Information processing method and device based on large model, electronic equipment and agent

Abstract

This disclosure provides a method, apparatus, electronic device, and intelligent agent for information processing based on a large model, relating to the fields of artificial intelligence technology, particularly large language models, computer vision, text processing, and AI intelligent assistants. The specific implementation of the information processing method is as follows: The large model is used to expand the information of the problem to be processed, generating target problem information; wherein the target problem information includes at least one related problem information, and the problem information to be processed and the related problem information have similar questioning intents; the large model is used to perform semantic analysis on the target problem information to generate a target processing strategy; wherein the target processing strategy includes the target logical relationship between various processing tasks used to answer the target problem information; and the large model is invoked to execute each processing task according to the target logical relationship, generating answer information for the target problem.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, particularly to the fields of large language models, computer vision, text processing, AI intelligent assistants, and human-computer interaction, specifically to information processing methods, devices, electronic devices, and intelligent agents based on large models. Background Technology

[0002] An AI agent is an intelligent entity capable of perceiving its environment, making decisions, and performing actions. Typically based on machine learning and artificial intelligence technologies, it uses a large model as its core brain, autonomously learning and improving in specific tasks or domains through processes similar to human thinking and execution, exhibiting autonomy and adaptability. Summary of the Invention

[0003] This disclosure provides an information processing method, apparatus, electronic device, and intelligent agent based on a large model.

[0004] According to one aspect of this disclosure, a large-scale model-based information processing method is provided, comprising: expanding the problem information to be processed using the large-scale model to generate target problem information; wherein the target problem information includes at least one related problem information, wherein the problem information to be processed and the related problem information have similar questioning intents; performing semantic analysis on the target problem information using the large-scale model to generate a target processing strategy; wherein the target processing strategy includes target logical relationships between various processing tasks for answering the target problem information; and generating response information for the target problem by calling the large-scale model to execute each processing task according to the target logical relationships.

[0005] According to another aspect of this disclosure, an information processing apparatus based on a large model is provided, comprising: a question expansion module, a strategy generation module, and a response generation module.

[0006] The problem expansion module is used to expand the problem information to be processed using the large model to generate target problem information. The target problem information includes at least one related problem information, wherein the problem information to be processed and the related problem information have similar questioning intentions.

[0007] The strategy generation module is used to perform semantic analysis on the target question information using a large model to generate target processing strategies; the target processing strategies include the target logical relationships between various processing tasks used to answer the target question information.

[0008] The response generation module is used to generate response information for the target question by calling the large model to execute various processing tasks according to the target logical relationship.

[0009] According to another aspect of this disclosure, an intelligent agent is provided, including: an input module, a processing module, and an output module.

[0010] The input module is used to receive information about problems to be processed.

[0011] The processing module is used to determine the target task based on the problem information to be processed received by the input module, determine the target large model based on the target task, and obtain the response information by calling the target large model to execute the above method.

[0012] The output module is used to output the response information obtained by the processing module.

[0013] According to another aspect of this disclosure, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method as described above.

[0014] According to another aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause the computer to perform the method described above.

[0015] According to another aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements the method described above.

[0016] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0017] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein:

[0018] Figure 1 This illustration schematically shows an exemplary system architecture for applying large-model-based information processing methods and apparatus according to embodiments of the present disclosure;

[0019] Figure 2 A flowchart illustrating an information processing method based on a large model according to an embodiment of the present disclosure is shown schematically.

[0020] Figure 3A The illustration schematically shows an extension of the problem to be processed using a large model according to an embodiment of the present disclosure;

[0021] Figure 3BThe illustration schematically shows an extension of the problem to be processed using a large model according to another embodiment of the present disclosure;

[0022] Figure 4 A schematic diagram illustrating an extension problem using a large model iterative loop according to an embodiment of the present disclosure is shown.

[0023] Figure 5 The illustration shows a schematic diagram of a semantic analysis and generation strategy for target problem information using a large model according to an embodiment of the present disclosure;

[0024] Figure 6A A schematic diagram illustrating a problem optimization according to an embodiment of the present disclosure is shown.

[0025] Figure 6B A schematic diagram illustrating a problem optimization according to another embodiment of the present disclosure is shown.

[0026] Figure 6C A schematic diagram illustrating a problem optimization according to yet another embodiment of the present disclosure is shown.

[0027] Figure 7 The illustration shows a schematic diagram of the execution of various processing tasks according to target logical relationships by invoking a large model according to an embodiment of the present disclosure;

[0028] Figure 8 This illustration schematically shows a diagram of optimizing the parameters of a large model based on response information feedback according to an embodiment of the present disclosure;

[0029] Figure 9 A block diagram of a large-model-based information processing apparatus according to embodiments of the present disclosure is schematically shown; and

[0030] Figure 10 A block diagram of an electronic device suitable for implementing a large-model-based information processing method according to an embodiment of the present disclosure is shown schematically. Detailed Implementation

[0031] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0032] AI agents can help users solve related problems in multiple fields through the thought chain model of large language models. With the widespread application of artificial intelligence technology in various fields, user needs are becoming increasingly diversified and complex. Therefore, there is an urgent need to improve the ability of large models to handle complex and diverse problems in order to meet user needs.

[0033] In the relevant examples, user questions are addressed primarily in the following ways:

[0034] 1. By summarizing and statistically analyzing a large number of question-and-answer documents, a database of question-and-answer documents is constructed. Then, based on rule matching between user questions and questions in the database, corresponding answers are generated.

[0035] While this rule-based matching method offers a fast response time, it lacks semantic analysis of the user's question and fails to understand the user's intent. When encountering questions with weak patterns, it can easily lead to irrelevant or off-topic answers.

[0036] 2. Train the NLP model using sample data, and then use the trained model to process user questions and generate corresponding answers.

[0037] While this approach takes into account the user's intent in asking the question, the generalization ability of the trained model is limited by the amount of sample data, resulting in poor model scalability. Furthermore, it can only output the final answer to the user's question, without demonstrating the model's thought process in addressing that question, leading to poor interpretability.

[0038] In view of this, embodiments of this disclosure provide an information processing method based on a large model. By leveraging the understanding and generation capabilities of the large model to extend the problem to be processed, the large model can combine related questions similar to the user's question intent to process the target problem, further improving the generalization of the large model. Simultaneously, semantic analysis is performed on the extended problem using the large model to generate logical relationships between various processing tasks for answering the question. This allows the large model to improve the accuracy of the answer information by fully understanding the question intent and combining logical reasoning capabilities, further enhancing the user experience.

[0039] Figure 1 The illustration schematically shows an exemplary system architecture for applying large-model-based information processing methods and apparatus according to embodiments of the present disclosure.

[0040] It is important to note that Figure 1The examples shown are merely examples of system architectures that can be applied to the embodiments of this disclosure, to help those skilled in the art understand the technical content of this disclosure, but do not mean that the embodiments of this disclosure cannot be used in other devices, systems, environments, or scenarios. For example, in another embodiment, an exemplary system architecture for applying the information processing method and apparatus based on large models may include a terminal device, but the terminal device can implement the information processing method and apparatus based on large models provided by the embodiments of this disclosure without interacting with the server.

[0041] like Figure 1 As shown, the system architecture 100 according to this embodiment may include terminal devices 101, 102, and 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the terminal devices 101, 102, and 103 and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links, etc.

[0042] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, and 103, such as knowledge reading applications, web browser applications, search applications, instant messaging tools, email clients, and / or social platform software, etc. (for example only).

[0043] Terminal devices 101, 102, and 103 can be various electronic devices with displays and web browsing capabilities, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0044] Server 105 can be a server that provides various services, such as a backend management server that supports the content browsed by users using terminal devices 101, 102, and 103 (for example only). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0045] It should be noted that the information processing method based on a large model provided in this disclosure can generally be executed by terminal devices 101, 102, or 103. Correspondingly, the information processing device based on a large model provided in this disclosure can also be disposed in terminal devices 101, 102, or 103.

[0046] Alternatively, the large-model-based information processing method provided in this embodiment can generally be executed by server 105. Correspondingly, the large-model-based information processing apparatus provided in this embodiment can generally be located in server 105. The large-model-based information processing method provided in this embodiment can also be executed by a server or server cluster that is different from server 105 and capable of communicating with terminal devices 101, 102, 103 and / or server 105. Correspondingly, the large-model-based information processing apparatus provided in this embodiment can also be located in a server or server cluster that is different from server 105 and capable of communicating with terminal devices 101, 102, 103 and / or server 105.

[0047] For example, a user can input a question to be processed using terminal devices 101, 102, and 103, and transmit the question to server 105 via network 104. Upon receiving the question, server 105 uses LLM (Large Language Model) to sequentially expand the question, understand the intent, and execute the task, generating a response. Finally, the response is fed back to terminal devices 101, 102, and 103 via network 104.

[0048] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0049] In the technical solution disclosed herein, the collection, storage, use, processing, transmission, provision, disclosure, and application of user personal information comply with the provisions of relevant laws and regulations, necessary confidentiality measures have been taken, and there is no violation of public order and good morals.

[0050] In the technical solution disclosed herein, the user's authorization or consent is obtained before acquiring or collecting the user's personal information.

[0051] Figure 2 A flowchart illustrating a large-model-based information processing method according to an embodiment of the present disclosure is shown schematically.

[0052] like Figure 2 As shown, the method 200 includes operations S210 to S230.

[0053] In operation S210, the large model is used to expand the information of the problem to be processed and generate the target problem information.

[0054] In the S220 operation, a large model is used to perform semantic analysis on the target problem information to generate a target processing strategy.

[0055] In operation S230, by calling the large model, each processing task is executed according to the target logical relationship to generate response information for the target question.

[0056] According to embodiments of this disclosure, the large model can be a large language model. The question information to be processed can be a question input by the user. The user input methods include, but are not limited to, selecting, inputting, or interacting with the question on an interactive page. The form of the question information to be processed includes, but is not limited to, text, images, speech, and multimodal information containing at least two of the above modalities.

[0057] Users can use different ways to express the same question. For example, "Could you help me find documents from the last 3 days?" or "Could you help me find files from the last 3 days?"

[0058] Therefore, to improve the large model's ability to understand questions with similar intents but different expressions, the model can be expanded to generate at least one related question. The information in the question to be processed and the related question information have similar questioning intents, thus reducing the probability of irrelevant answers during human-computer interaction.

[0059] Generally, the clearer the intent of the question, the more accurate the response generated by the large model will be. For example: "Please help me book a flight to location A at 5:30 AM on September 3rd, XX year." However, in real-world applications, in most cases, simply relying on the question is insufficient to fully understand the user's intent. For example: "Please help me book the earliest flight to location A tomorrow."

[0060] Therefore, we can use a similar human thought process to organize the above questions, for example: first query tomorrow's date, then query the departure time of the flight to location A tomorrow, and finally sort them according to the departure time to generate the final answer.

[0061] According to embodiments of this disclosure, a large model can be used to perform semantic analysis on target question information, understand the user's questioning intent, and generate a target processing strategy. The target processing strategy may include the target logical relationships between various processing tasks used to answer the target question information.

[0062] According to embodiments of this disclosure, the target logical relationship between processing tasks can characterize the processing order between processing tasks.

[0063] For example, based on the user's question: "Please help me book the earliest flight to location A tomorrow," the target processing strategy could be: execute the following three processing tasks in the order of task identifiers: Task T1 queries the date, Task T2 queries the departure time of flights to location A based on the query results of Task T1, and Task T3 determines the earliest departure time based on the query results of Task T2.

[0064] According to embodiments of this disclosure, each processing task is executed by calling a large model according to the target logical relationship. For example, first query the date of tomorrow as September 3rd of XX year, then query the departure time of the flight to location A on September 3rd of XX year, and finally sort the queried departure times in chronological order to determine the earliest departure time, and use the earliest departure time as the reply information.

[0065] According to embodiments of this disclosure, by leveraging the understanding and generation capabilities of a large model to extend the problem to be processed, the large model can combine related questions similar to the user's question intent to process the target problem, further improving the generalization ability of the large model. Simultaneously, semantic analysis of the extended problem is performed using the large model to generate logical relationships between various processing tasks for answering the question. This allows the large model to improve the accuracy of the response information by fully understanding the question intent and combining it with logical reasoning capabilities, further enhancing the user experience.

[0066] The following is for reference. Figures 3A to 8 In conjunction with specific embodiments, Figure 2 The method shown will be further explained.

[0067] According to an embodiment of this disclosure, the above operation S210 may include the following operations: performing attribute analysis on the problem information to be processed to obtain initial attribute information; expanding the initial attribute information using a large model to generate target attribute information; and processing the problem information to be processed, the initial attribute information, and the target attribute information using the large model to generate target problem information.

[0068] According to embodiments of this disclosure, the initial attribute information may include attributes such as semantic type, keywords, and statement format. The semantic type can characterize the type of intent of the user's query; for example, setting an alarm or booking a flight may belong to the trip planning intent, while searching for documents may belong to the office management intent. The statement format may include the length of the query statement, the statement type, etc.

[0069] For example, a problem to be addressed could be "Please set my alarm for 8:00 AM tomorrow." By performing attribute analysis on the problem information, the semantic type of this problem is determined to be trip planning. Different keywords can be used for different semantic types; for example, for an intention to make a plan, the keywords could be time, location, or names.

[0070] According to embodiments of this disclosure, the initial attribute information can be expanded using a large model, which can be used to expand initial attributes such as semantic type, keywords, and sentence format.

[0071] For example, the problem to be solved could be "Check recent documents." Expanding the problem with keywords could generate related questions such as "Check yesterday's documents," "Check documents from the past 3 days," "Check last week's documents," and so on. The sentence format can also be expanded, for example, "Can you check my recent documents?" It can also be expanded by combining keywords and sentence format, for example, "Can you check my documents from the past 3 days?"

[0072] According to embodiments of this disclosure, the problem information to be processed can be used as a reference example to construct a problem extension Prompt based on the initial attribute information and the target attribute information. Then, the problem extension Prompt is input into the large model to output the target problem information.

[0073] According to embodiments of this disclosure, by expanding the problem to be processed, the diversity of the problem can be increased, enabling the large model to more accurately understand the problem raised by the user and improving the generalization of the model.

[0074] When expanding the initial attributes of a problem, related attributes can be expanded. For example, based on the initial attribute information, first hint information can be constructed; and the first hint information can be input into the large model to output the target related attribute information.

[0075] According to embodiments of this disclosure, the first prompt information includes an example of generating associated attributes based on a reference attribute. For example, the example could be generating associated attributes such as "yesterday," "today," "last week," and "last month" based on the reference attribute "recent."

[0076] Figure 3A The illustration schematically shows an extension of the problem to be processed using a large model according to an embodiment of the present disclosure.

[0077] like Figure 3A As shown, in embodiment 300A, attribute analysis is performed on the problem 310 to be processed to generate initial attributes 311. The initial attributes 311 may include initial intent type 311_1, initial keywords 311_2, and initial statement structure 311_3.

[0078] Then, the intent type, keywords, and statement structure can be extended to generate associated attributes 312. For example, Prompt A can be constructed based on the initial intent type 311_1, and Prompt A can be input into LLM301. Since PromptA includes examples of generating associated intent types based on reference intent types, LLM301 can output the extended intent type 312_1.

[0079] Similarly, based on the initial keyword 311_2, a Prompt B can be constructed that includes examples of generating related keywords based on reference keywords. LLM301 can then be used to expand the initial keyword 311_2, generating keyword types and values ​​312_2. Similarly, based on the initial statement structure 311_3, a Prompt C can be constructed that includes examples of generating related statement requirements based on reference statement requirements. LLM301 can then be used to expand the statement requirements of the initial statement structure 311_3, generating related statement requirements 312_3.

[0080] Next, the extended associated attribute 312 and the initial attribute 311 can be sampled to obtain the intent requirement and example 313_1, the keyword requirement and example 313_2, and the statement requirement and example 313_3, respectively.

[0081] Next, fill the intent requirements and example 313_1, keyword requirements and example 313_2, and statement requirements and example 313_3 into the Prompt template 314 to generate the intent Prompt, keyword Prompt, and statement requirement Prompt.

[0082] To further enhance the diversity of target problems, related prompt information can be constructed based on initial attribute information, target attribute information, pre-defined related attribute combination strategies, and problem information to be processed; and related prompt information can be input into a large model to output related problem information.

[0083] For example, according to the predetermined associated attribute combination strategy, at least two of the intent type requirements, keyword requirements, and statement requirements can be combined and then filled into the Prompt template 314 to generate intent + statement Prompt and intent + keyword Prompt, etc.

[0084] Finally, input Promp315 into LLM301, and output target problem information 316. Promp315 can include Prompts for single attributes and Prompts for combined attributes.

[0085] According to embodiments of this disclosure, by expanding the initial attributes with associated attributes and combining the expanded attributes, the diversity of questions can be rapidly increased when the number of questions to be processed is limited. This improves the large model's ability to understand complex and diverse questions, enhances the generalization of the large model, meets the questioning needs of different users, and further improves the user experience.

[0086] In improving the ability of large models to understand questions, in addition to expanding related attributes, interference attributes can also be expanded so that large models can reduce the probability of generating questions with ambiguous question intentions during the question expansion process, and further improve the accuracy of the intention of related questions.

[0087] According to embodiments of this disclosure, expanding the initial attribute information using a large model to generate target attribute information may further include the following operations: constructing second prompt information based on the initial attribute information; and inputting the second prompt information into the large model to output target interference attribute information.

[0088] According to embodiments of this disclosure, the target attribute information also includes target interference attribute information, which characterizes attributes that can interfere with the large model's accurate understanding of the question's intent. Examples include: the technical field being unrelated to the technical field involved in the initial intent, or missing keywords.

[0089] According to embodiments of this disclosure, the second prompt message includes an example of generating interference attributes based on reference attributes.

[0090] For example, the reference question could be "Look up the symptoms of allergic reactions to amoxicillin." This example could generate interfering attributes "amoxicillin", "amoxicillin", and "amoxicillin" based on the reference attribute "amoxicillin".

[0091] Figure 3B The illustration schematically shows an extension of the problem to be processed using a large model according to another embodiment of the present disclosure.

[0092] like Figure 3B As shown, Example 300B adds an expansion operation of interfering attributes to the initial attributes based on Example 300A. The expansion of interfering attributes can be carried out according to intent, keywords, and sentence structure, or the initial attributes 311 can be combined before expanding the interfering attributes.

[0093] For example, Prompt D can be constructed based on the initial property 311. Since Prompt D includes examples of generating interference properties based on the reference properties, inputting Prompt D into a large model can generate interference property 317.

[0094] Then, the interference attribute 317 is sampled to obtain negative example requirements and example 318. The negative example requirements and example 318 are filled into the Prompt template 314 to generate a negative example Prompt. The negative example Prompt can be merged with Prompt 315 in embodiment 300A to obtain Prompt 319.

[0095] It should be noted that negative examples and example 318 do not participate in the combination of associated attributes.

[0096] According to embodiments of this disclosure, by generating negative example prompts by extending the interference attributes, the large model can be prompted not to generate negative example problems when generating target problem information, thereby improving the accuracy of the large model's intention in generating problems.

[0097] When expanding initial attributes using a large model, some attribute information may be identical to the initial attribute information. Therefore, deduplication can be performed on the initial attribute information, target-related attribute information, and target interference attribute information to obtain the target attribute information. This reduces the probability of the same problem recurring in the target problem information and further improves the information processing efficiency of the large model.

[0098] The ability and scope of a large model to understand intent are closely related to the amount of input data. Therefore, in order to further improve the generalization ability of a large model, iterative loops can be used to generate target problem information.

[0099] Figure 4 A schematic diagram illustrating an extension problem using a large model iterative loop according to an embodiment of the present disclosure is shown.

[0100] like Figure 4 As shown, this embodiment 400 may include operations S411 to S415.

[0101] In operation S411, attribute analysis is performed on the problem information to be processed to obtain initial attribute information.

[0102] In operation S412, the initial attribute information is expanded using the large model to generate target attribute information.

[0103] In operation S413, the large model is used to process the problem information, initial attribute information and target attribute information to generate related problem information.

[0104] In operation S414, determine whether the predetermined number of cycles has been reached. If yes, then execute operation S415; otherwise, return to execute operation S411 for the associated problem information.

[0105] For example: return attribute analysis operation for related question information to generate initial attribute information for related questions; expand the initial attribute information for related questions using a large model to generate target attribute information for related questions; and return related question information, initial attribute information for related questions, and target attribute information for related questions to generate related question information, until a predetermined number of iterations are reached to generate target question information.

[0106] In operation S415, the associated problem information and the problem information to be processed obtained in each loop are merged to generate the target problem information.

[0107] For example, in the first iteration, three related questions can be generated based on the problem to be processed. The initial attributes of these three related questions can be expanded to generate target attribute information for the related questions. Then, based on the target attribute information, the initial attribute information, the predetermined combination of related attributes, and the problem information to be processed, new related prompts are constructed. These new prompts are then input into the larger model to generate new related questions. This process continues until a predetermined number of iterations is reached, for example, five iterations, to generate the target question information.

[0108] According to embodiments of this disclosure, the target attribute information of the associated questions generated in each loop process can also be arbitrarily combined with the initial attributes of the question to be processed or the target attribute information of the question to be processed to construct associated prompt information.

[0109] Similarly, before combining the extended attributes, deduplication can be used to improve the information processing efficiency of using large models to expand problems.

[0110] According to embodiments of this disclosure, based on the extended association problem, the association problem is further extended through iterative loops, which further improves the diversity of association problems and enhances the generalization ability of the large model.

[0111] Figure 5 The illustration shows a schematic diagram of a semantic analysis and generation processing strategy for target problem information using a large model, according to an embodiment of the present disclosure.

[0112] like Figure 5 As shown, this embodiment 500 may include operations S521 to S524.

[0113] In operation S521, determine whether the questioning intent of the target question information 316 is clear. If yes, then execute operation S522; otherwise, execute operation S523.

[0114] In operation S522, policy Prompt1 is constructed. Policy Prompt1 is then input into LLM301, and processing policy Ta1 is output.

[0115] For example, a first strategy prompt can be constructed based on the target problem information, the predetermined strategy format information, and the predetermined reference strategy example; and the first strategy prompt can be input into a large model to output the target processing strategy.

[0116] According to embodiments of this disclosure, the predetermined strategy format information may be in COT (Chain of thought) format.

[0117] According to embodiments of this disclosure, the first strategy prompt information may include an example of generating a reference processing strategy based on a reference question.

[0118] For example, a reference problem could be "Zhang has 10 apples. He ate 3 on the first day and 4 on the second day. How many apples are left?" The corresponding reference processing strategy could include at least the following: Calculation task T1: Zhang has 10 apples. He ate 3 on the first day. How many apples are left? Calculation task T2: Based on the number of apples left on the first day, and after eating 4 more, how many apples are left?

[0119] According to embodiments of this disclosure, the thought chain model enables large models to break down complex and comprehensive tasks into smaller, more human-like logical thinking patterns when faced with tasks, thereby improving the processing efficiency and accuracy of each task.

[0120] By operating S523 and optimizing the target problem, we obtain the optimized problem 522.

[0121] In operation S524, policy Prompt2 is constructed. Policy Prompt2 is then input into LLM301, and processing policy Ta2 is output.

[0122] For example: in response to the detection of ambiguity in the question intent of the target question information, the target question information is optimized to generate optimized question information; based on the optimized question information, the predetermined strategy format information and the predetermined reference strategy example, a second strategy prompt information is constructed; and the second strategy prompt information is input into the large model to output the target processing strategy.

[0123] Since there are many reasons that can lead to ambiguity in question intent, the following section mainly focuses on optimizing the target question information based on these reasons. This aims to improve the question intent of the optimized question, making it easier for the larger model to understand the question intent.

[0124] To address the ambiguity in question intent caused by a lack of relevant information, such as the question "Should we schedule a meeting with A at the time we agreed on last time?", the "time we agreed on last time" is clearly unclear. However, this scenario often occurs in multi-turn interaction scenarios. Relevant information related to the question can be obtained from historical interaction information, such as "Let's schedule all our meetings for Friday afternoons from now on," and the question can be optimized accordingly.

[0125] For example, optimizing target problem information to generate optimized problem information may include the following operations: obtaining historical problem information associated with the target problem information; and using a large model to merge the historical problem information with the target problem information to generate optimized problem information.

[0126] Figure 6AA schematic diagram illustrating a problem optimization according to an embodiment of the present disclosure is shown.

[0127] like Figure 6A As shown, in this embodiment 600A, firstly, a merged Prompt can be constructed according to the target problem information 316A and historical problem information 621, following the merged Prompt template. The merged Prompt may include examples of optimizing the problem based on historical related problems of the reference problem. Then, the merged Prompt is input into LLM 301, and the optimized problem 522A is output.

[0128] For example, the optimized problem 522A could be "Make an appointment with Party A for a meeting on Friday afternoon".

[0129] According to embodiments of this disclosure, by combining the correlation between historical questions and current questions, the current question is optimized, enabling the large model to accurately understand the user's question intent and efficiently formulate a processing strategy for the optimized question, thereby improving the response speed of the large model and further enhancing the user experience.

[0130] To address the ambiguity of intent in complex problems, such as: "I'm going on a business trip to location A tomorrow, then to location B two days later, and then staying in location B for one day before returning. Could you please book my flight?", while the user's intent is clear—booking a flight—the complexity increases due to the involvement of multiple addresses. Therefore, this complex problem can be optimized by breaking it down into clearly defined and independent sub-problems.

[0131] For example, optimizing the target problem information to generate optimized problem information includes the following operations: in response to detecting that the target problem information is a complex problem, using a large model to decompose the target problem information and generate optimized problem information.

[0132] Figure 6B A schematic diagram illustrating a problem optimization according to another embodiment of this disclosure is shown.

[0133] like Figure 6B As shown, in embodiment 600B, a decomposition Prompt can first be constructed based on the target problem information 316B. The decomposition Prompt can include examples of obtaining multiple independent sub-problems based on the decomposition of the reference problem. Then, the decomposition Prompt is input into LLM301, and the optimized problem 522B is output. This optimized problem 522B can contain sub-problems Query1, Query2, ..., Query... n .

[0134] For example: Sub-problem Query1 could be: help me book a flight from my current location to location A tomorrow. Sub-problem Query2 could be: help me book a flight from location A to location B two days after arriving in location A. Sub-problem Query3 could be: help me book a flight from location B back to my current location one day after arriving in location B.

[0135] According to embodiments of this disclosure, decomposing complex problems reduces the difficulty of understanding them, enabling large models to more efficiently formulate processing strategies for each of the decomposed sub-problems, thereby improving the accuracy of handling complex problems and further enhancing the user experience.

[0136] To address the ambiguity in intent caused by some fields in independent questions, the clarity of the question's intent can be improved by optimizing the ambiguous fields.

[0137] For example, optimizing target question information to generate optimized question information may include the following operations: in response to detecting that the target question information is an independent question, obtaining the target field used to characterize the questioning intent from the target question information; and using a large model to optimize the target field to generate optimized question information.

[0138] Figure 6C A schematic diagram illustrating a problem optimization according to yet another embodiment of the present disclosure is shown.

[0139] like Figure 6C As shown, in embodiment 600C, an intent field 622 representing the questioning intent or intent ambiguity can be extracted from the target question information 316C. Then, an optimized Prompt is constructed based on the intent field 622. This optimized Prompt may include examples of optimizing the questioning intent based on ambiguous fields in a reference question. Next, the optimized Prompt is input into LLM 301, which outputs the optimized question 522C.

[0140] To further improve the accuracy of the question's intent, the optimized question can be interacted with by the user through an interactive interface. After receiving the user's confirmation of the optimized question, the target processing strategy can be generated using LLM.

[0141] According to embodiments of this disclosure, the operations for the above-mentioned optimization problems can also be performed in combination. For example, after decomposing a composite problem, if the detected sub-problems contain intention ambiguity fields, the sub-problems can be optimized again using a large model.

[0142] According to embodiments of this disclosure, by optimizing the intent fuzzy field of independent questions, the accuracy of question intent is further improved, thereby improving the interaction efficiency with large models and enhancing the user experience.

[0143] The embodiments disclosed herein can perform parameter reasoning on the parameters required by the API (Application Programming Interface) corresponding to each processing task based on the Function Call function of the large model, thereby calling the corresponding API interface according to the target logical relationship between each processing task to obtain the processing results of each processing task.

[0144] According to embodiments of this disclosure, the processing tasks include N, where N is an integer greater than 1. By calling a large model to execute each processing task according to the target logical relationship, and generating response information for the target question, the following operations may be included: For the nth processing task, by calling the large model, target parameters are generated. The target parameters are used to call the target operation interface to execute the nth processing task, where n is greater than 1 and less than or equal to N; Based on the target parameters, the processing result of the nth processing task is obtained by calling the target operation interface; In response to n being less than N-1, the operation of generating the target parameters is returned, and n is incremented; In response to n being equal to N-1, the processing results of each processing task are processed using the large model to generate response information.

[0145] For example, based on the user's question: "Please help me book the earliest flight to location A tomorrow," the target processing strategy could be: Execute the following two processing tasks in the order of task identifiers: Task T1 is the query date, and Task T2 is to find and book the earliest departure flight to location A based on the query results of Task T1.

[0146] According to embodiments of this disclosure, information about a target operation interface can be obtained; parameter prompt information can be constructed based on the information about the target operation interface and the nth processing task; and the parameter prompt information can be input into a large model to generate target parameters.

[0147] According to embodiments of this disclosure, the parameter prompt information may include at least the name of the target operation interface, the parameter definitions required to call the target operation interface, and the nth processing task. The target parameters may be specific parameter values ​​used to call the target operation interface. Then, these specific parameter values ​​are used to call the target operation interface, execute the nth processing task, and obtain the processing result.

[0148] For example: First, for task T1, it can be determined that a calendar query API needs to be called. Using a large model, parameter inference is performed to generate the parameter values ​​required to call the calendar query API. Based on these parameter values, the calendar query API is called to execute the calendar query task, and the result can be a specific date for tomorrow, such as March 2nd, xx year.

[0149] Then, for task T2, it can be determined that a flight ticket query API needs to be called. Using a large model, parameter inference is performed to generate the parameter values ​​required to call the flight ticket query API. Based on these parameter values, the flight ticket query API is called to execute the flight ticket query task and book the earliest departure time. The processing result could be a booked flight for March 2nd, xx year, departing at 5:00 AM.

[0150] According to embodiments of this disclosure, parameter reasoning is performed using a large model for APIs required for each processing task. Compared to the method of combining external databases for information processing in related examples, different APIs can be flexibly called based on specific problem requirements, further improving the scalability and interpretability of the large model's processing capabilities.

[0151] Since each processing task is executed according to the target logical relationship, the processing results of the previous step can be used to comprehensively verify whether the logical relationship between each processing task is accurate.

[0152] According to embodiments of this disclosure, the above method may further include the following operations: in response to receiving a processing result for the nth processing task, by calling a large model, verifying the logical relationship between the nth processing task and the (n+1)th processing task based on the processing result, and generating a verification result; in response to the verification result indicating that the logical relationship is correct, returning the execution of the target parameter generation operation and incrementing n; in response to the verification result indicating that the logical relationship is incorrect, processing the processing results of the first n processing tasks using the large model, and generating result summary information.

[0153] According to embodiments of this disclosure, a verification prompt can be constructed based on the processing result of the nth processing task, the target processing strategy, and the target parameters of the nth processing task. This verification prompt is then input into a large model to verify the accuracy of the logical relationship between the nth and (n+1)th processing tasks. If accurate, the (n+1)th processing task continues execution. If inaccurate, task execution is terminated, and a result summary is generated based on the current processing result.

[0154] For example, the target processing strategy includes 5 tasks. When the 3rd task is completed, the large model can be used to verify the logical relationship between the 3rd and 4th tasks based on the processing result of the 3rd task. If the verification passes, the 4th task is executed; if the verification fails, the processing results of the first 3 tasks are summarized.

[0155] Figure 7 The illustration shows a schematic diagram of the execution of various processing tasks according to a target logical relationship by invoking a large model according to an embodiment of the present disclosure.

[0156] like Figure 7As shown, in embodiment 700, firstly, in operation S731, task T1 in processing strategy Ta1521 is executed: the parameter Prompt1 is constructed based on the parameters required by AP11 corresponding to task T1, and the parameter Prompt1 is input into LLM301. Parameter reasoning is performed using LLM301 to obtain the parameter value Va1731. Then, based on the parameter value Va1731, API1701 is called to generate the processing result Re1732.

[0157] Then, a verification prompt is constructed based on the processing strategy Ta1521 and the processing result Re1732, and the verification prompt is input into LLM301 to output the verification result. Then, operation S732 is executed to determine whether the verification result passes. If it fails, the LLM summary task processing result is called to generate the result summary 733.

[0158] Next, if successful, operation S733 is executed to determine task T2. Based on the parameters required by AP12 corresponding to task T2, parameter Prompt2 is constructed. Parameter Prompt2 is input into LLM301, and parameter inference is performed using LLM301 to obtain parameter value Va2734. Based on parameter value Va2734, API2702 is called to generate processing result Re2735.

[0159] Finally, a summary Prompt is constructed based on the processing results Re1732 and Re2735, and the summary Prompt is input into LLM301 to generate response information 736.

[0160] According to embodiments of this disclosure, a large model is used in a thought-guided manner to verify the logical relationships between adjacent tasks in the processing strategy, providing a step-by-step supervision mechanism for the target processing strategy. This improves the interpretability of the large model's decision-making process for generating response information, enabling users to clearly understand the large model's thought-making process and enhancing the user experience.

[0161] For specific application scenarios, in order to further improve the accuracy of responses from large models, a feedback optimization mechanism can be introduced to optimize the model parameters of large models.

[0162] For example, using a large model to perform quality checks on response information and generate detection results; and optimizing the model parameters of the large model based on the detection results.

[0163] According to embodiments of this disclosure, the detection result characterizes the correlation between the response information and the question intent. The correlation between the response information and the question intent can be calculated based on periodic user feedback, or the response information can be quality-checked based on a large model.

[0164] For example: extract intent category information and keyword information from target question information; construct detection prompt information based on intent category information, keyword information, target question information and response information; and input the detection prompt information into a large model to output detection results.

[0165] According to embodiments of this disclosure, the detection prompt information may include examples of generating quality detection results based on reference questions and reference answers. The quality detection results may be quality category information or a score used to characterize a quality level, without specific limitations herein.

[0166] According to embodiments of this disclosure, for responses that are detected as high-quality, the query corresponding to the high-quality response can be used as a positive sample to optimize model parameters. For responses that are detected as low-quality, the query corresponding to the low-quality response can be used as a negative sample to optimize model parameters.

[0167] Figure 8 The illustration shows a schematic diagram of optimizing the model parameters of a large model based on response information feedback according to an embodiment of the present disclosure.

[0168] like Figure 8 As shown, in embodiment 800, firstly, intent category 841 and keywords 842 are extracted from the target question information 316. Next, a detection prompt 843 is constructed based on the response information 736, intent category 841, and keywords 842. The detection prompt 843 is then input into LLM 301, and a detection result 844 is output. The detection result 844 can be a quality score.

[0169] Then, operation S840 is performed to determine whether the detection result is less than a predetermined threshold. If yes, operation S841 is performed; otherwise, operation S842 is performed.

[0170] During operation of S841, the target problem information was determined to be a negative sample.

[0171] During operation of S842, the target problem information was determined to be a positive sample.

[0172] When operating the S843, optimize the model parameters of the LLM.

[0173] According to embodiments of this disclosure, by performing quality checks on the response information and labeling positive and negative samples for optimizing model parameters based on the check results, the problem of relying on manual labeling to optimize model parameters in related examples is solved, and closed-loop iterative optimization of large models is realized.

[0174] Figure 9 A block diagram of a large-model-based information processing apparatus according to an embodiment of the present disclosure is shown schematically.

[0175] like Figure 9 As shown, the device 900 includes a question expansion module 910, a strategy generation module 920, and a response generation module 930.

[0176] The problem expansion module 910 is used to expand the problem information to be processed using the large model to generate target problem information; wherein, the target problem information includes at least one related problem information, and the problem information to be processed and the related problem information have similar questioning intent.

[0177] The strategy generation module 920 is used to perform semantic analysis on the target question information using a large model to generate a target processing strategy; wherein, the target processing strategy includes the target logical relationship between each processing task used to answer the target question information.

[0178] The response generation module 930 is used to generate response information for the target question by calling the large model to execute various processing tasks according to the target logical relationship.

[0179] According to embodiments of this disclosure, the problem expansion module includes: an attribute analysis submodule, an attribute expansion submodule, and a problem generation submodule.

[0180] The attribute analysis submodule is used to perform attribute analysis on the information of the problem to be processed to obtain initial attribute information.

[0181] The attribute extension submodule is used to extend the initial attribute information using the large model to generate target attribute information.

[0182] The problem generation submodule is used to process the problem information, initial attribute information and target attribute information of the large model to generate target problem information.

[0183] According to embodiments of this disclosure, the target attribute information includes target associated attribute information; the attribute extension submodule includes: a first construction unit and a first output unit.

[0184] The first construction unit is used to construct the first prompt information based on the initial attribute information; wherein the first prompt information includes an example of generating associated attributes based on reference attributes.

[0185] The first output unit is used to input the first prompt information into the large model and output the target associated attribute information.

[0186] According to embodiments of this disclosure, the target attribute information further includes target interference attribute information. The attribute extension submodule further includes: a second construction unit and a second output unit.

[0187] The second construction unit is used to construct the second prompt information based on the initial attribute information, wherein the second prompt information includes an example of generating interference attributes based on the reference attributes.

[0188] The second output unit is used to input the second prompt information into the large model and output the target interference attribute information.

[0189] According to embodiments of this disclosure, the attribute expansion submodule further includes a deduplication unit, used to perform deduplication processing on the initial attribute information, target associated attribute information and target interference attribute information to obtain target attribute information.

[0190] According to embodiments of this disclosure, the problem generation submodule includes: a first generation unit, an attribute analysis unit, an attribute expansion unit, and a second generation unit.

[0191] The first generation unit is used to process the problem information, initial attribute information and target attribute information of the large model to generate related problem information.

[0192] The attribute analysis unit is used to perform attribute analysis operations on the returned information of related questions, and generate the initial attribute information of related questions.

[0193] The attribute expansion unit is used to expand the initial attribute information of the association problem using the large model, and generate the target attribute information of the association problem.

[0194] The second generation unit is used to return and execute the generation operation of the associated question information based on the associated question information, the initial attribute information of the associated question, and the target attribute information of the associated question, until a predetermined number of loops are reached to generate the target question information.

[0195] According to embodiments of this disclosure, the first generation unit includes: a first construction subunit and a first output subunit.

[0196] The first construction subunit is used to construct associated prompt information based on initial attribute information, target attribute information, predetermined associated attribute combination strategy, and information on the problem to be processed.

[0197] The first output sub-unit is used to input the associated prompt information into the large model and output the associated question information.

[0198] According to embodiments of this disclosure, the strategy generation module includes a first construction submodule and a first output submodule.

[0199] The first construction submodule is used to construct a first strategy prompt message in response to the detection of a clear question intent regarding the target question information, based on the target question information, the predetermined strategy format information, and the predetermined reference strategy example.

[0200] The first output submodule is used to input the first strategy prompt information into the large model and output the target processing strategy.

[0201] According to embodiments of this disclosure, the strategy generation module further includes: an optimization submodule, a second construction submodule, and a second output submodule.

[0202] The optimization submodule is used to optimize the target question information in response to the detection of ambiguity in the question intent, and generate optimized question information.

[0203] The second construction submodule is used to construct the second strategy prompt information based on the optimized problem information, the predetermined strategy format information, and the predetermined reference strategy example.

[0204] The second output submodule is used to input the second strategy prompt information into the large model and output the target processing strategy.

[0205] According to embodiments of this disclosure, the optimization submodule includes: a first acquisition unit and a merging unit.

[0206] The first acquisition unit is used to acquire historical problem information associated with the target problem information.

[0207] The merging unit is used to combine historical problem information with target problem information using a large model to generate optimized problem information.

[0208] According to embodiments of this disclosure, the optimization submodule includes: a decomposition unit, configured to decompose the target problem information using a large model in response to detecting that the target problem information is a complex problem, thereby generating optimized problem information.

[0209] According to embodiments of this disclosure, the optimization submodule includes a field acquisition unit and an optimization unit.

[0210] The field acquisition unit is used to retrieve target fields that characterize the questioning intent from the target question information in response to the detection that the target question information is an independent question.

[0211] The optimization unit is used to optimize the target field using a large model and generate optimized problem information.

[0212] According to embodiments of this disclosure, the processing tasks include N, where N is an integer greater than 1; the response generation module includes: a parameter generation submodule, a first acquisition submodule, a first loop submodule, and a response generation submodule.

[0213] The parameter generation submodule is used to generate target parameters for the nth processing task by calling the large model. The target parameters are used to call the target operation interface to execute the nth processing task, where n is greater than 1 and less than or equal to N.

[0214] The first acquisition submodule is used to obtain the processing result of the nth processing task by calling the target operation interface based on the target parameters.

[0215] The first loop submodule is used to respond to n being less than N-1, return to perform the generation operation of the target parameters, and increment n.

[0216] The response generation submodule is used to process the processing results of each processing task in response to n equal to N-1, and generate response information.

[0217] According to embodiments of this disclosure, the parameter generation submodule includes: a second acquisition unit, a third construction unit, and a third output unit.

[0218] The second acquisition unit is used to acquire information about the target operation interface.

[0219] The third construction unit is used to construct parameter prompt information based on the information of the target operation interface and the nth processing task.

[0220] The third output unit is used to input parameter prompts into the large model and generate target parameters.

[0221] According to embodiments of this disclosure, the response generation module further includes: a verification submodule, a second loop submodule, and a result generation submodule.

[0222] The verification submodule is used to respond to the received processing result for the nth processing task by calling the large model to verify the logical relationship between the nth and (n+1)th processing tasks based on the processing result and generate a verification result.

[0223] The second loop submodule is used to respond to the verification result indicating that the logical relationship is correct, return to perform the generation operation of the target parameters, and increment n.

[0224] The result generation submodule is used to process the results of the first n processing tasks using a large model in response to a logical relationship error in the verification result, and generate result summary information.

[0225] According to embodiments of this disclosure, the above-described apparatus further includes a quality inspection module and an optimization module.

[0226] The quality inspection module is used to perform quality inspection on the response information using a large model and generate inspection results, where the inspection results represent the degree of correlation between the response information and the question intent.

[0227] The optimization module is used to optimize the model parameters of large models based on the detection results.

[0228] According to embodiments of this disclosure, the quality inspection module includes: a second acquisition submodule, a third construction submodule, and a third output submodule.

[0229] The second acquisition submodule is used to obtain intent category information and keyword information from the target question information.

[0230] The third submodule is used to construct detection prompts based on intent category information, keyword information, target question information, and response information.

[0231] The third output submodule is used to input detection prompts into the large model and output the detection results.

[0232] According to embodiments of this disclosure, this disclosure also provides an intelligent agent, an electronic device, a readable storage medium, and a computer program product.

[0233] According to embodiments of this disclosure, an intelligent agent includes: an input module, a processing module, and an output module.

[0234] The input module is used to receive information about problems to be processed.

[0235] The processing module is used to determine the target task based on the problem information to be processed received by the input module, determine the target large model based on the target task, and obtain the response information by calling the target large model to execute the above method.

[0236] The output module is used to output the response information obtained by the processing module.

[0237] According to an embodiment of the present disclosure, an electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method described above.

[0238] According to embodiments of the present disclosure, a non-transitory computer-readable storage medium stores computer instructions, wherein the computer instructions are used to cause a computer to perform the method described above.

[0239] According to an embodiment of this disclosure, a computer program product includes a computer program that, when executed by a processor, implements the method described above.

[0240] Figure 10A schematic block diagram of an example electronic device 1000 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0241] like Figure 10 As shown, device 1000 includes a computing unit 1001, which can perform various appropriate actions and processes according to a computer program stored in read-only memory (ROM) 1002 or a computer program loaded into random access memory (RAM) 1003 from storage unit 1008. The RAM 1003 may also store various programs and data required for the operation of device 1000. The computing unit 1001, ROM 1002, and RAM 1003 are interconnected via bus 1004. Input / output (I / O) interface 1005 is also connected to bus 1004.

[0242] Multiple components in device 1000 are connected to I / O interface 1005, including: input unit 1006, such as keyboard, mouse, etc.; output unit 1007, such as various types of monitors, speakers, etc.; storage unit 1008, such as disk, optical disk, etc.; and communication unit 1009, such as network card, modem, wireless transceiver, etc. Communication unit 1009 allows device 1000 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0243] The computing unit 1001 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1001 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1001 performs the various methods and processes described above, such as large model-based information processing methods. For example, in some embodiments, the large model-based information processing method can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 1008. In some embodiments, part or all of the computer program can be loaded and / or installed on device 1000 via ROM 1002 and / or communication unit 1009. When the computer program is loaded into RAM 1003 and executed by the computing unit 1001, one or more steps of the large model-based information processing method described above can be performed. Alternatively, in other embodiments, the computing unit 1001 may be configured to perform a large-model-based information processing method by any other suitable means (e.g., by means of firmware).

[0244] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0245] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0246] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0247] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0248] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0249] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, distributed system servers, or servers incorporating blockchain technology.

[0250] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0251] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. An information processing method based on a large model, comprising: The large model is used to perform attribute analysis on the information of the problem to be processed, and the initial attribute information is obtained. The initial attribute information is expanded using the large model to generate target attribute information; The large model is used to process the question information to be processed, the initial attribute information, and the target attribute information to generate target question information; wherein, the target question information includes at least one related question information, and the question information to be processed and the related question information have similar questioning intent; The target question information is semantically analyzed using the large model to generate a target processing strategy; wherein, the target processing strategy includes the target logical relationships between various processing tasks for responding to the target question information; and By invoking the large model to execute each of the processing tasks according to the target logical relationship, response information for the target question is generated.

2. The method according to claim 1, wherein, The target attribute information includes target associated attribute information; The step of expanding the initial attribute information using the large model to generate target attribute information includes: Based on the initial attribute information, a first prompt message is constructed; wherein, the first prompt message includes an example of generating associated attributes based on reference attributes; and The first prompt information is input into the large model, and the target associated attribute information is output.

3. The method according to claim 2, wherein, The target attribute information also includes target interference attribute information, which represents attributes that can interfere with the large model's accurate understanding of the question's intent. The step of expanding the initial attribute information using the large model to generate target attribute information further includes: Based on the initial attribute information, a second prompt message is constructed, wherein the second prompt message includes an example of generating interfering attributes based on reference attributes; and The second prompt information is input into the large model, and the target interference attribute information is output.

4. The method according to claim 3, further comprising: The initial attribute information, the target associated attribute information, and the target interference attribute information are deduplicated to obtain the target attribute information.

5. The method according to any one of claims 1-4, wherein, The process of using the large model to process the problem information to be processed, the initial attribute information, and the target attribute information to generate the target problem information includes: The large model is used to process the problem information to be processed, the initial attribute information, and the target attribute information to generate the associated problem information; The associated question information is returned to perform an attribute analysis operation to generate initial attribute information for the associated questions; The initial attribute information of the association problem is expanded using the large model to generate the target attribute information of the association problem; and For the associated question information, the initial attribute information of the associated question, and the target attribute information of the associated question, return to execute the associated question information generation operation until a predetermined number of loops are reached, and generate the target question information.

6. The method according to claim 5, wherein, The process of using the large model to process the problem information to be processed, the initial attribute information, and the target attribute information to generate the associated problem information includes: Based on the initial attribute information, target attribute information, predetermined associated attribute combination strategy, and the problem information to be processed, construct associated prompt information; and The associated prompt information is input into the large model, and the associated question information is output.

7. The method according to claim 1, wherein the step of performing semantic analysis on the target problem information using the large model to generate a target processing strategy includes: In response to the detection that the target question information is clearly intended to be asked, a first strategy prompt information is constructed based on the target question information, the predetermined strategy format information, and the predetermined reference strategy example; as well as The first strategy prompt information is input into the large model, and the target processing strategy is output.

8. The method according to claim 7, further comprising: In response to the detection that the question intent of the target question information is ambiguous, the target question information is optimized to generate optimized question information; Based on the optimized problem information, the predetermined strategy format information, and the predetermined reference strategy example, construct the second strategy prompt information; as well as The second strategy prompt information is input into the large model, and the target processing strategy is output.

9. The method according to claim 8, wherein, The optimization of the target problem information to generate optimized problem information includes: Obtain historical problem information associated with the target problem information; The historical problem information and the target problem information are merged using the large model to generate the optimized problem information.

10. The method according to claim 8, wherein, The optimization of the target problem information to generate optimized problem information includes: In response to the detection that the target problem information is a complex problem, the target problem information is decomposed using the large model to generate the optimized problem information.

11. The method according to claim 8, wherein, The optimization of the target problem information to generate optimized problem information includes: In response to detecting that the target question information is an independent question, a target field characterizing the questioning intent is obtained from the target question information; and The target field is optimized using the large model to generate the optimized problem information.

12. The method according to claim 1, wherein, The processing tasks include N, where N is an integer greater than 1; The step of generating response information for the target question by calling the large model to execute each of the processing tasks according to the target logical relationship includes: For the nth processing task, target parameters are generated by calling the large model. These target parameters are used to call the target operation interface to execute the nth processing task, where n is greater than 1 and less than or equal to N. Based on the target parameters, the processing result of the nth processing task is obtained by calling the target operation interface; If n is less than N-1, return to perform the target parameter generation operation and increment n; In response to n equaling N-1, the processing results of each of the processing tasks are processed using the large model to generate the response information.

13. The method according to claim 12, wherein, For the nth processing task, the target parameters are generated by calling the large model, including: Obtain information about the target operation interface; Based on the information of the target operation interface and the nth processing task, construct parameter prompt information; and The parameter prompt information is input into the large model to generate the target parameters.

14. The method according to claim 12, further comprising: In response to receiving the processing result for the nth processing task, the large model is invoked to verify the logical relationship between the nth processing task and the (n+1)th processing task based on the processing result, and a verification result is generated. In response to the verification result indicating that the logical relationship is correct, return to the operation of generating the target parameter and increment n; In response to the verification result indicating an error in the logical relationship, the processing results of the first n processing tasks are processed using the large model to generate summary information of the results.

15. The method according to claim 14, further comprising: The large model is used to perform quality checks on the response information, generating a detection result, wherein the detection result characterizes the correlation between the response information and the question intent; and The model parameters of the large model are optimized based on the detection results.

16. The method according to claim 15, wherein, The process of using the large model to perform quality checks on the response information and generating check results includes: Obtain intent category information and keyword information from the target question information; Based on the intent category information, the keyword information, the target question information, and the response information, a detection prompt information is constructed; and The detection prompt information is input into the large model, and the detection result is output.

17. An information processing device based on a large model, comprising: The problem expansion module includes: an attribute analysis submodule, used to perform attribute analysis on the problem information to be processed using a large model to obtain initial attribute information; an attribute expansion submodule, used to expand the initial attribute information using the large model to generate target attribute information; and a problem generation submodule, used to process the problem information to be processed, the initial attribute information, and the target attribute information using the large model to generate target problem information. The target question information includes at least one related question information, wherein the question information to be processed and the related question information have similar questioning intent; The strategy generation module is used to perform semantic analysis on the target question information using the large model to generate a target processing strategy; wherein, the target processing strategy includes the target logical relationship between various processing tasks for responding to the target question information; and The response generation module is used to generate response information for the target question by calling the large model to execute each of the processing tasks according to the target logical relationship.

18. The apparatus according to claim 17, wherein, The target attribute information includes target associated attribute information; The attribute extension submodule includes: A first construction unit is configured to construct first prompt information based on the initial attribute information; wherein the first prompt information includes an example of generating associated attributes based on reference attributes; and The first output unit is used to input the first prompt information into the large model and output the target associated attribute information.

19. The apparatus according to claim 18, wherein, The target attribute information also includes target interference attribute information, which represents attributes that can interfere with the large model's accurate understanding of the question's intent. The attribute extension submodule also includes: The second construction unit is configured to construct second prompt information based on the initial attribute information, wherein the second prompt information includes an example of generating interference attributes based on reference attributes; and The second output unit is used to input the second prompt information into the large model and output the target interference attribute information.

20. The apparatus of claim 19, wherein the attribute extension submodule further comprises: The deduplication unit is used to perform deduplication processing on the initial attribute information, the target associated attribute information, and the target interference attribute information to obtain the target attribute information.

21. The apparatus according to any one of claims 17-20, wherein, The problem generation submodule includes: The first generation unit is used to process the problem information to be processed, the initial attribute information, and the target attribute information using the large model to generate the associated problem information; The attribute analysis unit is used to perform attribute analysis operations on the returned associated question information to generate initial attribute information for the associated questions. The attribute expansion unit is used to expand the initial attribute information of the association problem using the large model to generate the target attribute information of the association problem; and The second generation unit is used to return and execute the generation operation of the associated question information for the associated question information, the initial attribute information of the associated question, and the target attribute information of the associated question, until a predetermined number of loops are reached, and generate the target question information.

22. The apparatus according to claim 21, wherein, The first generation unit includes: The first construction subunit is used to construct associated prompt information based on the initial attribute information, target attribute information, predetermined associated attribute combination strategy, and the problem information to be processed; and The first output subunit is used to input the associated prompt information into the large model and output the associated question information.

23. The apparatus of claim 17, wherein the strategy generation module comprises: The first construction submodule is used to construct a first strategy prompt message in response to the detection that the target question information is clearly intended to ask the question, based on the target question information, the predetermined strategy format information, and the predetermined reference strategy example; as well as The first output submodule is used to input the first strategy prompt information into the large model and output the target processing strategy.

24. The apparatus according to claim 23, wherein the strategy generation module further comprises: The optimization submodule is used to optimize the target question information in response to the detection that the question intent of the target question information is ambiguous, and generate optimized question information; The second construction submodule is used to construct the second strategy prompt information based on the optimized problem information, the predetermined strategy format information, and the predetermined reference strategy example; as well as The second output submodule is used to input the second strategy prompt information into the large model and output the target processing strategy.

25. The apparatus according to claim 24, wherein, The optimization submodule includes: The first acquisition unit is used to acquire historical problem information associated with the target problem information; The merging unit is used to merge the historical problem information and the target problem information using the large model to generate the optimized problem information.

26. The apparatus according to claim 24, wherein, The optimization submodule includes: The decomposition unit is used to decompose the target problem information into the optimized problem information in response to the detection that the target problem information is a complex problem, using the large model.

27. The apparatus according to claim 24, wherein, The optimization submodule includes: A field acquisition unit is configured to, in response to detecting that the target question information is an independent question, acquire a target field representing the questioning intent from the target question information; and The optimization unit is used to optimize the target field using the large model to generate the optimized problem information.

28. The apparatus according to claim 17, wherein, The processing tasks include N, where N is an integer greater than 1; The response generation module includes: The parameter generation submodule is used to generate target parameters for the nth processing task by calling the large model. The target parameters are used to call the target operation interface to execute the nth processing task, where n is greater than 1 and less than or equal to N. The first acquisition submodule is used to obtain the processing result of the nth processing task by calling the target operation interface based on the target parameters; The first loop submodule is used to respond to n being less than N-1, return to perform the generation operation of the target parameters, and increment n. The response generation submodule is used to process the processing results of each of the processing tasks using the large model in response to n equaling N-1, and generate the response information.

29. The apparatus according to claim 28, wherein, The parameter generation submodule includes: The second acquisition unit is used to acquire information about the target operation interface; The third construction unit is used to construct parameter prompt information based on the information of the target operation interface and the nth processing task; and The third output unit is used to input the parameter prompt information into the large model and generate the target parameters.

30. The apparatus of claim 28, wherein the response generation module further comprises: The verification submodule is used to respond to the received processing result for the nth processing task by calling the large model to verify the logical relationship between the nth processing task and the (n+1)th processing task based on the processing result, and generate a verification result. The second loop submodule is used to respond to the verification result indicating that the logical relationship is correct, return to perform the target parameter generation operation, and increment n; The result generation submodule is used to process the processing results of the first n processing tasks using the large model in response to the verification result indicating a logical relationship error, and to generate result summary information.

31. The apparatus of claim 30, further comprising: A quality inspection module is used to perform quality inspection on the response information using the large model, and generate inspection results, wherein the inspection results characterize the correlation between the response information and the question intent; and An optimization module is used to optimize the model parameters of the large model based on the detection results.

32. The apparatus according to claim 31, wherein, The quality inspection module includes: The second acquisition submodule is used to acquire intent category information and keyword information from the target question information; The third construction submodule is used to construct detection prompt information based on the intent category information, the keyword information, the target question information, and the response information; and The third output submodule is used to input the detection prompt information into the large model and output the detection result.

33. An intelligent agent, comprising: The input module is used to receive information about the problems to be processed. The processing module is configured to determine a target task based on the problem information to be processed received by the input module, determine a target large model based on the target task, and execute the method described in any one of claims 1-16 by calling the target large model to obtain response information; as well as An output module is used to output the response information obtained by the processing module.

34. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-16.

35. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-16.

36. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-16.

Images