Customer service intelligent dialogue method and system

Abstract

The invention relates to the technical field of intelligent customer service, and discloses a customer service intelligent dialogue method and system.The method comprises the steps that cross-channel semantic unified processing is conducted on user input texts of multiple target channels, and unified input data is obtained; inputting the unified input data into an expert knowledge enhanced large language model for field rule matching and case retrieval to obtain an intention analysis result; performing command template matching and parameter verification according to the intention analysis result to obtain a structured service command; and executing data query and expression optimization through the structured service command to obtain a reply text conforming to customer service specifications, the method solves the problem of information islands of a traditional system, ensures continuity and consistency of cross-channel services, realizes specialization and standardization of customer service replies, and improves the user experience. And converting the structured business data into natural language expression conforming to the understanding habit of the user.

Description

Technical Field

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

[0002] Traditional customer service systems often employ independent processing mechanisms when handling user inquiries from different channels such as WeChat, apps, and websites, lacking a unified semantic understanding and cross-channel information integration capability. This fragmented approach forces users to repeatedly provide information when switching between channels, resulting in a inconsistent service experience and increasing the workload of customer service personnel and operational costs for businesses. Existing intelligent customer service technologies rely on rule matching and simple machine learning models for intent recognition, which are insufficiently accurate when faced with complex business scenarios and technical terminology. They struggle to accurately understand users' true needs and perform poorly in handling complex queries, multi-turn dialogues, and personalized service requests, often requiring human intervention to complete the service. Summary of the Invention

[0003] This invention provides a customer service intelligent dialogue method and system. This invention solves the problem of information silos in traditional systems, ensures the continuity and consistency of cross-channel services, realizes the professionalism and standardization of customer service responses, and converts structured business data into natural language expressions that conform to users' understanding habits.

[0004] In a first aspect, the present invention provides a customer service intelligent dialogue method, the customer service intelligent dialogue method comprising: Perform cross-channel semantic unification processing on user input text from multiple target channels to obtain unified input data; The unified input data is fed into a large language model enhanced with expert knowledge for domain rule matching and case retrieval to obtain intent parsing results. Based on the intent parsing results, command template matching and parameter verification are performed to obtain structured business commands; The structured business commands are used to perform data queries and optimize the presentation, resulting in response text that conforms to customer service standards.

[0005] In conjunction with the first aspect, in a first implementation of the first aspect of the present invention, the step of performing cross-channel semantic unification processing on user input text from multiple target channels to obtain unified input data includes: Symbolic semantic mapping and lexical expansion are performed on user input text from multiple target channels to obtain standardized channel text; Based on the standardized channel text, user identifier, device identifier, and session identifier are extracted to obtain a metadata set; The metadata set is input into a semantic aligner for cross-channel intent expression mapping and semantic normalization calculation to obtain standard intent and channel weight coefficients; Based on the stated standard intent and the stated channel weight coefficient, historical dialogues are associated and session states are synchronized to obtain unified input data.

[0006] In conjunction with the first aspect, in a second implementation of the first aspect of the present invention, the step of inputting the metadata set into a semantic aligner for cross-channel intent expression mapping and semantic normalization calculation to obtain standard intent and channel weight coefficients includes: The user identifier, device identifier, and session identifier in the metadata set are subjected to channel semantic feature extraction and text similarity calculation to obtain a multi-dimensional semantic feature vector. Based on the multidimensional semantic feature vector, a cross-channel semantic mapping matrix is ​​constructed, and based on the cross-channel semantic mapping matrix, association matching and similarity measurement of the same intent in different channels are performed to obtain the semantic association matrix between channels; The user input text is normalized and its intent is standardized and mapped based on the semantic correlation matrix between the channels to obtain the standard intent. Based on the similarity values ​​in the semantic correlation matrix between channels, channel credibility is quantified and weights are assigned to obtain channel weight coefficients.

[0007] In conjunction with the first aspect, in a third implementation of the first aspect of the present invention, the step of constructing a cross-channel semantic mapping matrix based on the multi-dimensional semantic feature vector, and performing association matching and similarity measurement of the same intent across different channels based on the cross-channel semantic mapping matrix to obtain a semantic association matrix between channels, includes: The multidimensional semantic feature vectors are grouped and clustered according to channel type, and the semantic center vector and variance distribution within each channel group are calculated to obtain the channel distribution features. Based on the channel distribution characteristics, a semantic mapping relationship between channels is established by cosine similarity calculation and Euclidean distance metric, and a similarity matrix is ​​constructed to obtain the initial semantic mapping matrix; The initial semantic mapping matrix is ​​normalized and threshold filtered to retain mapping relationships with similarity higher than a preset threshold, thus obtaining a semantic association matrix between channels.

[0008] In conjunction with the first aspect, in the fourth implementation of the first aspect of the present invention, the step of inputting the unified input data into a large language model enhanced with expert knowledge for domain rule matching and case retrieval to obtain intent parsing results includes: The unified input data is input into the large language model enhanced by expert knowledge. The unified input data is then semantically matched and relevance calculated with the domain rule base, historical case base and expert pattern base to obtain expert knowledge matching data. Based on the expert knowledge matching data, high-dimensional semantic vector transformation and context relationship establishment are performed to obtain expert knowledge vectors; The intent classification and confidence assessment are performed on the expert knowledge vector and the unified input data to obtain the intent parsing result.

[0009] In conjunction with the first aspect, in the fifth implementation of the first aspect of the present invention, the step of inputting the unified input data into the expert knowledge-enhanced large language model, and performing semantic matching and correlation calculation between the unified input data and the domain rule base, historical case base, and expert pattern base to obtain expert knowledge matching data includes: The unified input data is input into a large language model enhanced with expert knowledge, and the standard intent in the unified input data is matched with the business judgment rules in the domain rule base for text semantic matching and rule condition verification to obtain the rule matching result. Based on the target business entities in the unified input data, similar case retrieval and semantic distance calculation are performed in the historical case database to obtain case retrieval results; Based on the intent features of the unified input data, thought pattern matching and processing flow recognition are performed in the expert pattern library to obtain the pattern matching result. The rule matching results, the case retrieval results, and the pattern matching results are used as expert knowledge matching data.

[0010] In conjunction with the first aspect, in the sixth implementation of the first aspect of the present invention, the step of performing similar case retrieval and semantic distance calculation in the historical case database based on the target business entity in the unified input data to obtain case retrieval results includes: The target business entities in the unified input data are matched with the case keywords in the historical case library by text matching and entity alignment to filter out candidate case data containing the same business entities. For each case in the candidate case data, the cosine similarity and edit distance with the unified input data are calculated using semantic vector encoding to obtain the case ranking result; The top N cases with the highest similarity are selected from the case ranking results, and the corresponding case processing results and satisfaction scores are extracted to obtain the case retrieval results.

[0011] In conjunction with the first aspect, in the seventh implementation of the first aspect of the present invention, the step of performing command template matching and parameter verification based on the intent parsing result to obtain a structured business command includes: Based on the intent category in the intent parsing result, template retrieval and interface matching are performed in the preset command template library to obtain the matching command template; Based on the target business entity in the intent parsing result, the placeholders in the matching command template are filled with entity values ​​and their data types are converted to obtain the first business command. The format and value range of the business parameters in the first business command are validated to obtain the second business command. By using the business rule diagram, the second business command is subjected to constraint checks and process compliance verification to obtain a structured business command.

[0012] In conjunction with the first aspect, in the eighth implementation of the first aspect of the present invention, the step of performing constraint checks and process compliance verification on the second business command through a business rule graph to obtain a structured business command includes: Map the business operation type and parameter value in the second business command to the corresponding node in the business rule graph, and identify the constraints and prerequisite dependencies that need to be met; The constraints and prerequisite dependencies are verified and checked item by item by using a graph traversal algorithm. When a violation of business rules is found, error information is recorded and correction suggestions are generated to obtain the rule verification results. Based on the verification results of the rules, the second business command that passes the verification is marked as executable and an execution priority identifier is added to obtain the structured business command.

[0013] In conjunction with the first aspect, in the ninth implementation of the first aspect of the present invention, the step of performing data query and expression optimization through the structured business command to obtain a response text conforming to customer service specifications includes: Based on the target business system identifier in the structured business command, available service discovery and interface matching are performed in the service registry to obtain a candidate service list; Health status detection and response time statistics are performed on each business service in the candidate service list to obtain service status data; Based on the service status data, the response time, load factor, and data consistency are calculated, and routing decision results are generated. The structured business command is sent to the target business service via an interface call according to the routing decision result, and fault tolerance processing and response data reception are performed to obtain query data results from the order system, user system or inventory system. Based on the query data results, semantic transformation and expression optimization are performed to obtain a response text that conforms to customer service specifications.

[0014] In conjunction with the first aspect, in the tenth implementation of the first aspect of the present invention, the step of initiating an interface call to the target business service according to the routing decision result and performing fault-tolerant processing and response data reception to obtain query data results from the order system, user system, or inventory system includes: Based on the optimal service address and calling parameters in the routing decision results, the structured business command is encapsulated into an interface request and its protocol format is converted to obtain a standard interface call request. The standard interface call request is sent to the target business service to initiate a network connection and the call status and response time are monitored to obtain call status information. Based on the failure identifier in the call status information, exponential backoff retry and degradation strategy switching are performed to obtain service response data; The service response data is parsed and converted to obtain query data results from the order system, user system, or inventory system.

[0015] In conjunction with the first aspect, in the eleventh implementation of the first aspect of the present invention, the step of performing semantic transformation and expression optimization based on the query data results to obtain a response text conforming to customer service specifications includes: The order status code, delivery time and amount value in the query data results are processed to generate status descriptions and naturalize the time to obtain readable business information; Based on the business scenario type of the readable business information, retrieve matching customer service response patterns from the expert expression style library and extract style features; Based on the style features, the readable business information and historical dialogue records are used to generate response content and perform style adaptation and conversion to obtain candidate response text; The candidate response texts are subjected to accuracy verification, completeness checks, and sensitive content filtering to obtain response texts that conform to customer service specifications.

[0016] In conjunction with the first aspect, in the twelfth implementation of the first aspect of the present invention, the step of retrieving a matching customer service response pattern from the expert expression style library and extracting style features based on the business scenario type of the readable business information includes: Based on the business scenario identifier in the readable business information, scenario classification matching and style type filtering are performed in the expert expression style library to obtain candidate style patterns; Sentence structure features, standard use of professional terms, and emotional expression tendencies are extracted from each style pattern in the candidate style patterns. The matching degree with the current dialogue context is calculated to obtain a style matching score. Select the style pattern with the highest style matching score, extract its language feature vector and template parameters to obtain style features.

[0017] Secondly, the present invention provides a customer service intelligent dialogue system, the customer service intelligent dialogue system comprising: The semantic unification processing module is used to perform cross-channel semantic unification processing on user input text from multiple target channels to obtain unified input data; The intent parsing module is used to input the unified input data into a large language model enhanced with expert knowledge for domain rule matching and case retrieval to obtain intent parsing results; The command template matching module is used to perform command template matching and parameter verification based on the intent parsing result to obtain structured business commands. The expression optimization module is used to perform data query and expression optimization through the structured business commands to obtain response text that conforms to customer service specifications.

[0018] The technical solution provided by this invention achieves unified processing of user input from multiple channels through a semantic aligner, solving the information silo problem of traditional systems and ensuring the continuity and consistency of cross-channel services. The expert knowledge fusion layer deeply integrates the business rules, historical cases, and processing patterns of customer service experts into a large language model, significantly enhancing the accuracy of intent understanding in complex business scenarios through knowledge vectorization and attention mechanisms. The intelligent command conversion engine adopts a combination of dynamic template generation and expert rule constraints, avoiding the limitations of fixed templates, and ensuring that generated commands conform to actual business processes through business rule graph verification. The multi-dimensional routing decision algorithm comprehensively considers factors such as response time, system load, and data consistency to achieve optimal service selection, integrating fault tolerance, retry mechanisms, and degradation strategies to ensure high availability. The expert expression style library and semantic conversion technology achieve professional and standardized customer service responses, converting structured business data into natural language expressions that conform to user understanding habits.

[0019] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of one embodiment of the intelligent customer service dialogue method in this invention; Figure 2 This is a schematic diagram of one embodiment of the intelligent customer service dialogue system in this invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] The terms "comprising" and "having," and any variations thereof, used in the embodiments of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the steps or units listed, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0024] To facilitate understanding of this embodiment, a customer service intelligent dialogue method disclosed in this embodiment will first be described in detail. For example... Figure 1 As shown, this method includes the following steps: 101. Perform cross-channel semantic unification processing on user input text from multiple target channels to obtain unified input data; 102. Input the unified input data into the expert knowledge-enhanced large language model for domain rule matching and case retrieval to obtain intent parsing results; 103. Based on the intent parsing results, perform command template matching and parameter validation to obtain structured business commands; 104. Execute data queries and optimize presentations using structured business commands to obtain response texts that conform to customer service specifications.

[0025] In one specific embodiment, the process of performing step 101 may specifically include the following steps: Symbolic semantic mapping and lexical expansion are performed on user input text from multiple target channels to obtain standardized channel text; User identifiers, device identifiers, and session identifiers are extracted from the standardized channel text to obtain a metadata set; The metadata set is input into the semantic aligner to perform cross-channel intent expression mapping and semantic normalization calculation, resulting in standard intent and channel weight coefficients; Based on the standard intent and channel weight coefficient, historical dialogues are associated and session states are synchronized to obtain unified input data.

[0026] Specifically, the system performs symbolic semantic mapping and vocabulary expansion operations on the raw input text submitted by users through heterogeneous channels such as WeChat, apps, web pages, or telephone voice calls. The symbolic semantic mapping module uses a pre-defined emoji-emotion tag lookup table to convert non-text elements in the input, such as emojis and special identifiers, into emotion or semantic tags. Simultaneously, for text expressions containing abbreviations, pinyin, or industry-standard abbreviations, the vocabulary expansion mechanism utilizes a custom thesaurus, common expression normalization rules, and a contextual understanding model to standardize non-standard terms, expanding them into standardized business expressions such as "order query" and "view order status," resulting in standardized channel text. Based on the standardized text, key identification information related to the channel is extracted, including three dimensions: user identifier, device identifier, and session identifier. Specifically, the openid is extracted as the user identifier for WeChat, the deviceid as the device identifier for the app, and the current user's session identifier is extracted based on cookies or session_tokens for the web page. These identifiers are used to establish a multi-dimensional metadata set at the user, device, and session levels. The metadata set is input into a semantic aligner for cross-channel intent expression mapping and normalization. Based on a pre-trained semantic embedding model and a channel semantic mapping table, the semantic aligner identifies variations of the same intent across different channels and maps non-standard sentences to standard intents through sentence vector matching and context alignment mechanisms. Simultaneously, it comprehensively calculates the semantic credibility and semantic fit of the current channel based on factors such as historical channel response quality, user stickiness, and device interaction preferences, generating corresponding channel weight coefficients. Based on the standard intent and channel weight coefficients, the user's historical dialogue records across multiple channels are semantically hierarchically associated to construct a cross-channel dialogue graph. A context state synchronization mechanism embeds the current input into the historical conversation flow. After determining whether the user is asking the same question, whether there is already a clear business context, and whether there are any unclosed transactions, the associated context state is restored, integrating historical conversation content with the current channel input to form unified input data.

[0027] In one specific embodiment, the process of inputting the metadata set into the semantic aligner for cross-channel intent representation mapping and semantic normalization calculation to obtain the standard intent and channel weight coefficients can specifically include the following steps: Channel semantic features are extracted and text similarity is calculated from the user identifier, device identifier and session identifier in the metadata set to obtain a multi-dimensional semantic feature vector; A cross-channel semantic mapping matrix is ​​constructed based on multi-dimensional semantic feature vectors, and the association matching and similarity measurement of the same intent in different channels are performed based on the cross-channel semantic mapping matrix to obtain the semantic association matrix between channels; Based on the semantic correlation matrix between channels, the user input text is normalized and the intent is standardized and mapped to obtain the standard intent. Channel credibility is quantified and weighted based on the similarity values ​​in the semantic correlation matrix between channels, resulting in channel weight coefficients.

[0028] Specifically, channel semantic features are extracted and text similarity is calculated for user identifiers, device identifiers, and session identifiers in the metadata set. A channel feature encoder nests and combines each identifier with its corresponding context input, and a semantic embedding model is used to extract its deep semantic information, constructing a multi-dimensional semantic feature vector. This vector includes static features such as user identity attributes, device type preferences, and session interaction trajectories, and incorporates dynamic features such as sentiment, keyword density, and historical response efficiency from the input text. Each input has a comparable high-dimensional semantic representation. Based on the multi-dimensional semantic feature vector, a cross-channel semantic mapping matrix is ​​constructed. This matrix measures the semantic similarity between channels expressing the same intent by calculating the dot product similarity or cosine similarity between semantic vectors from different channels. A general intent classification index is used to constrain and cluster the similarity relationships in the matrix, forming channel matching blocks with clear semantic mapping relationships within the matrix. A line-by-line analysis of the cross-channel semantic mapping matrix is ​​performed to extract sentence fragments expressing the same intent from different channels. Semantic-level intent merging and standardization are then performed based on similarity scores. Through normalization transformation, phrases such as "check order," "where is the order," and "order status" are uniformly mapped to standard intents like ORDER_STATUS_INQUIRY, eliminating intent classification bias caused by differences in channel expression. Based on the similarity scores of each channel's input contribution to the standard intent in the inter-channel semantic association matrix, a weighted calculation is performed. Taking into account multiple factors such as semantic clarity, contextual consistency, and historical response quality, a channel credibility score is generated for each channel. This score is then standardized into a weight vector, forming the channel weight coefficient.

[0029] In the construction of the cross-channel semantic mapping matrix, the multi-dimensional semantic feature vectors are grouped and clustered according to three types: WeChat channel, APP channel, and web page channel. Specifically, the K-means clustering algorithm is used to perform cluster analysis on the semantic feature vectors of each channel type, and the semantic center vector within each channel group is calculated. This center vector is obtained by averaging all feature vectors within the group. At the same time, the variance distribution is calculated to reflect the dispersion of the semantic expression of the channel, thus obtaining the channel distribution characteristics including the center coordinates and distribution range. Based on the channel distribution characteristics, the cosine value of the angle between the semantic vectors of different channels is calculated using the cosine similarity calculation formula, and the spatial distance between the vectors is calculated using the Euclidean distance metric formula. When the cosine similarity is greater than 0.7 and the Euclidean distance is less than a preset threshold, it is determined that there is a semantic mapping relationship between the two channels, and the similarity value is filled into the corresponding position in the similarity matrix, thereby constructing the initial semantic mapping matrix. The initial semantic mapping matrix is ​​normalized by scaling the values ​​in the matrix to the range of 0 to 1. A threshold of 0.6 is set to filter out weak association mappings with similarity below the threshold and retain strong associations with high similarity. Finally, a semantic association matrix between channels that can accurately reflect the semantic correspondence between different channels is obtained.

[0030] In one specific embodiment, the process of performing step 102 may specifically include the following steps: The unified input data is fed into the large language model enhanced with expert knowledge. The unified input data is then semantically matched and relevance calculated with the domain rule base, historical case base, and expert pattern base to obtain expert knowledge matching data. Based on expert knowledge matching data, high-dimensional semantic vector transformation and context relationship establishment are performed to obtain expert knowledge vectors; Intent classification and confidence assessment are performed on expert knowledge vectors and unified input data to obtain intent parsing results.

[0031] Specifically, the unified input data is fed into a large language model structure with an expert knowledge enhancement mechanism. Building upon its original language understanding capabilities, the large language model integrates three types of expert knowledge resources: a domain rule base built for specific business scenarios, a historical case library derived from human customer service interactions, and an expert pattern library abstracted from experienced customer service processing procedures. These three types of knowledge cover rule matching, experience-based response, and process logic, respectively. After receiving the unified input data, the model initiates a semantic matching and relevance calculation process. Using a pre-trained semantic encoder, it performs nested matching between the input text and the content from the three types of knowledge. Through mechanisms such as keyword Boolean filtering, sentence vector similarity analysis, and contextual fragment comparison, it determines the semantic consistency between the input and each rule, case, or pattern, and calculates the relevance score, forming a set of expert knowledge matching data that includes the degree of rule matching, case similarity level, and pattern alignment. The expert knowledge matching data is input into the knowledge vector conversion module. This module uses knowledge vectorization tools to encode unstructured text or graph knowledge into a unified vector representation. Simultaneously, it establishes contextual semantic links through entity associations, logical dependencies, and business constraints defined in the expert knowledge graph. Based on this, it constructs expert knowledge vectors with context-sensitive features and cross-knowledge source alignment features. Intent classification and confidence assessment are then performed on the expert knowledge vectors and the unified input data. The language model of the expert knowledge vectors and input data is embedded into the vector input intent classifier, performing multi-class intent label matching calculations. A multi-head attention mechanism focuses on key correspondences between input and knowledge to determine the most likely intent category. Simultaneously, the confidence assessment module comprehensively calculates the reliability score of the current intent judgment based on indicators such as matching strength, semantic consistency, and knowledge distribution density. The output includes intent parsing results containing intent category, confidence score, key entity information, and business context description.

[0032] In one specific embodiment, the process of inputting unified input data into a large language model enhanced with expert knowledge, and performing semantic matching and correlation calculation on the unified input data with the domain rule base, historical case base, and expert pattern base to obtain expert knowledge matching data can specifically include the following steps: The unified input data is fed into the large language model enhanced with expert knowledge, and the standard intent in the unified input data is matched with the business judgment rules in the domain rule base through text semantic matching and rule condition verification to obtain the rule matching result. Based on the target business entities in the unified input data, similar cases are retrieved and semantic distance is calculated in the historical case database to obtain case retrieval results; Based on the intent features of the unified input data, the thought patterns are matched and the processing flow is identified in the expert pattern library to obtain the pattern matching results. The rule matching results, case retrieval results, and pattern matching results are used as expert knowledge matching data.

[0033] Specifically, a unified input data system is integrated into a large language model system with an expert knowledge mechanism. This input data includes semantic components such as standard intent, key entity information, contextual state, and channel weights. After receiving the input, the model performs textual semantic matching and rule condition verification between the standard intent and the business judgment rules in a pre-built domain rule base. The domain rule base contains judgment conditions, triggering semantics, parameter constraints, and execution boundaries defined by business experts. The standard intent is transformed into a semantic vector through semantic embedding and compared with the rule items using sentence vector similarity. A logical condition parser performs structural parsing and value verification on the rule condition fields to confirm whether the input matches the rule definition and forms a rule matching result. Based on the target business entity information contained in the unified input data, a similar case retrieval operation is performed in the historical case base. The similar case base includes a large number of high-quality manually processed records. Each case includes the original user expression, processing path, execution result, and satisfaction rating. A semantic distance calculation algorithm is used to evaluate the matching degree between the current input and the entity attribute vectors of all cases in the case base. The optimal matching path is calculated by combining semantic closeness and contextual similarity, and the case retrieval result is output, identifying the historical response paradigm closest to the current input. Simultaneously, based on the intent features in the unified input data, processing thought pattern matching and processing flow recognition operations are performed in the expert pattern library. The expert pattern library contains typical thought paths and operation flow models of experienced customer service personnel handling different business intents, including sequential logic such as "identity verification—information retrieval—solution generation." The model semantically pairs the intent tags in the input with the process entry nodes defined in the pattern library, and identifies the processing path corresponding to the current intent based on the trajectory backtracking mechanism of the process state diagram, generating pattern matching results. The rule matching results, case retrieval results, and pattern matching results are then integrated to form expert knowledge matching data.

[0034] In the process of retrieving similar cases from the historical case database, entity recognition is performed on the unified input data to extract target business entities, such as order numbers, product names, and problem types. Then, the extracted target business entities are matched with the keywords of tens of thousands of historical cases stored in the database. A combination of exact and fuzzy string matching is used, along with entity alignment processing, to identify "order query" and "query order" as the same entity type. Candidate cases containing at least one identical business entity are selected, forming a candidate case dataset. Deep semantic analysis is performed on each case in the candidate case dataset. The BERT model is used to encode the case text into semantic vectors, generating a semantic vector representation. The cosine similarity between this semantic vector representation and the semantic vector of the unified input data is calculated, along with the edit distance between the two vectors in high-dimensional space. These two metrics are combined to score the similarity of the cases, and the cases are sorted from highest to lowest score to obtain the case ranking results. Select the top 5 cases with the highest similarity from the case ranking results, extract the case handling results recorded in these cases, including the customer service expert's handling method, solution, processing time, and user satisfaction rating of the handling results, and return this information as the case retrieval results.

[0035] In one specific embodiment, the process of performing step 103 may specifically include the following steps: Based on the intent category in the intent parsing results, template retrieval and interface matching are performed in the preset command template library to obtain the matching command template; Based on the target business entity in the intent parsing result, the placeholders in the matching command template are filled with entity values ​​and their data types are converted to obtain the first business command. The format and value range of the business parameters in the first business command are validated to obtain the second business command. By using the business rule diagram, the second business command is subjected to constraint checks and process compliance verification to obtain a structured business command.

[0036] Specifically, based on the intent category field in the intent parsing results, template retrieval and interface matching operations are performed in a pre-defined command template library. This library predefines interface structures for multiple business systems. Each template includes a service name, calling method, parameter placeholders, calling protocol format, and input / output conventions. By matching the intent category with the tag fields defined in the template, the standard interface template that best matches the current user intent is quickly located. The target business entity information in the intent parsing results is used to populate the parameter placeholders in the command template with entity values. By mapping the correspondence between entity identifiers and template variable fields, key fields such as order number, user ID, refund amount, and delivery time are extracted from the standard intent for parameter replacement. Furthermore, data type conversion operations are performed in conjunction with business interface specifications to ensure that text, numeric, boolean, and date parameters meet the interface requirements for format consistency and semantic correctness, resulting in the first business command. The parameter validation module performs format and numerical range validation on all completed business parameters in the first business command. This includes validating the character format of fields such as order number, phone number, and email address using regular expressions, validating the boundary validity of numerical fields such as refund amount and product quantity using upper and lower limit rules, and performing time window constraint judgment on the time field. This generates a second business command that passes the initial compliance validation. The second business command is then input into a business rule graph constructed based on expert knowledge for in-depth constraint checks and process compliance validation. The business rule graph expresses the transformation relationships and constraint rules between entity states in various business scenarios in a graph structure, such as "refund application requires order status to be completed" and "modification of delivery address must be within 48 hours before shipment." The graph traversal mechanism verifies the logical conditions between nodes along the current business path to ensure that the generated business command will not trigger abnormal situations such as business conflicts, permission out-of-bounds access, or process interruption in the current session state. After all validations pass, the command becomes the final structured business command.

[0037] In the business rule graph constraint check process, a graph structure is established that includes business entities, business operations, and business rules. Nodes in the graph represent business elements such as order status, user permissions, and time limits, while edges represent business rule constraint relationships. When the second business command is generated, the business operation type in the command, such as "order refund," is mapped to the "refund operation" node in the business rule graph. Parameter values, such as order status, application amount, and application time, are mapped to the corresponding attribute nodes. The constraints that the operation must meet are identified through the edge relationships between nodes, such as "order status must be completed," "application time must be within 30 days of order completion," and "refund amount must not exceed the total order amount." Prerequisite dependencies, such as "user identity must be verified first," are also identified. A depth-first search algorithm is used to traverse the business rule graph. Constraints and prerequisite dependencies are verified and checked item by item according to the topological sorting order. This verifies whether the order status meets requirements, the application time is within the validity period, and the refund amount is compliant. When a constraint is not met, specific error information such as "Order status is in transit, refund conditions not met" is immediately recorded, and a correction suggestion such as "Please wait for the order to be completed before applying for a refund" is generated based on the business rules. The verification results are summarized into rule verification results. For the second business command that passes all constraint verifications, its status is marked as "executable," and an execution priority identifier is added according to the business priority rules: urgent orders are marked as high priority, and ordinary orders are marked as normal priority, forming a complete structured business command.

[0038] In one specific embodiment, the process of performing step 104 may specifically include the following steps: Based on the target business system identifier in the structured business command, perform available service discovery and interface matching in the service registry to obtain a candidate service list; Health status detection and response time statistics are performed on each business service in the candidate service list to obtain service status data; Response time, load factor, and data consistency are calculated based on service status data, and routing decision results are generated. The structured business commands are sent to the target business service via interface calls according to the routing decision results, and fault tolerance processing and response data reception are performed to obtain query data results from the order system, user system or inventory system. Based on the query data results, semantic transformation and expression optimization are performed to obtain response text that conforms to customer service standards.

[0039] Specifically, based on the target business system identifier contained in the structured business command, the system performs service discovery and interface matching operations in the service registry. The service registry stores registration information for different business services such as the order system, user system, and inventory system, including service name, interface path, calling method, protocol type, response constraints, and performance metrics. By comparing the target business system identifier with each registry entry, a list of all candidate services that can meet the command call requirements is filtered out, ensuring structural consistency between the selected interface and the parameter fields in the command. Health status checks and response time statistics are performed on each business service in the candidate service list. Through heartbeat monitoring, interface call latency detection, error rate calculation, and resource usage analysis, the availability of each service at the current moment is evaluated, and a set of service status data is generated based on this. Performance metrics are calculated based on the service status data, yielding response time, load factor, and data consistency metrics. These performance metrics are combined into a core reference for routing decisions through a multi-dimensional weighted evaluation method, and the routing optimization engine generates the optimal routing decision result. After completing the routing decision, the structured business commands are submitted to the target business service to initiate interface calls according to the optimal routing decision result. Simultaneously, fault tolerance mechanisms are enabled to improve the robustness of the call process. These include automatic exponential backoff retries in case of call failure, triggering circuit breaker protection when a service remains abnormal, and automatically degrading to backup services or cached data in unavailability situations. This yields query data results from the order system, user system, or inventory system. Through a business data semanticization component, the query data results undergo semantic transformation and expression optimization, mapping structured fields to natural language expressions. Furthermore, the response is optimized using an expert expression style library to obtain response text that conforms to customer service specifications.

[0040] In one specific embodiment, the process of executing the structured business command to initiate an interface call to the target business service according to the routing decision result, and performing fault tolerance processing and response data reception to obtain query data results from the order system, user system, or inventory system can specifically include the following steps: Based on the optimal service address and calling parameters in the routing decision results, the structured business command is encapsulated into an interface request and its protocol format is converted to obtain a standard interface call request. The standard interface call request initiates a network connection to the target business service, and the call status is monitored and the response time is measured to obtain the interface call status information. Based on the failure flag in the API call status information, exponential backoff retry and degradation strategy switching are performed to obtain service response data; The service response data is parsed according to the protocol and converted into a data format to obtain query data results from the order system, user system or inventory system.

[0041] Specifically, based on the optimal service address and call parameters specified in the routing decision results, the generated structured business commands undergo interface request encapsulation and protocol format conversion. The encapsulation process maps the service identifier, method name, and parameter fields in the command to standardized interface request messages, and adapts the format according to the communication protocol used by the target service, such as HTTP, gRPC, or SOAP, to form a standard interface call request that conforms to the interface call specification. The standard request initiates a connection to the target business service through the network transmission module. During the call, a real-time monitoring mechanism is activated to measure the connection status, data packet round-trip latency, and interface response time throughout the request's lifecycle. It also records whether the call experiences timeouts, interruptions, or error returns, generating interface call status information containing latency, error codes, and availability indicators. Based on the failure indicators in the interface call status information, a fault tolerance mechanism is dynamically triggered. When call failure or response timeout is detected, an exponential backoff retry strategy is activated, gradually extending the retry interval to reduce the impact on the target service. If multiple retries fail, a degradation strategy is automatically executed, switching the call request to a backup service instance or returning cached data to ensure the availability and continuity of the overall business chain. After successfully obtaining the service response data, protocol parsing and data format conversion are performed. The parsing layer decomposes and maps the field content in the original message according to the interface definition of the target business system, while the data conversion layer unifies the encoding format, data type, or field naming used by different systems into a standardized format to obtain the query data results.

[0042] In one specific embodiment, the process of performing semantic transformation and expression optimization based on the query data results to obtain a response text that conforms to customer service specifications can specifically include the following steps: The order status code, delivery time and amount values ​​in the query data results are processed to generate status descriptions and naturalize the time to obtain readable business information. Based on the business scenario type of readable business information, retrieve matching customer service response patterns from the expert expression style library and extract target style features; Based on the target style features, readable business information and historical dialogue records are used to generate response content and perform style adaptation and conversion to obtain candidate response text; The candidate response texts are verified for accuracy, checked for completeness, and filtered for sensitive content to obtain response texts that conform to customer service standards.

[0043] Specifically, the query data results returned from the business system undergo semantic processing. Fields such as order status codes, delivery times, and amounts are processed with status descriptions and time-based naturalization to obtain readable business information. Based on the business scenario type corresponding to the readable business information, a retrieval and matching operation is performed in an expert expression style library. This library includes various customer service response patterns, such as order inquiries, refund applications, logistics tracking, and complaint feedback. Each scenario defines sentence structure, tone style, and professional terminology. The most suitable response pattern is determined by comparing semantic indexes with scenario tags, and the corresponding target style features are extracted. These target style features are applied to the generation of readable business information response content, and historical dialogue records are used to maintain contextual consistency. A language generation engine adapts the expression style while preserving information integrity, generating a set of candidate response texts. The candidate response texts undergo multi-layered verification, including accuracy verification to ensure consistency between the data in the response and the query results, integrity checks to confirm that the response covers the core points of concern to the user, and sensitive content filtering to remove expressions involving violations, offenses, or sensitive information, resulting in standardized customer service response texts.

[0044] In the process of searching and matching expert expression style databases, a database of expert expression styles is maintained, encompassing various business scenarios such as order inquiries, refund processing, complaint responses, and consultation answers. For each scenario, multiple validated customer service response patterns are stored. Once readable business information is generated, the business scenario identifier is extracted, for example, identified as an "order delivery inquiry scenario." Then, the scenario category is located in the expert expression style database, and all style patterns within that category are filtered out, excluding inapplicable formal complaint styles and marketing promotion styles, resulting in a candidate style pattern set. This set includes various patterns such as approachable, professional, and concise styles. Feature analysis is performed on each style pattern in the candidate style patterns. Sentence structure features, such as the usage ratio of subject-verb-object, interrogative, and imperative sentences, are extracted. Standard terminology usage, such as "in delivery" and "expected delivery," is extracted. Emotional expression tendencies, such as a positive, reassuring tone and patient explanation, are extracted. These features are then compared with the historical dialogue records of the current conversation to calculate contextual matching. If the user exhibits anxiety in the historical dialogue, the matching score for the friendly style is increased; if the user uses professional terminology, the matching score for the professional style is increased. A weighted calculation is used to obtain a style matching score for each style pattern. The style pattern with the highest style matching score is selected, and its linguistic feature vectors, including numerical features such as word frequency distribution, sentence length distribution, and emotional polarity, as well as template parameters, including salutation methods and closing format, are extracted from this pattern. This information is then integrated into the target style features.

[0045] The above describes the intelligent customer service dialogue method in the embodiments of the present invention. The following describes the intelligent customer service dialogue system in the embodiments of the present invention. Please refer to [link / reference]. Figure 2 One embodiment of the intelligent customer service dialogue system in this invention includes: The semantic unification processing module 201 is used to perform cross-channel semantic unification processing on user input text from multiple target channels to obtain unified input data. The intent parsing module 202 is used to input unified input data into a large language model enhanced with expert knowledge for domain rule matching and case retrieval to obtain intent parsing results; The command template matching module 203 is used to perform command template matching and parameter verification based on the intent parsing results to obtain structured business commands. The expression optimization module 204 is used to perform data queries and expression optimization through structured business commands to obtain response text that conforms to customer service specifications.

[0046] Through the collaborative efforts of the aforementioned components, a semantic aligner enables unified processing of user input from multiple channels, resolving the information silo problem inherent in traditional systems and ensuring the continuity and consistency of cross-channel services. The expert knowledge fusion layer deeply integrates customer service experts' business rules, historical cases, and processing patterns into a large language model, significantly enhancing the accuracy of intent understanding in complex business scenarios through knowledge vectorization and attention mechanisms. The intelligent command conversion engine employs a combination of dynamic template generation and expert rule constraints, avoiding the limitations of fixed templates and ensuring that generated commands conform to actual business processes through business rule graph verification. The multi-dimensional routing decision algorithm comprehensively considers factors such as response time, system load, and data consistency to achieve optimal service selection, integrating fault tolerance, retry mechanisms, and degradation strategies to guarantee high availability. An expert expression style library and semantic conversion technology professionalize and standardize customer service responses, converting structured business data into natural language expressions that conform to user understanding habits.

[0047] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0048] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0049] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A customer service intelligent dialogue method, characterized in that, include: Perform cross-channel semantic unification processing on user input text from multiple target channels to obtain unified input data; The unified input data is fed into a large language model enhanced with expert knowledge for domain rule matching and case retrieval to obtain intent parsing results. Based on the intent parsing results, command template matching and parameter verification are performed to obtain structured business commands; The structured business commands are used to perform data queries and optimize the presentation, resulting in response text that conforms to customer service standards.

2. The intelligent customer service dialogue method according to claim 1, characterized in that, The process of performing cross-channel semantic unification processing on user input text from multiple target channels to obtain unified input data includes: Symbolic semantic mapping and lexical expansion are performed on user input text from multiple target channels to obtain standardized channel text; Based on the standardized channel text, user identifier, device identifier, and session identifier are extracted to obtain a metadata set; The metadata set is input into a semantic aligner for cross-channel intent expression mapping and semantic normalization calculation to obtain standard intent and channel weight coefficients; Based on the stated standard intent and the stated channel weight coefficient, historical dialogues are associated and session states are synchronized to obtain unified input data.

3. The intelligent customer service dialogue method according to claim 2, characterized in that, The step of inputting the metadata set into a semantic aligner for cross-channel intent representation mapping and semantic normalization calculation to obtain standard intent and channel weight coefficients includes: The user identifier, device identifier, and session identifier in the metadata set are subjected to channel semantic feature extraction and text similarity calculation to obtain a multi-dimensional semantic feature vector. Based on the multidimensional semantic feature vector, a cross-channel semantic mapping matrix is ​​constructed, and based on the cross-channel semantic mapping matrix, association matching and similarity measurement of the same intent in different channels are performed to obtain the semantic association matrix between channels. The user input text is normalized and its intent is standardized and mapped based on the semantic correlation matrix between the channels to obtain the standard intent. Based on the similarity values ​​in the semantic correlation matrix between channels, channel credibility is quantified and weights are assigned to obtain channel weight coefficients.

4. The intelligent customer service dialogue method according to claim 3, characterized in that, The process of constructing a cross-channel semantic mapping matrix based on the multi-dimensional semantic feature vector, and performing association matching and similarity measurement of the same intent across different channels based on the cross-channel semantic mapping matrix to obtain a semantic association matrix between channels, includes: The multidimensional semantic feature vectors are grouped and clustered according to channel type, and the semantic center vector and variance distribution within each channel group are calculated to obtain the channel distribution features. Based on the channel distribution characteristics, a semantic mapping relationship between channels is established by cosine similarity calculation and Euclidean distance metric, and a similarity matrix is ​​constructed to obtain the initial semantic mapping matrix; The initial semantic mapping matrix is ​​normalized and threshold filtered to retain mapping relationships with similarity higher than a preset threshold, thus obtaining a semantic association matrix between channels.

5. The intelligent customer service dialogue method according to claim 1, characterized in that, The process of inputting the unified input data into a large language model enhanced with expert knowledge for domain rule matching and case retrieval to obtain intent parsing results includes: The unified input data is input into the large language model enhanced by expert knowledge. The unified input data is then semantically matched and relevance calculated with the domain rule base, historical case base and expert pattern base to obtain expert knowledge matching data. Based on the expert knowledge matching data, high-dimensional semantic vector transformation and context relationship establishment are performed to obtain expert knowledge vectors; The intent classification and confidence assessment are performed on the expert knowledge vector and the unified input data to obtain the intent parsing result.

6. The intelligent customer service dialogue method according to claim 5, characterized in that, The process involves inputting the unified input data into a large language model enhanced with expert knowledge, and performing semantic matching and correlation calculations between the unified input data and the domain rule base, historical case base, and expert pattern base to obtain expert knowledge matching data, including: The unified input data is input into a large language model enhanced with expert knowledge, and the standard intent in the unified input data is matched with the business judgment rules in the domain rule base for text semantic matching and rule condition verification to obtain the rule matching result. Based on the target business entities in the unified input data, similar case retrieval and semantic distance calculation are performed in the historical case database to obtain case retrieval results; Based on the intent features of the unified input data, thought pattern matching and processing flow recognition are performed in the expert pattern library to obtain the pattern matching result. The rule matching results, the case retrieval results, and the pattern matching results are used as expert knowledge matching data.

7. The intelligent customer service dialogue method according to claim 6, characterized in that, The process of retrieving similar cases and calculating semantic distance in the historical case database based on the target business entity in the unified input data to obtain case retrieval results includes: The target business entities in the unified input data are matched with the case keywords in the historical case library by text matching and entity alignment to filter out candidate case data containing the same business entities. For each case in the candidate case data, the cosine similarity and edit distance with the unified input data are calculated using semantic vector encoding to obtain the case ranking result; The top N cases with the highest similarity are selected from the case ranking results, and the corresponding case processing results and satisfaction scores are extracted to obtain the case retrieval results.

8. The intelligent customer service dialogue method according to claim 1, characterized in that, The step of performing command template matching and parameter verification based on the intent parsing result to obtain structured business commands includes: Based on the intent category in the intent parsing result, template retrieval and interface matching are performed in the preset command template library to obtain the matching command template; Based on the target business entity in the intent parsing result, the placeholders in the matching command template are filled with entity values ​​and their data types are converted to obtain the first business command. The format and value range of the business parameters in the first business command are validated to obtain the second business command. By using the business rule diagram, the second business command is subjected to constraint checks and process compliance verification to obtain a structured business command.

9. The intelligent customer service dialogue method according to claim 8, characterized in that, The second business command is subjected to constraint checks and process compliance verification through a business rule diagram to obtain a structured business command, including: Map the business operation type and parameter value in the second business command to the corresponding node in the business rule graph, and identify the constraints and prerequisite dependencies that need to be met; The constraints and prerequisite dependencies are verified and checked item by item by using a graph traversal algorithm. When a violation of business rules is found, error information is recorded and correction suggestions are generated to obtain the rule verification results. Based on the verification results of the rules, the second business command that passes the verification is marked as executable and an execution priority identifier is added to obtain the structured business command.

10. A customer service intelligent dialogue system, characterized in that, A customer service intelligent dialogue method for performing any one of claims 1-9, comprising: The semantic unification processing module is used to perform cross-channel semantic unification processing on user input text from multiple target channels to obtain unified input data; The intent parsing module is used to input the unified input data into a large language model enhanced with expert knowledge for domain rule matching and case retrieval to obtain intent parsing results; The command template matching module is used to perform command template matching and parameter verification based on the intent parsing result to obtain structured business commands. The expression optimization module is used to perform data query and expression optimization through the structured business commands to obtain response text that conforms to customer service specifications.

Images