Intelligent customer service problem decomposition method based on large language model, medium and electronic equipment

Abstract

The invention provides an intelligent customer service problem decomposition method based on a large language model, a medium and electronic equipment. According to the method, error correction and compliance cleaning can be carried out on the input information of the user, and the user intention is accurately recognized in combination with the historical dialogue context. When it is detected that a plurality of intentions are included in user input, the intentions are decomposed into independent sub-intentions (including question description, categories and key parameters) by using a preset dynamic rule base, and corresponding skill knowledge base tool instructions are matched for different sub-intentions; and when key information is lacked, automatic asking and complementing are carried out. And then, inputting the skill indication information corresponding to each sub-intention and the user associated data into the large language model to generate a reply for each sub-intention. And for a plurality of reply results, a content generation rule is applied to combine and embellish the reply results, and the results are uniformly fed back to the user. Through the above mode, according to the scheme of the invention, multiple rounds of contexts and vertical industry knowledge bases in the same session can be fused, and accurate disassembly and response to complex multi-intention user requests can be realized; the technical problems that in the prior art, intelligent customer service is lack of knowledge in specific fields such as hotels, catering, fitness, theaters and scenic spots, and user intentions and continuous dialogues cannot be refined are solved, reply accuracy and service experience are remarkably improved, and man-machine dialogues are more efficient and natural.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an intelligent customer service problem decomposition method, medium, and electronic device based on a large language model. Background Technology

[0002] With the rapid development of artificial intelligence and natural language processing technologies, large language models (LLMs) are increasingly being used in the field of intelligent customer service. Current intelligent customer service systems can handle some general questions using large language models, such as simple knowledge-based Q&A and routine business inquiries. However, in specific vertical sectors such as hotels, restaurants, fitness centers, theaters, and tourist attractions, the performance of intelligent customer service still shows significant shortcomings. The main reason is the lack of dynamic data and rule-based knowledge relevant to these industries. For example, in the hotel industry, this includes information on check-in dates, room numbers, check-out policies, and facility service information; in the restaurant industry, it includes menu details and booking processes; in the fitness industry, it includes venue information and class schedules; in the theater industry, it includes film and show information and ticketing policies; and in the tourist attraction industry, it includes tour guide information and special services. Without sufficient industry knowledge support, intelligent customer service often fails to accurately understand and handle personalized questions raised by users in these areas, resulting in poor service quality and user experience.

[0003] For example, when a hotel guest asks, "Can I postpone the check-out time for room 1234 until 3 PM?", traditional models, unable to link the room number with the hotel's check-out rules, may provide a vague or incorrect answer. Similarly, a user might submit multiple different requests in a single input (i.e., mixed-intent input), such as "Please deliver a toothbrush to my room and book my yoga class for 8 AM tomorrow." For such complex requests, existing intelligent customer service systems, lacking effective intent decomposition capabilities, often misinterpret them as single intents, leading to omissions or misunderstandings. The pain points of existing technologies mainly focus on: 1) insufficient granularity of intent recognition, failing to correctly decompose multiple intents in complex problems; 2) lack of vertical domain knowledge, lacking industry rules and data support that are dynamically updated with business needs; and 3) poor dialogue context coherence, making it difficult for the system to combine historical dialogue content to process continuous user requests. These shortcomings prevent intelligent customer service from meeting users' personalized and diverse service needs in many scenarios.

[0004] Therefore, there is an urgent need for an improved technical solution that can enhance the ability of intelligent customer service to identify user intent in a fine-grained manner and decompose problems for vertical industry scenarios, make full use of industry knowledge bases to accurately handle user problems, and ensure the contextual coherence of multi-turn dialogues, thereby improving service quality and user experience. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent customer service question decomposition method based on a large language model, so as to solve the technical problems of inaccurate intent recognition and question processing in existing intelligent customer service systems in vertical industry applications.

[0006] To achieve the above objectives, this invention provides a method for multi-intent decomposition in intelligent customer service, applied to an intelligent customer service system with a large language model. The method includes: The system acquires the user's current input information and preprocesses it, including spelling error correction and non-compliant content cleaning, to obtain standardized user input. The standardized current input is combined with the user's historical context dialogue information, and the user's intent is identified using a large language model. When the identified user intent contains multiple different requests, it is parsed according to preset question splitting rules, decomposing into multiple independent sub-intents. Each sub-intent contains a complete intent description, the corresponding intent category, and key parameters required to process the intent. For each sub-intent, skill instruction information corresponding to the sub-intent is matched and determined from a pre-built skill knowledge base. This skill instruction information, along with the sub-intent and related user association information, is input into the large language model to generate corresponding response content. When the processing of a sub-intent requires additional information and the user input does not provide this information, the large language model generates a prompt to ask the user to supplement the key parameters. If multiple sub-intents and corresponding response content exist, the multiple response content is input into the large language model for merging and language polishing to generate a natural and fluent final response. Finally, the final response content is fed back to the user.

[0007] Furthermore, after obtaining user input information, the preprocessing process includes: correcting errors in the input text using a preset spell-checking algorithm; filtering and cleaning the input content using a predefined sensitive word library and rules to block non-compliant content such as advertisements and abusive language, as well as user privacy information, and removing special symbols or format errors to ensure data standardization and consistency in subsequent processing. When user input contains superficial polite phrases (such as "thank you" or "okay") and the input does not constitute a substantive request, its intent can be determined based on the dialogue context: if the polite phrase is a confirmation response to the previous intelligent customer service question, it is identified as the user's confirmation intent; otherwise, the input can be ignored to simplify the dialogue. In addition, for user input consisting only of numbers, the preprocessing parses the input based on the number length: when the number of digits is less than a preset threshold (such as 5 digits), it is preferentially identified as a room number or the last few digits of a room number; when the number of digits is greater than the threshold, it is identified as contact information such as a phone number. This rule can improve the accuracy of understanding short inputs and ensure that user requests can be correctly processed in conjunction with the context.

[0008] Furthermore, the user's current input information can include one or more forms such as text, voice, images, and videos. In the case of non-text input, the preprocessing also includes converting voice information into text (e.g., through speech recognition technology) and extracting information from images or videos and converting it into text descriptions (e.g., through image recognition and OCR technology) to unify subsequent processing. After format normalization, the current input text and the user's historical dialogue context information in this session are input into the large language model for semantic analysis. By introducing historical context, the intelligent customer service system can understand the relationship between the current request and previous communication content, thereby more accurately identifying the user's intent. When the current input is closely related to the historical dialogue, the large language model will infer the user's true intent based on the context; if the current input is a completely new topic unrelated to the previous context, the model mainly relies on the current input itself to identify the intent. For example, if the intelligent customer service asked for a certain information in a previous conversation and the user replied "okay" now, the system can recognize it as the user confirming a previous question; or if the intelligent customer service asked the user to provide a mobile phone verification code in the conversation, and the user then entered a string of numbers, the system can determine from the context that the numbers are a verification code or part of a mobile phone number, rather than a new service request.

[0009] Once the large language model determines the user's intent, it enters the question decomposition stage. Based on the pre-defined question decomposition instructions (dynamic rule base) of this invention, the intelligent customer service system can break down complex user requests into one or more independent sub-questions and extract the key elements of each sub-question. The question decomposition rule base is derived from summarizing and refining numerous dialogue scenarios across various vertical industries, and is used to guide the large language model in chain-like intent analysis, including but not limited to the following rules: When user input contains multiple requests of different topics or types, it is split into different sub-intents and processed separately to ensure that each sub-intent contains complete subject-verb-object information, making the split expression clear and complete.

[0010] When a user requests multiple items related to room service, the request for each item will be broken down into separate sub-intents and processed one by one (for example, if a user requests both a toothbrush and a towel, it will be divided into two sub-intents).

[0011] When the user's current input is a brief response or supplement to a previous question (such as simply answering "yes" or "okay"), it is merged with the historical context to form a complete intent, without splitting it into a new sub-intent.

[0012] • If multiple statements currently entered by the user actually revolve around the same topic or belong to the same intent category (e.g., several consecutive sentences describing the same problem or complaint), they can be merged into a single sub-intent without repeated splitting, in order to avoid lengthy responses and duplicate information.

[0013] When a user's input contains contradictory requests (e.g., "I want to cancel my reservation but also want to keep the room"), the user's last expressed intention will be used for processing based on the business logic.

[0014] For inputs that cannot be recognized as having a valid meaning (such as only emojis or small talk without actual content), sub-intents can be ignored to improve dialogue efficiency.

[0015] Through the above rules, the intelligent customer service system can intelligently judge inputs of varying complexity, flexibly splitting or merging user intents to ensure that each sub-intent accurately reflects the user's true needs. This design avoids the errors and omissions that occur in the traditional "question-and-answer" mode when faced with multiple intent inputs, and prevents irrelevant answers due to mismatched historical context.

[0016] In this embodiment of the invention, a sub-intent consists of two parts: an intent category and an intent content. The intent category indicates the business scenario and type to which the user's request belongs, while the intent content is a description and parameter details of the specific request. For example, in a hotel scenario, the intent category could be "Accommodation Service - Item Delivery," while the intent content might be "Need a toothbrush"; in a fitness scenario, the intent category might be "Fitness Service - Reservation," with corresponding intent content such as "Reserve a yoga class for 8 AM tomorrow"; in a theater scenario, the intent category is "Theater Service - Ticket Purchase," and the intent content is "Purchase two tickets for tonight's 7 PM performance." By representing sub-intents in this structured way, the system can clearly define the processing type and required information for each request, thereby providing a useful basis for subsequent steps.

[0017] The intent categories can cover common service types across multiple vertical industries. In a preferred embodiment, intent categories include, but are not limited to: "inquiry," "item delivery," "reservation," and "customer feedback" in the accommodation service sector; "inquiry," "reservation," and "customer feedback" in the catering service sector; "inquiry," "reservation," "complaints and feedback," and "other" in the fitness service sector; "inquiry," "ticket purchase," "feedback," and "complaint" in the theater service sector; and "information inquiry," "ticketing-related," "travel feedback," and "special needs" in the scenic area service sector. Of course, these categories can be expanded or adjusted according to actual applications to adapt to more industry scenarios or more specific service types.

[0018] After dividing the sub-intents, the intelligent customer service system needs to select the appropriate processing method for each sub-intent. To this end, this invention pre-constructs a skill knowledge base (also known as a tool library), which includes business logic and response templates for handling different types of user intents in various industry scenarios. Each skill instruction can be understood as a prompt or constraint on the large language model, guiding it to output results according to specific rules. For example, in the hotel accommodation field, for the "item delivery" intent, there can be a skill instruction defining the rules and response format for room item delivery; for the "inquiry" intent, there can be skill instructions containing knowledge base content related to the hotel and response requirements.

[0019] In this embodiment of the invention, the process of obtaining corresponding skill instruction information based on sub-intent is as follows: First, the system checks whether a skill instruction tool has already been used for the same type of need in the current session. If the user has previously made a similar request and called a certain skill tool in this session, the tool is reused first to ensure consistency and reduce redundant loading. For example, if the user previously asked about restaurant reservations and called the "restaurant reservation tool," it can be used directly when the restaurant reservation intent appears again. For new intents (where the corresponding skill tool has not been used in the current session), the system will select the most matching skill instruction information from the skill knowledge base based on the category and content of the sub-intent, using a recommendation algorithm (such as collaborative filtering algorithm CF). This means that the system comprehensively considers the history of tool calls for other similar questions and business rules to recommend the best processing tool. If no skill instruction information matching the sub-intent can be found in the knowledge base, a preset fallback solution is activated, i.e., the "fallback skill" instruction information is called. Usually, this fallback instruction will guide the intelligent customer service to politely transfer the current request to a human service representative to ensure that the user's problem is responded to and not ignored by the system.

[0020] Each skill instruction is typically in text format, containing response logic, business rules, and precautions tailored to a specific intent. Once a skill instruction is selected, the system inputs it along with the sub-intent description and relevant user-related information into the large language model. The model then generates a compliant response based on the guidance in the skill instruction. Here, user-related information refers to business data related to the current user or request. For example, in a hotel setting, user-related information might include the user's name or title, check-in date, room number, membership level, and the availability and pricing of the requested item; in a fitness setting, it might include the user's appointment history and personal trainer information; and in a theater setting, it might include the user's ticket purchase history and membership discount information. With this user-related information, intelligent customer service can provide personalized and accurate responses based on the actual situation. For example, when a guest requests extra bottled water, the system checks the complimentary bottled water quota and inventory for the user's room type: if the inventory is sufficient and within the free quota, delivery is arranged and the water is provided free of charge; if the free quota is exceeded, the system informs the user of the charges; if the inventory is insufficient, the system apologizes and contacts relevant personnel to replenish the water. Similarly, when a user requests a late check-out, the system retrieves the user's check-in information to determine if the room has subsequent bookings for the day and whether the user qualifies for late check-out (e.g., membership level or payment status), thus providing an accurate response or solution. This user-related information, combined with a skills knowledge base, ensures that the large language model's responses not only contain general business knowledge but also reflect the user's current situation.

[0021] In practical implementation, to ensure that the responses from the large language model strictly adhere to the information scope provided by the skills knowledge base, this invention can also input intervention or control strategy information into the model. For example, strategy constraints can be added to the prompts, requiring the model to "only answer based on the provided tool content; if the question exceeds the content, it should politely refuse or be transferred to a human agent." Similarly, it can include error correction strategies (automatically detecting and avoiding logical contradictions or factual errors), recommendation optimization strategies (providing relevant suggestions to users to optimize the experience where possible), and automatic completion strategies (intelligently completing details when the user's expression is incomplete). These strategies work together to ensure that the responses generated by the model do not deviate from the business knowledge base, do not arbitrarily fabricate non-existent promises, and conform to customer service requirements in terms of tone and format. For example, for special requests not covered in the knowledge base, the model will respond with a safe answer such as "I'm sorry, I will notify the relevant staff to contact you," rather than making promises that cannot be guaranteed.

[0022] Through the above steps, the system generates corresponding response content for each sub-intent. If the user's original input contains only a single intent, a response is directly obtained for that intent; however, when the user input is split into multiple sub-intents, multiple independent response contents will be generated. At this point, the solution provided by this invention enters a stage of response merging and polishing to further improve the naturalness and simplicity of human-computer dialogue. Specifically, the intelligent customer service system analyzes and processes multiple responses according to preset content polishing instructions: checking whether there is information duplication or contradiction between the responses, and whether they can be semantically connected and merged. For example, if a user makes multiple delivery requests in a short period of time, the system can merge the responses to these requests into a single reply to reduce the sending of multiple messages; in multi-intent responses involving monetary calculations, the system ensures that the total price is calculated based on the quantity and unit price and provides a clear result during polishing; for multiple inquiries that require user confirmation or information, the system also tries to merge the questions. The content polishing process optimizes the wording and format of responses using a large language model combined with rules. The final merged response must be polite, professional, clear, and concise, avoiding repetition and redundancy, and maintaining consistency with the user's previous tone and style, thus providing a coherent and natural interactive experience. It's important to note that a fundamental principle is followed during polishing and merging: the essential meaning of each sub-response remains unchanged, and no content beyond business knowledge is added arbitrarily. If a sub-response is a clear processing instruction (e.g., "We will notify our staff to contact you as soon as possible"), this part remains unchanged during polishing, with sentence structure adjusted only as needed to connect the context.

[0023] After generating the final response, the electronic device sends or displays the response to the user, fulfilling the user's service needs. For example, regarding the aforementioned user request, "Please deliver a toothbrush to my room and book a yoga class at 8 AM tomorrow," the method of this invention breaks it down into two sub-intents: "Room Delivery - Toothbrush Delivery" and "Fitness Booking - Yoga Class Reservation." The system generates responses using a skill knowledge base: "Okay, we will deliver the toothbrush to you immediately." and "Your yoga class at 8 AM tomorrow has been booked. Have a great workout!" After merging and refining, the final response can be integrated into a single sentence or paragraph, encompassing both pieces of information, such as: "Okay, the toothbrush has been delivered to your room; in addition, your yoga class at 8 AM tomorrow has been successfully booked. Have a pleasant stay!" This way, when the user receives the response, they can immediately understand that both of their requests have been properly processed without needing to inquire about either one again. This fully demonstrates the advantages of this invention in improving the efficiency of multi-intent request processing and enhancing user experience.

[0024] In addition to the methods described above, this invention also provides a computer-readable storage medium storing computer-executable instructions that instruct a processor to perform the steps of the intelligent customer service problem decomposition method as described above. By loading these program instructions into the memory of an electronic device and having them executed by a processor, the software functionality of the intelligent customer service system can be upgraded, and the method of this invention can be used to process user requests.

[0025] Meanwhile, the present invention also provides an electronic device, such as a memory for storing program instructions and a processor for executing the instructions. When the program instructions stored in the memory are executed by the processor, the electronic device can realize the functional flow of the intelligent customer service problem decomposition method described above. For example, the electronic device can be a server, a cloud computing device, or a software terminal device deployed with an intelligent customer service system. When a user accesses the customer service system and submits a request through the electronic device, the device can intelligently parse and respond to the request according to the method of the present invention.

[0026] Compared with existing technologies, the proposed solution has significant advantages: First, by introducing multi-turn dialogue context and a preset rule base for user intent decomposition, the system can accurately handle complex multi-intent inputs, improving the granularity and accuracy of intent recognition. Second, by utilizing vertical industry knowledge bases and skill tools, the system can fully call upon the latest business rules and dynamic data to provide users with targeted professional responses, compensating for the shortcomings of traditional intelligent customer service in terms of industry knowledge. Third, the addition of parameter integrity verification and follow-up questioning mechanisms in dialogue processing reduces unnecessary repeated inquiries, ensuring that the system obtains key information in the first response, thus improving response efficiency. Furthermore, by merging and refining the response content, the coherence and conciseness of multi-intent responses are ensured, avoiding the problem of stiff and verbose machine responses, making human-computer dialogue more natural and smooth. Finally, this invention adopts a modular layered architecture design, with the main control module responsible for intent recognition and task allocation, and each business sub-module performing its own function, avoiding contextual confusion and facilitating system expansion and maintenance. When it is necessary to expand to new industry scenarios, only the corresponding rule base and skill tools need to be added or updated, demonstrating good scalability and adaptability. Attached Figure Description The above and other objects, advantages, and features of the present invention can be more clearly understood by referring to the following description of the illustrative drawings. Wherein: Figure 1 This is a schematic diagram of an intelligent customer service system architecture according to an embodiment of the present invention, showing the collaborative relationship between the main control module and various business skill sub-modules; Figure 2 This is a logical flowchart of a problem decomposition rule engine according to an embodiment of the present invention, used to illustrate the steps and strategies for multi-intent splitting and processing.

Claims

1. A large language model-based intelligent customer service question decomposition method, characterized in that, The method comprises the following steps: acquiring user input information; correcting errors and cleaning sensitive content of the user input information to obtain normalized current input; combining the normalized current input information with the user's historical dialogue context to identify the user's intention by using a large language model; when the user's intention contains multiple different requests, the user's intention is decomposed into multiple independent sub-intentions according to a preset dynamic rule base, each of the sub-intentions includes corresponding intention description, intention category and key parameters; for each of the sub-intentions, skill instruction information corresponding to the intention category and content of the sub-intention is determined from a pre-constructed skill knowledge base; if a certain sub-intention lacks key parameters required for processing, prompt information is generated to obtain the key parameters; the skill instruction information corresponding to each of the sub-intentions and user associated information are input into the large language model to obtain reply content of each of the sub-intentions; for the case where there are multiple reply contents, the multiple reply contents and a preset content polishing instruction are input into the large language model for merging and language polishing to obtain final reply content; the final reply content is fed back to the user.

2. The method of claim 1, wherein, The error correction of the user input information includes correcting errors of the input text by using a preset spelling checking and grammar correction model; the sensitive content cleaning of the user input information includes intercepting rule-violating words, private data or irrelevant content contained in the input information by using a keyword library and semantic rules.

3. The method of claim 1, wherein, The user input information includes one or more of text, voice and image; when the user input information is voice or image, the pre-processing process of converting voice into text and recognizing image into text is further included to generate the normalized current input.

4. The method of claim 1, wherein, The process of inputting the normalized current input and the user's historical dialogue context into the large language model to identify the user's intention includes: when the current input is associated with the historical context, the historical dialogue content is combined to assist in identifying the current user's intention; when the current input is not associated with the historical context, the user's intention is identified only according to the current input information; wherein for the current input containing only confirmation or polite language, if there is a pre-sequence dialogue to be confirmed, the current input is identified as a user confirmation intention, otherwise the input is ignored; for the current input containing only numbers, if the length of the numbers is less than a preset threshold, the numbers are treated as room number information, and if the length of the numbers is greater than the threshold, the numbers are treated as contact information for processing.

5. The method of claim 1, wherein, The intention category in the sub-intention includes at least one of the following: accommodation service-advising, accommodation service-item delivery, accommodation service-reservation, accommodation service-customer feedback, catering service-advising, catering service-reservation, catering service-customer feedback, fitness service-advising, fitness service-appointment, fitness service-complaint and feedback, fitness service-other, theater service-advising, theater service-ticket purchase, theater service-feedback, theater service-complaint, scenic area service-information advising, scenic area service-ticket related, scenic area service-play feedback, scenic area service-special needs.

6. The method of claim 1, wherein, The determining the skill instruction information corresponding to the sub-intent comprises: if there is used skill instruction information corresponding to the sub-intent in the current session, selecting the used skill instruction information; if there is no used skill instruction information, acquiring corresponding skill instruction information according to the intent category and intent content of the sub-intent through a preset recommendation algorithm; if the matched skill instruction information cannot be determined, calling a preset bottom skill instruction information to indicate the transfer to an artificial customer service processing.

7. The method of claim 1, wherein, The user association information comprises user personal and business related information and environmental business data; wherein the user personal and business related information at least comprises the user's check-in date and room number, and when the user belongs to a hotel scenario, further comprises the user's surname and membership level, and when the user belongs to a fitness scenario, comprises the user's reserved course information, and when the user belongs to a theater scenario, comprises the user's purchased ticket information; the environmental business data comprises article inventory information, which comprises the current inventory quantity of the requested article, the upper limit of the number of free rooms and the charging standard after exceeding the free quota.

8. The method of claim 1, wherein, In the process of inputting the skill instruction information into the large language model to generate the reply content, a dialogue control strategy information is further provided to limit the reply content generated by the large language model within the knowledge range of the skill instruction information, and the dialogue control strategy information comprises a prompt strategy, a correction strategy, a recommendation optimization strategy and an automatic completion strategy, so as to avoid the large language model generating a promise or irrelevant content beyond the preset knowledge base range. 9.A computer readable storage medium having computer readable instructions stored thereon, which, when executed by a processor, cause the processor to perform the method of any one of claims 1 to 8. 10.An electronic device comprising a memory and a processor, wherein the memory stores computer program instructions, and the processor is configured to execute the computer program instructions to cause the electronic device to perform the method of any one of claims 1 to 8.

Images