment of this application provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the steps of the intelligent integrated customer service model optimization method described above, achieving the same technical effect. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0180] Another embodiment of this application provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the intelligent integrated customer service model optimization method described above, and achieve the same technical effect. To avoid repetition, it will not be described again here.

[0181] 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 apparatus 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 apparatus. 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 apparatus that includes that element.

[0182] 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 ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0183] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A model optimization method for intelligent integrated customer service, characterized in that, include: When the session is determined to be over, the key indicator data corresponding to multiple key indicators in the session data is obtained, and the session data and the key indicator data are archived to the optimization data pool. The key indicators include at least one of the following: task completion rate, intent deviation rate, misleading answer rate, crash rate, user intervention rate, and user satisfaction. Based on the preset optimization cycle and the data in the optimization data pool, obtain the judgment result of whether to perform model optimization; If the determination result indicates that model optimization is required, then model optimization is performed based on the optimized data pool and Bayesian optimization-driven low-rank adaptive LoRA fine-tuning.

2. The method according to claim 1, characterized in that, The session is determined to end if at least one of the following conditions is met: The user input was detected as a session end message; No user input was detected within the preset time period.

3. The method according to claim 1, characterized in that, The step of obtaining a judgment result on whether to perform model optimization based on a preset optimization period and data in the optimization data pool includes: If any of the aforementioned optimization cycles are reached within a given timeframe, obtain the comprehensive service evaluation score for the current optimization cycle based on the aforementioned key indicators. Based on the multi-dimensional operating indicators obtained by the dynamic perception layer, the optimization trigger threshold for the current optimization cycle is obtained; If the overall service evaluation score is less than or equal to the optimization trigger threshold, then the judgment result is determined to be that model optimization needs to be performed; otherwise, the judgment result is determined to be that model optimization does not need to be performed.

4. The method according to claim 3, characterized in that, The process of obtaining the comprehensive service evaluation score for the current optimization cycle based on the key indicators includes: Based on the data type of the key indicators, post-processing is performed to obtain the indicator values ​​corresponding to each key indicator. The post-processing includes at least one of the following: inverting the inverse meaning of indicators, outlier detection and correction, and unified dimension normalization. The comprehensive service evaluation score is obtained based on the index values ​​corresponding to each of the key indicators.

5. The method according to claim 4, characterized in that, The step of obtaining the comprehensive service evaluation score based on the indicator values ​​corresponding to each of the key indicators includes: Construct a comprehensive service evaluation decision vector and a multi-objective function for the key indicators; The multi-objective function is sorted using non-dominated methods to obtain multiple non-dominated solution sets; The comprehensive service evaluation score is calculated based on the non-dominated solution set, the first preset algorithm, and the initialized reference point.

6. The method according to claim 3, characterized in that, The step of obtaining the optimization trigger threshold for the current optimization cycle based on the multi-dimensional operational metrics obtained from the dynamic perception layer includes: Construct the state space of the current optimization cycle based on the multidimensional operating indicators; The optimization trigger threshold of the previous optimization cycle is adjusted according to the state space to obtain the optimization trigger threshold of the current optimization cycle.

7. The method according to claim 1, characterized in that, The model optimization based on the optimized data pool and Bayesian optimization-driven low-rank adaptive LoRA fine-tuning includes: The session data in the optimized data pool is preprocessed, and the training set and validation set are determined; The objective function of the Bayesian optimization is modeled based on the Gaussian process, and the objective function is the fine-tuned performance gain. Construct a joint hyperparameter space that includes LoRA structural parameters and hyperparameters related to fine-tuning; Based on the Bayesian optimization, an iterative search is performed in the hyperparameter space to obtain the target hyperparameter combination; The model is optimized based on the target hyperparameter combination, and the model is fine-tuned, deployed, and replaced.

8. The method according to claim 7, characterized in that, According to the Bayesian optimization, an iterative search is performed in the hyperparameter space to obtain the target hyperparameter combination, including: Initialize the objective function of Gaussian regression, and initialize the acquisition function of the Bayesian optimization; Hierarchical sampling is performed in the hyperparameter space to obtain the optimized sample set for the current iteration, where each sample in the optimized sample set is a hyperparameter combination; Perform LoRA fine-tuning training once for each sample in the optimized sample set to obtain the corresponding target performance gain; Gaussian regression probability modeling is performed based on the optimized sample set; Obtain the function value of the acquisition function corresponding to each sample in the optimized sample set, and determine the target sample corresponding to the maximum value among the function values; If the number of iterations corresponding to the current iteration is greater than the preset number of iterations threshold, then the target sample is determined to be the target hyperparameter combination; otherwise, the number of iterations is incremented by one, and the process returns to the step of performing stratified sampling in the hyperparameter space to obtain the optimized sample set for the current iteration.

9. A control device, characterized in that, include: The first processing module is used to, upon determining that the session has ended, acquire key indicator data corresponding to multiple key indicators from the session data, and archive the session data and the key indicator data into an optimized data pool. The key indicators include at least one of the following: task completion rate, intent deviation rate, misleading answer rate, crash rate, user intervention rate, and user satisfaction. The second processing module is used to obtain a judgment result on whether to perform model optimization based on the preset optimization cycle and the data in the optimization data pool; The third processing module is used to perform model optimization based on LoRA fine-tuning driven by Bayesian optimization if the judgment result indicates that model optimization is required.

10. An electronic device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the model optimization method for intelligent integrated customer service as described in any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the model optimization method for intelligent integrated customer service as described in any one of claims 1 to 8.

12. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the steps of the model optimization method for intelligent integrated customer service as described in any one of claims 1 to 8.

Images