Customer service robot interaction method, device, equipment, medium and program product

Abstract

The application relates to the technical field of artificial intelligence, and provides an interactive method, device, equipment, medium and program product of a customer service robot, the method comprising the following steps: determining a target operation process matched with a user demand indicated by a conversation context based on the conversation context; the conversation context is determined based on a user query to be replied and historical interaction data belonging to the same conversation as the user query; inputting process content corresponding to the conversation context and the target operation process into a node detection model to obtain a target process node to be executed output by the node detection model; the node detection model is used for determining a target process node most matched with the conversation context from a plurality of process nodes through semantic understanding capability; and generating reply information of the user query based on the target process node. The application solves the problem that a process execution is rigid and cannot adapt to user jump expression, realizes nonlinear and dynamic execution of an operation process, and further improves the interactive experience of the customer service robot.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an interaction method, apparatus, device, medium, and program product for a customer service robot. Background Technology

[0002] With the rapid development of mobile internet and artificial intelligence technologies, intelligent customer service robots have been widely used in e-commerce, finance, home appliance after-sales service, and other fields that require interaction with a large number of consumers, undertaking important functions such as answering user inquiries and guiding business transactions. Compared with traditional human customer service, intelligent customer service robots have significant advantages such as 24 / 7 online availability, fast response speed, low operating costs, and high service quality. In practical applications, in order to guide users through complex business transactions or troubleshooting (e.g., product returns and exchanges, home appliance malfunction reports), it is usually necessary to design standard operating procedures (SOPs). How customer service robots execute SOP processes is a technical requirement that urgently needs to be addressed.

[0003] Currently, most customer service robots employ a linear guided approach based on rules or finite state machines when executing Standard Operating Procedures (SOPs). Specifically, a strict order of steps is predefined (e.g., step A to step B to step C). The robot can only ask the user questions step-by-step according to the preset path, and the user must answer precisely on the current question to trigger the system to move to the next step. However, when faced with the complex and varied natural language expressions of end users, this rigid and inflexible process results in a poor user experience and can even cause user resentment. Therefore, improving the interactive experience of customer service robots is a pressing technical problem that needs to be solved. Summary of the Invention

[0004] This invention aims to at least solve one of the technical problems existing in the prior art. To this end, this invention proposes an interaction method for customer service robots, which can solve the problems of rigid process execution and inability to adapt to the user's jumpy expressions in the prior art, realize the non-linear and dynamic execution of the work process, and thus improve the interactive experience of customer service robots.

[0005] The present invention also provides an interactive device for a customer service robot, an electronic device, a non-transitory computer-readable storage medium, and a computer program product.

[0006] An interaction method for a customer service robot according to a first aspect embodiment of the present invention includes: Based on the session context, a target job flow matching the user needs indicated by the session context is determined; the session context is determined based on the user query to be answered and historical interaction data belonging to the same session as the user query. The session context and the process content corresponding to the target job process are input into the node detection model to obtain the target process node to be executed output by the node detection model; the process content includes multiple process nodes with logical dependencies, and the node detection model is used to determine the target process node that best matches the session context from the multiple process nodes through semantic understanding capabilities. Based on the target process node, the response information for the user query is generated.

[0007] According to the interaction method of the customer service robot of the present invention, the conversation context is determined based on the user query to be replied to and the historical interaction data belonging to the same conversation as the user query. Based on the conversation context, the target work process matching the user needs indicated by the conversation context is accurately determined. The conversation context and the process content corresponding to the accurate target work process are input into the node detection model to obtain the target process node to be executed output by the node detection model. The process content includes multiple process nodes with logical dependencies. The node detection model is used to determine the target process node that best matches the conversation context from multiple process nodes through semantic understanding capabilities. This enables the customer service robot to flexibly understand the user context, intelligently judge the progress of the work process, and flexibly jump between process nodes within the target work process, just like a human customer service representative. This not only greatly improves the intelligence level of the customer service robot in handling complex and long-link business (such as troubleshooting), but also effectively reduces the effort required for users in human-computer interaction, significantly improving the user's customer service experience and problem-solving efficiency. That is, it solves the problem of rigid process execution and inability to adapt to the user's jumpy expression in the prior art, realizes the non-linear and dynamic execution of the work process, and thus improves the interactive experience of the customer service robot.

[0008] According to one embodiment of the present invention, determining the target job flow that matches the user requirements indicated by the session context based on the session context includes: Retrieve search results from the knowledge base that match the user's query; Based on the session context and the search results, the target job flow is determined.

[0009] According to one embodiment of the present invention, retrieving search results from the knowledge base that match the user query includes: Retrieve the first candidate search result with the highest similarity to the user's query and the second candidate search result with the second highest similarity to the user's query from the knowledge base; If the similarity of the first candidate search result is greater than a preset similarity threshold, or the similarity difference is greater than a preset similarity difference, the first candidate search result is determined as the search result; the similarity difference is the difference between the similarity of the first candidate search result and the similarity of the second candidate search result. If the similarity of the first candidate search result is less than or equal to a preset similarity threshold, and the similarity difference is less than or equal to a preset similarity difference, the first candidate search result, the second candidate search result, and the user query are input into the classification search model to obtain the search result output by the classification search model; the classification search model is used to determine the search result most similar to the user query from the first candidate search result and the second candidate search result.

[0010] According to one embodiment of the present invention, retrieving from the knowledge base a first candidate search result with the highest similarity to the user query and a second candidate search result with the second highest similarity to the user query includes: The user query is rewritten to obtain multiple alternative queries; The similarity between the multiple candidate queries and each knowledge segment in the knowledge base is calculated to obtain multiple similarity calculation results; Based on the multiple similarity calculation results, the first candidate search result and the second candidate search result are retrieved from the knowledge base.

[0011] According to an embodiment of the present invention, determining the target job flow based on the session context and the search results includes: Obtain the user profile information of the user who inputs the query; The target job process is determined based on the session context, the search results, and the user profile information.

[0012] According to an embodiment of the present invention, the process content is determined based on the following method: The target workflow is input into the content generation model to obtain the workflow content output by the content generation model; the content generation model is trained based on sample workflows; or, The process content corresponding to the target operation process is determined from the template library; each process content template in the template library is generated by the content generation model.

[0013] According to an embodiment of the present invention, generating the response information for the user query based on the target process node includes: The target process node and the search results are input into the large language model to obtain the anthropomorphic response information output by the large language model; the search results are the results retrieved from the knowledge base that match the user's query.

[0014] According to one embodiment of the present invention, it further includes: Emotion recognition is performed on the user query to obtain the emotion recognition result; If the user's negative emotion is determined to be of the first level of negative emotion based on the emotion recognition result, the step of determining the target job process that matches the user's needs indicated by the conversation context is executed. If the user's negative emotion is determined to be of the second level of negative emotion based on the emotion recognition result, a reassuring message is output, and the step of determining the target operation process that matches the user's needs indicated by the conversation context is executed. If the user's negative emotion is determined to be at the third level of negative emotion based on the emotion recognition results, the human customer service interaction process is triggered. Wherein, the degree of the first negative emotion is less than the degree of the second negative emotion, and the degree of the second negative emotion is less than the degree of the third negative emotion.

[0015] According to one embodiment of the present invention, it further includes: If the user requirement is determined to be complete based on the session context, the step of determining the target job flow that matches the user requirement indicated by the session context is executed. If the user's request is determined to be incomplete based on the session context, the question content is output based on the session context to obtain a new user query input by the user based on the question content.

[0016] According to one embodiment of the present invention, it further includes: If it is determined based on the user query that interaction with the user is required, the step of determining the target job flow that matches the user needs indicated by the session context is executed. If it is determined that no interaction with the user is required based on the user query, the content corresponding to the user intent indicated by the user query is output.

[0017] An interactive device for a customer service robot according to a second aspect of the present invention includes: The process determination module is used to determine a target job process that matches the user needs indicated by the session context, based on the session context; the session context is determined based on the user query to be answered and historical interaction data belonging to the same session as the user query. The node detection module is used to input the session context and the process content corresponding to the target job process into the node detection model to obtain the target process node to be executed output by the node detection model; the process content includes multiple process nodes with logical dependencies, and the node detection model is used to determine the target process node that best matches the session context from the multiple process nodes through semantic understanding capabilities. The response generation module is used to generate response information for the user query based on the target process node.

[0018] An electronic device according to a third aspect of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the interaction method of any of the above-described customer service robots.

[0019] According to a fourth aspect of the present invention, a non-transitory computer-readable storage medium is provided thereon storing a computer program that, when executed by a processor, implements the interaction method of the customer service robot as described above.

[0020] A computer program product according to a fifth aspect of the present invention includes a computer program that, when executed by a processor, implements the interaction method of any of the above-described customer service robots.

[0021] The above-described one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects: The conversation context is determined based on the user query to be answered and historical interaction data belonging to the same conversation. Based on the conversation context, the target workflow matching the user's needs indicated by the conversation context is accurately determined. The conversation context and the corresponding workflow content are input into the node detection model, resulting in the output of the current target workflow node to be executed. The workflow content includes multiple workflow nodes with logical dependencies. The node detection model uses semantic understanding to determine the target workflow node that best matches the conversation context from among these multiple workflow nodes. This allows the customer service robot to flexibly understand the user's context, intelligently judge the progress of the workflow, and flexibly jump between workflow nodes within the target workflow, much like a human customer service representative. This not only greatly improves the intelligence level of the customer service robot in handling complex, long-chain businesses (such as troubleshooting), but also effectively reduces the effort required for users in human-computer interaction, significantly improving the user's customer service experience and problem-solving efficiency. It solves the problems of rigid workflow execution and inability to adapt to users' jumpy expressions in existing technologies, achieving non-linear and dynamic execution of the workflow, thereby improving the interactive experience of the customer service robot.

[0022] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is one of the flowcharts illustrating the interaction method of the customer service robot provided in this embodiment of the invention.

[0025] Figure 2 This is the second flowchart illustrating the interaction method of the customer service robot provided in this embodiment of the invention.

[0026] Figure 3 This is a schematic diagram of the interactive device of the customer service robot provided in an embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0029] Currently, most customer service robots employ a linear guided model based on rules or finite state machines when executing Standard Operating Procedures (SOPs). Specifically, a strict order of steps is predefined (e.g., step A to step B to step C). The robot can only ask the user questions step-by-step according to the preset path, and the user must answer precisely on the current question before the system can move to the next step. However, when faced with the complex and varied natural language expressions of end users, this rigid execution lacks flexibility, resulting in a poor user experience and even user resentment. For example, when inquiring about a problem, a user might describe multiple symptoms at once or directly state the results of attempts at certain intermediate steps (e.g., "I've cleaned the filter, but it's still not cooling"). Existing technology cannot understand this contextual information spanning multiple steps and will still mechanically require the user to perform the completed steps (e.g., continuing to ask "Please check if the filter is cleaned"), leading to a very poor user experience.

[0030] To address the shortcomings of existing technologies, even relying on keyword matching to trigger process nodes cannot accurately determine the user's current actual state based on deep semantics within a long context. When the user's description is not standardized or contains emotional expressions, the system is prone to misjudging process nodes, leading to irrelevant answers or getting stuck in a dead loop. In other words, the understanding of intent remains inaccurate, and the location of process nodes remains inaccurate.

[0031] To address the aforementioned problems, the present invention proposes the following embodiments. The interaction method of the customer service robot provided by the embodiments of the present invention is described below with reference to the accompanying drawings. The executing entity of this customer service robot interaction method can be the customer service robot's interaction system, a server, a desktop computer, a laptop computer, or a user's terminal, including but not limited to mobile phones, tablet computers, in-vehicle terminals, and smart home appliances.

[0032] Figure 1 This is one of the flowcharts illustrating the interaction method of the customer service robot provided in this embodiment of the invention, such as... Figure 1 As shown, the interaction method of the customer service robot includes steps 110, 120 and 130.

[0033] Step 110: Based on the session context, determine the target job flow that matches the user requirements indicated by the session context.

[0034] The session context is determined based on the user query to be responded to and the historical interaction data belonging to the same session as the user query.

[0035] Here, a user query refers to the natural language text or speech-to-text text that the user currently sends to the customer service chatbot. For example, the user query can be the first question at the start of the conversation (e.g., "What should I do if the air conditioner is not cooling"), or it can be the user's response to a question asked by the customer service chatbot during multiple rounds of dialogue (e.g., "I have already cleaned the filter").

[0036] Here, the conversation context is the key data foundation for understanding the user's true intent. Since users often omit subjects or background information in multi-turn conversations (e.g., only answering "useless" in the second round), relying solely on the current single-sentence user query often cannot accurately represent the user's needs; therefore, this embodiment of the invention limits the conversation context to be determined based on the user query to be responded to and the historical interaction data belonging to the same conversation as the user query.

[0037] In one specific embodiment, historical interaction data includes all historical text previously entered by the user in the current session (i.e., previously entered historical user queries) and all historical response information previously output by the customer service robot.

[0038] In one embodiment, the session context includes user queries and historical interaction data. That is, the session context can be represented as sequence data containing the complete dialogue history.

[0039] In another embodiment, key information is extracted from user queries and historical interaction data to obtain various key information items, and the session context includes these key information items. That is, the session context can be represented as a truncated or summarized context window to adapt to input length limitations. For example, if a user query is "My KFR-35GW / N8KS1-1P air conditioner has very poor cooling performance; the room is still not cool after running for half an hour, what's wrong?", the corresponding session context could be "the product model the user is inquiring about (KFR-35GW / N8KS1-1P), the problem (poor cooling performance), and the usage scenario (running for half an hour)".

[0040] Here, the target work process refers to a set of standardized operating procedures or guidelines predefined to address a specific type of user need; for example, it may be a target SOP. For instance, for the user need of "air conditioner not cooling," the corresponding target work process may include a series of process steps such as filter cleaning, temperature setting check, refrigerant check, and repair registration.

[0041] In one specific embodiment, the session context is input into the process recognition model to obtain the target job flow output by the process recognition model. This process recognition model can be a large model or other deep learning models. The process recognition model is trained based on sample session context.

[0042] Furthermore, user profile information of the user who entered the query is obtained, and the target workflow is determined based on the session context and user profile information.

[0043] It should be understood that by matching the target job flow based on the conversation context including historical interaction data, the embodiments of the present invention can effectively solve the problems of unclear referencing or intent drift in multi-turn dialogues. It not only focuses on what the user is saying at the moment, but also combines previous dialogue content (such as confirmed device models and troubleshooting faults), thereby more accurately locking the target job flow required by the user, avoiding entering the wrong job flow due to incorrect intent recognition, improving the accuracy of job flow matching, and thus improving the interactive experience of the customer service robot.

[0044] Step 120: Input the session context and the process content corresponding to the target job process into the node detection model to obtain the target process node to be executed currently output by the node detection model.

[0045] The process content includes multiple process nodes with logical dependencies. This process content includes specific information contained in the target work process, which is typically composed of multiple process nodes with logical dependencies. Each process node may include a node name (e.g., "Check the filter"), node message (e.g., "Please check if the filter is dusty"), jump conditions, and other information.

[0046] The node detection model is used to determine the target process node that best matches the session context from the multiple process nodes through semantic understanding capabilities. This node detection model is a machine learning model with deep semantic understanding capabilities. It can be a finely tuned large language model or a deep learning model with a small to medium number of parameters that has undergone specific training.

[0047] It should be noted that the session context (representing the user's current state and intent) and the process content (representing the optional solution path) are used together as input data and provided to the node detection model so that the node detection model can determine the target process node that matches the current state.

[0048] Furthermore, it should be noted that multiple process nodes with logical dependencies refer to processes in the target workflow where there is a business sequence or causal relationship between them (e.g., simple faults are usually checked first, then repairs are reported). However, logical dependencies do not necessarily imply mandatory linearity in execution. Therefore, the node detection model analyzes the input data through semantic understanding, examining the implicit state information contained in the session context (e.g., a user saying "It still doesn't work after cleaning" implies "The filter cleaning step has been completed" and "The fault has not been resolved"). The node detection model then semantically compares this implicit state information with the definitions of each process node, calculating the target process node that best matches the current state from among multiple process nodes.

[0049] For example, taking a C-end user in the home appliance industry inquiring about "poor air conditioning cooling effect" as an example, the process includes three process nodes: node A (cleaning the filter), node B (checking the temperature setting), and node C (guiding repair requests). The logical dependency is usually A->B->C. The user query is "I just washed the filter and adjusted the temperature to 20 degrees, but it's still not cooling." If following linear logic, the system in the initial state will forcibly execute node A and reply "Please try cleaning the filter first." However, after reading the session context, the node detection model of this embodiment understands that "the filter has been washed" (node ​​A has been completed) and "the temperature has been adjusted" (node ​​B has been completed). Through semantic matching, it finds that the current state best matches the triggering condition of node C (i.e., the previous investigation is invalid). Therefore, the node detection model directly outputs node C as the target process node to be executed.

[0050] For example, user queries provide real-time feedback, allowing users to dynamically adjust process nodes based on their feedback, such as dynamically adjusting investigation steps, i.e., dynamically selecting nodes based on user shortcut buttons or natural language interaction feedback.

[0051] It should be understood that the node detection model is no longer limited by preset linear jump rules; the node detection model can automatically identify the process nodes that the user has completed or the most urgent needs based on semantic understanding, thereby realizing the "jump" or "dynamic pruning" of process nodes. This solves the problem of rigid process execution and inability to adapt to the user's jumpy expression in the existing technology, realizes the non-linear and dynamic execution of the work process, and thus improves the interactive experience of the customer service robot.

[0052] It should be understood that, compared to keyword matching, the semantic understanding capability of this invention can handle more generalized and ambiguous user expressions. Even without matching specific keywords, as long as the semantic intent is consistent, the node detection model can accurately locate process nodes, significantly improving the task completion rate in complex scenarios such as troubleshooting, thereby enhancing the adaptability to complex scenarios and improving the interactive experience of customer service robots.

[0053] Step 130: Based on the target process node, generate the response information for the user query.

[0054] In one embodiment, a pre-set standard script in the target process node is directly invoked as the reply information.

[0055] In another embodiment, the target process node is input into a large language model to obtain a human-like response from the model's output. This can make the tone of the response more natural and gentle, and can dynamically incorporate the user's personalized address or reassuring phrases.

[0056] Furthermore, a quality analysis is performed on the target workflow, such as evaluating the quality of use through citation frequency and user satisfaction; at the same time, natural language processing is used to detect ambiguous expressions and node conflicts in the response information, and optimization suggestions are output to optimize the execution flow and network model of the embodiments of the present invention.

[0057] It should be understood that this invention breaks away from the limitations of traditional customer service systems that rely on preset rules or state machines for linear guidance. Leveraging the powerful semantic understanding capabilities of the node detection model, it performs a deep comparison and matching between the current session context and the content of the target workflow. Instead of mechanically executing workflow nodes sequentially, it intelligently locates the most appropriate target workflow node among multiple logically dependent workflow nodes based on explicit information or implicit states expressed by the user in the session context. This mechanism allows for non-linear workflow transitions (e.g., skipping completed steps), thereby achieving a more human-like, flexible, and efficient customer service interaction experience. In other words, by generating response information based on precisely located target workflow nodes, it ensures that the customer service robot's responses always closely align with the user's current actual business progress, thus avoiding irrelevant answers.

[0058] The customer service robot interaction method provided in this invention uses a conversation context determined based on the user query to be responded to and historical interaction data belonging to the same conversation as the user query. Based on the conversation context, it accurately determines the target workflow matching the user's needs indicated by the conversation context. The conversation context and the corresponding workflow content are input into a node detection model to obtain the currently executing target workflow node output by the node detection model. The workflow content includes multiple workflow nodes with logical dependencies. The node detection model uses semantic understanding to determine the target workflow node that best matches the conversation context from among the multiple workflow nodes. This enables the customer service robot to flexibly understand the user's context, intelligently judge the progress of the workflow, and flexibly jump between workflow nodes within the target workflow, much like a human customer service representative. This not only greatly improves the intelligence level of the customer service robot when handling complex, long-chain businesses (such as troubleshooting), but also effectively reduces the effort required for users in human-computer interaction, significantly improving the user's customer service experience and problem-solving efficiency. It solves the problems of rigid workflow execution and inability to adapt to users' jumpy expressions in existing technologies, achieving non-linear and dynamic execution of the workflow, thereby improving the interactive experience of the customer service robot.

[0059] Based on any of the above embodiments, in this method, step 110 includes: Retrieve search results from the knowledge base that match the user's query; Based on the session context and the search results, the target job flow is determined.

[0060] Here, a knowledge base refers to a pre-built structured or unstructured database that stores a large amount of domain knowledge. For customer service robot scenarios, a knowledge base may include, but is not limited to, at least one of the following: product manuals, frequently asked questions (FAQs), historical dialogues (such as historical fault case libraries, solutions to similar problems with the same model of air conditioner), work process libraries (such as SOP libraries), etc.

[0061] Furthermore, the knowledge fragments in the knowledge base are preprocessed into vector form and stored in a vector database to facilitate efficient semantic retrieval.

[0062] Here, the search results refer to the knowledge fragments that are most semantically relevant to the current user's query, retrieved from the knowledge base through search algorithms (such as vector similarity calculation, keyword matching, etc.).

[0063] In one specific embodiment, the user query (such as "the air conditioner is not cold") is first converted into a query vector, then the similarity between the query vector and the vectors of each knowledge segment in the knowledge base is calculated, and finally the knowledge segment with the highest similarity is selected as the retrieval result.

[0064] Furthermore, the user query is rewritten to obtain multiple alternative queries; the similarity between the multiple alternative queries and each knowledge segment in the knowledge base is calculated to obtain multiple similarity calculation results; based on the multiple similarity calculation results, the search result that best matches the user query is retrieved from the knowledge base.

[0065] In one specific embodiment, the session context and retrieval results are input into the process recognition model to obtain the target job flow output by the process recognition model. This process recognition model can be a large model or other deep learning models. The process recognition model is trained based on sample session context and sample retrieval results.

[0066] Considering that users' expressions are often unprofessional or vague, it is easy to make mistakes if the target work process is determined solely based on the conversation context. Therefore, the search results are used as an auxiliary basis for determining the target work process. The rich knowledge content retrieved is integrated with the original conversation context and used as input information to make a decision on the selection of the target work process, thereby improving the accuracy of the determination of the target work process.

[0067] For example, the professional descriptions in the search results (such as "typical symptoms of refrigerant leakage") are concatenated with the conversation context to form an enhanced semantic representation. Then, a process identification model is used to predict the most matching target workflow based on this enhanced semantic representation. For instance, if a user says "the air conditioner is only blowing air," the process identification model may have difficulty distinguishing between "incorrect mode setting" and "malfunction." In this case, the knowledge base retrieves relevant cases such as "the air conditioner is only blowing air and the outdoor unit is not turning, which may be due to a compressor starting capacitor failure." Combining this search result, the target workflow can be more likely to be identified as a "hardware fault troubleshooting SOP" rather than a simple "remote control setting SOP."

[0068] For example, considering that repeated retrieval of high-frequency questions will lead to low interaction efficiency, and that the lack of an efficient caching mechanism will further exacerbate response latency, a knowledge base includes a common questions database, historical dialogues, or work process databases. Based on this, search results including work processes (such as SOPs) can be quickly retrieved using the knowledge base, allowing for rapid determination of the target work process and thus improving interaction efficiency. For instance, if the knowledge base reveals that "poor cooling effect" is a high-frequency question for this model of air conditioner, and cached results exist in the knowledge base, combining this with real-time retrieval will yield the latest solution (i.e., the work process included in the search results).

[0069] It should be understood that for some long-tail questions or obscure expressions not covered in the training corpus, directly determining the target workflow based on the conversation context may be inaccurate. However, by searching the knowledge base, similar historical cases or document descriptions can often be found. These search results, as prompts, can guide the correct association with the corresponding target workflow, thereby improving the interaction accuracy of the customer service robot.

[0070] The customer service robot interaction method provided in this embodiment of the invention retrieves search results from the knowledge base that match the user's query, thereby potentially obtaining richer and more professional domain knowledge and background information than the user's query. By using these search results to disambiguate and enhance the user's intent, it can significantly reduce the error in matching the workflow caused by the user's ambiguous expression, ensure that the customer service robot enters the correct interaction process, thereby improving the interaction accuracy of the customer service robot and ultimately improving the interaction experience of the customer service robot.

[0071] Based on any of the above embodiments, in this method, retrieving the search results matching the user query from the knowledge base includes: Retrieve the first candidate search result with the highest similarity to the user's query and the second candidate search result with the second highest similarity to the user's query from the knowledge base; If the similarity of the first candidate search result is greater than a preset similarity threshold, or the similarity difference is greater than a preset similarity difference, the first candidate search result is determined as the search result. If the similarity of the first candidate search result is less than or equal to a preset similarity threshold, and the similarity difference is less than or equal to a preset similarity difference, the first candidate search result, the second candidate search result, and the user query are input into the classification search model to obtain the search result output by the classification search model.

[0072] Here, the first candidate search result is the result corresponding to TOP1, and the second candidate search result is the result corresponding to TOP2.

[0073] Here, the similarity corresponding to the first candidate search result is the similarity calculation result between the first candidate search result and the user query; the similarity corresponding to the second candidate search result is the similarity calculation result between the second candidate search result and the user query.

[0074] Here, the preset similarity threshold can be set according to actual needs, and the embodiments of the present invention do not impose specific limitations.

[0075] The similarity difference is the difference between the similarity scores of the first candidate search result and the second candidate search result. For example, the difference between the similarity scores of the first and second candidate search results can be defined as the similarity difference. It should be understood that if the similarity difference is large, it indicates that the first candidate search result is significantly better than the second candidate search result, and there is no confusion. Therefore, the first candidate search result can be directly determined as the search result.

[0076] Here, the preset similarity difference can be set according to actual needs, and the embodiments of the present invention do not impose specific limitations.

[0077] The classification retrieval model is used to determine the retrieval result most similar to the user's query from the first candidate retrieval result and the second candidate retrieval result. This classification retrieval model can be a large model with strong logical reasoning capabilities, or it can be other deep learning models. The classification retrieval model primarily performs a classification or ranking task. It deeply understands the semantic details of the user's query and compares the content of the two candidate retrieval results to determine which one truly meets the user's needs. In other words, for similar content, the classification capability of the classification retrieval model achieves accurate retrieval.

[0078] In one specific embodiment, the first candidate search result, the second candidate search result, and the user query are concatenated into a single input data set, which is then fed into the classification retrieval model. For example, the first candidate search result, the second candidate search result, and the user query are assembled into a Prompt, which is then input into the classification retrieval model.

[0079] Furthermore, the classification retrieval model uses a streaming output method, meaning that the retrieval result is determined as soon as the first token is streamed, thus improving the efficiency of result determination and ultimately enhancing interaction efficiency. For example, if the complete result output by the classification retrieval model is A.top1 or B.top2, if the first streaming output is A, the retrieval result is determined as the first candidate result; if the first streaming output is B, the retrieval result is determined as the second candidate result.

[0080] It should be understood that the embodiments of the present invention avoid calling expensive classification retrieval model computing resources for all user queries, thus ensuring response speed and improving interaction efficiency. For example, a small model can be used for rapid retrieval to obtain preliminary similar first and second candidate retrieval results. For similar content, the classification capability of a large model (classification retrieval model) is then used to achieve accurate retrieval.

[0081] The customer service robot interaction method provided in this invention introduces the deep semantic understanding capability of a classification retrieval model for arbitration when similarity retrieval is difficult to distinguish. This effectively solves the problem of "answering the wrong question" caused by high noise in recall results in traditional RAG (Retrieval-Augmented Generation), thereby improving retrieval accuracy, which in turn improves the accuracy of determining the target workflow, and ultimately improves the interaction accuracy of the customer service robot. At the same time, the classification retrieval model is not called for all user queries. Instead, an admission mechanism is set by setting a preset similarity threshold and a preset similarity difference. For simple and clear user queries, the retrieval results are returned directly, achieving millisecond-level response. The classification retrieval model is only called when necessary. This strategy significantly reduces the consumption of computing resources and the average response latency while ensuring high overall accuracy, thereby improving retrieval efficiency, which in turn improves the efficiency of determining the target workflow, and ultimately improves the interaction efficiency of the customer service robot.

[0082] Based on any of the above embodiments, in this method, retrieving from the knowledge base the first candidate search result with the highest similarity to the user query and the second candidate search result with the second highest similarity to the user query includes: The user query is rewritten to obtain multiple alternative queries; The similarity between the multiple candidate queries and each knowledge segment in the knowledge base is calculated to obtain multiple similarity calculation results; Based on the multiple similarity calculation results, the first candidate search result and the second candidate search result are retrieved from the knowledge base.

[0083] Considering the high diversity and arbitrariness of users' natural language expressions, directly using the original user query for retrieval may be limited by specific keywords or sentence structures, resulting in a narrow search scope. Therefore, the user query is rewritten to broaden the search to obtain multiple alternative queries.

[0084] For example, by using synonym replacement, non-standard terms in the user query are replaced with professional terms (e.g., expanding "air conditioner is not cooling" to "air conditioner has poor cooling effect"); by using intent completion, the omitted subject or object in the user query is completed based on the session context (e.g., completing "how to clean" to "KFR-35GW model air conditioner filter cleaning method"); and multiple variants that are semantically similar to the original user query but have different expressions are generated (e.g., for "no response when turning on", alternative queries such as "cannot start" and "power indicator light is not on" are generated).

[0085] It should be noted that multiple alternative queries may or may not include the original user query.

[0086] In one specific embodiment, the user query is rewritten using a large language model to obtain multiple alternative queries.

[0087] Specifically, each candidate query is used as an independent retrieval input, and its vector similarity is calculated with the massive amount of knowledge fragments in the knowledge base. The candidate query can obtain its corresponding query vector through a single vector model, while the vectors of the knowledge fragments in the knowledge base can be prepared in advance, i.e., their corresponding vectors are obtained beforehand through the vector model.

[0088] Furthermore, multiple candidate queries are compared with each knowledge fragment in the knowledge base in parallel (concurrent) to obtain multiple similarity calculation results, thereby improving retrieval efficiency and thus improving the efficiency of determining the target workflow. For example, if three candidate queries are generated, three vector retrieval requests will be initiated in parallel.

[0089] It should be understood that by rewriting the semantic boundaries of user queries, it is possible to capture knowledge fragments that are not entirely matched with the original user query keywords but are highly semantically related. This effectively solves the problem of missed detections caused by inaccurate user expressions or inappropriate terminology, thereby improving retrieval recall, i.e., improving retrieval accuracy, and consequently improving the accuracy of determining the target workflow.

[0090] It should be understood that even if a particular alternative query performs poorly, other alternative queries (especially those that have been standardized or completed) may still yield correct results. This multi-path complementary mechanism enables stable retrieval performance when faced with various long-tail and colloquial user queries, thereby enhancing retrieval robustness and improving the accuracy of determining the target workflow.

[0091] The customer service robot interaction method provided in this embodiment of the invention rewrites the user query to obtain multiple alternative queries, and performs similarity calculation based on the multiple alternative queries, thereby improving the retrieval accuracy, thereby improving the accuracy of determining the target operation process, thereby improving the interaction accuracy of the customer service robot, and ultimately improving the user experience.

[0092] Based on any of the above embodiments, in this method, determining the target job flow based on the session context and the search results includes: Obtain the user profile information of the user who inputs the query; The target job process is determined based on the session context, the search results, and the user profile information.

[0093] Here, user profile information refers to stored attribute data and historical behavioral tags associated with the current user account. In customer service robot scenarios, user profile information is crucial for understanding users' potential needs. This user profile information may include, but is not limited to, at least one of the following: purchased product information, user identity tags, and historical behavioral preferences. Purchased product information includes, for example, the device model bound to the user's name (e.g., KFR-35GW / N8KS1-1P), purchase time, warranty status, etc.; user identity tags include, for example, new users, VIP members, users who have reported repairs multiple times, etc.; historical behavioral preferences include, for example, the user's preferred service channel (self-service / human assistance), and high-frequency words in historical consultation records, etc.

[0094] In one specific embodiment, user profile information can be retrieved in real time from the enterprise's customer relationship management system or order database using the user ID.

[0095] In one specific embodiment, the session context, search results, and user profile information are input into the process recognition model to obtain the target job flow output by the process recognition model. This process recognition model can be a large model or other deep learning models. The process recognition model is trained based on sample session context, sample search results, and sample user profile information.

[0096] For example, the process identification model can use user profile information to automatically fill in the key slots required for job process matching. For instance, job process selection often depends on "equipment model". If the user profile shows that the user only owns a "wall-mounted air conditioner", then even if the user only mentioned "air conditioner" in the query, the "equipment type" can be automatically locked as "wall-mounted", thereby excluding job processes related to "central air conditioning".

[0097] For example, if a user queries "how to remove the filter", and there is no user profile information, the air conditioner type cannot be determined. Usually, the customer service robot will ask, "Is your air conditioner a wall-mounted or floor-standing unit?" (adding an extra round of interaction); if there is user profile information, and the user profile shows that the user purchased a "KFR-35GW wall-mounted unit", the robot will directly locate "wall-mounted unit filter removal SOP" and directly output graphic and text instructions for that model.

[0098] The customer service robot interaction method provided in this invention can proactively use known user profile information (such as device model) to complete missing items in user queries, avoiding the customer service robot repeatedly asking users for known basic information. This significantly reduces the number of interaction rounds, reduces user effort, and improves user experience. Furthermore, different types of users (such as different device owners) may have completely different solutions to the same problem. By combining user profile information, the method can accurately match the most suitable target workflow for a specific user's current situation, avoiding the push of generic but ineffective solutions. This improves the personalized matching of target workflows, thereby enhancing the interaction accuracy of the customer service robot and ultimately improving user experience.

[0099] Based on any of the above embodiments, the process content in this method is determined in the following way: The target workflow is input into the content generation model to obtain the workflow content output by the content generation model.

[0100] Alternatively, the process content can be determined based on the following method: The process content corresponding to the target work process is determined from the template library.

[0101] Considering that the process content relies on manual writing and maintenance, which is inefficient and prone to ambiguity, a content generation model is used to automatically generate structured and standardized process content.

[0102] The content generation model is trained based on a sample job workflow. This content generation model can be a large model or other deep learning models.

[0103] In one embodiment, sample workflows can be derived from high-quality customer service dialogue records accumulated over time (by mining frequently encountered solutions), standard troubleshooting manuals written by experts, or industry knowledge bases. Through learning from these samples, the content generation model acquires the ability to transform an abstract business intent (such as "troubleshooting air conditioner not cooling") into structured, clearly defined operational instructions. For example, by deeply mining high-quality customer service dialogue data and accumulating relevant guidance, the content is clustered and categorized into various workflows, thereby determining the corresponding workflow content based on the workflow, significantly reducing manual maintenance costs.

[0104] For example, once the ID or name of the target workflow is determined, it is input as an instruction to the content generation model, which then generates workflow content containing specific dialogue and node logic in real time.

[0105] The process content templates in the template library are generated by the content generation model. That is, process content templates are pre-generated offline, and then, before the system goes live or during periodic maintenance, the content generation model is used to generate process content templates for various job processes in batches. This method offers faster response times and more controllable content quality.

[0106] The customer service robot interaction method provided in this invention automatically generates process content using a content generation model, freeing manual labor from tedious writing work and greatly reducing manual maintenance costs. Simultaneously, the content generation model, trained on a large number of samples, generates process content that is often superior to manually written content in terms of clarity and logical rigor. This avoids the problem of inconsistent process content quality due to differences in personnel skill levels, thus improving the standardization and quality of process content. This, in turn, enhances the accuracy of identifying target process nodes, thereby improving the interaction accuracy of the customer service robot and ultimately improving the user experience.

[0107] Based on any of the above embodiments, considering that current responses are mechanical and lack contextual coherence, and that the difference between customer service robots and human agents often lies in the tone of voice, users prefer a more lively and emotionally engaging customer service style compared to the rigid, formal responses of traditional robots; therefore, in this method, step 130 includes: The target process node and the search results are input into the large language model to obtain the anthropomorphic response information output by the large language model.

[0108] Here, the target process node specifies the business action that must be performed at this time (such as "requesting the user to provide the serial number") or core guidance (such as "instructions for cleaning the filter").

[0109] The search results are those retrieved from the knowledge base that match the user's query. As mentioned above, the search results are the most relevant results retrieved from the knowledge base that match the user's query. They contain detailed domain knowledge that is not pre-defined in the target process node. For example, the target process node may only indicate "information on repair policy," while the search results may include specific details such as "24-hour on-site service commitment" and "fee standards."

[0110] Here, the large language model leverages its powerful language generation capabilities to reorganize and refine the input information to produce a more human-like response. For example, it transforms rigid, directive language in the target process nodes (such as "Execute step 1") into conversational, gentler expressions (such as "First, we need to execute step 1"). Based on a pre-set persona or detected user emotions, it adds modal particles, emoticons, or empathetic phrases (such as "I'm sorry for the inconvenience"). Appropriate conjunctions (such as "then," "in addition") naturally connect the target process nodes with the search results, forming a coherent response rather than a mechanically pieced-together message. For instance, by incorporating conversational expressions, modal particles, logical connectors, emoticon tags, and naturally segmenting long texts using the large language model, and supporting customizable interaction styles, it avoids mechanical responses, thereby enhancing user intimacy and improving the user experience.

[0111] For example, the reply message might be: "Regarding your KFR-35GW / N8KS1-1P air conditioner, please follow these steps: 1. Clean the filter: Open the indoor unit panel, remove the filter, rinse it with clean water, dry it, and then reinstall it (a clogged filter will affect airflow); 2. Check the temperature setting: It is recommended to set the temperature to 22-26℃ and close doors and windows to reduce heat inflow; 3. If the above steps do not improve the situation, you can submit a fault report through the APP's 'My - After-sales Repair' section. We will arrange for an engineer to contact you within 24 hours. Feel free to contact me anytime if you have any questions!"

[0112] The customer service robot interaction method provided in this invention, through the above-mentioned approach and in combination with search results, compensates for the lack of specific knowledge details in the target process nodes, enabling the customer service robot to both guide the process and answer detailed questions, thereby improving the completeness and practicality of the response information, that is, improving the interaction accuracy of the customer service robot, and ultimately improving the user experience; at the same time, it gets rid of the mechanical response style of traditional robots, and the anthropomorphic response information brings the robot closer to the user, reduces the user's resistance to the customer service robot service, and improves the friendliness and satisfaction of the service, that is, improves the user experience.

[0113] Based on any of the above embodiments Figure 2 This is a second flowchart illustrating the interaction method of the customer service robot provided in this embodiment of the invention. Figure 2 As shown, the method includes steps 210, 220, 230, 240, 120, and 130.

[0114] Step 210: Perform emotion recognition on the user query to obtain the emotion recognition result.

[0115] In one specific embodiment, a user query is input into an emotion recognition model, and the emotion recognition result output by the model is obtained. This emotion recognition model is trained based on sample user queries.

[0116] Furthermore, multi-dimensional emotion recognition is performed on user queries to accurately determine the emotion recognition results. This analysis dimension may include, but is not limited to, at least one of the following: text semantic features (such as interjections, contextual semantics, etc.), punctuation features (such as consecutive exclamation marks "!!!"), interactive behavior features (such as typing speed, repetitive input, etc.), and acoustic features (if it is voice input, such as volume, speech rate, tremor, etc.).

[0117] Wherein, the degree of the first negative emotion is less than the degree of the second negative emotion, and the degree of the second negative emotion is less than the degree of the third negative emotion.

[0118] In one embodiment, the emotion recognition results are quantified into different levels of negative emotion (e.g., levels 1-5). In this embodiment of the invention, they are divided into three progressively increasing levels: a first level of negative emotion (mild, such as doubt or mild anxiety), a second level of negative emotion (moderate, such as dissatisfaction or complaint), and a third level of negative emotion (severe, such as anger or threatening complaints).

[0119] Step 220: If the user's negative emotion is determined to be of the first negative emotion level based on the emotion recognition result, the step of determining the target job process that matches the user's needs indicated by the conversation context is executed.

[0120] It should be noted that at this point, the user's emotions are relatively stable, and no additional intervention is needed. Business processes should be carried out directly according to the normal interaction flow, with the primary goal of efficiently resolving the problem.

[0121] Step 230: If the user's negative emotion is determined to be of the second level of negative emotion based on the emotion recognition result, output a soothing message and execute the step of determining the target operation process that matches the user's needs indicated by the conversation context based on the conversation context.

[0122] It's important to note that the priority is to quickly and empathetically reassuringly respond with a message (such as "I completely understand your anxiety, please don't worry, I'll investigate for you right away") to calm the user and prevent complaints. Simultaneously, the matching and retrieval logic of the target workflow continues, and once a response is generated, a business solution is then provided. This strategy adds an emotional dimension without disrupting the business process, thereby improving the user experience.

[0123] Furthermore, different levels of negative emotions correspond to different comforting phrases, meaning that personalized comforting phrases can be output.

[0124] Step 240: If the user's negative emotion is determined to be at the third level of negative emotion based on the emotion recognition result, the human customer service interaction process is triggered.

[0125] It's important to note that when users are extremely agitated, any automated response from the customer service chatbot (even a reassuring one) may be perceived as perfunctory. Therefore, directly triggering a "transfer to human agent" command seamlessly transfers the current conversation (along with the conversation context and emotion recognition results) to a human agent, or generates a work order for urgent follow-up. Based on this, a specific handling process should be established for severely negative emotions, allowing users to be promptly transferred to a human agent or prioritized when dissatisfied, thereby improving user experience and reducing the complaint rate.

[0126] The customer service robot interaction method provided in this invention, through the above-described approach, maintains automated processing for mild negative emotions (first level of negative emotion), allocating valuable human resources only to key users who most need reassurance, thus achieving the best balance between service cost and service quality; for moderate negative emotions (second level of negative emotion), it provides both reassurance and business response, simulating the service logic of excellent human customer service—"addressing emotions first, then addressing the issue"—making users feel valued and understood, thereby improving overall satisfaction, i.e., enhancing user experience and avoiding complaints; through timely identification and human intervention for severe negative emotions (third level of negative emotion), it can intervene before the conflict escalates, effectively reducing complaint rates and user churn risk, i.e., enhancing user experience.

[0127] Based on any of the above embodiments, the method further includes: If the user requirement is determined to be complete based on the session context, the step of determining the target job flow that matches the user requirement indicated by the session context is executed.

[0128] Prior to step 110 above, the method further includes: If the user's request is determined to be incomplete based on the session context, the question content is output based on the session context to obtain a new user query input by the user based on the question content.

[0129] In one specific embodiment, a natural language understanding model or a large language model is used to perform semantic completeness analysis on the current conversation context.

[0130] For example, "key elements" or "necessary slots" required for various business scenarios are predefined. For instance, for a "troubleshooting" scenario, necessary elements might include "equipment type" and "fault phenomenon"; for an "order inquiry" scenario, necessary elements might include "order number" or "phone number". The session context is scanned to check whether these necessary elements have all been filled in, i.e., whether they are complete. For example, the session context contains enough information to uniquely identify a target workflow, such as a user saying "My KFR-35GW air conditioner (equipment) is not cooling (phenomenon)"; or, for example, the session context lacks key information, resulting in unclear intent, such as a user only saying "My air conditioner is broken" (lacking specific fault phenomenon) or "How to clean it" (lacking equipment type, because wall-mounted and floor-standing units have different cleaning methods).

[0131] It should be noted that if all the prerequisites for decision-making are met, the steps for determining the target workflow can continue. If specific missing elements are identified, targeted questions (i.e., clarification statements) can be generated accordingly. For example, if "fault phenomenon" is identified as missing, the question "What specific problem is your air conditioner experiencing? Is it not cooling at all or is the cooling effect poor?" can be generated; if "equipment type" is identified as missing, the question "Do you mean a wall-mounted unit or a floor-standing unit?" can be generated.

[0132] For example, if a user queries "My KFR-35GW / N8KS1-1P air conditioner has a very poor cooling effect, and the room is still not cool even after running for half an hour, what's wrong?", the conversation context is "the product model (KFR-35GW / N8KS1-1P), problem (poor cooling effect), and usage scenario (running for half an hour)". The user has provided the product model, the problem, and the usage time, and the user's needs are clear and complete, requiring no further clarification.

[0133] It should be noted that the user's answer to the question (i.e., the new user query) will be incorporated into the session context, and the integrity check will be performed again based on the updated session context until the user's request is complete before performing step 110 above.

[0134] The customer service robot interaction method provided in this embodiment of the invention avoids blindly matching the target workflow when user needs are insufficient, thereby reducing irrelevant answers or incorrect guidance, improving the accuracy of target workflow determination, and thus improving the interaction accuracy of the customer service robot. Simultaneously, when user needs are incomplete, timely clarification questions can quickly guide users to supplement key information, reducing ineffective interaction rounds and wasted computing power. Furthermore, it transforms passive response into proactive guidance, helping users with weaker communication skills to quickly clarify their needs, improving overall service efficiency, and ultimately enhancing the user experience.

[0135] Based on any of the above embodiments, the method further includes: If it is determined based on the user query that interaction with the user is required, the step of determining the target job flow that matches the user needs indicated by the session context is executed. If it is determined that no interaction with the user is required based on the user query, the content corresponding to the user intent indicated by the user query is output.

[0136] Specifically, user queries are analyzed to identify intent and determine whether the corresponding business type exhibits characteristics of multi-turn interaction. Interaction-required scenarios typically refer to complex business scenarios that rely on user feedback, step-by-step troubleshooting, or information confirmation, such as "troubleshooting" (requiring inquiry about symptoms and operational guidance) and "product recommendation" (requiring inquiry about preferences). Non-interaction-required scenarios (direct access) refer to deterministic business scenarios where the intent is clear, requiring no further clarification or guidance, and can be completed directly by calling external tools or providing standardized forms, such as "I need repair" (directly providing a form entry point), "Check progress" (directly returning the status), and "Download manual" (directly providing a link).

[0137] It's important to note that the content outputting the user's query intent directly triggers pre-defined atomic services or external tools. For example, if the user query is "I need to report a repair," the system recognizes that no multi-turn dialogue is needed and directly calls the repair service interface to generate a pre-filled repair card or link as the response. The user can then submit the request with a single click. If the user query is "Check my package," the system directly calls the logistics API and returns the current logistics tracking card. Based on this, there's no need to match the target workflow; the content is output directly.

[0138] It should be understood that for tool-type tasks with clear intent (such as reporting repairs or making inquiries), redundant dialogue guidance steps are eliminated, and executable tools or results are provided directly, greatly improving the user's efficiency. Furthermore, for direct tasks that do not require complex semantic understanding, the expensive node detection model inference process can be bypassed, and lightweight external tools can be directly called, reducing the overall computational load and response latency, thus improving the service efficiency of the customer service robot.

[0139] The customer service robot interaction method provided in this embodiment of the invention can guide complex troubleshooting like an expert and respond to instructions quickly like a tool, enabling the customer service robot to play the dual roles of consultant and assistant, thereby improving the service flexibility and professionalism of the customer service robot and thus improving the user experience.

[0140] To facilitate understanding of the above embodiments, a specific embodiment will be described here. Let's take a consumer inquiry from a home appliance manufacturer regarding "poor air conditioning cooling performance" as an example. Specifically, the user initiates the inquiry through the home appliance brand's application, i.e., the user's query is "My KFR-35GW / N8KS1-1P air conditioner has very poor cooling performance; the room is still not cool after half an hour. What's wrong?" Then, the system identifies expressions such as "very poor" and "still not cool" in the user's query, and, combined with the question form, determines the user's emotion as "dissatisfaction," with an intensity level of 3. Level 2 (Second level of negative emotion): First, output corresponding soothing remarks (soothing remarks matching dissatisfaction: "We completely understand your frustration with using the air conditioner. Poor cooling performance does affect the experience. We'll help you troubleshoot simple issues first."; more examples of emotional soothing: "I want to complain"—can directly trigger a transfer to a human operator); construct the conversation context, recording the product model (KFR-35GW / N8KS1-1P), problem (poor cooling performance), and usage scenario (running for half an hour); then, perform a user demand completeness check on the conversation context, since the product model, problem description, and usage scenario have already been provided. Since the duration is short, the user's needs are clear and require no clarification, so the process proceeds to the SOP matching stage. Next, the core requirement "Troubleshooting poor cooling performance of KFR-35GW / N8KS1-1P air conditioner" is extracted, along with the related information "running for half an hour" and the scenario "product troubleshooting." Then, search results matching the user's query are retrieved from the knowledge base, such as data related to cooling malfunctions of the same model of air conditioner. Next, the "Air Conditioner Cooling Fault Troubleshooting SOP" is matched, with the corresponding process content being "execute according to the steps 'filter cleaning → temperature setting → fault reporting'." Finally, the search results and target process nodes are combined to generate a response message.

[0141] The interaction device of the customer service robot provided by the present invention will be described below. The interaction device of the customer service robot described below can be referred to in correspondence with the interaction method of the customer service robot described above.

[0142] Figure 3 This is a schematic diagram of the interactive device of the customer service robot provided in an embodiment of the present invention, as shown below. Figure 3 As shown, the interactive device of the customer service robot includes: The process determination module 310 is used to determine a target job process that matches the user needs indicated by the session context, based on the session context; the session context is determined based on the user query to be answered and historical interaction data belonging to the same session as the user query. The node detection module 320 is used to input the session context and the process content corresponding to the target job process into the node detection model to obtain the target process node to be executed output by the node detection model; the process content includes multiple process nodes with logical dependencies, and the node detection model is used to determine the target process node that best matches the session context from the multiple process nodes through semantic understanding capabilities. The response generation module 330 is used to generate response information for the user query based on the target process node.

[0143] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4 As shown, the electronic device may include a processor 410, a communications interface 420, a memory 430, and a communication bus 440, wherein the processor 410, communications interface 420, and memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute the interaction method of the customer service robot. This method includes: determining a target workflow matching the user's needs indicated by the conversation context based on the conversation context; the conversation context is determined based on the user query to be answered and historical interaction data belonging to the same conversation as the user query; inputting the conversation context and the workflow content corresponding to the target workflow into a node detection model to obtain the currently executed target workflow node output by the node detection model; the workflow content includes multiple workflow nodes with logical dependencies, and the node detection model is used to determine the target workflow node that best matches the conversation context from the multiple workflow nodes through semantic understanding; and generating response information for the user query based on the target workflow node.

[0144] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0145] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the interaction method of the customer service robot provided by the above methods. The method includes: determining a target workflow that matches the user needs indicated by the session context based on the session context; the session context is determined based on the user query to be answered and historical interaction data belonging to the same session as the user query; inputting the process content corresponding to the session context and the target workflow into a node detection model to obtain the target workflow node to be executed currently output by the node detection model; the process content includes multiple workflow nodes with logical dependencies, and the node detection model is used to determine the target workflow node that best matches the session context from the multiple workflow nodes through semantic understanding capabilities; and generating reply information for the user query based on the target workflow node.

[0146] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements an interaction method for a customer service robot provided by the above methods. This method includes: determining a target workflow matching the user's needs indicated by the session context based on a session context; the session context being determined based on a user query to be answered and historical interaction data belonging to the same session as the user query; inputting the session context and the workflow content corresponding to the target workflow into a node detection model to obtain a target workflow node to be executed currently output by the node detection model; the workflow content including multiple workflow nodes with logical dependencies, the node detection model being used to determine the target workflow node that best matches the session context from the multiple workflow nodes through semantic understanding capabilities; and generating response information for the user query based on the target workflow node.

[0147] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0148] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0149] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0150] The above embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Although the invention has been described in detail with reference to the embodiments, those skilled in the art should understand that various combinations, modifications, or equivalent substitutions of the technical solutions of the invention do not depart from the spirit and scope of the invention and should be covered within the scope of the claims of the invention.

Claims

1. An interaction method for a customer service robot, characterized in that, include: Based on the session context, determine the target job flow that matches the user requirements indicated by the session context; The session context is determined based on the user query to be responded to and the historical interaction data belonging to the same session as the user query; The session context and the process content corresponding to the target job process are input into the node detection model to obtain the target process node to be executed output by the node detection model. The process content includes multiple process nodes with logical dependencies. The node detection model is used to determine the target process node that best matches the session context from the multiple process nodes through semantic understanding capabilities. Based on the target process node, the response information for the user query is generated.

2. The interaction method for a customer service robot according to claim 1, characterized in that, The step of determining the target job flow that matches the user needs indicated by the session context based on the session context includes: Retrieve search results from the knowledge base that match the user's query; Based on the session context and the search results, the target job flow is determined.

3. The interaction method for a customer service robot according to claim 2, characterized in that, The step of retrieving search results from the knowledge base that match the user's query includes: Retrieve the first candidate search result with the highest similarity to the user's query and the second candidate search result with the second highest similarity to the user's query from the knowledge base; If the similarity of the first candidate search result is greater than a preset similarity threshold, or the similarity difference is greater than a preset similarity difference, the first candidate search result is determined as the search result; the similarity difference is the difference between the similarity of the first candidate search result and the similarity of the second candidate search result. If the similarity of the first candidate search result is less than or equal to a preset similarity threshold, and the similarity difference is less than or equal to a preset similarity difference, the first candidate search result, the second candidate search result, and the user query are input into the classification search model to obtain the search result output by the classification search model; the classification search model is used to determine the search result most similar to the user query from the first candidate search result and the second candidate search result.

4. The interaction method for a customer service robot according to claim 3, characterized in that, The step of retrieving the first candidate search result with the highest similarity to the user's query and the second candidate search result with the second highest similarity to the user's query from the knowledge base includes: The user query is rewritten to obtain multiple alternative queries; The similarity between the multiple candidate queries and each knowledge segment in the knowledge base is calculated to obtain multiple similarity calculation results; Based on the multiple similarity calculation results, the first candidate search result and the second candidate search result are retrieved from the knowledge base.

5. The interaction method for a customer service robot according to claim 2, characterized in that, The step of determining the target job flow based on the session context and the search results includes: Obtain the user profile information of the user who inputs the query; The target job process is determined based on the session context, the search results, and the user profile information.

6. The interaction method for a customer service robot according to claim 1, characterized in that, The process content is determined based on the following method: The target workflow is input into the content generation model to obtain the workflow content output by the content generation model; the content generation model is trained based on the sample workflow. or, The process content corresponding to the target operation process is determined from the template library; each process content template in the template library is generated by the content generation model.

7. The interaction method for a customer service robot according to claim 1, characterized in that, The step of generating the response information for the user query based on the target process node includes: The target process node and the search results are input into the large language model to obtain the anthropomorphic response information output by the large language model; the search results are the results retrieved from the knowledge base that match the user's query.

8. The interaction method of the customer service robot according to any one of claims 1 to 7, characterized in that, Also includes: Emotion recognition is performed on the user query to obtain the emotion recognition result; If the user's negative emotion is determined to be of the first level of negative emotion based on the emotion recognition result, the step of determining the target job process that matches the user's needs indicated by the conversation context is executed. If the user's negative emotion is determined to be of the second level of negative emotion based on the emotion recognition result, a reassuring message is output, and the step of determining the target operation process that matches the user's needs indicated by the conversation context is executed. If the user's negative emotion is determined to be at the third level of negative emotion based on the emotion recognition results, the human customer service interaction process is triggered. Wherein, the degree of the first negative emotion is less than the degree of the second negative emotion, and the degree of the second negative emotion is less than the degree of the third negative emotion.

9. The interaction method of the customer service robot according to any one of claims 1 to 7, characterized in that, Also includes: If the user requirement is determined to be complete based on the session context, the step of determining the target job flow that matches the user requirement indicated by the session context is executed. If the user's request is determined to be incomplete based on the session context, the question content is output based on the session context to obtain a new user query input by the user based on the question content.

10. The interaction method of the customer service robot according to any one of claims 1 to 7, characterized in that, Also includes: If it is determined based on the user query that interaction with the user is required, the step of determining the target job flow that matches the user needs indicated by the session context is executed. If it is determined that no interaction with the user is required based on the user query, the content corresponding to the user intent indicated by the user query is output.

11. An interactive device for a customer service robot, characterized in that, include: The process determination module is used to determine, based on the session context, a target job process that matches the user requirements indicated by the session context; The session context is determined based on the user query to be responded to and the historical interaction data belonging to the same session as the user query; The node detection module is used to input the session context and the process content corresponding to the target job process into the node detection model to obtain the target process node to be executed output by the node detection model. The process content includes multiple process nodes with logical dependencies. The node detection model is used to determine the target process node that best matches the session context from the multiple process nodes through semantic understanding capabilities. The response generation module is used to generate response information for the user query based on the target process node.

12. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the interaction method of the customer service robot as described in any one of claims 1 to 10.

13. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the interaction method of the customer service robot as described in any one of claims 1 to 10.

14. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the interaction method of the customer service robot as described in any one of claims 1 to 10.

Images