Interaction method and system for service consultation

Abstract

The embodiment of the invention provides an interaction method and system for service consultation, electronic equipment, a readable storage medium and a computer program product, and relates to the technical field of big data processing. The method comprises the following steps: identifying intention classification of a service consultation request; based on the identified intention classification, generating a routing instruction to route the service consultation request to a target knowledge base determined by at least one of a plurality of predefined private domain knowledge bases, the plurality of private domain knowledge bases including a conventional business knowledge base, a policy term knowledge base, a legal assistance knowledge base and a service supervision knowledge base; executing a retrieval enhancement generation (RAG) retrieval to retrieve at least one document fragment related to the service consultation request from the target knowledge base; and generating a natural language reply to the service consultation request through a large language model based on the retrieved at least one document segment.

Description

Technical Field

[0001] Embodiments of this disclosure relate to the field of computer technology, and more particularly to interactive methods and systems for service consultation, electronic devices, readable storage media, and computer program products. Background Technology

[0002] In service consultation fields such as civil aviation, existing solutions mainly rely on traditional human customer service or automated systems based on rule / keyword matching. The human model has inherent drawbacks such as low efficiency and high cost; while automated systems can handle high-frequency and standardized service consultations, their semantic understanding capabilities are weak and they cannot deeply analyze the complex and cross-domain compound intentions of users, resulting in poor response accuracy.

[0003] Furthermore, while general-purpose large language models possess powerful natural language generation capabilities, they exhibit significant limitations in specialized fields. On one hand, the model's accuracy is difficult to improve, potentially leading to factual errors in responses to specific service inquiries. On the other hand, the model lacks the ability to perceive external knowledge, making it difficult to ensure consistency between responses and external information such as industry standards, policies, and regulations. These deficiencies in accuracy and perception can ultimately lead to serious compliance risks in responses, potentially generating misleading answers that contradict external knowledge.

[0004] Therefore, there is an urgent need in this field for a highly efficient, accurate, and professional interactive technology solution for service consultation. Summary of the Invention

[0005] Embodiments of this disclosure provide interactive methods and systems, electronic devices, readable storage media, and computer program products for service consultation that can at least partially solve the problems described above or other problems in the art.

[0006] According to a first aspect of this disclosure, an interactive method for service consultation is provided, the method comprising: identifying an intent classification of a service consultation request; generating a routing instruction based on the identified intent classification to route the service consultation request to a target knowledge base determined by at least one of a plurality of predefined private domain knowledge bases, wherein the plurality of private domain knowledge bases include a general business knowledge base, a policy clause knowledge base, a legal aid knowledge base, and a service supervision knowledge base; performing a Retrieval-Augmented Generation (RAG) retrieval to retrieve at least one document fragment related to the service consultation request from the target knowledge base; and generating a natural language response to the service consultation request based on the retrieved at least one document fragment using a large language model.

[0007] In some embodiments of this disclosure, the method further includes: assessing the confidence level of the natural language response; and outputting the natural language response in response to the confidence level of the natural language response being greater than or equal to a predetermined threshold.

[0008] In some embodiments of this disclosure, the method further includes: assessing the confidence level of the natural language response; and, in response to the confidence level of the natural language response being less than a predetermined threshold, manually reviewing the natural language response or regenerating the natural language response.

[0009] In some embodiments of this disclosure, the method further includes: determining the confidence level of the natural language response based at least on the degree of consistency between the natural language response and at least one document fragment.

[0010] In some embodiments of this disclosure, generating a routing instruction based on the identified intent classification includes: matching the identified intent classification with intent classifications in a predefined business rule base, and generating a routing instruction based on the matching result.

[0011] In some embodiments of this disclosure, generating routing instructions based on the identified intent classification further includes: in response to a mismatch between the intent classification and the intent classification in the business rule base, generating routing instructions based on the context of the service inquiry request through a large language model.

[0012] In some embodiments of this disclosure, the target knowledge base includes multiple private domain knowledge bases, and generating routing instructions through a large language model includes: generating routing instructions for coordinating and invoking multiple private domain knowledge bases in logical order through a large language model; and the method further includes: routing service inquiry requests to multiple private domain knowledge bases in logical order according to the routing instructions.

[0013] In some embodiments of this disclosure, generating a natural language response to a service inquiry request based on at least one retrieved document fragment using a large language model includes: integrating document fragments retrieved from multiple private domain knowledge bases using a large language model; and generating a natural language response based on the integrated document fragments.

[0014] In some embodiments of this disclosure, identifying the intent classification of a service inquiry request includes: in response to receiving a service inquiry request, performing semantic parsing and entity recognition on the service inquiry request using a large language model to obtain the intent of the service inquiry request and at least one entity; and identifying the intent classification based on the intent, at least one entity, and the context of the service inquiry request.

[0015] In some embodiments of this disclosure, at least one entity includes at least one of the following: airline name, flight number, legal provision number, and business type keyword.

[0016] In some embodiments of this disclosure, the method further includes: using a large language model fine-tuned for civil aviation text to perform semantic parsing and entity recognition on the service inquiry request to obtain intent and at least one entity, wherein the fine-tuning includes fine-tuning using Low-Rank Adaptation of Large Language Models (LoRA) technique.

[0017] In some embodiments of this disclosure, the context of a service inquiry request includes: historical dialogue state information generated during a continuous multi-turn dialogue, wherein the historical dialogue state information includes: historical service inquiry requests, natural language responses to historical service inquiry requests, at least one identified entity, and intent classification of historical service inquiry requests.

[0018] In some embodiments of this disclosure, the method further includes: storing and configuring industry standard information, including flight schedules, baggage check-in, refund and change procedures, and boarding procedures, as a regular business knowledge base; storing and configuring policy documents formulated by airlines or airports as a policy clause knowledge base; storing and configuring laws and regulations issued by national legislative bodies and related to civil aviation services as a legal aid knowledge base; and / or storing and configuring historical user complaint data as a service supervision knowledge base.

[0019] In some embodiments of this disclosure, performing a search-enhanced generative retrieval (RAG) to retrieve at least one document fragment related to a service inquiry request from a target knowledge base includes: converting the service inquiry request into a query vector; and retrieving a first set of document fragments as at least one document fragment by calculating the similarity between the query vector and a pre-stored document fragment vector in the target knowledge base, wherein the method for calculating the similarity includes at least one of a cosine similarity calculation method, an inner product similarity calculation method, and an Euclidean distance similarity calculation method.

[0020] In some embodiments of this disclosure, performing a search-enhanced generative algorithm (RAG) retrieval to retrieve at least one document fragment related to a service inquiry request from a target knowledge base further includes: converting the service inquiry request into a query vector; retrieving a first set of document fragments by calculating the similarity between the query vector and a pre-stored document fragment vector in the target knowledge base; extracting keywords from the service inquiry request; performing a full-text search in the target knowledge base based on the keywords to retrieve a second set of document fragments; and merging and reordering the first set of document fragments and the second set of document fragments to filter out at least one document fragment.

[0021] In some embodiments of this disclosure, the method further includes a step of constructing a private domain knowledge base. The construction step includes: preprocessing heterogeneous data from different sources, wherein the heterogeneous data includes structured data, semi-structured data, and unstructured data, and the preprocessing includes at least one of parsing processing, format conversion processing, and information extraction processing; standardizing and cleaning the preprocessed data, wherein the standardization and cleaning includes at least one of denoising processing, deduplication processing, format unification processing, and data normalization processing; and converting the cleaned data into a vector representation and establishing an index for retrieval of the vector representation to construct the private domain knowledge base.

[0022] In some embodiments of this disclosure, the method further includes: collecting user feedback data on natural language responses; and fine-tuning a large language model based on historical interaction data including service inquiry requests, at least one document fragment, natural language responses, and feedback data.

[0023] According to a second aspect of this disclosure, an interactive system for service consultation is provided, the system comprising: an intent recognition module configured to recognize an intent classification of a service consultation request; a routing decision module configured to generate a routing instruction based on the identified intent classification to route the service consultation request to a target knowledge base determined by at least one of a plurality of predefined private domain knowledge bases, wherein the plurality of private domain knowledge bases include a general business knowledge base, a policy clause knowledge base, a legal aid knowledge base, and a service supervision knowledge base; a retrieval module configured to perform a retrieval enhancement generation (RAG) retrieval to retrieve at least one document fragment related to the service consultation request from the target knowledge base; and a response generation module configured to generate a natural language response to the service consultation request based on the retrieved at least one document fragment using a large language model.

[0024] In some embodiments of this disclosure, the system further includes: a confidence assessment module configured to assess the confidence of a natural language response; and to output a natural language response in response to the confidence of the natural language response being greater than or equal to a predetermined threshold.

[0025] In some embodiments of this disclosure, the system further includes: a confidence assessment module configured to assess the confidence of a natural language response; and in response to a natural language response having a confidence level less than a predetermined threshold, manually reviewing the natural language response or regenerating the natural language response.

[0026] In some embodiments of this disclosure, the confidence assessment module is further configured to determine the confidence level of the natural language response based at least on the degree of consistency between the natural language response and at least one document fragment.

[0027] In some embodiments of this disclosure, the routing decision module is further configured to: match the identified intent categories with intent categories in a predefined business rule base, and generate routing instructions based on the matching results.

[0028] In some embodiments of this disclosure, the routing decision module is also configured to: generate routing instructions based on a large language model in response to a mismatch between the intent classification and the intent classification in the business rule base.

[0029] In some embodiments of this disclosure, the target knowledge base includes multiple private domain knowledge bases, and the routing decision module is further configured to: generate routing instructions for coordinating and invoking the multiple private domain knowledge bases in logical order through a large language model; and route service inquiry requests to the multiple private domain knowledge bases in logical order according to the routing instructions.

[0030] In some embodiments of this disclosure, the response generation module is further configured to: integrate document fragments retrieved from multiple private domain knowledge bases using a large language model; and generate a natural language response based on the integrated document fragments.

[0031] In some embodiments of this disclosure, the intent recognition module is further configured to: in response to receiving a service inquiry request, perform semantic parsing and entity recognition on the service inquiry request using a large language model to obtain the intent of the service inquiry request and at least one entity; and identify intent classification based on the intent, at least one entity, and the context of the service inquiry request.

[0032] In some embodiments of this disclosure, at least one entity includes at least one of the following: airline name, flight number, legal provision number, and business type keyword.

[0033] In some embodiments of this disclosure, the intent recognition module is further configured to: perform semantic parsing and entity recognition on service inquiry requests using a large language model fine-tuned for civil aviation text, to obtain intent and at least one entity, wherein fine-tuning includes fine-tuning using low-rank adaptive LoRA technology.

[0034] In some embodiments of this disclosure, the context of a service inquiry request includes: historical dialogue state information generated during a continuous multi-turn dialogue, wherein the historical dialogue state information includes: historical service inquiry requests, natural language responses to historical service inquiry requests, at least one identified entity, and intent classification of historical service inquiry requests.

[0035] In some embodiments of this disclosure, the routine business knowledge base includes industry standard information on flight schedules, baggage check-in, refund and change procedures, and boarding procedures; the policy and clause knowledge base includes policy documents developed by airlines or airports; the legal aid knowledge base includes laws and regulations issued by national legislatures that are related to civil aviation services; and / or the service supervision knowledge base includes historical user complaint data.

[0036] In some embodiments of this disclosure, the retrieval module is further configured to: convert a service inquiry request into a query vector; and retrieve a first set of document fragments as at least one document fragment by calculating the similarity between the query vector and a pre-stored document fragment vector in the target knowledge base, wherein the method for calculating the similarity includes at least one of the following: cosine similarity calculation method, inner product similarity calculation method, and Euclidean distance similarity calculation method.

[0037] In some embodiments of this disclosure, the retrieval module is further configured to: convert a service inquiry request into a query vector; retrieve a first set of document fragments by calculating the similarity between the query vector and a pre-stored document fragment vector in a target knowledge base; extract keywords from the service inquiry request; perform a full-text search in the target knowledge base based on the keywords to retrieve a second set of document fragments; and merge and reorder the first set of document fragments and the second set of document fragments to filter out at least one document fragment.

[0038] In some embodiments of this disclosure, the system further includes a private domain knowledge base construction module, configured to: preprocess heterogeneous data from different sources, wherein the heterogeneous data includes structured data, semi-structured data, and unstructured data, and the preprocessing includes at least one of parsing processing, format conversion processing, and information extraction processing; perform standardization cleaning on the preprocessed data, wherein the standardization cleaning includes at least one of denoising processing, deduplication processing, format unification processing, and data normalization processing; and convert the cleaned data into a vector representation and build an index for retrieval on the vector representation to construct a private domain knowledge base.

[0039] In some embodiments of this disclosure, the system also includes a fine-tuning module configured to: collect user feedback data on natural language responses; and fine-tune the large language model based on historical interaction data including service inquiry requests, at least one document fragment, natural language responses, and feedback data.

[0040] According to a third aspect of this disclosure, an electronic device is provided, the electronic device including a processor that can be used to implement the interactive method for service consultation in the first aspect and any possible implementation of the first aspect.

[0041] According to a fourth aspect of this disclosure, a non-volatile computer-readable storage medium is provided, on which computer program instructions are stored, which, when executed by a processor, implement the interactive method for service consultation described in the first aspect and any possible implementation thereof.

[0042] According to a fifth aspect of this disclosure, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the steps of the interactive method for service consultation described in the first aspect and any possible implementation of the first aspect.

[0043] The interactive methods and systems, electronic devices, readable storage media, and computer program products for service consultation provided by the embodiments of this disclosure achieve accuracy, compliance, professionalism, and efficiency in responding to service consultations in professional service consultation fields such as civil aviation by constructing an architecture that deeply integrates modular business routing and RAG.

[0044] It should be understood that the description in this section is not intended to identify key or important features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description.

[0045] Other features and aspects of this disclosure will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0046] Other features, objects, and advantages of this disclosure will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0047] Figures 1 to 9 These are flowcharts of an interactive method for service consultation provided according to exemplary embodiments of this disclosure;

[0048] Figure 10 This is a block diagram of an interactive system for service consultation provided according to exemplary embodiments of this disclosure; and

[0049] Figure 11 This is a schematic block diagram of an electronic device provided according to an exemplary embodiment of the present disclosure. Detailed Implementation

[0050] The various exemplary embodiments, features, and aspects of this disclosure will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0051] The term “exemplary” as used herein means “serving as an example, implementation method, or illustration.” Any implementation method described herein as “exemplary” is not necessarily to be construed as superior to or better than other implementation methods.

[0052] Furthermore, to better illustrate this disclosure, numerous specific details are set forth in the following detailed description. Those skilled in the art will understand that this disclosure can be practiced without certain specific details. In some instances, systems, means, components, and circuits well known to those skilled in the art have not been described in detail in order to highlight the main points of this disclosure.

[0053] In highly specialized service consulting fields such as civil aviation, railways, hotels, and finance, existing solutions mainly rely on traditional human customer service responses or automated intelligent response systems. Human response models are inherently inefficient and costly; while existing automated intelligent response systems, although capable of handling high-frequency, standardized service inquiries, have weak semantic understanding capabilities and struggle to deeply analyze complex, cross-domain user intents, resulting in insufficient response accuracy.

[0054] Existing automated intelligent response systems are mainly based on two technical approaches. The first approach is based on rule engines or keyword matching mechanisms. While these systems can efficiently handle structured, high-frequency service inquiries, they lack deep semantic understanding capabilities and cannot effectively parse complex user requests that contain multiple intentions and cross-domain issues. For example, when faced with complex inquiry scenarios involving "flight delay compensation standards," "airline internal policies," and "the applicability of the Consumer Rights Protection Law," these systems struggle to accurately identify the user's multiple intentions, resulting in narrow response coverage, low accuracy, and severely impacting user experience.

[0055] The second technical approach directly calls a general-purpose Large Language Model (LLM) to generate responses. While this type of model possesses excellent natural language generation and contextual understanding capabilities, it has significant limitations in specialized service consulting scenarios: First, the knowledge stored in the model parameters may be incomplete or outdated, easily leading to "hallucination" and resulting in factual errors in the generated responses; second, the model lacks the ability to effectively perceive and align with external knowledge such as industry standards, policies and regulations, and historical user complaint data, making it difficult for its generated content to maintain consistency with authoritative knowledge sources.

[0056] Therefore, existing technological solutions struggle to maintain the efficiency of natural language interaction while simultaneously ensuring factual accuracy, domain expertise, and regulatory compliance when dealing with specialized service inquiries. This has become a core technological bottleneck restricting the in-depth application of intelligent customer service technology in professional fields. There is an urgent need for an innovative technological solution that can deeply integrate the cognitive capabilities of large language models with domain-specific, real-time updated authoritative knowledge sources to systematically address the aforementioned problems.

[0057] To at least address the aforementioned issues, the embodiments of this disclosure provide interactive methods and systems, electronic devices, readable storage media, and computer program products for service consultation. By constructing an architecture that deeply integrates modular business routing with RAG, it achieves accuracy, compliance, professionalism, and efficiency in responding to service consultations in professional service consultation fields such as civil aviation.

[0058] Some embodiments of this disclosure provide an interactive method for service consultation. Figure 1 This is a flowchart of an interactive method 1000 for service consultation provided according to an exemplary embodiment of this disclosure.

[0059] like Figure 1 As shown, the interaction method 1000 for service consultation includes:

[0060] Step S1: Identify the intent classification of the service inquiry request.

[0061] Step S2: Based on the identified intent classification, a routing instruction is generated to route the service consultation request to a target knowledge base determined by at least one of a plurality of predefined private domain knowledge bases, including a general business knowledge base, a policy and provision knowledge base, a legal aid knowledge base, and a service supervision knowledge base.

[0062] Step S3: Perform a RAG search to retrieve at least one document fragment related to the service inquiry request from the target knowledge base.

[0063] Step S4: Based on at least one retrieved document fragment, generate a natural language response to the service inquiry request using a large language model.

[0064] In the service consultation interaction scheme defined in the embodiments of this disclosure, accurate intent classification ensures that the service consultation request can be correctly understood. Therefore, based on the identified intent classification, the service consultation request is routed to different predefined private domain knowledge bases, which can avoid irrelevant answers caused by misunderstanding of intent from the source.

[0065] By managing authoritative data from different fields and sources, such as routine business knowledge, policy and regulation knowledge, legal aid knowledge, and service supervision knowledge, in different predefined private domain knowledge bases, or by building mutually independent domain-specific knowledge bases, "routine business consultations" can be routed to the routine business knowledge base, and "policy and regulation consultations" can be routed to at least one of the policy and regulation knowledge bases and the legal aid knowledge base. This ensures that every service consultation request is answered by the most "professional" knowledge system and avoids inefficient full-scale retrieval in a single, complex knowledge base.

[0066] By introducing RAG technology, the large language model is endowed with the ability to retrieve prior knowledge, enabling it to efficiently search private domain knowledge bases outside the model itself. This improves the accuracy and relevance of responses to service inquiry requests. RAG technology combines generative models and information retrieval mechanisms, allowing for the reference of real-time background information during response generation, providing more targeted answers for each service inquiry request. Furthermore, responses to service inquiry requests can be confined to a pre-defined, audited private knowledge base, mitigating the risk of generating content that contradicts current policies and regulations based on a general large language model from the data source.

[0067] In addition, based on at least one retrieved document fragment, a natural language response to the service inquiry request is generated through a large language model. Natural language is characterized by fluency, structure, and ease of understanding. Therefore, the large language model generates a natural language response while remaining faithful to the retrieved document fragment. This inherits the professionalism and accuracy advantages brought by the aforementioned steps, and can also transform lengthy and obscure original text into a direct and clear response, enabling users to quickly obtain the information they need and improving the overall efficiency of the consultation process.

[0068] Figure 2 This is a flowchart of an interactive method 1000 for service consultation provided according to an exemplary embodiment of this disclosure.

[0069] Specifically, such as Figure 2 As shown, in some embodiments of this disclosure, the interactive method 1000 for service consultation may further include:

[0070] Step S5: Assess the confidence level of the natural language response; and output the natural language response in response to the confidence level of the natural language response being greater than or equal to a predetermined threshold.

[0071] To ensure the accuracy and reliability of the generated responses, some embodiments of this disclosure employ a confidence assessment mechanism. During the response generation phase, the large language model can generate a natural language response based on the service inquiry request and retrieved relevant document fragments, and determine the credibility of the generated response through a self-assessment mechanism; this can be understood as the confidence level of the natural language response.

[0072] Alternatively, the confidence level of a natural language response can be determined at least based on the degree of consistency between the natural language response and the retrieved relevant document fragments. For example, deep semantic analysis technology can be used to quantify the degree of consistency between the natural language response and the retrieved relevant document fragments, output a confidence score, and output the natural language response in response to a confidence score greater than or equal to a predetermined threshold. This confidence assessment mechanism effectively ensures that interactive technologies used for service consultation can provide high-quality, compliant responses when faced with complex and professional service consultation requests, avoiding potential legal risks or user misunderstandings, and improving the reliability and compliance of the entire technical solution.

[0073] Building upon this, the confidence level of a natural language response can be determined based on its interpretability and completeness. Optionally, in some implementations, a natural language response may have a high degree of interpretability due to its explicit citation of specific historical regulations or policy clause numbers. Furthermore, the completeness of the natural language response can be determined by assessing whether it comprehensively covers all aspects of the service inquiry request and conforms to preset response specifications and structured requirements. By comprehensively evaluating the confidence level, interpretability, and completeness of the natural language response, a unified confidence score can ultimately be generated.

[0074] Figure 3 This is a flowchart of an interactive method 1000 for service consultation provided according to an exemplary embodiment of this disclosure.

[0075] like Figure 3 As shown, in some embodiments of this disclosure, the interactive method 1000 for service consultation may further include:

[0076] Step S6: Assess the confidence level of the natural language response; and in response to the confidence level of the natural language response being less than a predetermined threshold, manually review the natural language response or regenerate the natural language response.

[0077] To ensure the accuracy and reliability of the generated response, in some embodiments of this disclosure, in response to a natural language response score being less than a predetermined threshold, the natural language response is manually reviewed, or steps S1 to S4 are re-executed to regenerate the natural language response.

[0078] As an option, a predetermined threshold can be set based on experimental data. For example, when the confidence score of the generated natural language response is detected to be below 60%, a mechanism for automatic manual review or regeneration of the natural language response will be automatically triggered. If the predetermined threshold is not reached after n rounds, no response will be provided to ensure that the final response is accurate and compliant, where n is a positive integer greater than 1.

[0079] By assessing the confidence level of natural language responses, the generation of erroneous or illusory answers can be prevented. This confidence assessment mechanism effectively ensures that interactive technology solutions used for service inquiries can provide high-quality, compliant responses when faced with complex and professional service inquiries, avoiding potential legal risks or user misunderstandings, and improving the reliability and compliance of the technology solutions.

[0080] Optionally, this predetermined threshold can be derived through multiple rounds of experimentation and data analysis to flexibly adapt to the needs of different fields and service consultation request types. Furthermore, the evaluation accuracy can be continuously optimized through historical consultation data and user feedback, thereby adjusting the predetermined threshold in real time according to actual conditions to ensure high-quality responses are provided to users.

[0081] Figure 4 This is a flowchart of an interactive method 1000 for service consultation provided according to an exemplary embodiment of this disclosure.

[0082] like Figure 1 and Figure 4 As shown, in some embodiments of this disclosure, step S1, identifying the intent classification of the service inquiry request, may include:

[0083] Step S1-1: In response to receiving a service inquiry request, the service inquiry request is semantically parsed and entity identified using a large language model to obtain the intent of the service inquiry request and at least one entity.

[0084] Step S1-2: Identify the intent classification based on the intent, at least one entity, and the context of the service inquiry request.

[0085] By leveraging the powerful natural language understanding capabilities of large language models, deep semantic analysis is performed on users' service inquiry requests. This process goes beyond simple keyword matching, achieving a complex decoding of the user's true intent. For example, when a user enters the service inquiry request, "My flight is delayed by 4 hours, what can I get?", the interaction method provided by the embodiments of this disclosure can understand that their intent is "to inquire about flight delay compensation standards," rather than remaining at the superficial level of "flight delay." Therefore, in the service inquiry interaction scheme defined by the embodiments of this disclosure, accurate intent classification ensures that the service inquiry request can be correctly understood, avoiding irrelevant answers caused by misunderstanding of intent from the outset.

[0086] Entity recognition can be understood as accurately locating and extracting key and meaningful objective information fragments from the original service consultation request text. These objective information fragments can be understood as entities from the service consultation request, which constitute the core parameters for refining and defining the service consultation request.

[0087] Furthermore, taking the civil aviation consulting vertical as an example, entities have very specific and professional connotations. Entities may include at least one of the following: airline name, flight number, legal article number, and business type keywords. Optionally, business type keywords may include industry terms related to specific businesses, such as "baggage check-in," "refund and change," and "delay compensation."

[0088] By identifying highly specialized entities such as regulatory clause numbers, we can improve our understanding of relevant technical terms when generating responses to service inquiries. This ensures that we accurately capture the core elements of each service inquiry, guaranteeing the professionalism of subsequent processing and the accuracy of our answers from the outset.

[0089] Furthermore, in some embodiments of this disclosure, a large language model fine-tuned for civil aviation text can be used to perform semantic parsing and entity recognition on service inquiry requests to obtain intent and at least one entity, wherein fine-tuning may include fine-tuning using low-rank adaptive LoRA technology.

[0090] Based on LoRA technology, when a large language model learns a new task or adapts to a new domain, such as learning civil aviation text, the changes required to the model's weight matrix are "low-rank." During fine-tuning, LoRA does not modify any of the original parameters of the large language model (i.e., it freezes the pre-trained weights). Instead, it endows the large language model with new capabilities by injecting and training these tiny "LoRA adapter parameters." Therefore, when deploying interactive technology solutions for service consultation, combining the pre-trained base large language model with trained LoRA weights can yield a model specialized for a specific domain.

[0091] Optionally, in some embodiments of this disclosure, the context of a service inquiry request may include: historical dialogue state information generated in a continuous multi-turn dialogue, wherein the historical dialogue state information includes: historical service inquiry requests, natural language responses to historical service inquiry requests, at least one identified entity, and intent classification of historical service inquiry requests.

[0092] The interactive technology solution for service consultation provided by the embodiments of this disclosure can structurally track and maintain the entire lifecycle of a conversation. When a user uses pronouns or ellipses in a new round of dialogue, their true meaning can be deciphered by querying "historical dialogue status information." This avoids users repeatedly using key information, providing a smooth, natural, and efficient user experience, making the interaction closer to human-to-human dialogue. Furthermore, when a user's intent changes significantly, the solution can sensitively capture the shift in topic by comparing the current intent with "historical intent categories," and dynamically adjust the processing path in subsequent processes. This ensures that the technical solution can flexibly follow the user's thought process while accurately maintaining the main line of dialogue, achieving context-aware adaptive interaction. Based on this, it provides a foundation for deep reasoning and personalized services for complex consultations. The complete context of the service consultation request allows the response to the request to provide a more in-depth and personalized answer based on the user's entire consultation process, rather than simply answering the current request in isolation.

[0093] Figure 5 This is a flowchart of an interactive method 1000 for service consultation provided according to an exemplary embodiment of this disclosure.

[0094] like Figure 5 As shown, in some embodiments of this disclosure, the interactive method 1000 for service consultation may further include:

[0095] Step S0: Predefine a private domain knowledge base.

[0096] Specifically, industry standard information including flight schedules, baggage check-in, refund and change procedures, and boarding procedures can be stored and configured as a regular business knowledge base; policy documents formulated by airlines or airports can be stored and configured as a policy clause knowledge base; laws and regulations issued by national legislative bodies and related to civil aviation services can be stored and configured as a legal aid knowledge base; and / or historical user complaint data can be stored and configured as a service supervision knowledge base.

[0097] By managing authoritative data from different fields and sources, such as routine business knowledge, policy and regulation knowledge, legal aid knowledge, and service supervision knowledge, in different predefined private domain knowledge bases, or by building mutually independent domain-specific knowledge bases, "routine business consultations" can be routed to the routine business knowledge base, and "policy and regulation consultations" can be routed to at least one of the policy and regulation knowledge bases and the legal aid knowledge base. This ensures that every service consultation request is answered by the most "professional" knowledge system and avoids inefficient full-scale retrieval in a single, complex knowledge base.

[0098] "Multiple independent domain-specific knowledge bases" can be understood as logically isolated and independently managed. For example, different private domain knowledge bases may correspond to different physical storage addresses; or, different private domain knowledge bases may be isolated from each other by logical identifiers within the same physical storage address. It should be noted that this disclosure does not limit the specific implementation of "multiple independent domain-specific knowledge bases," and can be based on the two or more specific physical implementation methods mentioned above. For example, each private domain knowledge base may include multiple sub-bases, which are stored independently of each other.

[0099] The isolation architecture ensures clear "data boundaries." When a user inquires about "Air China's baggage policy," the routing mechanism will search only within the "Air China" data partition of the "Policy Terms Knowledge Base," without mixing in "China Eastern's" policies. It will also never mistakenly return the provisions of the Consumer Rights Protection Law as direct regulations from Air China. Therefore, by building mutually independent domain-specific knowledge bases, cross-contamination between knowledge from different domains and sources is completely avoided, ensuring the high accuracy and domain professionalism of the generated responses.

[0100] Furthermore, the construction of a private domain knowledge base may include: preprocessing heterogeneous data from different sources, wherein the heterogeneous data includes structured data, semi-structured data, and unstructured data, and the preprocessing includes at least one of parsing, format conversion, and information extraction; standardizing and cleaning the preprocessed data, wherein the standardization and cleaning includes at least one of denoising, deduplication, format unification, and data normalization; and converting the cleaned data into a vector representation and building an index for retrieval of the vector representation, thereby constructing a private domain knowledge base.

[0101] Authoritative data from different domains and sources can be managed in different predefined private domain knowledge bases. This authoritative data can include heterogeneous data from various sources, such as structured, semi-structured, and unstructured data. To ensure that the information in the private domain knowledge base is reliable, consistent, and uniformly formatted, the heterogeneous data can undergo the aforementioned preprocessing and standardization cleaning during the construction of the private domain knowledge base, and the cleaned data can be converted into vector representations. This minimizes the interference of junk, erroneous, and contradictory information on the generation process from the data source, directly improving the factual accuracy and reliability of the final generated response.

[0102] Furthermore, data vectorization and index building can upgrade the existing "keyword matching" to "semantic similarity matching." By calculating the distance in the semantic space between the query vector converted from the service request and the document vector converted from authoritative data, the most relevant document fragments can be found, even if the keywords in the service request do not appear in the most relevant document fragments. Therefore, converting the cleaned data into a vector representation can turn the private domain knowledge base into a vector database. By improving retrieval speed through indexes used for searching, the efficiency of responding to service requests and the accuracy of the answers can be ensured.

[0103] Figure 6 This is a flowchart of an interactive method 1000 for service consultation provided according to an exemplary embodiment of this disclosure.

[0104] like Figure 6 As shown, in some embodiments of this disclosure, step S2 generates a routing instruction based on the identified intent classification to route the service inquiry request to a target knowledge base determined by at least one of a plurality of predefined private domain knowledge bases, wherein the plurality of private domain knowledge bases may include a general business knowledge base, a policy and provision knowledge base, a legal aid knowledge base, and a service regulatory knowledge base, which may include:

[0105] Step S2-1: Match the identified intent categories with the intent categories in the predefined business rule base, and generate routing instructions based on the matching results.

[0106] Specifically, taking the vertical field of civil aviation consulting as an example, a predefined business rule base can be established based on the rules of civil aviation business logic. This predefined business rule base can explicitly define the mapping relationship between at least one of multiple intent categories, keywords, and patterns and a private domain knowledge base. Once an intent category is identified, it can be searched within the business rule base. In response to a match between the identified intent category and an intent category in the predefined business rule base, a routing instruction can be generated, sending the request to the private domain knowledge base (i.e., the target knowledge base) specified by the rule. For example, a predefined rule in the business rule base can specify that, in response to the identification of the keyword "baggage check-in," the identified intent category should be assigned to the regular business knowledge base.

[0107] By introducing a deterministic rule logic predefined business rule base, an efficient, stable, and reliable processing baseline can be established, complementing the probabilistic large language model decision-making, resulting in significant improvements in response efficiency, service reliability, and resource utilization.

[0108] In some embodiments of this disclosure, step S2 generates a routing instruction based on the identified intent classification to route the service inquiry request to a target knowledge base determined by at least one of a plurality of predefined private domain knowledge bases, wherein the plurality of private domain knowledge bases may include a general business knowledge base, a policy clause knowledge base, a legal aid knowledge base, and a service supervision knowledge base, and may further include:

[0109] Step S2-2: In response to the mismatch between the intent classification and the intent classification in the business rule base, a routing instruction is generated based on the context of the service consultation request through the large language model.

[0110] While the business rule base is predefined, real-world service inquiry requests are often open-ended. Therefore, when dealing with novel or complex inquiries that have never been encountered before, routing instructions can be generated based on the context of the service inquiry request using a large language model. The large language model fully considers the "context of the service inquiry request" during decision-making, ensuring coherent routing decisions. By fully understanding the evolution of referents and intents during multiple turns of dialogue with the user, it generates routing instructions consistent with the dialogue history logic, thus adapting to changes in business requirements.

[0111] Based on this, when the target knowledge base includes multiple private domain knowledge bases, generating routing instructions through a large language model may include: generating routing instructions through the large language model to coordinate and invoke multiple private domain knowledge bases in logical order; and routing service inquiry requests to multiple private domain knowledge bases in logical order according to the routing instructions.

[0112] For example, suppose a user submits a service inquiry request: "I missed my connecting international flight due to a flight delay. Who should I seek compensation from, and should it be based on domestic regulations or international standards?" The reasoning and routing instructions generated by the large language model can include two aspects: firstly, determining the responsible party and the domestic basis for compensation; and secondly, understanding the special characteristics of international connecting flights. Based on this, the large language model can generate a logical sequence and call multiple private domain knowledge bases according to this logical sequence. For example, the large language model infers that the response to the above request must first clarify the airline's basic responsibilities and obligations under domestic law before further judging how this responsibility evolves and accumulates in the context of international connecting flights. Therefore, the logical sequence is determined to be a typical progressive logical sequence. Subsequently, according to the progressive logical sequence, the policy clause knowledge base can be called first to retrieve the specific compensation policies for "flight delays" and "missed connecting flights" in the General Conditions of Transport of domestic airlines. After obtaining the results, the legal aid knowledge base can be called to retrieve the mandatory provisions of the Civil Aviation Law and relevant international conventions regarding the rights of international air passengers and the responsibilities of airlines. This ensures that the final response is not a simple compilation of information, but a logically coherent, progressively developing, and rigorously argued in-depth solution, greatly improving the accuracy and professionalism of responses in complex professional scenarios.

[0113] Alternatively, step S4, which generates a natural language response to a service inquiry request based on at least one retrieved document fragment using a large language model, may include: integrating document fragments retrieved separately from multiple private domain knowledge bases using a large language model; and generating a natural language response based on the integrated document fragments.

[0114] For example, when the target knowledge base includes multiple private domain knowledge bases, the large language model can route service inquiry requests to multiple private domain knowledge bases in a logical order according to routing instructions, achieving database-specific retrieval. RAG retrieval can be performed in each private domain knowledge base to retrieve document fragments related to the service inquiry request from each knowledge base. Then, the large language model performs semantic integration and association on the document fragments retrieved from the multiple knowledge bases to generate a coherent, complete, and natural language response specifically for the service inquiry request. This enables deep fusion of cross-domain knowledge; by collaboratively utilizing multiple authoritative knowledge sources and performing deep semantic integration by the large language model, the comprehensiveness and high professionalism of the response can be ensured.

[0115] Figure 7 This is a flowchart of an interactive method 1000 for service consultation provided according to an exemplary embodiment of this disclosure.

[0116] like Figure 7As shown, in some embodiments of this disclosure, step S3, which performs retrieval enhancement to generate a RAG retrieval to retrieve at least one document fragment related to the service inquiry request from the target knowledge base, may include:

[0117] Step S3-1: Convert the service consultation request into a query vector; and retrieve the first set of document fragments as at least one document fragment by calculating the similarity between the query vector and the document fragment vectors pre-stored in the target knowledge base. The method for calculating the similarity includes at least one of the following: cosine similarity calculation method, inner product similarity calculation method, and Euclidean distance similarity calculation method.

[0118] Traditional keyword matching cannot understand synonyms, near-synonyms, or the same meaning expressed in different ways. For example, if a user asks "What are the requirements for checked baggage?", while the knowledge base states "Regulations for checked baggage", keyword matching may fail because "checked baggage" and "checked baggage" are different. However, vector retrieval can understand that the two are highly semantically related, thus accurately retrieving the correct information.

[0119] Therefore, by understanding the deep semantics of the query rather than the surface words, the obstacle of word mismatch can be overcome, which greatly improves the recall and accuracy of the retrieval, lays a solid factual foundation for generating high-quality and highly relevant responses, and thus directly improves the accuracy of the final response.

[0120] In other words, data vectorization and index building can upgrade the existing "keyword matching" to "semantic similarity matching." By calculating the distance in the semantic space between the query vector converted from the service request and the document vector converted from authoritative data, the most relevant document fragment can be found, even if the keyword in the service request does not appear in the most relevant document fragment. Therefore, converting the cleaned data into a vector representation can turn the private domain knowledge base into a vector database. By improving retrieval speed through indexes used for searching, the efficiency of responding to service requests and the accuracy of the answers can be ensured.

[0121] Figure 8 This is a flowchart of an interactive method 1000 for service consultation provided according to an exemplary embodiment of this disclosure.

[0122] like Figure 8 As shown, in some other embodiments of this disclosure, step S3, which performs retrieval enhancement to generate a RAG retrieval to retrieve at least one document fragment related to the service inquiry request from the target knowledge base, may include:

[0123] Step S3-2: Convert the service inquiry request into a query vector; retrieve the first set of document fragments by calculating the similarity between the query vector and the pre-stored document fragment vectors in the target knowledge base; extract keywords from the service inquiry request; perform a full-text search in the target knowledge base based on the keywords to retrieve the second set of document fragments; and merge and reorder the first set of document fragments and the second set of document fragments to filter out at least one document fragment.

[0124] This implementation method employs a hybrid retrieval and re-ranking strategy. By performing semantic and keyword retrieval in parallel, and then intelligently fusing and filtering them, it creatively solves the inherent shortcomings of single retrieval technologies. This combination enables the interactive solution provided by the implementation method of this disclosure to comprehensively improve the breadth, accuracy, robustness, and compliance compared to a single technical path when dealing with diverse and professional service inquiries, thus ensuring the accuracy, professionalism, and reliability of the responses.

[0125] Figure 9 This is a flowchart of an interactive method 1000 for service consultation provided according to an exemplary embodiment of this disclosure.

[0126] like Figure 9 As shown, in some other embodiments of this disclosure, the interaction method 1000 for service consultation further includes: collecting user feedback data on natural language responses; and fine-tuning a large language model based on historical interaction data including service consultation requests, at least one document fragment, natural language responses, and feedback data.

[0127] Specifically, upon receiving a service inquiry request, the intent category of the request can be identified. Based on the identified intent category, the large language module can generate routing instructions to route the service inquiry request to a target knowledge base, which can be at least one of multiple predefined private domain knowledge bases. For example, multiple private domain knowledge bases may include a general business knowledge base, a policy clause knowledge base, a legal aid knowledge base, and a service regulation knowledge base. After routing to the target knowledge base, a retrieval enhancement generation (RAG) retrieval can be performed to retrieve at least one document fragment related to the service inquiry request from the target knowledge base; and based on the retrieved document fragment, a natural language response to the service inquiry request is generated using the large language model.

[0128] Optionally, after generating a natural language response, a confidence level assessment can be performed on the natural language response. If the confidence level of the natural language response is greater than or equal to a predetermined threshold, the natural language response is output. Alternatively, if the confidence level of the natural language response is high, the natural language response can be output. If the confidence level of the natural language response is less than the predetermined threshold, the natural language response is manually reviewed or regenerated. Alternatively, if the confidence level of the natural language response is low, the natural language response is manually reviewed or regenerated.

[0129] Optionally, after performing the above steps, user feedback data on natural language responses can be collected, and the large language model can be fine-tuned based on historical interaction data including service inquiry requests, at least one document fragment, natural language responses, and feedback data. This self-optimizing closed loop brings dynamic growth capabilities to the interactive technology solution for service inquiries provided by the embodiments of this disclosure, continuously reinforcing and optimizing itself over time.

[0130] Some embodiments of this disclosure provide an interactive system for service consultation. Figure 10 This is a block diagram of an interactive system 2000 for service consultation provided according to an exemplary embodiment of this disclosure.

[0131] like Figure 10 As shown, the interactive system 2000 for service consultation provided by the embodiments of this disclosure may include: an intent recognition module 100, a routing decision module 200, a retrieval module 300, and a response generation module 400, wherein the intent recognition module 100 is configured to identify the intent category of the service consultation request; the routing decision module 200 is configured to generate a routing instruction based on the identified intent category to route the service consultation request to a target knowledge base determined by at least one of a plurality of predefined private domain knowledge bases, wherein the plurality of private domain knowledge bases include a general business knowledge base, a policy clause knowledge base, a legal aid knowledge base, and a service supervision knowledge base; the retrieval module 300 is configured to perform a retrieval enhancement generation (RAG) retrieval to retrieve at least one document fragment related to the service consultation request from the target knowledge base; and the response generation module 400 is configured to generate a natural language response to the service consultation request based on the retrieved at least one document fragment using a large language model.

[0132] In the service consultation interaction scheme defined in the embodiments of this disclosure, accurate intent classification ensures that the service consultation request can be correctly understood. Therefore, based on the identified intent classification, the service consultation request is routed to different predefined private domain knowledge bases, which can avoid irrelevant answers caused by misunderstanding of intent from the source.

[0133] By managing authoritative data from different fields and sources, such as routine business knowledge, policy and regulation knowledge, legal aid knowledge, and service supervision knowledge, in different predefined private domain knowledge bases, or by building mutually independent domain-specific knowledge bases, "routine business consultations" can be routed to the routine business knowledge base, and "policy and regulation consultations" can be routed to at least one of the policy and regulation knowledge bases and the legal aid knowledge base. This ensures that every service consultation request is answered by the most "professional" knowledge system and avoids inefficient full-scale retrieval in a single, complex knowledge base.

[0134] By introducing RAG technology, the large language model is endowed with the ability to retrieve prior knowledge, enabling it to efficiently search private domain knowledge bases outside the model itself. This improves the accuracy and relevance of responses to service inquiry requests. RAG technology combines generative models and information retrieval mechanisms, allowing for the reference of real-time background information during response generation, providing more targeted answers for each service inquiry request. Furthermore, responses to service inquiry requests can be confined to a pre-defined, audited private knowledge base, mitigating the risk of generating content that contradicts current policies and regulations based on a general large language model from the data source.

[0135] In addition, based on at least one retrieved document fragment, a natural language response to the service inquiry request is generated through a large language model. Natural language is characterized by fluency, structure, and ease of understanding. Therefore, the large language model generates a natural language response while remaining faithful to the retrieved document fragment. This inherits the professionalism and accuracy advantages brought by the aforementioned steps, and can also transform lengthy and obscure original text into a direct and clear response, enabling users to quickly obtain the information they need and improving the overall efficiency of the consultation process.

[0136] Specifically, in some embodiments of this disclosure, the interactive system 2000 for service consultation may further include: a confidence assessment module (not shown) configured to assess the confidence of a natural language response; and to output a natural language response in response to the confidence of the natural language response being greater than or equal to a predetermined threshold.

[0137] In addition, in some other embodiments of this disclosure, the confidence assessment module responds to a natural language response whose confidence level is less than a predetermined threshold by manually reviewing the natural language response or by regenerating the natural language response.

[0138] During the response generation phase, the large language model generates a natural language response based on the service inquiry and relevant document fragments retrieved. It then uses a self-evaluation mechanism to assess the credibility of the generated response, which can be understood as the confidence level of the natural language response. By evaluating the confidence level of the natural language response, the accuracy and reliability of the generated response can be ensured.

[0139] Alternatively, the confidence assessment module can also determine the confidence level of the natural language response based at least on the degree of consistency between the natural language response and the retrieved relevant document fragments. Furthermore, it can determine the confidence level of the natural language response based on its interpretability and completeness.

[0140] Optionally, the aforementioned predetermined thresholds can be derived through multiple rounds of experimentation and data analysis to flexibly adapt to the needs of different fields and service consultation request types. Furthermore, the evaluation accuracy can be continuously optimized through historical consultation data and user feedback, thereby adjusting the predetermined thresholds in real time according to actual circumstances to ensure high-quality responses are provided to users.

[0141] Optionally, in some embodiments of this disclosure, the intent recognition module 100 may, in response to receiving a service inquiry request, perform semantic parsing and entity recognition on the service inquiry request using a large language model to obtain the intent of the service inquiry request and at least one entity; and identify the intent classification based on the intent, at least one entity, and the context of the service inquiry request.

[0142] By leveraging the powerful natural language understanding capabilities of large language models, deep semantic analysis is performed on users' service inquiry requests. This process goes beyond simple keyword matching, achieving a complex decoding of the user's true intent. Therefore, in the service inquiry interaction scheme defined in the embodiments of this disclosure, accurate intent classification ensures that service inquiry requests are correctly understood, thus avoiding irrelevant answers caused by misunderstandings of intent from the outset.

[0143] Entity recognition can be understood as accurately locating and extracting key and meaningful objective information fragments from the original service consultation request text. These objective information fragments can be understood as entities from the service consultation request, which constitute the core parameters for refining and defining the service consultation request.

[0144] Furthermore, taking the civil aviation consulting vertical as an example, entities have very specific and professional connotations. Entities may include at least one of the following: airline name, flight number, legal article number, and business type keywords. Optionally, business type keywords may include industry terms related to specific businesses, such as "baggage check-in," "refund and change," and "delay compensation."

[0145] By identifying highly specialized entities such as regulatory clause numbers, we can improve our understanding of relevant technical terms when generating responses to service inquiries. This ensures that we accurately capture the core elements of each service inquiry, guaranteeing the professionalism of subsequent processing and the accuracy of our answers from the outset.

[0146] Furthermore, in some embodiments of this disclosure, a large language model fine-tuned for civil aviation text can be used to perform semantic parsing and entity recognition on service inquiry requests to obtain intent and at least one entity, wherein fine-tuning may include fine-tuning using low-rank adaptive LoRA technology.

[0147] Based on LoRA technology, when a large language model learns a new task or adapts to a new domain, such as learning civil aviation text, the changes required to the model's weight matrix are "low-rank." During fine-tuning, LoRA does not modify any of the original parameters of the large language model (i.e., it freezes the pre-trained weights). Instead, it endows the large language model with new capabilities by injecting and training these tiny "LoRA adapter parameters." Therefore, when deploying interactive technology solutions for service consultation, combining the pre-trained base large language model with trained LoRA weights can yield a model specialized for a specific domain.

[0148] Optionally, in some embodiments of this disclosure, the context of a service inquiry request may include: historical dialogue state information generated in a continuous multi-turn dialogue, wherein the historical dialogue state information includes: historical service inquiry requests, natural language responses to historical service inquiry requests, at least one identified entity, and intent classification of historical service inquiry requests.

[0149] The interactive technology solution for service consultation provided by the embodiments of this disclosure can structurally track and maintain the entire lifecycle of a conversation. When a user uses pronouns or ellipses in a new round of dialogue, their true meaning can be deciphered by querying "historical dialogue status information." This avoids users repeatedly using key information, providing a smooth, natural, and efficient user experience, making the interaction closer to human-to-human dialogue. Furthermore, when a user's intent changes significantly, the solution can sensitively capture the shift in topic by comparing the current intent with "historical intent categories," and dynamically adjust the processing path in subsequent processes. This ensures that the technical solution can flexibly follow the user's thought process while accurately maintaining the main line of dialogue, achieving context-aware adaptive interaction. Based on this, it provides a foundation for deep reasoning and personalized services for complex consultations. The complete context of the service consultation request allows the response to the request to provide a more in-depth and personalized answer based on the user's entire consultation process, rather than simply answering the current request in isolation.

[0150] In some embodiments of this disclosure, the interactive system 2000 for service consultation may further include a private domain knowledge base construction module (not shown). The private domain knowledge base includes: a routine business knowledge base, a policy and clause knowledge base, a legal aid knowledge base, and a service supervision knowledge base. Optionally, the routine business knowledge base may include industry standard information on flight schedules, baggage check-in, refund and change procedures, and boarding procedures; the policy and clause knowledge base may include policy documents developed by airlines or airports; the legal aid knowledge base may include laws and regulations issued by national legislatures and related to civil aviation services; and / or the service supervision knowledge base may include historical user complaint data.

[0151] By managing authoritative data from different fields and sources, such as routine business knowledge, policy and regulation knowledge, legal aid knowledge, and service supervision knowledge, in different predefined private domain knowledge bases, or by building mutually independent domain-specific knowledge bases, "routine business consultations" can be routed to the routine business knowledge base, and "policy and regulation consultations" can be routed to at least one of the policy and regulation knowledge bases and the legal aid knowledge base. This ensures that every service consultation request is answered by the most "professional" knowledge system and avoids inefficient full-scale retrieval in a single, complex knowledge base.

[0152] Optionally, the private domain knowledge base construction module can be configured to: preprocess heterogeneous data from different sources, wherein the heterogeneous data includes structured data, semi-structured data, and unstructured data, and the preprocessing includes at least one of parsing, format conversion, and information extraction; perform standardization cleaning on the preprocessed data, wherein the standardization cleaning includes at least one of denoising, deduplication, format unification, and data normalization; and convert the cleaned data into a vector representation and build an index for retrieval on the vector representation to construct the private domain knowledge base.

[0153] Authoritative data from different domains and sources can be managed in different predefined private domain knowledge bases. This authoritative data can include heterogeneous data from various sources, such as structured, semi-structured, and unstructured data. To ensure that the information in the private domain knowledge base is reliable, consistent, and uniformly formatted, the heterogeneous data can undergo the aforementioned preprocessing and standardization cleaning during the construction of the private domain knowledge base, and the cleaned data can be converted into vector representations. This minimizes the interference of junk, erroneous, and contradictory information on the generation process from the data source, directly improving the factual accuracy and reliability of the final generated response.

[0154] Furthermore, data vectorization and index building can upgrade the existing "keyword matching" to "semantic similarity matching." By calculating the distance in the semantic space between the query vector converted from the service request and the document vector converted from authoritative data, the most relevant document fragments can be found, even if the keywords in the service request do not appear in the most relevant document fragments. Therefore, converting the cleaned data into a vector representation can turn the private domain knowledge base into a vector database. By improving retrieval speed through indexes used for searching, the efficiency of responding to service requests and the accuracy of the answers can be ensured.

[0155] Alternatively, the routing decision module 200 can also be configured to match the identified intent categories with intent categories in a predefined business rule base and generate routing instructions based on the matching results. By introducing a deterministic rule logic predefined business rule base, an efficient, stable, and reliable processing baseline can be established, complementing the probabilistic large language model decision-making, resulting in significant improvements in response efficiency, service reliability, and resource utilization.

[0156] Furthermore, in some embodiments of this disclosure, the routing decision module 200 may also be configured to: generate routing instructions based on a large language model in response to a mismatch between the intent classification and the intent classification in the business rule base.

[0157] Based on this, the routing decision module 200 can generate routing instructions for coordinating and calling multiple private domain knowledge bases in logical order through a large language model; and route service consultation requests to multiple private domain knowledge bases in logical order according to the routing instructions.

[0158] While the business rule base is predefined, real-world service inquiry requests are often open-ended. Therefore, when dealing with novel or complex inquiries that have never been encountered before, routing instructions can be generated based on the context of the service inquiry request using a large language model. The large language model fully considers the "context of the service inquiry request" during decision-making, ensuring coherent routing decisions. By fully understanding the evolution of referents and intents during multiple turns of dialogue with the user, it generates routing instructions consistent with the dialogue history logic, thus adapting to changes in business requirements.

[0159] Therefore, the response generation module 400 can also be configured to: integrate document fragments retrieved from multiple private domain knowledge bases through a large language model; and generate a natural language response based on the integrated document fragments.

[0160] Specifically, when the target knowledge base includes multiple private domain knowledge bases, the large language model can route service inquiry requests to these knowledge bases in a logical order according to routing instructions, achieving database-specific retrieval. RAG retrieval can be performed in each private domain knowledge base to retrieve document fragments related to the service inquiry request. Then, the large language model performs semantic integration and association on the document fragments retrieved from the multiple knowledge bases to generate a coherent, complete, and specific natural language response to the service inquiry request. This enables deep fusion of cross-domain knowledge; by collaboratively utilizing multiple authoritative knowledge sources and performing deep semantic integration by the large language model, the comprehensiveness and high professionalism of the response can be ensured.

[0161] Alternatively, the retrieval module 300 can also be configured to: convert service inquiry requests into query vectors; and retrieve a first set of document fragments as at least one document fragment by calculating the similarity between the query vector and the document fragment vectors pre-stored in the target knowledge base, wherein the method for calculating the similarity includes at least one of the following: cosine similarity calculation method, inner product similarity calculation method, and Euclidean distance similarity calculation method.

[0162] Data vectorization and index building can upgrade the existing "keyword matching" to "semantic similarity matching." By calculating the distance in the semantic space between the query vector converted from the service inquiry and the document vector converted from authoritative data, the most relevant document fragments can be found, even if the keywords in the service inquiry do not appear in the most relevant document fragments. Therefore, converting the cleaned data into a vector representation can turn the private domain knowledge base into a vector database. By improving retrieval speed through indexes used for searching, the efficiency of responding to service inquiry requests and the accuracy of the answers can be ensured.

[0163] Alternatively, the retrieval module 300 can also be configured to: convert service inquiry requests into query vectors; retrieve a first set of document fragments by calculating the similarity between the query vectors and pre-stored document fragment vectors in the target knowledge base; extract keywords from the service inquiry requests; perform full-text search in the target knowledge base based on the keywords to retrieve a second set of document fragments; and merge and reorder the first and second sets of document fragments to filter out at least one document fragment.

[0164] This implementation method employs a hybrid retrieval and re-ranking strategy. By performing semantic and keyword retrieval in parallel, and then intelligently fusing and filtering them, it creatively solves the inherent shortcomings of single retrieval technologies. This combination enables the interactive solution provided by the implementation method of this disclosure to comprehensively improve the breadth, accuracy, robustness, and compliance compared to a single technical path when dealing with diverse and professional service inquiries, thus ensuring the accuracy, professionalism, and reliability of the responses.

[0165] In addition, the interactive system 2000 for service consultation may also include a fine-tuning module (not shown), which is configured to: collect feedback data from users in response to natural language replies; and fine-tune the large language model based on historical interaction data including service consultation requests, at least one document fragment, natural language replies, and feedback data.

[0166] After receiving a natural language response to a service inquiry request, the fine-tuning module can collect user feedback data on the natural language response and fine-tune the large language model based on historical interaction data including the service inquiry request, at least one document fragment, the natural language response, and the feedback data. This self-optimizing closed loop brings dynamic growth capabilities to the interactive technology solution for service inquiries provided by the embodiments of this disclosure, continuously reinforcing and optimizing itself over time.

[0167] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0168] Figure 11 A schematic block diagram of an example electronic device 3000 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobility methods, such as personal digital processing, cellular phones, smartphones, wearable devices, and other similar computing methods. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0169] like Figure 11As shown, device 3000 includes a computing unit 301, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 302 or a computer program loaded from storage unit 308 into random access memory (RAM) 303. The RAM 303 may also store various programs and data required for the operation of device 3000. The computing unit 301, ROM 302, and RAM 303 are interconnected via bus 304. Input / output (I / O) interface 305 is also connected to bus 304.

[0170] Multiple components in device 3000 are connected to I / O interface 305, including: input unit 306, such as keyboard, mouse, etc.; output unit 307, such as various types of monitors, speakers, etc.; storage unit 308, such as disk, optical disk, etc.; and communication unit 309, such as network card, modem, wireless transceiver, etc. Communication unit 309 allows device 3000 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0171] The computing unit 301 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 301 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 301 performs the various systems and processes described above, such as interactive methods for service consultation. For example, in some embodiments, the interactive methods for service consultation may be implemented as computer software programs tangibly contained in a machine-readable medium, such as storage unit 308. In some embodiments, part or all of the computer program may be loaded and / or installed on device 3000 via ROM 302 and / or communication unit 309. When the computer program is loaded into RAM 303 and executed by the computing unit 301, one or more steps of the interactive methods for service consultation described above may be performed. Alternatively, in other embodiments, computing unit 301 may be configured to perform interactive methods for service consultation by any other suitable means (e.g., by means of firmware).

[0172] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include: implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input method, and at least one output method, and transferring data and instructions to the storage system, the at least one input method, and the at least one output method.

[0173] The program code used to implement the systems disclosed herein can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing method, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0174] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, method, or apparatus. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, methods, or apparatuses, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0175] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display method for showing information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing method (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of methods can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0176] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0177] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other.

[0178] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, system products, and computer programs according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0179] Various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical applications, or technological improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

[0180] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0181] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. An interactive method for service consultation, characterized in that, include: Identify the intent category of service inquiry requests; Based on the identified intent classification, a routing instruction is generated to route the service consultation request to a target knowledge base determined by at least one of a plurality of predefined private domain knowledge bases, wherein the plurality of private domain knowledge bases include a general business knowledge base, a policy clause knowledge base, a legal aid knowledge base, and a service supervision knowledge base; Perform a search enhancement-generated RAG search to retrieve at least one document fragment related to the service inquiry request from the target knowledge base; as well as Based on the retrieved document fragment, a natural language response to the service inquiry request is generated using a large language model.

2. The method according to claim 1, wherein, The method further includes: The confidence level of the natural language response is assessed; and In response to a confidence level greater than or equal to a predetermined threshold, the natural language response is output.

3. The method according to claim 1, wherein, The method further includes: The confidence level of the natural language response is assessed; and If the confidence level of the natural language response is less than a predetermined threshold, the natural language response is manually reviewed or the natural language response is regenerated.

4. The method according to claim 2 or 3, wherein, The method further includes: The confidence level of the natural language response is determined based at least on the degree of consistency between the natural language response and the at least one document fragment.

5. The method according to claim 1, wherein, Based on the identified intent classification, the generated routing instructions include: The identified intent category is matched with intent categories in a predefined business rule base, and the routing instruction is generated based on the matching result.

6. The method according to claim 5, wherein, Based on the identified intent classification, the generated routing instructions further include: In response to a mismatch between the intent classification and the intent classification in the business rule base, the routing instruction is generated using the large language model based on the context of the service inquiry request.

7. An interactive system for service consultation, characterized in that, include: The intent recognition module is configured to classify the intent of service inquiry requests; The routing decision module is configured to generate routing instructions based on the identified intent classification to route the service inquiry request to a target knowledge base determined by at least one of a plurality of predefined private domain knowledge bases, wherein the plurality of private domain knowledge bases include a general business knowledge base, a policy clause knowledge base, a legal aid knowledge base, and a service supervision knowledge base. The retrieval module is configured to perform a retrieval enhancement-generated RAG retrieval to retrieve at least one document fragment related to the service consultation request from the target knowledge base; as well as The response generation module is configured to generate a natural language response to the service inquiry request based on the retrieved document fragments using a large language model.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 6.

9. A non-volatile computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 6.

10. A computer program product comprising a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1 to 6.

Images