System fine tuning method and device based on dialogue interaction model, equipment and medium

Abstract

The invention relates to a system fine tuning method and device based on a dialogue interaction model, equipment and a medium, and the method comprises the steps: receiving demand information in a natural language form, the demand information being used for indicating the adjustment of the target performance of a target system; semantic analysis is carried out on the demand information, a performance adjustment instruction is automatically generated, and the performance adjustment instruction carries target performance and an adjustment target; according to the adjustment target, determining a parameter object used for influencing target performance in the target system and adjustment information for the parameter object; and adjusting the parameter object based on the adjustment information to correspondingly obtain an updated target system. According to the method, the parameters of the target system can be automatically adjusted based on the natural language dialogue form, the flexibility and the response speed of the system can be improved, the method does not strongly depend on experience and technical ability of research and development personnel, the requirement change of a user can be quickly responded in practical application, and the development or operation and maintenance efficiency of the target system is improved.

Description

Technical Field

[0001] The present disclosure relates to the fields of artificial intelligence and system development, and in particular to a system fine-tuning method, apparatus, device, and medium based on a conversational interaction model. Background Art

[0002] When developing or maintaining various intelligent systems, developers often rely on their experience to adjust the underlying framework, operational logic, or algorithm parameters. This consumes significant debugging time and heavily relies on the developer's personal experience, making it difficult to quickly respond to changing user needs.

[0003] For example, taking the intelligent driving system of a vehicle as an example, during the development stage or operation and maintenance stage, there will be demands for performance improvement or improvement of some defects, such as the need to improve the accuracy of obstacle recognition, the need to improve the smoothness of obstacle avoidance operations, and the need to improve the execution reliability of safety response logic under collision events. Solutions to these needs generally require personnel from multiple departments to jointly think and find better improvement directions, and R&D personnel such as engineers or programmers to manually locate specific codes, and then try to adjust them by adding new programming codes or modifying the original codes. This is not only time-consuming, but also highly dependent on the experience and technical capabilities of the R&D personnel. In actual applications, it is difficult to quickly respond to changes in user needs. Summary of the Invention

[0004] To overcome the problems existing in the related art, the present disclosure provides a system fine-tuning method, apparatus, device and medium based on a conversational interaction model.

[0005] According to a first aspect of an embodiment of the present disclosure, a system fine-tuning method based on a conversational interaction model is provided. The method comprises: receiving demand information in natural language format, the demand information being used to instruct adjustments to target performance of a target system; performing semantic analysis on the demand information to automatically generate performance adjustment instructions, the performance adjustment instructions containing target performance and adjustment targets; determining, based on the adjustment targets, parameter objects in the target system that affect the target performance and adjustment information for the parameter objects; and adjusting the parameter objects based on the adjustment information to obtain an updated target system.

[0006] In some embodiments, determining, based on the adjustment target, a parameter object in the target system for influencing the target performance and adjustment information for the parameter object includes:

[0007] Determine, based on the constructed mapping relationship library, the target performance, and the adjustment target, a parameter object that matches the adjustment target of the target performance and adjustment information for the parameter object; wherein the mapping relationship stored in the mapping relationship library includes: a correspondence between the performance of the target system and an influencing parameter; the influencing parameter is a system parameter that affects the operation logic or model structure of the target system; or

[0008] An analysis is performed based on at least one of the historical training records and the experiential knowledge base of the above-mentioned target system to obtain a parameter object that matches the adjustment target of the above-mentioned target performance and adjustment information for the above-mentioned parameter object; wherein the above-mentioned historical training records cover the corresponding change relationship between system performance and influencing parameters, and the above-mentioned experiential knowledge base covers prior experiential knowledge on how to adjust the influencing parameters corresponding to the system performance adjustment.

[0009] In some embodiments, the above-mentioned parameter object that matches the adjustment target of the above-mentioned target performance and the adjustment information for the above-mentioned parameter object are determined based on the constructed mapping relationship library, the above-mentioned target performance and the above-mentioned adjustment target, including: based on the above-mentioned target performance, matching in the above-mentioned mapping relationship library to obtain matching influencing parameters; according to the above-mentioned adjustment target, the above-mentioned matching influencing parameters are screened to obtain the parameter object that matches the above-mentioned adjustment target; according to the historical impact relationship of the above-mentioned parameter object on the above-mentioned target performance or according to the corresponding change relationship of how each performance stored in the above-mentioned mapping relationship is affected by the corresponding influencing parameters, preliminary adjustment information that matches the above-mentioned adjustment target is determined; according to the above-mentioned preliminary adjustment information, adjustment information for the above-mentioned parameter object is determined.

[0010] In some embodiments, an analysis is performed based on at least one of the historical training records and the experience knowledge base of the above-mentioned target model to obtain a parameter object that matches the adjustment target of the above-mentioned target performance and adjustment information for the above-mentioned parameter object, including: an analysis is performed based on at least one of the historical training records and the experience knowledge base of the above-mentioned target system to obtain a matching influencing parameter for influencing the above-mentioned target performance; based on the corresponding change relationship in the above-mentioned historical training record and at least one of the prior experience knowledge in the above-mentioned experience knowledge base, the parameter object that matches the above-mentioned adjustment target in the above-mentioned matching influencing parameter and the preliminary adjustment information that matches the above-mentioned adjustment target are determined; based on the above-mentioned preliminary adjustment information, the adjustment information for the above-mentioned parameter object is determined.

[0011] In some embodiments, determining adjustment information for the parameter object based on the preliminary adjustment information; and adjusting the parameter object based on the adjustment information to obtain an updated target system includes:

[0012] Determine the preliminary adjustment information as adjustment information for the parameter object; adjust the parameter object based on the preliminary adjustment information to obtain an adjusted target system; perform performance verification on the adjusted target system based on at least one of user feedback information and the model's own verification strategy, and iteratively optimize the preliminary adjustment information based on the performance verification result until the obtained target adjustment information meets the adjustment target, and use the adjusted target system corresponding to the target adjustment information as the updated target system; or

[0013] Based on the above preliminary adjustment information, a parameter adjustment test is performed on the above target system, and the above preliminary adjustment information is iterated according to the performance verification result of the target system updated after the parameter adjustment test, until the obtained target adjustment information meets the above adjustment target, and the above target adjustment information is determined as the adjustment information for the above parameter object; and based on the above target adjustment information, the above parameter object is adjusted to obtain the updated target system; wherein, during the iterative process of the parameter adjustment test, each adjustment step is recorded and the step undo operation is supported, so as to cancel all parameter adjustment operations for the above target system during the parameter adjustment test after the completion of the above parameter adjustment test.

[0014] In some embodiments, based on a preset target function module, according to the above adjustment target, the parameter objects in the above target system used to affect the above target performance and the adjustment information for the above parameter objects are determined; the above preset target function module and the above target system are independent entities, or the above preset target function module is integrated into the above target system. The above preset target function module can call the sub-model in the above target system, and by calling the parameter adjustment analysis sub-model in the above sub-model, according to the above adjustment target, the parameter objects in the above target system used to affect the above target performance and the adjustment information for the above parameter objects are determined; or, the above preset target function module has a built-in parameter adjustment semantic analysis module, and based on the above parameter adjustment semantic analysis module, according to the above adjustment target, the parameter objects in the above target system used to affect the above target performance and the adjustment information for the above parameter objects are determined; or, the above preset target function module and the parameter adjustment semantic analysis module are two independent function modules, and by calling the parameter adjustment semantic analysis module, according to the above adjustment target, the parameter objects in the above target system used to affect the above target performance and the adjustment information for the above parameter objects are determined. Among them, the above-mentioned parameter adjustment semantic analysis module has at least one of the following functions: the above-mentioned parameter adjustment semantic analysis module can analyze the above-mentioned target system and construct a mapping relationship library, the above-mentioned parameter adjustment semantic analysis module can obtain the historical training records of the above-mentioned target system, and the above-mentioned parameter adjustment semantic analysis module is configured with an experience knowledge base.

[0015] In some embodiments, the target system is an intelligent model system or the conversational interaction model itself. Parameters affecting the target system's performance include system parameters that influence the target system's operational logic, iteration performance, or model structure. These parameters may include hyperparameters during model training or configuration parameters within the model structure.

[0016] According to the second aspect of the embodiment of the present disclosure, a system fine-tuning device based on a conversational interaction model is provided. The above-mentioned system fine-tuning device includes: a demand receiving module, a semantic analysis module, an adjustment information determination module and a system adjustment module. The above-mentioned demand receiving module is used to receive demand information in the form of natural language, and the above-mentioned demand information is used to indicate the adjustment of the target performance of the target system. The above-mentioned semantic analysis module is used to perform semantic analysis on the above-mentioned demand information and automatically generate performance adjustment instructions, and the above-mentioned performance adjustment instructions carry: target performance and adjustment target. The above-mentioned adjustment information determination module is used to determine the parameter objects in the above-mentioned target system that are used to affect the above-mentioned target performance and the adjustment information for the above-mentioned parameter objects according to the above-mentioned adjustment target. The above-mentioned system adjustment module is used to send the above-mentioned adjustment information to the target system for autonomous modification, or to adjust the parameter objects of the above-mentioned target system based on the above-mentioned adjustment information, and obtain the corresponding updated target system.

[0017] According to a third aspect of an embodiment of the present disclosure, an electronic device is provided, comprising: a processor; a memory for storing executable instructions of the processor; the processor, configured to read the executable instructions from the memory and execute the instructions to implement the system fine-tuning method based on the conversation interaction model provided in the first aspect of the present disclosure.

[0018] According to a fourth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided, on which computer program instructions are stored. When the program instructions are executed by a processor, the steps of the system fine-tuning method based on the dialogue interaction model provided in the first aspect of the present disclosure are implemented.

[0019] The technical solutions provided by the embodiments of the present disclosure may have the following beneficial effects:

[0020] Based on the dialogue interaction model, it is possible to perform semantic analysis on the demand information in the natural language form that indicates the adjustment of the target performance of the target system, automatically generate performance adjustment instructions, determine the parameter objects used to affect the target performance in the target system and the adjustment information for the above-mentioned parameter objects based on the performance adjustment instructions, and adjust the parameter objects to obtain an updated target system; overall, the parameters of the target system are automatically adjusted based on the natural language dialogue form. Compared with the related technology that requires R&D personnel to manually modify the code, this improves the flexibility and response speed of the system, does not strongly rely on the experience and technical capabilities of the R&D personnel, and can quickly respond to changes in user needs in actual applications, thereby improving the efficiency of development or operation and maintenance of the target system.

[0021] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0023] Figure 1 The present invention is a flowchart of a method for fine-tuning a system based on a conversation interaction model according to an exemplary embodiment.

[0024] Figure 2 The figure is a schematic diagram of the sequential interactions between a user, a conversation interaction model, and a target system in a system fine-tuning method based on a conversation interaction model according to an exemplary embodiment.

[0025] Figure 3A is a detailed implementation flowchart of step S130 according to an exemplary embodiment.

[0026] Figure 3B is a detailed implementation flowchart of step S130 according to another exemplary embodiment.

[0027] Figure 4 The present invention is a block diagram showing a system fine-tuning device based on a conversation interaction model according to an exemplary embodiment.

[0028] Figure 5 It is a block diagram of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0029] Exemplary embodiments will be described in detail below with reference to the accompanying drawings.

[0030] It should be pointed out that the relevant embodiments and drawings are only for describing exemplary embodiments provided by the present disclosure, rather than all embodiments of the present disclosure, and it should not be understood that the present disclosure is limited to the relevant exemplary embodiments.

[0031] It should be noted that the terms "first", "second", etc. used in this disclosure are only used to distinguish different steps, devices or modules, etc. The relevant terms neither represent any specific technical meanings nor indicate the order or interdependence between them.

[0032] It should be noted that the modifications of the terms "one", "a plurality of", and "at least one" used in the present disclosure are illustrative rather than restrictive. Unless otherwise clearly indicated in the context, they should be understood as "one or more".

[0033] It should be noted that the term "and / or" used in this disclosure to describe an association relationship between associated objects generally indicates the existence of at least three types of association relationships. For example, "A and / or B" can represent at least three types of association relationships: the existence of A alone, the existence of both A and B, and the existence of B alone.

[0034] It should be noted that the steps described in the method embodiments of the present disclosure may be performed in different orders and / or in parallel. Unless otherwise specified, the scope of the present disclosure is not limited by the order in which the steps are described in the relevant embodiments.

[0035] It should be noted that all actions of acquiring signals, information or data in the present disclosure are carried out in compliance with the corresponding data protection laws and policies of the country where they are located and with the authorization given by the owner of the corresponding device.

[0036] Exemplary Methods

[0037] Figure 1 The present invention is a flowchart of a method for fine-tuning a system based on a conversation interaction model according to an exemplary embodiment. Figure 2 The figure is a schematic diagram of the sequential interactions between a user, a conversation interaction model, and a target system in a system fine-tuning method based on a conversation interaction model according to an exemplary embodiment.

[0038] Reference Figure 1 As shown, the system fine-tuning method based on the dialogue interaction model provided by the embodiment of the present disclosure can be applied to an electronic device installed with a dialogue interaction model. The above method includes the following steps: S110, S120, S130 and S140.

[0039] In step S110 , demand information in a natural language format is received, where the demand information is used to instruct adjustment of target performance of the target system.

[0040] Reference Figure 2As shown, in some implementation scenarios, a user initiates a conversation with an electronic device installed with a conversation interaction model. In the electronic device, based on the conversation interaction model, the user's demand information for adjusting the target performance of the target system expressed in natural language is received (corresponding to step S110).

[0041] Requirements information can include direct input of performance-related parameter adjustment instructions or performance adjustment instructions. These requirements are generally expressed as performance adjustment instructions. These requirements can cover, but are not limited to, one or more performance dimensions such as system response speed, algorithm accuracy, and resource utilization.

[0042] In some specific scenarios, receiving demand information in natural language may include: receiving the demand information in natural language via a dialog box within a conversational interaction model; and importing the demand information in natural language into the dialog box in at least one of the modalities of text, voice, image, or video. In multimodal scenarios, the conversational interaction model may include functional modules or sub-models capable of modal conversion or multimodal processing.

[0043] In some embodiments, the target system is an intelligent model system or the conversational interaction model itself. Parameters affecting the target system's performance include system parameters that influence the target system's operational logic, iteration performance, or model structure. These parameters may include hyperparameters during model training or configuration parameters within the model structure.

[0044] In some application scenarios of the present disclosure, the conversational interaction model and the target system to be tuned are different entities, such as an intelligent model system. In such application scenarios, the conversational interaction model can understand the performance adjustment requirements, determine the parameter objects and adjustment information to be adjusted, and interact with the target system to adjust and update the parameters of the target system, thereby enabling the conversational interaction model to control the parameter updates of the external system. The conversational interaction model uses the target system's system architecture and algorithms to learn about the target system, and generates adjustment instructions based on the requirements (such instructions can be for a third-party code tool to automatically intervene in the target system to modify parameters, directly provide reference parameters, or replace the code involved in parameter adjustment in the target system). The adjustment instructions are then analyzed to determine the parameter objects in the target system that affect the target performance and the adjustment information for these parameter objects. This translates the user requirements into the underlying execution logic of the target system, thereby controlling the target system to adjust and update parameters based on the adjustment information.

[0045] In other application scenarios, the conversational interaction model is integrated with the target system to be performance-tuned. In these application scenarios, the conversational interaction model, as a functional module integrated into the target system, can understand performance tuning requirements, determine the parameters and information to be adjusted, and interact with the target system to adjust and update the parameters of the target system. This means that the conversational interaction model can control parameter updates of its own target system. The target system is the conversational interaction model itself, or the two are in a contained relationship. In this scenario, the large language model itself contains some algorithmic parameters (such as structural parameters) that can be dynamically modified during regular training. Furthermore, parameters that are generally not included in regular training (such as hyperparameters) require manual modification by industry professionals. Based on the solutions provided in the embodiments of this disclosure, ports / interfaces or functions that allow code parameters to be modified are added to the large language model, enabling automated, self-tuning of the large language model.

[0046] The parameters that can be adjusted are not limited to the structural parameters of the model. They can also be used to adjust hyperparameters, that is, system parameters that would not be affected by normal dialogue and training (these hyperparameters are usually manually adjusted in the existing technology, and include, but are not limited to: the model's learning rate, the number of heads in the multi-head attention mechanism, the model dimension, etc.). Of course, the solution disclosed in this disclosure also has the ability to adjust common parameters (such as weights and biases, which can be automatically adjusted through dataset training).

[0047] In the two exemplary application scenarios described above, the relevant knowledge about the target system learned based on the conversational interaction model can be used to verify the relevant code of the target system after the parameters are modified.

[0048] The system fine-tuning method provided in the embodiments of the present disclosure can be applied to various fields and various broad scenarios or segmented processing scenarios. For example, the target system can be but is not limited to: image processing systems, video processing systems, vehicle intelligent driving systems, operating systems of smart terminals (such as mobile phones, tablets, wearable devices, smart watches, smart bracelets, etc.), control systems of intelligent robots, etc.

[0049] The aforementioned conversational interaction model can be any natural language understanding model, such as, but not limited to, a large language model, a semantic recognition model, or a multimodal processing AI model. The aforementioned target system can be an artificial intelligence model (such as a large language model, a graphics processing model, a video processing model, or an analytical computing model), or any data processing functional module, third-party tool, or data processing system.

[0050] Taking the automotive sector as an example, scenarios for performance fine-tuning of target systems include, but are not limited to, improving the accuracy and efficiency of file mapping matching according to the AUTOSAR specification, improving the accuracy of obstacle recognition, reducing the computational cost of obstacle recognition, improving the smoothness of obstacle avoidance operations, improving the accuracy of vehicle collision detection, and improving the reliability of safety response logic in vehicle collision events. Specific performance fine-tuning scenarios in other sectors can be referenced and are not listed here.

[0051] For example, in one exemplary scenario, the performance improvement requirement is to improve the accuracy of file mapping matching according to the AUTOSAR specification. The target system is a third-party tool, such as a vehicle AUTOSAR mapping tool. The dialogue interaction model is a large language model, which is capable of communicating and interacting with the vehicle AUTOSAR mapping tool. Specifically, the large language model can be given permission to directly modify the AUTOSAR mapping tool's code, or the large language model can send parameter adjustment commands to the AUTOSAR mapping tool, which then updates itself according to these commands.

[0052] In another exemplary scenario, the target system is a large language model, and the conversation interaction model includes: a first target function module and a second target function module integrated in the target system, steps S110 and S120 are executed based on the first target function module, and step S130 is executed based on the second target function module. In some embodiments, the second target function module is configured with permissions to modify the large language model; or the second target function module sends relevant instructions for parameter adjustment to the function functions / modules / interfaces in the large language model that can modify parameters, and the function functions / modules / interfaces in the large language model that can modify parameters perform self-update of the large language model (i.e., the target system). In an exemplary use scenario of the system fine-tuning method provided in an embodiment of the present disclosure, by integrating relevant target function modules that support fine-tuning of system parameters in the large language model, relevant functions that support automatic fine-tuning of system parameters of the large language model are realized.

[0053] In step S120, semantic analysis is performed on the demand information to automatically generate a performance adjustment instruction. The performance adjustment instruction carries: target performance and adjustment target.

[0054] In some embodiments, reference Figure 2 As shown, the conversational interaction model performs semantic analysis on the demand information and automatically generates a performance adjustment instruction. The performance adjustment instruction carries: target performance and adjustment target. For example, the conversational interaction model includes: a first target function module and a second target function module.

[0055] In step S120, semantic analysis is performed on the demand information based on the first target function module to automatically generate performance adjustment instructions; the first target function module and the target system are independent entities, or the first target function module is integrated into the target system.

[0056] The first target function module has a built-in semantic analysis function, which performs semantic analysis on the demand information based on the semantic analysis function and automatically generates performance adjustment instructions; for example, the first target function module is a pre-trained semantic analysis model; or

[0057] The first target function module can call the sub-model in the target system, perform semantic analysis on the demand information by calling the semantic analysis sub-model in the sub-model, and automatically generate performance adjustment instructions; or,

[0058] The first target function module can call a semantic analysis service, perform semantic analysis on the demand information based on the semantic analysis service, and automatically generate performance adjustment instructions; wherein the semantic analysis service is a service provided by a target object different from the target system.

[0059] Target performance can be expressed in different granularity forms, and adjustment targets can be generalized or refined. As a limited example, in one scenario, the target performance is the accuracy of AUTOSAR file mapping, and the adjustment target is to improve accuracy. In another scenario, the target performance is the accuracy of AUTOSAR file mapping, and the adjustment target is to increase accuracy to above 90%. In another scenario, the target performance is the fuzzy matching performance during the AUTOSAR file mapping process, and the adjustment target is to achieve the maximum fuzzy matching. In another scenario, the target performance is to improve the smoothness of obstacle avoidance operations, and the adjustment target is to improve smoothness.

[0060] In step S130, according to the adjustment target, a parameter object in the target system for influencing the target performance and adjustment information for the parameter object are determined.

[0061] In some embodiments, reference Figure 2 As shown, the conversational interaction model determines, based on the adjustment target, parameter objects in the target system that affect the target performance and adjustment information for the parameter objects. Specifically, a preset target function module in the conversational interaction model implements step S130, and the preset target function module is, for example, a second target function module.

[0062] In some embodiments, based on a preset target function module (eg, a second target function module), according to the adjustment target, a parameter object in the target system for influencing the target performance and adjustment information for the parameter object are determined.

[0063] The preset target function module and the target system are independent entities, or the preset target function module is integrated into the target system.

[0064] The preset target function module can call the sub-model in the target system, and by calling the parameter adjustment analysis sub-model in the sub-model, determine the parameter objects in the target system that are used to affect the target performance and the adjustment information for the parameter objects according to the adjustment target; or

[0065] The preset target function module has a built-in parameter adjustment semantic analysis module, which determines the parameter objects in the target system that affect the target performance and the adjustment information for the parameter objects based on the parameter adjustment semantic analysis module and the adjustment target; or

[0066] The above-mentioned preset target function module and parameter adjustment semantic analysis module are two independent function modules. By calling the above-mentioned parameter adjustment semantic analysis module, according to the above-mentioned adjustment target, the parameter objects in the above-mentioned target system used to affect the above-mentioned target performance and the adjustment information for the above-mentioned parameter objects are determined.

[0067] Among them, the above-mentioned parameter adjustment semantic analysis module has at least one of the following functions: the above-mentioned parameter adjustment semantic analysis module can analyze the above-mentioned target system and construct a mapping relationship library, the above-mentioned parameter adjustment semantic analysis module can obtain the historical training records of the above-mentioned target system, and the above-mentioned parameter adjustment semantic analysis module is configured with an experience knowledge base.

[0068] Figure 3A is a detailed implementation flowchart of step S130 according to an exemplary embodiment.

[0069] In some embodiments, in the above step S130, based on the above adjustment target, the parameter object in the above target system used to affect the above target performance and the adjustment information for the above parameter object are determined, including: based on the constructed mapping relationship library, the above target performance and the above adjustment target, determining the parameter object that matches the adjustment target of the above target performance and the adjustment information for the above parameter object.

[0070] Among them, reference Figure 2As indicated by the dashed arrow, the mapping relationship library can be constructed by pre-learning the target system using the conversational interaction model. The mapping relationships stored in the mapping relationship library include: the correspondence between the performance of the target system and influencing parameters; these influencing parameters include, but are not limited to, system parameters that affect the target system's operational logic or model structure.

[0071] In some embodiments, reference Figure 3A As shown, in the above step S130, based on the constructed mapping relationship library, the above target performance and the above adjustment target, the parameter object matching the adjustment target of the above target performance and the adjustment information for the above parameter object are determined, including the following steps: S311, S312, S313 and S314.

[0072] In step S311, based on the above target performance, matching is performed in the mapping relationship library to obtain matching influencing parameters.

[0073] In some embodiments, the matching influence parameters obtained for the target performance may involve results in one or more adjustment directions. For example, the matching results for the target performance of accuracy may include: a first matching influence parameter that improves accuracy but increases cost, a second matching influence parameter that reduces accuracy but improves reliability, a third matching influence parameter that improves accuracy and can improve reliability, etc.

[0074] In step S312, the matching influencing parameters are screened according to the adjustment target to obtain parameter objects that match the adjustment target.

[0075] For example, if the adjustment target is to improve accuracy, then by screening the matching influence parameters, the parameter objects that match the adjustment target are: the first matching influence parameter that improves accuracy but increases cost, and the third matching influence parameter that improves accuracy and can improve reliability.

[0076] In step S313, preliminary adjustment information matching the adjustment target is determined based on the historical impact relationship of the parameter object on the target performance or the corresponding change relationship of how each performance is affected by the corresponding impact parameter stored in the mapping relationship.

[0077] In some embodiments, the historical impact relationship of each parameter object on the above-mentioned target performance is stored in a designated database; or stored in the above-mentioned mapping relationship library, that is, the ternary correspondence relationship stored in the mapping relationship library also belongs to a type of mapping relationship, and the ternary correspondence relationship is: target performance-influencing parameters of target performance-corresponding change relationship of how one or some influencing parameters cause a certain target performance to change (for example, it can be a quantitative change relationship).

[0078] By matching in historical impact relationships or corresponding change relationships, preliminary adjustment information that matches the adjustment target is determined.

[0079] For example, in the first matching influencing parameter that improves accuracy but increases cost, different parameter values ​​correspond to different degrees of improvement in accuracy. Based on the historical impact relationship or the corresponding change relationship, the specific parameter to be adjusted and the specific adjustment range are determined. This information serves as preliminary adjustment information. In the absence of explicit user instructions, certain strategies can be preset for the adjustment target of improving accuracy. If the adjustment target is achieved but other performance is compromised, a solution that achieves the adjustment target with less detrimental effects on other performance can be adopted. In other words, this strategy can be employed in the process of determining the specific parameter to be adjusted and the specific adjustment range based on the historical impact relationship or the corresponding change relationship.

[0080] In step S314, adjustment information for the parameter object is determined based on the preliminary adjustment information.

[0081] In some embodiments, the preliminary adjustment information can be directly determined as the adjustment information for the parameter object. In other embodiments, after performing system parameter iterative testing or verification based on the preliminary adjustment information, the adjustment information for the above parameter object is obtained, and the final adjustment information may be different from the preliminary adjustment information.

[0082] In the embodiment including steps S311 to S314, matching influencing parameters related to the target performance are first obtained based on the mapping relationship library; then the matching influencing parameters are screened according to the adjustment target; then preliminary adjustment information is determined; and then the adjustment information is determined based on the preliminary adjustment information; overall, the mapping relationship library learned and constructed in advance is utilized to realize the determination of the adjustment information.

[0083] Figure 3B is a detailed implementation flowchart of step S130 according to another exemplary embodiment.

[0084] In other embodiments, in the above-mentioned step S130, based on the above-mentioned adjustment target, the parameter object in the above-mentioned target system used to affect the above-mentioned target performance and the adjustment information for the above-mentioned parameter object are determined, including: analyzing at least one of the historical training records and the experience knowledge base of the above-mentioned target system to obtain the parameter object that matches the adjustment target of the above-mentioned target performance and the adjustment information for the above-mentioned parameter object.

[0085] The above historical training records include the corresponding change relationship between system performance and influencing parameters, and the above experience knowledge base includes prior experience knowledge on how to adjust the influencing parameters in response to system performance adjustments.

[0086] In some embodiments, reference Figure 3B As shown, in the above step S130, analysis is performed based on at least one of the historical training records and the experience knowledge base of the above target model to obtain a parameter object that matches the adjustment target of the above target performance and adjustment information for the above parameter object, including the following steps: S321, S322 and S323.

[0087] In step S321, an analysis is performed based on at least one of the historical training records and the experience knowledge base of the target system to obtain a matching influencing parameter for influencing the target performance.

[0088] In step S322, based on the corresponding change relationship in the above historical training records and at least one of the prior experience knowledge in the above experience knowledge base, the parameter object in the above matching influencing parameters that matches the above adjustment target and the preliminary adjustment information that matches the above adjustment target are determined.

[0089] In step S323, adjustment information for the parameter object is determined based on the preliminary adjustment information.

[0090] In the embodiment including steps S321 to S323, matching influencing parameters related to the target performance are obtained based on at least one of historical training records and an empirical knowledge base; the matching influencing parameters are then screened based on the adjustment target; preliminary adjustment information is then determined; and adjustment information is then determined based on the preliminary adjustment information. Overall, the adjustment information is determined using at least one of historical training records and an empirical knowledge base.

[0091] In step S140, the parameter objects are adjusted based on the adjustment information to obtain an updated target system.

[0092] Reference Figure 2 As shown by the dot-dash arrow in the figure, in some implementation scenarios, the conversational interaction model sends parameter adjustment control instructions carrying adjustment information to the target system directly or indirectly (for example, through some tools, interfaces, or intermediate transmission media). The target system adjusts its own parameter objects based on the adjustment information to obtain an updated target system.

[0093] Reference Figure 2 As shown by the double-dotted arrow in the middle, in other implementation scenarios, the conversational interaction model is assigned the authority to adjust parameters of the target system. Therefore, the conversational interaction model adjusts the parameter objects of the target system according to the adjustment information to obtain an updated target system.

[0094] In some embodiments, the implementation details of S314 or S323, S140 in the above step S130 are as follows:

[0095] One exemplary approach involves adjusting system parameters directly based on preliminary adjustment information; then verifying the system based on user feedback or the system's own validation strategy, and iterating to ultimately achieve an updated target system that meets the adjustment objectives. This approach, which validates the effects of parameter adjustments based on actual conditions (such as user feedback) or the system's own validation strategy, helps explore and iterate to achieve a target system with improved performance after parameter adjustments.

[0096] Another exemplary approach is to conduct a system parameter adjustment test based on preliminary adjustment information, and continuously iteratively update the parameters and the corresponding system during the test process until the system effect meets the adjustment target. After the test is completed, all parameter adjustment operations during the test period will be canceled; the optimal parameters obtained from the iterative test will be actually adjusted to obtain an updated target system that ultimately meets the adjustment target. This approach first adjusts the system parameters and verifies the effect based on the test simulation. Finally, the previous parameter adjustment operations are canceled after passing the test, realizing a goal-oriented experimental process and applying the iterative results to the parameter adjustment of the target system, thereby improving the reliability of the performance of the target system that meets user needs.

[0097] For example, in some embodiments, determining adjustment information for the parameter object based on the preliminary adjustment information; and adjusting the parameter object based on the adjustment information to obtain an updated target system includes:

[0098] Determine the preliminary adjustment information as adjustment information for the parameter object; adjust the parameter object based on the preliminary adjustment information to obtain an adjusted target system; perform performance verification on the adjusted target system based on at least one of user feedback information and the model's own verification strategy, and iteratively optimize the preliminary adjustment information based on the performance verification result until the obtained target adjustment information meets the adjustment target, and use the adjusted target system corresponding to the target adjustment information as the updated target system; or

[0099] Based on the above preliminary adjustment information, a parameter adjustment test is performed on the above target system, and the above preliminary adjustment information is iterated according to the performance verification result of the target system updated after the parameter adjustment test, until the obtained target adjustment information meets the above adjustment target, and the above target adjustment information is determined as the adjustment information for the above parameter object; and based on the above target adjustment information, the above parameter object is adjusted to obtain the updated target system; wherein, during the iterative process of the parameter adjustment test, each adjustment step is recorded and the step undo operation is supported, so as to cancel all parameter adjustment operations for the above target system during the parameter adjustment test after the completion of the above parameter adjustment test.

[0100] In summary, in the embodiment including the above steps S110 to S140, based on the dialogue interaction model, the demand information in the natural language form indicating the adjustment of the target performance of the target system can be semantically analyzed, and performance adjustment instructions can be automatically generated. Based on the performance adjustment instructions, the parameter objects used to affect the target performance in the target system and the adjustment information for the above parameter objects are determined, and the parameter objects are adjusted to obtain an updated target system; as a whole, the parameters of the target system are automatically adjusted based on the natural language dialogue form. Compared with the related technology that requires R&D personnel to manually modify the code, the flexibility and response speed of the system are improved, and it does not strongly rely on the experience and technical capabilities of the R&D personnel. In actual applications, it can quickly respond to changes in user needs and improve the efficiency of development or operation and maintenance of the target system.

[0101] Exemplary devices

[0102] Figure 4 The present invention is a block diagram showing a system fine-tuning device based on a conversation interaction model according to an exemplary embodiment.

[0103] Reference Figure 4 As shown, the system fine-tuning device 400 based on the dialogue interaction model provided by the embodiment of the present disclosure includes: a demand receiving module 410, a semantic analysis module 420, an adjustment information determination module 430 and a system adjustment module 440.

[0104] The demand receiving module 410 is used to receive demand information in a natural language format, where the demand information is used to instruct adjustment of target performance of the target system.

[0105] The semantic analysis module 420 is used to perform semantic analysis on the demand information and automatically generate a performance adjustment instruction. The performance adjustment instruction carries: target performance and adjustment target.

[0106] The adjustment information determination module 430 is configured to determine, according to the adjustment target, a parameter object in the target system that is used to affect the target performance and adjustment information for the parameter object.

[0107] The demand receiving module 440 is used to receive demand information in a natural language format, where the demand information is used to instruct adjustment of target performance of the target system.

[0108] The system adjustment module 440 is configured to send the adjustment information to the target system for autonomous modification, or to adjust the parameter objects of the target system based on the adjustment information, thereby obtaining an updated target system.

[0109] Reference Figure 2As shown by the dot-dash arrow in the figure, in some implementation scenarios, the conversational interaction model sends parameter adjustment control instructions carrying adjustment information to the target system directly or indirectly (for example, through some tools, interfaces, or intermediate transmission media). The target system adjusts its own parameter objects based on the adjustment information to obtain an updated target system.

[0110] Reference Figure 2 As shown by the double-dotted arrow in the middle, in other implementation scenarios, the conversational interaction model is assigned the authority to adjust parameters of the target system. Therefore, the conversational interaction model adjusts the parameter objects of the target system according to the adjustment information to obtain an updated target system.

[0111] The system fine-tuning device provided in this embodiment can perform semantic analysis on demand information in natural language form indicating adjustment of the target performance of the target system based on a dialogue interaction model, automatically generate performance adjustment instructions, determine the parameter objects used to affect the target performance in the target system and adjustment information for the above-mentioned parameter objects based on the performance adjustment instructions, and adjust the parameter objects to obtain an updated target system; overall, the parameters of the target system are automatically adjusted based on the natural language dialogue form. Compared with the related technology that requires R&D personnel to manually modify the code, the flexibility and response speed of the system are improved, and it does not strongly rely on the experience and technical capabilities of the R&D personnel. In actual applications, it can quickly respond to changes in user needs and improve the efficiency of development or operation and maintenance of the target system.

[0112] More details of this embodiment can be found in the description of the first embodiment, which will not be repeated here. Each functional module of this embodiment can be implemented by software, hardware, or a combination of software and hardware, and each functional module can be split or merged.

[0113] Exemplary electronic devices

[0114] Figure 5 FIG. 5 is a block diagram of an electronic device according to an exemplary embodiment. The electronic device 500 may be a vehicle controller, a vehicle terminal, a vehicle computer, or other types of electronic devices.

[0115] Reference Figure 5 As shown, electronic device 500 may include at least one processor 510 and memory 520. Processor 510 may execute instructions stored in memory 520. Processor 510 is communicatively connected to memory 520 via a data bus. In addition to memory 520, processor 510 may also be communicatively connected to input device 530, output device 540, and communication device 550 via the data bus.

[0116] The processor 510 may be any conventional processor, such as a commercially available CPU. The processor may also include a graphics processing unit (GPU), a field programmable gate array (FPGA), a system on chip (SOC), an application specific integrated circuit (ASIC), or a combination thereof.

[0117] The memory 520 may be implemented by any type of volatile or non-volatile memory device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk, or optical disk.

[0118] In an embodiment of the present disclosure, executable instructions are stored in the memory 520, and the processor 510 can read the executable instructions from the memory 520 and execute the instructions to implement all or part of the steps of system fine-tuning based on the dialogue interaction model described in any of the above exemplary embodiments.

[0119] Exemplary computer-readable storage media

[0120] In addition to the above methods and apparatuses, exemplary embodiments of the present disclosure may also be a computer program product or a computer-readable storage medium storing the computer program product. The computer product includes computer program instructions that can be executed by a processor to implement all or part of the steps described in any of the methods in the above exemplary embodiments.

[0121] The computer program product can be written in any combination of one or more programming languages ​​to write program code for performing the operations of the embodiments of the present application, including object-oriented programming languages ​​such as Java, C++, etc., as well as conventional procedural programming languages ​​such as "C" or similar programming languages ​​and scripting languages ​​(e.g., Python). The program code can be executed entirely on the user computing device, partially on the user computing device, as a standalone software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0122] The computer-readable storage medium can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can include, for example, but is not limited to, a system, device or component of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination thereof. More specific examples of readable storage media include: a static random access memory (SRAM) electrically connected with one or more wires, an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk or an optical disk, or any suitable combination thereof.

[0123] Other embodiments of the present disclosure will readily occur to those skilled in the art after considering the specification and practicing the present disclosure. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, with the true scope and spirit of the present disclosure being indicated by the following claims.

[0124] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes can be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A system fine-tuning method based on a conversational interaction model, characterized in that: include: receiving demand information in a natural language format, wherein the demand information is used to indicate adjustment of target performance of a target system; Performing semantic analysis on the demand information to automatically generate a performance adjustment instruction, wherein the performance adjustment instruction carries: target performance and adjustment target; determining, according to the adjustment target, a parameter object in the target system for influencing the target performance and adjustment information for the parameter object; Based on the adjustment information, the parameter object is adjusted to obtain an updated target system.

2. The system fine-tuning method according to claim 1, characterized in that: Determining, according to the adjustment target, a parameter object in the target system for influencing the target performance and adjustment information for the parameter object, including: Determine, based on the constructed mapping relationship library, the target performance, and the adjustment target, a parameter object that matches the adjustment target of the target performance and adjustment information for the parameter object; wherein the mapping relationship stored in the mapping relationship library includes: a correspondence between the performance of the target system and an influencing parameter; the influencing parameter is a system parameter that affects the operation logic or model structure of the target system; or An analysis is performed based on at least one of the historical training records and the experiential knowledge base of the target system to obtain a parameter object that matches the adjustment target of the target performance and adjustment information for the parameter object; wherein the historical training records cover the corresponding change relationship between system performance and influencing parameters, and the experiential knowledge base covers prior experiential knowledge on how to adjust the influencing parameters in response to system performance adjustments.

3. The system fine-tuning method according to claim 2, characterized in that: The determining, based on the constructed mapping relationship library, the target performance, and the adjustment target, a parameter object matching the adjustment target of the target performance and adjustment information for the parameter object includes: Based on the target performance, matching is performed in the mapping relationship library to obtain matching influencing parameters; According to the adjustment target, the matching influencing parameters are screened to obtain parameter objects that match the adjustment target; Determining preliminary adjustment information that matches the adjustment target based on a historical influence relationship of the parameter object on the target performance or a corresponding change relationship of how each performance is affected by the corresponding influencing parameter stored in the mapping relationship; Adjustment information for the parameter object is determined according to the preliminary adjustment information.

4. The system fine-tuning method according to claim 2, characterized in that: Analyzing at least one of the historical training records and the experience knowledge base of the target model to obtain a parameter object that matches the adjustment target of the target performance and adjustment information for the parameter object, including: Analyzing at least one of a historical training record and an experience knowledge base of the target system to obtain a matching influencing parameter for influencing the target performance; Determining, based on at least one of the corresponding change relationship in the historical training record and the prior experience knowledge in the experience knowledge base, a parameter object in the matching influencing parameter that matches the adjustment target and preliminary adjustment information that matches the adjustment target; Adjustment information for the parameter object is determined according to the preliminary adjustment information.

5. The system fine-tuning method according to claim 3 or 4, characterized in that: determining adjustment information for the parameter object based on the preliminary adjustment information; Adjusting the parameter object based on the adjustment information to obtain an updated target system includes: Determine the preliminary adjustment information as adjustment information for the parameter object; adjust the parameter object based on the preliminary adjustment information to obtain an adjusted target system; perform performance verification on the adjusted target system based on at least one of user feedback information and a model's own verification strategy, and iteratively optimize the preliminary adjustment information according to the performance verification result until the obtained target adjustment information meets the adjustment target, and use the adjusted target system corresponding to the target adjustment information as the updated target system; or A parameter adjustment test is performed on the target system based on the preliminary adjustment information, and the preliminary adjustment information is iterated according to the performance verification result of the target system updated after the parameter adjustment test, until the obtained target adjustment information meets the adjustment target, and the target adjustment information is determined as the adjustment information for the parameter object; and based on the target adjustment information, the parameter object is adjusted to obtain the updated target system; wherein, during the iterative process of the parameter adjustment test, each adjustment step is recorded and a step undo operation is supported, so as to cancel all parameter adjustment operations on the target system during the parameter adjustment test after the parameter adjustment test is completed.

6. The system fine-tuning method according to any one of claims 1 to 4, characterized in that: Based on a preset target function module, and according to the adjustment target, determining a parameter object in the target system for affecting the target performance and adjustment information for the parameter object; the preset target function module and the target system are independent entities, or the preset target function module is integrated into the target system; The preset target function module can call the sub-model in the target system, and by calling the parameter adjustment analysis sub-model in the sub-model, determine the parameter objects in the target system that are used to affect the target performance and the adjustment information for the parameter objects according to the adjustment target; or The preset target function module has a built-in parameter adjustment semantic analysis module, which determines, based on the parameter adjustment semantic analysis module and according to the adjustment target, a parameter object in the target system that is used to affect the target performance and adjustment information for the parameter object; or The preset target function module and the parameter adjustment semantic analysis module are two independent function modules. By calling the parameter adjustment semantic analysis module, according to the adjustment target, the parameter object in the target system that affects the target performance and the adjustment information for the parameter object are determined; Among them, the parameter adjustment semantic analysis module has at least one of the following functions: the parameter adjustment semantic analysis module can analyze the target system and construct a mapping relationship library, the parameter adjustment semantic analysis module can obtain the historical training records of the target system, and the parameter adjustment semantic analysis module is configured with an experience knowledge base.

7. The system fine-tuning method according to any one of claims 1 to 4, wherein the target system is an intelligent model system or a conversational interaction model itself; Parameters affecting the performance of the target system include: system parameters that affect the operation logic, iteration effect or model structure of the target system; The types of influencing parameters include: hyperparameters during model training, or configuration parameters in the model structure.

8. A system fine-tuning device based on a conversational interaction model, characterized in that: include: a demand receiving module, configured to receive demand information in a natural language format, wherein the demand information is used to instruct adjustment of target performance of a target system; A semantic analysis module, configured to perform semantic analysis on the demand information and automatically generate a performance adjustment instruction, wherein the performance adjustment instruction carries: target performance and adjustment target; an adjustment information determination module, configured to determine, according to the adjustment target, a parameter object in the target system for influencing the target performance and adjustment information for the parameter object; The system adjustment module is used to send the adjustment information to the target system for autonomous modification, or to adjust the parameter objects of the target system based on the adjustment information, thereby obtaining an updated target system.

9. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to read the executable instructions from the memory and execute the instructions to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the program instructions are executed by a processor, the steps of the method described in any one of claims 1 to 7 are implemented.

Images