Dialogue method and device based on large language model, medium and equipment

Abstract

The invention provides a dialogue method and device based on a large language model, a medium and equipment, and relates to the field of large language models and the field of financial science and technology. The method comprises the following steps: firstly, acquiring a current input instruction and context information of a user in a dialog box, and sensing the context information of the user by utilizing a context sensing module; then inputting the sensing result and the current input instruction into a large language model for analysis to obtain a target demand, and determining a target function corresponding to the target demand in a pre-constructed business function knowledge graph by using the large language model; and finally, determining a target service module corresponding to the target demand based on the large language model, calling the target service module to generate response information according to an execution result of the target function, and returning the response information to the user in a dialog box. According to the embodiment, the dialogue accuracy and naturalness are remarkably improved, the user experience is enhanced, and meanwhile the interaction quality and efficiency are greatly improved.

Description

Technical Field

[0001] The present disclosure relates to the field of large language models and financial technology, and specifically, to a dialogue method, device, medium and equipment based on a large language model. Background Art

[0002] With the rapid integration of financial technology and the insurance industry, natural language processing technology, especially large language models such as GPT-3, BERT and RoBERTa, has shown great potential in optimizing the business processes of insurance and other financial institutions. These models are pre-trained on large-scale corpora and can efficiently handle text generation, understanding and analysis tasks, providing strong technical support for improving customer service experience and automated business processing.

[0003] With the advancement of technology, people have begun to explore the application of these large language models in more complex scenarios, such as enterprise-level business dialogue scenarios such as insurance. In particular, when dealing with complex insurance business processes and function calls, the relevant technologies still face significant challenges. First, the business logic in enterprise-level business scenarios is highly professional and complex. Taking the insurance enterprise scenario as an example, its business logic covers multiple links such as insurance product design, risk assessment, and claims processing. However, the context perception ability of existing models is limited, and it is difficult to fully capture and understand the key information in these specific business contexts, resulting in the generated responses often failing to accurately match the actual needs of users. Secondly, the business environment of insurance is dynamic and changeable, and new insurance products, policies, and regulations are frequently introduced, requiring the dialogue system to have a high degree of flexibility and adaptability. However, the current function call mechanism mostly relies on predefined rules or plug-ins, lacking sufficient flexibility to cope with these changes, which not only limits the efficiency of automated processing, but also increases the risk of operational errors. In addition, with the continuous expansion and deepening of insurance business, the demand for fine-tuning and updating of models is increasing, but this process is not only costly, but also requires a large amount of labeled data and computing resources, which puts tremendous pressure on the IT operation and maintenance of enterprises. Summary of the invention

[0004] The disclosed embodiments at least provide a conversation method, apparatus, medium and equipment based on a large language model, which accurately perceives user context information and combines it with large language model analysis to accurately determine user target needs and corresponding business functions, and efficiently call the target business module to generate responses, thereby significantly improving the accuracy and naturalness of the conversation, enhancing the user experience, and greatly improving the quality and efficiency of interaction.

[0005] The present disclosure provides a conversation method based on a large language model, including:

[0006] Obtain the user's current input instructions and context information in the dialog box, and use the context awareness module to perceive the user's context information;

[0007] Input the perception result and the current input instruction into the large language model for parsing to obtain the target demand, and use the large language model to determine the target function corresponding to the target demand in the pre-built business function knowledge graph;

[0008] A target business module corresponding to the target requirement is determined based on the large language model, and the target business module is called to generate response information according to the target function execution result, and the response information is returned to the user in the dialog box.

[0009] The present disclosure provides a conversation device based on a large language model, including:

[0010] An information perception module is used to obtain the current input instruction and context information of the user in the dialog box; and use the context perception module to perceive the user's context information, and input the perception result and the current input instruction to the large language model for processing;

[0011] A function determination module, configured to parse the perception result and the current input instruction based on the large language model to obtain a target requirement; and to determine a target function corresponding to the target requirement in a pre-built business function knowledge graph using the large language model;

[0012] A function calling module is used to determine a target business module corresponding to the target demand based on the large language model, and to call the target business module to generate response information according to the target function execution result, and return the response information to the user in the dialog box.

[0013] An embodiment of the present disclosure provides a computer device, including: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the computer device is running, the processor and the memory communicate via the bus, and when the machine-readable instructions are executed by the processor, the dialogue method based on the large language model as described in any possible implementation manner described above is performed.

[0014] An embodiment of the present disclosure provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the conversation method based on a large language model as described in any possible implementation manner described above is implemented.

[0015] The conversation method, device, medium and equipment based on the large language model provided in the embodiments of the present disclosure deeply mine the user's intention and determine the target demand through the combination of the context perception module and the large language model; then, quickly locate and execute the corresponding target function in the pre-built business function knowledge graph, and accurately identify and call the target business module based on the large language model to generate response information, and finally return the response information to the user. In this way, by accurately perceiving the user's context information and combining it with the large language model analysis, accurately determining the user's target demand and the corresponding business function, and efficiently calling the target business module to generate a response, the accuracy and naturalness of the conversation are significantly improved, the user experience is enhanced, and the quality and efficiency of the interaction are greatly improved.

[0016] In order to make the above-mentioned objectives, features and advantages of the present disclosure more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following briefly introduces the drawings required to be cited in the embodiments. The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present disclosure and are used together with the specification to illustrate the technical solutions of the present disclosure. It should be understood that the following drawings only illustrate certain embodiments of the present disclosure and should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0018] Figure 1 A schematic diagram of an application environment of a conversation method based on a large language model provided by an embodiment of the present disclosure is shown;

[0019] Figure 2 A flowchart of a conversation method based on a large language model provided by an embodiment of the present disclosure is shown;

[0020] Figure 3 A flowchart of a method for constructing a business module in a conversation method based on a large language model provided in an embodiment of the present disclosure is shown;

[0021] Figure 4 A schematic diagram of the structure of a dialogue device based on a large language model provided by an embodiment of the present disclosure is shown;

[0022] Figure 5 A schematic diagram showing the structure of another large language model-based dialogue device provided by an embodiment of the present disclosure;

[0023] Figure 6 A schematic diagram of the structure of a computer device provided by an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical scheme and advantages of the embodiments of the present disclosure clearer, the technical scheme in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all of the embodiments. The components of the embodiments of the present disclosure generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present disclosure provided in the drawings is not intended to limit the scope of the present disclosure for protection, but merely represents the selected embodiments of the present disclosure. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present disclosure.

[0025] It is understandable that before using the technical solutions disclosed in the embodiments of the present disclosure, the types, scope of use, usage scenarios, etc. of the personal information involved in the present disclosure should be informed to the user and the user's authorization should be obtained in an appropriate manner in accordance with relevant laws and regulations.

[0026] For example, in response to receiving an active request from a user, a prompt message is sent to the user to clearly prompt the user that the operation requested to be performed will require obtaining and using the user's personal information. Thus, the user can autonomously choose whether to provide personal information to software or hardware such as an electronic device, application, server, or storage medium that performs the operation of the technical solution of the present disclosure according to the prompt message.

[0027] As an optional but non-limiting implementation, in response to receiving an active request from the user, the prompt information may be sent to the user in the form of a pop-up window, in which the prompt information may be presented in text form. In addition, the pop-up window may also carry a selection control for the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0028] It is understandable that the above notification and the process of obtaining user authorization are merely illustrative and do not constitute a limitation on the implementation of the present disclosure. Other methods that meet the relevant laws and regulations may also be applied to the implementation of the present disclosure.

[0029] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0030] The term "and / or" herein only describes an association relationship, indicating that three relationships may exist. For example, A and / or B may represent the following three situations: A exists alone, A and B exist at the same time, and B exists alone. In addition, the term "at least one" herein represents any combination of at least two of any one or more of a plurality of. For example, including at least one of A, B, and C may represent including any one or more elements selected from the set consisting of A, B, and C.

[0031] As the integration of financial technology and the insurance industry accelerates, natural language processing technology, especially large language models such as GPT-3, BERT and RoBERTa, has great potential in optimizing business processes for insurance and other financial institutions. These models can effectively handle text generation, understanding and analysis tasks through large-scale corpus pre-training, improve customer service experience and automated business processing efficiency.

[0032] However, in enterprise-level business dialogue scenarios, such as insurance in complex insurance business processes and function calls in insurance enterprise-level business scenarios, the application of these models still faces challenges. First, the insurance business is highly professional and complex, covering multiple links such as product design, risk assessment, and claims processing, while the existing models have limited context perception capabilities and it is difficult to fully understand the key information in a specific business context, resulting in responses that do not match user needs. Secondly, the business environment is changeable, and new products, policies, and regulations are frequently updated, requiring dialogue systems to be highly flexible and adaptable. However, the current function call mechanism mostly relies on predefined rules or plug-ins, lacks flexibility, limits the efficiency of automated processing, and increases the risk of operational errors. In addition, the demand for model fine-tuning and updating increases, the cost is high, and a large amount of labeled data and computing resources are required, which puts pressure on enterprise IT operations and maintenance.

[0033] Based on the above research, the present disclosure provides a dialogue method, apparatus, medium and device based on a large language model. Specifically, first, the current input instructions and context information of the user in the dialog box are obtained, and the context information of the user is perceived by using a context perception module; then the perception result and the current input instructions are input into the large language model for parsing to obtain the target requirement, and the large language model is used to determine the target function corresponding to the target requirement in a pre-built business function knowledge graph; finally, the target business module corresponding to the target requirement is determined based on the large language model, and the target business module is called according to the execution result of the target function to generate response information, and the response information is returned to the user in the dialog box.

[0034] In the disclosed embodiments, by accurately perceiving user context information and combining it with large language model analysis, the user's target needs and corresponding business functions are accurately determined, and the target business module is efficiently called to generate a response, which significantly improves the accuracy and naturalness of the conversation, enhances the user experience, and greatly improves the quality and efficiency of the interaction.

[0035] To facilitate understanding of this embodiment, the execution subject of the conversation method based on a large language model provided by the embodiment of the present disclosure is first introduced in detail. The conversation method based on a large language model provided by the embodiment of the present invention can be applied to Figure 1 In an application environment, the client communicates with the server through a network. The client can be a mobile device, a user terminal, a terminal, a handheld device, a computing device, a vehicle-mounted device, a wearable device, etc. The server can be an independent physical server, or a server cluster or a distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud storage, big data and artificial intelligence platforms. Optionally, the method can also be applied to an implementation environment composed of a client and a server.

[0036] To facilitate understanding of this embodiment, the application scenarios of the present invention are exemplified below. The dialogue method based on the large language model provided in the embodiment of the present disclosure can be applied to business scenarios at the financial insurance enterprise level, such as for improving the intelligent service level of vehicle insurance software and personal health insurance software. Specifically, in vehicle insurance software, the technology can deeply understand the user's driving behavior and vehicle information, and automatically generate personalized car insurance recommendations and renewal reminders by dynamically calling functions related to auto insurance, thereby improving user experience and sales conversion rate. In personal health insurance software, the model can combine the user's health data and living habits, intelligently call health insurance-related service interfaces, and provide users with customized health insurance products and health management suggestions. In some other embodiments, the present disclosure can also be applied to scenarios in the education field (i.e., education APP or online learning platform) and e-commerce industry where enterprise-level business dialogues can be conducted, which are not specifically limited here.

[0037] The following is a detailed description of the conversation method based on the large language model provided in the embodiment of the present application in conjunction with the accompanying drawings. Figure 2 As shown, it is a flowchart of a dialogue method based on a large language model provided by an embodiment of the present disclosure, and the method includes the following S201-S203:

[0038] S201, obtaining the current input instruction and context information of the user in the dialog box, and using a context perception module to perceive the user's context information.

[0039] It is understandable that context information refers to the historical interaction record between the user and the large language model, which may include information such as conversation history, user preferences and situational information. For example, in the insurance field, when a user consults a chatbot about the purchase of family health insurance, the context information may include key information such as the health status, age distribution, and medical history of family members mentioned by the user, and may also include the user's previous inquiry history and preferences for insurance products (such as preference for products with high cost performance or high coverage). At this time, the context-aware module uses a multi-task learning mechanism to deeply mine and integrate this information, thereby realizing knowledge sharing between various information, avoiding overfitting of the model on a single task, and improving the adaptability of the model to complex business scenarios. On this basis, the context-aware module can also adaptively assign weights according to the importance of different tasks. For example, for family members with poor health, the relevant medical expenses coverage weight may be increased. This way of dynamically adjusting weights enables the large language model to understand the complex needs of users more accurately, provide personalized solutions that fit the actual situation of users, and thus significantly improve user experience and satisfaction.

[0040] Specifically, the context-aware module proposed in the present disclosure is embedded in the encoding layer of the large language model, aiming to significantly improve the performance of the model in practical applications by capturing the business process context information in the dialogue window. This module uses a variety of advanced technologies during the training process, including mixed precision training, data parallelism, learning rate scheduler, and gradient accumulation. The combination of these technologies not only improves the efficiency and stability of training, but also significantly shortens the training time and reduces the consumption of computing resources, so that the model can achieve better performance in a shorter time.

[0041] S202, input the perception result and the current input instruction into the large language model for parsing to obtain the target requirement, and use the large language model to determine the target function corresponding to the target requirement in a pre-built business function knowledge graph.

[0042] Here, in the process of using the large language model to parse the perception results and current input instructions, the natural language processing technology of the large language model can be used to deeply understand the user's intentions and contextual information, and then obtain the user's target needs. Among them, natural language understanding technology can enable the model not only to recognize keywords and phrases, but also to capture potential emotions and context changes, so as to interpret user needs more comprehensively. In addition, the large model can also use this technology to reason about ambiguous or incomplete information and generate more accurate responses. In this way, not only the accuracy of understanding is improved, but also a basis for subsequent business logic decisions is provided.

[0043] For example, in the scenario of insurance business, when customers ask about insurance products, apply for or claim in the dialog box, the large language model first uses its natural language processing technology to deeply analyze the input information. This process is not just a simple matching of keywords and phrases, but can deeply understand the user's intentions and the context in which they are located. For example, when a user mentions "I want to buy a medical insurance for my 60-year-old parents", the large language model can not only identify the core demand of "buying medical insurance", but also capture the key information of "60-year-old parents". The context information may also contain information such as their living habits. Then the large language model can understand that the user may be concerned about the medical insurance plan exclusively for the elderly, as well as the details such as premiums, coverage, and claims process that may be involved. Furthermore, natural language understanding technology enables the model to perceive the potential emotions and context changes in the user's speech. For example, if the user says "I am really worried about my parents' health and want to give them the best protection", the model can capture the user's concerns and attention to the health of his parents, so as to recommend insurance products that better meet their emotional needs, or provide more intimate and detailed service suggestions. Faced with vague or incomplete information provided by users, the large language model also demonstrates strong reasoning capabilities. For example, the user only mentions "I need an insurance policy that covers cancer." Based on this information, combined with common insurance knowledge and the user's possible intentions, the model can infer that the user may be looking for major disease insurance that covers cancer treatment costs, and then provide more specific and detailed product recommendations. Through such a comprehensive and in-depth understanding, the large language model not only improves the accuracy of grasping user needs, but also provides a solid foundation for insurance companies' subsequent business logic decisions. For example, based on personalized information such as the user's age, health status, and protection needs, the model can assist insurance companies in quickly generating customized insurance plans, optimizing user experience, and improving service efficiency and customer satisfaction.

[0044] Specifically, after obtaining the user's target demand, the large language model can find the target function corresponding to the target demand in the pre-built knowledge graph according to the user's target demand, ensuring that the subsequent solutions provided meet the actual needs of the user and achieve efficient interaction and service. For example, when a user asks about the car insurance business in the vehicle insurance customer service system, the large language model will find the target function related to the car insurance business in the pre-built knowledge graph, which can be "car: XXX, name: xxx; phone: yyy; type: YYY"; or, when a user asks about the personal insurance business in the personal insurance customer service system, the large language model will find the target function related to the personal insurance business in the pre-built knowledge graph, such as: "name: nnn; phone: mmm; age: NNN; demand type: MMM" and so on. By determining the target function corresponding to the target demand, it can be used to achieve the structured output of the target demand, convert the user's needs into actionable information, thereby achieving efficient customer service and problem solving. In addition, the present disclosure also introduces parallel calling technology, which can enable multiple independent function calls to be executed simultaneously, reducing the delay of the business process and improving the overall efficiency.

[0045] The business function knowledge graph is constructed through the following steps (1) to (2):

[0046] (1) Acquiring enterprise business information; wherein the enterprise business information includes a plurality of enterprise business functions and semantic information corresponding to each enterprise business;

[0047] (2) Using entity relationship extraction technology, based on the semantic information corresponding to each enterprise business, identify and extract the relationship information between each enterprise business function in the enterprise business data; and import the relationship information between the multiple enterprise business functions, the multiple enterprise businesses and the enterprise business functions into a graph database to form the business function knowledge graph.

[0048] It can be understood that a business function refers to a unit that performs a specific business service, and the semantic information corresponding to each enterprise business may be business description, preconditions, post-effects, etc. Taking the insurance industry as an example, business data may contain specific terms, rate structures, customer qualification requirements, etc. of different types of insurance products (such as auto insurance, life insurance, etc.). Specifically for auto insurance business, the data will list in detail key information such as the name of the insured, contact number, vehicle type (such as sedan, truck), insurance type (such as compulsory traffic insurance, commercial full insurance), etc., which provides the basis for subsequent relationship extraction. Among them, entity relationship extraction technology aims to automatically extract semantic relationships between entity pairs from unstructured text. This technology can extract entities (such as name, phone number, vehicle type) and their associations from a large amount of text data (for example: if a user wants to recommend auto insurance, he needs to know his vehicle type. If he wants to buy auto insurance, he also needs to know his name, phone number, etc.). Once these relationships are accurately identified, they are imported into the graph database together with the enterprise business function itself and the enterprise business name. With its powerful node-edge structure, the graph database naturally represents entities and relationships, forming a business function knowledge graph. In this graph, each business function, business instance and their logical relationships are clearly displayed, making it easier to query, reason and analyze complex business scenarios.

[0049] In this way, the business function knowledge graph obtains enterprise business information and uses entity relationship extraction technology to identify and extract relationship information between business functions, imports it into the graph database, and forms a graph that intuitively displays business functions, business instances and their logical relationships, which is beneficial for efficient query, reasoning and analysis in complex business scenarios.

[0050] S203, determining a target business module corresponding to the target requirement based on the large language model, calling the target business module to generate response information according to the target function execution result, and returning the response information to the user in the dialog box.

[0051] For example, before determining the target business module corresponding to the target demand based on the large language model, it is necessary to build in advance the business module corresponding to the business information of each department in the enterprise-level business, referring to Figure 3 As shown, the method includes the following S301 to S303:

[0052] S301, obtaining enterprise business information; wherein the enterprise business information includes business information corresponding to multiple business departments.

[0053] It is understandable that the business information corresponding to each business department may include but is not limited to the scope of responsibilities, core business processes, key performance indicators (KPIs), staffing, and interactions with other departments of each business department.

[0054] S302, constructing business modules corresponding to each of the business departments respectively; and for each of the business modules, constructing a business adjustment module corresponding to the business module.

[0055] Here, after obtaining the business information corresponding to multiple business departments, a corresponding business module can be constructed for each business department, and a corresponding business adjustment module can be constructed for each business module for one-to-one adjustment of the corresponding business module.

[0056] S303 : For each of the business modules, based on a business adjustment module corresponding to the business module, adjust the business module according to the business information corresponding to the business department.

[0057] Specifically, after the business module and the business adjustment module are constructed, the business adjustment module corresponding to the business module can be used to adjust the business module according to the business information corresponding to the business department. These adjustments may involve multiple aspects such as establishment of business processes, data processing, and user information storage.

[0058] It is understandable that the above-mentioned multiple business modules include a global shared layer and a business logic layer corresponding to each business department. Here, the global shared layer is to provide some public services and data support across business modules, such as user authentication, data storage and permission management, etc., which ensures information sharing and consistency between modules, and can improve the overall efficiency and maintainability of the system. The business logic layer of each business module is mainly responsible for processing the business rules and logic unique to the module. In the process of adjusting the business module for each business module based on the business adjustment module corresponding to the business module according to the business information corresponding to the business department, the gradient freezing technology can be used (a strategy for training deep learning models, the core of which is to allow the weights of some layers to remain unchanged, focusing on optimizing specific business logic layers without affecting the learning process of other layers), that is, the adjustment of each business module is based on the business adjustment module corresponding to the business module according to the business information of the business department corresponding to the business module, freezing other irrelevant business logic layers, ensuring that only the parameters of the business logic layer and the global shared layer related to the module can be updated, so as to adapt to business changes more accurately.

[0059] In addition, if any of the above business modules has an update requirement, in response to the update instruction of the business module, the update business information corresponding to the update instruction is obtained, and the global shared layer and other business logic layers except the business logic layer corresponding to the business module are frozen, and the business adjustment module corresponding to the business module is used to update it according to the updated business information. For example: in the insurance enterprise-level business scenario, there is a "claim management" business module. If the module receives an update requirement, such as optimizing the claim process to shorten the processing time, first respond to the update instruction and obtain the updated business information corresponding to the instruction. At this time, the global shared layer and all other business logic layers are frozen to prevent data inconsistency or system failure during the update process; then, the business adjustment module corresponding to the "claim management" module is used to perform corresponding process optimization and system adjustment based on the updated business information obtained. In this way, not only the flexibility of the system is improved, but also it is ensured that when the module is updated, the business logic of other parts will not be affected, thereby ensuring the continuity and reliability of the overall business.

[0060] In the disclosed embodiment, by systematically constructing business modules corresponding to the business information of each department, the efficiency and consistency of the business process are ensured. At the same time, the setting of the business adjustment module allows independent fine-tuning of different business modules. Through this design, the modules can be quickly optimized and updated according to the business needs of different departments of the enterprise without retraining the entire large language model.

[0061] It is understandable that when determining the target business module corresponding to the target demand based on the large language model, and calling the target business module to generate response information according to the execution result of the target function, the target business module corresponding to the target demand can be determined based on the large language model first. Then, the target function determined by the large language model in the pre-established knowledge graph is executed, and the corresponding execution result is obtained, which usually contains the information required for decision-making. On this basis, the execution result of the target function is input into the global sharing layer to achieve more comprehensive analysis and response, thereby providing more accurate and richer information for the subsequent business logic layer. Once the global sharing layer is processed, its output data is input into the target business logic layer. The target business logic layer will use the globally shared data and combine its own business logic to conduct in-depth analysis and processing to ensure that the generated content is highly relevant to the user's needs. Finally, based on the output result of the target business logic layer, the large language model will generate the final response information.

[0062] For example, in the insurance business industry, when a customer asks about the health insurance claims process, the target business logic layer corresponding to the demand is first identified as the "claims process processing layer" based on the large language model. Subsequently, the large language model executes a specific objective function in the pre-built insurance knowledge graph, such as "querying standard claims steps", and obtains the execution results containing detailed steps. These results are then input into the global shared layer, which combines other relevant information (such as customer historical data, policy terms, etc.) for comprehensive analysis to provide a more detailed context for the claims logic layer. After receiving the data from the global shared layer, the claims logic layer further refines the claims process according to the insurance business rules defined internally, ensuring that personalized and accurate claims guidance is provided to customers. Finally, based on the processing results of the claims logic layer, the large language model generates a detailed claims process response information that complies with the insurance policy and meets customer needs.

[0063] Here, in order to monitor whether there is a conceptual deviation between the output results of the large language model and the user's needs (i.e., the question is not the answer), the present disclosure proposes that after the response information is returned to the user in the dialog box, it is also possible to determine whether the response information matches the current input instruction; if it is found that the response information does not match the instruction, the function call tasks performed by the large language model within a specific preset time period should be recorded, and relevant business features should be extracted from them. These business features will be used to analyze the behavior of the model during this period of time to help identify factors that may cause deviation. Then, based on the extracted business features, the context-aware module in the large language model can be updated. This update not only helps to improve the response accuracy of the model in similar tasks, but also enhances its sensitivity to user needs, thereby reducing the possibility of concept deviation in the future. Through this feedback mechanism, the model can continuously learn and optimize to ensure that the expectations and needs of users are better met. Ultimately, this process will promote a more efficient interactive experience and improve user satisfaction.

[0064] It is understandable that in order to ensure the privacy and security of the data of the enterprise during use, and meet the high requirements of modern enterprises for data protection, the present disclosure also proposes to use a differential privacy protection mechanism to protect the privacy of the data in the function call process in the large language model; and use a preset encryption technology to encrypt the data information in the large language model to achieve encrypted storage and encrypted transmission of data information in the large language model. Among them, differential privacy (DP) is a technology for protecting the privacy of personal data, which blurs the information of a specific user by introducing noise, so that the user of the data cannot determine the specific information of a certain user. This method ensures that even when using a data set for analysis or training a model, the sensitive information of a single user cannot be inferred from the output results. In the function call process of the large language model, differential privacy can be applied to the input and output data of the model, so that even after calling the model multiple times, the user's personal data cannot be identified. By adding appropriate random noise to the data, the learning process of the model can be effectively prevented from leaking the user's private information, thereby meeting the high requirements of modern enterprises for data protection. The preset encryption technology may include but is not limited to using AES-256 encryption technology to realize encrypted storage of data information and using TLS1.3 protocol to encrypt transmission of data information.

[0065] In some other embodiments, an access control system based on RBAC (Role-Based Access Control) can be introduced to strictly manage access to sensitive data during the operation of large language models, and integrate Elasticsearch audit log management to achieve transparency and traceability of data use and ensure enterprise data security and compliance.

[0066] The conversation method, device, medium and equipment based on the large language model provided in the embodiments of the present disclosure deeply mine the user's intention and determine the target demand through the combination of the context perception module and the large language model; then, quickly locate and execute the corresponding target function in the pre-built business function knowledge graph, and accurately identify and call the target business module based on the large language model to generate response information, and finally return the response information to the user. In this way, by accurately perceiving the user's context information and combining it with the large language model analysis, accurately determining the user's target demand and the corresponding business function, and efficiently calling the target business module to generate a response, the accuracy and naturalness of the conversation are significantly improved, the user experience is enhanced, and the quality and efficiency of the interaction are greatly improved.

[0067] Those skilled in the art will appreciate that, in the above method of specific implementation, the order in which the steps are written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of the steps should be determined by their functions and possible internal logic.

[0068] Based on the same inventive concept, the embodiments of the present disclosure also provide a dialogue device based on a large language model corresponding to the dialogue method based on a large language model. Since the principle of solving the problem by the device in the embodiments of the present disclosure is similar to the above-mentioned dialogue method based on a large language model in the embodiments of the present disclosure, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be repeated.

[0069] Reference Figure 4 FIG. 4 is a schematic diagram of a conversation device 400 based on a large language model provided by an embodiment of the present disclosure, wherein the device includes:

[0070] The information perception module 401 is used to obtain the current input instruction and context information of the user in the dialog box; and use the context perception module to perceive the user's context information, and input the perception result and the current input instruction to the large language model for processing;

[0071] A function determination module 402 is used to parse the perception result and the current input instruction based on the large language model to obtain a target requirement; and to determine a target function corresponding to the target requirement in a pre-built business function knowledge graph using the large language model;

[0072] The function calling module 403 is used to determine the target business module corresponding to the target demand based on the large language model, and call the target business module to generate response information according to the target function execution result, and return the response information to the user in the dialog box.

[0073] In some possible embodiments, the function calling module 403 is further used to:

[0074] Acquire enterprise business information; wherein the enterprise business information includes business information corresponding to multiple business departments;

[0075] Constructing business modules corresponding to each of the business departments respectively; and constructing business adjustment modules corresponding to each of the business modules;

[0076] For each of the business modules, the business module is adjusted based on the business adjustment module corresponding to the business module according to the business information corresponding to the business department.

[0077] In some possible embodiments, the multiple business modules include a global sharing layer and a business logic layer corresponding to each business department; the function calling module 403 is specifically used to:

[0078] Based on the large language model, a target business logic layer corresponding to the target requirement is determined, the target function execution result is input into the global shared layer, the output data of the global shared layer is input into the target business logic layer, and the response information is generated based on the output result of the target business logic layer through the large language model.

[0079] In some possible embodiments, the function calling module 403 is further used to:

[0080] For each of the business modules, based on the business adjustment module corresponding to the business module and the business information of the business department corresponding to the business module, freezing other business logic layers except the business logic layer corresponding to the business module, and adjusting parameters of the global sharing layer and the business logic layer corresponding to the business module;

[0081] The function calling module 403 is also used for:

[0082] In response to an update instruction of any business module, updated business information corresponding to the update instruction is obtained, and the global shared layer and other business logic layers except the business logic layer corresponding to any business module are frozen, and the business adjustment module corresponding to any business module is used to update the any business module according to the updated business information.

[0083] In some possible embodiments, the function calling module 403 is further used to:

[0084] Determining whether the response information matches the current input instruction;

[0085] In the case that the response information does not match the current input instruction, the function call tasks completed by the large language model within a preset time period are obtained, and business features are extracted from the function call tasks completed within the preset time period, and the context awareness module in the large language model is updated according to the business features.

[0086] In some possible embodiments, the function determination module 402 is further configured to:

[0087] Acquire enterprise business data; wherein the enterprise business data includes multiple enterprise business functions and semantic information corresponding to each enterprise business;

[0088] Utilizing entity relationship extraction technology, the relationship information between each enterprise business function is identified and extracted in the enterprise business data according to the semantic information corresponding to each enterprise business; and the relationship information between the multiple enterprise business functions, multiple enterprise businesses and the enterprise business functions is imported into a graphic database to form the business function knowledge graph.

[0089] In some possible embodiments, reference Figure 5 As shown, the device also includes a data encryption module 404, which is used to use a differential privacy protection mechanism to protect the privacy of data in the function call process in the large language model; and use a preset encryption technology to encrypt the data information in the large language model to achieve encrypted storage and encrypted transmission of the data information in the large language model.

[0090] Based on the same technical concept, the embodiment of the present disclosure also provides a computer device. Figure 6 , which is a schematic diagram of the structure of a computer device 600 provided in an embodiment of the present disclosure, including a processor 601, a memory 602, and a bus 603. The memory 602 is used to store execution instructions, including a memory 6021 and an external memory 6022; the memory 6021 is also called an internal memory, which is used to temporarily store the operation data in the processor 601 and the data exchanged with the external memory 6022 such as a hard disk. The processor 601 exchanges data with the external memory 6022 through the memory 6021.

[0091] In the embodiment of the present application, the memory 602 is specifically used to store the application code for executing the solution of the present application, and the execution is controlled by the processor 601. That is, when the computer device 600 is running, the processor 601 communicates with the memory 602 through the bus 603, so that the processor 601 executes the application code stored in the memory 602, and then executes the method described in any of the above embodiments.

[0092] Among them, the memory 602 can be, but is not limited to, random access memory (Random Access Memory, RAM), read only memory (Read Only Memory, ROM), programmable read-only memory (Programmable Read-Only Memory, PROM), erasable programmable read-only memory (Erasable Programmable Read-Only Memory, EPROM), electrically erasable read-only memory (Electric Erasable Programmable Read-Only Memory, EEPROM), etc.

[0093] Processor 601 may be an integrated circuit chip with signal processing capabilities. The above-mentioned processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The disclosed methods, steps and logic block diagrams in the embodiments of the present invention may be implemented or executed. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0094] It is to be understood that the structure illustrated in the embodiment of the present application does not constitute a specific limitation on the computer device 600. In other embodiments of the present application, the computer device 600 may include more or fewer components than shown in the figure, or combine some components, or split some components, or arrange the components differently. The components shown in the figure may be implemented in hardware, software, or a combination of software and hardware.

[0095] The present disclosure also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the large language model-based dialogue method described in the above method embodiment are executed. The storage medium can be a volatile or non-volatile computer-readable storage medium.

[0096] The embodiments of the present disclosure also provide a computer program product, which carries a program code. The instructions included in the program code can be used to execute the steps of the dialogue method based on the large language model described in the above method embodiment. For details, please refer to the above method embodiment, which will not be repeated here.

[0097] The computer program product may be implemented in hardware, software or a combination thereof. In one optional embodiment, the computer program product is implemented as a computer storage medium. In another optional embodiment, the computer program product is implemented as a software product, such as a software development kit (SDK).

[0098] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, the specific working process of the system and device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. In the several embodiments provided in the present disclosure, it should be understood that the disclosed system and method can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0099] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0100] In addition, each functional unit in each embodiment of the present disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0101] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present disclosure, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present disclosure. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0102] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present disclosure, which are used to illustrate the technical solutions of the present disclosure, rather than to limit them. The protection scope of the present disclosure is not limited thereto. Although the present disclosure is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed in the present disclosure, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure, and should be included in the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be based on the protection scope of the claims.

Claims

1. A conversation method based on a large language model, characterized in that: include: Obtain the user's current input instructions and context information in the dialog box, and use the context awareness module to perceive the user's context information; Input the perception result and the current input instruction into the large language model for parsing to obtain the target demand, and use the large language model to determine the target function corresponding to the target demand in the pre-built business function knowledge graph; A target business module corresponding to the target requirement is determined based on the large language model, and the target business module is called to generate response information according to the target function execution result, and the response information is returned to the user in the dialog box.

2. The method according to claim 1, characterized in that Before determining the target service module corresponding to the target demand based on the large language model, the method includes: Acquire enterprise business information; wherein the enterprise business information includes business information corresponding to multiple business departments; Constructing business modules corresponding to each of the business departments respectively; and constructing business adjustment modules corresponding to each of the business modules; For each of the business modules, the business module is adjusted based on the business adjustment module corresponding to the business module according to the business information corresponding to the business department.

3. The method according to claim 2, characterized in that The multiple business modules include a global sharing layer and a business logic layer corresponding to each business department; the target business module corresponding to the target demand is determined based on the large language model, and the target business module is called to generate response information according to the target function execution result, including: Based on the large language model, a target business logic layer corresponding to the target requirement is determined, the target function execution result is input into the global shared layer, the output data of the global shared layer is input into the target business logic layer, and the response information is generated based on the output result of the target business logic layer through the large language model.

4. The method according to claim 3, characterized in that: For each of the business modules, adjusting the business module based on the business adjustment module corresponding to the business module according to the business information corresponding to the business department includes: For each of the business modules, based on the business adjustment module corresponding to the business module and the business information of the business department corresponding to the business module, freezing other business logic layers except the business logic layer corresponding to the business module, and adjusting parameters of the global sharing layer and the business logic layer corresponding to the business module; The method further comprises: In response to an update instruction of any business module, updated business information corresponding to the update instruction is obtained, and the global shared layer and other business logic layers except the business logic layer corresponding to any business module are frozen, and the business adjustment module corresponding to any business module is used to update the any business module according to the updated business information.

5. The method according to claim 1, characterized in that After returning the response information to the user in the dialog box, the method further includes: Determining whether the response information matches the current input instruction; In the case that the response information does not match the current input instruction, the function call tasks completed by the large language model within a preset time period are obtained, and business features are extracted from the function call tasks completed within the preset time period, and the context awareness module in the large language model is updated according to the business features.

6. The method according to claim 1, characterized in that The business function knowledge graph is constructed by the following steps: Acquire enterprise business data; wherein the enterprise business data includes multiple enterprise business functions and semantic information corresponding to each enterprise business; Utilizing entity relationship extraction technology, the relationship information between each enterprise business function is identified and extracted in the enterprise business data according to the semantic information corresponding to each enterprise business; and the relationship information between the multiple enterprise business functions, multiple enterprise businesses and the enterprise business functions is imported into a graphic database to form the business function knowledge graph.

7. The method according to any one of claims 1 to 6, characterized in that: The method further comprises: A differential privacy protection mechanism is used to protect the privacy of data in the function call process in the large language model; and a preset encryption technology is used to encrypt the data information in the large language model to achieve encrypted storage and encrypted transmission of the data information in the large language model.

8. A conversation device based on a large language model, characterized in that: include: The information perception module is used to obtain the user's current input instructions and context information in the dialog box; And using the context perception module to perceive the context information of the user, and inputting the perception result and the current input instruction into the large language model for processing; A function determination module, configured to parse the perception result and the current input instruction based on the large language model to obtain a target requirement; and to determine a target function corresponding to the target requirement in a pre-built business function knowledge graph using the large language model; A function calling module is used to determine a target business module corresponding to the target demand based on the large language model, and to call the target business module to generate response information according to the target function execution result, and return the response information to the user in the dialog box.

9. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

10. A computer device comprising a storage medium, a processor, and a computer program stored in the storage medium and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

Images