Interaction method based on large model, method for training large model and intelligent agent

Abstract

The invention provides an interaction method based on a large model, a method for training the large model and an intelligent agent, and relates to the technical field of artificial intelligence, in particular to the technical fields of large models, intelligent search, intelligent medical treatment, intelligent e-commerce and the like. The interaction method based on the large model comprises the following steps: receiving demand information of a target object; performing semantic understanding on the demand information by using the first large model, outputting a clarification prompt text which is used for describing a demand intention with an intention semantic fuzzy attribute in the demand information, and prompting the target object to feed back clarification demand content corresponding to the demand intention; executing the retrieval task according to clarified demand content input by the target object to obtain reply information matched with the demand intention; and pushing reply information to the target object.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, particularly to the fields of large models, intelligent search, smart healthcare, and smart e-commerce. Background Technology

[0002] With the rapid development of artificial intelligence technology, large models can understand users' search intent by analyzing the information they input, and retrieve the data that users need based on that intent. Summary of the Invention

[0003] This disclosure provides an interaction method based on a large model, a method for training a large model, and an intelligent agent.

[0004] According to one aspect of this disclosure, a large-scale model-based interaction method is provided, comprising: receiving demand information from a target object; performing semantic understanding on the demand information using a first large-scale model, outputting clarification prompt text, wherein the clarification prompt text is used to describe the demand intent with semantic ambiguity in the demand information, and prompting the target object to provide clarification demand content corresponding to the demand intent; performing a retrieval task based on the clarification demand content input by the target object, obtaining response information matching the demand intent; and pushing the response information to the target object.

[0005] According to another aspect of this disclosure, a method for training a large model is provided, comprising: receiving training data, the training data including sample demand information and label clarification prompt text, wherein the sample demand information represents the demand intent of a sample object and has semantic ambiguity attributes, and the label clarification prompt text is used to prompt the sample object to provide clarification demand content corresponding to the demand intent, so as to perform a retrieval task based on the clarification demand content to obtain retrieval results matching the sample object; performing semantic understanding of the sample demand information using a first large model to be trained, and outputting sample clarification prompt text; and training the first large model based on the label clarification prompt text and the sample clarification prompt text to obtain the trained first large model.

[0006] According to another aspect of this disclosure, a large-model-based interactive device is provided, comprising: a first receiving module for receiving demand information of a target object; a clarification prompt text obtaining module for performing semantic understanding of the demand information using the first large-model and outputting clarification prompt text, wherein the clarification prompt text describes the demand intent with semantic ambiguity in the demand information and prompts the target object to provide clarification demand content corresponding to the demand intent; a response information obtaining module for performing a retrieval task based on the clarification demand content input by the target object and obtaining response information matching the demand intent; and a push module for pushing the response information to the target object.

[0007] According to another aspect of this disclosure, an apparatus for training a large model is provided, comprising: a second receiving module for receiving training data, the training data including sample demand information and label clarification prompt text, wherein the sample demand information represents the demand intent of a sample object and has semantic ambiguity, and the label clarification prompt text is used to prompt the sample object to provide clarification demand content corresponding to the demand intent, so as to perform a retrieval task based on the clarification demand content to obtain retrieval results matching the sample object; a sample clarification prompt text obtaining module for performing semantic understanding of the sample demand information using a first large model to be trained, and outputting sample clarification prompt text; and a training module for training the first large model based on the label clarification prompt text and the sample clarification prompt text, to obtain the trained first large model.

[0008] According to another aspect of this disclosure, an artificial intelligence agent is provided, comprising: an input module for receiving input information; a processing module for determining a target task based on the input information received by the input module, determining a large model based on the target task, and obtaining output information by calling the large model to execute the method provided in the embodiments of this disclosure; and an output module for outputting the output information obtained by the processing module.

[0009] 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 a method provided according to an embodiment of this disclosure.

[0010] 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 a method provided according to an embodiment of this disclosure.

[0011] 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 provided according to embodiments of this disclosure.

[0012] 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

[0013] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein:

[0014] Figure 1The illustration schematically depicts an exemplary system architecture for applying large-model-based interaction methods and apparatus according to embodiments of the present disclosure.

[0015] Figure 2 A flowchart illustrating a large-model-based interaction method according to an embodiment of the present disclosure is shown schematically.

[0016] Figure 3 The diagram illustrates an application scenario of the interaction method based on a large model provided according to an embodiment of the present disclosure.

[0017] Figure 4 A flowchart illustrating a method for training a large model according to an embodiment of this disclosure is shown schematically.

[0018] Figure 5 A schematic diagram illustrating a method for training a large model according to an embodiment of the present disclosure is shown.

[0019] Figure 6 A block diagram of a large-model-based interactive device according to an embodiment of the present disclosure is shown schematically.

[0020] Figure 7 A block diagram of an apparatus for training a large model according to an embodiment of the present disclosure is shown schematically.

[0021] Figure 8 A schematic block diagram of an artificial intelligence agent according to an embodiment of the present disclosure is shown.

[0022] Figure 9 A schematic block diagram of an example electronic device is shown, which can be used to implement embodiments of the present disclosure and methods for training large models for interaction. Detailed Implementation

[0023] 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.

[0024] In the technical solution disclosed herein, the acquisition, storage, 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.

[0025] The inventors discovered that during dialogues with users using large models, the information input by users may have ambiguities such as missing intent or unclear semantic expression of needs. This makes it difficult to accurately retrieve the information needed by the user based on the input information, or requires a large amount of computing resources to perform redundant retrieval so that the redundant information output by the large model contains the information that the user needs.

[0026] This disclosure provides a large-model-based interaction method, a method for training the large-model, and an intelligent agent. The large-model-based interaction method includes: receiving demand information from a target object; using a first large-model to perform semantic understanding on the demand information and outputting clarification prompt text, wherein the clarification prompt text describes the demand intent with semantic ambiguity in the demand information and prompts the target object to provide clarification content corresponding to the demand intent; performing a retrieval task based on the clarification content input by the target object to obtain response information matching the demand intent; and pushing the response information to the target object.

[0027] According to embodiments of this disclosure, by utilizing a first large model to perform semantic understanding on demand information, the system can accurately identify semantically ambiguous demand intentions within the demand information. This allows the large model to output clarification prompt text based on the identified demand intentions with ambiguous semantic attributes. The target object can then respond with clear clarification of its demand intentions based on the clarification prompt text. Furthermore, by using the large model to accurately identify and prompt semantically ambiguous demand intentions within the demand information, the system reduces the frequency of interaction where the target object needs to input clarification responses during the interaction process. It also avoids the computational resource redundancy overhead caused by expanding excessive redundant search information to perform search tasks, thus ensuring that the response information retrieved based on the clarification of demand content accurately matches the target object's needs. This achieves the technical effects of improving interaction efficiency and user experience while reducing redundant computational overhead.

[0028] Figure 1 The illustration schematically depicts an exemplary system architecture for applying large-model-based interaction methods and apparatus according to embodiments of the present disclosure.

[0029] It is important to note that Figure 1 The examples shown are merely examples of system architectures that can be applied to embodiments of this disclosure, intended to help those skilled in the art understand the technical content of this disclosure. However, they do not imply that 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 large-model-based interaction method and apparatus may include a terminal device. However, the terminal device may implement the large-model-based interaction method and apparatus provided by embodiments of this disclosure without interacting with a server.

[0030] like Figure 1 As shown, the system architecture 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing communication links between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links, etc.

[0031] Users can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 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).

[0032] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0033] Server 105 can be a server that provides various services, such as a backend management server that supports the content browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). 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.

[0034] It should be noted that the large-model-based interaction method provided in this disclosure can generally be executed by the first terminal device 101, the second terminal device 102, or the third terminal device 103. Correspondingly, the large-model-based interaction device provided in this disclosure can also be disposed in the first terminal device 101, the second terminal device 102, or the third terminal device 103.

[0035] Alternatively, the large-model-based interaction method provided in this embodiment can generally be executed by server 105. Correspondingly, the large-model-based interaction device provided in this embodiment can generally be located in server 105. The large-model-based interaction 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 the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Correspondingly, the large-model-based interaction device 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 the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105.

[0036] 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.

[0037] Figure 2 A flowchart illustrating a large-model-based interaction method according to an embodiment of the present disclosure is shown schematically.

[0038] like Figure 2 As shown, the interaction method based on the large model includes operations S210~S240.

[0039] In operation S210, the requirement information of the target object is received.

[0040] When operating S220, the first major model is used to perform semantic understanding of the demand information and output clarification prompt text.

[0041] In operation S230, a retrieval task is performed based on the clarification request content input by the target object, and a response information matching the request intent is obtained.

[0042] In operation S240, a response message is pushed to the target object.

[0043] According to embodiments of this disclosure, the demand information of the target object may include data of any modality such as voice data and text data. Embodiments of this disclosure do not limit the specific data modality type of the demand information.

[0044] According to embodiments of this disclosure, the first large model can be any type of large model, such as a trained large speech model or a multimodal large model. The large model can have hundreds of millions or even billions of model parameters. Based on the relatively powerful semantic understanding capabilities of the large model, the demand intention represented by the demand information can be semantically understood to perform the text generation task and obtain the clarification prompt text.

[0045] According to embodiments of this disclosure, the clarification prompt text is used to describe the requirement intent with semantic ambiguity in the requirement information, and to prompt the target object to provide clarification of the requirement content corresponding to the requirement intent. The requirement intent with semantic ambiguity can be represented by semantic information such as textual semantics and visual semantics in the target object's requirement information, indicating a requirement intent with semantic ambiguity or unclear content representation.

[0046] For example, given the requirement "Where is Park A?", the first large model, through semantic understanding of the requirement information, can determine that the keyword "Park A," representing the intent of the subject name, could refer to multiple subjects such as "Industrial Park A" or "Science and Technology Park AA." This allows the first large model to understand the semantically ambiguous intent in the requirement information and generate clarifying prompt text to describe the semantically ambiguous subject name intent. This prompt text then guides the target user to provide clarification on the subject name intent based on the clarification prompt text. Thus, after reviewing the clarification prompt text indicating semantic ambiguity regarding the subject name intent, the target user can directly and accurately input the clarification on the subject name intent, avoiding multiple rounds of questioning the target user's intent, which increases interaction frequency, and reducing redundant search results generated by the large model's expansion of the requirement information. Therefore, performing the search task based on the intent expressed in the clarification prompt text can more accurately retrieve the information the target user needs, improving the convenience and efficiency of information retrieval.

[0047] In some embodiments, the semantically ambiguous intent of a demand can also be related to missing demand content in the demand information. For example, if the input demand information is "Calculate the amount of pension payments," the first model, through semantic understanding of the demand information, can determine that the demand information contains missing content related to the intent of pension payments, such as the years of contribution and the region of contribution, making it difficult to accurately output a response that satisfies the intent of pension payments. Therefore, based on the semantic understanding results and text generation capabilities of the first model, clarification prompt text can be generated. This clarification prompt text describes the missing demand content related to the demand intent in the demand information, and accurately prompts the target audience to provide feedback on the missing clarification demand content. This allows for a more accurate retrieval of information matching the target audience's actual needs while reducing the frequency of interaction with the target audience, and ensures that the response information more precisely meets the target audience's actual demand intent, thus improving the user experience.

[0048] In some embodiments, performing a retrieval task based on the clarification requirement and obtaining response information may include performing a retrieval based on search terms determined by the clarification requirement, obtaining retrieval results that match the requirement intent, and pushing the retrieval results as response information to the target object to meet the target object's actual needs.

[0049] For example, based on the clarification requirement being "AA Science and Technology Park", the search term is "Where is Park A?" and the clarification requirement being "AA Science and Technology Park" to obtain the address of AA Science and Technology Park. The address of AA Science and Technology Park is then pushed to the target audience to meet their actual needs.

[0050] In some embodiments, semantic understanding of the requirement information is performed using the first major model, including performing the following operations using the first major model: semantically understanding the business requirement intent of the requirement content in the requirement information based on the business requirement conditions corresponding to the specified business domain, and determining the fuzzy business requirement intent with semantic fuzziness attributes; performing a text generation task based on the fuzzy business requirement intent and the requirement information to obtain a clarification prompt text.

[0051] According to embodiments of this disclosure, business requirement conditions characterize attribute dependencies among multiple business requirement attributes, and fuzzy business requirement intent is matched with at least one business requirement attribute.

[0052] In some embodiments, the attribute dependency relationship between multiple business requirement attributes may include logical relationships such as "OR" or "AND", or the attribute dependency relationship may also include nested relationships between multiple business requirement attributes.

[0053] For example, for a price inquiry regarding product A, the business requirement conditions can include multiple attributes such as the online shopping discount rate related to product A, the purchase location, the discount rate Y1 corresponding to model A1 of product A at the purchase location, and the discount rate Y2 corresponding to model A2 of product A at the purchase location. Among these, the purchase location and the discount rate Y1 corresponding to the purchase location are different business requirement attributes with a nested relationship.

[0054] With the first model performing semantic understanding of the demand information and context, it can be determined that the fuzzy business demand intent includes the intent to purchase method, the intent to purchase location, and the intent to purchase product model. Therefore, the clarification prompt text can be output as: "You need to provide whether you purchased product A online or at a designated store. If purchasing at a designated store, you need to provide the product model of product A you wish to purchase, as models A1 and A2 each have different discount rates." Thus, the clarification prompt text can precisely ask questions about the business demand attributes corresponding to the fuzzy demand intent, thereby avoiding multiple rounds of interaction with the target audience through follow-up questions and reducing the complexity of the target audience's interactive behavior during the information acquisition process.

[0055] According to embodiments of this disclosure, business requirement conditions corresponding to a designated business domain can represent the information content needed to determine specified business information. For example, if the designated business domain is pension accounting, the business requirement conditions for determining the pension payment amount can include multiple business requirement attributes such as "years of contribution," "contribution amount," and "city of pension payment." When the requirement content corresponding to each of the multiple business attributes is obtained, the pension payment amount that accurately meets the actual needs of the target object can be determined. However, if the requirement content corresponding to any business requirement attribute is semantically unclear, missing, or the attribute value of the requirement content does not match the attribute value corresponding to the business requirement attribute, it is difficult to accurately calculate the pension payment amount, and the actual needs of the target object that input the requirement information cannot be met.

[0056] For example, if the requirement content corresponding to "contribution period" in the requirement information is "80 years", it does not match the attribute threshold "less than 50 years" corresponding to the business requirement attribute "contribution period". Therefore, it is determined that the requirement intent corresponding to the business requirement attribute "contribution period" is an ambiguous requirement intent.

[0057] In some embodiments, the first major model determines the specified business domain matching the input requirement information by performing semantic understanding based on object-related information such as the input requirement information and the dialogue context with the target object. It then utilizes the first major model to perform semantic understanding on the requirement content corresponding to the business requirement conditions in the requirement information based on multiple business requirement attributes with dependencies, determining the requirement content corresponding to each business requirement attribute. If it detects that the requirement content corresponding to the target business requirement attribute is semantically unclear, the requirement content is missing, or the attribute of the requirement content does not match the attribute corresponding to the business requirement attribute, the business requirement intent corresponding to the target business requirement attribute is determined to be an ambiguous business requirement intent.

[0058] According to embodiments of this disclosure, clarification prompt text is used to prompt the target object to provide clarification content corresponding to the vague business requirement intent. When the first model understands the vague requirement intent corresponding to the requirement information, the business attribute type of the vague business requirement intent corresponding to the target business requirement intent can be determined based on the target business requirement attribute corresponding to the vague requirement intent. This allows the target object to be prompted to accurately fill the slot corresponding to the vague requirement intent with clarification content through questions in the clarification prompt text. This enables the clarification request from the target object to more accurately meet the requirements of the target business requirement attribute, avoiding multiple rounds of interaction to input clarification content through follow-up questions. Furthermore, it enables the retrieval task to be performed based on the clarification content and the complete business requirement conditions represented by the requirement information, obtaining a response that accurately meets the target object's actual needs. This reduces the complexity of the target object's interaction, reduces the amount of redundant information in the retrieval task, improves the user experience, and reduces redundant computational overhead caused by excessive recall of irrelevant information.

[0059] In some embodiments, the first model performs a thinking task based on the demand information, obtains the thinking result of the thinking task, and performs a text generation task based on the thinking result to obtain clarification prompt text.

[0060] For example, a thinking task can include multiple tasks with dependencies, and these dependencies can correspond to attribute dependencies between multiple business requirement attributes. The thinking result can indicate whether there are semantically ambiguous or missing content in the requirement content related to the requirement information, or other semantically unclear intents. Based on the thinking result, fuzzy business requirement attributes are determined, and the corresponding defect types are described. Therefore, the first major model performs a text generation task based on the fuzzy business requirement intent indicated by the thinking result and the corresponding defect types of the fuzzy business requirement attributes. This ensures that the clarification prompt text clearly represents the business requirement attributes and specific defect types corresponding to the fuzzy business requirement intent, accurately prompting the target object to input clarification content for the fuzzy business requirement intent. This reduces the frequency of interaction operations where the target object accurately inputs requirement content that meets the business requirement conditions, and avoids redundant computational overhead and redundant storage resource consumption caused by full information retrieval based on multiple business requirement attributes, thus improving computational efficiency and user experience.

[0061] According to embodiments of this disclosure, performing a retrieval task based on the clarification request content input by the target object to obtain response information matching the request intent may include: using a first model to perform semantic fusion on the clarification request content and the request information to obtain target request information; performing a retrieval task based on the target request information to obtain target retrieval results matching the clarification request content; and determining response information based on the target retrieval results.

[0062] In some embodiments, semantic fusion of the clarification requirement content and requirement information using the first major model may include semantically fusing the clarification requirement content and the requirement content in the requirement information based on the contextual content related to the requirement information, thereby obtaining target retrieval execution parameters for performing the retrieval task. These target retrieval execution parameters are then sent to a retrieval agent with retrieval capabilities, and the designated retrieval agent invokes retrieval tools based on the target retrieval execution parameters to perform the retrieval task, resulting in the target retrieval results.

[0063] According to embodiments of this disclosure, target retrieval results can be pushed as response information to the target object to meet the target object's actual information retrieval needs. Alternatively, a first model can be used to perform semantic fusion on the information content in the target retrieval results based on the demand information and clarifying the demand content to perform a generative task and obtain response information. Thus, response information based on natural language expression can be output through dialogue with the target object to improve the target object's information acquisition experience.

[0064] According to embodiments of this disclosure, target demand information represents multiple target demand intentions with intent dependencies, and the target retrieval results match all multiple target demand intentions.

[0065] For example, multiple target requirement intents can correspond to multiple business requirement attributes with attribute dependencies in the business requirement conditions. The target requirement information also includes the requirement content corresponding to multiple target requirement intents. Therefore, a retrieval task can be performed based on the requirement content corresponding to multiple target requirement intents to obtain target retrieval results corresponding to multiple target requirement intents. This allows for accurate retrieval of target retrieval results that meet the needs of the target object, provided that the target object accurately responds to the clarification prompt text with clarification requirement content that matches the target requirement intent.

[0066] In some embodiments, multiple target search results correspond to multiple business requirement attributes indicated by business requirement conditions. Therefore, by leveraging a large model to semantically fuse multiple target search results based on the requirement intent indicated by the requirement information and the business requirement conditions, response information that satisfies the needs of the target object can be obtained.

[0067] It should be noted that the response information may include data of any modality such as text, images, and charts. The embodiments of this disclosure do not limit the specific data modality type of the response information.

[0068] In some embodiments, semantic fusion of clarifying requirement content and requirement information using the first major model may include: based on cue words representing the object preference attributes of the target object, using the first major model to understand the target object's requirement intent based on the clarifying requirement content and requirement information, and outputting the target requirement information.

[0069] For example, the target object's object preference attributes can be obtained under the condition of obtaining authorization from the target object. The first main model can deeply understand the target object's actual need intent based on the object preferences represented by prompt words, and output target need information that can represent multiple complete target need intents of the target object, as well as the target need content corresponding to multiple target need intents. Therefore, based on the target need information, precise retrieval can be performed according to the target object's needs, obtaining multiple target retrieval results that accurately meet the target object's information acquisition needs. Thus, the first main model can be used to generate response information by semantically fusing multiple target retrieval results, thereby reducing the frequency of retrieval interactions for the target object and reducing the computational and storage resources required for executing retrieval tasks and generating response information by the main model.

[0070] Figure 3 The diagram illustrates an application scenario of the interaction method based on a large model provided according to an embodiment of the present disclosure.

[0071] like Figure 3 As shown, the display interface 300 can display the conversation content between the target object and the large model. The target object asks questions to the large model by inputting requirement information 311 in order to obtain relevant information about studying in country A.

[0072] The first major model uses semantic understanding of the business intent expressed in the demand information based on business demand conditions related to the study abroad business field. This determines that fuzzy business demand intents can be matched with business demand attributes such as "study abroad budget," "study abroad major," "language test scores," and "study abroad duration." Among these, "study abroad duration," "language test scores," and "study abroad budget" have a nested attribute relationship. Therefore, a text generation task can be performed based on multiple fuzzy business demand intents corresponding to multiple business demand attributes, as well as the attribute dependencies between these attributes, resulting in clarification prompt text 321. Clarification prompt text 321 uses natural language to provide prompts regarding the fuzzy demand intents corresponding to "study abroad budget," "study abroad major," "language test scores," and "study abroad duration."

[0073] The target user inputs clarification request content 312 by browsing the clarification prompt text 321. The first major model can then perform semantic fusion based on the clarification request content 312 and the request content in the request information to obtain the target request information: "2025 postgraduate admission information for Country A, requiring postgraduate students in architecture. If postgraduate admission information in architecture is found, retrieve the following requirements for postgraduate students in architecture at that school: tuition fees, language proficiency requirements, postgraduate study duration, and student living expenses. If the search result for any requirement does not match the clarification request content, the search task will stop." This yields multiple target search results corresponding to the target request information. The first major model, through semantic fusion of multiple target search results, can obtain the response information 322 pushed to the target user. This allows the target user to input the clarification request content only once, retrieving accurate and complete target search results for their intention to study postgraduate programs in Country A, and avoiding searching for admission information from multiple schools in Country A, thus reducing the amount of resource data retrieved by the search task and lowering the computational and storage resource consumption required for the search task.

[0074] Figure 4 A flowchart illustrating a method for training a large model according to an embodiment of this disclosure is shown schematically.

[0075] like Figure 4 As shown, the method for training a large model includes operations S410~S430.

[0076] The S410 is used to receive training data.

[0077] Operate S420 to perform semantic understanding of sample demand information using the first large model to be trained, and output sample clarification prompt text.

[0078] Operate S430 to train the first large model based on the label clarification prompt text and the sample clarification prompt text, and obtain the trained first large model.

[0079] According to embodiments of this disclosure, the training data includes sample demand information and label clarification prompt text. The sample demand information represents the semantic ambiguity of the demand intent of the sample object. The label clarification prompt text is used to prompt the sample object to provide clarification of the demand intent corresponding to the demand intent, so as to perform a retrieval task based on the clarification of the demand content to obtain retrieval results that match the sample object.

[0080] According to embodiments of this disclosure, the label clarification prompt text describes the demand intent with semantic ambiguity in the demand information, thereby enabling the sample object to input the corresponding clarification demand content based on the demand intent with semantic ambiguity described in the label clarification prompt text, so as to achieve accurate retrieval based on the clarification demand content and the demand content in the demand information, and obtain accurate retrieval results.

[0081] According to embodiments of this disclosure, by acquiring sample demand information and clarification prompt text, the sample demand information represents the semantically ambiguous query information of the sample object's demand intent, and the clarification prompt text represents prompting the sample object to input clarification content regarding the demand intent with semantically ambiguous attributes. Thus, by using a first-level model to perform semantic understanding of the sample demand information and output sample clarification prompt text, the first-level model learns the ability to recognize semantically ambiguous demand intents in the sample demand information. By training the first-level model with clarification prompt text and sample clarification prompt text, the first-level model can accurately identify semantically ambiguous demand intents and accurately generate prompt information for such intents. This allows the sample object to accurately input demand content based on the clarification prompt text to clarify the target object's actual demand intent, thereby reducing the frequency of interactions required for intent clarification. Furthermore, by accurately obtaining clear demand content, redundant retrieval data is reduced, lowering the computational overhead required by the search engine and subsequent response generation.

[0082] It should be noted that the technical terms involved in the embodiments of this disclosure, including but not limited to sample requirement information and label clarification prompt text, have the same or similar meanings as the technical terms involved in the interaction method based on a large model provided in the embodiments of this disclosure, including but not limited to requirement information and clarification prompt text. The embodiments of this disclosure will not be repeated here.

[0083] The first trained large model determined by the method for training a large model according to embodiments of this disclosure can be applied to the large model-based interaction method according to embodiments of this disclosure. For example, the trained first large model can be used to perform semantic understanding of demand information and output clarifying prompt text. Embodiments of this disclosure will not be described in detail here.

[0084] In some embodiments, the training data is determined based on the following operation: semantically perturbing at least one sample target demand content in the sample target demand information using a second large model to obtain the sample demand information.

[0085] According to embodiments of this disclosure, multiple sample target requirement contents in the sample target requirement information represent multiple target requirement intentions with intent dependencies, and the sample target requirement contents satisfy a preset semantic clarity condition.

[0086] For example, the sample target demand information could be text related to studying abroad in country A: "Study abroad in country A in 2025, majoring in architecture, monthly budget of 4xxxx yuan, language test score of A, expected completion of studies within 3 years." Multiple target demand contents in the sample target demand information are matched with target demand intentions corresponding to multiple business demand attributes such as "year of study abroad," "country of study abroad," "budget of study abroad," "major of study abroad," "language test score," and "duration of study abroad." Among these, "duration of study abroad," "language test score," and "budget of study abroad" have a nested attribute relationship. The second major model semantically perturbs at least one sample target demand content to make the perturbated sample demand content represent ambiguous semantics or semantically missing meanings. This allows the sample demand content in the obtained sample demand information to express semantically ambiguous attributes, facilitating the training of the first model to identify sample demand content with semantically ambiguous attributes and learn to generate clarifying prompt text to indicate the ambiguous intent attributes corresponding to the sample demand content. This clarification prompt text reduces the frequency of interaction for the target object to input clarifying demand content, allowing for precise retrieval based on the target object's needs.

[0087] In some embodiments, the second major model determines training data by performing the following operations: performing intent understanding on the sample target demand information to determine the sample target demand content corresponding to the target demand intent; rewriting the sample target demand information based on the fuzzy demand intent determined from the target demand intent to obtain sample demand information; and performing semantic understanding based on the fuzzy demand intent and sample demand information to obtain clarification prompt text.

[0088] According to embodiments of this disclosure, the sample demand information representation has semantic ambiguity attributes for ambiguous label demand intent.

[0089] In some embodiments, the second major model can rewrite the sample target demand content in the sample target demand information corresponding to the fuzzy demand intent by deleting, replacing, or other operations using the fuzzy demand intent determined from the target demand intent, so that the sample demand content in the sample demand information corresponding to the fuzzy demand intent has an fuzzy semantic attribute.

[0090] In some embodiments, the second model, upon determining the ambiguous demand intent, can perform a text generation task on the sample demand content within the sample demand information. This generates a semantic ambiguity defect type describing the sample demand content and a label clarification prompt text for questioning the ambiguous demand intent. This allows the target object to accurately respond with clarification of the demand content based on the label clarification prompt text. Thus, the second model can generate sample demand information and corresponding label clarification prompt text under the condition of accurately understanding the ambiguous demand intent, thereby improving the efficiency and accuracy of constructing training data and enhancing the training efficiency of the first model to accurately output clarification prompt text.

[0091] In some embodiments, the sample target demand information is determined based on the following operations: acquiring multiple historical interaction data; and using a second major model to perform semantic fusion on the multiple historical interaction data to obtain the sample target demand information.

[0092] According to embodiments of this disclosure, a sample object performs information retrieval through multiple historical interaction data. Historical interaction data may include historical query information input by the sample object and historical search results determined based on the historical query information. For example, historical search results may represent resource information such as page resources from which the sample object performed interactive operations such as browsing, copying, and replying among the recalled search results.

[0093] For example, within a specified historical period, the sample object performs multiple searches by inputting multiple historical query information, obtaining historical search results corresponding to each query. These multiple historical query information can be query terms with semantic hierarchical relationships. Therefore, the query process employed by the sample object within a specified historical period to obtain accurate historical search results can be represented based on these multiple historical query information.

[0094] The second major model performs semantic understanding on multiple historical interaction data representing the query process of a sample object in a specified historical period to accurately obtain historical search results. This enables the output sample target demand information to describe the query information needed by the sample object to obtain accurate historical search results. Thus, the sample target demand information clearly represents the demand intent, and historical search results can be retrieved by executing the retrieval task based on the sample target demand information. Therefore, by rewriting the sample target demand information, the ambiguous demand intent and corresponding tag clarification prompt text can be determined, thereby improving the training efficiency of the first major model.

[0095] In some embodiments, multiple historical interaction data are related to a designated business domain. The designated business domain could be, for example, medical assistance advice, study abroad consulting, or legal information consulting. The embodiments of this disclosure do not limit the specific business type of the designated business domain.

[0096] In some embodiments, semantic fusion of multiple historical interaction data using a second major model may include: performing intent understanding on multiple historical interaction data based on business requirement conditions corresponding to a specified business domain using a second major model.

[0097] According to embodiments of this disclosure, business requirement conditions characterize the attribute dependencies between multiple business requirement attributes, and the intent dependencies between multiple target requirement intents with intent dependencies are matched with the attribute dependencies. Historical query information in multiple historical interaction information can be matched with the target requirement intents corresponding to multiple business requirement attributes, and historical search results retrieved based on historical query information can satisfy the actual requirement intents of the sample object. Therefore, the second major model can perform semantic understanding on multiple historical interaction data based on the attribute dependencies between multiple business requirement attributes to determine the intent dependencies between multiple target requirement intents that performed the retrieval process in the historical period of the sample object. This allows multiple sample target requirement contents in the sample target requirement information to correspond to the intent dependencies between multiple corresponding target requirement intents, thereby improving the accuracy of the sample target requirement information in accurately retrieving historical search results related to a specified business domain.

[0098] In some embodiments, semantic fusion of multiple historical interaction data using a second model enables the second model to semantically understand multiple historical demand intentions represented by multiple historical interaction data, as well as the temporal and semantic relationships and other intent dependencies among these historical demand intentions. This allows for semantic understanding of multiple related historical query information and historical search results, enabling multiple sample target demand contents in the sample target demand information to represent multiple target demand intentions generated by the sample object in historical periods to obtain historical search results for a specified business domain. This achieves the accurate generation of multiple target demand contents representing the need for precise retrieval in a specified business domain based on the second model, thereby improving the accuracy and completeness of the representation of multiple business demand attributes in the specified business domain by the sample target demand information. Furthermore, it improves the accuracy and logical rationality of subsequently generated ambiguous label demand intentions and label clarification prompt text, enhancing the training effect of the training data on the first model, enabling the first model to output accurate clarification prompt text for demand information with semantically ambiguous intentions of the target object.

[0099] In one embodiment, multiple historical interaction data may include multiple historical query information, such as "approximate cost of treatment for disease A", "which Class A drugs will be included in medical insurance in 2025", "are Class A drugs for diseases A reimbursed", "are A1 drugs in Class A reimbursed by medical insurance in city A", "how much of a 100,000 yuan expense for a disease A can be reimbursed", and "what should be done if a Class C change occurs in 30cm of the esophagus".

[0100] The second major model can understand the intent of multiple historical queries based on business needs related to the fields of medical assistance advice and regulations, and obtain sample target demand information: "Combined with the medical insurance policy of a specified year, determine whether Class A drugs (such as A1 drugs) related to Class A diseases are included in medical insurance and the reimbursement situation in a certain city, estimate the reimbursable ratio of costs, and explain the significance of the treatment method and the basis for the judgment of adopting the corresponding treatment method."

[0101] By processing the sample target demand information using the second major model, semantic understanding of the sample target demand information can be achieved, and the content of the sample target demand can be perturbed to obtain structured output information. The output information may include:

[0102] The disturbed sample demand information is as follows: "Based on recent medical insurance policies, explain whether Class A drugs related to a certain type of disease are included in medical insurance and the reimbursement situation in a certain city, estimate the reimbursement ratio, and explain the significance of the treatment method and the basis for determining the corresponding treatment method."

[0103] Designated business areas: medical assistance advice and regulatory areas.

[0104] Vague tagging of demand intent: time attribute intent, region attribute intent.

[0105] Example 1 of label clarification prompt text: Please confirm the region and time range you are interested in, such as which cities the treatment will be conducted in and which year the drug reimbursement policy is in.

[0106] In one embodiment, the sample requirement information may be: "Based on the medical insurance policy of a specified year, determine whether the drug is included in medical insurance and the reimbursement situation in a certain city, estimate the reimbursable ratio of the cost, and explain the significance of the treatment method and the basis for the judgment of adopting the corresponding treatment method."

[0107] Example 2 of label clarification prompt text: Are you focusing on a specific type A disease and related drugs? If not, please enter an example of the disease name and drug name.

[0108] According to embodiments of this disclosure, obtaining multiple historical interaction data may further include: determining intermediate historical interaction data that semantically matches the business keywords from multiple initial historical interaction data received, based on business keywords related to a specified business domain; and determining multiple historical interaction data from multiple intermediate interaction data based on the semantic difference between the multiple intermediate interaction data.

[0109] In some embodiments, intermediate historical interaction data can be represented as historical query information semantically similar to business keywords. By calculating the semantic distance between multiple intermediate historical interaction data, the semantic difference between them can be obtained. Thus, intermediate historical interaction data whose semantic difference meets a preset difference threshold can be identified as historical interaction data to be processed by the second-largest model. This ensures that the generated sample target demand information exhibits diverse query characteristics specific to the designated business domain, avoiding excessive duplication in training data due to determining sample target demand information based on repetitive query information, thereby improving the quality of the training data.

[0110] In some embodiments, multiple initial historical interaction data can be scored based on a scoring mechanism corresponding to a specified business domain to determine the initial scores of the initial historical interaction data across multiple scoring dimensions. The multiple initial scores are then weighted and fused to obtain the target score for the initial historical interaction data. Based on the target score, the multiple initial historical interaction data are then filtered to obtain the historical interaction data.

[0111] For example, the initial historical interaction data can be scored based on formula (1).

[0112] The target score is calculated as follows: w1 Professional Confidence Score + w2 Semantic Ambiguity Score + w3 Time Efficacy Score + w4 Entity Quantity Score + w5 Semantic Distance Score – w6 Compliance Risk Score. Based on the target scores of multiple initial historical interaction data, the top N initial interaction historical data can be selected as historical interaction data.

[0113] According to embodiments of this disclosure, multiple initial historical interaction data are scored and filtered according to a specified business domain. A second model is then used to perform target demand intent recognition and semantic fusion on the query process represented by the filtered historical interaction data to obtain sample target demand information. This allows the sample target demand information to accurately represent the target demand content corresponding to multiple target demand intents. Therefore, by perturbing the sample target demand content in the sample target demand information, sample demand information with semantically ambiguous intent attributes and corresponding tag clarification prompt text can be generated. This allows the tag clarification prompt text to more accurately describe the demand intent with semantically ambiguous attributes in the sample demand information, achieving the effect of accurately prompting the target object to provide feedback on the demand content. Thus, the first model can be trained based on the difference between the sample clarification prompt text and the tag clarification prompt text obtained by processing the sample demand information output. This enables the first model to more accurately identify and understand ambiguous demand intents with semantically ambiguous intents, and by outputting prompt content that accurately describes the demand intent that the target object needs to clarify, the frequency of interaction operations required for the target object to input clarification content is reduced, thus reducing the complexity of the target object's interaction operations.

[0114] In some embodiments, the training data also includes structured labeled fuzzy demand intentions, where the sample demand information representation has semantically fuzzy attributes for the labeled fuzzy demand intentions. For example, the sample demand content corresponding to the labeled fuzzy demand intention may be missing in the sample demand information, or the sample demand content corresponding to the labeled fuzzy demand intention may have semantic ambiguity, or the semantic attributes may not match the attribute thresholds corresponding to the business demand attributes.

[0115] According to embodiments of this disclosure, training a first master model based on label clarification prompt text and sample clarification prompt text may include: determining first loss information based on the difference between the label clarification prompt text and the sample clarification prompt text; determining second loss information based on the difference between the label ambiguity demand intent and the sample thinking result; and training the first master model based on the first loss information and the second loss information.

[0116] According to embodiments of this disclosure, the sample thinking result is determined by the first model performing a thinking task based on the sample requirement information. The sample thinking result represents the identification result of the requirement intent with semantically ambiguous attributes in the sample requirement information. For example, the sample thinking result can represent the sample fuzzy requirement intent obtained by the first model through fuzzy requirement intent identification of the sample requirement information based on business requirement conditions. When the sample requirement information has requirement content corresponding to multiple business requirement attributes, all of which have semantically ambiguous intent attributes, the first model can perform the thinking task based on the attribute dependency relationship between multiple business requirement attributes. The resulting thinking result can represent multiple sample fuzzy requirement intents with intent dependency relationships, thereby realizing the mining of sample fuzzy requirement intents with semantically ambiguous attributes in the sample requirement information according to business requirement conditions.

[0117] According to embodiments of this disclosure, a second loss information determined based on the difference between the label's ambiguous demand intent and the sample's thinking result can be used to represent the first model's ability to identify ambiguous demand intents in sample demand information. Thus, the first model can be trained based on the target loss information determined by the first and second loss information. This allows the trained first model to accurately describe the defects in the demand content corresponding to the ambiguous demand intent while precisely identifying it. Consequently, the clarification prompt text generated by the trained first model can accurately prompt the target object to input the corresponding clarification demand content, thereby improving the accuracy of the response information and reducing the frequency of interaction for the target object during the interaction process.

[0118] In some embodiments, a second loss information is obtained by identifying the difference between the fuzzy demand intent and the sample thinking results. This second loss information is then used to adjust the model parameters of the first master model, thereby refining its thinking process strategy. This allows the intermediate first master model, trained with the second loss information, to accurately determine the fuzzy business demand intent that matches the business domain's demand intent by combining the thinking process with the corresponding business demand conditions. Conversely, using the first loss information to train the intermediate first master model enables it to output clarification prompt text more accurately, improving the prediction accuracy and matching degree of the clarification prompt text with the business domain's demand conditions, while reducing interaction frequency and redundant recall information.

[0119] Figure 5 A schematic diagram illustrating a method for training a large model according to an embodiment of the present disclosure is shown.

[0120] like Figure 5As shown, the first main model can process sample demand information and business demand conditions in a specified domain to identify ambiguous demand intentions within the sample demand information. Taking these ambiguous demand intentions as the result, a loss function is used to process the ambiguous demand intentions of the samples and the ambiguous demand intentions of the labels, yielding second loss information. The first main model is then trained using this second loss information, resulting in an intermediate first main model. This intermediate first model processes the sample demand information and business demand conditions to obtain sample clarification prompt text. A loss function is then used to process this sample clarification prompt text and the label clarification prompt text to obtain first loss information. This first loss information is then used to train the intermediate first main model, resulting in the trained first main model.

[0121] In some embodiments, the training data further includes labeled demand content corresponding to the fuzzy demand intent. The labeled demand content can be determined by a second model through information extraction of sample target demand information based on the fuzzy demand intent. For example, the fuzzy demand intent could be a location-fuzzy intent, and the labeled demand content could be: in city A.

[0122] In some embodiments, determining the second loss information based on the difference between the label fuzzy demand intent and the sample thinking result may include: determining intent recognition loss information based on the difference between the label fuzzy demand intent and the sample fuzzy intent recognition result in the sample thinking result; determining content recognition loss information based on the semantic difference between the label demand content and the sample clarification keywords in the sample thinking result; and determining the second loss information based on the intent recognition loss information and the content recognition loss information.

[0123] According to embodiments of this disclosure, sample clarification keywords are used to prompt the sample object to provide clarification of the required content corresponding to the ambiguous tag request intent. In some embodiments, sample clarification keywords can represent keywords related to the sample target request content. For example, the sample clarification keyword can be "which city," and the tag request content can be "City A." Thus, based on the semantic difference between the tag request content and the sample clarification keywords in the sample thinking results, content recognition loss can be obtained. This allows the first model to learn to accurately prompt the target object to focus on the semantically related request content of the clarification keywords in the output clarification prompt text, avoiding situations where the semantic difference between the attribute value indicated by the business requirement attribute is too large, making it difficult to accurately retrieve the target search results for the requested content provided by the target object. For example, the first model can output clarification prompt text containing the clarification keyword "city" based on content recognition loss information, avoiding situations where the output clarification prompt text only contains keywords similar to location requests such as "location" or "region," leading to requests from the target object containing requests such as "North District A" or "Community A," where the city location is difficult to determine.

[0124] According to embodiments of this disclosure, a second loss information can be determined based on the sum of intent recognition loss information and content recognition loss information, so as to supervise the first model to output more accurate clarification prompt text, and enable the trained first model to reduce the frequency of interaction of the target object to input clarification request content during the interaction with the target object, thereby improving the user experience.

[0125] The method for training a large-scale model provided in this disclosure can enable the clarification prompt text output by the trained large-scale model to address issues such as "where clarification is needed, which key points need clarification, and whether the clarification covers key information." Furthermore, by designing loss information, the accuracy of the clarification prompt text output by the first model can be precisely improved. In addition, by determining the target demand information of the sample objects from real historical interaction data over historical periods, the generation of ambiguous demand intent and clarification prompt text can be precisely controlled, thereby accurately adjusting the accuracy of the first large-scale model in generating ambiguous demand intent and clarification prompt text. Furthermore, by testing the trained first large-scale model, the performance of the trained large-scale model can be tested by determining indicators such as the ambiguous demand intent missing rate, the redundancy of ambiguous demand intent, and the retrieval accuracy of the clarification prompt text output by the trained large-scale model. This achieves deep linkage between training and evaluation, reducing the waste of computational resources during interaction with the target object. This provides an operational basis for products to choose a "ask fewer questions but ask the right questions" strategy, reducing infinite loop jumps and resource waste, reducing redundant retrieval of web page resources, and reducing the token processing volume and network communication costs of the large-scale model. In addition, for specific business requirements in a particular domain, the first model can be adjusted and optimized by replacing the corresponding business requirements in each domain or by using a business keyword list. This allows the first model to adapt to various types of business domains and enhance its ability to interact with target objects.

[0126] Figure 6 A block diagram of a large-model-based interactive device according to an embodiment of the present disclosure is shown schematically.

[0127] like Figure 6 As shown, the interactive device 600 based on a large model includes: a first receiving module 610, a clarifying prompt text obtaining module 620, a reply information obtaining module 630, and a push module 640.

[0128] The first receiving module 610 is used to receive the requirement information of the target object.

[0129] The clarification prompt text acquisition module 620 is used to perform semantic understanding of the demand information using the first major model and output clarification prompt text. The clarification prompt text is used to describe the demand intent with semantic ambiguity in the demand information and prompt the target object to provide feedback on the clarification demand content corresponding to the demand intent.

[0130] The response information acquisition module 630 is used to perform a retrieval task based on the clarification request content input by the target object, and obtain response information that matches the request intent.

[0131] The push module 640 is used to push reply information to the target object.

[0132] According to embodiments of this disclosure, the clarification prompt text acquisition module is configured to perform the following operations using a first major model: based on business requirement conditions corresponding to a specified business domain, semantically understand the business requirement intent of the requirement content in the requirement information to determine a fuzzy business requirement intent with semantic fuzziness attributes, wherein the business requirement conditions characterize the attribute dependency relationship between multiple business requirement attributes, and the fuzzy business requirement intent matches at least one business requirement attribute; perform a text generation task based on the fuzzy business requirement intent and the requirement information to obtain clarification prompt text, which is used to prompt the target object to provide feedback on the clarification requirement content corresponding to the fuzzy business requirement intent.

[0133] According to embodiments of this disclosure, the response information acquisition module includes: a target demand information acquisition unit, a retrieval task execution unit, and a response information determination unit.

[0134] The target requirement information acquisition unit is used to perform semantic fusion of the clarified requirement content and requirement information using the first major model to obtain the target requirement information.

[0135] The retrieval task execution unit is used to perform retrieval tasks based on target demand information and obtain target retrieval results that match the content of the clarification demand.

[0136] The response information determination unit is used to determine the response information based on the target search results.

[0137] According to embodiments of this disclosure, the target requirement information acquisition unit includes a first output subunit.

[0138] The first output subunit is used to understand the target object's intention based on the prompt words representing the target object's object preference attributes, and to use the first main model to understand the target object's intention based on the clarified requirement content and requirement information, and output the target requirement information.

[0139] According to embodiments of this disclosure, target demand information represents multiple target demand intentions with intent dependencies, and the target retrieval results match all multiple target demand intentions.

[0140] Figure 7 A block diagram of an apparatus for training a large model according to an embodiment of the present disclosure is shown schematically.

[0141] like Figure 7 As shown, the apparatus 700 for training a large model includes: a second receiving module 710, a sample clarification prompt text acquisition module 720, and a training module 730.

[0142] The second receiving module 710 is used to receive training data, which includes sample demand information and label clarification prompt text. The sample demand information represents the demand intention of the sample object and has semantic ambiguity. The label clarification prompt text is used to prompt the sample object to provide clarification of the demand intention corresponding to the demand intention, so as to perform a retrieval task based on the clarification of the demand content to obtain retrieval results that match the sample object.

[0143] The sample clarification prompt text acquisition module 720 is used to perform semantic understanding of sample demand information using the first large model to be trained, and output sample clarification prompt text.

[0144] Training module 730 is used to train the first large model based on the label clarification prompt text and the sample clarification prompt text, thus obtaining the trained first large model.

[0145] According to embodiments of this disclosure, the training data further includes structured labeled fuzzy demand intent, and the sample demand information representation has semantic fuzzy attributes for labeled fuzzy demand intent; wherein, the training module includes: a first loss information determination unit, a second loss information determination unit, and a training unit.

[0146] The first loss information determination unit is used to determine the first loss information based on the difference between the label clarification prompt text and the sample clarification prompt text.

[0147] The second loss information determination unit is used to determine the second loss information based on the difference between the label fuzzy demand intention and the sample thinking result. The sample thinking result is determined by the first model performing a thinking task based on the sample demand information. The sample thinking result represents the recognition result of demand intention with semantic fuzzy attributes in the sample demand information.

[0148] The training unit is used to train the first large model based on the first loss information and the second loss information.

[0149] According to embodiments of this disclosure, the training data also includes label demand content corresponding to the label ambiguity demand intent; wherein, the second loss information determination unit includes: a first determination subunit, a second determination subunit, and a third determination subunit.

[0150] The first determining subunit is used to determine the intent recognition loss information based on the difference between the label fuzzy demand intent and the sample fuzzy intent recognition result in the sample thinking result;

[0151] The second determining subunit is used to determine content recognition loss information based on the semantic differences between the label requirement content and the sample clarification keywords in the sample thinking results. The sample clarification keywords are used to prompt the sample object to provide clarification requirement content corresponding to the label's ambiguous requirement intent; and

[0152] The third determining subunit is used to determine the second loss information based on the intent recognition loss information and the content recognition loss information.

[0153] According to an embodiment of this disclosure, the training data is determined based on the following operation: semantically perturbing at least one sample target demand content in the sample target demand information using a second large model to obtain sample demand information, wherein multiple sample target demand contents in the sample target demand information represent multiple target demand intentions with intent dependency relationships, and the sample target demand contents satisfy a preset semantic clarity condition.

[0154] According to embodiments of this disclosure, the sample target demand information is determined based on the following operations: acquiring multiple historical interaction data, and the sample object performing information retrieval through the multiple historical interaction data; using a second major model to perform semantic fusion on the multiple historical interaction data to obtain the sample target demand information.

[0155] According to embodiments of this disclosure, multiple historical interaction data are related to a specified business domain; semantic fusion of the multiple historical interaction data is performed using a second major model, including: based on business requirement conditions corresponding to the specified business domain, intent understanding is performed on the multiple historical interaction data using the second major model, wherein the business requirement conditions represent the attribute dependency relationship between multiple business requirement attributes, and the intent dependency relationship between multiple target requirement intents with intent dependency relationship is matched with the attribute dependency relationship.

[0156] According to embodiments of this disclosure, obtaining multiple historical interaction data includes: determining intermediate historical interaction data semantically matched with business keywords from multiple received initial historical interaction data based on business keywords related to a specified business domain; and determining multiple historical interaction data from multiple intermediate interaction data based on the semantic difference between the multiple intermediate interaction data.

[0157] According to embodiments of this disclosure, the second model determines training data by performing the following operations: performing intent understanding on sample target demand information to determine sample target demand content corresponding to the target demand intent; rewriting the sample target demand information based on the fuzzy demand intent determined from the target demand intent to obtain sample demand information, wherein the sample demand information represents a semantically fuzzy attribute for the fuzzy demand intent; and performing semantic understanding based on the fuzzy demand intent and the sample demand information to obtain a clarification prompt text.

[0158] Figure 8 A schematic block diagram of an artificial intelligence agent according to an embodiment of the present disclosure is shown.

[0159] In embodiments of this disclosure, such as Figure 8 As shown, the AI ​​agent 800 may include an input module 810, a processing module 820, and an output module 830.

[0160] Input module 810 is used to receive input information;

[0161] The processing module 820 is used to determine the target task based on the input information received by the input module, determine the large model based on the target task, and obtain output information by calling the large language model to execute the interaction method based on the large model provided in the embodiments of this disclosure, or by calling the large model to execute the training method of the large model provided in the embodiments of this disclosure.

[0162] Output module 830 is used to output the output information obtained by the processing module.

[0163] According to embodiments of this disclosure, the input module 810 is responsible for receiving or sensing information such as queries, requests, instructions, signals, or data from the outside world (e.g., users or the external environment), and converting it into a format that the AI ​​agent 800 can understand and process. The input module 810 is the primary link for the AI ​​agent 800 to interact with the outside world, enabling the AI ​​agent 800 to efficiently and accurately obtain necessary "sensory" information from the outside world and respond to this information.

[0164] In the example, input module 810 can input the requirement information described above, sample requirement information, etc.

[0165] In the example, processing module 820 is the core support for the AI ​​agent 800's ability to handle complex tasks. Processing module 820 can execute the interaction methods and training methods based on large models described above.

[0166] In the example, the performance of the processing module 820 is closely related to the large model on which the AI ​​agent 800 is based. To fully leverage the capabilities of the large model, the internal structure of the processing module 820 can be designed to be highly configurable and scalable to handle various types of tasks and requirements in real-world scenarios.

[0167] In the example, after the AI ​​agent 800 acquires the request voice, the processing module 820 can use the first big model to process the request information to obtain the clarification prompt text, and then pass the clarification prompt text to the output module 830.

[0168] Understandably, while large language models possess excellent language understanding and generation capabilities, like humans, their ability to solve tasks is limited without the aid of any tools. However, once the AI ​​agent 800 is given the ability to invoke tools, it can perform tasks such as using a calculator to complete mathematical calculations, using Python to perform data analysis, and using a search engine to create weather forecasts.

[0169] In the example, output module 830 can output the clarification prompt text described above or the first trained model.

[0170] The AI ​​agent 800 according to embodiments of this disclosure can simply and effectively improve the level of intelligence, as well as enhance flexibility and versatility.

[0171] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0172] 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.

[0173] 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.

[0174] 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.

[0175] Figure 9A schematic block diagram of an example electronic device is shown, illustrating an interaction method based on a large model and a method for training a large model that can be used to implement embodiments of the present disclosure. 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.

[0176] like Figure 9 As shown, the electronic device 900 includes a computing unit 901, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 902 or a computer program loaded from a storage unit 908 into a random access memory (RAM) 903. The RAM 903 may also store various programs and data required for the operation of the electronic device 900. The computing unit 901, ROM 902, and RAM 903 are interconnected via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.

[0177] Multiple components in electronic device 900 are connected to I / O interface 905, including: input unit 906, such as keyboard, mouse, etc.; output unit 907, such as various types of displays, speakers, etc.; storage unit 908, such as disk, optical disk, etc.; and communication unit 909, such as network card, modem, wireless transceiver, etc. Communication unit 909 allows electronic device 900 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0178] The computing unit 901 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 901 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 901 performs the various methods and processes described above, such as large model-based interaction methods and methods for training large models. For example, in some embodiments, the large model-based interaction methods and methods for training large models can be implemented as computer software programs tangibly contained in a machine-readable medium, such as storage unit 908. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 900 via ROM 902 and / or communication unit 909. When the computer program is loaded into RAM 903 and executed by the computing unit 901, one or more steps of the large model-based interaction methods and methods for training large models described above can be performed. Alternatively, in other embodiments, computing unit 901 may be configured by any other suitable means (e.g., by means of firmware) to perform large model-based interactive methods and methods for training large models.

[0179] 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), payload-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.

[0180] 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.

[0181] 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.

[0182] 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).

[0183] 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.

[0184] 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.

[0185] 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.

[0186] 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 interaction method based on a large model, comprising: Receive the target object's requirements information; The first major model is used to perform semantic understanding on the demand information and output clarification prompt text. The clarification prompt text is used to describe the demand intent with semantic ambiguity in the demand information and prompt the target object to provide clarification demand content corresponding to the demand intent. Based on the clarification request input by the target object, a retrieval task is performed to obtain response information matching the requested intent; and The response information is pushed to the target object.

2. The method according to claim 1, wherein, The semantic understanding of the demand information using the first major model includes performing the following operations using the first major model: Based on the business requirement conditions corresponding to the specified business domain, the semantic understanding of the business requirement intent of the requirement content in the requirement information is performed to determine the fuzzy business requirement intent with semantic fuzziness attributes, wherein the business requirement conditions represent the attribute dependency relationship between multiple business requirement attributes, and the fuzzy business requirement intent matches at least one of the business requirement attributes. Based on the fuzzy business requirement intent and the requirement information, a text generation task is performed to obtain the clarification prompt text. The clarification prompt text is used to prompt the target object to provide feedback on the clarification requirement content corresponding to the fuzzy business requirement intent.

3. The method according to claim 1, wherein, The step of performing a retrieval task based on the clarification request input by the target object to obtain response information matching the request intent includes: The first large model is used to perform semantic fusion on the clarified requirement content and the requirement information to obtain the target requirement information; Based on the target demand information, the retrieval task is executed to obtain target retrieval results that match the clarified demand content; and The response information is determined based on the target search results.

4. The method according to claim 3, wherein, The step of semantically fusing the clarification requirement content and the requirement information using the first large model includes: Based on the cue words representing the object preference attributes of the target object, the first large model is used to understand the target object's intention based on the clarification requirement content and the requirement information, and outputs the target requirement information.

5. The method according to claim 3, wherein, The target demand information represents multiple target demand intentions with intent dependencies, and the target retrieval result matches all of the multiple target demand intentions.

6. A method for training large models, comprising: The system receives training data, which includes sample demand information and label clarification prompt text. The sample demand information represents the demand intention of the sample object and has a semantic ambiguity attribute. The label clarification prompt text is used to prompt the sample object to provide clarification demand content corresponding to the demand intention, so as to perform a retrieval task based on the clarification demand content to obtain retrieval results that match the sample object. The first large model to be trained is used to perform semantic understanding of the sample demand information and output sample clarification prompt text. as well as The first large model is trained based on the label clarification prompt text and the sample clarification prompt text to obtain the trained first large model.

7. The method according to claim 6, wherein, The training data also includes structured labeled ambiguous demand intentions, and the sample demand information characterizes the label ambiguous demand intentions as having semantic ambiguity attributes. The step of training the first large model based on the label clarification prompt text and the sample clarification prompt text includes: The first loss information is determined based on the difference between the label clarification prompt text and the sample clarification prompt text; A second loss information is determined based on the difference between the fuzzy demand intent and the sample thinking result, wherein the sample thinking result is determined by the first large model performing a thinking task based on the sample demand information, and the sample thinking result represents the recognition result for demand intents with semantically fuzzy attributes in the sample demand information; and The first large model is trained based on the first loss information and the second loss information.

8. The method according to claim 7, wherein, The training data also includes label requirement content corresponding to the label ambiguity requirement intent; The step of determining the second loss information based on the difference between the label's ambiguous demand intent and the sample's thinking result includes: Based on the difference between the fuzzy demand intent of the label and the fuzzy intent recognition result of the sample thinking result, the intent recognition loss information is determined; Based on the semantic differences between the tag requirement content and the sample clarification keywords in the sample thinking results, content recognition loss information is determined. The sample clarification keywords are used to prompt the sample object to provide clarification content corresponding to the ambiguity of the tag requirement intent; and The second loss information is determined based on the intent recognition loss information and the content recognition loss information.

9. The method according to claim 6, wherein, The training data was determined based on the following operations: The second model is used to semantically perturb at least one sample target demand content in the sample target demand information to obtain the sample demand information. The sample target demand content in the sample target demand information represents multiple target demand intentions with intent dependency relationship, and the sample target demand content satisfies the preset semantic clarity condition.

10. The method according to claim 9, wherein, The target demand information for the sample is determined based on the following operations: The sample object acquires multiple historical interaction data points and uses these historical interaction data points to retrieve information. The second major model is used to perform semantic fusion on multiple historical interaction data to obtain the target demand information of the sample.

11. The method according to claim 10, wherein, The aforementioned historical interaction data are related to a specific business area; The step of using the second major model to perform semantic fusion on multiple historical interaction data includes: Based on the business requirement conditions corresponding to the specified business domain, the second major model is used to perform intent understanding on multiple historical interaction data, wherein the business requirement conditions represent the attribute dependency relationship between multiple business requirement attributes, and the intent dependency relationship between multiple target requirement intents with intent dependency relationship matches the attribute dependency relationship.

12. The method according to claim 10 or 11, wherein, The acquisition of multiple historical interaction data includes: Based on business keywords related to a specified business domain, determine intermediate historical interaction data that semantically matches the business keywords from multiple received initial historical interaction data; and Based on the semantic difference between the multiple intermediate interaction data, a plurality of historical interaction data are determined from the multiple intermediate interaction data.

13. The method according to claim 6 or 9, wherein, The second major model determines the training data by performing the following operations: The sample target demand information is subjected to intent understanding to determine the sample target demand content corresponding to the target demand intent; Based on the fuzzy demand intent determined from the target demand intent, the sample target demand information is rewritten to obtain the sample demand information, which represents that it has a semantic fuzzy attribute for the fuzzy demand intent. as well as Based on the ambiguous demand intent of the label and the sample demand information, semantic understanding is performed to obtain the label clarification prompt text.

14. An interactive device based on a large model, comprising: The first receiving module is used to receive the requirement information of the target object; The clarification prompt text acquisition module is used to perform semantic understanding of the demand information using the first major model and output clarification prompt text. The clarification prompt text is used to describe the demand intent with semantic ambiguity in the demand information and prompt the target object to provide feedback on the clarification demand content corresponding to the demand intent. The response information acquisition module is used to perform a retrieval task based on the clarification request content input by the target object, and obtain response information that matches the request intent; and The push module is used to push the reply information to the target object.

15. An apparatus for training large models, comprising: The second receiving module is used to receive training data, which includes sample demand information and label clarification prompt text. The sample demand information represents the demand intention of the sample object and has semantic ambiguity. The label clarification prompt text is used to prompt the sample object to provide clarification demand content corresponding to the demand intention, so as to perform a retrieval task based on the clarification demand content to obtain retrieval results that match the sample object. The sample clarification prompt text acquisition module is used to perform semantic understanding of the sample demand information using the first large model to be trained, and output sample clarification prompt text. as well as The training module is used to train the first large model based on the label clarification prompt text and the sample clarification prompt text, so as to obtain the trained first large model.

16. An intelligent agent of artificial intelligence, comprising: The input module is used to receive input information; The processing module is configured to determine a target task based on the input information received by the input module, determine a large model based on the target task, and execute the method of any one of claims 1 to 13 by calling the large model to obtain output information; An output module is used to output the output information obtained by the processing module.

17. 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 to 13.

18. 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 to 13.

19. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1 to 13.

Images