Task-oriented dialog method, and storage medium and computer device

Abstract

The present invention relates to the technical field of the Internet. Disclosed are a task-oriented dialog method, and a storage medium and a computer device. The method comprises: in response to receiving dialog information initiated by a user, a client sending the dialog information to a server; in response to receiving the dialog information, the server acquiring user information of the user, and on the basis of the dialog information and the user information, determining intent information of the user by means of a large model; on the basis of the user information and the intent information, the server generating a planning scheme by means of the large model, and executing execution steps in the planning scheme, so as to obtain a scheme execution result; the server sending the scheme execution result to the client; and the client receiving and displaying the scheme execution result. The method can effectively improve the accuracy of intent recognition and the accuracy of scheme planning, reduce the workload of scheme enumeration, and increase the degree of matching between a scheme execution result and a user intent, thus satisfying the actual requirements of users.

Description

Task-oriented dialogue methods, storage media, and computer equipment Technical Field

[0001] This invention relates to the field of Internet technology, and in particular to a task-oriented dialogue method, storage medium, and computer device. Background Technology

[0002] Task-based dialogue refers to interactive conversations where users have specific goals or tasks, and the AI ​​system strives to assist them in achieving these goals or tasks. In task-based dialogues, the AI ​​system attempts to understand the user's intent and provides relevant information, suggestions, or performs corresponding actions to meet the user's needs. For example, if a user enters in a task-based dialogue, "Please help me book a meeting room for 3 PM tomorrow," the AI ​​system will understand the user's intent and automatically complete the task of booking the meeting room.

[0003] Task-oriented dialogue is commonly used in areas such as automated assistants for businesses or individuals, customer service, and smart home control. It can help improve work efficiency, enhance user experience, and reduce human intervention. Existing technologies for implementing task-oriented dialogue primarily rely on intent recognition, slot filling, and knowledge graphs to process natural language information. Specifically, it first identifies the intent within the dialogue information through text classification, then fills in slots based on entity recognition, and finally normalizes synonyms or entities using a knowledge graph to pinpoint the user's true intent and assist the user in completing the corresponding goal or task.

[0004] However, identifying intent in dialogue information through text classification or finding solutions through knowledge graphs requires the system to enumerate as many possible user questions and corresponding solutions as possible. However, these questions and solutions are difficult to apply to the large number of personalized user needs, have significant time lag, low solution matching accuracy, and fail to meet the actual needs of users.

[0005] Summary of the Invention

[0006] In view of this, the present invention provides a task-oriented dialogue method, storage medium and computer device, the main purpose of which is to solve the technical problems of low matching accuracy and large workload of manual enumeration of solutions in existing task-oriented dialogue schemes.

[0007] According to a first aspect of the present invention, a task-oriented dialogue method is provided, the method comprising: a client, in response to receiving dialogue information initiated by a user, sending the dialogue information to a server; the server, in response to receiving the dialogue information, obtaining user information of the user, and determining the user's intent information based on the dialogue information and the user information through a large model; the server, based on the user information and the intent information, generating a planning scheme through the large model, and executing the execution steps in the planning scheme to obtain a scheme execution result; the server sending the scheme execution result to the client; and the client receiving and displaying the scheme execution result.

[0008] According to a second aspect of the present invention, a task-oriented dialogue method is provided, the method comprising: in response to receiving dialogue information initiated by a user, sending the dialogue information; receiving and displaying a scheme execution result, wherein the scheme execution result is obtained by executing execution steps in a planning scheme, the planning scheme is generated by a large model based on the user's user information and intent information corresponding to the dialogue information, and the intent information is determined by the large model based on the dialogue information and the user information.

[0009] According to a third aspect of the present invention, a task-oriented dialogue method is provided, the method comprising: in response to receiving dialogue information initiated by a user, obtaining user information of the user, and determining the user's intent information through a large model based on the dialogue information and the user information; generating a planning scheme through the large model based on the user information and the intent information, and executing the execution steps in the planning scheme to obtain a scheme execution result; and sending the scheme execution result to the client so that the client receives and displays the scheme execution result.

[0010] According to a fourth aspect of the present invention, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described task-oriented dialogue method.

[0011] According to a fifth aspect of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described task-oriented dialogue method.

[0012] This invention provides a task-oriented dialogue method, storage medium, and computer device. Upon receiving dialogue information initiated by a user, the method first acquires the user's information. Then, based on the dialogue information and user information, it determines the user's intent information using a large-scale model. Furthermore, based on the user's information and intent information, it generates a planning scheme using the large-scale model. Finally, it executes the steps in the planning scheme to obtain the scheme execution result. This method utilizes the text analysis and text generation capabilities of the large-scale model to clarify the intent of the dialogue content and plan and execute solutions. This makes the intent recognition results closer to the user's true intent in the dialogue, and also allows the generated planning schemes to go beyond existing pre-defined schemes, generating new ones. This reduces the workload of scheme enumeration, improves the matching accuracy between the scheme execution result and the user's intent, and thus meets the user's actual requirements. Attached Figure Description

[0013] Figure 1 shows a flowchart of a task-oriented dialogue method provided by an embodiment of the present invention;

[0014] Figure 2 shows a flowchart of a task-oriented dialogue method provided by an embodiment of the present invention;

[0015] Figure 3 shows a flowchart of a task-oriented dialogue method provided by an embodiment of the present invention;

[0016] Figure 4 shows a flowchart of a task-oriented dialogue method provided by an embodiment of the present invention;

[0017] Figure 5 shows a flowchart of a task-oriented dialogue method provided by an embodiment of the present invention;

[0018] Figure 6 shows a schematic diagram of the storage structure of a task solution provided in an embodiment of the present invention;

[0019] Figure 7 shows a schematic diagram of the storage structure of a task solution provided by an embodiment of the present invention;

[0020] Figure 8 shows a scenario diagram of a task-oriented dialogue method provided by an embodiment of the present invention;

[0021] Figure 9 shows a scenario diagram of a task-oriented dialogue method provided by an embodiment of the present invention;

[0022] Figure 10 shows a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0023] The present invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present invention can be combined with each other.

[0024] In one embodiment, as shown in Figure 1, a task-oriented dialogue method is provided. Taking the application of this method to a client and a server as an example, the method includes the following steps:

[0025] 101. Upon receiving the dialogue information initiated by the user, the client sends the dialogue information to the server.

[0026] Specifically, when a user enters and sends any dialogue information through the client's information input box, the client can receive the user-initiated dialogue information and send it to the server. In this embodiment, the information input box can be any component capable of receiving text information; for example, it can be a chat input box, a search input box, etc. The dialogue information can be any text information entered by the user, such as a greeting "Hello," a question "Are you a robot?", a command "Can you help me make a travel plan?", etc. This embodiment does not impose specific limitations on this.

[0027] 102. Upon receiving the dialogue information, the server obtains the user's information and, based on the dialogue information and user information, determines the user's intent information through a large model.

[0028] Specifically, upon receiving a dialogue message initiated by a user, the server can obtain the user's information and input the obtained dialogue information and user information as parameters into a pre-trained large model. This allows the large model to respond to the user's dialogue messages based on the user information and determine the user's intent through the dialogue information. In this embodiment, the user information can be user feature information obtained with the user's authorization, such as user profile information, or the user's recent operations and status information, as well as historical dialogue information between the user and the artificial intelligence and the context information of the current dialogue, etc. The large model can be a generative language processing model such as a large language model or a multimodal model. This embodiment does not specifically limit the type of user information and the type of large model.

[0029] In this embodiment, the large model can determine the user's intent information through one or more rounds of dialogue. Specifically, when the intent in the user's initiated dialogue is ambiguous, the large model can only output the response information and wait for the user to initiate the next round of dialogue. When the intent in the user's initiated dialogue is clear, the large model can identify the intent information in the dialogue, then generate a planning scheme based on the intent information, execute the planning scheme, and send the execution result to the client for display. In addition, during the waiting time for the large model to plan and execute the scheme, the server can also output the response information of the dialogue to the client for display, so as to achieve timely response to the dialogue information.

[0030] Understandably, when the intent in the dialogue is ambiguous, the server does not need to generate a planning solution and execute it; therefore, steps 103 to 105 are unnecessary in this scenario. Conversely, when the intent in the dialogue is clear, the server can generate a planning solution and execute it, then send the execution result to the client for display. For example, when the dialogue is something like "Hello" or "Are you a robot?"—chat messages without a clear intent—the large model only needs to respond; no actual solution needs to be generated. However, when the dialogue is something like "What positive reviews have I recently received?" or "Can I pin a positive review?"—chat messages with a clear intent—the large model can use the dialogue and user information to clarify the user's intent, such as "What positive reviews have been added recently?" or "Can I pin a positive review myself?" This allows the large model to generate a corresponding solution based on the intent and output the solution to the client for display, thus answering the user's question.

[0031] It should be noted that, compared to existing methods that rely on text classification for intent recognition, the task-oriented dialogue method provided in this embodiment is more appropriate in its intent recognition approach. In this embodiment, by leveraging the text analysis and generation capabilities of a large model, the user's intent information is determined based on the user's dialogue information and user information. This overcomes the limitations of traditional text matching and text classification, thereby making the determined intent information closer to the user's true intent.

[0032] 103. Based on user information and intent information, the server generates a planning scheme through a large model, executes the execution steps in the planning scheme, and obtains the scheme execution result.

[0033] Specifically, after determining the user's intent information, the server can input the user information and intent information into another large model (hereinafter referred to as the large model agent). In this embodiment, the large model agent in step 103 differs somewhat from the large model in step 102. The large model in step 102 can be used to provide everyday conversations, while the large model agent in step 103 is not intended for everyday conversations. In this embodiment, the large model agent is mainly used in a targeted knowledge domain, given a complete question, to produce a solution through self-thinking and scheduling tools. The given question is the intent information determined in step 102.

[0034] In this embodiment, the large model agent can generate a planning scheme based on user information and intent information using a pre-defined prompt template. The planning scheme can be extracted from an existing scheme, modified from an existing scheme, or directly generated based on user information and intent information. After generating the planning scheme, the large model agent can plan and execute each execution step according to the execution steps in the scheme to obtain the execution result.

[0035] It should be noted that, compared to existing methods that use knowledge graphs for solution matching, the task-oriented dialogue method provided in this embodiment is more flexible in generating planning solutions. In this embodiment, by utilizing the text analysis and text generation capabilities of a large model agent, planning solutions can be generated. This not only allows for the recall of existing planning solutions but also the modification of existing plans, or the direct generation of previously unseen planning solutions based on intent information. This results in a more accurate match between the generated planning solutions and the user's intent information.

[0036] 104. The server sends the execution result of the solution to the client.

[0037] 105. The client receives and displays the execution results of the solution.

[0038] Specifically, after generating the execution result, the server can send the result to the client. Upon receiving the result, the client can display it. For example, the client can display the result through various pages such as instant messaging pages or search results pages. In this embodiment, when the server outputs the execution result through the large model agent, it can also output other information, such as user dialogue information, user intent information, and the large model agent's thought process information. In this scenario, the server can send only a portion of the information output by the large model agent to the client for display; for example, the server can send only the execution result to the client and record other information in a log for future analysis and processing.

[0039] In one embodiment, prior to step 104, the above-described task-oriented dialogue method may further include the following steps (not shown in Figure 1):

[0040] 106. Based on the intent information, the server uses a large model to determine whether the execution result of the solution matches the intent information.

[0041] 107. When the execution result of the plan does not match the intention information, the server re-executes the execution steps in the planning plan to obtain the corrected execution result.

[0042] Specifically, referring to the flowchart in Figure 2, after receiving the user's input dialogue information (including the user's input question), the server can extract the user information from the global memory. Then, based on the dialogue information and the user information, it performs intent clarification to obtain intent information, which represents the user's request. During intent clarification, the server can ask follow-up questions to the user during the request confirmation phase. If the user rejects the question, the server directly outputs the answer. If the user's intent is clear, after confirming the user's request, the server performs global planning based on the intent information to obtain a planning scheme. This planning scheme can be extracted from the planning memory, modified from the extracted scheme, or generated directly. Further, after generating the planning scheme, the server can call a tool to execute the scheme, obtaining the execution result, i.e., the answer to the question. Then, the server judges and corrects the execution result (this judgment and correction process can be performed once or multiple times), i.e., steps 106 and 107, to obtain the corrected execution result. Finally, the corrected execution result (i.e., the final answer) is output to the client for display.

[0043] In this embodiment, steps 106 and 107 can be implemented using an agent constructed from a large model. Specifically, steps 106 and 107 can be implemented using two separate large model agents, or using a single large model agent. During implementation, a preset prompt template can be input into the large model agent to improve the accuracy of judging and correcting the execution results.

[0044] The above embodiments can effectively improve the accuracy of the solution execution results by judging and correcting them, avoid deviations in the solution execution results, and thus improve the solution's effectiveness.

[0045] In one embodiment, the method for determining intent information based on dialogue information and user information in step 102 can be implemented in the following way:

[0046] 201. Based on dialogue information and user information, the server generates response information and response identifiers for the response information through a large model.

[0047] 202. When the response identifier is a problem location identifier, the server determines the intent information based on the dialogue information.

[0048] Specifically, referring to the flowchart in Figure 3, when performing intent recognition, the large model can generate response information and response identifiers based on the user's input dialogue information and the recalled user information. In this embodiment, the large model for intent recognition needs to be pre-trained on a dataset from a predetermined knowledge domain so that the large model can complete daily dialogues and output response information from the predetermined knowledge domain, as well as response identifiers for the output response information. The response identifier can be used to indicate the type of response information.

[0049] In this embodiment, the response flag can be designed according to the actual functional needs of the large model. For example, the response flag can be set as three types of flags: question location flag, question follow-up flag, and question rejection flag. These three flags can be used to indicate three situations: the current dialogue information can be converted into explicit intent information, the current dialogue information cannot be converted into explicit intent information, and the current dialogue information does not meet the requirements of the predetermined knowledge domain and cannot be answered. In this embodiment, when the response flag output by the large model is the question location flag, it indicates that the current dialogue information can be converted into explicit intent information. At this time, the user's intent information can be determined based on the user's dialogue information.

[0050] For example, as shown in the flowchart of Figure 8, when the large model determines that the intent of the user's input dialogue information is unclear, it can output the question follow-up indicator "Reply" and the corresponding response information; when the large model determines that the intent of the user's input dialogue information is clear, it can output the question location indicator "Clarified" and the corresponding response information, and determine the user's intent information as "XXXXX solution" based on the user's dialogue information and context information, thereby clarifying the user's intent.

[0051] The above embodiments output response information and response identifiers of the response information through a large model, and judge the response identifiers. This enables the large language model to quickly determine whether the intent information in the dialogue information can be identified through the current dialogue content, thereby improving the efficiency of user intent recognition.

[0052] In one embodiment, the method for determining intent information based on dialogue information in step 202 can be implemented as follows: when the response identifier is a problem location identifier, the server converts the dialogue information into problem information in a preset domain based on the context information in the user information, and uses the problem information in the preset domain as intent information.

[0053] Specifically, when the response identifier is a question location identifier, the server can convert the dialogue information into question information within a preset domain based on the context information in the user information, and then use this preset domain question information as intent information. In this embodiment, converting dialogue information into preset domain question information can be achieved through prompts set in the large model. For example, the large model can convert the user-input dialogue information "What are some recent positive reviews?" into the merchant service knowledge domain-specific question information "What are some recently added positive reviews?", and then use this question information as intent information. In this embodiment, the converted question information is more standardized than the user-input question, and therefore, it is easier to find an accurate solution in the corresponding knowledge domain.

[0054] In one embodiment, the method for determining intent information based on dialogue information in step 202 can also be implemented in the following way: when the response identifier is a problem location identifier, the server determines the problem type of the dialogue information, converts the dialogue information into problem information in a preset domain, extracts context information from user information and dialogue information, and then obtains intent information based on the problem type of the dialogue information, the problem information in the preset domain, and the context information.

[0055] Specifically, when the response identifier is a question location identifier, the server can determine the user's intent information through various information. In this embodiment, the server can first determine the question type of the dialogue information, then convert the dialogue information into question information of a preset domain, and then extract context information from the user information and dialogue information. Finally, based on the question type of the dialogue information, the question information of the preset domain, and the context information, the intent information is obtained. The question type can be set according to actual functional needs, for example, it can be set as a knowledge query question and an operation question, or as a simple question type and a complex question type, etc., etc., which is not specifically limited in this embodiment. The conversion of question information and the extraction of context information can be determined based on the training results of a large model, and there are no fixed setting rules.

[0056] In this embodiment, intent information can be generated with the help of prompts set in the large model. For example, the prompts in the large model can include the following:

[0057] When Action == Clarified, four items [Type, Problem, Info, Answer] must be output. Note that all four items must be output.

[0058] Type: Problem type, must be within [Consult, Operate];

[0059] Problem: Generate a question posed from the merchant's perspective, ensuring no information is missing;

[0060] Info: Summarize key contextual information, such as date, order details, numbers, positive reviews, etc.

[0061] Answer: Output the answer to the question;

[0062] In this embodiment, by incorporating an intent information generation step into the prompts of the large model, the generated intent information can be made clearer and more explicit. Based on the aforementioned prompts, the large model can output the following content in response to the user's input dialogue information "What positive reviews have I received recently?":

[0063] Question: What positive reviews have I received recently?

[0064] Thought: Clearly define the merchant's question; answers can be searched for, but cannot be directly answered.

[0065] Action: Clarified;

[0066] Response: Thank you so much for your patience. I need a little time to process this, it will be done soon.

[0067] nType: Consult\nProblem: What are the new positive reviews?\nInfo: Time: Recent;\nAnswer: Your recent positive reviews include: 1. Good service attitude, 2. Good food taste;

[0068] The above embodiments, by setting the question type in the dialogue information within the intent information, facilitate the large model agent in identifying the difficulty of generating the planning scheme or the type of tools required for the planning scheme. This allows the large model agent to match a more accurate planning scheme and obtain a more accurate execution result. Furthermore, by setting preset domain question information and contextual information in the intent information, the intent information can be made more complete and closer to the user's true intent. This improves the recall efficiency and accuracy of existing solutions, thereby enhancing the effectiveness of solution execution.

[0069] In one embodiment, the above-described task-oriented dialogue method may further include the following steps:

[0070] 203. When the response identifier is a follow-up question identifier and / or a rejection question identifier, the server generates context information based on the dialogue information and the response information, and adds the context information to the user information.

[0071] Specifically, referring to the flowchart in Figure 3, when the response identifier output by the large model is a follow-up question identifier, it indicates that the current dialogue information cannot be converted into explicit intent information, and / or, when the response identifier output by the large model is a question rejection identifier, it indicates that the current dialogue information does not meet the requirements of the predetermined knowledge domain and therefore cannot be answered. In this scenario, the dialogue information input by the user and the response information output by the large model can be added to the user information as contextual information so that it can be used as reference information for response information in the next round of dialogue. At this time, it is not necessary to determine the intent information based on the user's dialogue information.

[0072] In this embodiment, when the large model is pre-trained in a specific knowledge domain for daily conversations, the prompt set in the large model constrains it to avoid responding to questions from other knowledge domains, thus preventing the large model from outputting inaccurate information. Based on this, when a user inputs a question from another knowledge domain, the large model, pre-trained in the pre-defined knowledge domain, will output a question rejection flag and corresponding response information. Conversely, when the user's input question makes it impossible for the large model to determine the user's intent, the large model can output information such as casual conversation, counter-questions, follow-up questions, greetings, or introductions, in which case it will output a follow-up question flag.

[0073] For example, when a large model is pre-trained using a dataset from the merchant service knowledge domain, and the large model's prompt is set to only answer questions from the merchant service knowledge domain, and the large model is instructed to output response information and response identifiers, the user and the large model can engage in the following dialogue:

[0074] Question: Hello;

[0075] Action: Reply;

[0076] Response: I am a merchant assistant, here to help you solve problems encountered during store operation. How can I help you?

[0077] Question: Can you help me make a travel plan?

[0078] Action: Reject;

[0079] Response: I'm sorry, as a merchant assistant, I mainly provide help related to merchant services. For travel plans, I suggest you search for travel guides on XX, where you will find many good suggestions.

[0080] In one embodiment, the above-described task-oriented dialogue method may further include the following steps:

[0081] 204. The server sends the response information to the client.

[0082] 205. The client receives and displays the response information.

[0083] Specifically, referring to the flowchart in Figure 3, after the server obtains the dialogue information and user information, it can respond to the user's dialogue information through a large model. Furthermore, based on different response results, the large model will also output a response identifier for the response information. In this embodiment, regardless of which pre-set identifier the response identifier is, the large model can output the response information corresponding to the dialogue information and send the response information to the client for display. In this embodiment, referring to the scenario diagram in Figure 8, the client can only display the response information and not the response identifier.

[0084] In this embodiment, when the response identifier is a follow-up question identifier and / or a rejection question identifier, the large model can directly output the response information to improve the response speed of the dialogue information; when the response identifier is a problem location identifier, the large model can first output a response message prompting the user to wait, such as "Thank you very much for your patience, I need a little time to process, it will be done soon", and then generate and execute the planning scheme during the user's waiting time, and finally send the execution result of the planning scheme to the client for display.

[0085] In one embodiment, the user information obtained in step 102 may include at least one of the following: user profile information, function information of preset tools, dialogue constraint information, user history chat information, user history operation information, user history state information, and context information.

[0086] Specifically, user information plays a crucial role throughout the dialogue process between the user and the large model, preventing the large model from forgetting historical information and serving as important reference information for dialogue. In this embodiment, user information can be categorized into three types: long-memory information, short-memory information, and contextual information. Long-memory information includes important reference information that needs to be stored long-term, such as user profile information, function information of basic tools, and dialogue constraint information. Short-memory information includes important reference information that needs to be updated periodically, such as user chat history, user operation history, and user status history. Contextual information includes important reference information that needs to be updated in real-time, such as user feedback context information and system feedback context information.

[0087] In this embodiment, user profile information can be determined based on the user's identity. For example, when the user is a consumer, the user profile information can be a profile of the user's consumption behavior; when the user is a merchant, the user profile information can be a profile of the merchant, a profile of the store, etc. By adding user profile information to the user information, it is convenient to conduct dialogues according to the user's personality during the conversation. The function information of the preset tools refers to the definition of various basic functions and parameter definitions needed during the dialogue, such as the function definition information of the order query function and the evaluation query function. The dialogue constraint information refers to the restrictive conditions that limit the output of certain information by the large model. The dialogue constraint information can help the large model avoid answering questions that are irrelevant to the preset knowledge domain during the dialogue, and can also prevent the large model from having illusions and outputting inaccurate information. The user's historical chat information refers to the chat history information between the user and the large model. This chat history information can be the chat history information of the previous hour, the previous day, or the previous N days, etc., and this embodiment does not make specific limitations. The user's historical operation information can be the operation information related to the preset knowledge domain initiated by the user, such as the user complaint operation, the user refund operation, etc. The user's historical status information can be the user's status flags, such as the evaluation status, the order status, etc. Contextual information mainly includes recent chat history between the user and the large model, as well as intermediate state information such as planning schemes, execution steps, and execution results.

[0088] In this embodiment, incorporating multiple types of information into the user information increases its length. Therefore, this embodiment can store all user information at once. Then, when inputting the user information into the large model, a set of configuration rules is set to filter or extract the user information in a structured manner. The filtered or extracted information is then input into the large model, enabling it to provide more accurate and user-relevant responses based on the input user information during the generation of response information, planning schemes, and execution of planning schemes.

[0089] In one embodiment, the above-described task-oriented dialogue method may further include the following steps.

[0090] 301. The server retrieves historical chat log information according to the preset long memory generation cycle.

[0091] 302. Based on historical chat log information, the server generates user behavior summary information in at least one dimension through a large model, and obtains user profile information based on the user behavior summary information.

[0092] Specifically, the server can generate user behavior summary information in at least one dimension based on user information and historical chat history using a large model, according to a preset long memory generation cycle. This summary is then used to generate user profile information. In this embodiment, the length of the long memory generation cycle can be set according to actual conditions, and different cycles can be set for different users. For example, users who log in frequently can be updated weekly, users who log in occasionally can be updated monthly, and users who log in infrequently can be updated annually, and so on.

[0093] Furthermore, the type and content of user profiles will vary depending on the scenario and target audience of the task-based dialogue. For example, in the domain of merchant service knowledge, user profile information can be merchant profiles and store profiles. A merchant profile can include the merchant's name, personality traits exhibited during the chat, and frequently asked questions, etc.; a store profile can include basic information about the store across multiple dimensions, such as store name, store description, store status, store location, store brand, main products, target audience, popular products, popular reviews, and so on. In this embodiment, user profile information needs to consider multiple dimensions of feature information, thus resulting in a longer word count. Therefore, when generating user profile information, a large model can be used to summarize and simplify the user profile information to ensure that the user profile information retains multi-dimensional and concise user feature information.

[0094] In one embodiment, the above-described task-oriented dialogue method may further include the following steps:

[0095] 303. The server obtains historical chat log information and / or historical operation details according to the preset short memory generation cycle.

[0096] 304. The server performs structured information extraction on historical chat logs and / or historical operation details to obtain at least one of the following: user historical chat information, user historical operation information, and user historical status information.

[0097] Specifically, the server can extract structured information from historical chat logs and / or historical operation details according to a preset short memory generation cycle to obtain at least one of the following: user historical chat information, user historical operation information, and user historical status information. In this embodiment, the duration of the short memory generation cycle can be set according to actual conditions; for example, it can be set to update in units of one day or one hour. In this embodiment, short memory information such as user historical chat information, user historical operation information, and user historical status information can be stored in a detailed information format and summarized according to certain specifications during the information extraction process to obtain complete and concise user information. For example, the extraction and summarization format of the extracted user historical chat information, user historical operation information, and user historical status information can refer to the following:

[0098] User historical operation information:

[0099] 2023-11-25 12:00:03, appeal submitted; 2023-11-26 14:32:03, refund approved;

[0100] User historical status information:

[0101] 2023-11-25 12:00:03, Order xxx successful; 2023-11-26 14:32:03, Order xxx positive review; 2023-11-26 18:12:03, New task accepted;

[0102] User chat history: 2023-11-25 12:00:03, {"user": "ddd", "assistant": "dddd"}; 2023-11-25 14:32:03, {"user": "ddd", "assistant": "dddd"}; 2023-11-26 18:12:03, {"user": "ddd", "assistant": "dddd"};

[0103] In one embodiment, the method for generating a planning scheme based on user information and intent information in step 103 can be implemented through the following steps:

[0104] 401. The server performs text matching and / or similarity matching between the intent information and the names of the basic plans in the planning scheme library to obtain a pre-selected planning scheme that matches the intent information.

[0105] 402. The server determines the planning scheme based on user information, intent information, and pre-selected planning schemes through a large model.

[0106] Specifically, referring to the flowchart in Figure 5, when generating a planning scheme, the server can extract existing basic schemes from the planning memory based on the intent information, and then generate a planning scheme corresponding to the intent information based on the extracted basic schemes. In this embodiment, the server can recall basic schemes matching the intent information from the planning scheme library based on text matching or similarity matching. The process of matching the intent information with the scheme names of the basic schemes can be achieved using a pre-trained natural language processing model. For example, a natural language processing model can be used to score the similarity between the intent information and the scheme names of the basic schemes, and then a certain number of basic schemes can be recalled as pre-selected planning schemes based on the similarity score results. Furthermore, after selecting pre-selected planning schemes that match the intent information, a large model can be used as a supervisor to reconfirm the matching between the pre-selected planning schemes and the intent information, to determine whether to select a particular pre-selected planning scheme as the final planning scheme, or to adaptively modify a pre-selected planning scheme to obtain the final planning scheme, or to regenerate a new planning scheme, etc., thereby obtaining the final planning scheme.

[0107] The above embodiments, by performing text matching and / or similarity matching between intent information and the scheme name of the basic scheme, obtain a pre-selected planning scheme that matches the intent information. This can improve the recall speed of the pre-selected planning scheme and reduce server overhead. Furthermore, by utilizing a large model to determine the final planning scheme based on the pre-selected planning schemes, the generation speed of the planning scheme and the accuracy of the scheme content can be improved.

[0108] In one embodiment, step 402 can be implemented in at least one of the following ways:

[0109] 4021. The server uses a large model to select the target planning scheme that matches the intent information from the pre-selected planning schemes, based on user information and intent information.

[0110] 4022. The server selects the target planning scheme that matches the intent information from the pre-selected planning schemes through the large model, and modifies at least one execution step in the target planning scheme based on user information and intent information to obtain the planning scheme.

[0111] Specifically, referring to the flowchart in Figure 5, after selecting a pre-selected planning scheme, the server can reconfirm the scheme to determine whether to choose it as the final plan or to adaptively modify it to obtain the final plan. This process can be implemented using a correction task prompt set in the large model agent, as shown below:

[0112] Action definition:

[0113] Usable: One of the plans given for reference is available;

[0114] Modify: The given reference plan needs to be modified before it can be used;

[0115] Output format:

[0116] Thought: Please read and understand the Problem, Goal, Tools, and Milestone sections of each plan in the reference plan clearly and accurately. Consider in depth whether this reference plan meets your needs, especially when the problem closely matches yours; in such cases, the reference plan is likely what you need. If the reference plan meets your needs, then Action = Usable, and the Final Answer will output the reference plan number. If the reference plan does not meet your needs, you need to generate a new plan based on your original problem and the reference plan. This new plan must have reasonable justifications and solutions, clearly explaining why such modifications are necessary.

[0117] Action: You can only choose one action [Usable, Modify];

[0118] Final Answer:

[0119] If you select Action=Usable, the output will be the number of the reference plan you selected, such as 1, 2, 3;

[0120] If you selected Action=Modify, then you need to output your new plan, including Step / Goal / Tools / Milestone / Risk;

[0121] In one embodiment, the above-described task-oriented dialogue method may further include the following steps:

[0122] 403. If the server fails to select a target planning scheme that matches the intent information from the pre-selected planning schemes, it executes steps 404 and 405 according to the preset number of loops.

[0123] 404. The server performs text matching and / or similarity matching between the intent information and the names of the basic schemes in the planning scheme library to obtain at least one pre-selected planning scheme that matches the intent information, wherein the intent information is the original intent information and / or the modified intent information.

[0124] 405. The server determines the planning scheme based on user information, intent information, and pre-selected planning schemes through a large model.

[0125] Specifically, referring to the flowchart in Figure 5, when the server fails to select a target planning scheme that matches the intent information from the pre-selected planning schemes, it can again perform text matching and / or similarity matching between the intent information and the scheme names of the basic schemes in the planning scheme library to recall a larger number of pre-selected planning schemes, and determine the final planning scheme based on the re-recalled pre-selected planning schemes. In this embodiment, during the multiple recall process, the number of pre-selected planning schemes recalled can be increased, or the description of the intent information can be changed to recall a wider range of pre-selected planning schemes, thereby facilitating the selection of a scheme that better matches the intent information as the final planning scheme. In addition, for time considerations, the number of loop executions can be limited, for example, the number of loop executions can be set to 2 or 3 times, etc., to avoid the user waiting time being too long. It is understood that the specific implementation methods of steps 404 and 405 can refer to the relevant descriptions in steps 401 and 402, and will not be elaborated further here.

[0126] The above embodiments can improve the recall range of pre-selected planning schemes by changing the description of intent information and / or expanding the recall of basic schemes when no target planning scheme is selected in the large model. This can effectively improve the matching accuracy of planning schemes and the effect of scheme execution.

[0127] In one embodiment, the method for generating a planning scheme based on user information and intent information in step 103 can also be implemented through the following steps:

[0128] 406. When the preset scheme generation conditions are met, the server generates a planning scheme based on user information, intent information and preset prompt template through a large model. The scheme generation conditions include not obtaining a pre-selected planning scheme that matches the intent information, and / or the pre-selected planning scheme does not contain a target planning scheme that matches the intent information.

[0129] Specifically, when the server fails to obtain a pre-selected planning scheme that matches the intent information, and / or when the pre-selected planning schemes do not include a target planning scheme that matches the intent information, it can stop the process of recalling pre-selected planning schemes and selecting a target planning scheme. Instead, it can directly enter the planning scheme generation process. Then, through a large model agent, a planning scheme is generated based on user information and intent information. In this way, even when the number of basic schemes in the planning scheme library is insufficient, new planning schemes can be generated based on the user's actual needs. This overcomes the limitations of existing basic schemes, resulting in a planning scheme that better matches the intent information and obtains accurate scheme execution results.

[0130] In one embodiment, the method for obtaining the execution result of the planning scheme by executing the execution steps in step 103 can be implemented through the following steps:

[0131] 501. The server executes each execution step in the order of execution to obtain the execution result of the solution. For any execution step, the execution method includes: determining the target tool corresponding to the execution step through the large model, calling the target tool to execute the execution step, and obtaining the execution result of the execution step.

[0132] Specifically, referring to the flowchart in Figure 2, after generating the planning scheme, the server can execute each execution step sequentially according to the execution order to obtain the scheme execution result. In this embodiment, the target tool corresponding to each execution step can be one or more, or the target tool can be omitted, and the text analysis and text generation capabilities of the large model agent can be directly utilized. This embodiment obtains the execution result of the execution step by calling the target tool to execute the execution step, which can make the execution result of the execution step more stable and less prone to deviation from the scheme execution result.

[0133] In one embodiment, the method of calling the target tool to perform the execution step in step 501 to obtain the execution result of the execution step can be implemented through the following steps:

[0134] 5011. Based on the content of the execution steps, the server searches for the target tool in the preset tool library, calls the target tool, obtains the call result of the target tool, and obtains the execution result of the execution steps based on the call result of the target tool.

[0135] 5012. When the server does not find the target tool, it calls the search tool to search for the retrieval results corresponding to the execution steps in the preset knowledge base, and obtains the execution results of the execution steps based on the retrieval results.

[0136] Specifically, the server can, according to the logic in the prompt set in the large model agent, select the required tool as the target tool during the execution step based on the function information of the pre-set tools, and then obtain the execution result of the execution step through the call result of the target tool. Alternatively, if the target tool is not found, a search tool can be invoked to search for the corresponding retrieval results in the pre-set knowledge base, and the execution result of the execution step can be obtained based on the retrieval results. For example, the prompt set in the large model agent can refer to the following:

[0137] Objective: The task is to answer the merchant's questions by using the tool or to continue reasoning about the merchant's questions to complete the task given by the merchant. If no tool is available, the knowledge_retrieval_tool must be used to search for the answer.

[0138] The following API tools can be used: {tool_descs};

[0139] Output format:

[0140] Question: Merchant input question;

[0141] Thought: Begin or continue thinking about what to do, which tools to use, how to answer, and how to politely decline to answer. If no tools are available, do not answer the question directly; you need to search for the answer.

[0142] Action: Use the tool, which must be within [{tool_names}];

[0143] Action Input: Input tool parameters to view tool description;

[0144] Observation: Tool output;

[0145] (This Thought / Action / Action Input / Observation can be repeated zero or more times)

[0146] Thought: Continue thinking based on the above steps;

[0147] Final Answer: The final answer to the original input question. Note that the perspective of the answer should be directed at the merchant; do not include any personal thoughts or opinions.

[0148] Merchant information in the dialogue:

[0149] {memory};

[0150] In this embodiment, by restricting the large model agent to searching for target tools only in a preset tool library during the execution steps, and by calling a search tool to search for retrieval results corresponding to the execution steps in a preset knowledge base when no target tool is found, the execution results of the execution steps can be limited to the preset knowledge domain, thereby greatly avoiding the problem of large models generating illusions.

[0151] In one embodiment, the execution step may include a task description and / or a tool description. Based on this, the method in step 501 of invoking the target tool to execute the execution step and obtain the execution result can also be implemented through the following steps:

[0152] 5013. Based on the tool description in the execution step, the server determines the target tool corresponding to the execution step, and calls the target tool to execute the execution step according to the parameter information recorded in the task description, thereby obtaining the execution result of the execution step.

[0153] Specifically, when the server invokes a target tool, it can do so based on the task description and / or tool description in the execution steps, thereby improving the selection efficiency and execution accuracy of the target tool. In this embodiment, step 5013 can be applied to scenarios where a basic plan matching the intent information is selected from the planning scheme library, or to scenarios where a new planning scheme is generated based on a pre-selected planning scheme. For example, assuming the planning scheme includes the following two execution steps: "1. Compare the merchant's data for the last 7 days with industry data" and "2. Summarize the data", then based on the above two execution steps, the following task description and tool description can be obtained:

[0154] Node 1:

[0155] Task Description: Retrieve the merchant's data for the past 7 days and the industry's data for the past 7 days, based on the merchant ID;

[0156] Tool description: search_QA (specific information parameters are stored in another database);

[0157] Node 2:

[0158] Task Description: Summarize based on the extracted user and industry data;

[0159] Tool description: None;

[0160] In the example above, by obtaining the task descriptions and tool descriptions for each execution step, it can be determined that the target tool to be called in the first execution step is `search_QA`, and the tool parameters of the target tool are the merchant ID, the time period of the last 7 days, and the data types of the merchant data and industry data. Then, the actions in the task description can be executed based on the target tool to obtain the step execution result. Furthermore, in the second execution step, the large model agent can use the data obtained in the first step, along with its own text analysis and text generation capabilities, to obtain the solution execution result, namely, the comparative analysis results between the merchant's data and industry data from the last 7 days.

[0161] In one embodiment, the intent information may include the question type, question information, and context information. Based on this, the method in step 501 of invoking the target tool to execute the execution steps and obtain the execution result of the execution steps can also be implemented through the following steps:

[0162] 5014. Based on the problem type and / or problem information in the intent information, the server determines the target tool corresponding to the execution step, determines the parameter information of the target tool according to the problem information and / or context information, and calls the target tool to execute the execution step through the parameter information of the target tool, thereby obtaining the execution result of the execution step.

[0163] Specifically, when invoking a target tool, the server can use various information such as the question type, question information, and context information from the intent information to improve the selection efficiency and execution accuracy of the target tool. In this embodiment, step 5014 is applicable to scenarios where the intent information is relatively clear and can be directly executed to obtain the solution execution result, or scenarios where the planning solution is relatively simple. For example, suppose the intent information determined by the large model based on dialogue information and user information is as follows:

[0164] Type: Consult; Problem: What are the new positive reviews?; Info: Time: Recent;

[0165] Based on the aforementioned intent information, the large model agent can determine that the question type is a knowledge query type, the target tool is an evaluation query tool, and the parameter information of the target tool is a merchant ID and a time period of the most recent N days. Using this information, the evaluation query tool is invoked, and the execution result is as follows:

[0166] Answer: Your recent positive reviews include: 1. Good service attitude, 2. Good food taste;

[0167] In one embodiment, the method of calling the target tool to perform the execution step in step 501 to obtain the execution result of the execution step can also be implemented through the following steps:

[0168] 5015. The server determines the target tool corresponding to the execution step, calls the target tool, executes the execution step at least once, and obtains at least one call result. In this case, calling the target tool once can obtain one or more call results.

[0169] 5016. The server summarizes the results of at least one call to obtain the execution results of the execution steps.

[0170] Specifically, during the execution of the steps, the server can invoke the target tool at least once to obtain one or more invocation results. Then, the large model can summarize based on at least one invocation result to obtain a streamlined execution result. In this embodiment, this process can be implemented through a prompt set in the large model agent. For example, the relevant content set in the prompt is as follows:

[0171] Output format:

[0172] Question: Merchant input question;

[0173] Thought: Begin or continue thinking about what to do, which tools to use, how to answer, and how to politely decline to answer. If no tools are available, do not answer the question directly; you need to search for the answer.

[0174] Action: Use the tool, which must be within [{tool_names}];

[0175] Action Input: Input tool parameters to view tool description;

[0176] Observation: Tool output;

[0177] (This Thought / Action / Action Input / Observation can be repeated zero or more times)

[0178] Thought: Continue thinking based on the above steps;

[0179] Final Answer: The final answer to the original input question. Note that the perspective of the answer should be directed at the merchant; do not include any personal thoughts or opinions.

[0180] Notice:

[0181] The search yields multiple answers; first, the question and search results are matched (question), and then the results (answer) are summarized.

[0182] Make assumptions are prohibited; important information from tool results is prohibited; and improvisation is prohibited when no tool results are available.

[0183] After using the tools multiple times, summarize the answer in the final form, with a word count not exceeding 300 words;

[0184] In one embodiment, the method for the client to display the execution result of the scheme in step 105 can be implemented in the following way:

[0185] 601. The client displays the solution execution result through a preset component and displays at least one solution evaluation control, wherein the solution evaluation control is used to mark the solution execution result as solved or unsolved in response to the trigger operation.

[0186] Specifically, referring to the scenario diagram in Figure 9, the client can display the solution execution result and at least one solution evaluation control through a composite preset component. In this embodiment, when the user clicks the solution evaluation control, the client can respond to the trigger operation by marking the solution execution result as resolved or unresolved. For example, when the user clicks "Unresolved" in the preset component, the solution execution result can be marked as unresolved; when the user clicks "Resolved" in the preset component, the solution execution result can be marked as resolved. It is understood that in other embodiments, the preset component and solution evaluation control can also be displayed in various forms, which will not be exemplified here.

[0187] The above embodiments display the execution results and evaluation controls of the scheme through preset components, which makes it convenient for users to evaluate the effectiveness of the scheme in a timely manner, thereby helping the system to generate more accurate planning schemes.

[0188] In one embodiment, the above-described task-oriented dialogue method may further include the following steps:

[0189] 701. The server obtains the historical chat history information of at least one user, and extracts the planning schemes corresponding to the execution results of the solutions marked as resolved in the historical chat history information to obtain at least one basic solution.

[0190] 702. The server stores the basic scheme in the planning scheme library with a preset structure, wherein the preset structure includes at least one of the following structures: chain structure, tree structure and network structure.

[0191] Specifically, referring to the flowcharts in Figures 4 and 5 and the scenario diagram in Figure 9, the server can extract the planning schemes corresponding to the execution results of solutions marked as resolved from the historical chat records of multiple users. After organizing these planning schemes according to a preset format, they are stored in a planning scheme library with a preset structure. This method improves the efficiency of generating basic solutions and reduces the workload of manually generating them.

[0192] In this embodiment, the server can store the extracted basic solutions according to at least one of the following structures: a chain structure (as shown in Figure 6(A)), a tree structure (as shown in Figure 6(B)), and a network structure (as shown in Figure 6(C)). Furthermore, for each basic solution, it can be stored in the manner shown in Figure 7. That is, when storing the planning scheme, the task description and tool description corresponding to each execution step in the planning scheme can be stored simultaneously, thereby improving the efficiency of tool selection and step execution when the basic solution is recalled. For example, the basic solution can be stored in the following format:

[0193] Task Description:

[0194] When users inquire about abnormal changes in order sales behavior, we can search for data, behavioral data, and user data to address the issue.

[0195] Further analysis:

[0196] Basic solution:

[0197] 1. Compare the merchant's data for the past 7 days with industry data; 2. Summarize the data; 3. Review the merchant's recent operations and activities; 4. Summarize the behavior; 5. Review the merchant's order reviews for the past 7 days; 6. Create an overall review; 7. Summarize xxxx; Node 1: Task Description: Retrieve the merchant's data for the past 7 days and the industry's data for the past 7 days based on the merchant ID; Tool Description: search_QA (specific information parameters are stored in another database);

[0198] Node 2:

[0199] Task Description: Summarize based on the extracted user and industry data;

[0200] Tool description: None;

[0201] The above embodiments, by accumulating and organizing the planning schemes corresponding to the execution results of solutions marked as resolved in historical chat records, can improve the speed of basic solution collection, thereby helping to improve the recall speed and accuracy of planning schemes. Furthermore, by utilizing various structures such as chain structures, tree structures, and network structures to store basic solutions, the space occupied by basic solution storage can be reduced, and the efficiency of basic solution retrieval can be improved.

[0202] In one embodiment, the user initiating the conversation is a merchant, and the target tool includes at least one of the following: searching a knowledge base, querying orders, querying data, analyzing data, image parsing, creating a store, canceling an order, and sending instant messages to the user.

[0203] Specifically, the task-oriented dialogue method proposed in any of the above embodiments can be applied to the knowledge domain of merchant services. In this scenario, the large model, acting as a merchant service assistant, can help merchants complete various simple or complex tasks such as knowledge search, order query, data query, data analysis, image parsing, store creation, order cancellation, and sending instant messages to users. This can improve the efficiency of solving merchant problems, reduce the workload of human customer service, and enhance the interactive experience for merchants. By limiting the application scenario of this case to the merchant assistant scenario, the above embodiments can improve the efficiency of solving merchant problems in this scenario.

[0204] 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, stored data, displayed data, etc.) involved in this invention are all information and data authorized by the user or fully authorized by all parties. Furthermore, the labels corresponding to each step in the above embodiments are only for identification purposes and are not intended to limit the order of execution of the steps. The order of execution of the steps in each embodiment can be set according to the actual situation.

[0205] Furthermore, as a specific implementation of the method shown in Figures 1 to 9, this embodiment provides a task-oriented dialogue device, which includes a dialogue information receiving module and a solution result display module. The dialogue information receiving module is used to send the dialogue information in response to receiving dialogue information initiated by a user; the solution result display module is used to receive and display the solution execution result, which is obtained by executing the execution steps in the planning scheme. The planning scheme is generated by a large model based on the user's user information and the intent information corresponding to the dialogue information, and the intent information is determined by the large model based on the dialogue information and the user information.

[0206] It should be noted that other corresponding descriptions of the functional units involved in the task-oriented dialogue device provided in this embodiment can be found in Figures 1 to 9 and the corresponding descriptions in any of the above-mentioned task-oriented dialogue method embodiments, and will not be repeated here.

[0207] Furthermore, as a specific implementation of the method shown in Figures 1 to 9, this embodiment provides a task-oriented dialogue device, which includes: an intent information determination module, a solution result generation module, and a solution result sending module. The intent information determination module can be used to, in response to receiving dialogue information initiated by a user, obtain the user's user information, and determine the user's intent information based on the dialogue information and the user information using a large model. The solution result generation module can be used to, based on the user information and the intent information, generate a planning scheme using a large model, and execute the execution steps in the planning scheme to obtain the scheme execution result. The solution result sending module can be used to send the scheme execution result to the client, so that the client receives and displays the scheme execution result.

[0208] It should be noted that other corresponding descriptions of the functional units involved in the task-oriented dialogue device provided in this embodiment can be found in Figures 1 to 9 and the corresponding descriptions in any of the above-mentioned task-oriented dialogue method embodiments, and will not be repeated here.

[0209] Based on the methods shown in Figures 1 to 9, this embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the task-based dialogue method shown in Figures 1 to 9.

[0210] Based on this understanding, the technical solution of the present invention can be embodied in the form of a software product. The software product to be identified can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, or portable hard drive), and includes several instructions to cause a computer device (such as a personal computer, server, or network device) to execute the methods described in various implementation scenarios of the present invention.

[0211] Based on the method embodiments shown in Figures 1 to 9 above, in order to achieve the above objectives, as shown in Figure 10, this embodiment also provides a task-oriented dialogue computer device, which can be a personal computer, server, smartphone, tablet computer, smartwatch, or other network device, etc. The computer device includes a storage medium and a processor; the storage medium is used to store computer programs and operating systems; the processor is used to execute computer programs to implement the methods shown in Figures 1 to 9 above.

[0212] Optionally, the computer device may also include internal memory, a communication interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, a display screen, and input devices such as a keyboard. The communication interface may also include a USB interface, a card reader interface, etc. The network interface may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0213] Those skilled in the art will understand that the computer device structure for task-oriented dialogue provided in this embodiment does not constitute a limitation on the computer device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0214] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the aforementioned computer hardware and the software resources to be identified, supporting the operation of information processing programs and other software and / or programs to be identified. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software in the information processing computer device.

[0215] Through the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented using software plus necessary general-purpose hardware platforms, or it can be implemented in hardware. By applying the technical solution of the present invention, after receiving dialogue information initiated by a user, the user's user information is first obtained. Based on the dialogue information and user information, the user's intent information is determined through a large model. Then, based on the user's user information and intent information, a planning scheme corresponding to the intent information is generated through the large model. Finally, the execution steps in the planning scheme are executed to obtain the scheme execution result. Compared with the prior art, the above method can effectively improve the accuracy of intent recognition and scheme planning, reduce the workload of scheme enumeration, improve the matching degree between the scheme execution result and the user's intent, and meet the user's actual requirements.

[0216] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention. Those skilled in the art will understand that the modules in the apparatus of the embodiment can be distributed within the apparatus of the embodiment as described, or they can be located in one or more apparatuses different from this embodiment, with corresponding changes. The modules of the above-described embodiment can be combined into one module, or further divided into multiple sub-modules.

[0217] The serial numbers used above are for descriptive purposes only and do not represent the superiority or inferiority of the implementation scenarios. The above disclosures are merely a few specific implementation scenarios of the present invention; however, the present invention is not limited thereto, and any variations conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A task-oriented dialogue method, characterized in that, The method includes: Upon receiving a dialogue message initiated by the user, the client sends the dialogue message to the server. In response to receiving the dialogue information, the server obtains the user's user information and, based on the dialogue information and the user information, determines the user's intent information through a large model. Based on the user information and the intent information, the server generates a planning scheme through a large model, executes the execution steps in the planning scheme, and obtains the scheme execution result; The server sends the execution result of the solution to the client; The client receives and displays the execution result of the scheme.

2. The method according to claim 1, characterized in that, Based on the dialogue information and the user information, the server determines the user's intent information using a large model, including: Based on the dialogue information and the user information, the server generates response information and a response identifier for the response information using a large model. When the response identifier is a problem location identifier, the server determines the intent information based on the dialogue information.

3. The method according to claim 2, characterized in that, When the response identifier is a problem location identifier, the server determines the intent information based on the dialogue information, including: When the response identifier is the problem location identifier, the server, based on the context information in the user information, converts the dialogue information into problem information in a preset domain, and uses the problem information in the preset domain as the intent information; or When the response identifier is the problem location identifier, the server determines the problem type of the dialogue information, converts the dialogue information into problem information of a preset domain, extracts context information from the user information and the dialogue information, and obtains the intent information based on the problem type of the dialogue information, the problem information of the preset domain, and the context information.

4. The method according to claim 2 or 3, characterized in that, The method further includes: When the response identifier is a follow-up question identifier and / or a rejection question identifier, the server generates context information based on the dialogue information and the response information, and adds the context information to the user information. The server sends the response information to the client; The client receives and displays the response information.

5. The method according to claim 1, characterized in that, The user information includes user profile information; therefore, the method further includes: The server retrieves historical chat log information according to a preset long memory generation cycle; Based on the historical chat log information, the server generates user behavior summary information in at least one dimension through a large model, and obtains the user profile information based on the user behavior summary information.

6. The method according to claim 1 or 5, characterized in that, The user information further includes at least one of the following: user history chat information, user history operation information, and user history status information; therefore, the method further includes: The server acquires historical chat logs and / or historical operation details according to a preset short memory generation cycle. The server performs structured information extraction on the historical chat history information and / or historical operation details to obtain at least one of the user's historical chat information, the user's historical operation information, and the user's historical status information.

7. The method according to claim 1, characterized in that, Based on the user information and the intent information, the server generates a planning scheme through a large model, including: The server performs text matching and / or similarity matching between the intent information and the names of basic schemes in the planning scheme library to obtain a pre-selected planning scheme that matches the intent information. The server determines the planning scheme based on the user information, the intent information, and the pre-selected planning scheme using the large model.

8. The method according to claim 7, characterized in that, The server determines the planning scheme based on the user information, the intent information, and the pre-selected planning scheme using the large model, including: The server, using the large model, selects a target planning scheme that matches the intent information from the pre-selected planning schemes as the planning scheme based on the user information and the intent information; and / or The server uses the large model to select a target planning scheme that matches the intent information from the pre-selected planning schemes, and modifies at least one execution step in the target planning scheme based on the user information and the intent information to obtain the planning scheme.

9. The method according to claim 8, characterized in that, The method further includes: When the server fails to select a target planning scheme that matches the intent information from the pre-selected planning schemes, it executes the following steps according to a preset number of loops: The server performs text matching and / or similarity matching between the intent information and the names of the basic schemes in the planning scheme library to obtain at least one pre-selected planning scheme that matches the intent information, wherein the intent information is the original intent information and / or the corrected intent information. The server determines the planning scheme based on the user information, the intent information, and the pre-selected planning scheme using the large model.

10. The method according to claim 1 or 7, characterized in that, The server generates a planning scheme based on the user information and the intent information using a large model, and also includes: When the preset scheme generation conditions are met, the server generates the planning scheme through a large model based on the user information, the intent information, and the preset prompt template. The scheme generation conditions include not obtaining a pre-selected planning scheme that matches the intent information, and / or the pre-selected planning scheme does not include a target planning scheme that matches the intent information.

11. The method according to claim 1, characterized in that, The server executes the execution steps in the planning scheme to obtain the scheme execution result, including: The server executes each execution step sequentially according to the execution order to obtain the execution result of the scheme. For any given execution step, the execution method includes: The target tool corresponding to the execution step is determined by the large model, and the target tool is called to execute the execution step to obtain the execution result of the execution step.

12. The method according to claim 11, characterized in that, The server determines the target tool corresponding to the execution step, and calls the target tool to execute the execution step, obtaining the execution result of the execution step, including: Based on the content of the execution steps, the server searches for a target tool in a preset tool library, calls the target tool, obtains the call result of the target tool, and obtains the execution result of the execution steps based on the call result of the target tool. When the server does not find the target tool, it calls a search tool to search for the retrieval results corresponding to the execution step in a preset knowledge base, and obtains the execution result of the execution step based on the retrieval results.

13. The method according to claim 11, characterized in that, The execution steps include a task description and / or a tool description; the server determines the target tool corresponding to the execution step, and calls the target tool to execute the execution step to obtain the execution result of the execution step, and further includes: Based on the tool description in the execution step, the server determines the target tool corresponding to the execution step, and calls the target tool to execute the execution step according to the parameter information recorded in the task description, so as to obtain the execution result of the execution step.

14. The method according to claim 11, characterized in that, The intent information includes the question type, question information, and context information; the server determines the target tool corresponding to the execution step, and calls the target tool to execute the execution step to obtain the execution result of the execution step, and also includes: The server determines the target tool corresponding to the execution step based on the question type and / or the question information in the intent information, determines the parameter information of the target tool according to the question information and / or the context information, and calls the target tool to execute the execution step using the parameter information of the target tool, thereby obtaining the execution result of the execution step; and / or The server determines the target tool corresponding to the execution step, calls the target tool, executes the execution step at least once, and obtains at least one call result, wherein calling the target tool once can obtain one or more call results; based on at least one call result, the execution result of the execution step is obtained.

15. The method according to any one of claims 11 to 14, characterized in that, The user information includes at least one of the following: user profile information, function information of preset tools, dialogue constraint information, user chat history information, user operation history information, user state history information, and context information. and / or The user initiating the dialogue is a merchant, and the target tool includes at least one of the following: searching a knowledge base, querying orders, querying data, analyzing data, image parsing, creating a store, canceling an order, and sending instant messages to the user.

16. The method according to claim 1, characterized in that, The client displays the execution results of the scheme, including: The client displays the solution execution result through preset components and displays at least one solution evaluation control, wherein the solution evaluation control is used to mark the solution execution result as solved or unsolved in response to a trigger operation; and / or The method further includes: The server obtains historical chat logs of at least one user and extracts planning schemes corresponding to the execution results of solutions marked as resolved from the historical chat logs to obtain at least one basic solution; the basic solution is stored in a planning scheme library with a preset structure, the preset structure including at least one of chain structure, tree structure, and network structure; and / or The method further includes: Based on the intent information, the server uses a large model to determine whether the execution result of the scheme matches the intent information; if the execution result of the scheme does not match the intent information, the server re-executes the execution steps in the planning scheme to obtain a corrected execution result.

17. A task-oriented dialogue method, characterized in that, The method includes: In response to receiving a dialogue message initiated by the user, the dialogue message is sent. Receive and display the execution result of the scheme, wherein the execution result is obtained by executing the execution steps in the planning scheme, the planning scheme is generated by the large model based on the user's user information and the intent information corresponding to the dialogue information, and the intent information is determined by the large model based on the dialogue information and the user information.

18. A task-oriented dialogue method, characterized in that, The method includes: In response to receiving a dialogue message initiated by a user, the system obtains the user's user information and, based on the dialogue message and the user information, determines the user's intent information using a large model. Based on the user information and the intent information, a planning scheme is generated through a large model, and the execution steps in the planning scheme are executed to obtain the scheme execution result; The execution result of the solution is sent to the client so that the client can receive and display the execution result.

19. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 18.

20. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 18.

Images