Large model-based scene interaction system and method

Abstract

The invention discloses a scene interaction system and method based on a large model, relates to the technical field of artificial intelligence, and solves the problems that in the prior art, customer service performance is not considered, dynamic adjustment is achieved, personalized summary suggestions and diversified problem description are lacked, customer service training time is long, the training effect is poor, and user experience is poor. And the efficiency of the interaction system is relatively low. Generating question and answer data according to the background data; constructing a figure portrait according to the historical dialogue data and generating question data according to the figure portrait; generating customer service performance indexes according to the customer service answer data and adjusting question levels; a summary report is generated according to question data and customer service answer data, multi-character figure portraits are constructed by means of historical dialogue data, multi-modal and multi-character question data can be generated, customer service personnel are assisted to train dialogue content and modes in advance, expression indexes are generated according to the customer service answer data, and the difficulty of the next round of questions is dynamically adjusted. Customer service training is more efficient and accurate, and the efficiency of an interaction system is improved.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, and specifically relates to a scenario interaction system and method based on a large model. Background Art

[0002] The customer service dialogue training system is an intelligent system designed specifically to improve the dialogue ability and service level of customer service personnel. It combines advanced data analysis, artificial intelligence, and natural language processing technologies, aiming to help customer service personnel better master communication skills, product knowledge, and problem-solving abilities by simulating real dialogue scenarios, thereby improving customer satisfaction and the overall efficiency of the interaction system. By refining problems and interacting with customer service in different scenarios, the purpose of training customer service is achieved.

[0003] In the process of realizing customer service training through scenario interaction in the prior art, a fixed-question dialogue training method is usually adopted, lacking consideration for dynamically adjusting questions according to the performance of customer service in answering questions and providing personalized summary suggestions, and lacking diversification in problem descriptions, resulting in a long customer service training time and poor training effects, and further leading to low efficiency of the scenario interaction system; therefore, further improvement of the scenario interaction system is still required. Summary of the Invention

[0004] This application aims to solve at least one of the technical problems existing in the prior art; for this purpose, this application proposes a scenario interaction system and method based on a large model, which is used to solve the technical problems in the prior art that lack consideration for dynamically adjusting problems according to customer service performance, lack of personalized summary suggestions, and lack of diversification in problem descriptions, resulting in a long customer service training time and poor training effects, and low efficiency of the interaction system.

[0005] To achieve the above object, the first aspect of this application provides a scenario interaction system based on a large model, including: a data collection module, a data analysis module, an early warning module, and a database;

[0006] The data collection module: obtains background data and historical dialogue data through data collection devices;

[0007] The data analysis module: generates question-and-answer data according to background data; constructs a user profile according to historical dialogue data; generates question data according to the user profile and question-and-answer data; obtains customer service answer data in real time, generates a customer service performance index according to the customer service answer data and adjusts the question level; generates a summary report according to the question data and the customer service answer data;

[0008] The early warning module: makes a prompt according to the alarm signal and contacts the management personnel;

[0009] The database is used to store the data of each module and the historical data required for training the model.

[0010] Through the above steps, this application constructs a portrait of a person with diverse personality characteristics by leveraging historical conversation data; in the question generation phase, it can generate multimodal question data that fits different personalities, which is of great benefit to the customer service in pre-training the conversation content and communication methods; in addition, based on the customer service answer data, a customer service performance index is generated, and according to this index, the difficulty level of the next round of questions is dynamically adjusted, enabling customer service personnel to master relevant skills more quickly and accurately during training, thereby improving the operating efficiency of the interaction system.

[0011] Further, generating the Q&A data according to the background data includes:

[0012] Obtain background data; the background data includes background tags and background content;

[0013] Input the background tags and background content into an entity relationship extraction model to obtain a number of entities and a number of relationships; the entity relationship extraction model is constructed through a pre-trained large model;

[0014] Input the number of entities and the number of relationships into a graph database to obtain a background knowledge graph;

[0015] Define template questions; the template questions include attribute questions and relationship questions;

[0016] Input the number of entities, the number of relationships, and the template questions into a large model to obtain a number of Q&A data; the large model includes the GPT-4 model; the Q&A data includes questions and their corresponding answers.

[0017] Further, the entity relationship extraction model is constructed through a pre-trained large model, including:

[0018] Obtain a pre-trained large model, as well as a number of historical background tags, historical background content, and their corresponding historical entities and relationships;

[0019] Divide the number of historical background tags, historical background content, and their corresponding historical entities and relationships into training data, validation data, and test data; perform data preprocessing on the training data, validation data, and test data to obtain a training set, a validation set, and a test set;

[0020] Select the pre-trained large model as the base model;

[0021] Train the base model with the training set, and adjust the learning rate and hyperparameters on the validation set to obtain a pre-trained model;

[0022] Verify the pre-trained model on the test set, and finally obtain an entity relationship extraction model that takes the input background tags and background content and outputs a number of entities and a number of relationships.

[0023] Further, constructing a user portrait based on historical conversation data includes:

[0024] Obtaining historical conversation data; the historical conversation data includes historical user questions and historical customer service answers;

[0025] Performing desensitization data processing on historical user questions to obtain desensitized historical user questions;

[0026] Extracting a number of user personality characteristics through a feature extraction method;

[0027] Calculating a number of personality weights according to the ratio of the number of a number of user personality characteristics to the sum of the number of a number of user personality characteristics;

[0028] Quantifying user personality characteristics through the OCEAN model to obtain quantified user personality characteristics;

[0029] Using the DBSCAN algorithm to cluster user personality characteristics to obtain clustered user personality characteristics;

[0030] Creating a virtual user portrait; inputting user personality characteristics into the virtual user portrait to obtain a user portrait.

[0031] Further, generating question data based on the user portrait and Q&A data includes:

[0032] Obtaining the user portrait, Q&A data, and a number of personality weights; the Q&A data includes questions, answers, and question levels; the user portrait includes a number of personality characteristics;

[0033] Creating a candidate set XJ; the candidate set includes a number of personality characteristics;

[0034] Creating a weight set QJ; the weight set is the personality weights corresponding to a number of personality characteristics;

[0035] Obtaining the finally selected personality characteristics through the roulette wheel algorithm;

[0036] Inputting the personality characteristics, questions, and question levels into a question generation model to obtain a number of question data;

[0037] Among them, the question generation model is constructed through a multi-task learning model, including:

[0038] Obtaining a number of historical personality characteristics, historical questions, historical question levels, and their corresponding historical question levels and historical question data; the historical question data is multi-modal data, including voice and text;

[0039] Divide a number of historical personality traits, historical problems, historical problem levels, and their corresponding historical problem levels and historical problem data into training data, validation data, and test data; perform data preprocessing on the training data, validation data, and test data to obtain a training set, a validation set, and a test set;

[0040] Select a multi-task learning model as the basic model;

[0041] Train the basic model with the training set, and adjust the learning rate and hyperparameters on the validation set to obtain a pre-trained model;

[0042] Verify the pre-trained model on the test set, and finally obtain a problem generation model with personality traits, problems, and problem levels as inputs and problem levels and their corresponding problem data as outputs.

[0043] Further, the generating a customer service performance index based on the customer service answer data and adjusting the problem level includes:

[0044] Real-time obtain the customer service answer data, its corresponding problem data, problem level, and answer;

[0045] Obtain the semantic similarity YXD through the BERT model; obtain the number of several entities in the customer service answer data and the answer through the BERT-NER model, and obtain the corresponding comprehensive matching rate F1 through non-linearly combining the number of several entities;

[0046] Through the formula Calculate the response speed XS; where BS is the standard speed, SS is the actual speed, and max() represents the maximum value operation;

[0047] Obtain the emotional matching degree QPD through the large model;

[0048] Through the formula KBZ = ln(1 + YXD×F1)×e ω×XS×QPD Calculate the customer service performance index KBZ; where ω is the service quality influence coefficient, ω>0;

[0049] Adjust the problem level according to the customer service performance index.

[0050] This application calculates various parameter values required for evaluating the customer service performance through multiple methods, analyzes the service quality when the customer service answers questions through multiple parameter values, can more intuitively feel the customer service performance, provides a strong judgment standard for the problem generation of subsequent conversations, and provides data support for the generation of subsequent comprehensive customer service performance reports.

[0051] Further, the adjusting the problem level according to the customer service performance index includes:

[0052] Obtain the customer service performance index and the problem level WD; the problem level refers to the problem level corresponding to the current customer service's answer to the question;

[0053] When the customer service performance index for N consecutive times is greater than the first upgrade threshold; determine whether the problem level is the highest level. If so, do nothing. If not, increase the problem level by one level;

[0054] Otherwise, when the customer service performance index is greater than the second upgrade threshold; determine whether the problem level is the highest level. If so, do nothing. If not, increase the problem level by one level;

[0055] Otherwise, when the customer service performance index is less than the second downgrade threshold; determine whether the problem level is the lowest level. If so, do nothing. If not, decrease the problem level by one level and generate a warning signal for poor customer service performance;

[0056] Otherwise, when the customer service performance index for M consecutive times is less than the first downgrade threshold; determine whether the problem level is the lowest level. If so, do nothing. If not, decrease the problem level by one level and generate a warning signal for poor customer service performance; where M and N are integers, both M and N are greater than 0; and the second upgrade threshold is greater than the first upgrade threshold, and the second downgrade threshold is less than the first downgrade threshold.

[0057] Furthermore, generating a summary report based on the question data and the customer service answer data includes:

[0058] Obtain the question data and its corresponding answers, customer response data, customer service performance index, and several parameter values composed thereof; the several parameter values include semantic similarity, comprehensive matching rate, response speed, and emotional matching degree;

[0059] Integrate the question data and its corresponding answers, customer response data, customer service performance index, and several parameter values composed thereof into comprehensive data;

[0060] Input the comprehensive data into the analysis and summary large model to obtain a summary report; the summary report includes an answer summary and answer optimization suggestions; the analysis and summary large model is constructed through an artificial intelligence model.

[0061] Furthermore, the analysis and summary large model is constructed through an artificial intelligence model, including:

[0062] Obtain several historical comprehensive data and their corresponding historical summary reports;

[0063] Divide several historical comprehensive data and their corresponding historical summary reports into training data, validation data, and test data; perform data preprocessing on the training data, validation data, and test data to obtain a training set, a validation set, and a test set;

[0064] Select an artificial intelligence model as the basic model;

[0065] Train the basic model with the training set, and adjust the learning rate and hyperparameters on the validation set to obtain the pre-trained model;

[0066] Verify the pre-trained model on the test set, and finally obtain an analysis and summary large model with comprehensive data as input and a summary report as output.

[0067] This application evaluates various parameters of the customer service in continuous conversations, and thus obtains a summary report through the pre-trained analysis and summary large model. The summary report includes a summary of the continuous conversation content, and more importantly, provides improvement suggestions for the deficiencies of the customer service when answering questions, enabling the customer service to learn and self-improve targeted according to its own shortcomings, which helps to speed up the training progress of the customer service, improve the efficiency of customer service training, and enhance the efficiency of the interaction system.

[0068] Another aspect of the present invention provides a scenario interaction method based on a large model, including:

[0069] S0: Obtain background data and historical conversation data;

[0070] S1: Generate Q&A data according to the background data; construct a portrait of the person based on the historical conversation data;

[0071] S2: Generate question data according to the portrait of the person and the Q&A data; obtain the customer service answer data in real time, generate a customer service performance index according to the customer service answer data and adjust the question level;

[0072] S3: Generate a summary report according to the question data and the customer service answer data;

[0073] S4: Make a prompt according to the alarm signal and contact the management staff.

[0074] Compared with the prior art, the beneficial effects of this application are:

[0075] 1. This application generates Q&A data according to the background data; constructs a portrait of the person based on the historical conversation data; generates question data according to the portrait of the person and the Q&A data; obtains the customer service answer data in real time, generates a customer service performance index according to the customer service answer data and adjusts the question level; generates a summary report according to the question data and the customer service answer data, constructs a portrait of a person with multiple personality characteristics based on the historical conversation data, and can generate multi-modal and multi-personality question data when generating questions, which helps to help customer service personnel train the conversation content and methods in advance; and generates a customer service performance index based on the customer service answer data, and dynamically adjusts the difficulty level of the next round of questions accordingly, enabling the customer service to master more quickly and accurately during training, and improving the efficiency of the interaction system.

[0076] 2. This application extracts the user's personality traits during each conversation through historical conversation data, and constructs a user profile based on this. The user profile includes various common personality traits. Combining the user profile with the Q&A data can make the Q&A data more in line with the actual situation, which is more helpful for training the answering ability and skills of customer service, and improving the efficiency of the interaction system.

[0077] 3. This application evaluates the status of the customer service performance index to achieve the adaptive generation of the next round of question levels. The system can accurately match the customer service ability level, achieve the goals of self-adaptive training difficulty and quantifiable ability improvement, speed up the training progress of customer service, and improve the efficiency of the interaction system. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0079] Figure 1 Schematic diagram of the principle of a scenario interaction system based on a large model according to the present application;

[0080] Figure 2 Flowchart of a scenario interaction method based on a large model according to the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0081] The following will clearly and completely describe the technical solutions of the present application in conjunction with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0082] Please refer to Figure 1 , an embodiment of the first aspect of the present application provides a scenario interaction system based on a large model, including: a data acquisition module, a data analysis module, an early warning module, and a database;

[0083] Data acquisition module: Obtain background data and historical conversation data through data acquisition devices; the data acquisition devices include several sensors, etc.;

[0084] Data analysis module: Generate Q&A data based on background data, where the Q&A data are the questions and answers used for conversations with customer service representatives; construct a user profile based on historical conversation data, and the user profile is data including various personality characteristics of a person, and questions with various personality characteristics can be generated using the user profile; generate question data based on the user profile and the Q&A data, and the question data refers to questions with personality characteristics of a person; obtain customer service answer data in real time, generate a customer service performance index based on the customer service answer data and adjust the question level, where the customer service performance index refers to the degree of performance of a customer service representative when answering the question data provided by the system, and adjusting the question level means adjusting the difficulty of questions in the next round of conversation according to the historical performance of the customer service representative; generate a summary report based on the question data and the customer service answer data; the summary report includes an answer summary and answer optimization suggestions;

[0085] Early warning module: Make a prompt according to the alarm signal and contact the management personnel; the alarm signals include early warning signals of poor customer service performance, etc.;

[0086] The database is used to store the data of each module and the historical data required for training the model.

[0087] In this embodiment, generating Q&A data based on background data includes:

[0088] Obtain background data; the background data includes background tags and background content;

[0089] Input the background tags and background content into an entity relationship extraction model to obtain a number of entities and a number of relationships; the entity relationship extraction model is constructed through a pre-trained large model;

[0090] Input the number of entities and the number of relationships into a graph database to obtain a background knowledge graph; the graph database includes Neo4j database, etc.;

[0091] Define template questions; the template questions include attribute questions and relationship questions; the attribute questions are expressed as "What is the {attribute name} of {entity}?", etc., and the relationship questions are expressed as "Which {entities} will cause {problem}?", etc.;

[0092] Input the number of entities, the number of relationships and the template questions into a large model to obtain a number of Q&A data; the large model includes GPT-4 model; the Q&A data includes questions and their corresponding answers.

[0093] In this embodiment, by defining different template questions when extracting Q&A data corresponding to background data, the question data generated from the background data is more diversified, not limited to using simple questions in the process of training customer service representatives, and provides rich training data for customer service training.

[0094] In this embodiment, the entity relationship extraction model is constructed through a pre-trained large model, including:

[0095] Obtain a pre-trained large model, as well as a number of historical background tags, historical background content, and their corresponding historical entities and relationships;

[0096] Divide a number of historical background tags, historical background content, and their corresponding historical entities and relationships into training data, validation data, and test data; perform data preprocessing on the training data, validation data, and test data to obtain a training set, a validation set, and a test set; the ratio among the training set, the test set, and the validation set is 7:2:1;

[0097] Select the pre-trained large model as the base model; the pre-trained large model includes models such as the BiLSTM-CRF model;

[0098] Train the base model with the training set, and adjust the learning rate and hyperparameters on the validation set to obtain a pre-trained model;

[0099] Verify the pre-trained model on the test set, and finally obtain an entity relationship extraction model that takes background tags and background content as input and outputs a number of entities and a number of relationships.

[0100] Constructing a user portrait based on historical conversation data in this embodiment includes:

[0101] Obtain historical conversation data; the historical conversation data includes historical user questions and historical customer service answers;

[0102] Perform desensitization data processing on the historical user questions to obtain desensitized historical user questions; performing desensitization data processing is to prevent the generation of some sensitive words and cause adverse consequences;

[0103] Extract a number of user personality characteristics through feature extraction; use the BERT model fine-tuned on the Personality20 dataset for feature extraction;

[0104] Calculate a number of personality weights according to the ratio of the number of a number of user personality characteristics to the sum of the number of a number of user personality characteristics;

[0105] Quantify the user personality characteristics through the OCEAN model to obtain quantified user personality characteristics;

[0106] Use the DBSCAN algorithm to cluster the user personality characteristics to obtain clustered user personality characteristics;

[0107] Create a virtual user portrait; input the user personality characteristics into the virtual user portrait to obtain the user portrait.

[0108] Generating question data based on the user portrait and Q&A data in this embodiment includes:

[0109] Obtain a portrait of a person, Q&A data, and several personality weights; the Q&A data includes questions, answers, and question levels; the portrait of the person includes several personality traits, etc.; the personality traits include impatient type, cautious type, casual type, professional type, etc.

[0110] Create a candidate set XJ; the candidate set includes several personality traits.

[0111] Create a weight set QJ; the weight set is the personality weights corresponding to several personality traits.

[0112] Obtain the finally selected personality trait through the roulette algorithm; the personality trait selected in this way is selected based on the personality weights corresponding to several historical personality traits. The higher the personality weight, the greater the probability of being selected, so that the finally generated question data conforms to the historical personality trait distribution.

[0113] Input the personality trait, question, and question level into the question generation model to obtain several question data; the higher the question level, the more difficult the generated question data.

[0114] Among them, the question generation model is constructed through a multi-task learning model, including:

[0115] Obtain several historical personality traits, historical questions, historical question levels, and their corresponding historical question levels and historical question data; the historical question data is multi-modal data, including voice and text.

[0116] Divide several historical personality traits, historical questions, historical question levels, and their corresponding historical question levels and historical question data into training data, validation data, and test data; perform data preprocessing on the training data, validation data, and test data to obtain a training set, a validation set, and a test set; the ratio among the training set, test set, and validation set is 7:2:1.

[0117] Select a multi-task learning model as the basic model; the multi-task learning model is constructed by combining the ChatGLM3 model and the StyleTTS2 model. Among them, the ChatGLM3 model is used to output question data in text modality, and the StyleTTS2 model is used to output question data in voice modality.

[0118] Train the basic model with the training set, and adjust the learning rate and hyperparameters on the validation set to obtain a pre-trained model.

[0119] Verify the pre-trained model on the test set, and finally obtain a question generation model with personality traits, questions, and question levels as inputs and question levels and their corresponding question data as outputs.

[0120] In this embodiment, by leveraging historical conversation data, the personality traits demonstrated by the user during each conversation are accurately extracted, and then a comprehensive and detailed user profile is constructed. This user profile covers various common personality trait types and can highly restore the true characteristics of the user. By deeply integrating such a rich user profile with the Q&A data, the Q&A data can be made more in line with the actual conversation scenario, which helps to train the answering ability and communication skills of customer service personnel targeted, improve the operation efficiency of the interaction system, and bring a better and more efficient interaction experience to users.

[0121] This embodiment generates a customer service performance index based on customer service answer data and adjusts the question level, including:

[0122] Real-time obtain customer service answer data and its corresponding question data, as well as the question level and answer;

[0123] Obtain the semantic similarity YXD through the BERT model; obtain the number of several entities in the customer service answer data and the answer through the BERT-NER model, and obtain the corresponding comprehensive matching rate F1 through non-linearly combining the number of several entities;

[0124] The specific calculation formula of the comprehensive matching rate F1 is expressed as: Among them, Precision represents the ratio between the number of correctly matched entities and the total number of entities in the customer service answer data, and Recall represents the ratio between the number of correctly matched entities and the total number of entities in the answer;

[0125] Through the formula Calculate the response speed XS; where BS is the standard speed, and the standard data is set according to experience; SS is the actual speed, and max() represents the operation of taking the maximum value; as the actual speed corresponding to the customer service answer data is faster, the corresponding value less than the speed value is larger;

[0126] In another embodiment, the standard speed can be set according to the question level; the lower the question level, the smaller the corresponding standard speed, and the higher the question level, the larger the corresponding standard speed;

[0127] Obtain the emotion matching degree QPD through a large model; the large model includes the VADER model and the wav2vec2.0+LSTM model, etc.; among them, the wav2vec2.0+LSTM model is used to output the emotion matching degree corresponding to the voice question, and the VADER model is used to output the emotion matching degree corresponding to the text question;

[0128] Through the formula KBZ = ln(1 + YXD × F1) × e ω×XS×QPDCalculate the customer service performance index KBZ; where ω is the service quality impact coefficient, ω > 0, and the specific value is set according to experience. In this embodiment, ω is set to 2; the setting of ω is to amplify the positive impact of service quality. When the semantic similarity and comprehensive matching degree of the customer service's answer data are higher, it indicates that the correctness of the customer service's answer to the question is higher. When the response speed corresponding to the customer's answer data is faster and the emotional matching degree is higher, it indicates that the customer service's answering attitude is better. Therefore, the customer service performance index will increase accordingly;

[0129] Adjust the question level according to the customer service performance index.

[0130] Adjusting the question level according to the customer service performance index in this embodiment includes:

[0131] Obtain the customer service performance index and the question level WD; the question level refers to the question level corresponding to the current customer service's answer to the question;

[0132] When the customer service performance index for N consecutive times is greater than the first upgrade threshold; determine whether the question level is the highest level. If it is, do nothing. If not, increase the question level by one level; considering the good historical performance of the customer service when answering questions, it can be considered to increase the question level during the next conversation, which helps to better train the customer service's conversation level;

[0133] Otherwise, when the customer service performance index is greater than the second upgrade threshold; determine whether the question level is the highest level. If it is, do nothing. If not, increase the question level by one level; by only judging that the customer service performance index reaches an excellent state in this way, the question level of the next conversation question can be increased, enabling the customer service to quickly master the training requirements;

[0134] Otherwise, when the customer service performance index is less than the second downgrade threshold; determine whether the question level is the lowest level. If it is, do nothing. If not, lower the question level by one level and generate a warning signal for poor customer service performance; considering that the customer service performs very poorly when answering questions, the question level needs to be lowered to allow the customer service to gradually adapt to the training content;

[0135] Otherwise, when the customer service performance index for M consecutive times is less than the first downgrade threshold; determine whether the question level is the lowest level. If it is, do nothing. If not, lower the question level by one level and generate a warning signal for poor customer service performance; where M and N are integers, both M and N are greater than 0, and the specific values are set according to experience. In this embodiment, both M and N are set to 3; and the second upgrade threshold is greater than the first upgrade threshold, and the second downgrade threshold is less than the first downgrade threshold; the specific values of the first upgrade threshold, the second upgrade threshold, the first downgrade threshold, and the second downgrade threshold are set according to experience.

[0136] This embodiment adopts a method of comprehensively evaluating the customer service performance index to achieve an adaptive generation mechanism for the next-level problem level, which can highly accurately match the actual ability level of customer service personnel, and effectively achieve the dual goals of adaptive adjustment of training difficulty and quantifiable evaluation of ability improvement. It not only greatly speeds up the training progress of customer service personnel, enabling them to master the required skills more quickly, but also effectively improves the overall operation efficiency of the interaction system, providing users with a smoother and more efficient service experience.

[0137] Generating a summary report based on the problem data and the customer service answer data in this embodiment includes:

[0138] Obtaining the problem data and its corresponding answers, customer response data, customer service performance index, and several parameter values that make it up; the several parameter values include semantic similarity, comprehensive matching rate, response speed, and emotional matching degree;

[0139] Integrating the problem data and its corresponding answers, customer response data, customer service performance index, and several parameter values that make it up into comprehensive data;

[0140] Inputting the comprehensive data into an analysis and summary large model to obtain a summary report; the summary report includes answer summary and answer optimization suggestions; the analysis and summary large model is constructed through an artificial intelligence model.

[0141] The analysis and summary large model is constructed through an artificial intelligence model, including:

[0142] Obtaining several historical comprehensive data and their corresponding historical summary reports;

[0143] Dividing the several historical comprehensive data and their corresponding historical summary reports into training data, validation data, and test data; performing data preprocessing on the training data, validation data, and test data to obtain a training set, a validation set, and a test set; the ratio among the training set, the test set, and the validation set is 7:2:1;

[0144] Selecting an artificial intelligence model as the basic model; the artificial intelligence model includes the GPT-4 model, etc.;

[0145] Training the basic model with the training set and adjusting the learning rate and hyperparameters on the validation set to obtain a pre-trained model;

[0146] Finally obtaining an analysis and summary large model with comprehensive data as the input and a summary report as the output by validating the pre-trained model on the test set.

[0147] Please refer to Figure 2 , another embodiment of this application provides a scenario interaction method based on a large model, including:

[0148] S0: Obtaining background data and historical conversation data;

[0149] S1: Generate Q&A data based on background data; construct a user profile based on historical conversation data;

[0150] S2: Generate question data based on the user profile and Q&A data; obtain the customer service response data in real time, generate a customer service performance index based on the customer service response data, and adjust the question level;

[0151] S3: Generate a summary report based on the question data and the customer service response data;

[0152] S4: Make a prompt according to the alarm signal and contact the management personnel.

[0153] Some of the data in the above formula is calculated by removing the dimension and taking its numerical value. The formula is obtained by software simulation of a large amount of collected data to get a formula closest to the actual situation; the preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

[0154] The working principle of this application: By obtaining background data and historical conversation data; generating Q&A data based on the background data; constructing a user profile based on the historical conversation data; generating question data based on the user profile and Q&A data; obtaining the customer service response data in real time, generating a customer service performance index based on the customer service response data, and adjusting the question level; generating a summary report based on the question data and the customer service response data; making a prompt according to the alarm signal and contacting the management personnel. By constructing a multi-personality user profile with historical conversation data, multi-modal and multi-personality question data can be generated, which helps customer service personnel to pre-train the conversation content and methods. Generate a performance index based on the customer service response data, dynamically adjust the difficulty of the next round of questions, make customer service training more efficient and accurate, improve the efficiency of the interaction system, and avoid the problems in the prior art that lack consideration of customer service performance to achieve dynamic adjustment of questions, lack personalized summary suggestions and diverse question descriptions, resulting in a long customer service training time and poor training effect, and low efficiency of the interaction system.

[0155] The above embodiments are only used to illustrate the technical method of this application and not to limit it. Although this application has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical method of this application can be modified or equivalently replaced without departing from the spirit and scope of the technical method of this application.

Claims

1. A scene interaction system based on a large model, characterized in that, Including: A data acquisition module, a data analysis module, an early warning module, and a database; The data acquisition module: obtains background data and historical conversation data through data acquisition devices; The data analysis module: generates Q&A data based on the background data; constructs a user profile based on the historical conversation data; generates question data based on the user profile and the Q&A data; Obtains the customer service answer data in real time, generates a customer service performance index based on the customer service answer data and adjusts the question level; generates a summary report based on the question data and the customer service answer data.

2. The scene interaction system based on a large model according to claim 1, wherein, The generation of Q&A data based on the background data includes: Obtains the background data; the background data includes background tags and background content; Inputs the background tags and background content into an entity relationship extraction model to obtain a number of entities and a number of relationships; the entity relationship extraction model is constructed through a pre-trained large model; Inputs the number of entities and the number of relationships into a graph database to obtain a background knowledge graph; Defines template questions; the template questions include attribute questions and relationship questions; Inputs the number of entities, the number of relationships, and the template questions through a large model to obtain a number of Q&A data; the large model includes the GPT-4 model; the Q&A data includes questions and their corresponding answers.

3. The scene interaction system based on a large model according to claim 2, wherein, The entity relationship extraction model is constructed through a pre-trained large model, including: Obtains a pre-trained large model, as well as a number of historical background tags, historical background content, and their corresponding historical entities and relationships; Divides the number of historical background tags, historical background content, and their corresponding historical entities and relationships into training data, validation data, and test data; performs data preprocessing on the training data, validation data, and test data to obtain a training set, a validation set, and a test set; Selects the pre-trained large model as the base model; Trains the base model through the training set, and adjusts the learning rate and hyperparameters on the validation set to obtain a pre-trained model; Verifies the pre-trained model on the test set, and finally obtains an entity relationship extraction model that inputs background tags and background content and outputs a number of entities and a number of relationships.

4. The scenario interaction system based on a large model according to claim 1, characterized in that, The construction of a user profile based on the historical conversation data includes: Obtains the historical conversation data; the historical conversation data includes historical user questions and historical customer service answers; Performs desensitization data processing on the historical user questions to obtain desensitized historical user questions; Extracts a number of user personality characteristics through feature extraction methods; Calculates a number of personality weights based on the ratio of the number of a number of user personality characteristics to the sum of the number of a number of user personality characteristics; Quantifies the user personality characteristics through the OCEAN model to obtain quantified user personality characteristics; Uses the DBSCAN algorithm to cluster the user personality characteristics to obtain clustered user personality characteristics; Creates a virtual user profile; inputs the user personality characteristics into the virtual user profile to obtain a user profile.

5. The scene interaction system based on a large model according to claim 1, wherein The generation of question data based on the user profile and the Q&A data includes: Obtains the user profile, the Q&A data, and a number of personality weights; the Q&A data includes questions, answers, and question levels; the user profile includes a number of personality characteristics; Creates a candidate set XJ; the candidate set includes a number of personality characteristics; Create a weight set QJ; the weight set is the personality weights corresponding to a number of personality traits; Obtain the finally selected personality trait through the roulette algorithm; Input the personality trait, the question, and the question level into the question generation model to obtain a number of question data; Among them, the question generation model is constructed through a multi-task learning model, including: Obtain a number of historical personality traits, historical questions, historical question levels, and their corresponding historical question levels and historical question data; the historical question data is multi-modal data, including voice and text; Divide a number of historical personality traits, historical questions, historical question levels, and their corresponding historical question levels and historical question data into training data, validation data, and test data; perform data preprocessing on the training data, validation data, and test data to obtain a training set, a validation set, and a test set; Select a multi-task learning model as the basic model; Train the basic model with the training set, and adjust the learning rate and hyperparameters on the validation set to obtain a pre-trained model; Verify the pre-trained model on the test set, and finally obtain a question generation model with personality traits, questions, and question levels as inputs and question levels and their corresponding question data as outputs.

6. The scene interaction system based on a large model according to claim 1, characterized in that, The generating the customer service performance index according to the customer service answer data and adjusting the question level includes: Real-time obtain the customer service answer data, its corresponding question data, question level, and answer; Obtain the semantic similarity YXD through the BERT model; obtain the number of entities in the customer service answer data and the answer through the BERT-NER model, and obtain the corresponding comprehensive matching rate F1 through non-linearly combining the number of entities; Calculate the response speed XS through the formula where BS is the standard speed, SS is the actual speed, and max() represents the maximum value operation; Obtain the emotional matching degree QPD through a large model; Calculate the customer service performance index KBZ through the formula KBZ = ln(1 + YXD × F1) × e ω×XS×QPD where ω is the service quality impact coefficient, ω > 0; Adjust the question level according to the customer service performance index.

7. The scene interaction system based on a large model according to claim 6, characterized in that, The adjusting the question level according to the customer service performance index includes: Obtain the customer service performance index and the question level WD; the question level refers to the question level corresponding to the current customer service answering the question; When the customer service performance index for N consecutive times is greater than the first upgrade threshold; determine whether the question level is the highest level, if so, do nothing, otherwise, increase the question level by one level; Otherwise, when the customer service performance index is greater than the second upgrade threshold; determine whether the question level is the highest level, if so, do nothing, otherwise, increase the question level by one level; Otherwise, when the customer service performance index is less than the second downgrade threshold; determine whether the question level is the lowest level, if so, do nothing, otherwise, lower the question level by one level and generate a warning signal for poor customer service performance; Otherwise, when the customer service performance index for M consecutive times is less than the first downgrade threshold; determine whether the question level is the lowest level, if so, do nothing, otherwise, lower the question level by one level and generate a warning signal for poor customer service performance; where M and N are integers, both M and N are greater than 0; and the second upgrade threshold is greater than the first upgrade threshold, and the second downgrade threshold is less than the first downgrade threshold.

8. A scenario interaction system based on a large model according to claim 1, characterized in that, The generating a summary report according to the question data and the customer service answer data includes: Obtain the question data, its corresponding answer, customer answer data, customer service performance index, and several parameter values composed thereof; the several parameter values include semantic similarity, comprehensive matching rate, response speed, and emotional matching degree; Integrate the problem data, its corresponding answers, customer response data, customer service performance index, and several parameter values it consists of into comprehensive data; Input the comprehensive data into the analysis and summary large model to obtain a summary report; the summary report includes answer summaries and answer optimization suggestions; the analysis and summary large model is constructed through an artificial intelligence model.

9. The scene interaction system based on a large model according to claim 8, wherein The analysis and summary large model is constructed through an artificial intelligence model, including: Obtain a number of historical comprehensive data and their corresponding historical summary reports; Divide a number of historical comprehensive data and their corresponding historical summary reports into training data, validation data, and test data; perform data preprocessing on the training data, validation data, and test data to obtain a training set, a validation set, and a test set; Select an artificial intelligence model as the basic model; Train the basic model with the training set and adjust the learning rate and hyperparameters on the validation set to obtain a pre-trained model; Finally, obtain an analysis and summary large model with comprehensive data as the input and a summary report as the output through validating the pre-trained model on the test set.

10. A scene interaction method based on a large model, applied to a scene interaction system based on a large model according to any one of claims 1-9, characterized in that, Including: S0: Obtain background data and historical conversation data; S1: Generate Q&A data according to the background data; construct a persona based on the historical conversation data; Construct a persona based on the historical conversation data; S2: Generate problem data according to the persona and Q&A data; obtain customer service response data in real time, generate a customer service performance index according to the customer service response data, and adjust the problem level; S3: Generate a summary report according to the problem data and the customer service response data.

Images