Electric power customer service training method and system based on large model semantic abstract understanding capability

Abstract

The invention relates to an electric customer service training method and system based on large model semantic abstract understanding capability. The electric customer service training method comprises the following steps: simulating a customer role; the seat replies the recommendation; quality evaluation is replied; according to the invention, a customer role simulation and reply recommendation module based on a large model can generate standardized and diversified customer dialogue scenes and normative reply verbal skills, so that the consistency and stability of a training process are ensured; the reply content of the employee is automatically scored and analyzed through the reply quality evaluation module, targeted promotion suggestions are provided, and the training content and difficulty can be adjusted according to the actual performance of the employee; customized training schemes can be provided according to ability levels and learning progresses of different employees, and the employees are helped to rapidly improve customer service ability; and the adopted large model technology can process complex multiple rounds of dialogues, generate diversified customer roles and business scenes, and adapt to the actual requirements of electric customer service businesses.

Description

Technical Field

[0001] The present invention relates to an educational or training simulation device, and more particularly to a power customer service training method and system based on the semantic understanding and summarization ability of large models. Background Art

[0002] As an important bridge for communication between power enterprises and users, the service quality of power customer service directly affects user satisfaction and the enterprise image. With the rapid development of the power industry and the increasing diversification of user needs, power customer service employees need to possess higher professional qualities and communication skills. However, traditional customer service training methods have many limitations and cannot meet the requirements of modern power customer service operations. Currently, the main training methods for power customer service are as follows:

[0003] (1) Manual simulation training: Experienced trainers play the role of customers and have conversations with new employees for practice. Although this method can provide certain practical opportunities, it has problems such as low efficiency, high cost, and unstable training quality.

[0004] (2) Online course training: By recording video courses or writing text teaching materials, employees are taught customer service knowledge and skills. However, this method lacks interactivity and practicality and is difficult to help employees truly master customer service skills.

[0005] (3) Rule-based dialogue systems: Some enterprises have tried to use rule-based dialogue systems for customer service training, but these systems can usually only handle simple dialogue scenarios and cannot adapt to complex customer problems and diverse business requirements.

[0006] The disadvantages of the existing technologies are as follows:

[0007] Manual simulation training has low efficiency; since trainers need to guide new employees one by one, the training progress is slow and cannot meet the needs of large-scale training. The training quality is unstable; manual simulation of customer conversations is subjective and inconsistent, and the training effects of different trainers vary greatly, making it difficult to ensure the stability of training quality. There is a lack of personalized training; existing training methods cannot provide personalized training content and feedback according to the individual differences and learning progress of employees, which is not conducive to employees quickly improving their customer service capabilities. The training cost is high; traditional training requires a large amount of human and material resources, including the employment of trainers, the rental of training venues, etc., increasing the operating costs of enterprises; rule-based dialogue systems have great limitations; they can only handle simple dialogue scenarios and cannot adapt to complex customer problems and diverse business requirements, making it difficult to meet the actual needs of power customer service operations. Summary of the Invention

[0008] The present invention aims to provide a power customer service staff training system and method based on large model technology. Through processes such as customer role simulation, seat reply recommendation, and reply quality evaluation, it realizes efficient, stable, and personalized customer service training, reduces training costs, improves training effects, and enhances the professional qualities and service quality of power customer service staff.

[0009] In a first aspect, the present invention provides a power customer service training method based on the semantic understanding and summarization ability of a large model, including the following processes:

[0010] S100: Customer role simulation; generating a diverse interactive process between simulated real customers and customer service staff based on the large model; improving the authenticity and diversity of training;

[0011] S200: Seat reply recommendation; using the multi-round dialogue generation ability of the large model to provide real-time reply recommendations and business handling process guidance for customer service staff; helping employees quickly master standard speech patterns and business operations;

[0012] S300: Reply quality evaluation; constructing a customer service reply quality evaluation system based on the large model, automatically scoring and analyzing the reply content of employees, providing targeted improvement suggestions, and realizing quantitative evaluation and personalized improvement of training effects.

[0013] Preferably, the S100 includes the following processes:

[0014] S110: Construction of the first data set; specifically including the following steps:

[0015] S111: Collecting dialogue data for specific application business scenarios;

[0016] S112: Processing the samples;

[0017] S113: Coding and labeling the training samples according to the requirements of multi-round role-playing, enabling the large model to learn the call techniques of users; directly constructing labels including all the speech contents of users in multi-round conversations, which not only fully utilizes all the speech information of users but also does not involve duplicate calculations, being very efficient;

[0018] The coding and labeling of the training samples are as follows:

[0019] inputs = <user><seat><user><seat><user><seat>;

[0020] labels = <user><-100><user><-100><user><-100>;

[0021] Under this identification, the large model only learns the user's speech information, quickly masters the user's habitual speech, and enables it to generate customer conversations that meet the business requirements of power customer service; <-100> means that this data information only participates in the forward inference of the model and does not participate in the update of the model training gradient;

[0022] The model training is SFT training, and the LORA method is selected;

[0023] S120: Construct a set of prompt words for multi-role playing and multi-business scenarios;

[0024] Prompt word template: For example, you are a {age}{personality} {occupation}, {description of business scenario}, and you need to call the customer service specialist to ask about the situation. The starting inquiry words are {scenario speech};

[0025] S130: Role simulation Q&A; Perform role simulation Q&A based on the trained model; The specific process is as follows:

[0026] S131: Randomly select a prompt word from the set of prompt words for multi-role playing and multi-business scenarios, or select any prompt word for a specific business scenario based on business requirements;

[0027] S132: Send the prompt word and conversation context to the large model for a service request;

[0028] S133: Add the agent's reply content to the conversation context.

[0029] S134: Repeat S132 and S133 until the large model replies with the final business handling or query result.

[0030] Preferably, the S200 includes the following process:

[0031] S210: Construction of the second data set; The specific steps are as follows:

[0032] S211: Collect conversation data for specific application business scenarios;

[0033] S212: Perform sample processing;

[0034] S213: Encode and label the training samples according to the agent's reply requirements, so that the large model can learn the agent's reply speech;

[0035] The encoding and labeling of the training samples are as follows:

[0036] inputs = <user><agent><user><agent><user><agent>;

[0037] labels = <-100><agent><-100><agent><-100><agent>;

[0038] The model training is SFT training, and the LORA method is selected;

[0039] S220: Collaborate with business experts to set the business scenario configuration information set. Different scenarios have different slots and business request parameters, and are associated with the core business system to complete the feedback of business results;

[0040] S230: Recommend agent replies based on the trained model; The process includes the following:

[0041] S231: Based on the business scenario of the user's incoming call, that is, the training exercise type, select the corresponding scenario configuration information from the business scenario configuration information set;

[0042] S232: Based on the multi-round dialogue interaction information with the user, recommend the corresponding reply words and phrases, and update the slot values in the scenario configuration information in real time;

[0043] S233: Repeat S232 until the slot is filled and then reply "The business is being processed. Please wait a moment."

[0044] Preferably, the S300 includes the following process:

[0045] S310: Build a quality evaluation standard system, and the evaluation dimensions include:

[0046] Degree of problem solving, whether it has been solved or not;

[0047] Service perception, whether the user is satisfied or not;

[0048] Service attitude, positive or negative;

[0049] Service professionalism, professional, average or unfamiliar;

[0050] Context question and answer matching degree, which level belongs to high, medium or low;

[0051] S320: Evaluate the reply quality based on the trained model, which specifically includes the following process:

[0052] S321: For the agent's real-time question and answer, evaluate the reply quality and return the context question and answer matching degree and service professionalism information;

[0053] S322: For the single-call session context after the session ends, use the model to perform information element summarization and perform multi-dimensional rule scoring based on the element results;

[0054] S323: Statistically analyze the training evaluation data of employees and output subsequent training suggestions and improvement measures.

[0055] In a second aspect, the present invention provides a power customer service training system based on the semantic understanding and summarization ability of a large model, including:

[0056] A customer role simulation module: generating diverse customer roles and conversation scenarios based on a large model to simulate the interaction process between real customers and customer service employees;

[0057] A seat reply recommendation module: using the multi-turn conversation generation ability of a large model to provide real-time reply recommendations and business handling process guidance for customer service employees;

[0058] A reply quality evaluation module: constructing a customer service reply quality evaluation system based on a large model to automatically score and analyze the reply content of employees and provide targeted improvement suggestions.

[0059] The advantages of the present invention over the prior art are as follows:

[0060] The present invention realizes customer role simulation and seat reply recommendation through large model technology, and can support multiple employees for training at the same time. The present invention can significantly improve the training efficiency, reduce the dependence on manual trainers, realize large-scale and high-concurrency training, and quickly improve the customer service ability of employees.

[0061] Based on the customer role simulation and reply recommendation modules of the large model, the present invention can generate standardized and diverse customer conversation scenarios and standardized reply scripts to ensure the consistency and stability of the training process. The present invention can provide high-quality and standardized training content, reduce the influence of human factors on the training quality, and ensure the stability and reliability of the training effect.

[0062] The present invention automatically scores and analyzes the reply content of employees through the reply quality evaluation module and provides targeted improvement suggestions, and can adjust the training content and difficulty according to the actual performance of employees. The present invention can realize personalized training, provide customized training programs according to the ability levels and learning progress of different employees, and help employees quickly improve their customer service ability.

[0063] Based on large model technology, the present invention realizes automated training through a software system, reduces the dependence on manual trainers, and reduces the requirements for training venues and equipment. The present invention can significantly reduce the training cost, reduce the human and material input of enterprises, and improve the economy of training.

[0064] The large model technology adopted by the present invention can handle complex multi-turn conversations, generate diverse customer roles and business scenarios, and adapt to the actual needs of power customer service business. The present invention can effectively solve the limitations of the prior art, provide a more flexible and intelligent training solution, and meet the diverse needs of power customer service business. Description of the Drawings

[0065] Figure 1 It is a schematic diagram of the process of the power customer service training method of the present invention.

[0066] Figure 2 It is a schematic diagram of the application of the power customer service training system of the present invention. Detailed implementation manner

[0067] A power customer service training method based on the semantic understanding and summarization ability of a large model includes the following processes:

[0068] S100: Customer role simulation; generating a diverse simulation of the interaction process between real customers and customer service employees based on a large model; improving the authenticity and diversity of training; including the following processes:

[0069] S110: Construction of the first data set; specifically including the following steps:

[0070] S111: Collecting dialogue data for specific application business scenarios, and the data format is shown in Table 1;

[0071] S112: Processing the samples into the sample training format shown in Table 2;

[0072] S113: Coding and labeling the training samples according to the requirements of multi-round role-playing, so that the large model can learn the calling techniques of users; directly constructing labels including all the speech contents of users in multi-round conversations, which not only fully utilizes all the speech information of users, but also does not have duplicate calculations, which is very efficient;

[0073] The coding and labeling of the training samples are as follows:

[0074] inputs = <user><agent><user><agent><user><agent>

[0075] labels = <user><-100><user><-100><user><-100>

[0076] Under this kind of labeling, the large model only learns the speech information of users, quickly masters the habitual speech of users, and enables it to generate customer conversations that meet the requirements of power customer service business; <-100> means that this data information only participates in the forward inference of the model and does not participate in the update of the model training gradient;

[0077] The model training is SFT training, and the LORA method is selected;

[0078] Table 1 Sample of dialogue data for business scenarios

[0079]

[0080] Table 2 Sample of sample training format

[0081]

[0082] S120: Construct a set of prompts for multi-role playing and multi-business scenarios;

[0083] Prompt template: For example, you are a {age}{personality} {occupation}, {description of business scenario}, and you need to call the customer service specialist to ask about the situation. The starting inquiry statement is {scenario statement};

[0084] Sample prompt 1: For example, you are a 30-year-old IT programmer and your home has lost power. You need to call the customer service specialist to ask about the situation. The starting inquiry statement is "Hello, my home has lost power".

[0085] The age includes different age groups such as young, middle-aged, and old, such as 30 years old, 80 years old, etc.

[0086] The personality includes being cheerful, enthusiastic, introverted, nervous, etc.

[0087] The occupations include programmers, farmers, designers, teachers, etc.

[0088] The descriptions of business scenarios include power outage information query, electricity bill query, fault repair, etc.

[0089] The scenario statements include: My home has lost power, I want to query my electricity bill, My electricity meter has tripped, etc.

[0090] S130: Role-playing Q&A; Conduct role-playing Q&A based on the trained model; The specific process is as follows:

[0091] S131: Randomly select a prompt from the set of prompts for multi-role playing and multi-business scenarios, or based on business requirements, select any prompt for a specific business scenario;

[0092] S132: Send the prompt and the conversation context to the large model to make a service request;

[0093] S133: Add the agent's reply content to the conversation context.

[0094] S134: Repeat S132 and S133 until the large model replies with the final business handling or query result.

[0095] S200: Agent reply recommendation; With the help of the large model's multi-turn conversation generation ability, provide real-time reply recommendations and business handling process guidance for customer service staff; Help employees quickly master standard statements and business operations;

[0096] Based on the business process handling and multi-turn conversation generation technology for high-frequency power customer service scenarios, perform multi-turn interaction through preset slots, complete information filling in the slots, and request results, and then return the final reply statement;

[0097] Model Selection: Considering factors such as scenario response latency and computing power resources, large models with better capabilities such as Tongyi Qianwen 32B and 01.AI are selected. After SFT training, they can generate accurate and standardized reply recommendations and business processing procedures according to employees' inputs;

[0098] It includes the following processes:

[0099] S210: Construction of the second dataset; specifically including the following steps:

[0100] S211: Collect dialogue data for specific application business scenarios, and the data format is shown in Table 3;

[0101] S212: Perform processing on the samples and process them into the sample training format shown in Table 4;

[0102] S213: According to the requirements of the agent's reply, encode and label the training samples so that the large model can learn the agent's reply words; directly construct labels including all the agent's speaking content in multi-round conversations, which not only fully utilizes all the agent's speaking information but also does not involve duplicate calculations, which is very efficient;

[0103] The encoding and labeling of the training samples are as follows:

[0104] inputs = <User><Agent><User><Agent><User><Agent>;

[0105] labels = <-100><Agent><-100><Agent><-100><Agent>;

[0106] Under this kind of labeling, the large model only learns the agent's words information, quickly masters the agent's habitual words, and enables it to generate agent conversations that meet the business requirements of the power customer service. <-100> means that this data information only participates in the forward inference of the model and does not participate in the update of the model training gradient;

[0107] The model training is SFT training, and the LORA method is selected;

[0108] Table 3 Example of dialogue data for business scenarios

[0109]

[0110] Table 4 Example of sample training format

[0111]

[0112] S220: Collaborate with business experts to set the business scenario configuration information set, configure different slots and business request parameters for different scenarios, and associate with the core business system to complete the feedback of business results.

[0113] Configuration information for the electricity bill inquiry scenario: Electricity bill inquiry scenario_slot {household number, mobile phone number, month}_service request URL_return result.

[0114] The service request URL is like: https: / / www.bdu.com / ?Num=101&month=1

[0115] The return result is like: {"res": 98 yuan}

[0116] S230: Perform agent reply recommendation based on the trained model; the process includes the following:

[0117] S231: Based on the business scenario of the user's incoming call, that is, the training practice type, select the corresponding scenario configuration information from the business scenario configuration information set;

[0118] S232: Based on the multi-round conversation interaction information with the user, recommend the corresponding reply words and phrases, and update the slot values in the scenario configuration information in real time;

[0119] S233: Repeat S232 until the slot is filled and then reply "The business is being processed, please wait a moment";

[0120] S234: Complete the business system request, based on the return result, use the large model to rewrite and polish the words and phrases, and reply to the user.

[0121] S300: Reply quality evaluation; build a customer service reply quality evaluation system based on the large model, automatically score and analyze the employee's reply content, provide targeted improvement suggestions, and achieve quantitative evaluation and personalized improvement of the training effect;

[0122] Model selection: Considering factors such as scenario response latency and computing power resources, select inference large models such as deepseek, perform session summarization on the multi-round conversation text through prompt design, output the core content, and output the score and subsequent improvement suggestions according to the evaluation standard system;

[0123] The process includes the following:

[0124] S310: Build a quality evaluation standard system, and the evaluation dimensions include:

[0125] Degree of problem solving, solved or not solved;

[0126] Service perception, satisfied or dissatisfied by the user;

[0127] Service attitude, positive or negative;

[0128] Service professionalism, professional, average or unfamiliar;

[0129] Context Q&A matching degree. Which category does high, medium, and low belong to?

[0130] S320: Conduct reply quality evaluation based on the trained model, which specifically includes the following process:

[0131] S321: For the agent's real-time Q&A, conduct reply quality evaluation and return context Q&A matching degree and service professionalism information;

[0132] S322: For the single-pass conversation context after the conversation ends, use the large model to perform information element summary and conduct multi-dimensional rule scoring based on the element results;

[0133] S323: Statistically analyze the training evaluation data of employees and output subsequent training suggestions and improvement measures.

[0134] An electric power customer service training system based on the semantic understanding and summarization ability of a large model, including:

[0135] Customer role simulation module: Generate diverse customer roles and conversation scenarios based on the large model to simulate the interaction process between real customers and customer service employees;

[0136] Agent reply recommendation module: Utilize the multi-turn conversation generation ability of the large model to provide real-time reply recommendations and business handling process guidance for customer service employees;

[0137] Reply quality evaluation module: Build a customer service reply quality evaluation system based on the large model to automatically score and analyze the employee's reply content and provide targeted improvement suggestions.

[0138] Training management module: Responsible for functions such as the formulation of training plans, tracking of training progress, and statistical analysis of training effects, to achieve comprehensive management of the training process.

[0139] The present invention proposes a large model training sample construction scheme. In the electric power customer service scenario, based on multi-turn conversation data, perform training sample encoding and identification, only perform model training on the user's speech, and mask the agent's reply speech, so that the large model can generate user common conversations that meet the business requirements of electric power customer service, making the speech of the user role more anthropomorphic.

[0140] The present invention proposes a large model prompt word design method. Add business scenario information such as complaint, report, business application and other scenarios and the first sentence of the call in the prompt word, and user role information such as age, occupation, personality, etc., to ensure that the model interaction is personalized and diverse, making the training practical exercise more real and the training more efficient.

[0141] The present invention proposes a large-model training sample construction scheme. In the electric power customer service scenario, training sample encoding and identification are performed based on multi-round conversation data. Model training is only performed on the agent's reply words, and the user's speech content is masked, so that the large model can generate agent reply words that meet the business needs of electric power customer service, and the large model reply has knowledge of the electric power customer service business scenario, which is more in line with business training needs.

[0142] The present invention proposes a multi-round dialogue technology based on a big model and power customer service business scenario configuration information, which realizes real-time recommendation of agent speech through big model question and answer and information extraction and slot filling based on the big model.

[0143] The present invention constructs a customer service response quality evaluation standard system, including different dimensions such as problem solving degree, service perception, service attitude, service professionalism, contextual question and answer matching, etc. It can support real-time scoring and statistical scoring.

[0144] The present invention uses large model technology to perform real-time agent response quality scoring, single-pass conversation information summary and rule statistical scoring in power customer service training scenarios, statistically analyze multi-pass conversation scoring results and output employee service improvement suggestions.

Claims

1. A power customer service training method based on the semantic understanding and summarization ability of large models, characterized in that It includes the following processes: S100: Customer role simulation; generating diverse simulated interactions between real customers and customer service staff based on a large model; improving the authenticity and diversity of training; S200: Agent reply recommendation; With the multi-round dialogue generation ability of the large model, providing real-time reply recommendations and business process guidance for customer service staff; helping employees quickly master standard scripts and business operations; S300: Reply quality assessment; constructing a customer service reply quality assessment system based on the large model, automatically scoring and analyzing the reply content of employees, providing targeted improvement suggestions, and realizing quantitative assessment and personalized improvement of training effects.

2. The power customer service training method based on the semantic understanding and summarization ability of the large model according to claim 1, wherein, The S100 includes the following processes: S110: First dataset construction; specifically including the following steps: S111: Collecting dialogue data for specific application business scenarios; S112: Processing the samples; S113: Encoding and labeling the training samples according to the requirements of multi-round role-playing, enabling the large model to learn the call techniques of users; directly constructing labels including all the speaking content of users in multi-round conversations, which not only fully utilizes all the speaking information of users but also does not involve duplicate calculations, being very efficient; The encoding and labeling of training samples are as follows: inputs = <user><agent><user><agent><user><agent>; labels = <user><-100><user><-100><user><-100>; Under this kind of labeling, the large model only learns the speaking information of users, quickly masters the habitual speaking patterns of users, enabling it to generate customer conversations that meet the requirements of power customer service business; <-100> means that this data information only participates in the forward inference of the model and does not participate in the update of the model training gradient; The model training is SFT training, and the LORA method is selected; S120: Constructing a set of prompt words for multi-role-playing and multi-business scenarios; Prompt word template: For example, you are a {age}{personality} {occupation}, {description of business scenario}, and you need to call the customer service specialist to ask about the situation. The starting inquiry script is {scenario script}; S130: Role simulation Q&A; conducting role simulation Q&A based on the trained model; specifically including the following processes: S131: Randomly selecting a prompt word from the set of prompt words for multi-role-playing and multi-business scenarios, or based on business requirements, selecting any prompt word for a specific business scenario; S132: Sending the prompt word and the dialogue context to the large model for a service request; S133: Adding the agent's reply content to the dialogue context; S134: Repeating S132 and S133 until the large model replies with the final business handling or query result.

3. The power customer service training method based on the semantic understanding and summarization ability of the large model according to claim 1, wherein The S200 includes the following processes: S210: Second dataset construction; specifically including the following steps: S211: Collecting dialogue data for specific application business scenarios; S212: Processing the samples; S213: Encoding and labeling the training samples according to the agent's reply requirements, enabling the large model to learn the agent's reply scripts; The encoding and labeling of training samples are as follows: inputs = <user><agent><user><agent><user><agent>; labels=<-100><agent><-100><agent><-100><agent>; The model training is SFT training, and the LORA method is selected; S220: Cooperate with business experts to set a business scenario configuration information set, configure different slots and business request parameters for different scenarios, and associate with the core business system to complete the feedback of business results; S230: Recommending agent responses based on the trained model; including the following process: S231: based on the business scenario of the incoming call from the user, that is, the type of exercise for training, selecting corresponding scenario configuration information from the business scenario configuration information set; S232: Based on the multi-round dialogue interaction information with the user, recommend corresponding reply words, and update the slot value in the scenario configuration information in real time; S233: Repeat S232 until the slot is completely filled and then reply "Business is being processed, please wait".

4. The power customer service training method based on the semantic understanding and summarization ability of the large model according to claim 3, characterized in that, The S233 is followed by the following process: S234: Complete the business system request, rewrite and polish the speech based on the returned results using the big model, and reply to the user.

5. The power customer service training method based on the semantic understanding and summarization ability of the large model according to claim 1, characterized in that, The S300 includes the following process: S310: Construct a quality assessment standard system, with the following assessment dimensions: The degree of problem resolution, whether it has been resolved or not; Service perception, whether the user is satisfied or dissatisfied; Service attitude, positive or negative; The level of professionalism of the service: professional, average, or unfamiliar; The matching degree of contextual question and answer, high, medium, or low; S320: Perform response quality assessment based on the trained model, which specifically includes the following process: S321: For the real-time questions and answers of the agents, the quality of the responses is evaluated, and the matching degree of the contextual questions and answers and the service professionalism information is returned; S322: For the single-pass conversation context after the conversation ends, use the big model to perform information element summary, and perform multi-dimensional rule scoring based on the element results; S323: Collect employees’ training evaluation data and output follow-up training suggestions and improvement measures.

6. A power customer service training system based on the semantic understanding and summarization ability of large models, characterized in that, include: Customer role simulation module: Generates a variety of customer roles and dialogue scenarios based on the big model, simulating the interaction process between real customers and customer service staff; Agent response recommendation module: This module uses the multi-round dialogue generation capability of the large model to provide customer service staff with real-time response recommendations and business process guidance. Response quality assessment module: Build a customer service response quality assessment system based on a large model, automatically score and analyze employees' responses, and provide targeted improvement suggestions.

7. The power customer service training system based on the semantic understanding and summarization ability of the large model according to claim 6, characterized in that, Also includes: Training management module: responsible for the formulation of training plans, tracking of training progress, statistical analysis of training results, etc., to achieve comprehensive management of the training process.

Images