AI model training method based on training data acquisition

Abstract

The invention discloses an AI model training method based on training data acquisition, relates to the technical field of intelligent AI, and aims to accurately identify user requirements and problem types by a system through deep analysis and feature extraction of interaction data between a user and a customer service system, so as to realize more accurate problem classification and response. And secondly, the system can dynamically optimize the automatic response strategy Rj through fitting of the feature vector F and the problem category clustering center Ck, so that complex and non-standardized problems are effectively processed, and the dependence on manual customer service is greatly reduced. Furthermore, the system can adapt to user feedback and problem changes in real time through continuous evaluation and regulation and control of the automatic response strategy Rj, the method not only solves the defects of a traditional intelligent customer service system in the aspects of processing complex problems and improving response efficiency, but also effectively reduces customer loss and operation cost, and the method is suitable for popularization and application. And the overall user experience is obviously improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent AI technology, and specifically to an AI model training method based on training data collection. Background Art

[0002] Artificial intelligence involves using computer systems to simulate the functions of human intelligence, while machine learning focuses on automatic learning and prediction through algorithms and data. How to optimize and improve the performance of AI models through a large amount of training data. This approach is particularly suitable for application scenarios that require processing large amounts of data and making intelligent decisions, such as intelligent customer service systems. In the telecommunications industry, intelligent customer service systems use this method to continuously optimize their service capabilities to better meet user needs.

[0003] Current intelligent customer service systems often encounter the following defects in practical applications: First, the existing system has limited ability to understand complex problems and cannot effectively handle users' non-standardized problems and complaints. Second, the system's self-service function often relies on rules and templates, lacks flexibility and intelligence, and results in user problems not being satisfactorily resolved. In addition, the system's analysis of user interaction data is not in-depth enough to accurately identify high-frequency problems and potential user needs. Therefore, the demand for manual customer service remains high, and the system's self-service function has not been able to fully realize its potential. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides an AI model training method based on training data collection, which solves the problems mentioned in the background technology.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: an AI model training method based on training data collection comprises the following steps:

[0006] S1. Collect the communication data between users and the customer service system, including voice records, text chat records, user feedback, waiting time and resolution time, and process and extract the communication data to form an interaction data group UID;

[0007] S2, extracting features from the interaction data group UID, obtaining user interaction feature information, question type feature information, feedback score feature information, user status feature information, and solved problem feature information, and simultaneously performing preprocessing, including feature information standardization and feature information conversion, to form a feature vector F;

[0008] S3, preliminarily classify the user's intercommunication data by using the feature vector F, and obtain the proportion of high-frequency problems and the proportion of low-frequency problems, then perform cluster analysis on the feature vector F, obtain the problem category cluster center Ck, and simultaneously iteratively regulate the problem category cluster center Ck;

[0009] S4, fitting the feature vector F and the problem category cluster center Ck, identifying the intercommunication data, and triggering the corresponding automatic response strategy Rj;

[0010] S5. Evaluate the automatic response strategy Rj, count the performance of the automatic response strategy Rj in different time periods and different problem types, obtain the system performance evaluation index P, and match it with the preset system operation performance status evaluation threshold YZ to obtain the system operation status evaluation result;

[0011] S6. Trigger iterative control of the automatic response strategy Rj according to the system operation status evaluation result.

[0012] Preferably, S1, collects the communication data between the user and the customer service system, including voice records, text chat records, user feedback, waiting time and resolution time, and processes them using speech-to-text and natural language processing technology to form the communication data;

[0013] The voice recording data is specifically generated by capturing the voice interaction content between the user and the customer service system to generate a voice file, including the audio file of the user's questions and the customer service's responses, and the timestamp of each interaction marked in the voice recording, including recording and acquisition through a language recording device;

[0014] Text chat records: The chat record function of the customer service system records all text conversations between users and customer service in real time, including the messages entered by users, the customer service replies, and the timestamp of each message;

[0015] User feedback: Collect user feedback on the customer service experience through questionnaires or rating systems after each interaction, including collecting user ratings as a manifestation of user feedback;

[0016] The waiting time and resolution time are automatically tracked by marking the time from the user initiating the request to the first response, as well as the total time for problem resolution, to obtain the waiting time WT and resolution time ST, and form the interaction data group UID, specifically UID = {voice record, text chat record, user feedback, WT, ST}.

[0017] Preferably, S2 further includes S21 and S22;

[0018] S21, extracting features from the interaction data group UID to obtain user interaction feature information, question type feature information, feedback score feature information, user status feature information, and solved problem feature information;

[0019] S22, perform feature information standardization and feature information conversion on user interaction feature information, question type feature information, feedback score feature information, user status feature information and solved problem feature information, obtain total response TAT, question category QTY, user evaluation EVA, customer feedback CTF and problem solving rate RSR, and form a feature vector F, specifically F = {TAT, QTY, EVA, CTF, RSR};

[0020] The total response TAT is obtained by processing the user interaction feature information using TAT=WT+ST;

[0021] The question category QTY is obtained by analyzing the question type feature information using natural language processing technology, and the question category information is obtained specifically through the following S221, S222, S223 and S224;

[0022] The user evaluation EVA extracts the rating data from the user feedback data in the feedback rating feature information, and performs a normalization process to adjust the range of the rating data to between 0 and 1. The normalization process is specifically performed through the following calculation formula:

[0023]

[0024] In the formula, UR represents the actual score given by the user, MaxR represents the upper limit score of the system, and MinR represents the lower limit score of the system;

[0025] Customer Feedback CTF analyzes user status feature information by using natural language processing technology and sentiment analysis models to obtain user feedback degree values ​​ranging from 0 to 1. Sentiment analysis models include VADER, TextBlob, and BERT models.

[0026] The problem resolution rate (RSR) is obtained by extracting the problem resolution status from the feature information of the resolved problems, including resolved and unresolved, and obtaining the subscript values ​​1 and 0, and by counting the ratio of the total number of problems to the total number of resolved problems.

[0027] Preferably, the contents of obtaining the question categories in S221, S222, S223 and S224 are specifically:

[0028] S221, text preprocessing, performing word segmentation, stop word removal and stem extraction on the text chat records;

[0029] S222, feature extraction, using the bag-of-words model and word embedding to convert text into feature vectors;

[0030] S223, classifier training, applying supervised learning algorithms, including SVM, random forest and deep learning models, to the labeled training data to identify the problem category;

[0031] S224, category labeling, input the feature vector of each user interaction text into the classifier to obtain the question category QTY.

[0032] Preferably, S3 also includes S31 and S32;

[0033] S31, preliminarily classifying the user's communication data by using the feature vector F, and obtaining the proportion of high-frequency problems and the proportion of low-frequency problems, and then performing cluster analysis on the feature vector F to obtain the problem category cluster center Ck;

[0034] Among them, cluster analysis is performed through the following calculation formula:

[0035]

[0036] In the formula, Ck represents the cluster center of the problem category, specifically the cluster center of the k-th category, Nk represents the number of samples belonging to the k-th category, Cluster, k represents the sample set of the k-th category, and Fi represents the feature vector of the i-th sample in the feature vector F.

[0037] Preferably, S32, iteratively control the problem category cluster center Ck to obtain an updated cluster center Ck, new and calculate the distance d from the sample i to the k category center;

[0038] The updated cluster center Ck,new is obtained by the following calculation formula:

[0039]

[0040] In the formula, Wi represents the weight value of the i-th sample;

[0041] The distance d is obtained by the following calculation formula:

[0042]

[0043] Where, d i，k represents the distance d from sample i to the center of category k, and T represents the transpose of the sentence, which is specifically used to convert a column vector into a row vector.

[0044] Preferably, S4 further includes S41 and S42;

[0045] S41, fit the feature vector F and the problem category cluster center Ck using the AI ​​model, identify and communicate data, obtain the classification result CTi of the sample, and trigger the corresponding automatic response strategy Rj, synchronously assign the automatic response strategy Rj content to the classification result CTi of the sample, and store it in the database;

[0046] The classification result CTi is obtained by the following calculation formula:

[0047]

[0048] In the formula, λ represents the regularization factor, which is used to control the weight value of the distance d from sample i to the center of category k, Fi-Ck 2 It represents the sum of squares of the feature vector of the i-th sample in the feature vector F and the cluster center of the k-th category, which is specifically used to measure the fit between the feature vector Fi and the cluster Ck;

[0049] The automatic response strategy Rj is allocated by the following calculation formula:

[0050] Rj=Strategy map (CTi);

[0051] Where Rj represents the jth automatic response strategy R, Strategy map Represents a strategy mapping function, which selects the corresponding automatic response strategy Rj according to the classification result CTi.

[0052] Preferably, S42, performing a secondary evaluation on the triggering of the corresponding automatic response strategy Rj according to the obtained allocation sample category Rj, obtaining the secondary evaluation triggering result, and triggering the execution of the automatic response strategy Rj according to the secondary evaluation triggering result;

[0053] The secondary evaluation trigger result is triggered by the following trigger judgment formula:

[0054]

[0055] Wherein, TG represents the return signal of the trigger judgment formula, specifically, TG return signal 1 represents triggering, TG return signal 0 represents non-triggering, TJL represents the preset trigger judgment threshold;

[0056] The execution of the automatic response strategy Rj includes sending FAQs, transferring to manual customer service, providing service status updates, sending questionnaires, providing additional help links, sending problem confirmations and system self-diagnosis.

[0057] Preferably, S5 further includes S51 and S52;

[0058] S51, evaluating the automatic response strategy Rj, counting the performance of the automatic response strategy Rj in different time periods and different problem types, and obtaining a system performance evaluation index P;

[0059] The system performance evaluation index P is obtained by the following calculation formula:

[0061]

[0062] In the formula, EL represents the success rate of response strategy execution, TL represents the average response time of the response strategy, and SL represents the user satisfaction score of the response strategy execution;

[0063] Among them, the execution success rate EL is The calculation formula is obtained, where Nsuccess represents the number of successfully resolved problems, and Ntotal represents the total number of automatic response strategies;

[0064] Average response time TL passed The calculation formula is obtained, where T Rj，w represents the response time of the jth automatic response strategy R for the wth user;

[0065] User satisfaction rating SL passed Calculation formula obtained, Nfeedback represents the total number of user feedback collected, SL Rj，w The user feedback score S52 representing the jth automatic response strategy R of the wth user is matched with the preset system operation performance status evaluation threshold YZ according to the calculated system performance evaluation index P to obtain the system operation status evaluation result;

[0066] The system operation status evaluation result is Matching method acquisition;

[0067] When the Status match return signal is 1, the system operation status evaluation result is qualified;

[0068] When the Status match return signal is 0, the evaluation result of the system operation status is unqualified.

[0069] Preferably, S6, triggering the iterative regulation of the automatic response strategy Rj according to the system operation status evaluation result is triggered by the following method:

[0070] When the system operation status evaluation result is qualified and the Status match return signal is 1, iterative control is not triggered;

[0071] When the system operation status evaluation result is unqualified and the Status match return signal is 0, iterative regulation is triggered, including adjusting response rules, optimizing trigger conditions, updating strategy parameters, and adjusting model hyperparameters;

[0072] Adjusting model hyperparameters includes regulating the AI ​​model training data category, implementing cross-validation technology and data enhancement technology to adjust the effectiveness of the AI ​​model’s training data.

[0073] The present invention provides an AI model training method based on training data collection, which has the following beneficial effects:

[0074] (1) Through S1 to S6, the service quality and efficiency can be significantly improved. First, through in-depth analysis and feature extraction of the interaction data between users and the customer service system, the system can accurately identify user needs and problem types, thereby achieving more accurate problem classification and response. Secondly, through the fitting of the feature vector F and the problem category cluster center Ck, the system can dynamically optimize the automatic response strategy Rj and effectively handle complex and non-standardized problems, which greatly reduces the dependence on manual customer service. Furthermore, through the continuous evaluation and regulation of the automatic response strategy Rj, the system can adapt to user feedback and problem changes in real time, optimize service performance, and improve user satisfaction. This method not only solves the shortcomings of traditional intelligent customer service systems in handling complex problems and improving response efficiency, but also effectively reduces customer churn and operating costs, and significantly improves the overall user experience.

[0075] (2) Through cluster analysis and iterative regulation, the problem category cluster center can be dynamically updated, thereby improving the recognition accuracy and response effect of user problems. Furthermore, by fitting the feature vector F and the problem category cluster center Ck using the AI ​​model, the most appropriate response strategy can be automatically assigned according to real-time data, and the appropriate strategy can be accurately triggered through the secondary evaluation mechanism. This method not only enhances the system's ability to handle complex user problems, but also significantly improves the flexibility of the service and user experience through an efficient and accurate automatic response mechanism, further reducing operating costs and the need for manual intervention. Unlike traditional methods that rely on the limitations of static rules, this dynamic, data-driven optimization strategy makes the intelligent customer service system more intelligent and adaptive, and can maintain efficient and stable service performance in the ever-changing user needs.

[0076] (3) The performance of the automatic response strategy is comprehensively evaluated through detailed performance evaluation indicators, including the execution success rate EL, average response time TL, and user satisfaction score SL of the response strategy. These evaluation results not only provide the system with an accurate evaluation of the operating status, but also ensure real-time monitoring and optimization of service quality. In particular, by matching the preset system operating performance status evaluation threshold YZ, the system can flexibly trigger necessary iterative control measures, such as adjusting response rules, optimizing trigger conditions, and updating policy parameters. This dynamic adjustment mechanism ensures that the system can quickly make adaptive improvements in the face of various service challenges, improving the accuracy and efficiency of problem solving. The unique benefit of this approach lies in its real-time, data-driven adjustment capability, which not only ensures efficient service delivery, but also significantly enhances the adaptability and stability of the system in a changing environment, further optimizing the user experience and reducing long-term operating costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 Schematic diagram of the steps of the AI ​​model training method based on training data collection of the present invention. DETAILED DESCRIPTION

[0078] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0079] Example 1

[0080] The present invention provides an AI model training method based on training data collection, please refer to Figure 1 , including the following steps:

[0081] S1. Collect the communication data between users and the customer service system, including voice records, text chat records, user feedback, waiting time and resolution time, and process and extract the communication data to form an interaction data group UID;

[0082] S2, extracting features from the interaction data group UID, obtaining user interaction feature information, question type feature information, feedback score feature information, user status feature information, and solved problem feature information, and simultaneously performing preprocessing, including feature information standardization and feature information conversion, to form a feature vector F;

[0083] S3, preliminarily classify the user's intercommunication data by using the feature vector F, and obtain the proportion of high-frequency problems and the proportion of low-frequency problems, then perform cluster analysis on the feature vector F, obtain the problem category cluster center Ck, and simultaneously iteratively regulate the problem category cluster center Ck;

[0084] S4, fitting the feature vector F and the problem category cluster center Ck, identifying the intercommunication data, and triggering the corresponding automatic response strategy Rj;

[0085] S5. Evaluate the automatic response strategy Rj, count the performance of the automatic response strategy Rj in different time periods and different problem types, obtain the system performance evaluation index P, and match it with the preset system operation performance status evaluation threshold YZ to obtain the system operation status evaluation result;

[0086] S6. Trigger iterative control of the automatic response strategy Rj according to the system operation status evaluation result.

[0087] In this embodiment, through S1 to S6, its service quality and efficiency can be significantly improved. First, through in-depth analysis and feature extraction of the interaction data between users and the customer service system, the system can accurately identify user needs and problem types, thereby achieving more accurate problem classification and response. Secondly, through the fitting of the feature vector F and the problem category cluster center Ck, the system can dynamically optimize the automatic response strategy Rj and effectively handle complex and non-standardized problems, which greatly reduces the dependence on manual customer service. Furthermore, through the continuous evaluation and regulation of the automatic response strategy Rj, the system can adapt to user feedback and problem changes in real time, optimize service performance, and improve user satisfaction. This method not only solves the shortcomings of traditional intelligent customer service systems in handling complex problems and improving response efficiency, but also effectively reduces customer churn and operating costs, and significantly improves the overall user experience.

[0088] Example 2

[0089] This embodiment is explained in Example 1, please refer to Figure 1 Specifically: S1, collects the communication data between users and customer service systems, including voice records, text chat records, user feedback, waiting time and resolution time, and processes them using speech-to-text and natural language processing technology to form communication data;

[0090] The voice recording data is specifically generated by capturing the voice interaction content between the user and the customer service system to generate a voice file, including the audio file of the user's questions and the customer service's responses, and the timestamp of each interaction marked in the voice recording, including recording and acquisition through a language recording device;

[0091] Text chat records: The chat record function of the customer service system records all text conversations between users and customer service in real time, including the messages entered by users, the customer service replies, and the timestamp of each message;

[0092] User feedback: Collect user feedback on the customer service experience through questionnaires or rating systems after each interaction, including collecting user ratings as a manifestation of user feedback;

[0093] The waiting time and resolution time are automatically tracked by marking the time from the user initiating the request to the first response, as well as the total time for problem resolution, to obtain the waiting time WT and resolution time ST, and form the interaction data group UID, specifically UID = {voice record, text chat record, user feedback, WT, ST}.

[0094] Among them, S2 also includes S21 and S22;

[0095] S21, extracting features from the interaction data group UID to obtain user interaction feature information, question type feature information, feedback score feature information, user status feature information, and solved problem feature information;

[0096] S22, perform feature information standardization and feature information conversion on user interaction feature information, question type feature information, feedback score feature information, user status feature information and solved problem feature information, obtain total response TAT, question category QTY, user evaluation EVA, customer feedback CTF and problem solving rate RSR, and form a feature vector F, specifically F = {TAT, QTY, EVA, CTF, RSR};

[0097] The total response TAT is obtained by processing the user interaction feature information using TAT=WT+ST;

[0098] The question category QTY is obtained by analyzing the question type feature information using natural language processing technology, and the question category information is obtained specifically through the following S221, S222, S223 and S224;

[0099] The user evaluation EVA extracts the rating data from the user feedback data in the feedback rating feature information, and performs a normalization process to adjust the range of the rating data to between 0 and 1. The normalization process is specifically performed through the following calculation formula:

[0100]

[0101] In the formula, UR represents the actual score given by the user, MaxR represents the upper limit score of the system, and MinR represents the lower limit score of the system;

[0102] Customer Feedback CTF analyzes user status feature information by using natural language processing technology and sentiment analysis models to obtain user feedback degree values ​​ranging from 0 to 1. Sentiment analysis models include VADER, TextBlob, and BERT models.

[0103] The problem resolution rate (RSR) is obtained by extracting the problem resolution status from the feature information of the resolved problems, including resolved and unresolved, and obtaining the subscript values ​​1 and 0, and by counting the ratio of the total number of problems to the total number of resolved problems.

[0104] Among them, S221, S222, S223 and S224 are used to obtain the content of the question category:

[0105] S221, text preprocessing, performing word segmentation, stop word removal and stem extraction on the text chat records;

[0106] S222, feature extraction, using the bag-of-words model and word embedding to convert text into feature vectors;

[0107] S223, classifier training, applying supervised learning algorithms, including SVM, random forest and deep learning models, to the labeled training data to identify the problem category;

[0108] S224, category labeling, input the feature vector of each user interaction text into the classifier to obtain the question category QTY.

[0109] Example 3

[0110] This embodiment is explained in Example 2. Please refer to Figure 1 , specifically: wherein S3 also includes S31 and S32;

[0111] S31, preliminarily classifying the user's communication data by using the feature vector F, and obtaining the proportion of high-frequency problems and the proportion of low-frequency problems, and then performing cluster analysis on the feature vector F to obtain the problem category cluster center Ck;

[0112] Among them, cluster analysis is performed through the following calculation formula:

[0113]

[0114] In the formula, Ck represents the cluster center of the problem category, specifically the cluster center of the k-th category, Nk represents the number of samples belonging to the k-th category, Cluster, k represents the sample set of the k-th category, and Fi represents the feature vector of the i-th sample in the feature vector F.

[0115] S32, iteratively control the problem category cluster center Ck, obtain the updated cluster center Ck, new and calculate the distance d from sample i to the k category center;

[0116] The updated cluster center Ck,new is obtained by the following calculation formula:

[0117]

[0118] In the formula, Wi represents the weight value of the i-th sample;

[0119] The distance d is obtained by the following calculation formula:

[0120]

[0121] Where, d i，k represents the distance d from sample i to the center of category k, and T represents the transpose of the sentence, which is specifically used to convert a column vector into a row vector.

[0122] Among them, S4 also includes S41 and S42;

[0123] S41, fit the feature vector F and the problem category cluster center Ck using the AI ​​model, identify and communicate data, obtain the classification result CTi of the sample, and trigger the corresponding automatic response strategy Rj, synchronously assign the automatic response strategy Rj content to the classification result CTi of the sample, and store it in the database;

[0124] The classification result CTi is obtained by the following calculation formula:

[0125]

[0126] In the formula, λ represents the regularization factor, which is used to control the weight value of the distance d from sample i to the center of category k, Fi-Ck 2 It represents the sum of squares of the feature vector of the i-th sample in the feature vector F and the cluster center of the k-th category, which is specifically used to measure the fit between the feature vector Fi and the cluster Ck;

[0127] The automatic response strategy Rj is allocated by the following calculation formula:

[0128] Rj=Strategy map (CTi);

[0129] Where Rj represents the jth automatic response strategy R, Strategy map Represents a strategy mapping function, which selects the corresponding automatic response strategy Rj according to the classification result CTi.

[0130] S42, performing a secondary evaluation on the triggering of the corresponding automatic response strategy Rj according to the obtained allocation sample category Rj, obtaining the secondary evaluation triggering result, and triggering the execution of the automatic response strategy Rj according to the secondary evaluation triggering result;

[0131] The secondary evaluation trigger result is triggered by the following trigger judgment formula:

[0132]

[0133] Wherein, TG represents the return signal of the trigger judgment formula, specifically, TG return signal 1 represents triggering, TG return signal 0 represents non-triggering, TJL represents the preset trigger judgment threshold;

[0134] The execution of the automatic response strategy Rj includes sending FAQs, transferring to manual customer service, providing service status updates, sending questionnaires, providing additional help links, sending problem confirmations and system self-diagnosis.

[0135] In this embodiment, through cluster analysis and iterative regulation, the problem category cluster center can be dynamically updated, thereby improving the recognition accuracy and response effect of user problems. Furthermore, by fitting the feature vector F and the problem category cluster center Ck using the AI ​​model, the most appropriate response strategy can be automatically allocated according to real-time data, and the appropriate strategy can be accurately triggered through the secondary evaluation mechanism. This method not only enhances the system's ability to handle complex user problems, but also significantly improves the flexibility of the service and user experience through an efficient and accurate automatic response mechanism, further reducing operating costs and the need for manual intervention. Unlike traditional methods that rely on the limitations of static rules, this dynamic, data-driven optimization strategy makes the intelligent customer service system more intelligent and adaptive, and can maintain efficient and stable service performance in the ever-changing user needs.

[0136] Example 4

[0137] This embodiment is explained in Example 3, please refer to Figure 1 , specifically: S5 also includes S51 and S52;

[0138] S51, evaluating the automatic response strategy Rj, counting the performance of the automatic response strategy Rj in different time periods and different problem types, and obtaining a system performance evaluation index P;

[0139] The system performance evaluation index P is obtained by the following calculation formula:

[0141]

[0142] In the formula, EL represents the success rate of response strategy execution, TL represents the average response time of the response strategy, and SL represents the user satisfaction score of the response strategy execution;

[0143] Among them, the execution success rate EL is The calculation formula is obtained, where Nsuccess represents the number of successfully resolved problems, and Ntotal represents the total number of automatic response strategies;

[0144] Average response time TL passed The calculation formula is obtained, where T Rj , w represents the response time of the jth automatic response strategy R of the wth user;

[0145] User satisfaction rating SL passed Calculation formula obtained, Nfeedback represents the total number of user feedback collected, SL Rj，w represents the user feedback score of the jth automatic response strategy R for the wth user

[0146] S52, matching the calculated system performance evaluation index P with a preset system operation performance status evaluation threshold YZ to obtain a system operation status evaluation result;

[0147] The system operation status evaluation result is Matching method acquisition;

[0148] When the Status match return signal is 1, the system operation status evaluation result is qualified;

[0149] When the Status match return signal is 0, the evaluation result of the system operation status is unqualified.

[0150] Among them, S6, triggering the iterative regulation of the automatic response strategy Rj according to the system operation status evaluation result is triggered by the following method:

[0151] When the system operation status evaluation result is qualified and the Status match return signal is 1, iterative control is not triggered;

[0152] When the system operation status evaluation result is unqualified and the Status match return signal is 0, iterative regulation is triggered, including adjusting response rules, optimizing trigger conditions, updating strategy parameters, and adjusting model hyperparameters;

[0153] Adjusting model hyperparameters includes regulating the AI ​​model training data category, implementing cross-validation technology and data enhancement technology to adjust the effectiveness of the AI ​​model’s training data.

[0154] In this embodiment, the performance of the automatic response strategy is comprehensively evaluated through detailed performance evaluation indicators, including the execution success rate EL, average response time TL and user satisfaction score SL of the response strategy. These evaluation results not only provide the system with an accurate operating status evaluation, but also ensure real-time monitoring and optimization of service quality. In particular, by matching the preset system operating performance status evaluation threshold YZ, the system can flexibly trigger necessary iterative control measures, such as adjusting response rules, optimizing trigger conditions and updating policy parameters. This dynamic adjustment mechanism ensures that the system can quickly make adaptive improvements in the face of various service challenges, improving the accuracy and efficiency of problem solving. The unique benefit of this method lies in its real-time, data-driven adjustment capabilities, which not only ensures efficient service delivery, but also significantly enhances the adaptability and stability of the system in a changing environment, further optimizes the user experience and reduces long-term operating costs.

[0155] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI model training method based on training data collection, characterized in that: The following steps are involved: S1. Collect the communication data between users and the customer service system, including voice records, text chat records, user feedback, waiting time and resolution time, and process and extract the communication data to form an interaction data group UID; S2. Extract features from the interaction data group UID to obtain user interaction feature information, question type feature information, feedback score feature information, user status feature information, and solved problem feature information, and simultaneously perform preprocessing, including feature information standardization and feature information conversion. Form the feature vector F; S3, preliminarily classify the user's intercommunication data by using the feature vector F, and obtain the proportion of high-frequency problems and the proportion of low-frequency problems, then perform cluster analysis on the feature vector F, obtain the problem category cluster center Ck, and simultaneously iteratively regulate the problem category cluster center Ck; S4, fitting the feature vector F and the problem category cluster center Ck, identifying the intercommunication data, and triggering the corresponding automatic response strategy Rj; S5. Evaluate the automatic response strategy Rj, count the performance of the automatic response strategy Rj in different time periods and different problem types, obtain the system performance evaluation index P, and match it with the preset system operation performance status evaluation threshold YZ to obtain the system operation status evaluation result; S6. Trigger iterative control of the automatic response strategy Rj according to the system operation status evaluation result.

2. The AI ​​model training method based on training data collection according to claim 1, characterized in that: S1 collects the communication data between users and the customer service system, including voice records, text chat records, user feedback, waiting time and resolution time, and processes them using speech-to-text and natural language processing technology to form communication data; The voice recording data is specifically generated by capturing the voice interaction content between the user and the customer service system to generate a voice file, including the audio file of the user's questions and the customer service's responses, and the timestamp of each interaction marked in the voice recording, including recording and acquisition through a language recording device; Text chat records: The chat record function of the customer service system records all text conversations between users and customer service in real time, including the messages entered by users, the customer service replies, and the timestamp of each message; User feedback: Collect user feedback on the customer service experience through questionnaires or rating systems after each interaction, including collecting user ratings as a manifestation of user feedback; The waiting time and resolution time are automatically tracked by marking the time from the user initiating the request to the first response, as well as the total time for problem resolution, to obtain the waiting time WT and resolution time ST, and form the interaction data group UID, specifically UID = {voice record, text chat record, user feedback, WT, ST}.

3. The AI ​​model training method based on training data collection according to claim 2 is characterized in that: in, S2 also includes S21 and S22; S21, extracting features from the interaction data group UID to obtain user interaction feature information, question type feature information, feedback score feature information, user status feature information, and solved problem feature information; S22, perform feature information standardization and feature information conversion on user interaction feature information, question type feature information, feedback score feature information, user status feature information and solved problem feature information, obtain total response TAT, question category QTY, user evaluation EVA, customer feedback CTF and problem solving rate RSR, and form a feature vector F, specifically F = {TAT, QTY, EVA, CTF, RSR}; The total response TAT is obtained by processing the user interaction feature information using TAT=WT+ST; The question category QTY is obtained by analyzing the question type feature information using natural language processing technology, and the question category information is obtained specifically through the following S221, S222, S223 and S224; The user evaluation EVA extracts the rating data from the user feedback data in the feedback rating feature information, and performs a normalization process to adjust the range of the rating data to between 0 and 1. The normalization process is specifically performed through the following calculation formula: In the formula, UR represents the actual score given by the user, MaxR represents the upper limit score of the system, and MinR represents the lower limit score of the system; Customer Feedback CTF analyzes user status feature information by using natural language processing technology and sentiment analysis models to obtain user feedback degree values ​​ranging from 0 to 1. Sentiment analysis models include VADER, TextBlob, and BERT models. The problem resolution rate (RSR) is obtained by extracting the problem resolution status from the feature information of the resolved problems, including resolved and unresolved, and obtaining the subscript values ​​1 and 0, and by counting the ratio of the total number of problems to the total number of resolved problems.

4. The AI ​​model training method based on training data collection according to claim 3 is characterized in that: in, S221, S222, S223 and S224 are used to obtain the content of the question category: S221, text preprocessing, performing word segmentation, stop word removal and stem extraction on the text chat records; S222, feature extraction, using the bag-of-words model and word embedding to convert text into feature vectors; S223, classifier training, applying supervised learning algorithms, including SVM, random forest and deep learning models, to the labeled training data to identify the problem category; S224, category labeling, input the feature vector of each user interaction text into the classifier to obtain the question category QTY.

5. The AI ​​model training method based on training data collection according to claim 1, characterized in that: in, S3 also includes S31 and S32; S31, preliminarily classifying the user's communication data by using the feature vector F, and obtaining the proportion of high-frequency problems and the proportion of low-frequency problems, and then performing cluster analysis on the feature vector F to obtain the problem category cluster center Ck; Among them, cluster analysis is performed through the following calculation formula: In the formula, Ck represents the cluster center of the problem category, specifically the cluster center of the k-th category, Nk represents the number of samples belonging to the k-th category, Cluster, k represents the sample set of the k-th category, and Fi represents the feature vector of the i-th sample in the feature vector F.

6. The AI ​​model training method based on training data collection according to claim 5 is characterized in that: S32, iteratively control the problem category cluster center Ck, obtain the updated cluster center Ck, new and calculate the distance d from sample i to the k category center; The updated cluster center Ck,new is obtained by the following calculation formula: In the formula, Wi represents the weight value of the i-th sample; The distance d is obtained by the following calculation formula: Where, d i，k represents the distance d from sample i to the center of category k, and T represents the transpose of the sentence, which is specifically used to convert a column vector into a row vector.

7. The AI ​​model training method based on training data collection according to claim 6 is characterized in that: in, S4 also includes S41 and S42; S41, fit the feature vector F and the problem category cluster center Ck using the AI ​​model, identify and communicate data, obtain the classification result CTi of the sample, and trigger the corresponding automatic response strategy Rj, synchronously assign the automatic response strategy Rj content to the classification result CTi of the sample, and store it in the database; The classification result CTi is obtained by the following calculation formula: In the formula, λ represents the regularization factor, which is used to control the weight value of the distance d from sample i to the center of category k, ||Fi-Ck|| 2 It represents the sum of squares of the feature vector of the i-th sample in the feature vector F and the cluster center of the k-th category, which is specifically used to measure the fit between the feature vector Fi and the cluster Ck; The automatic response strategy Rj is allocated by the following calculation formula: Rj=Strategy map (CTi); Where Rj represents the jth automatic response strategy R, Strategy map Represents a strategy mapping function, which selects the corresponding automatic response strategy Rj according to the classification result CTi.

8. The AI ​​model training method based on training data collection according to claim 7 is characterized in that: S42, performing a secondary evaluation on the triggering of the corresponding automatic response strategy Rj according to the obtained allocation sample category Rj, obtaining the secondary evaluation triggering result, and triggering the execution of the automatic response strategy Rj according to the secondary evaluation triggering result; The secondary evaluation trigger result is triggered by the following trigger judgment formula: Wherein, TG represents the return signal of the trigger judgment formula, specifically, TG return signal 1 represents triggering, TG return signal 0 represents non-triggering, TJL represents the preset trigger judgment threshold; The execution of the automatic response strategy Rj includes sending FAQs, transferring to manual customer service, providing service status updates, sending questionnaires, providing additional help links, sending problem confirmations and system self-diagnosis.

9. The AI ​​model training method based on training data collection according to claim 7, characterized in that: in, S5 also includes S51 and S52; S51, evaluating the automatic response strategy Rj, counting the performance of the automatic response strategy Rj in different time periods and different problem types, and obtaining a system performance evaluation index P; The system performance evaluation index P is obtained by the following calculation formula: In the formula, EL represents the success rate of response strategy execution, TL represents the average response time of the response strategy, and SL represents the user satisfaction score of the response strategy execution; Among them, the execution success rate EL is The calculation formula is obtained, where Nsuccess represents the number of successfully resolved problems, and Ntotal represents the total number of automatic response strategies; Average response time TL passed The calculation formula is obtained, where T Rj，w represents the response time of the jth automatic response strategy R for the wth user; User satisfaction rating SL passed Calculation formula obtained, Nfeedback represents the total number of user feedback collected, SL Rj，w represents the user feedback score of the jth automatic response strategy R for the wth user S52, matching the calculated system performance evaluation index P with a preset system operation performance status evaluation threshold YZ to obtain a system operation status evaluation result; The system operation status evaluation result is Matching method acquisition; When the Status match return signal is 1, the system operation status evaluation result is qualified; When the Status match return signal is 0, the evaluation result of the system operation status is unqualified.

10. The AI ​​model training method based on training data collection according to claim 7 is characterized in that: S6. Triggering the iterative control of the automatic response strategy Rj according to the evaluation result of the system operation status is triggered in the following way: When the system operation status evaluation result is qualified and the Status match return signal is 1, iterative control is not triggered; When the system operation status evaluation result is unqualified and the Status match return signal is 0, iterative regulation is triggered, including adjusting response rules, optimizing trigger conditions, updating strategy parameters, and adjusting model hyperparameters; Adjusting model hyperparameters includes regulating the AI ​​model training data category, implementing cross-validation technology and data enhancement technology to adjust the effectiveness of the AI ​​model’s training data.

Images