Intelligent customer service method and system

Abstract

The invention discloses an intelligent customer service method and system, and the method comprises the steps: carrying out the intention recognition and slot filling of user input, and obtaining structured intention information; executing dynamic intention management on the structured intention information to obtain a confirmation intention and knowledge enhancement information; executing the business process based on the confirmation intention, and combining the knowledge enhancement information in the execution process to obtain a business execution result; according to the method, the service logic is solidified through the preset service processing atlas in the execution engine, and then the illusion content is eradicated through accurate scheduling of the intelligent agent, so that the reply content is 100% accurate and reliable; business results, dialogue history and enhanced knowledge are fused, reply information is complete and deep, expert-level complete reply is provided, and user experience and professional degree are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent customer service technology, and in particular to an intelligent customer service method and system. Background Technology

[0002] Currently, intelligent customer service systems based on Large Language Models (LLM) are gradually becoming the mainstream in the industry. These systems can leverage the powerful natural language understanding and generation capabilities of LLM to engage in fluent and human-like dialogues with users, significantly improving the interactive experience.

[0003] However, it has inherent technical drawbacks: First, LLM responses are unpredictable and unstable, and may produce "illusionary" content that contradicts established business rules or facts, making it difficult to guarantee the accuracy and reliability of customer service responses; Second, LLM is essentially a closed text model, lacking the ability to interact with the external environment, and cannot proactively call internal enterprise API interfaces (such as order inquiries and business processing) or access real-time, structured business databases, resulting in its inability to handle substantive user requests involving specific business operations, thus greatly limiting its practicality. Summary of the Invention

[0004] To address the technical problems existing in the background art, this invention proposes an intelligent customer service method and system.

[0005] The present invention proposes an intelligent customer service method, comprising: S1. Perform intent recognition and slot filling on user input to obtain structured intent information; S2. Perform dynamic intent management on the structured intent information to obtain confirmed intent and knowledge enhancement information; S3. Execute business processes based on confirmed intentions, and combine knowledge-enhanced information during execution to obtain business execution results; S4. Generate a response based on the business execution results and output it to the user.

[0006] Preferably, the intent recognition specifically involves: performing semantic analysis on user input using a natural language processing model to identify the user intent category and obtain intent classification results; extracting entity information from the user input based on the semantic analysis, mapping the entity information to predefined slots to generate slot key-value pairs; and combining the intent classification results and the slot key-value pairs into initial intent data.

[0007] Preferably, the slot filling specifically involves: verifying the completeness of the required slots in the initial intent data based on predefined business rules; when there are missing required slots, generating a slot clarification query based on the missing required slots and obtaining supplementary user information; merging the supplementary user information with the slot key-value pairs in the initial intent data to generate complete structured intent information.

[0008] Preferably, step S2 specifically involves: determining the clarity of intent based on the structured intent information; when the intent is unclear, conducting an intent clarification dialogue to obtain a confirmed intent; when the intent is clear, using the intent in the structured intent information as the confirmed intent; and based on the confirmed intent, converting the user input into a query vector, and obtaining relevant knowledge fragments from the knowledge base through semantic retrieval to obtain knowledge enhancement information.

[0009] Preferably, the construction steps of the knowledge base are as follows: document slicing of internal enterprise documents to form several knowledge fragments; vectorizing the several knowledge fragments to generate vectorized representations of the several knowledge fragments; and storing the vectorized representations of the several knowledge fragments in a database to obtain the knowledge base.

[0010] Preferably, the semantic retrieval specifically involves: calculating the cosine similarity between the query vector and the vectorized representation of each knowledge segment; obtaining a similarity score for each knowledge segment based on the cosine similarity; sorting the knowledge segments based on the similarity scores; selecting a predetermined number of knowledge segments with the highest similarity scores as retrieval results; and outputting the retrieval results as knowledge enhancement information.

[0011] Preferably, step S3 specifically involves: selecting a business processing graph from a preset business processing graph based on the confirmed intent and knowledge enhancement information; performing logical judgment on the confirmed intent according to the selected business processing graph to obtain an execution path; calling business interfaces or operating business data according to the execution path to obtain raw execution data; and summarizing the raw execution data to obtain the business execution result.

[0012] Preferably, the logical judgment of confirming intent specifically involves: extracting key parameters from the structured intent information; and performing conditional branch judgments in the selected business processing graph based on the key parameters to determine the execution path.

[0013] Preferably, the execution path includes a parameter supplementation path and a direct execution path. The conditional branch judgment specifically involves: determining whether the key parameters include the required parameters for executing the current business operation; when the required parameters are missing, selecting the parameter supplementation path and obtaining the missing parameters by calling the query interface; when all required parameters are included, selecting the direct execution path and continuing to execute the corresponding business interface call or business data operation based on the existing key parameters.

[0014] This invention also provides an intelligent customer service system, comprising: an intent acquisition module, an intent management module, an execution module, and a response module. The intent acquisition module is used to identify and fill slots in user input to obtain structured intent information. The intent management module is used to perform dynamic intent management on the structured intent information to obtain confirmation intent and knowledge enhancement information. The execution module is used to execute business processes based on the confirmation intent and, during execution, incorporates knowledge enhancement information to obtain business execution results. The response module is used to generate a response based on the business execution results and output it to the user.

[0015] In this invention, business logic is solidified by pre-setting a business processing graph in the execution engine (Workflow engine). Then, through precise scheduling of agents, "illusory" content is eliminated, ensuring 100% accurate and reliable responses. By integrating business results, dialogue history, and enhanced knowledge, the responses are complete and in-depth, providing expert-level answers and improving user experience and professionalism. This invention, by setting up functions that can call business interfaces or manipulate business data along the execution path, endows the system with substantial business processing capabilities, truly achieving end-to-end automated service from consultation to processing. Combined with the text processing capabilities of the LLM model, this invention achieves separation of responsibilities, with each component performing its specific function, leveraging the combined advantages. Attached Figure Description

[0016] Figure 1 This is a flowchart of the intelligent customer service method proposed in this invention. Detailed Implementation

[0017] Reference Figure 1 The present invention proposes an intelligent customer service method, comprising: S1. Perform intent recognition and slot filling on the user input to obtain structured intent information; this step is divided into two processes: intent recognition and slot filling. When the system detects a natural language statement for input, it first performs intent recognition processing: S1.1.1 The agent uses a natural language processing model (such as LLM) to perform semantic analysis on the natural language input by the user; the LLM model, based on its powerful language understanding ability, determines which predefined business category the user's statement best fits, and obtains the intent classification result; For example, categorize user input such as "I want to cancel my flight insurance" as "Insurance Cancellation" and "What flights are available from Beijing to Shanghai today?" as "Flight Status Inquiry".

[0018] S1.1.2. While understanding the semantics of the entire sentence, the LLM model identifies and extracts key information fragments, i.e., entity information, from the sentence. This entity information is then mapped to predefined slots related to the current business intent, and then the slot key-value pairs are output.

[0019] For example, when a user inputs "I want to cancel the insurance for order 12345", after recognizing the intent as "cancel insurance", the model will extract the entity "12345" and map it to the predefined slot "order number", generating a key-value pair {"order number": "12345"}.

[0020] It is worth noting that this predefined slot also includes "Date" and "Business Type".

[0021] S1.1.3. Combine the intent classification results and the slot key-value pairs into initial intent data, mainly used to integrate the data output from the first two steps. For example, combine the intent classification results representing "what to do" (such as "cancel insurance") with the slot key-value pairs representing "what parameters to use" (such as {"order number": "12345"}). This is a preliminary structured data object, providing clear input for subsequent slot verification, intent clarification, and business process scheduling.

[0022] Next, the slot filling process will begin: S1.2.1 Verify the completeness of the required slots in the initial intent data based on predefined business rules; wherein, the predefined business rules are a set of coded and structured enterprise business knowledge systems, specifically including: an execution condition list, a logic judgment tree, and data constraints and permission specifications. The execution condition list specifies the parameters required for each execution operation, the logic judgment tree specifies the logical order of each operation, and the data constraints and permission specifications are used to determine whether the operation is allowed or legal.

[0023] Each business intent (such as "cancel insurance") has predefined parameters required for its execution, i.e., mandatory slots. The system checks the slot key-value pairs in the "Initial Intent Data" to determine whether these mandatory slots have been filled with valid values.

[0024] For example, for the intent to "cancel insurance," the order number is usually a required field. If a user says "I want to cancel insurance" (without providing an order number), the validation result will be "Missing required field: order number."

[0025] When a required slot is missing, the system, through the agent, proactively initiates multi-turn dialogues based on this missing slot, generating clarification queries. For example, if an order number is missing, the system might ask, "What is your order number?" This process can be driven by a simple template or generated by an LLM model for more natural-sounding questions. Waiting for the user's re-entry, the customer might answer, "The order number is 12345." The system can then retrieve the additional information from the user's re-entry.

[0026] S1.2.2. The obtained supplementary user information is parsed and filled into the missing slots. Then, it is merged with the existing slot key-value pairs in the "Initial Intent Data" to generate complete structured intent information. This is the final product of the slot filling step, which contains all the necessary parameters required to execute the current confirmation intent.

[0027] S2. Based on the structured intent information, determine the clarity of the intent. When the intent is unclear, the system uses the Agent to conduct an intent clarification dialogue to obtain a confirmed intent; when the intent is clear, the intent in the structured intent information is used as the confirmed intent. The determination is mainly based on the following three aspects: (1) Confidence judgment: If the confidence score of the primary intent is lower than a preset threshold (e.g., lower than 0.7), the intent is determined to be unclear; if the confidence scores of the first two or more identified intents are very close (e.g., the difference is less than 0.1), the intent is determined to be unclear. For example, if a user enters "flight," the system might simultaneously identify both "checking flight status" (confidence level 0.5) and "booking a ticket" (confidence level 0.45). Since both confidence levels are similar and not high, the system will determine that the intent is unclear and initiate a clarification process, such as asking, "Do you want to check flight status or book a new ticket?"

[0028] (2) Judgment of the completeness and rationality of slot filling: Even if the intent itself is clear, if the parameters necessary to execute the intent (mandatory slots) are missing or obviously unreasonable, the intent is "unclear" or "unexecutable" in the current context. Specifically, the operation can be based on the predefined business rules in step S1.2.1 to check whether the mandatory slots are empty, or to extract the slot values ​​with logical contradictions, which will determine that the intent is unclear.

[0029] for example: Regarding the situation where a required field is missing: A user says, "I want a refund." The intent (refund) is clear, but the required field (order number) is missing. The system determines this as "requiring clarification" and asks, "Which ticket do you want to refund? Please provide the order number."

[0030] Regarding situations where there is a logical contradiction in the slot value: a user says, "I'm looking for flights from Beijing to Shanghai tomorrow," but the system detects from the context or user profile that the user ID's "frequently used departure city" is "Guangzhou." The system might cautiously clarify by asking, "I'm looking for flights from Beijing to Shanghai tomorrow; is the departure point Beijing correct?"

[0031] (3) Judging based on the coherence with the dialogue context: As the memory of multi-turn dialogues, the Agent will judge whether the current intent is smoothly connected with the previous dialogue history. A sudden, unrelated intent switch may mean that the user has started a new topic, or that the previous understanding was wrong. In this case, the system can directly judge whether it is reasonable to switch from one intent to another based on the predefined business logic diagram. If it is not reasonable, clarification is required.

[0032] For example, if the conversation has been discussing "changing flight tickets," and the user suddenly asks, "What about insurance?" Although "insurance" might refer to canceling or purchasing insurance, the system will consider the context and prioritize understanding that it is related to the current order, clarifying: "Do you want to know the insurance information for the current flight, or do you want to cancel or purchase insurance for this trip?"

[0033] After confirming the intent, the system will convert the user input into a query vector. This operation is mainly to transform the natural language question into a mathematical representation that the computer can use for similarity calculation. Then, relevant knowledge fragments are obtained from the knowledge base through semantic retrieval to obtain knowledge enhancement information.

[0034] The specific steps for building the knowledge base are as follows: First, the company's complete long documents (such as product manuals, business rules, and FAQ documents) are cut into smaller, semantically independent knowledge fragments by chapter, paragraph, or fixed token length. Next, a text embedding model is used to convert each knowledge fragment from text into a numerical vector in a high-dimensional space (i.e., "vectorized representation"). Finally, the vectorized representations of all knowledge fragments, along with their corresponding original text, are systematically stored in a dedicated database (i.e., a vector database). This database is the constructed knowledge base. Building such a vector database essentially creates an "external brain" independent of the LLM (Limited Least Model) that can be updated and expanded in real time. The company can update the system's knowledge at any time by updating document slices and re-vectorizing them without retraining the expensive LLM model.

[0035] The specific operation of semantic retrieval is as follows: The cosine similarity between the query vector and the vectorized representation of each knowledge fragment stored in the knowledge base is calculated. Cosine similarity is a method that evaluates the similarity between two vectors by measuring the cosine of the angle between them in a multidimensional space. Cosine similarity measures the difference in direction between vectors, i.e., semantic consistency. It is insensitive to text length; a short query and a long document fragment can achieve a high similarity score as long as their core semantics are consistent. This makes it particularly suitable for text semantic matching.

[0036] Each cosine value calculated in the previous step is the similarity score of the corresponding knowledge segment. The system sorts all knowledge segments according to their similarity scores from high to low. From the sorted list, the system selects the top-ranked knowledge segments of a predetermined number (K) as the search results. The search results are then output as knowledge enhancement information.

[0037] The top-K knowledge fragments selected at the end are output as "knowledge enhancement information" and passed to the response generation module in the system.

[0038] Here, K is an adjustable hyperparameter. For example, K can be set to 3 or K=5. However, it should be noted that if the value of K is too small, it may not cover all aspects of the problem, resulting in incomplete information; if the value of K is too large, it may introduce irrelevant noise information, interfering with the LLM model's response generation.

[0039] S3. Based on the confirmed intent and knowledge enhancement information, select a corresponding selected business processing graph from the preset business processing graph; The WorkFlow engine receives a "confirmation intent" and its accompanying "structured intent information" (i.e., a set of intent and slot key-value pairs) from the Agent. Then, the WorkFlow engine parses and extracts the variables necessary to execute the current business process from this structured data packet; these variables are the "key parameters". The WorkFlow engine takes the extracted key parameters and makes conditional branch judgments at the decision nodes of the "selected business processing graph". Based on the judgment results, it selects the next path to take, that is, determines the execution path. The execution path includes the parameter supplementation path and the direct execution path. Specifically, the conditional branch judgment is as follows: The system determines whether the "key parameters" extracted from the structured intent information contain essential inputs required to perform the business operation, i.e., mandatory parameters; these parameters are mandatory requirements for calling downstream business interfaces. For example, calling the "insurance cancellation interface" requires a valid order number; calling the "flight status query interface" requires a flight number or (departure point, destination, and date).

[0040] Based on the judgment result, the following two different execution paths will be implemented: 1) Parameter supplementation path When the judgment result indicates that a required parameter is missing, the parameter supplementation path is selected, and the Workflow engine calls another predefined API, such as the "Order List Query Interface". This call typically uses other existing parameters (such as user ID) as input. The interface returns structured order list data. Subsequently, the system enters the "Information Summary" step to obtain the business execution results. The list is then formatted using an LLM model and presented to the user, guiding the user to make selections and confirmations.

[0041] 2) Direct execution path When all required parameters are included, the direct execution path is selected. The WorkFlow engine continues to execute the corresponding business interface calls or business data operations based on the existing key parameters. Then, after information aggregation by the LLM model, the business execution results are directly output to the customer. This process will again utilize the generation capabilities of the LLM model, inputting raw data and a carefully designed prompt (instructions and examples) into the LLM model, allowing the LLM model to autonomously organize the language and return it to the user.

[0042] For example, a user enters "cancel insurance for order 12345". The system recognizes the intention to cancel insurance, and the key parameter, order number "12345", is complete. The Workflow engine directly selects the "direct execution path", calls the "insurance cancellation interface" and passes in the order number to complete the business operation. This is the most ideal and smoothest execution flow. The system requires no additional preparation and directly performs core business operations.

[0043] This process optimizes the user experience by transforming rigid, structured data into a friendly and natural dialogue, and achieves separation of responsibilities: WorkFlow is responsible for "accurate execution," while the LLM model is responsible for "elegant expression," each fulfilling its function and leveraging the combined advantages.

[0044] S4. Generate a response based on the business execution results and output it to the user. During this process, the Agent is responsible for collecting and integrating all relevant information from previous steps to form a complete context, serving as a resource library for generating the response. This relevant information includes: Business execution results: Structured or semi-structured data returned from the WorkFlow engine after "information aggregation". This is the core factual basis for the response. For example, confirmation information of successful insurance cancellation, a list of queried orders, flight status data, etc.

[0045] Conversation history: The Agent maintains the entire session context as a "memory of multi-turn conversations." This ensures that responses flow naturally from previous conversations, maintaining coherence.

[0046] Knowledge-enhancing information: Relevant knowledge fragments retrieved from the knowledge base in step S2. This knowledge provides background explanations, rule descriptions, or factual supplements to the response, increasing its professionalism and depth.

[0047] The agent uses the fused information as a new prompt, invokes the natural language generation capabilities of the large language model, and organizes it into a final response. Finally, the LLM-generated, "security-checked" natural language text is output to the user through a user interface (such as a web chat window, speech synthesis, etc.).

[0048] Now, taking "canceling insurance" as an example, we will demonstrate the specific operation process of step S4: Input 1 (Business Execution Result): After the WorkFlow engine executes the insurance cancellation operation, it returns a success message; Input 2 (dialogue history): The user's previous question: "I want to cancel my insurance," and the dialogue in which the system guided the user to select an order.

[0049] Input 3 (Knowledge Enhancement Information): May contain a knowledge fragment about "the refund will be returned to the original payment method within 3-5 business days after the policy is cancelled".

[0050] S4 Processing: After integrating this information, the Agent calls the LLM model. The LLM model will generate a response similar to the following: "Hello! Your flight insurance for order 12345 has been successfully cancelled. The refund will be returned to your payment account within 3-5 business days. Please check your account. Is there anything else I can help you with?"

[0051] This invention is strictly based on results returned by external business systems and retrieved enterprise knowledge, eliminating "illusory" content and ensuring 100% accurate and reliable responses. It integrates business results, dialogue history, and enhanced knowledge, providing complete and in-depth responses with expert-level expertise, thus improving user experience and professionalism. It achieves a closed loop between dialogue and business, upgrading intelligent customer service from a "question-answering machine" to a "business processing assistant," multiplying its value.

[0052] The present invention also provides an intelligent customer service system, including: an intent acquisition module, an intent management module, an execution module, and a response module.

[0053] The intent acquisition module comprises an analysis and identification unit, a slot mapping unit, and an initial intent generation unit. Specifically, the analysis and identification unit uses a natural language processing model (such as a Large Language Model, LLM) to perform semantic analysis on user input, identify the user intent category, and obtain intent classification results. The slot mapping unit extracts entity information from the user input based on the semantic analysis, maps the entity information to predefined slots, and generates slot key-value pairs. The initial intent generation unit combines the intent classification results and the slot key-value pairs into initial intent data; verifies the completeness of required slots in the initial intent data based on predefined business rules; when there are missing required slots, it generates slot clarification queries based on the missing required slots and obtains supplementary user information; finally, it merges the supplementary user information with the slot key-value pairs in the initial intent data to generate complete structured intent information.

[0054] The intent management module comprises a clarification judgment unit, an intent clarification management unit, and a knowledge retrieval unit. Specifically, the clarification judgment unit determines the clarification of intent based on the structured intent information (e.g., by evaluating intent confidence, slot completeness, and contextual coherence). The intent clarification management unit initiates an intent clarification dialogue when the intent is unclear to obtain a confirmed intent; when the intent is clear, it uses the intent from the structured intent information as the confirmed intent. This unit also serves as the memory for multi-turn dialogues, maintaining the dialogue context. The knowledge retrieval unit converts user input into a query vector based on the confirmed intent; calculates the cosine similarity between the query vector and the vectorized representation of each knowledge fragment in the knowledge base to obtain a similarity score for each knowledge fragment; sorts the knowledge fragments based on the similarity scores, selects a predetermined number of knowledge fragments with the highest similarity scores as retrieval results, and outputs them as knowledge enhancement information.

[0055] The execution module comprises a business graph selection unit, a logic judgment unit, an interface call and data operation unit, and an information aggregation unit. Specifically, the business graph selection unit selects a preferred business processing graph from a preset set of graphs based on the confirmed intent and knowledge enhancement information. The logic judgment unit performs logical judgments on the confirmed intent according to the selected business processing graph to obtain the execution path. The interface call and data operation unit calls business interfaces or operates business data according to the execution path to obtain the raw execution data. In the parameter supplementation path, it calls the query interface to retrieve missing parameters; in the direct execution path, it directly calls the target business interface. The information aggregation unit aggregates information from the raw execution data (usually by calling an LLM and constraining its output format) to obtain a user-understandable business execution result.

[0056] The response module comprises a multi-source information fusion unit and a natural language generation unit. Specifically, the multi-source information fusion unit receives and integrates business execution results, dialogue history, and knowledge enhancement information from the aforementioned modules to form a complete context for generating a response. The natural language generation unit utilizes the generation capabilities of the Large Language Model (LLM) to organize an accurate, natural, and complete final response based on the fused multi-source information, under specific prompt word constraints, and outputs it to the user.

[0057] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An intelligent customer service method, characterized in that, include: S1. Perform intent recognition and slot filling on user input to obtain structured intent information; S2. Perform dynamic intent management on the structured intent information to obtain confirmed intent and knowledge enhancement information; S3. Execute business processes based on confirmed intentions, and combine knowledge-enhanced information during execution to obtain business execution results; S4. Generate a response based on the business execution results and output it to the user.

2. The intelligent customer service method according to claim 1, characterized in that, The intent recognition specifically involves: performing semantic analysis on user input using a natural language processing model to identify the user intent category and obtain intent classification results; extracting entity information from the user input based on the semantic analysis, mapping the entity information to predefined slots, and generating slot key-value pairs; The intent classification results and the slot key-value pairs are combined to form the initial intent data.

3. The intelligent customer service method according to claim 2, characterized in that, The slot filling process specifically involves: verifying the completeness of required slots in the initial intent data based on predefined business rules; when there are missing required slots, generating a slot clarification query based on the missing required slots and obtaining supplementary user information; merging the supplementary user information with the slot key-value pairs in the initial intent data to generate complete structured intent information.

4. The intelligent customer service method according to claim 1, characterized in that, S2 specifically involves: determining the clarity of intent based on the structured intent information; when the intent is unclear, conducting an intent clarification dialogue to obtain a confirmed intent; when the intent is clear, using the intent in the structured intent information as the confirmed intent. Based on the stated confirmation intent, the user input is transformed into a query vector, and relevant knowledge fragments are retrieved from the knowledge base through semantic retrieval to obtain knowledge enhancement information.

5. The intelligent customer service method according to claim 4, characterized in that, The specific steps for constructing the knowledge base are as follows: slice the internal documents of the enterprise to form several knowledge fragments; vectorize the knowledge fragments to generate vectorized representations of the knowledge fragments; and store the vectorized representations of the knowledge fragments in the database to obtain the knowledge base.

6. The intelligent customer service method according to claim 5, characterized in that, The semantic retrieval specifically involves: calculating the cosine similarity between the query vector and the vectorized representation of each knowledge segment; obtaining a similarity score for each knowledge segment based on the cosine similarity; sorting the knowledge segments based on the similarity scores; selecting a predetermined number of knowledge segments with the highest similarity scores as retrieval results; and outputting the retrieval results as knowledge enhancement information.

7. The intelligent customer service method according to claim 1, characterized in that, Specifically, S3 involves: selecting a business processing graph from a preset business processing graph based on the confirmed intent and knowledge enhancement information; performing logical judgment on the confirmed intent according to the selected business processing graph to obtain the execution path; Based on the execution path, call the business interface or operate the business data to obtain the raw execution data; summarize the information of the raw execution data to obtain the business execution result.

8. The intelligent customer service method according to claim 7, characterized in that, The specific steps of performing logical judgment on the confirmed intent are as follows: extracting key parameters from the structured intent information; and performing conditional branch judgment in the selected business processing graph based on the key parameters to determine the execution path.

9. The intelligent customer service method according to claim 8, characterized in that, The execution path includes a parameter supplementation path and a direct execution path. The conditional branch judgment is as follows: determine whether the key parameters include the required parameters for executing the current business operation; when the required parameters are missing, select the parameter supplementation path and obtain the missing parameters by calling the query interface; when all required parameters are included, select the direct execution path and continue to execute the corresponding business interface call or business data operation based on the existing key parameters.

10. An intelligent customer service system, characterized in that, include: Intent acquisition module: used to identify user input and fill in slots to obtain structured intent information; Intent Management Module: Used to perform dynamic intent management on the structured intent information to obtain confirmed intent and knowledge enhancement information; Execution module: Used to execute business processes based on confirmed intentions, and to combine knowledge-enhanced information during execution to obtain business execution results; Response module: Used to generate responses based on business execution results and output them to the user.

Images