Intelligent customer service question and answer processing method and device, computer device and storage medium

Abstract

The application relates to an intelligent customer service question and answer processing method and device, computer equipment and a storage medium. The method comprises the following steps: receiving text information of a user question, and obtaining a tenant identifier; performing intent recognition based on the text information to obtain an intent label and an intent recognition score; if the intent recognition score is greater than a preset threshold value and the intent label is a static business rule type label, obtaining matched target business text information, and generating response content; if the intent recognition score is greater than the preset threshold value and the intent label is a dynamic business data type label, matching a target intelligent agent tool, obtaining business data, and generating response content; and if the intent recognition score is less than or equal to the preset threshold value, performing feature extraction and vector conversion processing on the text information to obtain a user semantic feature vector, screening out a target semantic feature vector, obtaining target business text information, and generating response content. The method can improve the accuracy and response efficiency of a customer service system in responding to user questions.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, and in particular to an intelligent customer service question-and-answer processing method, apparatus, computer equipment, and storage medium. Background Technology

[0002] In the ride-hailing industry, customer service systems are a crucial bridge connecting the platform with users (including passengers and drivers). Traditional customer service systems are generally rule-based, which, while capable of responding to preset questions through keyword matching and achieving basic answering functions, have significant technical limitations. For example, they are strict in their requirements for user question wording, unable to accommodate synonyms, near-synonyms, and colloquial expressions, easily leading to incorrect or ineffective responses; they struggle to handle complex issues involving multiple business processes, lacking semantic understanding and logical reasoning capabilities; and with business expansion, the workload of maintaining and updating the rule base is enormous, and rule conflicts are prone to occur. Consequently, with the continuous expansion of business scale, the number of user inquiries and complaints has exploded, and the traditional rule-based customer service model can no longer meet users' core needs for efficient, accurate, and timely service, impacting the user's consultation experience.

[0003] To overcome the shortcomings of rule-based customer service systems, machine learning-based intelligent customer service systems have evolved. While these systems improve efficiency and accuracy to some extent, their performance is limited by the quality and quantity of data used to train the machine learning models. If the training data is biased or insufficient, the performance of the machine learning models will be severely affected. For example, for some rare or complex questions, because they appear infrequently in the training data, the machine learning models struggle to accurately identify and handle them. Furthermore, machine learning models lack contextual understanding capabilities, failing to accurately understand the user's intent based on previous conversations in multi-turn dialogues. This results in inconsistent and unfocused responses, negatively impacting the user's consultation experience. In addition, existing customer service systems often employ multi-tenant architectures, lacking flexibility in model deployment and updates. When model updates or optimizations are needed, separate configuration and adjustments may be required for each tenant's system, increasing operational and time costs. Summary of the Invention

[0004] Therefore, it is necessary to provide an intelligent customer service question-and-answer processing method, device, computer equipment, and storage medium to address the aforementioned technical issues. This would improve the accuracy and response efficiency of customer service systems in answering user questions, thereby enhancing the user's consultation experience and reducing the operation and maintenance costs of multi-tenant architectures.

[0005] According to a first aspect of certain exemplary embodiments of this disclosure, an intelligent customer service question-and-answer processing method is provided, comprising: receiving text information of a user question input by a user, and obtaining a tenant identifier corresponding to the user; performing intent recognition based on the text information to obtain an intent tag and intent recognition score corresponding to the text information; if the intent recognition score is greater than a preset threshold, and the intent tag is a static business rule type tag, then obtaining matching target business text information from the knowledge base corresponding to the tenant identifier based on the intent tag, and generating response content for the user question based on the target business text information; if the intent recognition score is greater than a preset threshold, and the intent tag is a dynamic business data type tag, then obtaining matching target business text information from the knowledge base corresponding to the tenant identifier based on the intent tag. The system uses a centralized intelligent agent tool to match target intelligent agent tools, calls the target intelligent agent tools to obtain business data that matches the intent label, and generates the response content for the user's question based on the business data. If the intent recognition score is less than or equal to a preset threshold, feature extraction and vector transformation are performed on the text information of the user's question to obtain the user's semantic feature vector. Multiple semantic feature vectors are obtained from the knowledge base corresponding to the tenant identifier, and each semantic feature vector is matched with the user's semantic feature vector. Based on the matching results, the target semantic feature vector is selected, and the target business text information of the target semantic feature vector is obtained from the knowledge base corresponding to the tenant identifier. The response content for the user's question is generated based on the target business text information.

[0006] According to certain exemplary embodiments of this disclosure, the knowledge base obtains multiple semantic feature vectors by performing feature extraction and vector transformation on multiple pieces of business text information, establishes the association between each semantic feature vector and the corresponding business text information, and stores the multiple pieces of business text information, multiple semantic feature vectors and the association relationship in the vector library of the corresponding tenant identifier.

[0007] According to certain exemplary embodiments of this disclosure, obtaining matching target business text information from a knowledge base corresponding to a tenant identifier based on an intent tag includes: performing feature extraction and vector transformation processing on the intent tag to obtain an intent semantic feature vector; identifying the knowledge base corresponding to the tenant identifier based on the tenant identifier, and obtaining multiple semantic feature vectors from the knowledge base corresponding to the tenant identifier; calculating a first similarity value between the intent semantic feature vector and each semantic feature vector; determining the semantic feature vector whose first similarity value satisfies a first preset matching condition as the target semantic feature vector; and querying the business text information corresponding to the target semantic feature vector based on the association relationship stored in the knowledge base corresponding to the tenant identifier, wherein the target business text information includes the business text information corresponding to the target semantic feature vector.

[0008] According to certain exemplary embodiments of this disclosure, each intelligent agent tool in the intelligent agent toolset is associated with one or more preset intent tags. Matching a target intelligent agent tool from the intelligent agent toolset corresponding to a tenant identifier based on the intent tags includes: identifying the intelligent agent toolset corresponding to the tenant identifier based on the tenant identifier; obtaining one or more preset intent tags associated with each intelligent agent tool in the intelligent agent toolset corresponding to the tenant identifier; obtaining a second similarity value between the intent tags and each preset intent tag; determining the preset intent tag whose second similarity value satisfies a second preset matching condition as the target intent tag; and determining the intelligent agent tool associated with the target intent tag as the target intelligent agent tool.

[0009] According to certain exemplary embodiments of this disclosure, invoking a target agent tool to obtain business data matching an intent label includes: sending a tool invocation request for the target agent tool to the agent toolset corresponding to the tenant identifier through a multi-tenant agent; and in response to the tool invocation request, invoking the target agent tool to obtain target business data matching the business data type, wherein the business data matching the intent label includes the target business data.

[0010] According to certain exemplary embodiments of this disclosure, each semantic feature vector is matched with a user semantic feature vector, and a target semantic feature vector is selected based on the matching results. The target business text information of the target semantic feature vector is obtained from the knowledge base corresponding to the tenant identifier. This includes: calculating a third similarity value between each semantic feature vector and the user semantic feature vector; determining the semantic feature vector whose third similarity value satisfies a third preset matching condition as the target semantic feature vector; and querying the business text information corresponding to the target semantic feature vector based on the association relationship stored in the knowledge base corresponding to the tenant identifier. The target business text information includes the business text information corresponding to the target semantic feature vector.

[0011] According to certain exemplary embodiments of this disclosure, an intelligent customer service question-and-answer processing method further includes: if the intent recognition score is less than or equal to a preset threshold and the number of selected target semantic feature vectors is zero, then triggering a transfer step with human customer service; during the service process of human customer service, collecting the conversation text information between the user and human customer service in real time, and performing feature extraction and vector transformation processing on the conversation text information to obtain supplementary semantic feature vectors; matching the supplementary semantic feature vectors with each semantic feature vector in the knowledge base corresponding to the tenant identifier, and selecting the matching reference semantic feature vectors based on the matching results; obtaining the reference business text information corresponding to the reference semantic feature vectors from the knowledge base corresponding to the tenant identifier, and pushing the reference business text information to human customer service to assist human customer service in answering user questions.

[0012] According to certain exemplary embodiments of this disclosure, an intelligent customer service question-and-answer processing method further includes: collecting historical conversation text information of human customer service corresponding to a target tenant identifier within a historical time period of a preset duration; filtering out new conversation text information with a consultation frequency higher than a preset frequency threshold from the historical conversation text information; performing intent recognition on the new conversation text information to identify new intent tags corresponding to the new conversation text information; performing feature extraction and vector transformation processing on the new conversation text information to obtain new semantic feature vectors; establishing new association relationships between the new semantic feature vectors, new conversation text information, and new intent tags; and storing the new semantic feature vectors, new conversation text information, new intent tags, and new association relationships in a knowledge base corresponding to the target tenant identifier.

[0013] According to a second aspect of certain exemplary embodiments of this disclosure, an intelligent customer service question-and-answer processing apparatus is provided, characterized in that the apparatus includes: a first acquisition module, configured to receive text information of a user question input by a user and acquire a tenant identifier corresponding to the user; a first intent recognition module, configured to perform intent recognition based on the text information to obtain an intent tag and intent recognition score corresponding to the text information; a first generation module, configured to, if the intent recognition score is greater than a preset threshold and the intent tag is a static business rule type tag, then obtain matching target business text information from the knowledge base corresponding to the tenant identifier based on the intent tag, and generate response content for the user question based on the target business text information; and a second generation module, configured to, if the intent recognition score is greater than a preset threshold and the intent tag is a dynamic business rule type tag, then obtain matching target business text information from the knowledge base corresponding to the tenant identifier ... and generate response content for the user question; and a second generation module, configured to, if the intent recognition score is greater than a preset threshold and the intent tag is a dynamic business rule type tag, then obtain matching target business text information from the knowledge base corresponding to the tenant identifier and generate response content for the user question; and a second generation module, configured to, if the intent recognition score is greater than a preset threshold and the intent tag is a dynamic business rule type tag, then obtain matching target business text information from the knowledge base corresponding to the tenant identifier and generate response content for the user question; and a second generation module, configured to, if the intent recognition score is greater than a preset threshold and the intent tag is a dynamic business rule type tag, then obtain matching target business text information from the knowledge base corresponding to the tenant identifier Based on the class label, the system matches the target intelligent agent tool from the intelligent agent tool set corresponding to the tenant identifier according to the intent label, calls the target intelligent agent tool to obtain business data that matches the intent label, and generates the response content of the user question based on the business data; the third generation module is used to perform feature extraction and vector transformation processing on the text information of the user question if the intent recognition score is less than or equal to a preset threshold, to obtain the user semantic feature vector, obtain multiple semantic feature vectors from the knowledge base corresponding to the tenant identifier, match each semantic feature vector with the user semantic feature vector, filter out the target semantic feature vector based on the matching result, obtain the target business text information of the target semantic feature vector from the knowledge base corresponding to the tenant identifier, and generate the response content of the user question based on the target business text information.

[0014] According to a third aspect of certain exemplary embodiments of the present disclosure, a computer device is provided, including 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 steps of any of the above methods.

[0015] According to a fourth aspect of certain exemplary embodiments of the present disclosure, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0016] The aforementioned intelligent customer service question-and-answer processing method, apparatus, computer equipment, and storage medium receive text information of user questions input by the user and obtain the tenant identifier corresponding to the user; perform intent recognition based on the text information to obtain the intent tag and intent recognition score corresponding to the text information; if the intent recognition score is greater than a preset threshold and the intent tag is a static business rule type tag, then obtain matching target business text information from the knowledge base corresponding to the tenant identifier based on the intent tag, and generate the response content for the user question based on the target business text information; if the intent recognition score is greater than a preset threshold and the intent tag is a dynamic business data type tag, then obtain the response content for the user question from the intelligent agent corresponding to the tenant identifier based on the intent tag. The toolset matches the target intelligent agent tool, which is then invoked to obtain business data that matches the intent tag. Based on the business data, the response content for the user's question is generated. If the intent recognition score is less than or equal to a preset threshold, feature extraction and vector transformation are performed on the text information of the user's question to obtain the user's semantic feature vector. Multiple semantic feature vectors are obtained from the knowledge base corresponding to the tenant identifier. Each semantic feature vector is matched with the user's semantic feature vector. Based on the matching results, the target semantic feature vector is selected. The target business text information of the target semantic feature vector is obtained from the knowledge base corresponding to the tenant identifier. Based on the target business text information, the response content for the user's question is generated.

[0017] Therefore, this application's solution first performs intent recognition on the text information of user-input questions, identifying the corresponding intent tags and intent recognition scores. Then, based on the intent recognition scores, it determines the clarity of the intent and distinguishes tag types in conjunction with the intent tags. This facilitates the precise matching of differentiated response processing logic for different types of user questions, fundamentally solving the technical shortcomings of traditional customer service systems that offer only a single response method and cannot specifically handle various types of questions. On one hand, for user questions with clear intent (i.e., intent recognition scores greater than a preset threshold), the response path is further precisely adapted based on the tag type. Specifically, for static business rule tags with standardized and fixed answers, the matching target business text information is directly retrieved from the corresponding tenant's dedicated knowledge base. Based on this target business text information, standardized and fixed response content for the user question can be generated. For dynamic business data tags that require real-time data acquisition and dynamically changing results, real-time and accurate business data is obtained by calling the target tools in the corresponding tenant's intelligent agent toolset. Then, response content tailored to the actual business scenario is generated based on the acquired business data. This approach, which targets scenarios with clear intent and differentiates responses based on demand categories, ensures both efficiency in handling standardized user questions and real-time accuracy in responding to user questions requiring dynamic business data. Compared to traditional rule-based customer service systems, which require manual keyword matching and cannot access real-time business data, and machine learning-based customer service systems, which suffer from lag and insufficient accuracy in dynamic responses, this approach effectively avoids response bias and delays, improving the accuracy and efficiency of customer service system responses to user questions and enhancing the user's consultation experience.

[0018] On the other hand, for user questions with ambiguous intent (i.e., intent recognition scores less than or equal to a preset threshold), the user question is first transformed into a semantic feature vector through feature extraction and vector transformation. Then, it is matched with the semantic feature vector in the tenant's exclusive knowledge base for similarity. Based on the matching results, the target business text information is selected. Finally, based on the target business text, accurate and relevant response content is generated to meet the user's real needs. This method of matching directly based on the semantic feature vector of the user question breaks through the limitation of traditional customer service systems that cannot effectively respond to ambiguous intent questions. It can accurately capture the core semantics of user questions and achieve effective compatibility with diverse, colloquial, and ambiguous user questions. It significantly improves the efficiency of customer service systems in adapting to various user questions and the comprehensiveness of problem handling. It effectively solves the technical problems of traditional rule-based customer service systems that have strict requirements for user question expression and are prone to answer errors or failures. At the same time, it makes up for the technical defects of machine learning-based customer service systems that lack contextual understanding, have inconsistent responses in multi-turn dialogues, and lack targeting. It can continuously capture the core needs of users in multi-turn ambiguous dialogues, maintain the consistency and targeting of responses, and further improve the user consultation experience.

[0019] Furthermore, this application's solution invokes the corresponding dedicated knowledge base or intelligent agent tool through tenant identifiers, which not only achieves the isolation and reuse of knowledge and resources among multiple tenants, ensuring the differentiated service needs of different tenants, but also leverages a large model base shared by multiple tenants. When updating or optimizing the model, there is no need to adjust the system configuration of each tenant separately. It can be quickly synchronized to each tenant simply by completing the iteration based on a unified base. This effectively solves the problems of low flexibility in updating existing multi-tenant architecture models and high operation and maintenance costs, and significantly reduces the overall operation and maintenance costs and time costs of multi-tenant architecture customer service systems. Attached Figure Description

[0020] Figure 1 This is a schematic diagram illustrating the application environment of an intelligent customer service question-and-answer processing method as one of the exemplary embodiments of this disclosure. Figure 2 This is a flowchart illustrating an intelligent customer service question-and-answer processing method as one of the exemplary embodiments of this disclosure; Figure 3 This is a flowchart illustrating a method for retrieving matching target business text information from a knowledge base corresponding to a tenant identifier based on intent tags, as described in some exemplary embodiments of this disclosure. Figure 4 This is a flowchart illustrating a method for matching a target agent tool from an agent toolset corresponding to a tenant identifier based on an intent tag, as shown in some exemplary embodiments of this disclosure. Figure 5 This is a flowchart illustrating a method for invoking a target intelligent agent tool to obtain business data matching an intent tag, as shown in some exemplary embodiments of this disclosure. Figure 6 This is a flowchart illustrating a method for obtaining target business text information of a target semantic feature vector in some exemplary embodiments of this disclosure; Figure 7 This is a flowchart illustrating a human customer service transfer method in some exemplary embodiments of this disclosure; Figure 8 This is a flowchart illustrating a knowledge base self-updating method based on human conversation in some exemplary embodiments of this disclosure; Figure 9 This is a flowchart illustrating an example of an intelligent customer service question-and-answer processing method as shown in some exemplary embodiments of this disclosure; Figure 10 This is a structural block diagram of an intelligent customer service question-and-answer processing device in some other exemplary embodiments of this disclosure; Figure 11 This is a diagram illustrating the internal structure of a computer device in some other exemplary embodiments of this disclosure. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0022] The following detailed descriptions are provided to aid the reader in gaining a comprehensive understanding of the methods, apparatus, computer devices, storage media, and / or computer program products described herein. However, upon understanding the disclosure of this disclosure, various changes, modifications, and equivalents of the methods, apparatus, storage media, and / or computer program products described herein will become apparent. For example, the order of operations described herein is merely illustrative and is not limited to those orders set forth herein, but may be changed as will become clear upon understanding the disclosure of this disclosure, except for operations that must occur in a specific order. Furthermore, for clarity and conciseness, descriptions of features known in the art may be omitted.

[0023] The features described herein may be implemented in different forms and should not be construed as limited to the examples described herein. Rather, the examples described herein are provided only to illustrate some of the many feasible ways of implementing the methods, computer devices, and / or storage media described herein, many of which will become clear upon understanding this disclosure.

[0024] The terminology used herein is for the purpose of describing various examples only and is not intended to limit disclosure. Unless the context clearly indicates otherwise, the singular form is intended to include the plural form as well. The terms “comprising,” “including,” and “having” indicate the presence of the described features, quantities, operations, components, elements, and / or combinations thereof, but do not exclude the presence or addition of one or more other features, quantities, operations, components, elements, and / or combinations thereof. Unless otherwise stated, “ / ” means “or,” for example, A / B can mean A or B; “and / or” in the text is merely a description of the relationship between related objects, indicating that three relationships can exist, for example, A and / or B can mean: A alone, A and B simultaneously, and B alone. Furthermore, in the description of embodiments of the invention, “multiple” means two or more.

[0025] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains upon understanding this disclosure. Unless expressly defined herein, terms (such as those defined in a general dictionary) shall be interpreted as having a meaning consistent with their meaning in the context of the relevant field and in this disclosure, and shall not be interpreted in an idealized or overly formalistic manner.

[0026] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in some of the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0027] Furthermore, in the description of the examples, detailed descriptions of well-known related structures or functions will be omitted when it is believed that such detailed descriptions would lead to a vague interpretation of this disclosure.

[0028] In the following description, embodiments will be described in detail with reference to the accompanying drawings. However, embodiments may be implemented in various forms and are not limited to the examples described herein.

[0029] First, the specific technical problem that this application aims to solve is stated. The existing technologies closest to this technical solution mainly include rule-based customer service systems, machine learning-based intelligent customer service systems, and multi-tenant architecture intelligent customer service systems. All three types of customer service systems have technical shortcomings in practical applications, as detailed below: I. Rule-Based Customer Service System: This system responds to user inquiries using a keyword matching algorithm based on a pre-defined mapping relationship between keywords and response templates. After receiving user input, the system scans the text for pre-defined keywords, triggering the corresponding fixed response template to return the result. This rule-based customer service system has the following technical issues: 1. Poor expression adaptability: It only supports fixed keyword matching and cannot adapt to colloquial language, synonyms / near-synonyms and vague questions, which easily leads to incorrect or invalid responses; 2. Lack of ability to handle complex problems: unable to perform semantic understanding and logical reasoning on complex inquiries involving multiple business processes, making it difficult to provide comprehensive answers; 3. High maintenance costs: Business expansion and changes in demand require manual addition and modification of rules, which can easily lead to conflicts and redundancy in the rule base, resulting in a large workload for maintenance.

[0030] II. Machine Learning-Based Intelligent Customer Service System: This system collects labeled historical dialogue data and trains a machine learning model using algorithms such as Support Vector Machines and Naive Bayes. The machine learning model is then used to achieve intent recognition and automatic response. This machine learning-based intelligent customer service system faces the following technical challenges: 1. Strong data dependence: The model performance depends on the quality and quantity of training data. Bias in training data, incomplete coverage, rare problems that occur infrequently, and new business inquiries can easily lead to the inability to accurately identify intent and provide effective responses. 2. Lack of contextual understanding: In multi-turn conversations, they are unable to connect information from previous contexts, and their responses lack coherence and relevance; 3. Long iteration cycle: When updating business, a large amount of data needs to be re-labeled to train the model, which cannot quickly adapt to business changes.

[0031] III. Multi-tenant Architecture Intelligent Customer Service System: By sharing core resources and configuring independent business data, multi-tenant resource sharing and isolation are achieved, meeting the differentiated needs of ride-hailing services across multiple regions and business lines. Current multi-tenant architecture intelligent customer service systems have the following technical problems: 1. Insufficient update flexibility: When upgrading or optimizing a large model, individual configuration adjustments are required for each tenant, making it impossible to achieve synchronous updates for all tenants. 2. High operation and maintenance costs: Adapting to each tenant individually increases the workload and cycle of operation and maintenance, and is prone to configuration errors and version inconsistencies.

[0032] The abbreviations and key terms in this disclosure are explained as follows: 1. Large Models (LM): These typically refer to deep learning models with a huge number of parameters (usually 1 billion+), which can cover multimodal tasks.

[0033] 2. Retrieval-Augmented Generation (RAG): This is a technical framework that combines information retrieval with generative artificial intelligence (AI). Its core objective is to improve the accuracy, reliability, and relevance of the generative model's output by incorporating external knowledge. The RAG model first retrieves information relevant to the user's query from a pre-defined external knowledge base, then uses the retrieved valid knowledge as contextual input to assist the model in generating answers. This effectively avoids the "illusion" problem caused by knowledge lag and memory bias in the generative model, making the output more targeted and credible.

[0034] 3. Embedding: Embedding is a technique that transforms discrete data (such as text, words, images, audio, etc.) into continuous vectors (high-dimensional numerical arrays). Its core goal is to map abstract, unstructured information into low- or high-dimensional mathematical spaces, enabling computers to understand and process this data. Embedding models (i.e., models used to implement embedding techniques) learn the semantic features and relationships of the data, transforming different forms of discrete data into computable and comparable vector forms. The distance between vectors corresponds to the semantic similarity of the original data, providing core support for subsequent tasks such as information retrieval, semantic matching, and cluster analysis. It is also a key technological foundation for knowledge base retrieval and matching in RAG models.

[0035] 4. Intelligent Agent: This refers to an intelligent system developed based on a large model, capable of autonomously perceiving its environment, making decisions, and executing actions to achieve specific goals. It can be a software program, a robot, or other form of intelligent entity, with its core characteristics being autonomy and goal orientation.

[0036] First, the application environment of the intelligent customer service question-and-answer processing method provided in this application is introduced. For example, taking the application of an intelligent customer service question-and-answer processing method in the ride-hailing industry as an example, such as... Figure 1 As shown, the intelligent customer service question-and-answer processing method provided in this application can be applied to, for example... Figure 1In the application environment shown, the intelligent customer service question-and-answer processing system 100 is used to execute an intelligent customer service question-and-answer processing method according to this application. For example, the intelligent customer service question-and-answer processing system 100 can be used to receive and process text information of various user questions input by users (such as drivers and / or passengers), and interact with the user's terminal device 102 via a network. The user's terminal device 102 can be a mobile terminal, such as a mobile phone or tablet computer. Specifically, the intelligent customer service question-and-answer processing system 100 receives text information of user questions input by the user from the terminal device 102 and obtains the tenant identifier corresponding to the user; performs intent recognition based on the text information to obtain the intent tag and intent recognition score corresponding to the text information; if the intent recognition score is greater than a preset threshold and the intent tag is a static business rule type tag, then the system obtains matching target business text information from the knowledge base corresponding to the tenant identifier based on the intent tag, and generates the response content for the user question based on the target business text information; if the intent recognition score is greater than a preset threshold and the intent tag is a dynamic business data type tag, then the system obtains the response content for the user question from the intelligent entity corresponding to the tenant identifier based on the intent tag. The system has a centralized target intelligent agent tool. It calls the target intelligent agent tool to obtain business data that matches the intent label, and generates the response content for the user's question based on the business data. If the intent recognition score is less than or equal to a preset threshold, the text information of the user's question is processed by feature extraction and vector transformation to obtain the user semantic feature vector. Multiple semantic feature vectors are obtained from the knowledge base corresponding to the tenant identifier. Each semantic feature vector is matched with the user semantic feature vector. The target semantic feature vector is selected based on the matching result. The target business text information of the target semantic feature vector is obtained from the knowledge base corresponding to the tenant identifier. The response content for the user's question is generated based on the target business text information.

[0037] In one embodiment, such as Figure 2 As shown, an intelligent customer service question-and-answer processing method is provided, which can be applied to... Figure 1 Taking the intelligent customer service question and answer processing system 100 in China as an example, the following steps are included: Step 201: Receive the text information of the user's question input by the user, and obtain the tenant identifier corresponding to the user.

[0038] Specifically, users initiate inquiries through the driver / passenger app or backend terminal. The intelligent customer service Q&A processing system 100 receives user questions in the form of spoken text or voice. At the same time, it automatically identifies and obtains the user's corresponding tenant identifier through the user's login account, device binding information, or business affiliation, clarifying the user's business area, business line, or partner tenant, providing a basis for subsequent tenant-specific resource calls.

[0039] Step 202: Perform intent recognition based on text information to obtain the intent label and intent recognition score corresponding to the text information.

[0040] Specifically, the intelligent customer service question-and-answer processing system 100 performs intent recognition on the text information of received user questions, identifying the intent tags and intent recognition scores corresponding to the text information. For example, the intelligent customer service question-and-answer processing system 100 can perform intent recognition using a pre-trained intent recognition model. This intent recognition model can be built based on a large model (LM) architecture, obtained through pre-training on a dataset containing multiple sets of historical dialogue data and the corresponding intent tags (such as "order billing consultation," "order status consultation," "material consultation," etc.). By using a precisely labeled dataset covering all business scenarios of the tenant and combining it with the contextual understanding and semantic reasoning capabilities of the large model, the system performs targeted training and optimization to improve the large model's adaptability to tenant-specific business terminology, consultation scenarios, and user expression habits. Simultaneously, data augmentation techniques (such as synonym replacement, sentence transformation, and colloquial rewriting) are introduced to expand the dataset coverage, enabling the trained intent recognition model to possess accurate and comprehensive semantic understanding capabilities, accurately parsing users' colloquial expressions, synonyms / The model identifies the core needs expressed through synonyms and vague questions, outputting intent tags strongly relevant to the tenant's business scenario (such as "unable to contact passengers," "cancel order," "registration material consultation," etc.). Simultaneously, this intent recognition model calculates the intent matching degree between the user's question text and each preset intent tag, generating an intent recognition score from 0 to 100. A higher intent recognition score indicates a higher degree of matching between the user's intent and the intent tag, and a clearer intent; conversely, a lower score indicates a more vague user intent. This intent recognition score provides a precise quantitative basis for selecting differentiated response paths based on the clarity of the intent, ensuring the adaptability and rationality of subsequent response logic, and improving the accuracy of customer service responses from the source.

[0041] Step 203: If the intent recognition score is greater than the preset threshold and the intent tag is a static business rule type tag, then the matching target business text information is obtained from the knowledge base corresponding to the tenant identifier based on the intent tag, and the response content of the user question is generated based on the target business text information.

[0042] Specifically, a preset threshold is used to determine whether the intent expressed by the user in their question is clear. This preset threshold is adapted to the quantitative standard of intent recognition score (e.g., 0-100 points) in step 202. For example, the preset threshold can be set to 80 points. When the intent recognition score is higher than the preset threshold (e.g., 80 points), the intelligent customer service question-and-answer processing system 100 can determine that the intent expressed by the user in their question is clear. At the same time, combined with the intent tag identified in step 202, if the intent tag is identified as a static business rule tag, it means that the user is inquiring about fixed and unchanging business rules or common questions that do not require real-time calculation. Then, the intelligent customer service question-and-answer processing system 100 accurately locates the tenant's exclusive knowledge base based on the previously obtained tenant identifier. This knowledge base can be entered by operations or knowledge engineers through the multi-tenant backend page and stores all static fixed business knowledge exclusive to the tenant, including but not limited to tenant business rules, frequently asked questions (FAQ), service specifications, and other fixed static knowledge. This ensures that the response content is highly adapted to the tenant's business scenario and avoids confusion with the business knowledge of other tenants. Furthermore, the intelligent customer service question and answer processing system 100 uses the identified intent tags as the retrieval basis to search and obtain the target business text information that matches the intent tags in the knowledge base corresponding to the tenant identifier. Finally, based on the retrieved standardized and fixed target business text information, the intelligent customer service question and answer processing system 100 generates fixed response content that conforms to the tenant's business specifications and is accurately and uniformly expressed. This ensures both the efficient response of the intelligent customer service question and answer processing system 100 to standardized inquiries and the accuracy and standardization of the response content.

[0043] Step 204: If the intent recognition score is greater than the preset threshold and the intent label is a dynamic business data type label, then match the target intelligent agent tool from the intelligent agent tool set corresponding to the tenant identifier according to the intent label, call the target intelligent agent tool to obtain the business data that matches the intent label, and generate the response content of the user question based on the business data.

[0044] Specifically, when the intent recognition score exceeds a preset threshold (e.g., 80 points), and the intent tag is determined to be a dynamic business data type tag, it indicates that the user's question requires real-time updated business data to provide an accurate answer. In this case, it is necessary to use intelligent agent tools to acquire and process real-time data. First, the intelligent customer service question-and-answer processing system 100 will locate the tenant's exclusive intelligent agent toolset based on the user's corresponding tenant identifier. This toolset is configured by the tenant according to its own business scenario and includes various callable business tools, such as real-time order query tools, account balance statistics tools, business progress tracking tools, and data statistical analysis tools. Each tool is bound to a specific business intent tag to ensure the targeted nature of tool calls. Then, the intelligent customer service question-and-answer processing system 100 uses the intent tag as the matching basis to filter out the target intelligent agent tool that matches the current intent tag from the tenant's exclusive intelligent agent toolset. For example, if the user's question intent tag is "real-time query of current order progress," then the corresponding "real-time order progress query tool" will be matched; if the intent tag is "statistics on today's total account transaction amount," then the corresponding "account transaction data statistics tool" will be matched. To ensure the accuracy of tool matching, the intelligent customer service Q&A processing system 100 will further verify the binding relationship between intent tags and tools, and rule out situations such as binding anomalies or tool unavailability. If a tool is not matched or the matching fails, a preset fault tolerance mechanism will be triggered, which will promptly report to the system backend and prompt the tenant's operation and maintenance personnel to check the tool configuration. At the same time, the user will be given a temporary response of "The current query function is not ready yet, please try again later" to avoid the user waiting for timeout.

[0045] Furthermore, after tool matching is complete, the intelligent customer service question-and-answer processing system 100 sends a tool invocation request to the target intelligent agent tool to retrieve business data matching the intent tag. Based on the retrieved business data and combined with the user's question description, the intelligent customer service question-and-answer processing system 100 transforms the retrieved business data into natural language responses that fit the user's needs, thereby achieving accurate responses to such user questions. This method of retrieving real-time business data and generating dynamic responses through intelligent agent tool invocation avoids the technical shortcomings of traditional customer service systems, such as the inability to retrieve real-time data and the lag in dynamic responses. It ensures the real-time, accurate, and professional nature of responses to dynamic business data-based questions, further improving the accuracy of customer service system responses, meeting users' needs for real-time business consultation, and enhancing the user's consultation experience.

[0046] Step 205: If the intent recognition score is less than or equal to a preset threshold, feature extraction and vector transformation are performed on the text information of the user's question to obtain the user's semantic feature vector. Multiple semantic feature vectors are obtained from the knowledge base corresponding to the tenant identifier. Each semantic feature vector is matched with the user's semantic feature vector. The target semantic feature vector is selected based on the matching result. The target business text information of the target semantic feature vector is obtained from the knowledge base corresponding to the tenant identifier. The response content of the user's question is generated based on the target business text information.

[0047] Specifically, when the intent recognition score is less than or equal to a preset threshold (e.g., 80 points), it indicates that the user's question intent is ambiguous and the response path cannot be directly located through intent tags. In this case, semantic feature matching is needed to capture the core needs of the user's question. First, the intelligent customer service question-and-answer processing system 100 performs feature extraction and vector transformation on the user's input question text information, extracting the core semantic features from the text, removing irrelevant and redundant information, and converting the processed text information into user semantic feature vectors that can be used for similarity matching, ensuring that the vectors can accurately represent the core semantics of the user's question. Then, based on the user's corresponding tenant identifier, the intelligent customer service question-and-answer processing system 100 locates the tenant's exclusive knowledge base, retrieves multiple semantic feature vectors that have undergone pre-feature extraction and vector transformation from the knowledge base, and performs similarity matching between the transformed user semantic feature vectors and each semantic feature vector obtained from the knowledge base, calculating the semantic similarity value between the two. The higher the calculated similarity value, the more closely the corresponding business text information matches the core needs of the user's question. Furthermore, the intelligent customer service question-and-answer processing system 100, based on the matching results (i.e., the similarity value), selects semantic feature vectors whose similarity values ​​meet preset similarity conditions as target semantic feature vectors. Further, the intelligent customer service question-and-answer processing system 100, according to the association relationship between the semantic feature vectors stored in the knowledge base corresponding to the tenant identifier and the business text information, retrieves the target business text information corresponding to the target semantic feature vector from the knowledge base corresponding to the tenant identifier. Then, based on the target business text information and the ambiguous expression scenario and colloquial tone of the user's question, it appropriately optimizes and adjusts the information to generate a response that is tailored to the user's needs and is clearly and easily understood. This achieves accurate responses to user questions with ambiguous intent, effectively solving the shortcomings of traditional customer service systems in efficiently responding to questions with ambiguous intent, and overcoming the technical limitations of traditional rule-based customer service systems that have strict requirements on the expression of user questions and are prone to answering errors or failures. This improves the adaptability and accuracy of the customer service system to various types of user questions, enhancing the user consultation experience. Meanwhile, in multi-round fuzzy dialogue scenarios, the intelligent customer service question and answer processing system 100 can continuously capture the core needs of users based on the semantic feature vector of each round of user questions, optimize the response content by combining the historical dialogue context, maintain the consistency and relevance of the response, and further improve the user consultation experience.

[0048] The aforementioned intelligent customer service question-and-answer processing method first performs intent recognition on the text information of the user's input question, identifying the corresponding intent tags and intent recognition scores. Then, based on the intent recognition scores, it determines the clarity of the intent and distinguishes the tag types in conjunction with the intent tags. This allows for the precise matching of differentiated response processing logic to different types of user questions, fundamentally solving the technical shortcomings of traditional customer service systems that offer only a single response method and cannot specifically handle various types of questions. On one hand, for user questions with clear intent (i.e., intent recognition scores greater than a preset threshold), the response path is further precisely adapted based on the tag type. Specifically, for static business rule tags with standardized and fixed answers, the matching target business text information is directly retrieved from the corresponding tenant's dedicated knowledge base. Based on this target business text information, standardized and fixed response content for the user question can be generated. For dynamic business data tags that require real-time data acquisition and dynamically changing results, real-time and accurate business data is obtained by calling the target tools in the corresponding tenant's intelligent agent toolset. Then, response content that fits the actual business scenario is generated based on the acquired business data. This approach, which targets scenarios with clear intent and differentiates responses based on demand categories, ensures both efficiency in handling standardized user questions and real-time accuracy in responding to user questions requiring dynamic business data. Compared to traditional rule-based customer service systems, which require manual keyword matching and cannot access real-time business data, and machine learning-based customer service systems, which suffer from lag and insufficient accuracy in dynamic responses, this approach effectively avoids response bias and delays, improving the accuracy and efficiency of customer service system responses to user questions and enhancing the user's consultation experience.

[0049] On the other hand, for user questions with ambiguous intent (i.e., intent recognition scores less than or equal to a preset threshold), the user question is first transformed into a semantic feature vector through feature extraction and vector transformation. Then, it is matched with the semantic feature vector in the tenant's exclusive knowledge base for similarity. Based on the matching results, the target business text information is selected. Finally, based on the target business text, accurate and relevant response content is generated to meet the user's real needs. This method of matching directly based on the semantic feature vector of the user question breaks through the limitation of traditional customer service systems that cannot effectively respond to ambiguous intent questions. It can accurately capture the core semantics of user questions and achieve effective compatibility with diverse, colloquial, and ambiguous user questions. It significantly improves the efficiency of customer service systems in adapting to various user questions and the comprehensiveness of problem handling. It effectively solves the technical problems of traditional rule-based customer service systems that have strict requirements for user question expression and are prone to answer errors or failures. At the same time, it makes up for the technical defects of machine learning-based customer service systems that lack contextual understanding, have inconsistent responses in multi-turn dialogues, and lack targeting. It can continuously capture the core needs of users in multi-turn ambiguous dialogues, maintain the consistency and targeting of responses, and further improve the user consultation experience.

[0050] Furthermore, this application's solution invokes the corresponding dedicated knowledge base or intelligent agent tool through tenant identifiers, which not only achieves the isolation and reuse of knowledge and resources among multiple tenants, ensuring the differentiated service needs of different tenants, but also leverages a large model base shared by multiple tenants. When updating or optimizing the model, there is no need to adjust the system configuration of each tenant separately. It can be quickly synchronized to each tenant simply by completing the iteration based on a unified base. This effectively solves the problems of low flexibility in updating existing multi-tenant architecture models and high operation and maintenance costs, and significantly reduces the overall operation and maintenance costs and time costs of multi-tenant architecture customer service systems.

[0051] In some exemplary embodiments of this disclosure, based on the above embodiments, it is further explained that the knowledge base obtains multiple semantic feature vectors by performing feature extraction and vector transformation on multiple pieces of business text information, and establishes the association relationship between each semantic feature vector and the corresponding business text information. The multiple pieces of business text information, multiple semantic feature vectors and the association relationship are stored in the vector library of the corresponding tenant identifier.

[0052] Specifically, the knowledge base construction process in this application embodiment may include the following steps: First, acquiring multiple business text messages entered by operations or knowledge engineers through a multi-tenant backend page, targeting the business scenarios of the corresponding tenants. These business text messages may include frequently asked questions and business rules related to the tenants' daily business, ensuring that the entered information aligns with the tenants' actual consultation needs. Then, feature extraction and vector transformation are performed on each entered business text message. This can be achieved using a pre-trained text embedding model: first, the business text messages undergo standardized preprocessing (e.g., removing special symbols, standardizing text format, word segmentation), and then the preprocessed text is input into an embedding model (e.g., BERT, Sentence-BERT, Text2Vec, etc.). The embedding model automatically extracts the core semantic features of each text message, removes redundant and interfering information, and then converts the text into a high-dimensional semantic feature vector (e.g., 768-dimensional or 1024-dimensional vector), ensuring that each business text message has a unique corresponding semantic feature vector. Finally, the association between each semantic feature vector and the corresponding business text message is established. Finally, all the entered business text information, the converted multiple semantic feature vectors, and the established relationships are uniformly stored in the vector library corresponding to the tenant identifier (such as Milvus, Pinecone, FAISS, and other dedicated vector databases) to complete the construction of the vector library in the knowledge base, provide data support for the semantic matching of user fuzzy intent questions, and realize the isolated storage of vector libraries of different tenants to meet the differentiated business needs of multi-tenants.

[0053] Furthermore, when a new tenant (such as a clothing retailer or a catering service provider) has a need to build an intelligent customer service system, there is no need to build knowledge from scratch or customize the knowledge retrieval process. The tenant only needs to enter its own exclusive business text information (such as the return and exchange rules for clothing retailers or the delivery specifications for catering service providers) through the multi-tenant backend page. The intelligent customer service question and answer processing system 100 will automatically reuse the preset general RAG retrieval framework and call the text embedding model, vector transformation, and relationship establishment process mentioned above to quickly complete the construction of the vector library of the tenant's exclusive knowledge base. This eliminates the need for tenants to invest a lot of manpower and resources in developing knowledge infrastructure. Compared with the traditional customized development model, the efficiency is greatly improved. It enables multi-tenant to quickly reuse knowledge infrastructure, reducing the cost and cycle of building intelligent customer service systems for tenants. At the same time, relying on the isolation characteristics of the tenant vector library, it ensures that the business knowledge of each tenant is stored independently and does not interfere with each other, taking into account both reuse efficiency and differentiated needs.

[0054] In some exemplary embodiments of this disclosure, based on the above embodiments, the specific implementation method of obtaining matching target business text information from the knowledge base corresponding to the tenant identifier based on the intent tag in step 203 is further described. For example... Figure 3 As shown, the steps for retrieving matching target business text information from the knowledge base corresponding to the tenant identifier based on intent tags can specifically include the following steps: Step 301: Perform feature extraction and vector transformation on the intent label to obtain the intent semantic feature vector.

[0055] Step 302: Identify the knowledge base corresponding to the tenant identifier based on the tenant identifier, and obtain multiple semantic feature vectors from the knowledge base corresponding to the tenant identifier.

[0056] Step 303: Calculate the first similarity value between the intent semantic feature vector and each semantic feature vector.

[0057] Step 304: Determine the semantic feature vector whose first similarity value satisfies the first preset matching condition as the target semantic feature vector.

[0058] Step 305: Based on the association relationship stored in the knowledge base corresponding to the target semantic feature vector and the tenant identifier, query the business text information corresponding to the target semantic feature vector. The target business text information includes the business text information corresponding to the target semantic feature vector.

[0059] Specifically, in step 203, the scenario where "the intent recognition score is greater than the preset threshold and the intent label is a static business rule type label" is used because the response content of such labels depends on standardized and fixed business text information (such as business rules, FAQs, service specifications, etc.) in the tenant's exclusive knowledge base, and the semantics of the intent label are clear. Therefore, the target business text information that highly matches the intent label can be retrieved from the tenant's exclusive knowledge base through accurate semantic matching. The specific implementation process is as follows: First, the intelligent customer service question and answer processing system 100 performs feature extraction and vector transformation processing on the identified intent label. This processing can be implemented through a pre-trained text embedding model, which is consistent with the text embedding model used for business text vector transformation in the knowledge base construction stage to ensure the uniformity of the vector space and avoid semantic matching deviation due to model differences. Secondly, the intelligent customer service question-and-answer processing system 100, based on the pre-acquired tenant identifier corresponding to the user, accurately identifies and locates the knowledge base exclusive to that tenant identifier. It then batch-retrieves all stored business text semantic feature vectors from the vector library of this knowledge base. These vectors are all obtained through the same embedding model during the knowledge base construction phase and have established associations with the corresponding business text information, providing basic data support for subsequent semantic matching. Next, it calculates the first similarity value between the intent semantic feature vector and each business text semantic feature vector. This process can be integrated into the retrieval module of the Retrieval Augmentation (RAG) model, serving as the core link for the RAG model to achieve accurate retrieval. For example, the retrieval module of the RAG model can use common vector similarity calculation methods such as cosine similarity algorithm and Euclidean distance algorithm, with cosine similarity algorithm (with a value range of [-1, 1]) being preferred. It calculates the cosine value of the angle between the intent semantic feature vector and each business text semantic feature vector. The closer this similarity value is to 1, the higher the semantic similarity between the two; the closer the similarity value is to -1, the greater the semantic difference. The retrieval module of the RAG model traverses all business text semantic feature vectors, calculates the similarity between each one and the intent semantic feature vector, obtains the first similarity value between the intent semantic feature vector and each semantic feature vector, and generates a similarity ranking list to facilitate subsequent selection of target vectors.

[0060] Furthermore, the intelligent customer service question-and-answer processing system 100, based on a preset first preset matching condition, filters out semantic feature vectors that meet the requirements from the similarity ranking list and determines them as target semantic feature vectors. This first preset matching condition can be flexibly configured according to the tenant's business needs. For example, a first similarity threshold (e.g., a cosine similarity threshold of 0.7) can be set, determining all semantic feature vectors with a first similarity value greater than or equal to 0.7 as target semantic feature vectors. If multiple similarity values ​​meet the threshold requirement, the top N vectors (e.g., N=3) in similarity value ranking can be further retained as target semantic feature vectors, ensuring that the selected vectors have the strongest semantic correlation with the intent tag.

[0061] Finally, the intelligent customer service question and answer processing system 100, based on the determined target semantic feature vector, combines the association between the semantic feature vector and business text information pre-established and stored in the knowledge base corresponding to the tenant identifier to quickly retrieve the business text information corresponding to the target semantic feature vector. This business text information is the target business text information that matches the current intent tag and can be directly used to generate subsequent response content. This not only ensures the accuracy of static business rule-based requirement responses but also improves response efficiency. At the same time, relying on the isolation characteristics of the tenant's exclusive knowledge base, it further ensures that the response content is highly adapted to the tenant's business scenario.

[0062] In some exemplary embodiments of this disclosure, based on the above embodiments, each intelligent agent tool in the intelligent agent toolset is associated with one or more preset intent tags, further illustrating the specific implementation method of matching the target intelligent agent tool from the intelligent agent toolset corresponding to the tenant identifier according to the intent tag in step 204 above. For example... Figure 4 As shown, the steps for matching the target agent tool from the agent toolset corresponding to the tenant identifier based on the intent label can specifically include the following steps: Step 401: Identify the intelligent agent tool set corresponding to the tenant identifier based on the tenant identifier, and obtain one or more preset intent tags associated with each intelligent agent tool in the intelligent agent tool set corresponding to the tenant identifier.

[0063] Step 402: Obtain the second similarity value between the intent tag and each preset intent tag.

[0064] Step 403: Determine the preset intent tag whose second similarity value satisfies the second preset matching condition as the target intent tag, and determine the intelligent agent tool associated with the target intent tag as the target intelligent agent tool.

[0065] Specifically, for the scenario in step 204 where "the intent recognition score is greater than the preset threshold and the intent label is a dynamic business data type label," since the response content for such requests depends on real-time updated business data, it cannot be directly responded to using static text in the tenant's exclusive knowledge base. It is necessary to call the tools corresponding to the tenant's exclusive intelligent agent toolset to obtain real-time data. Moreover, the semantics of the intent label are clear. Therefore, a target intelligent agent tool that highly matches the intent label can be matched from the tenant's exclusive intelligent agent toolset through precise semantic matching. The specific implementation process is as follows: First, the intelligent customer service question and answer processing system 100 accurately identifies and locates the intelligent agent toolset exclusive to the tenant based on the pre-acquired tenant identifier corresponding to the user. This toolset is configured independently by the tenant according to its own business scenario, and each intelligent agent tool is associated with a preset intent label. The intelligent customer service question and answer processing system 100 obtains one or more preset intent labels associated with all intelligent agent tools in batches from this intelligent agent toolset. These preset intent labels are all in line with the tenant's business scenario and are semantically consistent with the user intent label identified in step 202, providing basic data support for subsequent similarity calculation. Then, the intelligent customer service question and answer processing system 100 calculates the second similarity value between the identified intent tag and each preset intent tag. The calculation process is similar to step 303 above, and can be found in the relevant description in step 303 above. It will not be repeated here.

[0066] Furthermore, the intelligent customer service question and answer processing system 100, based on a preset second preset matching condition, filters out preset intent tags that meet the second preset matching condition, and determines the preset intent tag that meets the second preset matching condition as the target intent tag. Then, it determines the intelligent agent tool associated with the target intent tag as the target intelligent agent tool that matches the current intent tag, which can be directly used for subsequent business data retrieval. This not only ensures the real-time and accuracy of dynamic business data-type request responses, but also improves the efficiency of tool invocation. At the same time, relying on the isolation characteristics of the tenant-specific intelligent agent tool set, it further ensures that the invoked tools are highly adapted to the tenant's business scenario, meeting the differentiated dynamic data processing needs of multiple tenants. It should be noted that the second preset matching condition can be flexibly configured according to the tenant's business needs. The configuration logic is similar to that of the first preset matching condition. For example, a second similarity threshold can be set (e.g., a cosine similarity threshold of 0.75), which is slightly higher than the first similarity threshold to ensure the accuracy of intelligent agent tool matching. All preset intent tags with a second similarity value greater than or equal to 0.75 are determined as target intent tags. If multiple similarity values ​​meet the threshold requirements, the top N preset intent tags with the highest similarity values ​​(e.g., N=1) can be retained as target intent tags to avoid conflicts in multiple tool calls and ensure that the selected intelligent agent tools have the strongest correlation with the business data requirements corresponding to the intent tags.

[0067] Based on the above embodiments, the specific implementation method of calling the target intelligent agent tool to obtain business data matching the intent label in step 204 is further explained. For example... Figure 5 As shown, the steps for calling the target intelligent agent tool to obtain business data matching the intent label can specifically include the following steps: Step 501: Send a tool invocation request for the target agent tool to the agent tool set corresponding to the tenant identifier through the multi-tenant agent.

[0068] Step 502: In response to the tool invocation request, invoke the target intelligent agent tool to obtain target business data that matches the business data type.

[0069] Among them, the business data that matches the intent tag includes the target business data.

[0070] Specifically, after identifying the target intelligent agent tool, the intelligent customer service Q&A processing system 100 sends a tool invocation request to the intelligent agent toolset corresponding to the tenant identifier through a multi-tenant intelligent agent. The multi-tenant intelligent agent is a scheduling and management intelligent entity designed for multi-tenant scenarios, possessing the ability to autonomously perceive tenant identity, verify invocation permissions, and forward requests, aiming to achieve isolation and security of tool invocations across tenants. Its role is to uniformly manage tool invocation requests from different tenants, complete tenant affiliation verification and user permission verification, and prevent cross-tenant data access and unauthorized invocations. Through the multi-tenant intelligent agent, while ensuring data isolation and security, the standardization and stability of tool invocations can be improved, providing reliable support for the differentiated business needs of multi-tenants. Meanwhile, the tool call request can carry core parameters such as tenant identifier, tool identifier, and business data type. In response to the tool call request, the call steps of the target intelligent agent tool are initiated. The target business data is accurately retrieved through the preset data interface to connect to the tenant's business system and accurately retrieve the target business data that matches the business data type. This target business data is the business data that matches the intent tag, ensuring that the retrieval process is safe and standardized and that the data is accurate in real time. This lays the foundation for generating natural language response content that fits the user's needs based on the business data.

[0071] Based on the above embodiments, the specific implementation method of matching each semantic feature vector with the user semantic feature vector in step 205, filtering out the target semantic feature vector based on the matching results, and obtaining the target business text information of the target semantic feature vector from the knowledge base corresponding to the tenant identifier is further explained. For example... Figure 6 As shown, the steps of matching each semantic feature vector with the user semantic feature vector, filtering out the target semantic feature vector based on the matching results, and obtaining the target business text information of the target semantic feature vector from the knowledge base corresponding to the tenant identifier can specifically include the following steps: Step 601: Calculate the third similarity value between each semantic feature vector and the user's semantic feature vector.

[0072] Step 602: Determine the semantic feature vector that satisfies the third preset matching condition with the third similarity value as the target semantic feature vector.

[0073] Step 603: Based on the association relationship stored in the knowledge base corresponding to the target semantic feature vector and the tenant identifier, query the business text information corresponding to the target semantic feature vector.

[0074] The target business text information includes the business text information corresponding to the target semantic feature vector.

[0075] Specifically, in step 205, the scenario where "the intent recognition score is less than or equal to the preset threshold" refers to a situation where the user's question intent is ambiguous (e.g., non-standard colloquial expression, unclear core demands, or confusion of multiple intents). In this case, it is impossible to directly locate the response path through intent tags. The intelligent customer service question and answer processing system 100 first extracts features and transforms vectors from the user's question text (this process can be achieved by reusing the text embedding model) to obtain the user's semantic feature vector. Then, based on the tenant identifier, it batches multiple semantic feature vectors from the knowledge base exclusive to the corresponding tenant. Subsequently, through the semantic matching, feature vector filtering, and business text query process in steps 601-603, it accurately captures the core demands corresponding to the user's ambiguous intent and obtains the matched target business text information. The complete process of matching each semantic feature vector with the user's semantic feature vector and selecting the target semantic feature vector based on the matching results can be implemented through the retrieval module of the Retrieval Augmentation Generation (RAG) model. The specific implementation steps are as follows: First, the intelligent customer service question-and-answer processing system 100 calculates the third similarity value between each semantic feature vector and the user's semantic feature vector. This calculation process is consistent with the calculation logic of the first similarity value in step 303 and the second similarity value in step 402, reusing the same vector calculation standard corresponding to the same text embedding model to ensure the uniformity of the full-text technology implementation and avoid matching deviations caused by differences in calculation logic. For example, the cosine similarity algorithm (with a value range of [-1, 1]) is preferentially used to calculate the cosine value of the angle between each business text semantic feature vector retrieved from the knowledge base and the user's semantic feature vector; this cosine value is the third similarity value.

[0076] Then, the intelligent customer service question-and-answer processing system 100 determines the semantic feature vectors whose third similarity value meets the third preset matching condition as the target semantic feature vectors. The third preset matching condition can be flexibly configured according to the tenant's business needs and the accuracy of fuzzy intent processing. It is similar to the configuration logic of the first and second preset matching conditions and can be adjusted independently to adapt to fuzzy intent scenarios. For example, the third similarity threshold can be set to 0.65 (slightly lower than the first threshold of 0.7 and the second threshold of 0.75), taking into account both the comprehensiveness and accuracy of matching, and including semantic feature vectors with a third similarity value greater than or equal to 0.65 in the candidate range; if there are many candidate vectors, the top N (N is an integer greater than 0, such as 1) vectors can be further retained as target semantic feature vectors, which ensures the relevance of the screening results and avoids the risk of single vector matching deviation.

[0077] Finally, the intelligent customer service question-and-answer processing system 100 retrieves the business text information corresponding to the target semantic feature vector based on the relationships stored in the knowledge base corresponding to the target semantic feature vector and the tenant identifier. These relationships are established and stored synchronously with the semantic feature vector and business text information in the tenant's dedicated vector library during the knowledge base construction phase, ensuring convenient and accurate queries. By quickly matching related records with the target semantic feature vector, the corresponding business text information can be extracted. This information is the target business text information that matches the user's ambiguous intent, providing reliable support for subsequently generating clear and easy-to-understand responses adapted to the user's colloquial and ambiguous expressions. This effectively solves the technical shortcomings of traditional customer service systems, such as their inability to efficiently respond to ambiguous intents and their strict requirements on user question expression. It further improves the customer service system's adaptability and response accuracy to various user questions, enhancing the user's consultation experience. Simultaneously, in multi-round fuzzy dialogue scenarios, the intelligent customer service question-and-answer processing system 100 can recalculate the third similarity value and filter the target semantic feature vector based on the user's supplementary question text in each round, continuously optimizing the target business text information and maintaining the consistency and relevance of the response.

[0078] According to certain exemplary embodiments of this disclosure, based on the above embodiments, such as Figure 7 As shown, an intelligent customer service question-and-answer processing method further includes the following steps: Step 701: If the intent recognition score is less than or equal to a preset threshold and the number of selected target semantic feature vectors is zero, then the transfer to human customer service is triggered.

[0079] Step 702: During the service process of human customer service, collect the conversation text information between the user and human customer service in real time, and perform feature extraction and vector transformation on the conversation text information to obtain supplementary semantic feature vectors.

[0080] Step 703: Match the supplementary semantic feature vector with each semantic feature vector in the knowledge base corresponding to the tenant identifier, and select the matching reference semantic feature vector based on the matching results.

[0081] Step 704: Obtain the reference business text information corresponding to the reference semantic feature vector from the knowledge base corresponding to the tenant identifier, and push the reference business text information to human customer service to assist human customer service in answering user questions.

[0082] Specifically, in scenarios where the intent recognition score is less than or equal to a preset threshold and semantic matching fails (i.e., the number of target semantic feature vectors is zero), such as when the user's question intent is extremely vague, the colloquial expression is chaotic, or the consultation content exceeds the current tenant's existing knowledge base coverage, the intelligent customer service cannot capture the core needs and obtain the corresponding business text information through semantic matching. In this case, human customer service intervention is required to ensure that the user's needs are met. Simultaneously, the intelligent customer service question-and-answer processing system 100 assists human customer service in improving response efficiency and accuracy. First, when the intelligent customer service question-and-answer processing system detects that the intent recognition score is less than or equal to a preset threshold (e.g., 80 points) and the number of target semantic feature vectors obtained after the aforementioned semantic matching is zero, it immediately triggers the human customer service transfer mechanism. The intelligent customer service question-and-answer processing system 100 automatically synchronizes the text information of the current user's consultation session and the user's associated information (such as tenant identifier, user account, etc.) to the human customer service reception interface, while simultaneously providing the user with a standardized prompt of "Transferring you to human customer service, please wait," avoiding long waiting times and clearly informing the user of the progress, thus improving the user experience.

[0083] Subsequently, during the real-time conversation between the customer service representative and the user, the intelligent customer service question-and-answer processing system 100 collects the conversation text information from both parties in real time (including the user's supplementary inquiry statements and the customer service representative's inquiries and responses). The system then performs standardized preprocessing on the collected conversation text information (removing redundant interjections and special symbols, and standardizing terminology). It then reuses the pre-trained text embedding model described earlier to extract features and transform vectors from the preprocessed conversation text, obtaining supplementary semantic feature vectors that accurately represent the core semantics of the conversation. Furthermore, the intelligent customer service question-and-answer processing system 100 performs similarity matching between the generated supplementary semantic feature vectors and the semantic feature vectors in the tenant identifier-specific knowledge base corresponding to the user. The matching logic is consistent with the calculation logic of the third similarity value in step 601. This can be integrated into the RAG model retrieval module, using a cosine similarity algorithm to calculate the similarity value between the two vectors. Based on preset reference matching conditions (such as a similarity threshold ≥ 0.6), it filters out reference semantic feature vectors that match the supplementary semantic feature vectors, ensuring that the business text information corresponding to the filtered vectors highly matches the explicit user needs in the current conversation.

[0084] Finally, the intelligent customer service question and answer processing system 100, based on the selected reference semantic feature vectors and combined with the pre-established association between semantic feature vectors and business text information in the tenant's exclusive knowledge base, quickly retrieves the corresponding reference business text information (including relevant business rules, FAQs, service specifications, etc.) and synchronizes this reference business text information to the human customer service reception interface in the form of pop-ups, sidebar pushes, etc. Without the need for human customer service to manually search the knowledge base, standardized business reference content can be quickly obtained, assisting human customer service in answering user questions accurately and in a standardized manner, avoiding answering errors or inefficiencies caused by human customer service due to business memory biases or cumbersome knowledge base searches, thereby improving the user's consultation experience.

[0085] Based on certain exemplary embodiments of this disclosure, and building upon the above embodiments, further explanation is provided, such as... Figure 8 As shown, an intelligent customer service question-and-answer processing method further includes the following steps: Step 801: Collect historical conversation text information of human customer service corresponding to the target tenant identifier within a historical time period of a preset duration.

[0086] Step 802: Filter out new conversation text information from historical conversation text information whose consultation frequency is higher than a preset frequency threshold.

[0087] Step 803: Perform intent recognition on the newly added conversation text information to identify the new intent tag corresponding to the newly added conversation text information.

[0088] Step 804: Perform feature extraction and vector transformation on the newly added conversation text information to obtain the newly added semantic feature vector.

[0089] Step 805: Establish new associations between the newly added semantic feature vector, the newly added session text information, and the newly added intent tag, and store the newly added semantic feature vector, the newly added session text information, the newly added intent tag, and the newly added associations in the knowledge base corresponding to the target tenant identifier.

[0090] Specifically, firstly, the intelligent customer service Q&A processing system 100 automatically collects historical conversation text information of all human customer service representatives corresponding to the target tenant identifier, according to a preset time period (which can be flexibly configured by the tenant through the multi-tenant backend page, for example, 7 days, 30 days, etc.). This historical conversation text information all comes from the conversation scenarios in steps 701-704 where the intelligent customer service response failed and was transferred to human customer service, and includes a large number of consultation requests not covered by the tenant's corresponding knowledge base. During the collection process, the intelligent customer service Q&A processing system 100 preprocesses the historical conversation text information, including deduplication and noise reduction, and removes invalid conversations, duplicate conversations, and meaningless chatter, ensuring that the collected conversation text information has effective semantic value. At the same time, based on the multi-tenant isolation mechanism, the intelligent customer service Q&A processing system 100 only collects the conversation data corresponding to the target tenant identifier, without involving the conversation information of other tenants, effectively ensuring data isolation and security in the multi-tenant scenario. Then, the intelligent customer service Q&A processing system 100 uses a large model to statistically analyze the consultation frequency of the preprocessed historical conversation text information, calculates the number of times the core consultation needs corresponding to each conversation text information occur, and compares the statistical consultation frequency with a preset frequency threshold (this threshold can be flexibly configured by the tenant according to its own business scale and consultation volume; for example, it can be set to 5 times or more per month). New conversation text information with a consultation frequency higher than the preset frequency threshold is then filtered out. This new conversation text information refers to conversation text information corresponding to this type of consultation need that has not been included in the target tenant's current knowledge base, and has not formed corresponding standardized business text information and intent tags. It typically represents high-frequency consultation content that appears after the tenant adds new business or adjusts business rules, and is the core content to be added to the knowledge base.

[0091] Furthermore, the intelligent customer service question-and-answer processing system 100 performs intent recognition processing on the newly selected conversation text information. This processing can reuse the intent recognition model obtained by fine-tuning the large model base in step 202 above, ensuring the consistency and accuracy of the intent recognition logic. Specifically, the newly added conversation text information is first standardized and preprocessed (removing redundant embellishments, standardizing terminology, and eliminating irrelevant symbols). Then, the preprocessed text information is input into the intent recognition model. Through the model's contextual semantic understanding, core request extraction, and business terminology adaptation capabilities, the new intent tag corresponding to each new conversation text information is accurately identified. This new intent tag is consistent with the format and semantic specifications of the intent tags in the previous embodiment, fitting the business scenario of the target tenant, ensuring seamless compatibility with subsequent intelligent customer service intent matching, response generation, and other logic. Then, the intelligent customer service question and answer processing system 100 performs feature extraction and vector transformation on the newly added conversation text information. This process can reuse the pre-trained text embedding models (such as BERT, Sentence-BERT, etc.) used in the previous knowledge base construction and semantic matching process to ensure the uniformity of vector transformation standards and avoid subsequent semantic matching deviations due to model differences.

[0092] Finally, the intelligent customer service question-and-answer processing system 100 establishes new associations among the newly added semantic feature vectors, newly added conversation text information, and newly added intent tags. This ensures that the corresponding newly added conversation text information can be quickly retrieved through the newly added intent tags or newly added semantic feature vectors. After establishing the associations, the intelligent customer service question-and-answer processing system 100 stores the newly added semantic feature vectors, newly added conversation text information, newly added intent tags, and the newly added associations in a unified manner in the knowledge base corresponding to the target tenant identifier, and synchronously updates it to the tenant's dedicated vector library, completing the automatic supplementation and iterative update of the knowledge base. Therefore, this embodiment achieves dynamic iteration and autonomous optimization of the knowledge base by automatically collecting historical conversation text information of the target tenant's human customer service, filtering frequently added consultation content, performing intent recognition and vector conversion on the added content, establishing associations, and supplementing it to the tenant's dedicated knowledge base. This solves the technical defects of traditional knowledge base static storage, low efficiency of manual updates, and inability to adapt to the tenant's new business consultation needs in real time. At the same time, it adapts to multi-tenant isolation scenarios, ensuring that each tenant's knowledge base can be independently updated and iterated based on its own business scenarios, further improving the coverage and accuracy of the intelligent customer service system's responses and reducing the pressure on human customer service transfers.

[0093] For the intelligent customer service question-and-answer processing method disclosed herein, a further explanation of the specific implementation method of the intelligent customer service question-and-answer processing method is provided, refer to... Figure 9 As shown: The first part will be introduced, which is the knowledge infrastructure stage (i.e.) Figure 9In the process described in path ①, the tenant's operations or knowledge engineer enters the tenant's unique business text information through the multi-tenant backend page. This business text information includes, but is not limited to, static knowledge such as business rules and frequently asked questions. After receiving this business text information, the intelligent customer service question-and-answer processing system 100 inputs it into a pre-trained text embedding model for standardized preprocessing and feature extraction, generating a high-dimensional semantic feature vector that accurately represents the semantics of the text. Then, after establishing the association between the semantic feature vector and the original business text information, the intelligent customer service question-and-answer processing system 100 stores this content together in the exclusive vector knowledge base corresponding to the tenant's identifier, thus completing the construction of the knowledge base. This step, through a multi-tenant isolation mechanism, ensures that the business knowledge of different tenants is stored independently and does not interfere with each other, providing accurate and isolated knowledge data support for subsequent intelligent responses.

[0094] Next, we will introduce the intelligent response process for user inquiries (i.e.) Figure 9 In path ②), when a user initiates an inquiry through the driver's app, passenger's app, or backend terminal, the intelligent customer service question-and-answer processing system 100 first performs intent recognition on the user's input question text, generating corresponding intent tags and intent recognition scores to determine the user's intent and its clarity. Then, based on the intent recognition results, the intelligent customer service question-and-answer processing system 100 matches differentiated response paths: If the intent recognition score is greater than the preset threshold and the intent tag is a static business rule type requirement, the intelligent customer service question and answer processing system 100 will perform semantic matching in the tenant's exclusive vector knowledge base based on the identified intent tag by retrieving the enhanced generative model (RAG model), retrieving the matched target business text information, and generating standardized response content.

[0095] If the intent recognition score is greater than the preset threshold and the intent label is a dynamic business data type requirement, the intelligent customer service question and answer processing system 100 will use a multi-tenant intelligent agent to schedule the tenant's exclusive intelligent agent toolset, call the corresponding business tools to obtain real-time business data, and then generate dynamic response content based on this data.

[0096] If the intent recognition score is less than or equal to a preset threshold, i.e. the user's intent is ambiguous, the intelligent customer service question and answer processing system 100 will convert the user's input question into a semantic feature vector through a retrieval augmented generation model (RAG model), perform similarity matching in the vector knowledge base, filter out the target business text information that fits the user's core needs, and generate accurate response content.

[0097] Finally, we will introduce the dynamic iteration and self-evolution of the knowledge base (i.e.) Figure 9(Path ③) To continuously optimize the coverage of the knowledge base, the intelligent customer service question-and-answer processing system 100 periodically performs a knowledge mining process. First, the system periodically collects historical conversation text information from human customer service representatives within a preset timeframe. These conversations originate from scenarios where intelligent customer service responses fail and are then transferred to human agents, representing a knowledge blind spot not currently covered by the knowledge base. Next, the system analyzes this conversation text information using a large language model, filtering out new conversation text information with a consultation frequency exceeding a preset threshold. It then performs intent recognition and vector transformation on this new text, generating new intent tags and new semantic feature vectors. Finally, the system establishes the relationships between the new content and adds it to the tenant's dedicated vector knowledge base, completing the autonomous iteration of the knowledge base.

[0098] In summary, this disclosure provides an intelligent customer service question-and-answer processing method: 1. Improved adaptability and response accuracy: Through a dual mechanism of intent recognition and semantic matching, it can accurately parse users' colloquial and ambiguous questions, effectively solving the shortcomings of traditional rule-based customer service systems, such as "poor expression adaptability and lack of ability to handle complex issues." Simultaneously, in multi-turn dialogues, it can continuously capture users' core needs based on semantic feature vectors, compensating for the lack of contextual understanding in machine learning-based customer service systems, making responses more coherent and targeted.

[0099] 2. The efficiency of dynamic demand response is significantly improved. Through the multi-tenant intelligent agent scheduling dedicated toolset, dynamic business data can be retrieved in real time, solving the problem of "inability to obtain real-time data and delayed dynamic response" in traditional customer service systems.

[0100] 3. Improved adaptability to multi-tenant scenarios: By using tenant identifiers, knowledge and tools can be isolated and reused, which not only ensures the differentiated service needs of different tenants, but also avoids cross-tenant data confusion and leakage.

[0101] 4. Improved knowledge base iteration efficiency and reduced operation and maintenance costs: By automatically collecting human conversations, mining high-frequency new demands and supplementing them to the knowledge base, a closed-loop iteration mechanism of "knowledge entry - intelligent response - consultation conversation collection - knowledge supplementation" is constructed, which solves the problem of "static storage and low efficiency of manual updates" in traditional knowledge bases. In addition, multiple tenants share a large model base, and model upgrades do not require adaptation to each tenant individually. They only need to be uniformly iterated to be synchronized to all tenants, which greatly reduces the risks of "insufficient update flexibility and high operation and maintenance costs" in multi-tenant architectures. Furthermore, new tenants can quickly reuse the general RAG framework to build their own knowledge base without developing from scratch, which significantly reduces the cost and cycle of building tenant intelligent customer service.

[0102] It should be understood that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order constraint on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0103] In some exemplary embodiments of this disclosure, such as Figure 10 As shown, an intelligent customer service question-and-answer processing device is provided, including a first acquisition module 1001, a first intent recognition module 1002, a first generation module 1003, a second generation module 1004, and a third generation module 1004. The first acquisition module 1001 is used to receive text information of a user's question input by the user and acquire the tenant identifier corresponding to the user; the first intent recognition module 1002 is used to perform intent recognition based on the text information to obtain the intent tag and intent recognition score corresponding to the text information; the first generation module 1003 is used to, if the intent recognition score is greater than a preset threshold and the intent tag is a static business rule type tag, then obtain matching target business text information from the knowledge base corresponding to the tenant identifier based on the intent tag, and generate the response content of the user's question based on the target business text information; the second generation module 1004 is used to, if the intent recognition score is greater than a preset threshold and the intent tag is a dynamic business data type tag, then obtain the response content of the user's question from the knowledge base corresponding to the tenant identifier based on the intent tag. The corresponding intelligent agent tool set matches the target intelligent agent tool, calls the target intelligent agent tool to obtain business data that matches the intent label, and generates the response content of the user question based on the business data; the third generation module 1005 is used to perform feature extraction and vector transformation processing on the text information of the user question if the intent recognition score is less than or equal to a preset threshold, to obtain the user semantic feature vector, obtain multiple semantic feature vectors from the knowledge base corresponding to the tenant identifier, match each semantic feature vector with the user semantic feature vector, filter out the target semantic feature vector based on the matching result, obtain the target business text information of the target semantic feature vector from the knowledge base corresponding to the tenant identifier, and generate the response content of the user question based on the target business text information.

[0104] In one embodiment of this disclosure, the knowledge base obtains multiple semantic feature vectors by performing feature extraction and vector transformation on multiple pieces of business text information, establishes the association between each semantic feature vector and the corresponding business text information, and stores the multiple pieces of business text information, multiple semantic feature vectors and association relationships in a vector library with corresponding tenant identifiers.

[0105] In one embodiment of this disclosure, the first generation module 1003 is specifically used for: performing feature extraction and vector transformation processing on the intent tag to obtain an intent semantic feature vector; identifying the knowledge base corresponding to the tenant identifier based on the tenant identifier, and obtaining multiple semantic feature vectors from the knowledge base corresponding to the tenant identifier; calculating a first similarity value between the intent semantic feature vector and each semantic feature vector; determining the semantic feature vector whose first similarity value satisfies a first preset matching condition as the target semantic feature vector; and querying the business text information corresponding to the target semantic feature vector based on the association relationship stored in the knowledge base corresponding to the tenant identifier, wherein the target business text information includes the business text information corresponding to the target semantic feature vector.

[0106] In one embodiment of this disclosure, each intelligent agent tool in the intelligent agent toolset is associated with one or more preset intent tags. The second generation module 1004 is specifically used to: identify the intelligent agent toolset corresponding to the tenant identifier based on the tenant identifier; obtain one or more preset intent tags associated with each intelligent agent tool in the intelligent agent toolset corresponding to the tenant identifier; obtain a second similarity value between the intent tag and each preset intent tag; determine the preset intent tag whose second similarity value satisfies the second preset matching condition as the target intent tag; and determine the intelligent agent tool associated with the target intent tag as the target intelligent agent tool.

[0107] In one embodiment of this disclosure, the second generation module 1004 is specifically configured to: send a tool invocation request for the target intelligent agent tool to the intelligent agent toolset corresponding to the tenant identifier through a multi-tenant intelligent agent; and in response to the tool invocation request, invoke the target intelligent agent tool to obtain target business data that matches the business data type, wherein the business data that matches the intent tag includes the target business data.

[0108] In one embodiment of this disclosure, the third generation module 1005 is specifically used for: calculating a third similarity value between each semantic feature vector and the user semantic feature vector; determining the semantic feature vector whose third similarity value satisfies a third preset matching condition as the target semantic feature vector; and querying the business text information corresponding to the target semantic feature vector based on the association relationship stored in the knowledge base corresponding to the target semantic feature vector and the tenant identifier, wherein the target business text information includes the business text information corresponding to the target semantic feature vector.

[0109] In one embodiment of this disclosure, an intelligent customer service question-and-answer processing device further includes: a triggering module, configured to trigger a human customer service transfer step if the intent recognition score is less than or equal to a preset threshold and the number of selected target semantic feature vectors is zero; a first acquisition module, configured to collect the conversation text information between the user and the human customer service representative in real time during the service process of the human customer service representative, and perform feature extraction and vector transformation processing on the conversation text information to obtain supplementary semantic feature vectors; a matching module, configured to match the supplementary semantic feature vectors with each semantic feature vector in the knowledge base corresponding to the tenant identifier, and select the matched reference semantic feature vectors based on the matching results; and a second acquisition module, configured to acquire the reference business text information corresponding to the reference semantic feature vectors from the knowledge base corresponding to the tenant identifier, and push the reference business text information to the human customer service representative to assist the human customer service representative in answering user questions.

[0110] In one embodiment of this disclosure, an intelligent customer service question-and-answer processing device further includes: a second acquisition module, used to acquire historical conversation text information of human customer service corresponding to a target tenant identifier within a historical time period of a preset duration; a filtering module, used to filter out new conversation text information with a consultation frequency higher than a preset frequency threshold from the historical conversation text information; a second intent recognition module, used to perform intent recognition on the new conversation text information and identify the new intent tag corresponding to the new conversation text information; a processing module, used to perform feature extraction and vector transformation processing on the new conversation text information to obtain a new semantic feature vector; and an association module, used to establish a new association relationship between the new semantic feature vector, the new conversation text information, and the new intent tag, and store the new semantic feature vector, the new conversation text information, the new intent tag, and the new association relationship in the knowledge base corresponding to the target tenant identifier.

[0111] For specific limitations regarding an intelligent customer service question-and-answer processing device, please refer to the limitations of an intelligent customer service question-and-answer processing method described above, which will not be repeated here. Each module in the aforementioned intelligent customer service question-and-answer processing device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0112] In some exemplary embodiments of this disclosure, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 11As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores business data from each tenant's intelligent customer service question-and-answer system. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements an intelligent customer service question-and-answer processing method.

[0113] Those skilled in the art will understand that Figure 11 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0114] In some exemplary embodiments of this disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of an intelligent customer service question-and-answer processing method as described in any of the exemplary embodiments above.

[0115] In some exemplary embodiments of this disclosure, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of an intelligent customer service question-and-answer processing method as described in any of the exemplary embodiments above.

[0116] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0117] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0118] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for intelligent customer service question and answer processing, characterized in that, The method includes: Receive text information of user questions input by the user, and obtain the tenant identifier corresponding to the user; Based on the text information, intent recognition is performed to obtain the intent label and intent recognition score corresponding to the text information; If the intent recognition score is greater than a preset threshold and the intent tag is a static business rule type tag, then based on the intent tag, the matching target business text information is obtained from the knowledge base corresponding to the tenant identifier, and the response content of the user question is generated based on the target business text information. If the intent recognition score is greater than the preset threshold, and the intent tag is a dynamic business data type tag, then a target intelligent agent tool is matched from the intelligent agent tool set corresponding to the tenant identifier according to the intent tag, the target intelligent agent tool is called to obtain dynamic business data that matches the intent tag, and the response content of the user question is generated based on the dynamic business data; If the intent recognition score is less than or equal to the preset threshold, feature extraction and vector transformation are performed on the text information of the user question to obtain a user semantic feature vector. Multiple semantic feature vectors are obtained from the knowledge base corresponding to the tenant identifier. Each semantic feature vector is matched with the user semantic feature vector. Based on the matching results, a target semantic feature vector is selected. The target business text information of the target semantic feature vector is obtained from the knowledge base corresponding to the tenant identifier. The response content of the user question is generated based on the target business text information.

2. The method according to claim 1, characterized in that, The knowledge base obtains multiple semantic feature vectors by performing feature extraction and vector transformation on multiple pieces of business text information, and establishes the association between each semantic feature vector and the corresponding business text information. The multiple pieces of business text information, the multiple semantic feature vectors, and the association are stored in the vector library of the corresponding tenant identifier.

3. The method according to claim 2, characterized in that, The step of retrieving matching target business text information from the knowledge base corresponding to the tenant identifier based on the intent tag includes: The intent tag is subjected to feature extraction and vector transformation to obtain the intent semantic feature vector; Based on the tenant identifier, the knowledge base corresponding to the tenant identifier is identified, and multiple semantic feature vectors are obtained from the knowledge base corresponding to the tenant identifier; Calculate the first similarity value between the intent semantic feature vector and each semantic feature vector; The semantic feature vector whose first similarity value satisfies the first preset matching condition is determined as the target semantic feature vector; Based on the relationship stored in the knowledge base corresponding to the target semantic feature vector and the tenant identifier, the business text information corresponding to the target semantic feature vector is queried, and the target business text information includes the business text information corresponding to the target semantic feature vector.

4. The method according to claim 1, characterized in that, Each agent tool in the agent toolset is associated with one or more preset intent tags. The step of matching a target agent tool from the agent toolset corresponding to the tenant identifier based on the intent tags includes: Based on the tenant identifier, identify the intelligent agent tool set corresponding to the tenant identifier, and obtain one or more preset intent tags associated with each intelligent agent tool in the intelligent agent tool set corresponding to the tenant identifier; Obtain the second similarity value between the intent tag and each preset intent tag; The preset intent tag whose second similarity value satisfies the second preset matching condition is determined as the target intent tag, and the intelligent agent tool associated with the target intent tag is determined as the target intelligent agent tool.

5. The method according to claim 4, characterized in that, The step of invoking the target intelligent agent tool to obtain business data matching the intent label includes: The multi-tenant agent sends a tool invocation request for the target agent tool to the agent tool set corresponding to the tenant identifier; In response to the tool invocation request, the target intelligent agent tool is invoked to obtain target business data that matches the business data type, wherein the business data that matches the intent tag includes the target business data.

6. The method according to claim 2, characterized in that, The step of matching each semantic feature vector with the user semantic feature vector, filtering out target semantic feature vectors based on the matching results, and obtaining target business text information of the target semantic feature vectors from the knowledge base corresponding to the tenant identifier includes: Calculate the third similarity value between each semantic feature vector and the user semantic feature vector; The semantic feature vector whose third similarity value satisfies the third preset matching condition is determined as the target semantic feature vector; Based on the relationship stored in the knowledge base corresponding to the target semantic feature vector and the tenant identifier, the business text information corresponding to the target semantic feature vector is queried, and the target business text information includes the business text information corresponding to the target semantic feature vector.

7. The method according to claim 6, characterized in that, The method further includes: If the intent recognition score is less than or equal to the preset threshold, and the number of the selected target semantic feature vectors is zero, then the human customer service transfer step is triggered. During the service process of the human customer service, the conversation text information between the user and the human customer service is collected in real time, and the conversation text information is processed by feature extraction and vector transformation to obtain supplementary semantic feature vectors. The supplementary semantic feature vector is matched with each semantic feature vector in the knowledge base corresponding to the tenant identifier, and the matching reference semantic feature vector is selected based on the matching results. The reference business text information corresponding to the reference semantic feature vector is obtained from the knowledge base corresponding to the tenant identifier, and the reference business text information is pushed to the human customer service to assist the human customer service in answering the user's question.

8. An intelligent customer service question-and-answer processing device, characterized in that, The device includes: The first acquisition module is used to receive text information of user questions input by the user and to acquire the tenant identifier corresponding to the user; The first intent recognition module is used to perform intent recognition based on the text information to obtain the intent label and intent recognition score corresponding to the text information. The first generation module is used to obtain matching target business text information from the knowledge base corresponding to the tenant identifier based on the intent tag if the intent recognition score is greater than a preset threshold and the intent tag is a static business rule type tag, and generate the response content of the user question based on the target business text information. The second generation module is used to match a target intelligent agent tool from the intelligent agent tool set corresponding to the tenant identifier according to the intent tag if the intent recognition score is greater than a preset threshold and the intent tag is a dynamic business data type tag, call the target intelligent agent tool to obtain business data that matches the intent tag, and generate the response content of the user question based on the business data; The third generation module is used to perform feature extraction and vector transformation processing on the text information of the user question if the intent recognition score is less than or equal to the preset threshold, to obtain a user semantic feature vector, obtain multiple semantic feature vectors from the knowledge base corresponding to the tenant identifier, match each semantic feature vector with the user semantic feature vector, filter out the target semantic feature vector based on the matching result, obtain the target business text information of the target semantic feature vector from the knowledge base corresponding to the tenant identifier, and generate the response content of the user question based on the target business text information.

9. A computer 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 steps of the method according to any one of claims 1 to 7.

10. A 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 steps of the method according to any one of claims 1 to 7.

Images