Interaction method and device based on large model, intelligent agent and storage medium

Abstract

The invention provides an interaction method and device based on a large model, an intelligent agent and a storage medium, and particularly relates to the technical fields of large models, intelligent e-commerce, intelligent agents and the like. The interaction method based on the large model comprises the steps that session information is displayed, the session information comprises reply content for input information of a user object and a strategy mark corresponding to the reply content, the strategy mark is used for prompting a target object and a reply strategy corresponding to the reply content, and the input information is processed by the large model according to the reply strategy to obtain the reply content; according to an interaction behavior of the target object for the reply content, determining adjustment demand information corresponding to the reply strategy, the adjustment demand information describing an adjustment demand intention of the target object for the reply content; according to the session information and the adjustment demand information, a training sample is determined, the training sample is used for adjusting a reply strategy of the large model, and an updated large model is obtained.

Description

Technical Field

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

[0002] With the rapid development of artificial intelligence technology, related platforms can use large models to semantically understand the user's input and output responses that meet that need. For example, large models can be used to understand the questions users enter on e-commerce platforms and output response text that meets the user's needs for product information. Summary of the Invention

[0003] This disclosure provides an interaction method, apparatus, intelligent agent, and storage medium based on a large model.

[0004] According to one aspect of this disclosure, a large-model-based interaction method is provided, comprising: displaying session information, the session information including response content to input information from a user object and a strategy tag corresponding to the response content, the strategy tag being used to prompt the target object with a response strategy corresponding to the response content, the large model processing the input information according to the response strategy to obtain the response content; determining adjustment requirement information corresponding to the response strategy based on the target object's interaction behavior with the response content, the adjustment requirement information describing the target object's intention to adjust the response content; and determining training samples based on the session information and the adjustment requirement information, wherein the training samples are used to adjust the response strategy of the large model to obtain an updated large model.

[0005] According to another aspect of this disclosure, a large-model-based interaction method is provided, comprising: receiving target input information from a user object; performing semantic understanding on the target input information using a specified large model, outputting target response content, and pushing the target response content to the target object; wherein, the specified large model is determined by adjusting the response strategy of the large model based on training samples, and the training samples are determined according to the large-model-based interaction method provided in the embodiments of this disclosure.

[0006] According to another aspect of this disclosure, a large-model-based interactive device is provided, comprising: a display module for displaying session information, the session information including response content to input information from a user object and a strategy marker corresponding to the response content, the strategy marker being used to prompt the target object with a response strategy corresponding to the response content, the large model processing the input information according to the response strategy to obtain the response content; a first determining module for determining adjustment requirement information corresponding to the response strategy based on the target object's interactive behavior toward the response content, the adjustment requirement information describing the target object's intention to adjust the response content; and a second determining module for determining training samples based on the session information and the adjustment requirement information, wherein the training samples are used to adjust the large model's response strategy to obtain an updated large model.

[0007] According to another aspect of this disclosure, a large-model-based interactive device is provided, comprising: a receiving module for receiving target input information from a user object; and a target response content acquisition module for performing semantic understanding on the target input information using a specified large model, outputting target response content, and pushing the target response content to the target object; wherein the specified large model is determined by adjusting the response strategy of the large model based on training samples, and the training samples are determined according to the method provided in the embodiments of this disclosure.

[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 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.

[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 according to an embodiment of the present disclosure.

[0017] Figure 4A The diagram illustrates an application scenario of a large-model-based interaction method according to another embodiment of the present disclosure.

[0018] Figure 4B The diagram illustrates an application scenario of a large-model-based interaction method according to yet another embodiment of this disclosure.

[0019] Figure 4C The diagram illustrates an application scenario of a large-model-based interaction method according to another embodiment of the present disclosure.

[0020] Figure 5 A schematic diagram illustrating the principle of a large-model-based interaction method according to an embodiment of the present disclosure is shown.

[0021] Figure 6 The diagram illustrates an application scenario of a large-model-based interaction method according to another embodiment of the present disclosure.

[0022] Figure 7 A flowchart illustrating a large-model-based interaction method according to another embodiment of this disclosure is shown schematically.

[0023] Figure 8 A block diagram of a large-model-based interactive device according to an embodiment of the present disclosure is shown schematically.

[0024] Figure 9 A block diagram of a large-model-based interactive device according to another embodiment of the present disclosure is shown schematically.

[0025] Figure 10A schematic block diagram of an artificial intelligence agent according to an embodiment of the present disclosure is shown.

[0026] Figure 11 A schematic block diagram of an example electronic device that can be used to implement embodiments of the present disclosure based on a large model interaction method is shown. Detailed Implementation

[0027] 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.

[0028] 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.

[0029] The inventors discovered that in scenarios involving interaction between intelligent agents and users, agents built upon the model understanding capabilities of large models sometimes produce responses that are inconsistent with the user's actual needs, leading to reduced accuracy and difficulty in meeting user demands. In e-commerce marketing and similar scenarios, the response strategies of agents providing intelligent responses to customers are difficult to update in a timely manner. This makes it difficult to promptly meet the urgent product recommendation needs of merchants and the information acquisition needs of users. Furthermore, adjusting response strategies for intelligent agents is time-consuming and has low accuracy, making it difficult to output accurate responses during actual interactions. It also fails to accurately recommend resources that match both the merchant's recommendation needs and the user's information acquisition needs, thus requiring repeated experimentation to adjust the response strategies of the large model.

[0030] This disclosure provides an interaction method, apparatus, agent, and storage medium based on a large model. The interaction method includes: displaying session information, which includes response content to input information from a user object and a strategy tag corresponding to the response content. The strategy tag is used to prompt the target object with a response strategy corresponding to the response content. The large model processes the input information according to the response strategy to obtain the response content. Based on the target object's interaction behavior with the response content, adjustment requirement information corresponding to the response strategy is determined. The adjustment requirement information describes the target object's intention to adjust the response content. Training samples are determined based on the session information and the adjustment requirement information. The training samples are used to adjust the large model's response strategy to obtain an updated large model.

[0031] According to embodiments of this disclosure, by displaying the response content to the input information and the corresponding strategy markers in the displayed session information, the target object can clearly understand the response strategy adopted by the large model in the interaction context during the interaction between the large model and the user object. This allows the target object to intuitively and conveniently understand the defect types of each response content and the corresponding response strategies. Based on the context of the interaction scenario, the target object can more accurately perform interactive behaviors on the response content, thereby determining the adjustment requirements corresponding to the response strategy. This allows the target object to more accurately, finely, and conveniently adjust the performance output of the large model in real interaction scenarios. Furthermore, by obtaining training samples, the response strategy of the large model can be precisely adjusted based on the context represented by the session information in the training samples, the response content corresponding to the response strategy, and the adjustment requirements. This improves the accuracy of strategy adjustments for the large model, enabling the adjusted large model to meet the actual adjustment needs of the target object and the actual interaction needs of the user object. This achieves rapid, convenient, and accurate adjustment of the model strategy in the interaction scenario, improving the user's interactive experience.

[0032] 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.

[0033] 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.

[0034] 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.

[0035] 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).

[0036] 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.

[0037] 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.

[0038] 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.

[0039] 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.

[0040] 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.

[0041] Figure 2 A flowchart illustrating a large-model-based interaction method according to an embodiment of the present disclosure is shown schematically.

[0042] like Figure 2 As shown, this interaction method based on a large model can be applied to terminal devices or servers related to the target object. The interaction method based on a large model includes operations S210~S230.

[0043] When operating S210, session information is displayed.

[0044] In operation S220, based on the target object's interaction behavior with the response content, the adjustment requirement information corresponding to the response strategy is determined.

[0045] In operation S230, training samples are determined based on session information and adjustment requirements information.

[0046] The execution subject of the method provided according to the embodiments of this disclosure can be a server. The embodiments of this disclosure are described based on the server as the execution subject and are not intended to limit the execution subject of the interaction method based on the large model provided in the embodiments of this disclosure.

[0047] According to embodiments of this disclosure, conversation information can be information content generated during a dialogue between a large model and a user object. For example, conversation information may include questions raised by the user object regarding information needs such as products or news, and the response content output by the large model based on semantic understanding of the user object's questions.

[0048] In some embodiments, session information can represent session messages between a user object and a large model during a specified historical interaction period. For example, it can be user input information and the response content determined by the large model in response to the input information within one hour. There can be multiple session information sets, which can represent the dialogue process between the large model and the user object during multiple historical interaction periods.

[0049] According to embodiments of this disclosure, the session information includes response content to the input information of the user object and a policy tag corresponding to the response content. The policy tag is used to prompt the target object with the response policy corresponding to the response content. The large model processes the input information according to the response policy to obtain the response content.

[0050] According to embodiments of this disclosure, the response content may include information in any modality, such as text, images, voice, and product resource links. Embodiments of this disclosure do not limit the specific modality of the information included in the response content.

[0051] According to embodiments of this disclosure, the response strategy may include any type of execution strategy such as retrieval enhancement strategy, non-retrieval enhancement strategy, and deep thinking strategy. Embodiments of this disclosure do not limit the specific type of response strategy, as long as it can be executed by a large model. Strategy markers can be displayed using any type of display method, such as text or graphics, as long as they can be observed by the target object and the type of response strategy represented by the strategy marker can be determined.

[0052] In some embodiments, the session information can be historical dialogue information between the customer service big model specified by the merchant object and the user object. The reply content in the session information can be the reply text obtained by the big model performing information generation tasks based on model strategies such as retrieval-enhanced reply strategies. In the displayed session information, the reply text and the corresponding strategy tag are displayed in adjacent positions, for example, the reply text and the strategy tag are displayed in adjacent positions above and below each other in the same column. Thus, the strategy tag can be used to prompt the target object big model about the reply strategy used to generate the reply text.

[0053] According to embodiments of this disclosure, the adjustment request information describes the target object's intention to adjust the response content. The adjustment request intention can be expressed as a denial of the response content; for example, the target object might input "The language style lacks respect and needs to be adjusted" in the response content. The adjustment request information can then be expressed as an intention to adjust the language style of the response content.

[0054] In some embodiments, the interactive behavior may include performing an operation on the displayed adjustment option component to determine, for example, "block inappropriate words" or "incorrect business scope selection". This allows the target object to display the adjustment option component when selecting response content, prompting the target object to select the corresponding adjustment option component for the defect type of the response content, so as to determine the adjustment requirement information corresponding to the selected target adjustment option component.

[0055] In some embodiments, the interactive behavior may include input operations by the target object. For example, when the target object performs a selection operation on the response content, an input box corresponding to the selected response content can be displayed. The target object can input adjustment requirement information such as a description of the defect corresponding to the response content, an example of a correct response content, etc., in the input box through any type of interactive operation, such as voice interaction or touch screen interaction. This allows the target object to accurately express its intention to adjust the defects of the response content through the input text content.

[0056] According to embodiments of this disclosure, training samples are used to adjust the response strategy of the large model, resulting in an updated large model. For example, based on the user's input information in the training samples as the sample input information of the large model, the large model performs semantic understanding of the input information based on the context and adjustment demand information in the conversation information and outputs sample response content. The large model is trained based on the difference between the labeled response content in the training samples and the sample response content to obtain an updated large model. The labeled response content can be an example of the response content in the adjustment demand information input by the target object through interactive behavior, or it can be the output of the target large model based on the semantic understanding of the input information according to the context and adjustment demand information in the conversation information. The target large model can act as a teacher model to output labeled response content, so as to adjust the model strategy of the large model based on a knowledge distillation mechanism. This allows the updated large model to output response content that satisfies the adjustment demand intent of the target object and the actual demand intent of the user object, improving the matching degree between the target response content and the user object's demand intent. Furthermore, the large model's response strategy can be adjusted in a fine-grained manner according to the adjustment demand information corresponding to the response content based on the target object's interactive behavior, achieving timely optimization of the user's interactive experience according to the target object's needs.

[0057] In some examples, training samples may include user input information, associated response content, and adjustment request information from the conversation. Based on this adjustment request information, prompt words can be determined, and the large model can be guided to understand the adjustment intent represented by the adjustment request information. This allows for semantic understanding of the input information, resulting in the output of a target response that satisfies both the user's needs and the target object's adjustment intent. Thus, updating the prompt words can adjust the large model's response strategy, improving the ease of interaction for adjusting strategies used to interact with users.

[0058] In some embodiments, the target object's interaction with the reply text may include an interaction operation indicating confirmation or modification of the reply content, to determine adjustment requirement information corresponding to the type of interaction operation. For example, the target object can perform a confirmation operation on approved reply content to use the reply content as positive-labeled reply content in the training samples, so as to adjust the model strategy of the large model based on the positive-labeled reply content in the training samples. As another example, the target object can perform a denial operation on reply content deemed unacceptable in the conversation information to use the reply content as negative-labeled reply content. Thus, the large model can be trained based on negative-labeled reply content in the training samples, so that the semantics of the sample reply content output by the large model are far removed from negative-labeled reply information. This allows the updated large model to avoid outputting reply content that does not meet the target object's requirements, enabling the large model to output reply content that matches the needs and intentions of both the target object and the user object during interaction, thereby improving the user's interactive experience.

[0059] It should be noted that the embodiments of this disclosure do not limit the specific adjustment method of the response strategy of the large model, as long as the large model can be optimized based on the conversation information and adjustment requirement information in the training samples.

[0060] It should be noted that the information acquisition involved in any embodiment of this disclosure, including but not limited to session information and adjustment requirement information, was obtained under conditions authorized by relevant personnel or organizations. The purpose of acquiring the information was communicated beforehand, and necessary encryption or desensitization measures were taken for the acquired information, complying with relevant regulations.

[0061] According to embodiments of this disclosure, the response strategy of the large model is adjusted based on the following operations: the large model performs semantic understanding of the conversation information according to the adjustment requirements to obtain sample response content; and the large model is trained based on the difference between the sample response content and the tag response content.

[0062] In some embodiments, the tagged response content is determined based on prompts identified by the adjustment requirements, using a target large model to perform semantic understanding of the conversation information. The target large model may have a larger number of parameters than the large model to be adjusted, or the response content generation performance of the target large model may be higher than that of the large model to be adjusted.

[0063] By semantically understanding the context of conversational information based on prompts related to adjustment needs using a target-oriented big data model, tagged responses that satisfy the adjustment intent and match the intent of the input information can be output. Therefore, based on a knowledge distillation mechanism, the target-oriented big data model can be used as a teacher model to generate the expected response content for the target audience based on adjustment need information and conversational information. This allows the tagged response text to be used as ground truth, and the big data model to analyze the adjustment intent...

[0064] According to embodiments of this disclosure, the interaction method based on a large model further includes: displaying adjustment progress information. The adjustment progress information may represent the interaction behavior performed by the target object on multiple session information; for example, the adjustment progress information may represent the number of session information on which the target object has already performed an interaction behavior.

[0065] In some embodiments, the adjustment progress information includes at least one of sample utilization and response content interaction data.

[0066] According to embodiments of this disclosure, sample utilization represents the number of training samples that have been used during the process of adjusting the response strategy of a large model. Sample utilization can be, for example, the ratio of the number of used samples to the number of currently generated training samples, where used samples are those that have already been used to train the large model.

[0067] The response content interaction data represents the statistical results of adjusted response content among multiple responses, and the target object performs interactive behavior in response to the adjusted response content. For example, the response content interaction data can be the statistical number of response content for which the target object has performed interactive behavior. Alternatively, the response content interaction data can be the ratio between the number of adjusted response content and the total number of response content included in the collected session information. Therefore, the displayed response content interaction data can indicate the number of response content that the target object has adjusted, reducing the interaction complexity of the target object's interaction with session information adjustment requests.

[0068] In some embodiments, the interaction method based on the large model further includes: displaying session attribute information.

[0069] In some embodiments, session attribute information can represent any data related to the attributes of the session information, such as the content of the session information or the large model generated by the session information. Session attribute information characterizes at least one of the following attribute information corresponding to the session information: time attribute, response strategy, and adjustment request information. Specifically, the time attribute can represent time-related information such as the time period or moment in which the session information was generated; the response strategy can represent any type of response strategy data, such as the retrieval enhancement strategy included in the session information; and the adjustment request information can represent the content of adjustment request information generated by the target object's interaction with the session information, or it can also represent whether the target object has performed an interaction with the session information.

[0070] According to embodiments of this disclosure, the target object determines the displayed session information by selecting based on session attribute information. For example, the target object uses the displayed session attribute information to filter response content that requires targeted interactive actions or targeted review. This enables the target object to filter response content that requires interactive actions based on the displayed session attribute information, thereby reducing the interactive complexity of adjusting response strategies for large models and improving the computational efficiency of adjusting response strategies for large models.

[0071] Figure 3 The diagram illustrates an application scenario of the interaction method based on a large model according to an embodiment of the present disclosure.

[0072] like Figure 3 As shown, the display interface 310 displays session attribute information and adjustment progress information. The adjustment progress information is displayed in the progress statistics box 311 within the display interface 310. The adjustment progress information may include the adjustment ratio, number of sessions, number of adjusted sessions, model application rate, number of training samples, and number of model application samples. The adjustment ratio represents the ratio of the number of session information items that have already interacted with the target object to the total number of session information items collected. The number of sessions represents the total number of session information items collected, and the number of adjusted sessions represents the number of session information items that have already interacted with the target object. The model application rate represents the sample utilization rate, characterizing the ratio of the number of training samples used in adjusting the response strategy of the large model to the total number of training samples acquired. The number of training samples represents the total number of training samples acquired. The number of model application samples represents the number of training samples used in adjusting the response strategy of the large model.

[0073] The session attribute information includes the session number, session time, whether retrieval enhancement was performed, model application status, and number of adjusted messages. The session number serves as a unique identifier for the session information, and the session time indicates when the session information was generated. Whether retrieval enhancement was performed indicates whether the response content in the session information was generated based on a retrieval enhancement strategy. The model application status can be "Collected," "Applied," or "—". "Collected" indicates that training samples have been generated for this session information; "Applied" indicates that the training samples corresponding to this session information have been used to adjust the response strategy of the large model; "—" indicates that no interaction has been performed on this session information. The number of adjusted messages represents the number of response messages for which interaction has been performed. The target user can interact with the "View Session" option to view the input information and response content in the session information corresponding to the session number, allowing them to interact with the response content and input adjustment requirements based on the context of the session information.

[0074] According to embodiments of this disclosure, the strategy tagging includes retrieval enhancement tags, which characterize the retrieval result data in the response content determined by the large model based on the initial resources obtained from the retrieval.

[0075] In some embodiments, search enhancement tags can be displayed in a nearby location along with their corresponding search enhancement responses. This allows target users to quickly locate the search enhancement responses generated by executing search enhancement strategies based on a large model by browsing the search enhancement tags.

[0076] According to embodiments of this disclosure, the search result data in the response content is determined based on the resource data corresponding to the retrieved initial resource. For example, the search result data may be product description text, product sales amount data, etc., contained in the initial resource. Alternatively, the search result data may be obtained by processing the resource data in the retrieved initial resource through calculation, fusion, or other processing methods. For example, the search result data may be the calculation result obtained by performing calculation operations on the data in the input information using the calculation formula in the retrieved calculation standard page.

[0077] It should be noted that the embodiments of this disclosure do not limit the specific type of search result data, as long as it is determined based on the initial resources retrieved.

[0078] In some embodiments, the large-model-based interaction method may further include, based on a designated agent, invoking a retrieval tool to perform resource retrieval according to the semantic understanding results of the large model on the input information, thereby obtaining initial resources that match the semantic understanding results of the input information. The agent then transmits the resource data of the initial resources to the large model, which performs semantic fusion based on the resource data to generate the response content in the conversation information.

[0079] In some embodiments, determining the adjustment requirement information corresponding to the response strategy based on the target object's interaction behavior with the response content may include: determining the target resource based on the interaction behavior with the search result data; and determining the search adjustment requirement information based on the target resource and the search result data.

[0080] In some embodiments, the target resource can be a resource that meets the business needs of the target object. For example, the target resource could be the latest product resource that the target object needs to market in the future, or a specific promotional resource that the target object needs to promote. Alternatively, the target resource could be a corrected version of the initial resource in the response content that contains a recommendation error. For example, the initial resource could be a purchase page for Class A sneakers, and the search result data in the response content could be product description text for Class A sneakers. After browsing the contextual conversation content based on the conversation information, the target object believes that the search result data in the response content output by the large model should be product description text related to Class A rackets. Therefore, by performing interactive behavior on the response content, the target resource corresponding to the search result data can be determined. Thus, the purchase page for Class A rackets can be associated and stored with the search result data to obtain search adjustment requirement information.

[0081] According to embodiments of this disclosure, the retrieval adjustment requirement information indicates the target object's intention to adjust the resource data source of the retrieval result data in the response content. This retrieval adjustment requirement information is used to prompt the large model to determine new response content based on the target resources. This allows the target object to combine the dialogue context represented by the conversation content in the conversation information to perform timely and fine-grained interactive actions on the retrieval result data in the response content. This prompts the large model to prioritize generating target response content based on target resources that meet the target object's business needs for similar dialogue contexts or similar input information in future periods, thereby improving the accuracy of adjusting the large model's response strategy.

[0082] In some embodiments, displaying session information may further include displaying a source marker representing the initial resource at a location corresponding to the search result data.

[0083] According to embodiments of this disclosure, the source marker can represent search result data such as text, numerical values, images, and resource link components in the response content, and is determined based on resource data in the corresponding initial resource. The target object can determine the association between the initial resource and the search result data based on the source marker.

[0084] In some embodiments, the source marker can be a specified marker element, such as an underline, a color-highlighted pixel, or a superscript text element corresponding to the search result data. The embodiments of this disclosure do not limit the specific display format of the source marker.

[0085] In some embodiments, the response content and source marker in the session information can be determined based on the following operations: Input information from a user and contextual session content from the session information are received during a historical time period. A large model is used to perform semantic understanding on the input information and contextual session content to obtain retrieval execution data. An agent is used to invoke a retrieval tool to perform a retrieval based on the retrieval execution data, obtaining retrieval results of any type, such as document fragments, page text, and table data. Each retrieval result is assigned an initial resource identifier representing the data source as a source marker. The large model is used to perform semantic fusion on the retrieval results and source markers based on the input information to obtain response content, and the corresponding source marker is inserted into the retrieval result data side of the response content. This allows the large model, which performs the response content generation task based on a retrieval enhancement strategy, to insert source markers into the retrieval result data in the response content. This enables the target object to check or query the response content based on the associated source markers and retrieval result data, improving the ability to trace the source of the response content and reducing the interaction complexity for the target object in adjusting business needs. It allows the target object to directly adjust the large model's response strategy based on the dialogue context represented by the real session content, enabling the large model to meet both business needs and the actual needs of the user object after adjusting the response strategy.

[0086] In one example, the target user can perform a selection operation based on the displayed source marker. The display interface will then redirect to the resource page of the initial resource associated with the search result data, allowing the target user to promptly check whether the initial resource is suitable as the search result data in the response content within the current conversation context. When the target user determines that they need to adjust the response to the input information to the target resource, the search result data corresponding to the initial resource to be adjusted can be precisely determined based on the source marker. This enables the interactive behavior related to the search result data to determine the associated target resource and search result data in the search adjustment request information, and allows for timely adjustment of the large model's response strategy according to the target user's actual business needs. This ensures that the large model can match the user's business needs with the target user's resource recommendation needs, improving recommendation accuracy and the flexibility and precision of the large model's recommendation strategy adjustments.

[0087] Figure 4A The diagram illustrates an application scenario of a large-model-based interaction method according to another embodiment of the present disclosure.

[0088] like Figure 4A As shown, the interactive interface displays a session information box 410 and an interaction behavior box 420. The session information box 410 displays the input information of the target object and the response content generated by the large model based on the input information within the historical time period. The response content and strategy markers are displayed adjacent to each other; for example, the first response content 411 and the corresponding first search enhancement marker 412 are displayed adjacent to each other, so that the target object can quickly find the response content that allows for target resource adjustments based on the first search enhancement marker 412.

[0089] The search result data "consulting fee 1xxx yuan" in the first reply (content 411) was generated by the large model based on resource data from the initial resource. The initial resource could be the service product page for consulting service item A. The source marker could be the underscore "__" corresponding to "consulting fee 1xxx yuan". By browsing to the source marker, the target audience can confirm that the search result data was determined by the large model based on the resource data of the initial resource.

[0090] When the target audience interacts with the search result data "Consultation fee 1xxx yuan", the display interface can redirect to the service product page of Consultation Service A to facilitate the target audience's tracing of the search result data. When the target audience needs to adjust the large model's response strategy according to business requirements, they can interact with the search result data "Consultation fee 1xxx yuan". The interaction action box 420 in the display interface will show the optimization component corresponding to the first response content 411. The optimization component can include "Avoid blocked words", "Copywriting needs optimization", "Adjust business scope", and "Adjust recommended products". The target audience interacts with the optimization component and enters the product name of the target resource that meets their business requirements in the resource selection box 401 of the interaction action box. This identifies the target resource associated with the search result data "Consultation fee 1xxx yuan" in the first response content 411. This ensures that the training samples contain the associated input information "How much is the fee?", the search result data "Consultation fee 1xxx yuan" in the first response content 411, and the target resource that meets the target audience's business requirements. The large model can understand the dialogue context based on the contextual conversation content in the training samples, and adjust the response strategy of the large model in a timely and accurate manner based on the relevant search result data and target resources in the search adjustment demand information. This allows the large model to understand the target object's business need intention for resource recommendation more accurately, as well as the accurate recommendation intention combined with the contextual conversation content, thereby improving the recommendation accuracy of target resources.

[0091] According to embodiments of this disclosure, based on the target object's interactive behavior towards the interactive interface, the target object can accurately determine the search result data that needs adjustment based on the contextual session content in the session information. By generating related search result data and input information from the training samples, and adjusting the large model to prioritize the recommendation of target resources, the target object's business needs for recommending target resources can be accurately met. This enables the large model to match the user's business needs with the target object's resource recommendation needs based on the contextual session content and target resources, thereby improving recommendation accuracy and the flexibility and accuracy of the large model's recommendation strategy adjustment.

[0092] In some embodiments, the strategy tags include retrieval enhancement tags, which represent the retrieval result data in the response content determined by the large model based on the initial resource obtained from the retrieval. The session information also includes source tags, which represent the association between the retrieval result data and the initial resource.

[0093] In some embodiments, determining the adjustment requirement information corresponding to the response strategy based on the target object's interaction behavior with the response content may further include: determining the initial resource corresponding to the retrieval result data based on the selection operation for the source tag; and updating the resource data of the initial resource based on the update operation for the initial resource to obtain the target resource.

[0094] According to embodiments of this disclosure, the adjustment requirement information includes target resources. An update operation on the initial resource can represent adding, deleting, replacing, or modifying any resource data such as text, numerical values, or images of the initial resource.

[0095] For example, the target object can modify the product performance description text for Class A rackets on the initial resource page to obtain the target resource page. The content of this target resource page will then be used as the target resource in the adjustment requirements information.

[0096] In some embodiments, performing an update operation on the initial resource may further include modifying any information related to the initial resource, such as product resource description information and business type information, based on the update operation. This allows the target object to promptly and accurately determine erroneous search result data or search result data requiring adjustments to business needs based on the contextual session content in the actual session information. Simultaneously, by directly jumping from the session information browsing interface to the initial resource update interface through the update operation on the initial resource, it enables convenient and fine-grained updates to the initial resource to address errors in the response content. Thus, by promptly modifying erroneous data and temporary business adjustment data in the initial resource during the process of checking the quality of the large model's response content, the convenience and accuracy of meeting the target object's real-time business needs are improved, achieving rapid fulfillment of the needs of both the user and the target object.

[0097] In some embodiments, the adjustment requirement information includes at least one of the following: semantic style requirement information, content blocking requirement information, expected response content, and business type adjustment information;

[0098] According to embodiments of this disclosure, semantic style requirement information represents the expected expression style of the target object for the response content. For example, the target object performs semantic understanding of the context of the response content "Let me check" and inputs the semantic style requirement information "A response based on a two-dimensional style is required." This allows the response content in the training samples to be associated with the semantic style requirement information, and the model strategy of the large model can be adjusted through the training samples. This enables the large model to adjust the expression style of the response content based on context and input information similar to the conversation information in a future specified period. This achieves the goal of obtaining a large model that can meet the needs and intentions of the user and target object by finely adjusting the requirement information input, thereby improving the accuracy and timeliness of adjusting the response strategy of the large model.

[0099] According to embodiments of this disclosure, the content blocking requirement information can represent words, phrases, or business content that the target object indicates should be avoided in the response content. For example, the content blocking requirement information can be the blocked word "AA" input by the target object. This allows the training samples to contain the indicated blocked word "AA" to prompt the large model to avoid including "AA" in the output response content, enabling the updated large model to output response content based on synonyms or near-synonyms in a future specified time period.

[0100] According to embodiments of this disclosure, the expected response content represents the response that the target user expects in response to the input information. For example, in response to the input information "Which ones suit me?", the target user might think that the tone of the response, "The first one," is insufficient to meet the user's actual needs, and that the expression style is stiff, failing to satisfy the user's emotional needs. Therefore, the target user might input the expected response, "The first one is made of wool, and its design is very novel, meeting your requirements for fashionable coats."

[0101] For example, a resource call interface related to an agent with specific capabilities can be displayed. The target object inputs session information and rewrite requirements for the response content through the resource call interface. The agent then performs semantic understanding of the context in the session information based on the rewrite requirements and outputs the response content expected by the target object.

[0102] In some embodiments, the expected response content in the training samples can be used as the labeled response content for training the large model, so as to train the large model according to the difference between the labeled response content and the sample response content output by the large model to adjust the response strategy of the large model.

[0103] According to embodiments of this disclosure, the business type adjustment information describes a target business type that matches the response content determined based on the input information. For example, in the conversation information, the response to the input information "How is City A?" is "City A has many attractions; you can visit Garden A." This response content represents a tourism business. After browsing the contextual content in the conversation information, the target object can perform an interactive action based on the response content to input business type adjustment information representing "hotel booking business." Thus, based on the input information, response content, and business type adjustment information in the conversation information, the large model can be prompted to perform a generative task based on a response strategy related to "hotel booking business" to obtain target response content that meets the demand intent of "hotel booking business."

[0104] In one example, the target object interacts with the response content corresponding to the input information by responding to the context content in the conversation information, thus inputting business type adjustment information. This allows the target object to adjust the business type push target based on the content in the real interaction scenario and the context of the real conversation information. This enables the target object to finely adjust the response strategy of the large model by inputting business type adjustment information into the response content. The large model can then understand the target object's need for business type adjustment based on the context content in the real conversation information. This prompts the large model to understand the question semantics or context semantics related to the input information and prioritizes generating new target response content based on the adjusted target business type. This ensures that the target response content accurately meets the actual needs and intentions of both the target object and the user object, improving the accuracy of the target response content and timely meeting the target object's real-time needs for business type adjustment.

[0105] It should be noted that the adjustment requirement information may include any one or more of the following: semantic style requirement information, content blocking requirement information, expected response content, and business type adjustment information. The embodiments disclosed herein will not be described in detail here.

[0106] Figure 4B The diagram illustrates an application scenario of a large-model-based interaction method according to yet another embodiment of this disclosure.

[0107] like Figure 4BAs shown, the target object can browse the second response content 421 and non-retrieval enhancement markers 422 displayed in the session information box 410, thereby enabling the target object to determine that the second response content 421 was obtained by the large model performing a response content generation task without a retrieval enhancement strategy. The target object performs a selection operation on the second response content 421, and the interaction action box 420 can display the tuning item component corresponding to the second response content 421. The target object performs an interaction operation on the tuning component "Adjust Business Scope", and enters the new target business type corresponding to the second response content 421 in the input box corresponding to "Adjust Business Type", and enters the business description information corresponding to the target business type in the input box corresponding to "Business Content".

[0108] For example, the target user enters "Country A tourism business" in the input box corresponding to "Adjust Business Type" and enters descriptive information related to tourism products in Country A in the input box corresponding to "Business Content". This determines that the business type adjustment information includes the target business type and business product description information corresponding to the second response content 421. Through the associated business type adjustment information and the second response content 422 in the training samples, the large model can deeply understand the target user's actual business adjustment needs based on the contextual conversation content in the conversation information and adjust its response strategy accordingly. This allows the updated large model to generate response content related to tourism products in Country A based on similar contextual conversation content in future periods, thereby improving the target user's actual needs for business adjustments.

[0109] In some embodiments, the adjustment information may further include adjustments such as liking, quality control, or disapproval of the response content. The display interface can show the corresponding representation of the adjustment information to prompt the target object about the interaction results of the response content. Simultaneously, the target object can also query response content that has undergone interactive behavior based on adjustment information such as liking or disapproval, thereby improving the efficiency of training sample generation and training efficiency for large models.

[0110] In some embodiments, feedback results for the adjusted response content can be generated based on the feedback time of the large model processing the training samples. For example, if no feedback information is output within a preset time period during the process of the large model processing the adjusted response content and the corresponding adjustment requirement information, the response content is considered invalid.

[0111] In some embodiments, the training samples further include contextual session content and tag response content corresponding to the adjustment requirement information. The tag response content includes tag retrieval result data determined based on the target resource, and the contextual session content is determined by performing a selection operation on the response content in the session information based on the target object.

[0112] According to embodiments of this disclosure, the tagged response content can be an example of the response content input by the target object. This allows the expected response content of the target object to be represented based on the example response content, and the example response content can be used as tagged response content to improve the accuracy of model strategy adjustment for large models.

[0113] In some embodiments, the tagged response content can also be generated using a target big model based on semantic understanding of the contextual conversation content, input information, response content, and adjustment requirement information. This allows for the use of a knowledge distillation mechanism to understand the dialogue context based on the contextual conversation content selected by the target object, and to rewrite the response content based on the requirement intent represented by the input information and the business requirement intent for the target resource represented by the adjustment requirement information, thereby generating tagged response content that satisfies both the business requirement intent of the target object and the requirement intent of the user object.

[0114] In some embodiments, the response strategy of the large model is adjusted based on the following operations: based on the target resource, the large model performs semantic understanding of the contextual conversation content and adjustment requirement information to output the sample retrieval intent; based on the sample retrieval intent, the large model performs a response task on the input information to obtain the sample response content; and the large model is trained according to the difference between the sample response content and the tag response content to obtain an updated large model.

[0115] In some embodiments, sample retrieval intent can represent retrieval information used to retrieve sample retrieval resources. For example, sample retrieval information may include search terms. The sample retrieval resource can be a target resource, or, during the training of a large model, it can be a different resource than the target resource. The large model performs semantic understanding of the input information and executes a response task based on the sample retrieval resource, enabling the sample retrieval result data in the sample response content to be determined based on the sample retrieval resource.

[0116] Therefore, training a large-scale model based on the differences between sample response content and tag response content enables the model to accurately retrieve target resources from preset resources during the semantic understanding of contextual conversation content, adjustment needs information, and target resources. Furthermore, the model learns during training to perform response tasks based on resource data within the target resource and input information. It can combine the contextual conversation content selected by the target object to deeply understand the input needs and the target object's business needs for recommending the target resource. This allows for accurate retrieval of the target resource while simultaneously extracting accurate resource data from it to generate sample retrieval results that are semantically close to the tag retrieval results, ensuring semantic similarity between the sample response content and the tag response content. Furthermore, based on the contextual conversation content and target resources selected by the target object, the response strategy of the large model can be easily adjusted. This allows the updated large model to generate response content based on the target resources, according to the dialogue context similar to the contextual conversation content selected by the target object. This ensures that the search results data in the response content, based on the target resources, can simultaneously meet the business needs and intentions of both the target object and the user, thereby improving the user experience and enhancing the accuracy and timeliness of the target object's response strategy adjustment of the large model.

[0117] In some embodiments, the policy tag also indicates that the response content is the configuration response content output to the user object when the configuration keyword matches the specified keyword in the input information.

[0118] According to embodiments of this disclosure, the configuration response content may include responses to common questions such as the target object's product manufacturing address and merchant contact information.

[0119] In one example, if the user's input information contains the specified keyword "place of production", the configuration response content "the place of production is a high-altitude sterile area in province A" can be determined based on the matching results between the specified keyword and the configured keyword "production address".

[0120] According to embodiments of this disclosure, the interaction method based on a large model may further include: updating at least one of configuration keywords or configuration response content based on configuration update information indicated by the interaction behavior.

[0121] According to embodiments of this disclosure, the configuration update information can be obtained by the target object modifying either the configuration keyword or the configuration response content. The target object modifies at least one of the configuration keyword or the configuration response content by inputting the configuration update information to obtain the target configuration keyword or the target configuration response content. Therefore, in a future specified time period, in response to the matching of target input information input by the user object with the target configuration keyword, the updated target configuration response content can be pushed to the user object; alternatively, based on the matching result of the target input information and the configuration keyword, the target configuration response content corresponding to the configuration keyword can be pushed to the user object.

[0122] For example, if the configured keyword is "production location," and the target configured response content is "production location is a high-altitude sterile area in province B," then if the user's input information contains the specified keyword "production location," the target configured response content "production location is a high-altitude sterile area in province B" will be pushed to the user.

[0123] Figure 4C The diagram illustrates an application scenario of a large-model-based interaction method according to another embodiment of the present disclosure.

[0124] like Figure 4C As shown, by browsing the third response content 431 and the configuration question-and-answer strategy marker 432, the target object can determine the input information "Where is the address?" and directly generate the configuration response content "No. 33x, Street B, City A" on the interactive interface. If the target object's business address changes, the third response content 431 can be updated through interactive operations to input the updated configuration information "No. 11x, Street B, City A". This allows for the determination of the configuration keyword "address" and the target configuration response content "No. 11x, Street B, City A" to generate new configuration question-and-answer pair information. This ensures timely response to the user's target input information in future periods, pushing matching target configuration response content.

[0125] Figure 5 A schematic diagram illustrating the principle of a large-model-based interaction method according to an embodiment of the present disclosure is shown.

[0126] like Figure 5As shown, historical session information generated during historical periods in the interaction between user objects and the business system can be stored in a historical session dataset. The merchant object interaction module can be used to execute the large-model-based interaction method provided according to this disclosure. With the authorization of the relevant user object, the merchant object interaction module, from the massive offline session data stored in the historical session dataset, filters out session content that does not meet the relevant business requirements of the merchant object interaction module to reduce the data volume. The filtered session content includes user-initiated sessions and non-anti-fraud sessions. The merchant object interaction module calls an agent to annotate the response content in the session information with response strategies, obtaining strategy tags corresponding to each response. Strategy tags can represent retrieval enhancement strategies, non-retrieval enhancement strategies, and configured question-answering strategies.

[0127] The response content corresponding to the detection enhancement strategy can be the retrieval enhancement strategy session information, the response content corresponding to the non-retrieval enhancement strategy can be the non-retrieval enhancement strategy session information, and the configuration response content based on the configuration keyword is represented as configuration question and answer information. The display interface prompts the target object to perform interactive behaviors for optimizing the large model by showing the strategy tags corresponding to the response content, thereby obtaining the adjustment requirement information corresponding to the response content.

[0128] To address potential issues such as incorrect referenced resource data or inappropriate reference to initial resources when retrieving enhanced strategy session information, the target object can determine the target resource by performing interactive actions or modifying the resource data of the initial resource, so as to associate and store the target resource with the retrieval enhanced strategy session information.

[0129] Non-retrieval-enhanced conversation information is dynamically generated by a large model based on the context of the conversation or predetermined rules. It does not rely on external resource data retrieval enhancement. However, the generated response content may not match the context of the conversation, leading to user dissatisfaction. Examples include inappropriate sales-style responses or non-compliant responses related to system risk control. Target users can interact with flawed responses to input relevant adjustment requests.

[0130] Configuring question-and-answer information involves the interactive agent providing general, reasonable responses that maintain dialogue coherence and completeness, based on the context of the conversation or keywords in the input information. Examples of configuring question-and-answer information include responses with merchant addresses and answers to general questions.

[0131] The target user can interact with the above three types of responses to input adjustment requirements. For example, adjustment requirements might include primary and secondary defect reasons. A primary defect reason could be an error in the business description, indicating that the response generated by the large model does not match the user's actual needs and intentions, and contains factual or logical errors.

[0132] Level 2 defects include incorrect vocabulary in responses and the need for optimization of wording. Incorrect vocabulary indicates that the responses generated by the large model contain spelling errors, inappropriate word choice, or semantic ambiguity, affecting the quality and readability of the responses. The need for optimization of wording indicates that the responses output by the large model have shortcomings in expression, language style, and information completeness, requiring further optimization and improvement. It should be noted that the reasons for Level 1 and Level 2 feedback are not fixed and can be dynamically updated and improved based on factors such as the target audience's business needs and model performance optimization.

[0133] The training sample construction module constructs training samples based on adjustment requirements and session information. Training samples can include input information from the complete session, original response content, tagged response content, etc. Training samples can be input into the training sample set to train the large model. Discarded, unqualified responses are stored in the discard dataset to avoid training sample confusion. The target object inputs configuration update information based on the configuration question-and-answer information and stores the resulting target configuration keywords and target configuration response content in the configuration information repository. The target object can also configure response content generated based on non-retrieval enhancement strategies to obtain new configuration question-and-answer information, which is then stored in the configuration information repository.

[0134] Training samples can be used to fine-tune the parameters of a large model for training purposes. For example, in supervised training, input information from training samples can be used as training data, the expected response content of the target object can be used as labels, and resource data from the target resource can be used as prompts. The large model then outputs sample responses, and the differences between the labeled responses and the sample responses are used to train the large model multiple times, enabling the updated model to understand and correctly use new knowledge.

[0135] For example, for semantic style adjustment information and sentence structure modification information in the adjustment requirements, training can be performed based on preference alignment. The training samples include input information, expected response content, and original response content. The large model processes the input information and contextual conversation content of the training samples and outputs sample response content. The expected response content is used as the positive label, and the original response content is used as the negative label to fit the sample response content output by the large model towards the expected response style of the target object, thus obtaining an updated large model.

[0136] In some embodiments, the large model-based interaction method may further include: displaying an experience session interface; responding to experience input information input to the experience session interface, using an updated large model to perform semantic understanding of the experience input information, and outputting experience response content.

[0137] According to embodiments of this disclosure, the experience session interface can represent an interface simulating a user object's conversation with an updated large model. The experience input information can be simulated demand information input by the target object, or it can be demand information generated based on preset prompts from the large model.

[0138] According to embodiments of this disclosure, by using an updated large model in the experience session interface to output experience response content based on experience input information, the target audience can intuitively evaluate the performance of the updated large model after adjusting the response strategy. This allows for timely verification of the response content generation effect of the updated large model and enables further improvements.

[0139] Figure 6 The diagram illustrates an application scenario of a large-model-based interaction method according to another embodiment of the present disclosure.

[0140] like Figure 6 As shown, the interactive interface displays a business adjustment interface 610 and an experience session interface 620. The business adjustment interface 610 allows the target user to input adjustment requirements such as the business name, business scope, and description of business advantages, representing their business needs. After determining the updated large model, a conversation with the user can be simulated in the experience session interface, enabling the updated large model to perform a response task based on the user input information, such as "How much does it cost?" The large model outputs experience response content 621 based on a retrieval enhancement strategy. This allows the target user to detect whether the large model can respond based on similar contextual conversation content and resource data of the target resource, improving the accuracy of the target user's adjustments to the large model.

[0141] Figure 7 A flowchart illustrating a large-model-based interaction method according to another embodiment of this disclosure is shown schematically.

[0142] like Figure 7 As shown, the interaction method based on the large model can be applied to terminal devices related to user objects. The interaction method based on the large model includes operations S710~S720.

[0143] When operating the S710, the target input information of the user object is received.

[0144] When operating the S720, a specified large model is used to perform semantic understanding on the target input information, output the target response content, and push the target response content to the target object.

[0145] According to embodiments of this disclosure, the designated large model is determined based on adjusting the response strategy of the large model using training samples, and the training samples are determined according to the interaction method based on the large model provided in embodiments of this disclosure.

[0146] In some embodiments, the specified large model is determined based on the following operations: displaying session information, which includes response content to user input and a strategy tag corresponding to the response content, the strategy tag being used to prompt the target object with the response strategy corresponding to the response content; determining adjustment requirement information corresponding to the response strategy based on the target object's interaction behavior with the response content, the adjustment requirement information describing the target object's intention to adjust the response content; and adjusting the response strategy of the large model based on the training samples determined by the session information and the adjustment requirement information to obtain the specified large model.

[0147] The interaction method for a terminal device related to a user object provided in the embodiments of this disclosure can be executed based on an updated large model determined according to the embodiments of this disclosure. For example, the designated large model can be an updated large model trained based on the interaction method based on the large model in the above embodiments, which is used as the designated large model to execute the interaction method for the terminal device related to the user object. Further details are omitted here.

[0148] According to embodiments of this disclosure, semantic understanding of target input information is performed using a specified large model, including performing the following operations using the specified large model: making a response strategy decision based on historical interaction information related to the target object and the target input information to obtain a retrieval enhancement strategy; and performing a retrieval task based on the target input information according to the retrieval enhancement strategy to obtain target resources; and performing semantic understanding based on the target resources and the input information to output target response content.

[0149] According to embodiments of this disclosure, a retrieval enhancement strategy may include retrieval-related information such as search terms and search conditions for retrieving the target resource. The large model can send the retrieval-related information from the retrieval enhancement strategy to the retrieval agent, enabling the agent to retrieve the target resource by invoking a retrieval tool. This allows the large model to perform a response task based on the target resource and target input information, obtaining the target response content.

[0150] In some embodiments, a common session rendering component can be used to support the interaction logic of the merchant's interactive interface (where the target object executes interaction methods) and the user interface (where the user object executes interaction methods). This reduces the development cost of the interaction component and reproduces the realistic interaction and style when users use a large model for conversation. The common session rendering component renders uniformly on the user's page and displays the session information with the user on the merchant's page. The pages obtain interaction data through bidirectional communication via postMessage. By recognizing session information, the corresponding information content pathways between the user and merchant sides are connected. This enables the merchant object to use a real-time message pathway for conversation in real-world user interaction scenarios and in experience optimization scenarios. When configuration items change, the pathway is restarted to display the adjusted content.

[0151] Figure 8 A block diagram of a large-model-based interactive device according to an embodiment of the present disclosure is shown schematically.

[0152] like Figure 8 As shown, the interactive device based on the large model can be a first interactive device based on the large model 800. The first interactive device based on the large model 800 can be used for the terminal device of the target object. The first interactive device based on the large model 800 includes: a display module 810, a first determining module 820 and a second determining module 830.

[0153] The display module 810 is used to display session information, which includes the response content to the user object's input information and the strategy tag corresponding to the response content. The strategy tag is used to prompt the target object with the response strategy corresponding to the response content. The large model processes the input information according to the response strategy to obtain the response content.

[0154] The first determining module 820 is used to determine the adjustment requirement information corresponding to the response strategy based on the target object's interactive behavior towards the response content. The adjustment requirement information describes the target object's intention to adjust the response content.

[0155] The second determining module 830 is used to determine training samples based on session information and adjustment requirement information, wherein the training samples are used to adjust the response strategy of the large model to obtain an updated large model.

[0156] According to embodiments of this disclosure, the strategy tagging includes retrieval enhancement tags, which characterize the retrieval result data in the response content determined by the large model based on the initial resources obtained from the retrieval; wherein, the first determining module includes:

[0157] The first determining unit is used to determine the target resource based on the interaction behavior with the search result data.

[0158] The second determining unit is used to determine the retrieval adjustment requirements based on the target resources and retrieval result data. The retrieval adjustment requirements are used to prompt the large model to determine new response content based on the target resources.

[0159] According to embodiments of this disclosure, the display module includes a first display unit.

[0160] The first display unit is used to display a source marker representing the initial resource at a position corresponding to the search result data, wherein the target object determines the association between the initial resource and the search result data based on the source marker.

[0161] According to embodiments of this disclosure, the strategy tag includes a retrieval enhancement tag, which represents the retrieval result data in the response content determined by the large model based on the initial resource obtained from the retrieval. The session information also includes a source tag, which represents the association between the retrieval result data and the initial resource. The first determining module includes a third determining unit and an updating unit.

[0162] The third determining unit is used to determine the initial resource corresponding to the retrieval result data based on the selection operation for the source marker.

[0163] The update unit is used to update the resource data of the initial resource according to the update operation for the initial resource to obtain the target resource, wherein the adjustment requirement information includes the target resource.

[0164] According to embodiments of this disclosure, the training samples further include contextual conversation content and tag-based response content corresponding to adjustment requirement information. The tag-based response content includes tag retrieval result data determined based on the target resource. The contextual conversation content is determined by performing a selection operation on the response content in the conversation information based on the target object. The response strategy of the large model is adjusted based on the following operations: based on the target resource, the large model performs semantic understanding of the contextual conversation content and adjustment requirement information to output a sample retrieval intent; based on the sample retrieval intent, the large model performs a response task on the input information to obtain sample retrieval resources obtained by retrieving preset resources, thus obtaining sample response content; and the large model is trained based on the difference between the sample response content and the tag-based response content to obtain an updated large model.

[0165] According to embodiments of this disclosure, the adjustment requirement information includes at least one of the following: semantic style requirement information, masked content requirement information, expected response content, and business type adjustment information; wherein, the expected response content represents the response content that the target object expects in response to the input information, and the business type adjustment information describes the target business type that matches the response content determined based on the input information.

[0166] According to embodiments of this disclosure, the response strategy of the large model is adjusted based on the following operations: the large model performs semantic understanding of the conversation information according to the adjustment demand intention to obtain sample response content; and the large model is trained based on the difference between the sample response content and the tag response content, wherein the tag response content is determined by semantic understanding of the conversation information using the target large model based on prompt words determined by the adjustment demand information.

[0167] According to embodiments of this disclosure, the first large-model-based interactive device further includes: an experience session interface display module and an experience response content output module.

[0168] The Experience Session Interface Display Module is used to display the experience session interface.

[0169] The Experience Response Content Output Module is used to respond to the experience input information input into the experience session interface, utilize the updated large model to perform semantic understanding of the experience input information, and output the experience response content.

[0170] According to embodiments of this disclosure, the first large-model-based interactive device further includes a session attribute information display module.

[0171] The session attribute information display module is used to display session attribute information. The target object determines the session information to be displayed by selecting the session attribute information. The session attribute information represents at least one of the following attribute information corresponding to the session information: time attribute, response strategy, and adjustment requirement information.

[0172] According to embodiments of this disclosure, the strategy tag further characterizes the response content as the configuration response content output to the user object when the configuration keyword matches the specified keyword in the input information; wherein, the first large-model-based interaction device further includes an update module.

[0173] The update module is used to update at least one of the configuration keywords or configuration response content based on the configuration update information indicated by the interaction behavior.

[0174] According to embodiments of this disclosure, the first large-model-based interactive device further includes an adjustment progress information display module.

[0175] The adjustment progress information display module is used to display adjustment progress information, which includes at least one of the following: sample utilization rate, which represents the training samples used in the process of adjusting the response strategy of the large model; and response content interaction data, which represents the statistical results of the adjusted response content among multiple response contents, and the target object performs interactive behavior in response to the adjusted response content.

[0176] Figure 9 A block diagram of a large-model-based interactive device according to another embodiment of the present disclosure is shown schematically.

[0177] like Figure 9 As shown, the large-model-based interactive device may include a second large-model-based interactive device 900, which can be used on a user's terminal device. The second large-model-based interactive device 900 includes a receiving module 910 and a target response content acquisition module 920.

[0178] The receiving module 910 is used to receive target input information from the user object.

[0179] The target response content acquisition module 920 is used to perform semantic understanding on the target input information using a specified large model, output the target response content, and push the target response content to the target object; wherein the specified large model is determined based on the response strategy adjustment of the large model using training samples, and the training samples are determined according to the interaction method based on the large model provided in the embodiments of this disclosure.

[0180] According to embodiments of this disclosure, the target response content acquisition module is configured to perform the following operations using a specified large model: make a response strategy decision based on historical interaction information related to the target object and target input information to obtain a retrieval enhancement strategy; perform a retrieval task based on the retrieval enhancement strategy and target input information to obtain target resources; and perform semantic understanding based on target resources and input information to output target response content.

[0181] Figure 10 A schematic block diagram of an artificial intelligence agent according to an embodiment of the present disclosure is shown.

[0182] In embodiments of this disclosure, such as Figure 10 As shown, the AI ​​agent 1000 may include an input module 1010, a processing module 1020, and an output module 1030.

[0183] Input module 1010 is used to receive input information;

[0184] The processing module 1020 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 according to the embodiments of this disclosure.

[0185] Output module 1030 is used to output the output information obtained by the processing module.

[0186] According to embodiments of this disclosure, the input module 1010 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 1000 can understand and process. The input module 1010 is the primary link for the AI ​​agent 1000 to interact with the outside world, enabling the AI ​​agent 1000 to efficiently and accurately obtain necessary "sensory" information from the outside world and respond to this information.

[0187] In the example, input module 1010 can input the session information, target input information, etc., described above.

[0188] In the example, processing module 1020 is the core support for the AI ​​agent 1000's ability to handle complex tasks. Processing module 1020 can execute the large model-based interaction methods described above.

[0189] In the example, the performance of the processing module 1020 is closely related to the large model on which the AI ​​agent 1000 is based. To fully leverage the capabilities of the large model, the internal structure of the processing module 1020 can be designed to be highly configurable and scalable to handle various types of tasks and requirements in real-world scenarios.

[0190] In the example, after the AI ​​agent 1000 acquires the required voice, the processing module 1020 can use a specified large model to process the target input information to obtain the target response content, and then pass the response content to the output module 1030.

[0191] 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. Once the AI ​​agent 1000 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.

[0192] In the example, output module 1030 can output the target response content.

[0193] The AI ​​agent 1000 according to the embodiments of this disclosure can simply and effectively improve the level of intelligence, as well as enhance flexibility and versatility.

[0194] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0195] 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.

[0196] 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.

[0197] 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.

[0198] Figure 11 A schematic block diagram of an example electronic device for implementing a large-model-based interaction method of embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0199] like Figure 11 As shown, the electronic device 1100 includes a computing unit 1101, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 1102 or a computer program loaded into a random access memory (RAM) 1103 from a storage unit 1108. The RAM 1103 may also store various programs and data required for the operation of the electronic device 1100. The computing unit 1101, ROM 1102, and RAM 1103 are interconnected via a bus 1104. An input / output (I / O) interface 1105 is also connected to the bus 1104.

[0200] Multiple components in electronic device 1100 are connected to I / O interface 1105, including: input unit 1106, such as keyboard, mouse, etc.; output unit 1107, such as various types of displays, speakers, etc.; storage unit 1108, such as disk, optical disk, etc.; and communication unit 1109, such as network card, modem, wireless transceiver, etc. Communication unit 1109 allows electronic device 1100 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0201] The computing unit 1101 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1101 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 1101 performs the various methods and processes described above, such as the large model-based interaction method. For example, in some embodiments, the large model-based interaction method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 1108. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 1100 via ROM 1102 and / or communication unit 1109. When the computer program is loaded into RAM 1103 and executed by the computing unit 1101, one or more steps of the large model-based interaction method described above may be performed. Alternatively, in other embodiments, computing unit 1101 may be configured to perform large-model-based interactive methods by any other suitable means (e.g., by means of firmware).

[0202] 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.

[0203] 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.

[0204] 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.

[0205] 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).

[0206] 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.

[0207] 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.

[0208] 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.

[0209] 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. A large model-based interaction method, comprising: displaying conversation information, the conversation information including reply content of input information of a user object and a policy mark corresponding to the reply content, the policy mark being used to prompt a target object to adopt a reply policy corresponding to the reply content, and a large model processing the input information according to the reply policy to obtain the reply content; determining adjustment demand information corresponding to the reply policy according to an interaction behavior of the target object for the reply content, the adjustment demand information describing an adjustment demand intention of the target object for the reply content; determining a training sample according to the conversation information and the adjustment demand information, wherein the training sample is used to adjust a reply policy of the large model to obtain an updated large model.

2. The method of claim 1, wherein, the policy mark includes a retrieval enhancement mark, the retrieval enhancement mark representing retrieval result data in the reply content determined by the large model according to an initial resource obtained through retrieval; wherein the determining of the adjustment demand information corresponding to the reply policy according to the interaction behavior of the target object for the reply content comprises: determining a target resource according to the interaction behavior for the retrieval result data; and determining retrieval adjustment demand information according to the target resource and the retrieval result data, the retrieval adjustment demand information being used to prompt the large model to determine new reply content according to the target resource.

3. The method of claim 2, wherein, the displaying of the conversation information comprises: displaying a source mark representing the initial resource at a position corresponding to the retrieval result data, wherein the target object determines an association relationship between the initial resource and the retrieval result data according to the source mark.

4. The method of claim 1, wherein, the policy mark includes a retrieval enhancement mark, the retrieval enhancement mark representing retrieval result data in the reply content determined by the large model according to an initial resource obtained through retrieval, and the conversation information further includes a source mark representing an association relationship between the retrieval result data and the initial resource; wherein the determining of the adjustment demand information corresponding to the reply policy according to the interaction behavior of the target object for the reply content comprises: determining an initial resource corresponding to the retrieval result data according to a selection operation for the source mark; updating resource data of the initial resource according to an update operation for the initial resource to obtain a target resource, wherein the adjustment demand information includes the target resource.

5. The method of any one of claims 2 to 4, wherein, the training sample further includes context conversation content and label reply content corresponding to the adjustment demand information, the label reply content including label retrieval result data determined based on the target resource, and the context conversation content being determined based on a selection operation of the target object for reply content in conversation information; wherein the reply policy of the large model is adjusted based on the following operations: performing semantic understanding on the context conversation content and the adjustment demand information based on the target resource by using the large model to output a sample retrieval intention; retrieve a sample retrieval resource from a preset resource based on the sample retrieval intention, and perform a reply task on the input information by using the large model to obtain sample reply content; and train the large model according to the difference between the sample reply content and the label reply content to obtain an updated large model.

6. The method of any one of claims 1 to 4, wherein, The adjustment requirement information includes at least one of the following: semantic style requirement information, shielding content requirement information, expected reply content, and business type adjustment information. The expected reply content represents the expected reply content of the target object for the input information, and the business type adjustment information describes a target business type matched with the reply content determined based on the input information.

7. The method of claim 1, wherein, Adjust the reply strategy of the large model based on the following operations: perform semantic understanding on the conversation information according to the adjustment requirement intention by using the large model to obtain sample reply content; and train the large model according to the difference between the sample reply content and the label reply content, wherein the label reply content is a prompt word determined based on the adjustment requirement information and determined by performing semantic understanding on the conversation information by using the target large model.

8. The method of claim 1 or 7, further comprising: displaying an experience conversation interface; performing semantic understanding on experience input information input for the experience conversation interface by using the updated large model to output experience reply content in response to the experience input information.

9. The method of claim 1, further comprising: displaying conversation attribute information, wherein the target object determines the displayed conversation information by selecting the conversation attribute information, and the conversation attribute information represents at least one of the following attribute information corresponding to the conversation information: time attribute, reply strategy, and adjustment requirement information.

10. The method of claim 1, wherein, The strategy mark also represents that the reply content is configuration reply content output to the user object in a case where the configuration keyword matches a specified keyword in the input information. The method further comprises: updating at least one of the configuration keyword or the configuration reply content according to configuration update information indicated by the interaction behavior.

11. The method of claim 1, further comprising: displaying adjustment progress information, wherein the adjustment progress information includes at least one of the following: sample utilization rate, representing a training sample already used in the process of adjusting the reply strategy of the large model; reply content interaction data, representing a statistical result of an adjustment reply content in a plurality of the reply contents, wherein the target object performs an interaction behavior on the adjustment reply content.

12. A large model-based interaction method, comprising: receiving target input information of a user object; performing semantic understanding on the target input information by using a specified large model to output target reply content and push the target reply content to the target object; wherein the specified large model is determined based on reply strategy adjustment of a large model by using a training sample, and the training sample is determined according to the method of any one of claims 1 to 11.

13. The method of claim 12, wherein, The semantic understanding of the target input information by using the specified large model comprises the following operations performed by using the specified large model: making a reply strategy decision according to the historical interaction information related to the target object and the target input information, to obtain a retrieval enhancement strategy; performing a retrieval task according to the target input information based on the retrieval enhancement strategy, to obtain a target resource; and performing semantic understanding according to the target resource and the input information, to output target reply content.

14. An interactive device based on a large model, comprising: a display module configured to display conversation information, the conversation information comprising reply content for input information of a user object and a strategy mark corresponding to the reply content, the strategy mark being used to prompt a target object to use a reply strategy corresponding to the reply content, and a large model being used to process the input information according to the reply strategy to obtain the reply content; a first determination module configured to determine adjustment demand information corresponding to the reply strategy according to an interaction behavior of the target object for the reply content, the adjustment demand information describing an adjustment demand intention of the target object for the reply content; a second determination module configured to determine a training sample according to the conversation information and the adjustment demand information, wherein the training sample is used to adjust a reply strategy of the large model to obtain an updated large model.

15. An interactive device based on a large model, comprising: a receiving module configured to receive target input information of a user object; a target reply content obtaining module configured to perform semantic understanding of the target input information by using a specified large model, to output target reply content, and to push the target reply content to the target object; wherein the specified large model is determined based on a reply strategy adjustment of a large model by using a training sample, and the training sample is determined according to the method of any one of claims 1 to 11.

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

17. An electronic device, comprising: at least one processor; and a memory connected in communication with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the method of any one of claims 1 to 13.

18. A non-transitory computer readable storage medium having stored thereon computer instructions, wherein, The computer instructions are used to enable the computer to execute the method of any one of claims 1 to 13.

19. A computer program product comprising a computer program, the computer program being executed by a processor to implement the method of any one of claims 1 to 13.

19. A computer program product comprising a computer program, the computer program being executed by a processor to implement the method of any one of claims 1 to 13.

Images