Intelligent task-type dialogue method and system based on large language model, device, and program product

Abstract

An intelligent task-type dialogue method based on a large language model, comprising: acquiring a dialogue with a user; identifying and parsing input information of the user to obtain an identification and parsing result; on the basis of the identification and parsing result, synchronously updating the state of a dialogue state tracker; on the basis of the intelligent guidance of the tracker, entering a proper dialogue scene node; executing a preset Action to complete a specific task; and on the basis of the current dialogue node context and dialogue historical data, generating a feedback message of a robot. In the method, by combining the large language model, rapid adaptation and efficient configuration of a scene can be achieved, and by dynamically loading the configuration information of a specific scene, a round of dialogue process can be efficiently pushed on the basis of preset process logic. Also disclosed are an intelligent task-type dialogue system based on a large language model, an electronic device, and a computer program product.

Description

Intelligent task-oriented dialogue method, system, device and program product fusing large language model

[0001] This application claims priority to the Chinese patent application No. 202410658940.3, filed on May 24, 2024, and entitled "Intelligent task-oriented dialogue method, system, device and program product fusing large language model", the entire content of which is incorporated herein by reference. TECHNICAL FIELD

[0002] The present application relates to the field of artificial intelligence, and in particular to an intelligent task-oriented dialogue method and system fusing a large language model. BACKGROUND

[0003] In the field of natural language processing, intelligent task-oriented dialogue systems have always played a key role and are widely used in various industries to assist artificial customer service in efficiently handling various transactions. Such systems are usually focused on completing tasks in specific scenarios, such as flight booking, conference room booking, product promotion, etc., through real-time dialogue interaction with users, accurately capturing the specific information necessary to complete the task, and flexibly switching between different dialogue processes with high-precision user intent recognition technology, and then feeding back appropriate answer strategies. Traditional task-oriented dialogue frameworks, such as RASA, build a complete task processing structure by constructing scenario stories, defining user intents, identifying key entities, designing dialogue templates, and presetting execution actions. They also use labeled data to train intent recognition and entity extraction models to form a powerful natural language understanding (NLU) module. However, with the rapid development of large language model technologies such as ChatGPT, existing intelligent task-oriented dialogue systems such as the RASA framework also face some limitations and challenges.

[0004] Firstly, the dependence on data is relatively strong. Traditional task-oriented dialogue systems usually require a large amount of labeled data for model training so that the model can understand and process the intent and entity in the user's dialogue. For task-oriented dialogue systems in specific domains, data collection and labeling are needed to prepare for model training. Whether training two models for intent classification and entity recognition or training a joint model to complete both tasks, a large amount of training data is required. The models for intent classification and entity recognition can only handle the content contained in the training data. When the scenario needs to be expanded or new scenarios are introduced, new intent types and new entities are often introduced, which requires retraining or secondary training of the previously trained model, which is a time-consuming and labor-intensive step.

[0005] Secondly, in terms of dialogue generation, traditional methods often lack diversity. Systems usually respond based on pre-set dialogue templates, and although the diversity of responses can be increased by configuring different templates, the overall diversity is still limited. Users are likely to encounter repetitive response dialogues when trying multiple times, which reduces the user experience.

[0006] In addition, there is a lack of consistency and coherence in the dialogue. In traditional task-oriented dialogue frameworks, the system's response content mainly depends on the current dialogue flow position. For example, in the flight booking scenario, after asking for the departure location and destination, the system usually directly asks for the departure date without considering the user's previous dialogue content and attitude. This mechanical response, although it achieves multi-round dialogue interaction, has relatively weak coherence and fluency in the overall dialogue, affecting the natural communication experience between users and the system. SUMMARY

[0007] To address the above problems, the present application proposes an intelligent task-oriented dialogue method and system that integrates a large language model. This system draws on the design philosophy of traditional task-oriented dialogue systems and combines the powerful capabilities of large language models to achieve rapid adaptation and efficient configuration of scenarios. By dynamically loading configuration information for specific scenarios, the system can efficiently advance the dialogue process according to the pre-set flow logic, greatly simplifying the data labeling and model training steps.

[0008] The present application provides an intelligent task-oriented dialogue method that integrates a large language model, specifically including:

[0009] Obtain dialogue with the user;

[0010] Identify and analyze the user's input information;

[0011] Synchronously update the state of the dialogue state tracker according to the results of the identification and analysis;

[0012] Intelligently guide the tracker to the appropriate dialogue scenario node;

[0013] Execute pre-set actions to complete specific tasks;

[0014] Generate feedback dialogues for the robot based on the current dialogue node context and complete dialogue history data.

[0015] Preferably, to adapt to different task-oriented dialogue scenarios, the system also needs to be carefully configured, covering a series of key elements: customized dialogue node design, entity information collection node setting, flow jump logic planning, and task-related execution actions.

[0016] The application provides a smart task-oriented dialogue system of a large language model, which realizes a more flexible and intelligent smart task-oriented dialogue system by fusing a large language model and a task-oriented dialogue module, and mainly comprises an intent entity recognition module, a state tracking module, an action execution module and a dialogue generation module.

[0017] (1) Task-oriented dialogue module

[0018] For different task scenarios, corresponding scene configuration work needs to be carried out. The configuration process usually includes the following key parts: dialogue node, information collection node, flow jump logic, task-related execution action.

[0019] Dialogue node, that is, configuring appropriate dialogue reply strategies on different nodes; bearing the task of guiding and shaping the interactive process. Each dialogue node is finely configured to ensure that it can accurately execute the preset dialogue generation strategy in a specific context, thereby promoting the smoothness and goal orientation of the dialogue process. For the dialogue generation strategy, the key is to configure its corresponding prompt. An effective prompt can clearly convey the intent, guide the user into the relevant dialogue process, and maintain a high degree of naturalness. In the subsequent NLG dialogue generation module, it will play a crucial role.

[0020] Information collection node, according to the specific task requirements, asks the user questions to collect relevant entity information; responsible for collecting the entity information needed to complete the task. For each entity information, in addition to its name, the entity type, candidate and example of each entity also need to be configured, which will be used in the prompt construction of NLU entity recognition. For the information collection node itself, when the dialogue process jumps to this node, the dialogue strategy of this node is "according to the collected entity information and the entity information that is still empty, ask the user questions, so as to supplement all the required entities".

[0021] Flow jump logic, the key to ensure smooth connection and switching between different nodes. According to the type of node, the corresponding jump logic will be set. For example, after the user completes information collection, the system will perform a secondary confirmation on the collected information. In this process, the system will judge the user's intent according to the user's answer. If the user expresses the intent of "to modify" or "not to modify", the flow will jump according to this intent and guide the user to the corresponding node. In the information collection stage, the decision of flow jump will also be based on the completeness of the current information. When configuring these jump logics, a brief description or introduction will usually be written for each intent logic. It not only helps the clarity of flow design, but also serves as part of the prompt construction for natural language understanding (NLU) intent recognition, providing more comprehensive information reference for large models when performing specific tasks.

[0022] Task-related execution actions are a series of operations triggered at a specific position in the dialogue flow. These actions support access to various functional interfaces, such as querying databases, updating user account status, sending notification emails, etc., to meet the needs of different task scenarios. In practical applications such as booking restaurants, purchasing goods, and making appointments, action execution plays a key role in ensuring that the dialogue can directly drive the execution of business operations. By embedding execution actions in the flow of task-oriented dialogue, not only does it enrich the interaction mode of dialogue, but also makes dialogue no longer limited to simple information exchange, but can directly act on the promotion of business processes. With the continuous development and change of business needs, actions can be conveniently added, modified or deleted to connect new interface services to adapt to the evolving business logic. This flexibility enables the task-oriented dialogue system to continuously meet the diverse needs of users and provide a more efficient and convenient service experience.

[0023] (2) NLU intent entity recognition

[0024] The NLU module is mainly responsible for identifying the intent and entity information in the user's speech. In this invention, the module efficiently completes the tasks of intent recognition and entity recognition with the help of the capabilities of large language models: through the construction of prompts, appropriate instructions are given to drive large language models to complete specific tasks.

[0025] In the intent recognition part, the system determines the possible child nodes of the current node according to the current dialogue scenario and dialogue progress, obtains the intent logic linking the current node and the child nodes, and takes it as the current candidate intent set. Since the system supports highly free dialogue, it also includes "chitchat, repetition, transfer to manual" and other intents in the global intent. Each intent is configured with a corresponding intent description to clarify its meaning. For the construction of prompts, the following example template can be referred to: "You are an intent recognition assistant, you need to return the most appropriate intent name, when you cannot judge the correct intent, you can return the chitchat intent. The current dialogue topic is xxx. All candidate intent names and descriptions are: {need to book: indicates that the user has the idea or demand to book a conference room, do not need to book: indicates xxx,...} Please classify the following speech according to intent, and only return the intent name, do not return other content. The current speech is: xxx. The intent is: " Here, the dialogue topic is a brief description of the task scenario, and all candidate intents are the union of the current candidate intents and global intents. The large language model can return the most appropriate intent name according to the candidate intents and descriptions in the prompt, combined with the current speech, to achieve the purpose of intent recognition.

[0026] In the entity recognition part, the system mainly determines whether to update or modify the entity slot information based on the currently recognized entity slot information and the user's current utterance, so as to achieve the purpose of entity recognition. The prompt construction in this part is different from that of intent recognition. The template can be as follows: "You are an entity maintenance assistant in a multi-turn dialogue task. Please update and maintain the entity information based on the existing entity information and the latest utterance. For the entity information that has not been recognized, keep its original value. All the entity information to be recognized is as follows: xxx. The current entity information slot_info: xxx. Please recognize the entity information in the following utterance and update slot_info. The updated slot_info is: " The large language model can return the updated entity information based on the current entity information in the prompt and the user's current utterance. The advantage of this is that it can closely integrate entity recognition and entity update. In some special scenarios, users may modify the previously described entity information. This method can handle this situation well and maintain the entity slot information in the entire task-oriented dialogue.

[0027] (3) DST Tracker state tracking

[0028] The state tracker is an important component of the task-oriented dialogue framework, which is mainly responsible for recording and updating the user's dialogue state in multi-turn dialogue. It mainly includes two aspects: recording and updating.

[0029] The state tracker will record the following information in the process of task-oriented dialogue, including:

[0030] Scene information: For scenarios, including but not limited to scenario type, scenario description, complete flow information contained in the scenario, scenario-specific rules or constraints, etc., to ensure that the dialogue system can understand and respond to the user's utterance in the correct context.

[0031] Current node: In task-oriented dialogue, the dialogue process is usually modeled as a directed graph, and each node represents a stage or decision point of the dialogue. The state tracker will record the current dialogue progress to which node, which helps the dialogue system determine the next dialogue behavior and possible dialogue path.

[0032] Entity slot information: The state tracker will continuously record the recognized entity slot information in the current dialogue and the status of the unfilled slots, which helps the system ask the user for questions about the unfilled slots during the chat process of the task-oriented dialogue, thereby improving the efficiency and accuracy of information collection.

[0033] Dialogue history information: The state tracker saves the complete dialogue history of the user, which helps the large model to generate more natural and consistent chat scripts by referring to more complete dialogue information.

[0034] The state tracker also dynamically maintains and updates the above information during the process of the dialogue flow. When a specific user intent is recognized, appropriate node jumping is performed according to the current process position, and the information content of the current node is dynamically updated after node jumping. For entity slot information, the entity information recorded in the current state tracker is dynamically updated when the model recognizes specific entity information during the dialogue process and at specific information collection nodes. For dialogue history information, it is updated during the process of interacting with the user.

[0035] (4) Action action execution

[0036] The core function of the action execution module is to trigger and execute operations closely related to the current task at specific links of the task-oriented dialogue flow. These operations may involve querying databases, calling external interfaces, etc., aiming to immediately meet user needs and promote the smooth progress of the dialogue flow.

[0037] The action execution module has strong communication capabilities with external systems, including but not limited to databases, API interface services, and enterprise internal systems. It can obtain the results of action execution through pre-set action interaction mechanisms and quickly feed back these results to the dialogue system. In some scenarios, these execution results may also affect the state information of the state tracker, thereby affecting the strategy selection during dialogue generation.

[0038] The action execution module realizes efficient collaboration between the dialogue system and external resources by accurately triggering and executing various actions, greatly expanding the functional range of the dialogue system.

[0039] (5) NLG script generation

[0040] The NLG module is mainly responsible for script generation in the dialogue system, and the generated script will be directly displayed as a robot reply script to the user. In this invention, the module will use the powerful text generation capability of the large language model to generate system reply scripts. Through the construction of prompts, the large model is driven by reasonable instructions to generate the most appropriate reply content.

[0041] In the dialogue generation process, we mainly consider two key parts. First, the dialogue generation strategy of the current dialogue node, which is defined and refined in the scene configuration stage, and the language style, dialogue purpose, pre-knowledge and generation suggestions that should be followed when the system generates dialogue in a specific node. Second, the user's historical dialogue information is also an important reference for generating dialogue. When the NLG module receives these two parts of information, it will guide the large language model to generate dialogue according to the pre-configured strategy. Ensure that the generated dialogue content not only follows the set dialogue logic and business rules, but also maintains the coherence and consistency of the dialogue.

[0042] When constructing the input instruction prompt of the large language model, the template example is as follows:

[0043] "You are a multi-turn dialogue answer generation assistant in an xxx scenario, please refer to the current dialogue and the current requirement to generate a suitable latest bot dialogue. When the latest user dialogue is a question, you must first answer his question briefly, and then connect a segment of dialogue to meet the requirement.

[0044] Current dialogue content: [xxx, xxx, xxx, xxx]

[0045] Requirement: xxx.

[0046] Please generate the latest bot answer: "

[0047] Among them, the scene name is a brief introduction to the scene of the task-based dialogue, the current dialogue content is the complete historical dialogue information, and the requirement is the dialogue generation strategy configured for each node, for example, in the conference room reservation scenario, the greeting node can be configured as follows. Strategy: "Greet the user friendly, tell him that you are a conference room reservation assistant, and ask if he has a conference room reservation requirement." Through the configuration of the above prompt template, the large model can give the most appropriate response to the user's question and dialogue, and then pull the dialogue content back to the task theme. The generated reply dialogue not only maintains professionalism and consistency, but also can promote the process of task-based dialogue, significantly improving the interaction quality and efficiency of the dialogue system.

[0048] Based on the intelligent task-based dialogue method proposed above, in order to better realize the present application, an electronic device is further proposed, comprising a memory and a processor; the memory stores a computer program; when the computer program is executed on the processor, the intelligent task-based dialogue method of the large language model is realized.

[0049] In order to better achieve the present application, based on the intelligent task type dialogue method based on the large language model proposed above, the present application further provides a computer program product comprising computer programs / instructions which, when executed by a processor, implement the steps of the intelligent task type dialogue method based on the large language model. Advantages

[0050] The present application adopts a large language model as a replacement solution, without relying on specific training data set preparation and time-consuming model training process, but can guide the model to generate the required answer by means of instruction prompts. After experiencing large-scale data pre-training, such a large language model not only has a deep natural language knowledge base, but also has excellent human instruction execution capability. The reasoning mechanism of the large language model presents a zero-shot characteristic, without additional model training and a large amount of special training data, but only needs to specify the task category and related parameters when constructing the prompt, and assist with one or two typical examples as a reference, and the model can respond to new tasks. This method significantly reduces the data size and human labor cost required for model training, and for new intent categories or entity types that may appear in the future, only the prompt setting needs to be updated, which can achieve rapid adaptation and iterative optimization, greatly improving the flexibility and efficiency of the system.

[0051] In addition, for the intelligent reply module of the dialogue system, the powerful language dialogue capability of the large language model will be used to configure the prompt, so that the large language model can generate appropriate dialogue content at each scene node. The large language model can generate rich and diverse dialogue while keeping the original meaning of the dialogue unchanged. On the other hand, the large language model has strong long text processing capability, and when generating dialogue, the past several rounds of historical dialogue records are also transmitted to the large model as part of the prompt. This design ensures that the generated answer can closely connect with the previous dialogue context, forming a coherent and logically consistent dialogue flow. This way greatly enhances the authenticity of the dialogue and the user experience, making intelligent reply no longer limited to pre-set templates, and can flexibly cope with complex scenarios and changes in user problems. BRIEF DESCRIPTION OF DRAWINGS

[0052] FIG. 1 is a flowchart of an intelligent task type dialogue method and system based on a large language model;

[0053] FIG. 2 is a flowchart of a task type dialogue scene configuration example of a conference room reservation in an intelligent task type dialogue method and system based on a large language model. DETAILED DESCRIPTION

[0054] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0055] According to the flow shown in Figure 1, the present application provides an intelligent task-oriented dialogue method based on a large language model, which specifically comprises:

[0056] Obtaining a dialogue with a user;

[0057] Identifying and analyzing the input information of the user;

[0058] Synchronously updating the state of the dialogue state tracker according to the result of the identification and analysis;

[0059] Guiding to a proper dialogue scene node according to the tracker;

[0060] Executing a pre-set action to complete a specific task;

[0061] Generating a feedback script of a robot based on the current dialogue node context and complete dialogue history data.

[0062] Preferably, in order to adapt to different task-oriented dialogue scenes, the system also needs to be finely configured, covering a series of key elements: customized script node design, entity information collection node setting, flow jump logic planning, and task-related execution actions.

[0063] The present application provides an intelligent task-oriented dialogue system based on a large language model, which mainly realizes a more flexible and intelligent intelligent task-oriented dialogue system by fusing a large language model and a task-oriented dialogue module, which mainly comprises: an NLU (Natural Language Understanding) intent entity recognition module, a DST Tracker (Dialogue State Tracking) state tracking module, an Action action execution module, and an NLG (Natural Language Generation) script generation module.

[0064] (1) Task-oriented dialogue scene configuration

[0065] For different task scenarios, corresponding scene configuration work needs to be carried out. The configuration process usually includes the following key parts: one is the dialogue node, that is, configuring appropriate dialogue reply strategies on different nodes; two is the information collection node, according to the specific task demand, asking the user questions to collect relevant entity information; three is the flow jump logic, which clearly defines the jump logic between different nodes; and finally the task related execution action, that is, triggering the corresponding action at a specific scene position.

[0066] According to the flow shown in Figure 2, a conference room reservation scenario is set, and then each module is analyzed in detail.

[0067] First, the dialogue node carries the task of guiding and shaping the interaction process. Fine-grained configuration is carried out for each dialogue node to ensure that it can accurately execute the preset dialogue generation strategy in a specific context, thereby promoting the smoothness and goal orientation of the dialogue process. Take the greeting node as an example, which aims to create a friendly chat atmosphere while opportunistically introducing the core service subject of this scenario, i.e. conference room reservation. For dialogue generation strategy, the key is to configure its corresponding prompt, an effective prompt can clearly convey the intention, guide the user into the relevant dialogue process, while maintaining a high degree of naturalness. The NLG dialogue generation module will play a crucial role in the subsequent process.

[0068] Second, the information collection node is mainly responsible for collecting the entity information needed to complete the task. In the conference room reservation example, the information to be collected includes: "conference room name", "meeting time" and "attendees". For each entity information, in addition to its name, the entity type, candidate and example of each entity also need to be configured, which will be used in the construction of the prompt for NLU entity recognition. For the information collection node itself, when the dialogue flow jumps to this node, the dialogue strategy of this node is "according to the entity information already collected and the entity information still missing, ask the user questions, so as to supplement all the required entities".

[0069] Third, flow jump logic, which plays the role of a serial node. According to different node types, different types of jump logic will be configured. For example, after the user information collection is completed, the collected information will be confirmed again, and at this time, according to the user's answer, the intention will be judged, and when the intention is "to modify" or "not to modify", the flow will be jumped, and according to the user's intention, it will be jumped to the corresponding node. When it is an information collection node, the jump logic will be determined according to the completeness of the current information collection. In this example, only when the information collection related to the conference room reservation is completed, it will jump to a new node. When configuring the jump logic, a brief introduction will often be written for each intention logic, such as "need to reserve: indicates that the user has the idea or appeal of reserving a conference room", and such an introduction will also be part of the prompt construction for NLU intention recognition, aiming to enable the large model to complete specific tasks while referring to more comprehensive information.

[0070] Fourth, task-related execution actions, which refer to actions triggered at a specific position in the dialogue flow. Actions can support the access of various functional interfaces, such as querying databases, updating user account status, sending notification emails, etc. In various task scenarios such as restaurant reservation, product purchase, service reservation, etc., action execution can play a key role, realizing the deep integration of dialogue and business processes. In this example, the action of conference room reservation represents the trigger of the conference room reservation interface, and when the required information is collected, the corresponding interface can be called to complete the reservation process of the conference room. By adding execution actions in the flow of task-oriented dialogue, the entire dialogue flow can be more diverse and rich, not limited to information exchange, but also directly driving the execution of business operations. With the development and change of scene requirements, actions can be conveniently added, modified or deleted, and new interface services can be connected to adapt to the evolving business logic.

[0071] (2) NLU intention entity recognition

[0072] The NLU module is mainly responsible for identifying the intention and entity information in the user's speech, and in this invention, the module efficiently completes the tasks of intention recognition and entity recognition with the help of the large language model. The core method is to give appropriate instructions to drive the large language model to complete specific tasks through the construction of prompts.

[0073] For the intent recognition part, first we will get the sub-nodes of the current position according to the configured scene information and the current progress of the dialogue to a certain node position, and we get the intent logic between the current node and the sub-node as the current candidate intent set. At the same time, since our dialogue system is a system with high degree of freedom of dialogue, we will also add the intents of "chitchat, repeat, and transfer to manual" as global intents. For each intent, we will configure the corresponding intent description when configuring the scene, for example, "need to reserve": "indicates that the user has the idea or appeal of reserving a conference room"; "transfer to manual": "indicates that the user wants to involve the manual customer service". For the construction of the prompt, the following example template can be referred to: "You are an intent recognition assistant, you need to return the most appropriate intent name, when you cannot judge the correct intent, you can return the chitchat intent. The current dialogue topic is xxx. All candidate intent names and descriptions are: {need to reserve: indicates that the user has the idea or appeal of reserving a conference room, do not need to reserve: indicates xxx,...} Please classify the following dialogue as an intent, and only return the intent name, do not return other content. The current dialogue is: xxx. The intent is: " Here, the dialogue topic is a brief description of the task scene, and all candidate intents are the union of the current candidate intents and global intents. The large language model can return the most appropriate intent name according to the candidate intents and descriptions in the prompt, combined with the current dialogue, so as to achieve the purpose of intent recognition.

[0074] For the entity recognition part, it is similar to the intent recognition. Mainly according to the current recognized entity slot information and the user's current utterance, it is judged whether the entity slot information needs to be modified or updated, so as to achieve the purpose of entity recognition. The prompt construction of this part is different from that of intent recognition, which can refer to the following template: "You are an entity maintenance assistant in a multi-turn dialogue task, please update and maintain the entity information according to the existing entity information and the latest utterance, and keep the original value for the entity information that has not been recognized. All the entity information that needs to be recognized is as follows: xxx. The current entity information slot_info: xxx, please identify the entity information in the following utterance and update slot_info, and return the updated slot_info content. Utterance: xxx. The updated slot_info is: " All the entity information that needs to be recognized is the entity description information configured in the scene configuration, which includes entity name, type, candidate value, etc. The current entity information slot_info represents an entity information record table that records the recognition status of each entity, and None indicates that it has not been recognized. The large language model can return the updated entity information according to the current entity information in the prompt and the user's current utterance. The advantage of this is that it can closely combine entity recognition and entity update, and in some special scenarios, the user may modify the previously described entity information, and this method can handle this situation well, and can maintain the entity slot information in the entire task-based dialogue well.

[0075] (3) DST Tracker state tracking

[0076] The state tracker is an important component in the task-based dialogue framework, which is mainly responsible for recording and updating the user's dialogue state in multi-turn dialogue. Here we will expand the state tracker from recording and updating two aspects.

[0077] On the one hand, the state tracker will record the following information in the process of task-based dialogue: scene information, current node, entity slot information, dialogue history information.

[0078] Scene information: For the scene, including but not limited to scene type, scene description, complete flow information contained in the scene, scene-specific rules or constraints, etc., to ensure that the dialogue system can understand and respond to the user's utterance in the correct context.

[0079] Current node: In task-based dialogue, the dialogue process is usually modeled as a directed graph, and each node represents a stage or a decision point of the dialogue. The state tracker will record in real time which node the current dialogue has progressed to, which helps the dialogue system determine the next dialogue behavior and possible dialogue path.

[0080] Entity slot information: The state tracker continuously records the entity slot information identified in the current conversation and the status of unfilled slots, which helps the system ask the user questions about unfilled slots during the chat process of a task-oriented conversation, thereby improving the efficiency and accuracy of information collection.

[0081] Conversation history information: The state tracker saves the user's complete conversation history, which helps the large model generate chat tactics by referring to more complete conversation information, and the conversation content generated by combining historical records will be more natural and consistent.

[0082] On the other hand, the state tracker also dynamically maintains and updates the above-mentioned content during the conversation flow. When a specific user intent is recognized, appropriate node jumping is performed according to the current process position, and the information content of the current node is dynamically updated after node jumping. For entity slot information, when the model recognizes specific entity information during the conversation process and at specific nodes of information collection, the entity information recorded in the current state tracker is dynamically updated. For conversation history information, it is updated during the conversation interaction with the user.

[0083] The state tracker accurately records and timely updates the core elements of the conversation state, providing a stable and consistent conversation management foundation for the task-oriented conversation system, ensuring smooth progress of the conversation process and efficient achievement of the task goal.

[0084] (4) Action execution

[0085] The main purpose of the action execution module is to trigger and execute operations closely related to the current task at specific nodes of the task-oriented conversation process, such as querying databases, calling external interfaces, etc. The purpose is to achieve immediate satisfaction and smooth progress of user demand.

[0086] The action execution module has the ability to communicate with external systems (such as databases, API interface services, enterprise internal systems, etc.). It can obtain the results of action execution through pre-set action interactions and feed back the results to the conversation system, which in some scenarios will also affect the state information of the state tracker and the strategy during conversation generation. For example, in the shopping consultation scenario, it can query the inventory data in the database according to the product asked by the user, and return the inventory situation of the product to the user; in the travel reservation scenario, it can obtain the relevant reservation situation by calling the API interface of the third-party reservation platform; in the conference room reservation scenario, it can connect to the enterprise internal OA system to complete the action of reserving a conference room.

[0087] In summary, the action execution module triggers and executes various actions accurately, realizing the efficient linkage between the dialogue system and external resources, greatly expanding the functional boundaries of the dialogue system and enabling it to handle more complex business processes.

[0088] (5) NLG dialogue generation

[0089] The NLG module is mainly responsible for dialogue generation in the dialogue system, and the generated dialogue will be directly displayed to the user as a robot reply dialogue. In this invention, the module will generate system reply dialogue with the help of the powerful text generation capability of large language models. The core method is to generate the most suitable reply content by constructing prompts and driving large models with reasonable instructions.

[0090] When generating reply dialogue, the following two parts will be mainly considered. One part is the dialogue generation strategy configured by the current node, and the invention will define and refine the dialogue generation strategy for each dialogue node in the scene configuration stage. These strategies mainly stipulate that under certain nodes, the system should follow the language style, dialogue purpose, pre-known information, generation suggestions, etc. when generating dialogue. The other part is the historical dialogue information of the user. After receiving the above two parts of content, the NLG module will guide the large language model to generate dialogue according to the pre-configured strategy when generating dialogue, ensuring that the generated dialogue content follows the set dialogue logic and business rules, while maintaining the coherence and consistency of the dialogue. When constructing the input instruction prompt of the large language model, the following template can be referred to:

[0091] "You are a multi-turn dialogue answer generation assistant in an xxx scenario, please refer to the current dialogue and the current requirement to generate a suitable latest bot dialogue. When the latest user dialogue is a question, you must first answer his question briefly, and then connect a segment of dialogue to meet the requirement.

[0092] Current dialogue content: [xxx, xxx, xxx, xxx]

[0093] Requirement: xxx.

[0094] Please generate the latest bot answer: "

[0095] Wherein the scene name is a brief introduction of the scene of the task-oriented dialogue, the current dialogue content dialogue is complete historical dialogue information, and the requirement requirement is the configured dialogue generation strategy of each node. For example, in the conference room reservation scene, the greeting node can be configured as follows: "greet the user politely, tell the other party that you are a conference room reservation assistant, and ask whether the other party has a conference room reservation requirement." Through the configuration of the above prompt template, the large model can give the most appropriate response to the user's question and dialogue, and then pull the dialogue content back to the task theme. The generated reply dialogue can keep the professionalism and consistency while also advancing the process of the task-oriented dialogue, which can significantly improve the interaction quality and efficiency of the dialogue system.

[0096] The embodiment further provides an electronic device including a memory and a processor; the memory stores a computer program; when the computer program is executed on the processor, the intelligent task-oriented dialogue method of the fusion large language model is realized.

[0097] The embodiment further provides a computer program product including computer programs / instructions, which, when executed by a processor, realize the steps of the intelligent task-oriented dialogue method of the fusion large language model.

[0098] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those having ordinary knowledge in the art can still modify the technical solutions recorded in the foregoing embodiments, or equivalently replace some or all of the technical features thereof; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application, and they should be covered in the scope of the claims and the specification of the present application.

Claims

1. An intelligent task-oriented dialogue system fused with a large language model, characterized in that, By fusing a large language model and a task-oriented dialogue module, a task-oriented dialogue is realized. The task-oriented dialogue module is composed of an intent entity recognition module, a state tracking module, an action execution module, and a dialogue generation module. The intent entity recognition module is configured to recognize the intent and entity information in the user dialogue. The state tracking module records and updates the dialogue state of the user in multiple rounds of dialogue. The action execution module triggers and executes operations closely related to the current task at specific nodes in the task-oriented dialogue process. The dialogue generation module is configured to generate dialogue in the dialogue system, and the generated dialogue will be directly displayed as a robot reply dialogue to the user. 2.The intelligent task-oriented dialogue system of a fusion large language model according to claim 1, wherein, The task-oriented dialogue module performs corresponding scene configuration work for different task scenarios. The configuration work includes dialogue nodes, information collection nodes, flow jump logic, and task-related execution actions. 3.The intelligent task-oriented dialogue system of the fusion large language model according to claim 1, characterized in that, The intent entity recognition module uses the large language model to identify intent and entity, and constructs a prompt to give appropriate instructions to drive the large language model to complete specific tasks. The intent recognition module determines the sub-nodes of the current node based on the scenario and progress of the dialogue, obtains the intent logic linking the current node and the sub-nodes, and uses it as the current candidate intent set. 4.The intelligent task-oriented dialogue system of a fusion large language model according to claim 1, wherein, The state tracking module records include scene information, current node, entity slot information, and dialogue history information. 5.The intelligent task-oriented dialogue system of a fusion large language model according to claim 1 and 4, characterized in that, The state tracking module dynamically maintains and updates the state tracking module record content during the dialogue flow. 6.The intelligent task-oriented dialogue system of a fusion large language model according to claim 1, wherein, The action execution module operations include querying the database and calling external interfaces. 7.The intelligent task-oriented dialogue system of the fusion large language model according to claim 1, wherein, The dialogue generation strategy of the current dialogue node in the dialogue generation module is predefined and refined in the configuration work. 8.The intelligent task-oriented dialogue system of a fusion large language model according to claim 1, wherein, The dialogue generation module generates dialogue by taking the historical dialogue information as an important reference.

9. An intelligent task-oriented dialogue method based on a large language model, specifically comprising: Obtaining a dialogue with a user; Identifying and analyzing the input information of the user to obtain the identification and analysis results; Synchronously updating the state of the dialogue state tracker according to the identification and analysis results; Intelligently guiding the tracker to the appropriate dialogue scene node; Executing the pre-set Action to complete a specific task; Generating the feedback dialogue of the robot based on the current dialogue node context and the dialogue history data.

10. An electronic device, comprising: A computer program is stored in a memory and executed by a processor; when the computer program is executed on the processor, the intelligent task-oriented dialogue method based on a large language model is realized.

11. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to realize the steps of the method of claim 9. The computer program / instructions are executed by the processor to realize the steps of the method of claim 9.

Images