Heat supply charging customer service system based on large language model

Abstract

The invention discloses a heat supply charging customer service system based on a large language model, and relates to the technical field of instant messaging, and the system comprises a user interaction module, a large language model service layer, a cue word optimization module, a heat supply knowledge graph module, an emotion semantic recognition module and a service platform module. A large language model is introduced to be fused with a heat supply professional knowledge graph, so that heat supply charging business query, complaint processing and process execution under natural language multi-round interaction are realized; a cue word adaptive optimization algorithm is adopted, a language model probability guiding genetic algorithm is used for optimizing a cue template, and the ability of understanding complex semantics is enhanced; and the service platform gets through the charging system, the bank interface and the main data to realize real-time query and automatic service processing in the dialogue. The method has the remarkable advantages of intelligent interaction, data fusion, automatic process, efficient response and the like, the user experience is improved, and the operation and maintenance cost is reduced.

Description

Technical Field

[0001] The present invention relates to the field of instant messaging technology, and in particular to a heating fee customer service system based on a large language model. Background Art

[0002] With the advancement of smart cities and public service informatization, heating companies are facing growing customer service demands. As a core business, heating billing raises higher standards for customer service, including information accuracy, response efficiency, and user experience. Currently, most heating customer service systems rely on manual hotlines or pre-configured question-and-answer bots, which are unable to effectively address the increasingly diverse and complex user inquiry scenarios. Their low level of intelligence leads to low user satisfaction.

[0003] The existing heating customer service system suffers from the following major issues: Multiple, complex conversations fail to accurately identify user intent, often resulting in irrelevant answers. The customer service system is poorly integrated with billing, banking, and other business systems, lacking automated access capabilities, forcing users to undergo multiple verifications or be transferred to a human operator. Furthermore, the system lacks understanding and adaptation of heating industry terminology and regulations, resulting in unprofessional responses and potential misunderstandings. These issues severely hinder the improvement of heating companies' service levels, increase customer service costs, and increase the risk of user churn.

[0004] In response to the above problems, the present invention proposes a heating fee customer service system based on a large language model. By constructing modules such as emotional semantic recognition, prompt word optimization, heating knowledge graph, and business middle platform, it realizes customer service automation, specialization, and intelligence, and automatically calls the background system to complete data query and operation, effectively solving the problems of low intelligence and poor business integration of existing customer service systems. Summary of the Invention

[0005] In response to the above problems, the present invention provides a heating fee customer service system based on a large language model to solve the problem that the heating customer service system in the existing technology is not intelligent enough in terms of multi-round semantic understanding, business system integration and professional rule adaptation.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: a heating fee customer service system based on a large language model, comprising: A user interaction module is used to receive heating charge questions raised by users in natural language and display the response content generated by the system; The emotional semantics recognition module is used to identify the emotional state and business intent in user sentences. Business intent includes arrears appeals, heating complaints, and bill inquiries. The heating knowledge graph module is used to provide heating industry business rules and process nodes related to the identified user intent and convert them into structured knowledge that can be used as a reference for the language model; The prompt word optimization module performs adaptive optimization on the prompt words based on a genetic algorithm guided by the probability of the language model. The specific steps of the genetic algorithm are as follows: Construct a population of prompt word templates, each of which consists of multiple text fragments; Define a fitness function, and calculate the probability score of the answer generated by the language model and the weighted matching degree of the business semantics; Perform crossover and mutation operations on the high-fitness template to form the next generation template; After multiple iterations, the optimal template is selected to guide the large language model to generate answers; The large language model service module is used to generate natural language answers based on the optimized prompt words and user input. The large language model service module integrates knowledge graph information and supports multi-round dialogue context tracking; the business middle platform module is used to automatically call the enterprise's internal charging system interface, bank interface and master data management system according to user intention.

[0007] The user interaction module adopts the form of web pages, mobile apps and self-service terminals to provide users with a natural language input and output interface. When users enter questions and requests, the information is sent to the back-end processing flow. After receiving the reply generated by the system, the user interaction module is responsible for presenting the reply to the user.

[0008] After receiving user input from the user interaction module, the emotional semantic recognition module first performs preprocessing analysis and uses a pre-trained classification model and rule engine to extract the user's emotional state and semantic intention from the user's sentence; emotional state discrimination is based on the sentiment analysis model, dividing the user's sentence into positive, neutral and negative, and semantic intention discrimination is to identify which category of heating business the user topic belongs to.

[0009] The heating knowledge graph module is constructed in the form of a graph database, and the nodes include: charging standards, payment channels, heating time, business processes and exception handling rules; the edges represent rule dependencies and conditional transfer relationships.

[0010] The fitness function is specifically shown in the formula:

[0011] in, is the fitness function, T is the prompt template individual to be evaluated, Ask questions to users in the i-th test scenario, For the large language model, under the guidance of the prompt template T, The generated answer, Generate an answer for the language model given this prompt and question The probability score of is a business assessment score for the accuracy of the answer, , The business middle platform module includes: charging system interface, bank interface and master data management module; The charging system interface is used to obtain the user's chargeable fees, payment status, and bill details; The bank interface is used to verify user payment records, payment time, and transaction status; The master data management module is used to read user profile information, including user address, contract parameters, and heating area.

[0012] Compared with the prior art, the present invention has the following beneficial effects: The heating billing customer service system based on a large language model provided by this invention combines the powerful natural language processing capabilities of the large language model with professional knowledge and business system data in the heating field. It is optimized for the special needs of heating billing customer service scenarios and has the following unique advantages and beneficial effects: The present invention is specially customized for common consultation and complaint scenarios in the heating field. Through the emotional semantic recognition module and the heating knowledge graph, the system can identify the professional terms and specific expressions in the user's speech and understand the implicit specific demands. No matter what expression method the user uses, the system can determine their needs based on the context.

[0013] This invention uses the heating knowledge graph to embed verified business rules and professional knowledge in the answers, integrating the knowledge graph into the dialogue generation method, avoiding the random answers or biased suggestions that may appear in a simple large language model, and ensuring that customer service responses comply with corporate and industry standards under any circumstances.

[0014] The present invention introduces a prompt word optimization mechanism based on a genetic algorithm, automatically adjusts the dialogue prompts for the heating customer service scenario, and continuously evolves the optimal prompt template through a large number of simulated dialogues in the offline stage, so as to maximize the model's answer accuracy and user satisfaction. In the online stage, the wording is adjusted in real time according to the user's emotions and context, and the answer strategy can be dynamically adjusted according to the actual situation, reducing the situation of irrelevant answers or stiff tones. Especially in complex inquiries or abnormal situations, the optimized prompts can guide the model to output responses that are more in line with user expectations.

[0015] The present invention connects the heating fee system, bank payment system and user master data through the business middle-end module, and can complete the closed loop from user questions to data queries to answers in a single round of dialogue, so that users do not need to be directed to other channels to verify information, thereby improving the one-stop service experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following is a brief introduction to the drawings required for use in the embodiments. It is understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0017] Figure 1 It is a system architecture diagram of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, 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 part of the embodiments of the present invention, not all of 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. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention claimed for protection, but is merely for selected embodiments of the present invention.

[0019] Please refer to Figure 1 , Figure 1 This is a schematic diagram of a heating fee customer service system based on a large language model provided by an embodiment of the present invention, including: The user interaction module uses web pages, mobile apps, and self-service terminals to provide users with a natural language input and output interface. When users enter questions or requests, the information is sent to the back-end processing flow. After receiving the system-generated reply, the user interaction module is responsible for presenting the reply to the user. The user interaction module also includes a speech synthesis and recognition unit. If it is in a voice customer service scenario, speech and text conversion will be performed.

[0020] After receiving user input from the user interaction module, the emotional semantic recognition module first performs preprocessing analysis. Using a pre-trained classification model and rule engine, it extracts the user's emotional state and semantic intent from the user's sentence. Emotional state discrimination is based on the sentiment analysis model, classifying user sentences into positive, neutral, and negative categories. Semantic intent discrimination identifies which category the user's topic belongs to in the heating business. For example: When a user's statement includes "My heating was turned off because I didn't pay, but I clearly paid," it is identified as an overdue payment appeal scenario, indicating that the user is disputing the billing status. When a user asks, "The temperature at home never meets my expectations. What's going on?", this is identified as a heating quality complaint or substandard heating quality complaint, with the user's intention being to report issues with the heating quality. If the user asks "How much is my heating bill this year?", it is identified as a bill inquiry scenario, and the intention is to inquire about the bill amount.

[0021] The emotion semantic recognition module outputs a set of results, such as {emotion: negative (angry), intent: appeal for overdue bills}, {emotion: negative (complaint), intent: heating complaint}, or {emotion: neutral, intent: inquire about bills}. This result is then sent to the prompt word optimization module and the business middleware module. Emotion and intent information are crucial for subsequent steps: It helps the prompt word optimization module select appropriate dialogue strategies, and the business middleware module determines whether to call corresponding backend services based on the intent. For example, if the intent is to inquire about bills or overdue bills, the billing system interface or bank interface needs to be called to verify the data.

[0022] Heating knowledge graph module. After the emotional semantic recognition module determines the user's intention, the system will search for knowledge items or business processes related to the intention in the heating knowledge graph module. The knowledge graph stores a large number of knowledge points in the heating field and the relationships between them in the form of a graph database, including: charging policy regulations, charging standards, payment channels, frequently asked questions and answers, fault reporting procedures, complaint handling specifications, and professional term definitions.

[0023] For example, for the intent of "appealing for outstanding fees," the knowledge graph contains relevant process nodes, such as "querying user payment records," "verifying bank payment information," and "initiating an outstanding fee objection process," as well as the company's regulations for handling outstanding fee objections. For complaints about "heating not meeting standards," the knowledge graph provides corresponding rule nodes, including business processes such as "asking the user about the room temperature," "checking whether the room temperature is below standard," and "if so, generating a repair report and contacting maintenance personnel." After receiving the user's intent, the heating knowledge graph module outputs a set of knowledge fragments or rule processes that are highly relevant to the intent, which serve as the basis for subsequent answer generation. This knowledge graph-based retrieval ensures that the system's answers are professional and correct, and will not deviate from company regulations due to possible hallucinations of the large language model. The content provided by the knowledge graph usually exists in the form of structured data, which the system converts into prompt information fragments suitable for the large language model to understand. For example, relevant regulations or steps are expressed in natural language so that they can be embedded in the prompt template.

[0024] The prompt word optimization module dynamically adjusts and optimizes prompt templates based on the characteristics of the heating customer service scenario, guiding the large language model to produce more accurate responses. The prompt template can be understood as a piece of guidance content provided to the large language model, which contains instructions for the model, contextual information, and reference knowledge. This module maintains a set of prompt template libraries and optimization strategies, and will select corresponding template bases for filling and optimization based on different user intentions and situations.

[0025] For example, in the scenario of an overdue bill complaint, the template may include a soothing opening, such as "We apologize for the inconvenience, we will help you verify your payment status"; followed by a placeholder for the query results and an explanation based on the policy; in the scenario of substandard heating, the template may include an apology and understanding of the user's feelings, and then guide the user to provide room temperature information or inform that maintenance has been contacted for them.

[0026] The prompt word optimization module not only selects an appropriate initial template but also makes adaptive adjustments based on the conversation context and model feedback. This adaptation is mainly reflected in two stages: pre-conversation optimization and in-conversation optimization. Pre-conversation optimization utilizes the prompt word optimization algorithm based on language model probability guidance proposed in this paper, and iteratively improves the prompt template library offline through a genetic algorithm. In-conversation optimization makes subtle adjustments to the prompt based on the real-time conversation situation. For example, when it detects that the user is strongly dissatisfied, it adds more gentle interjections or more explanatory details to the prompt to improve user satisfaction.

[0027] The prompt word optimization algorithm is as follows: First, the prompt template is represented by an appropriate genetic code. Each prompt template is considered an individual, and its "chromosome" can be composed of several text segments or parameters. For example, a prompt template chromosome can be divided into several genetic segments, [tone segment, knowledge insertion segment, question-and-answer format segment], etc. Each segment corresponds to a type of function. For example, the tone segment determines the politeness of the answer, the knowledge insertion segment determines the type of knowledge representation to be embedded, and the question-and-answer format segment determines whether the answer is listed in paragraphs. The initial population consists of several prompt template individuals with different styles designed based on human experience. Then, for each individual, its fitness is evaluated by having the large language model generate and score answers in a set of representative heating customer service dialogue scenarios. The design of the fitness function comprehensively considers the probabilistic feedback of the language model and the business correctness indicator. The fitness function is defined as shown in the formula:

[0028] in, is the fitness function, T is the prompt template individual to be evaluated, Ask questions to users in the i-th test scenario, For the large language model, under the guidance of the prompt template T, The generated answer, Generate an answer for the language model given this prompt and question The probability score or confidence level of the key correct content reflects the probability guidance factor of the language model; It is a business evaluation score for the accuracy of the answer, such as the rule matching degree based on the knowledge graph or the semantic similarity of the expected answer; weight , Used to balance the relative importance of the language model's internal preferences and business correctness; through the above fitness function, quantitatively evaluate the quality of a prompt template when comprehensively considering model fluency and business correctness.

[0029] After evaluating the fitness, the algorithm selects several prompt templates with high fitness from the current population as "parents", and then performs crossover and mutation operations on the selected individuals to generate new candidate prompt templates, namely "offspring". The crossover strategy can reorganize the chromosomes of the two parent templates at certain segment boundaries, such as exchanging tone segments or knowledge insertion segments, thereby combining different prompt styles.

[0030] It should be noted that due to the semantic connections between different segments of the prompt text, to maintain the coherence of the new template, the present invention preferably uses a large language model to assist in the crossover process: the model generates a new prompt that combines the advantages of the two parent templates based on their respective characteristics. This is equivalent to intelligent crossover, ensuring the readability and effectiveness of the new template. The mutation strategy randomly introduces small changes to individual chromosomes, such as replacing certain words, adjusting sentence structures, adding or removing polite language, or using the large language model to perform synonymous rewriting of the template. The purpose of mutation is to explore new ways of writing prompts to increase population diversity. The language model can also be used to check the grammatical and semantic rationality of the mutation results to avoid generating prompts that are inconsistent with the business.

[0031] After a new generation of templates is generated through crossover and mutation, they undergo another iteration of fitness evaluation and selection. After multiple generations of evolution, the algorithm converges on a number of high-quality prompt templates. These templates significantly improve the accuracy and user satisfaction of the large language model in the heating customer service scenario. Ultimately, these optimized, high-quality prompt templates are deployed in the prompt word optimization module for use in online conversations. This mechanism enables the system to continuously fine-tune conversation guidance strategies based on the specific corpus and needs of heating billing customer service. This significantly differs from the fixed scripts of traditional rule-based dialogue systems, improving the adaptability and intelligence of the dialogue system.

[0032] The large language model service module receives the final prompt from the prompt word optimization module, as well as the user's question and related context, and uses the pre-loaded large language model to generate answers. During the generation process, the large language model refers to various information in the prompt, including user questions, user emotions and intention descriptions, knowledge fragments extracted from the knowledge graph, and real-time data obtained through the business middle platform. Since the prompt template has been optimized, the model can more accurately understand the user's question and accurately select knowledge points and data to answer. The output is a natural language text answer. In addition, to ensure that the answer is accurate and professional, the model follows the instructions in the prompt when generating, such as "give priority to the statements provided by the knowledge graph" or "directly quote the query results when giving the bill amount instead of making it up."

[0033] When multiple rounds of interaction are required, the large language model will also maintain the context based on the conversation history, so that questions and answers can be connected. For example, the user first asks "How much was my heating bill last year?" After the system answers, the user then asks "When did I pay it?" The model can continue to answer based on the results of the previous round. Through the above mechanism, the large language model service module realizes automatic response to heating customer service questions. In most cases, the generated content can be directly provided to users without human intervention.

[0034] During the conversation generation process, if the large language model service module detects the need to call real business data or perform business operations, this will be completed through the business middle platform module. The business middle platform accepts the request from the large language model service module, which includes parameters such as the user ID and the required business operation type, and connects to multiple internal information systems of the enterprise. In addition, the business middle platform module also includes the following functions: The charging system interface is called. When a user's question involves account fees, such as inquiring about the amount of outstanding fees, historical payment records, etc., the business middle platform calls the heating charging system to obtain the user's latest bill data or payment records; for example, if the user asks "How much did I pay this year", the system will request the charging system to return the user's payable and paid fees for this heating season.

[0035] When a user asks a question such as "When did I pay?" that requires verification of payment time or transaction details, the business middle platform queries the corresponding payment flow or bank receipt through the bank interface. The system retrieves the user's most recent heating fee payment transaction record based on the user ID, including payment date, amount, channel, etc., and provides it to the large language model to generate an accurate answer; when a user claims that he or she has paid but the system shows that the fee is in arrears, the business middle platform can immediately verify the transaction status with the third-party payment platform, determine whether there is a delay or failure, and feedback the conclusion to the customer service dialogue.

[0036] In the master data management module, when the conversation involves verifying the user's basic information or needs to execute branch logic based on user attributes, such as determining the start time of heating based on the user's city or community, the business middle platform obtains the corresponding information from the master data management module. The master data management centrally maintains user files, including name, address, account number, heating station area, contract plan, etc. This information can be used to enrich the context of the large language model and make the answer more personalized and accurate. For example, if a user asks "When will the heating start in my home?", the system can query the heating start and end dates in the area where the user's address belongs through the master data, and then the large language model will answer accordingly.

[0037] In summary, in the actual workflow, when a user initiates an inquiry or complaint, the user interaction module obtains their input, and the emotional semantic recognition module first analyzes the input to determine the user's intention and emotional state; then the prompt word optimization module generates optimized prompt information based on the recognition results and related content of the knowledge graph, and packages the user's intention, context and knowledge base information to form the input for the large language model; the large language model service layer generates an answer based on this; if it is necessary to query the user's payment record or perform business operations, the charging system or bank interface is accessed through the business middle-end module to obtain the required data for reference by the large language model and integrated into the answer; finally, the user interaction module feeds back the answer returned by the large language model to the user.

[0038] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A heating fee customer service system based on a large language model, characterized in that: include: A user interaction module is used to receive heating charge questions raised by users in natural language and display the response content generated by the system; The emotional semantics recognition module is used to identify the emotional state and business intent in user sentences. Business intent includes arrears appeals, heating complaints, and bill inquiries. The heating knowledge graph module is used to provide heating industry business rules and process nodes related to the identified user intent and convert them into structured knowledge that can be used as a reference for the language model; The prompt word optimization module performs adaptive optimization on the prompt words based on a genetic algorithm guided by the probability of the language model. The specific steps of the genetic algorithm are as follows: Construct a population of prompt word templates, each of which consists of multiple text fragments; Define a fitness function, and calculate the probability score of the answer generated by the language model and the weighted matching degree of the business semantics; Perform crossover and mutation operations on the high-fitness template to form the next generation template; After multiple iterations, the optimal template is selected to guide the large language model to generate answers; A large language model service module, which is used to generate natural language responses based on optimized prompt words and user input. The large language model service module integrates knowledge graph information and supports multi-round conversation context tracking. The business middle platform module is used to automatically call the enterprise's internal charging system interface, bank interface and master data management system according to user intentions.

2. A heating fee customer service system based on a large language model according to claim 1, characterized in that: The user interaction module adopts the form of web pages, mobile apps and self-service terminals to provide users with a natural language input and output interface. When users enter questions and requests, the information is sent to the back-end processing flow. After receiving the reply generated by the system, the user interaction module is responsible for presenting the reply to the user.

3. The heating fee customer service system based on a large language model according to claim 1 is characterized by: After receiving the user input from the user interaction module, the emotion semantic recognition module first performs pre-processing analysis and uses the pre-trained classification model and rule engine to extract the user's emotional state and semantic intention from the user's sentence; Emotional state discrimination is based on the sentiment analysis model, which divides user sentences into positive, neutral and negative. Semantic intent discrimination is to identify which category of heating business the user topic belongs to.

4. The heating fee customer service system based on a large language model according to claim 1 is characterized by: The heating knowledge graph module is constructed in the form of a graph database, and the nodes include: charging standards, payment channels, heating time, business processes and exception handling rules; the edges represent rule dependencies and conditional transfer relationships.

5. The heating fee customer service system based on a large language model according to claim 1 is characterized by: The fitness function is specifically shown in the formula: in, is the fitness function, T is the prompt template individual to be evaluated, Ask questions to users in the i-th test scenario, For the large language model, under the guidance of the prompt template T, The generated answer, Generate an answer for the language model given this prompt and question The probability score of is a business assessment score for the accuracy of the answer, , The weight is adjustable.

6. The heating fee customer service system based on a large language model according to claim 1 is characterized by: The business middle platform module includes: charging system interface, bank interface and master data management module; The charging system interface is used to obtain the user's chargeable fees, payment status, and bill details; The bank interface is used to verify user payment records, payment time, and transaction status; The master data management module is used to read user profile information, including user address, contract parameters, and heating area.

Images