Task-based dialogue robot system based on large language model

Abstract

The invention discloses a task-based dialogue robot system based on a large language model, and relates to the technical field of artificial intelligence and natural language processing, and the system comprises a process control module, an action execution module and a node module which are connected in sequence; the node module comprises a plurality of nodes; the flow control module is used for acquiring input information; inputting the input information into a large language model deployed in a process control module to obtain an initial decision; performing priority ranking and conflict processing on the initial decision to obtain a target decision, and sending the target decision to an action execution module; the action execution module is used for performing a corresponding conversation with the user according to a preset execution logic of each decision in the target decision, executing a corresponding behavior, and selecting a corresponding node in the node module; and each node in the node module is used for executing an action according to the corresponding execution rule. The problem that a task-based dialogue system is low in robustness, poor in cross-domain applicability and stiff in response is solved.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and natural language processing technology, and in particular to a task-oriented chatbot system based on a large language model. Background Technology

[0002] Task-oriented multi-turn dialogue is an artificial intelligence system based on natural language processing and deep learning technologies, designed to meet specific user needs through multi-turn interactions. This system determines user intent and generates appropriate responses through understanding and clarification. To achieve these functions, task-oriented dialogue systems typically employ techniques such as intent recognition, slot filling, and dialogue management. Furthermore, the system can call external knowledge bases or application programming interfaces (APIs) to obtain real-time data, helping users complete specific tasks, such as booking tickets or checking logistics, thereby improving interaction efficiency and user experience.

[0003] Traditional task-oriented dialogue systems consist of four modules: 1. Natural Language Understanding Module: This module primarily handles intent recognition and slot filling. It determines the intent and extracts key information by parsing user input. 2. Dialogue State Tracking Module: This module represents and updates the state information of multi-turn dialogues. It maintains the current dialogue state based on dialogue history, user intent, and slot information. 3. Dialogue Policy Learning Module: This module determines the next system action based on the current dialogue state. 4. Natural Language Generation Module: This module is responsible for converting system actions into natural language responses. When a user inputs a message, the system first matches the user's intent, fills the slots using an entity recognition model, updates the dialogue state using a reinforcement learning algorithm, and determines the next system action. Finally, it generates the corresponding response based on the determined system action.

[0004] Although the technical framework of task-oriented dialogue systems is relatively mature, it still faces some challenges in practice: 1) Low robustness: Existing task-oriented dialogue systems often can only run according to specific scripts and cannot handle ambiguous or vague user input. Users may use different words and expressions to describe the same intent, and their input may contain spelling and grammatical errors. This requires training and complex rule design to handle various non-standard expressions and accurately understand user intent and extract slots. 2) Poor cross-domain applicability: Classification models need to be trained in different domains, which requires a large amount of domain-specific labeled data. Sufficient domain data usually involves high costs, posing a challenge to cross-domain applicability. 3) Stiff responses: Because task-oriented dialogue systems often use template-based methods for responses, the replies are often rigid and cannot provide a sufficiently natural and smooth interactive experience. Summary of the Invention

[0005] The purpose of this application is to provide a task-oriented dialogue robot system based on a large language model to solve the problems of low robustness, poor cross-domain applicability, and stiff responses in task-oriented dialogue systems.

[0006] To achieve the above objectives, this application provides the following solution.

[0007] In a first aspect, this application provides a task-oriented dialogue robot system based on a large language model, comprising: a process control module, an action execution module, and a node module connected in sequence; the node module includes multiple nodes; The process control module is used for: Obtain input information; The input information is fed into the large language model deployed in the process control module to obtain the initial decision; The initial decisions are prioritized and conflict-handled to obtain the target decision, which is then sent to the action execution module. The action execution module is used to conduct corresponding dialogues with the user, execute corresponding behaviors, and select the corresponding nodes in the node module according to the preset execution logic of each decision in the target decision. Each node in the node module is used to perform actions according to the corresponding execution rules.

[0008] In one embodiment, the input information includes: control instructions, process information of candidate processes, process information of the current process, historical dialogue data, and user questions; the process information includes the process name, description, and slot information of the process; the slot information includes the name, description, type, and value of the slot.

[0009] In one embodiment, the initial decision includes one or more of the following: clarifying flow decision, activating flow decision, switching flow decision, canceling flow decision, filling slot decision, and responding to knowledge query decision.

[0010] In one implementation, the multiple nodes are: a collection node, a large language model judgment node, a rule judgment node, an interface call node, a response node, a work order node, and a task node.

[0011] In one embodiment, prioritizing and conflict-handling the initial decisions to obtain the target decision includes: Determine whether process control decisions and response knowledge query decisions coexist in the initial decision-making process to obtain a first determination result; process control decisions include at least one of clarifying flow decisions, activating flow decisions, switching flow decisions, and canceling flow decisions. If the first judgment result is yes, then the response knowledge query decision is removed from the initial decision to obtain a second decision, and the process control decision is checked for conflict according to the preset conflict handling method to obtain the target decision; If the first judgment result is negative, then it is determined whether there is a process control type decision in the initial decision, and a second judgment result is obtained; If the second judgment result is yes, then according to the preset conflict handling method, the process control decision is conflict checked to obtain the target decision; If the second judgment result is negative, then the initial decision is determined as the target decision.

[0012] In one embodiment, the preset conflict handling method includes: When both cancel flow decision and activate flow decision exist in the process control class, and the process name in the activate flow decision is the name of the current process, the cancel flow decision and activate flow decision will be removed from the process control class. When both cancel flow decision and activate flow decision exist in the process control category, and the process name in the activate flow decision is the name of the new process, the cancel flow decision and activate flow decision will be merged into a switch flow decision. When the process name in the clarifying flow decision in the process control class is the name of the current process, the clarifying flow decision will be removed from the process control class. When the process name in the flow switching decision in the process control class is the name of the current process, the flow switching decision will be removed from the process control class.

[0013] In one implementation, the preset execution logic of the clarification flow decision is as follows: using the large language model deployed in the action execution module, a prompt message is generated based on the process name and description of the process in the activation flow decision to guide the user to confirm the process; The default execution logic for activating flow decisions is as follows: notify the user of the flow name of the flow they have entered, and select the corresponding node in the node module according to the flow configuration. The flow configuration includes: flow name, description and corresponding node. The default execution logic for switching flow decisions is as follows: inform the user to switch to the new flow, update the current flow to the new flow, and select the corresponding node in the selected node module according to the flow configuration of the updated current flow. The default execution logic for canceling a flow decision is as follows: inform the user that the flow has been canceled, update the current flow to empty, and inquire about the user's subsequent needs; The default execution logic for filling slots is as follows: update slot information; if the value of a slot is detected to have been replaced, notify the user of the change in the slot value; implement a validation strategy for the slot value; if the validation passes, store the slot value and select the corresponding node in the node module. The default execution logic for responding to knowledge query decisions is as follows: directly answer the user's knowledge-based questions, bring the conversation back to the original process, and prompt the user to supplement the information that was not completed in the previous process.

[0014] In one implementation, the execution rule for the collection node is as follows: determine whether the value of the slot in the current process has been stored; if it has been stored, execute the next node sequentially; if it has not been stored, generate a collection prompt message based on the name and description of the slot in the current process and output it to the user. The execution rule of the large language model judgment node is as follows: input the user question, historical dialogue data, and the names and descriptions of the conditions configured in the large language model judgment node into the large language model deployed in the large language model judgment node, obtain the corresponding condition names, and execute the next node in sequence; The execution rule for the rule judgment node is as follows: determine whether the rule judgment condition is met, obtain the rule judgment result, and execute the next node sequentially based on the rule judgment result; the rule judgment condition is that the value of the slot in the current process has a value, the value of the slot in the current process is equal to the first preset value, the value of the slot in the current process is greater than the second preset value, the value of the slot in the current process is less than the third preset value, or the value of the slot in the current process contains the fourth preset value. The execution rules for the API call node are as follows: based on the configured API address, input parameters, and output parameters, the API is invoked and data is obtained. The output parameters are assigned to the corresponding slots, and the next node is executed sequentially. The execution rule for the response node is: if there is no next node, the task process ends; otherwise, return to the message configured for the node and execute the next node sequentially. The execution rules for the work order node are as follows: input historical dialogue data and user questions into the large language model deployed in the work order node to generate the corresponding work order title and work order summary, summarize the information to be collected for the work order to generate the work order, and execute the next node in sequence; The execution rules for task nodes are as follows: the task objective, task description, termination condition, slot information of the task node, historical dialogue data, and user questions are input into the large language model deployed in the task node to obtain the dialogue and output behavior; the output behavior includes: slot filling behavior, knowledge query behavior, session termination behavior, and tool invocation behavior.

[0015] In one embodiment, the slot filling behavior includes: if the slot value already exists, then storing the slot value; otherwise, storing the slot value of the slot in the slot information of the task node. The knowledge query process includes: generating a query question, retrieving relevant content from a pre-configured knowledge base based on the query question, and inputting the relevant content and the user question into a large language model deployed on a task node to obtain the corresponding answer; The session termination behavior includes: determining whether the preset termination conditions are met based on historical dialogue data; if so, terminating the task node and proceeding to the next node in sequence. The tool invocation behavior includes: using the large language model deployed in the task node to determine whether a tool needs to be invoked; if a tool needs to be invoked, returning the required function and parameters; after invoking the tool, sending the invocation result back to the large language model deployed in the task node; and the large language model deployed in the task node summarizing or planning subtasks based on the invocation result.

[0016] Secondly, this application provides a task-oriented chatbot method based on a large language model, implemented using the aforementioned task-oriented chatbot system based on a large language model. The task-oriented chatbot method based on a large language model includes: Obtain input information; The input information is fed into the large language model deployed in the process control module to obtain the initial decision; The initial decisions are prioritized and conflict-handled to obtain the target decision; The action execution module is used to conduct corresponding dialogues with the user, execute corresponding behaviors, and select the corresponding nodes in the node module according to the preset execution logic of each decision in the target decision. Each node in the node module performs an action according to its corresponding execution rules.

[0017] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application discloses a task-oriented dialogue robot system based on a large language model. This system utilizes a large language model to understand user intent. When a user's response does not match the task scenario, such as when a user asks a different question while collecting slot information, these special cases require no additional configuration and are automatically handled by the large language model. The same word can have multiple expressions; the large language model accurately extracts slots through context, simplifying slot rule configuration and increasing the robot's tolerance to human input, thus improving the robustness of the task-oriented dialogue system. With the help of a general large language model, intent recognition and slot extraction can be quickly performed on tasks in various domains without the need for data collection and model training, reducing the time and resource investment in domain adaptation and improving cross-domain applicability. The large language model understands the context of the dialogue and automatically generates response scripts, making the dialogue more natural and fluent. It can adjust dialogue strategies in a timely manner in the face of abnormal situations, avoiding rigid dialogue flows, improving the flexibility of interaction, and enhancing dialogue fluency. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A schematic diagram of a task-oriented dialogue robot system based on a large language model provided in an embodiment of this application; Figure 2 A detailed structural diagram of a task-oriented dialogue robot system based on a large language model provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] The purpose of this application is to provide a task-oriented dialogue robot system based on a large language model, which aims to solve the problems of low robustness, poor cross-domain applicability, and stiff responses in task-oriented dialogue systems.

[0022] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] In one exemplary embodiment, such as Figure 1 and Figure 2 As shown, a task-oriented dialogue robot system based on a large language model is provided, including: a process control module, an action execution module, and a node module connected in sequence; the node module includes multiple nodes; The process control module is used for: Obtain input information; The input information is fed into the large language model deployed in the process control module to obtain the initial decision; Prioritize and resolve conflicts in the initial decisions to obtain the target decision, and then send the target decision to the action execution module; The action execution module is used to engage in corresponding dialogues with the user, execute corresponding actions, and select the corresponding nodes in the node module according to the preset execution logic of each decision in the target decision. Each node in the node module is used to perform actions according to the corresponding execution rules.

[0024] As an optional implementation, the input information includes: control instructions, process information of candidate processes, process information of the current process, historical dialogue data, and user questions; process information includes the process name, description, and slot information of the process; slot information includes the name, description, type, and value of the slot.

[0025] As an optional implementation, the initial decision includes one or more of the following: clarifying flow decision, activating flow decision, switching flow decision, canceling flow decision, filling slot decision, and responding to knowledge query decision.

[0026] Specifically, the clarification flow decision is represented as: ClarifyFlow('flow name 1', 'flow name 2', ..., 'flow name n'), which means that when a user's information matches multiple flows and they do not know which flow to start, it is necessary to clarify to the user which flow to start; The activation flow decision is represented as: StartFlow('flow name'), which means starting a new flow; The switching flow decision is represented as ChangeFlow('process name'), which means switching from one process to another; The decision to cancel a flow is represented as: CancelFlow(), which cancels the current flow. The slot filling decision is represented as: SetSlot({'name': slot name, 'value': [collected value 1, collected value 2... collected value n]}); The response to a knowledge query decision is represented as SearchAndReply(), which directly answers the knowledge-based questions raised by the user.

[0027] As an optional implementation, the multiple nodes are: a collection node, a large language model judgment node, a rule judgment node, an interface call node, a response node, a work order node, and a task node.

[0028] As an optional implementation, the initial decisions are prioritized and conflict-handled to obtain the target decision, including: Determine whether process control decisions and response knowledge query decisions coexist in the initial decision-making process to obtain the first judgment result; process control decisions include at least one of the following: clarifying flow decisions, activating flow decisions, switching flow decisions, and canceling flow decisions. If the first judgment result is yes, the response knowledge query decision will be removed from the initial decision to obtain the second decision. Then, according to the preset conflict handling method, the process control decision will be conflict-checked to obtain the target decision. If the first judgment result is negative, then determine whether there is a process control type decision in the initial decision and obtain the second judgment result; If the second judgment result is yes, then according to the preset conflict handling method, the process control decision is conflict checked to obtain the target decision; If the second judgment result is negative, then the initial decision is determined as the target decision.

[0029] As an optional implementation method, a preset conflict handling method includes: When both cancel flow decision and activate flow decision exist in the process control class, and the process name in the activate flow decision is the name of the current process, the cancel flow decision and activate flow decision will be removed from the process control class. When both cancel flow decision and activate flow decision exist in the process control category, and the process name in the activate flow decision is the name of the new process, the cancel flow decision and activate flow decision will be merged into a switch flow decision. When the process name in the clarifying flow decision in the process control class is the name of the current process, the clarifying flow decision will be removed from the process control class. When the process name in the flow switching decision in the process control class is the name of the current process, the flow switching decision will be removed from the process control class.

[0030] As an optional implementation, the default execution logic of the clarification flow decision is as follows: using the large language model deployed in the action execution module, a prompt message is generated based on the process name and description of the process in the activation flow decision to guide the user to confirm the process; The default execution logic for activating flow decisions is as follows: notify the user of the flow name of the flow they have entered, and select the corresponding node in the node module according to the flow configuration. The flow configuration includes: flow name, description and corresponding node. The default execution logic for switching flow decisions is as follows: inform the user to switch to the new flow, update the current flow to the new flow, and select the corresponding node in the selected node module according to the flow configuration of the updated current flow. The default execution logic for canceling a flow decision is as follows: inform the user that the flow has been canceled, update the current flow to empty, and inquire about the user's subsequent needs; The default execution logic for filling slots is as follows: update slot information; if the value of a slot is detected to have been replaced, notify the user of the change in the slot value; implement a validation strategy for the slot value; if the validation passes, store the slot value and select the corresponding node in the node module. The default execution logic for responding to knowledge query decisions is as follows: directly answer the user's knowledge-based questions, bring the conversation back to the original process, and prompt the user to supplement the information that was not completed in the previous process.

[0031] Specifically, in the preset execution logic of slot filling decision, to ensure user experience, a preset number of pullbacks is set. When the number of times the dialogue is pulled back to the original process exceeds the preset number and no valid information is obtained, the process pullback is stopped to avoid excessively disturbing the user. In the preset execution logic of responding to knowledge query decision, when implementing the verification strategy, regular expression verification is used for enumerated values ​​and specific values.

[0032] As an optional implementation, the execution rule for the collection node is as follows: determine whether the value of the slot in the current process has been stored. If it has been stored, execute the next node sequentially. If it has not been stored, generate a collection prompt message based on the name and description of the slot in the current process and output it to the user. The execution rule of the large language model judgment node is as follows: input the user question, historical dialogue data, and the names and descriptions of the conditions configured in the large language model judgment node into the large language model deployed in the large language model judgment node, obtain the corresponding condition names, and execute the next node in sequence; The execution rule for the rule judgment node is as follows: determine whether the rule judgment condition is met, obtain the rule judgment result, and execute the next node sequentially based on the rule judgment result; the rule judgment condition is that the value of the slot in the current process has a value, the value of the slot in the current process is equal to the first preset value, the value of the slot in the current process is greater than the second preset value, the value of the slot in the current process is less than the third preset value, or the value of the slot in the current process contains the fourth preset value. The execution rules for the API call node are as follows: based on the configured API address, input parameters, and output parameters, the API is invoked and data is obtained. The output parameters are assigned to the corresponding slots, and the next node is executed sequentially. The execution rule for the response node is: if there is no next node, the task process ends; otherwise, return to the message configured for the node and execute the next node sequentially. The execution rules for the work order node are as follows: input historical dialogue data and user questions into the large language model deployed in the work order node to generate the corresponding work order title and work order summary, summarize the information to be collected for the work order to generate the work order, and execute the next node in sequence; The execution rules for task nodes are as follows: Input the task objective, task description, termination condition, task node slot information, historical dialogue data, and user questions into the large language model deployed in the task node to obtain the dialogue and output behavior; the output behavior includes: slot filling behavior, knowledge query behavior, session termination behavior, and tool invocation behavior.

[0033] As an optional implementation, the slot filling behavior includes: if the slot value already exists, then store the slot value; otherwise, store the slot value in the slot information of the task node. The knowledge query process includes: generating a query question, retrieving relevant content from a pre-configured knowledge base based on the query question, and inputting the relevant content and the user question into a large language model deployed on a task node to obtain the corresponding answer; The session termination behavior includes: determining whether the preset termination conditions are met based on historical dialogue data; if so, terminating the task node and proceeding to the next node in sequence. The tool invocation behavior includes: using the large language model deployed in the task node to determine whether a tool needs to be invoked; if a tool needs to be invoked, returning the required function and parameters; after invoking the tool, sending the invocation result back to the large language model deployed in the task node; and the large language model deployed in the task node summarizing or planning subtasks based on the invocation result.

[0034] The task-oriented dialogue robot system based on a large language model proposed in this application has the following significant advantages: First, it significantly enhances the flexibility and intelligence of the dialogue system. By introducing a large language model, the system can deeply understand users' diverse, colloquial, and even ambiguous expressions, reducing reliance on strict rule matching and greatly enhancing its ability to handle complex semantics and ambiguous intentions, making the dialogue process closer to natural interpersonal interaction; Secondly, it effectively reduces the complexity of system configuration and maintenance. Traditional task-oriented dialogue systems typically require a large amount of manual configuration of rules and states. This system, through a visual canvas and a composable node mechanism, supports flexible and intuitive process construction. In particular, through intelligent nodes driven by a large language model (such as LLM task nodes and large language model judgment nodes), it realizes automatic inference and dynamic decision-making of dialogue logic, reducing the burden of manually writing complex rule sets. Third, it achieves efficient guidance and robust handling of multi-round task dialogues. The system has stronger context maintenance and slot management capabilities, and can automatically handle complex scenarios such as information collection, error correction, process switching, and knowledge retrieval. Through built-in conflict resolution and priority mechanisms, it ensures the continuity of the main task line, while supporting seamless transfer to manual intervention or creation of work orders at appropriate times, balancing automation and manual intervention; Fourth, it possesses excellent scalability and applicability. The system architecture supports deep integration with external APIs and knowledge bases, capable of handling highly structured information registration tasks as well as open-domain intelligent question answering and complex business processing, demonstrating strong industry versatility and scenario adaptability.

[0035] In one exemplary embodiment, a task-oriented chatbot method based on a large language model is provided, implemented using the aforementioned task-oriented chatbot system based on a large language model. The task-oriented chatbot method based on a large language model includes: Obtain input information; The input information is fed into the large language model deployed in the process control module to obtain the initial decision; Prioritize and resolve conflicts in the initial decisions to arrive at the target decision; The action execution module is used to engage in corresponding dialogues with the user, execute corresponding actions, and select the corresponding nodes in the node module according to the preset execution logic of each decision in the target decision. Each node in the node module performs an action according to its corresponding execution rules.

[0036] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0037] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0038] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the system and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A task-oriented dialogue robot system based on a large language model, characterized in that, The task-oriented dialogue robot system based on a large language model includes: a process control module, an action execution module, and a node module connected in sequence; the node module includes multiple nodes. The process control module is used for: Obtain input information; The input information is fed into the large language model deployed in the process control module to obtain the initial decision; The initial decisions are prioritized and conflict-handled to obtain the target decision, which is then sent to the action execution module. The action execution module is used to conduct corresponding dialogues with the user, execute corresponding behaviors, and select the corresponding nodes in the node module according to the preset execution logic of each decision in the target decision. Each node in the node module is used to perform actions according to the corresponding execution rules.

2. The task-oriented dialogue robot system based on a large language model according to claim 1, characterized in that, The input information includes: control commands, process information of candidate processes, process information of the current process, historical dialogue data, and user questions; the process information includes the process name, description, and slot information of the process; the slot information includes the name, description, type, and value of the slot.

3. The task-oriented dialogue robot system based on a large language model according to claim 2, characterized in that, The initial decision includes one or more of the following: clarifying flow decision, activating flow decision, switching flow decision, canceling flow decision, filling slot decision, and responding to knowledge query decision.

4. The task-oriented dialogue robot system based on a large language model according to claim 3, characterized in that, The multiple nodes are: collection node, large language model judgment node, rule judgment node, interface call node, response node, work order node, and task node.

5. The task-oriented dialogue robot system based on a large language model according to claim 4, characterized in that, The initial decisions are prioritized and conflict-handled to obtain the target decision, including: Determine whether process control decisions and response knowledge query decisions coexist in the initial decision-making process to obtain a first determination result; process control decisions include at least one of clarifying flow decisions, activating flow decisions, switching flow decisions, and canceling flow decisions. If the first judgment result is yes, then the response knowledge query decision is removed from the initial decision to obtain a second decision, and the process control decision is checked for conflict according to the preset conflict handling method to obtain the target decision; If the first judgment result is negative, then it is determined whether there is a process control type decision in the initial decision, and a second judgment result is obtained; If the second judgment result is yes, then according to the preset conflict handling method, the process control decision is conflict checked to obtain the target decision; If the second judgment result is negative, then the initial decision is determined as the target decision.

6. The task-oriented dialogue robot system based on a large language model according to claim 5, characterized in that, The preset conflict handling method includes: When both cancel flow decision and activate flow decision exist in the process control class, and the process name in the activate flow decision is the name of the current process, the cancel flow decision and activate flow decision will be removed from the process control class. When both cancel flow decision and activate flow decision exist in the process control category, and the process name in the activate flow decision is the name of the new process, the cancel flow decision and activate flow decision will be merged into a switch flow decision. When the process name in the clarifying flow decision in the process control class is the name of the current process, the clarifying flow decision will be removed from the process control class. When the process name in the flow switching decision in the process control class is the name of the current process, the flow switching decision will be removed from the process control class.

7. The task-oriented dialogue robot system based on a large language model according to claim 4, characterized in that, The default execution logic for clarifying flow decisions is as follows: using the large language model deployed in the action execution module, a prompt message is generated based on the process name and description of the process in the activation flow decision to guide the user to confirm the process; The default execution logic for activating flow decision is: notify the user of the process name of the process they have entered, and select the corresponding node in the node module according to the process configuration. The process configuration includes: process name, description, and corresponding nodes; The default execution logic for switching flow decisions is as follows: inform the user to switch to the new flow, update the current flow to the new flow, and select the corresponding node in the selected node module according to the flow configuration of the updated current flow. The default execution logic for canceling a flow decision is as follows: inform the user that the flow has been canceled, update the current flow to empty, and inquire about the user's subsequent needs; The default execution logic for filling slots is as follows: update slot information; if the value of a slot is detected to have been replaced, notify the user of the change in the slot value; implement a validation strategy for the slot value; if the validation passes, store the slot value and select the corresponding node in the node module. The default execution logic for responding to knowledge query decisions is as follows: directly answer the user's knowledge-based questions, bring the conversation back to the original process, and prompt the user to supplement the information that was not completed in the previous process.

8. The task-oriented dialogue robot system based on a large language model according to claim 4, characterized in that, The execution rule for the collection node is: determine whether the value of the slot in the current process has been stored; if it has been stored, then execute the next node sequentially. If not stored, generate a collection prompt script based on the name and description of the slot in the current process and output it to the user; The execution rule of the large language model judgment node is as follows: input the user question, historical dialogue data, and the names and descriptions of the conditions configured in the large language model judgment node into the large language model deployed in the large language model judgment node, obtain the corresponding condition names, and execute the next node in sequence; The execution rule for the rule judgment node is as follows: determine whether the rule judgment condition is met, obtain the rule judgment result, and execute the next node sequentially based on the rule judgment result; The rule judgment conditions are: the value of the slot in the current process has a value, the value of the slot in the current process is equal to the first preset value, the value of the slot in the current process is greater than the second preset value, the value of the slot in the current process is less than the third preset value, or the value of the slot in the current process contains the fourth preset value. The execution rules for the API call node are as follows: based on the configured API address, input parameters, and output parameters, the API is invoked and data is obtained. The output parameters are assigned to the corresponding slots, and the next node is executed sequentially. The execution rule for the response node is: if there is no next node, the task process ends; otherwise, return to the message configured for the node and execute the next node sequentially. The execution rules for the work order node are as follows: input historical dialogue data and user questions into the large language model deployed in the work order node to generate the corresponding work order title and work order summary, summarize the information to be collected for the work order to generate the work order, and execute the next node in sequence; The execution rules for task nodes are as follows: Input the task objective, task description, termination condition, task node slot information, historical dialogue data, and user questions into the large language model deployed in the task node to obtain the speech and output behavior. The output behaviors include: filling slot behavior, knowledge query behavior, ending session behavior, and tool invocation behavior.

9. The task-oriented dialogue robot system based on a large language model according to claim 8, characterized in that, The slot filling behavior includes: if the slot value already exists, then store the slot value; otherwise, store the slot value from the slot information of the task node. The knowledge query process includes: generating a query question, retrieving relevant content from a pre-configured knowledge base based on the query question, and inputting the relevant content and the user question into a large language model deployed on a task node to obtain the corresponding answer; The session termination behavior includes: determining whether the preset termination conditions are met based on historical dialogue data; if so, terminating the task node and proceeding to the next node in sequence. The tool invocation behavior includes: using the large language model deployed in the task node to determine whether a tool needs to be invoked; if a tool needs to be invoked, returning the required function and parameters; after invoking the tool, sending the invocation result back to the large language model deployed in the task node; and the large language model deployed in the task node summarizing or planning subtasks based on the invocation result.

10. A task-oriented dialogue robot method based on a large language model, implemented using the task-oriented dialogue robot system based on a large language model as described in any one of claims 1-9, characterized in that, The task-oriented dialogue robot method based on a large language model includes: Obtain input information; The input information is fed into the large language model deployed in the process control module to obtain the initial decision; The initial decisions are prioritized and conflict-handled to obtain the target decision; The action execution module is used to conduct corresponding dialogues with the user, execute corresponding behaviors, and select the corresponding nodes in the node module according to the preset execution logic of each decision in the target decision. Each node in the node module performs an action according to its corresponding execution rules.

Images