Training method and device of classification model, vehicle-side control equipment, vehicle, storage medium and program product
By independently constructing training sample data and optimizing multi-classification models, the problems of low training efficiency and high cost of traditional text classification models are solved, achieving efficient and low-cost model training and real-time optimization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING JINKANG NEW ENERGY VEHICLE CO LTD
- Filing Date
- 2026-03-16
- Publication Date
- 2026-06-05
AI Technical Summary
Traditional text classification models are inefficient to train and require manual data annotation and training dataset construction, resulting in high training costs.
By generating a set of judgment instructions and inputting it into a large model for classification, and combining user instructions and the results of multi-classification models to construct target training sample data, the construction of autonomous training sample data and model optimization are achieved, avoiding manual intervention.
It improves the training efficiency of classification models, reduces training costs, ensures the accuracy of training data and real-time optimization of models, and enables autonomous training and online updates.
Smart Images
Figure CN122153451A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of intelligent vehicle technology, and in particular to a method, apparatus, vehicle-side control device, vehicle, storage medium, and program product for training a classification model. Background Technology
[0002] With the development of artificial intelligence technology, text classification models, as a core foundational task of intelligent models, have been widely applied in various fields such as internet services, information security, smart offices, smart furniture, and smart vehicles. Traditional methods for training text classification models include: first, technicians manually collect and label text data for specific classification task scenarios; second, optimizing the initial dataset using data generalization and augmentation techniques; and finally, inputting the augmented dataset into the text classification model to be trained, completing the model's training and deployment.
[0003] However, traditional text classification models suffer from low training efficiency. Summary of the Invention
[0004] Based on this, this application addresses the aforementioned technical problems by providing a training method, apparatus, vehicle-side control device, vehicle, storage medium, and program product for a classification model that can improve training efficiency.
[0005] Firstly, this application provides a method for training a classification model, including:
[0006] A judgment instruction set is generated based on the input user command, and the judgment instruction set is input into the first large model for classification to obtain the first classification result; the judgment instruction set includes multiple judgment instructions and the definition content corresponding to each judgment instruction; the definition content is used to explain the specific explanation and functional scope of the corresponding judgment instruction classification;
[0007] The user command is input into the target multi-classification model for classification, and a second classification result is obtained.
[0008] Target training sample data is constructed based on the user instructions, the first classification result, the second classification result, and the training sample data. The target multi-classification model is then trained based on the target training sample data to obtain an optimized multi-classification model.
[0009] The training method of the above classification model does not require manual data annotation, manual construction of training datasets, or manual initiation of model training. That is, the above training method realizes the autonomous construction of training sample data and the autonomous training of target multi-classification model, which greatly improves the training efficiency of classification model and reduces training manpower costs.
[0010] In an optional embodiment of the first aspect, constructing target training sample data based on the user instruction, the first classification result, the second classification result, and the training sample data includes:
[0011] The first classification result and the second classification result are compared to obtain the third classification result;
[0012] Target training sample data is constructed based on the user instructions, the third classification result, and the training sample data.
[0013] The training process of the above classification model does not require manual data annotation or manual construction of training datasets. That is, the above training method realizes the autonomous construction and expansion of training sample data, reduces the annotation cost of training samples, and thus reduces the training cost of classification model.
[0014] In an optional embodiment of the first aspect, comparing the first classification result and the second classification result to obtain a third classification result includes:
[0015] If the first classification result and the second classification result are consistent, then the first classification result or the second classification result shall be used as the third classification result;
[0016] If the first classification result and the second classification result are inconsistent, the third classification result is generated according to the user instruction, the first classification result and the second classification result.
[0017] The above method, by judging the consistency between the first classification result and the second classification result, selects accurate classification results to add to the preset database. This can ensure to a certain extent that the preset database stores accurate classification results, thereby improving the training effect of the classification model based on the training sample data in the preset database.
[0018] In an optional embodiment of the first aspect, generating the third classification result based on the user instruction, the first classification result, and the second classification result includes:
[0019] A new judgment instruction is generated based on the user instruction, the judgment instruction corresponding to the first classification result, the second classification result, and the judgment instruction corresponding to the second classification result.
[0020] The new judgment instruction is input into the second major model for classification to obtain the third classification result.
[0021] The above training method describes how to generate a third classification result when the first and second classification results are inconsistent. That is, by re-concatenating a new judgment instruction containing more effective information, the classification is performed again. Based on the new judgment instruction, an accurate third classification result can be obtained. This ensures that accurate classification results are stored in the subsequent preset database, which can avoid training failures caused by erroneous data to a certain extent. It can also improve the efficiency and quality of subsequent training of classification models based on data in the preset database to a certain extent.
[0022] In an optional embodiment of the first aspect, after training the target multi-classification model based on the target training sample data, the method further includes:
[0023] Add the version number of the target multi-classification model to the preset database.
[0024] The method described above adds the version number of the classification model to a preset database, which makes it easier to search for training data later by directly searching for the version number of the classification model. This improves the efficiency of searching for training data and, consequently, the efficiency of constructing training sample data.
[0025] In an optional embodiment of the first aspect, the method further includes:
[0026] The performance of the optimized multi-class classification model is compared with that of the target multi-class classification model;
[0027] If the performance of the optimized multi-class classification model is higher than that of the target multi-class classification model, then the optimized multi-class classification model replaces the target multi-class classification model.
[0028] If the performance of the optimized multi-class classification model is lower than that of the target multi-class classification model, then the target training sample data is cleaned.
[0029] The above method, after improving the performance of the optimized multi-class classification model after training, can automatically update the corresponding target multi-class classification model, achieving supervised self-improvement of multi-class classification model performance, i.e., autonomous training of the classification model. Compared with traditional offline classification model training methods, the above method enables the system where the target multi-class classification model is deployed to achieve real-time online training, ensuring that the target multi-class classification model on the system can maintain excellent performance in real time to a certain extent. In addition, the above training method also introduces a rollback mechanism, that is, after the model's performance declines, the current training sample data can be screened, re-labeled, and updated to update the training sample data, and retraining autonomously until the optimal classification model with the best performance is obtained. This method can improve the quality of the trained classification model.
[0030] Secondly, this application also provides a training device for a classification model, comprising:
[0031] The first classification module is used to generate a set of judgment instructions based on the input user instructions, and input the set of judgment instructions into the first large model for classification to obtain the first classification result; the set of judgment instructions includes multiple judgment instructions and the definition content corresponding to each judgment instruction; the definition content is used to explain the specific explanation and functional scope of the classification of the corresponding judgment instruction;
[0032] The second classification module is used to input the user question into the target multi-classification model for classification and obtain the second classification result.
[0033] The training module is used to construct target training sample data based on the user instruction, the first classification result, the second classification result, and the training sample data, and to train the target multi-classification model based on the target training sample data to obtain an optimized multi-classification model.
[0034] Thirdly, this application also provides a vehicle-side control device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods described above.
[0035] Fourthly, this application also provides a vehicle including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described above.
[0036] Fifthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in any of the above aspects.
[0037] Sixthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in any of the above aspects.
[0038] Regarding the beneficial effects of any of the technical solutions in the second to sixth aspects mentioned above, refer to the beneficial effects of the corresponding technical solutions in the first aspect; repeated examples will not be listed here. Attached Figure Description
[0039] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0040] Figure 1 This is a schematic diagram of an optional application environment for the training method of the classification model in the embodiments of this application;
[0041] Figure 2 This is one of the optional flowcharts illustrating the training method of the classification model in the embodiments of this application;
[0042] Figure 3 This is a second optional flowchart illustrating the training method of the classification model in the embodiments of this application;
[0043] Figure 4 This is a schematic diagram of an optional process for constructing target training sample data in an embodiment of this application;
[0044] Figure 5 This is a schematic diagram of an optional process for generating the third classification result in an embodiment of this application;
[0045] Figure 6 This is a third optional flowchart illustrating the training method of the classification model in the embodiments of this application;
[0046] Figure 7 This is a fourth optional flowchart illustrating the training method of the classification model in the embodiments of this application;
[0047] Figure 8 This is a schematic diagram of an optional structure of a training device for a classification model in an embodiment of this application;
[0048] Figure 9 This is a schematic diagram of an optional internal structure of the vehicle-side control device in an embodiment of this application. Detailed Implementation
[0049] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of this application.
[0050] The terms "first," "second," etc., used in this application may be used to describe various elements, but these elements are not limited by these terms. These terms are used only to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.
[0051] The training method for the classification model provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed on the cloud or other network servers. A business system is installed on terminal 102, on which at least one classification model is deployed to support various intelligent applications. In practical applications, the business system can collect various user instructions based on daily data, and be guided by these user instructions to autonomously construct a training dataset and guide the classification model to autonomously train based on the constructed training dataset. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, drones, low-altitude aircraft, IoT devices, and portable wearable devices. IoT devices can be smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, projection devices, etc. Portable wearable devices can be smartwatches, smart bracelets, head-mounted devices, etc. Head-mounted devices can be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. Server 104 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides cloud computing services.
[0052] In one exemplary embodiment, such as Figure 2 As shown, a training method for a classification model is provided, which can be applied to... Figure 1 Taking the terminal in the example, the explanation includes:
[0053] S201, Generate a judgment instruction set based on the input user command, and input the judgment instruction set into the first large model for classification to obtain the first classification result.
[0054] The user instruction can be a user-generated question to be categorized (e.g., ...). Figure 3 As shown in the image, user commands are represented using Query, such as "It's a bit hot inside the car." The set of judgment commands includes multiple judgment commands related to the user command and the definition content corresponding to each judgment command (using...). Figure 3 As shown, "Prompt" represents the judgment instruction and its corresponding definition. Each judgment instruction guides the primary model in making a judgment, and the corresponding definition includes a detailed explanation of the classification and functional scope of the judgment instruction. The set of judgment instructions is used to ask questions of the primary model, enabling it to classify the same user instruction from different perspectives. The primary model can be a "teacher model" that can perform single-class and multi-class judgments based on the input judgment instructions. For example, see... Figure 3The training process of the classification model shown includes steps S1-S2. In this step, a Prompt set is generated from the Query. The Prompt set includes Prompt_1, Prompt_2, ..., Prompt_n, and Prompt_ALL. Prompt_1 includes judgment instruction 1 and its corresponding definition content 1; Prompt_2 includes judgment instruction 2 and its corresponding definition content 2; ..., Prompt_n includes judgment instruction n and its corresponding definition content n; and Prompt_ALL includes judgment instruction n+1 and its corresponding definition content n+1. Judgment instruction 1, judgment instruction 2, ..., judgment instruction n are binary classification judgment instructions, used to determine whether the user instruction (Query) belongs to a certain class. Judgment instruction n+1 is a multi-class judgment instruction, specifically used to determine which class the user instruction (Query) belongs to from multiple classifications. For example, judgment instruction 1 is used to determine whether a user instruction (Query) belongs to class A, judgment instruction 2 is used to determine whether a user instruction (Query) belongs to class B, ..., judgment instruction n is used to determine whether a user instruction (Query) belongs to class C; judgment instruction n+1 is used to determine which class (A, B, ..., C) the user instruction (Query) belongs to. Classes A, B, ..., C can be determined according to predefined classification rules, which are classification rules related to the user instruction. For example, if the user instruction is "activate the vehicle's seat heating function," then the classification rule is a rule based on the relevant functions of the vehicle, and the corresponding categories determined by this classification rule can include navigation, entertainment, and vehicle control. Definition content 1 includes the specific explanation and functional scope of judgment instruction 1 corresponding to class A, definition content 2 includes the specific explanation and functional scope of judgment instruction 2 corresponding to class B, ..., definition content n includes the specific explanation and functional scope of judgment instruction n corresponding to class C. Definition content n+1 includes the specific explanation and functional scope of judgment instruction n+1 corresponding to all classes (classes A, B, and C). The first classification result can include multiple single-view classification results obtained from multiple different single-view judgments (e.g.) Figure 3In the context of the algorithm, Answer_1, Answer_2, ..., Answer_n, Answer_1 represents the single-view classification result (e.g., "yes" or "no") obtained after performing a single-view binary classification judgment on the user command (Query) based on Prompt_1; Answer_2 represents the single-view classification result (e.g., "yes" or "no") obtained after performing a single-view binary classification judgment on the user command (Query) based on Prompt_2; and Answer_n represents the single-view classification result (e.g., "yes" or "no") obtained after performing a single-view binary classification judgment on the user command (Query) based on Prompt_n. Optionally, the first classification result may include multiple single-view classification results obtained from multiple different single-view judgments (e.g., ...). Figure 3 The full-view classification results obtained from the full-view judgment (e.g., Answer_1, Answer_2, ..., Answer_n) and Answer_n in the full-view judgment (e.g., Figure 3 In the context of Answer_ALL, Answer_ALL represents the overall classification result obtained after performing multiple different single-perspective multi-classification judgments on the user command (Query) based on Prompt_ALL (e.g., the user command belongs to class A among classes A, B, and C); optionally, the first classification result may include the overall classification result (e.g., ...). Figure 3 (Answer_ALL in the middle).
[0055] For example, firstly, the terminal can collect user commands from real users based on daily data from the business system, or guide a large model to generate input user commands. For example, the collected or generated user command query could be "start the seat heating function in the vehicle", "go to the nearest charging station", "turn on the air conditioning", "play music", etc. Secondly, for the same user command, the terminal can generate a series of judgment commands related to the same user command based on predefined classification rules (the classification rules here include multiple predefined categories, such as navigation, entertainment, and vehicle control), forming a judgment command set. Optionally, the terminal can also combine the large model to analyze the same user command and obtain multiple judgment commands related to the user command. For example, assuming the user command query is "turn on the seat heating function in the vehicle", the corresponding generated judgment command set Prompt includes the first judgment command Prompt_1 and definition content 1, the first judgment command Prompt_2 and definition content 2, and the first judgment command Prompt_3 and definition content 3. The first judgment instruction, Prompt_1, determines whether the query belongs to the vehicle control category, including definition 1 which explains that vehicle control includes functions such as air conditioning adjustment and seat adjustment. The second judgment instruction, Prompt_2, determines whether the query belongs to the navigation and map category, including definition 2 which explains that navigation and map includes functions such as destination setting, route planning, and map building. The third judgment instruction, Prompt_3, determines whether the query belongs to the entertainment and information category, including definition 3 which explains that entertainment and information includes functions such as music playback and video playback. The fourth judgment instruction, Prompt_4, determines whether the query belongs to the air conditioning control, navigation and map, or entertainment and information category, including definition 4 which includes the aforementioned definition 1, definition 2, and definition 3. Finally, the terminal can directly input the judgment instruction set into the first large model for classification, allowing multiple different judgment instructions to "query" the large model, resulting in classification results from multiple judgment perspectives, i.e., obtaining the first classification result. (See example two.) Figure 3As shown, the Prompt set (Prompt_1, Prompt_2, ..., Prompt_n, Prompt_ALL) is input into the teacher model A for classification, resulting in the classification results corresponding to Prompt_1 (Answer_1), Prompt_2 (Answer_2), ..., Prompt_n (Answer_n), and Prompt_ALL (Answer_ALL). Based on Example 2 above, the final first classification result can include one classification result or multiple classification results. Optionally, when the first classification result includes multiple classification results, it can include Answer_1, Answer_2, ..., Answer_n; optionally, the first classification result can also include Answer_1, Answer_2, ..., Answer_n, Answer_ALL; optionally, when the first classification result includes one classification result, the first classification result can include Answer_ALL. The above steps can be found in [reference needed]. Figure 3 The process of S1, S2 and S3.
[0056] S202, input the user command into the target multi-classification model for classification, and obtain the second classification result.
[0057] The target multi-classification model is an intelligent model deployed on the terminal. It can be an intelligent model to be classified or optimized, used to determine which category a user command belongs to.
[0058] For example, the terminal can directly input user commands into the target multi-classification model to be trained. After classifying the command, the target multi-classification model outputs a second classification result that includes the first classification result. For instance, assuming the user command Query is "turn on the seat heating function in the vehicle," the target multi-classification model, after classifying the user command, can obtain a second classification result of "Query belongs to the control category of the smart cockpit in the vehicle." The above steps can be found in [reference needed]. Figure 3 The process of S4 in China.
[0059] S203: Construct target training sample data based on user instructions, first classification results, second classification results, and training sample data, and train the target multi-classification model based on the target training sample data.
[0060] The training sample data includes training data collected by the terminal within a historical time period (including user commands collected within the historical time period, classification results corresponding to user commands, etc.), or manually annotated training data containing user commands and classification results. The training sample data may also include the version number of the corresponding classification model trained using the aforementioned training data; that is, this version number is used to mark the "source" of the corresponding classification result. The training sample data can be stored in a pre-set database, allowing for direct extraction of training sample data for each model training iteration.
[0061] For example, when the first classification result and the second classification result are obtained based on the aforementioned steps, since the first classification result and the second classification result are obtained by different classification models, and the performance of different classification models is different, one of the first classification result and the second classification result may not be accurate, or neither may be accurate. Therefore, before constructing the target training sample data, it is necessary to further determine which of the first classification result and the second classification result is accurate. The specific method for determining this can be by using other classification models (for example, see...). Figure 3 The second major model in step S6 reclassifies the data based on the user command and corresponding judgment command, obtaining a new classification result. Then, referring to this new classification result, it selects one from the first and second classification results that matches the new classification result as the accurate classification result. The accurate classification result (either the first or second classification result) and the corresponding user command are added to a preset database for training data accumulation. Once the accumulated training data in the preset database reaches the standard amount for training the model, the standard amount of training data can be extracted from the preset database to construct the target training sample data. When the terminal has constructed the target training sample data, it can be used to train the target multi-classification model, resulting in an optimized multi-classification model.
[0062] It should be noted that when training the target multi-class model based on the target training sample data, if the target multi-class model is an initial small model, seed data (a small amount of data, generally 50-200 records) can be extracted from the target training sample data as sample data for small model training; if the target multi-class model is not an initial small model, the target multi-class model can be trained once using the target training sample data generated above, and then the model can be self-trained.
[0063] In the training method of the above classification model, a judgment instruction set is generated based on the input user instructions. This set is then input into the first main model for classification to obtain the first classification result. The user instructions are also input into the target multi-classification model for classification to obtain the second classification result. Then, target training sample data is constructed based on the user instructions, the first classification result, the second classification result, and the training sample data. The target multi-classification model is then trained using this target training sample data to obtain an optimized multi-classification model. This training process eliminates the need for manual data annotation, manual construction of the training dataset, and manual initiation of model training. In other words, this training method achieves autonomous construction of training sample data and autonomous training of the target multi-classification model, greatly improving the training efficiency of the classification model and reducing training labor costs.
[0064] In one exemplary embodiment, a method for constructing target training sample data is provided, such as... Figure 4 As shown, the "constructing target training sample data based on user instructions, first classification results, second classification results, and training sample data" in S203 above includes:
[0065] S301, compare the first classification result and the second classification result to obtain the third classification result.
[0066] For example, when the first classification result and the second classification result are obtained based on the aforementioned steps, the first classification result and the second classification result can be compared to obtain a comparison result, and then a third classification result can be determined based on the comparison result. The above comparison method can include two scenarios. The first comparison method is that when the first classification result includes one classification result and the second classification result includes one classification result, the terminal can directly compare the first classification result and the second classification result to determine whether the first classification result and the second classification result are consistent. If the first classification result and the second classification result are consistent, it means that both the first classification result and the second classification result are correct, and either the first classification result or the second classification result can be used as the third classification result. If the first classification result and the second classification result are inconsistent, it means that one of the classification results is inaccurate, and a third classification result can be generated according to the user instruction, the first classification result, and the second classification result.
[0067] The method for generating a third classification result based on user instructions, the first classification result, and the second classification result includes: first, analyzing the user instructions and corresponding judgment instructions to determine which of the first and second classification results is the accurate classification result, and then using the accurate classification result as the third classification result. The analysis method includes using other classification models (e.g., see...). Figure 3The second major model in step S6 reclassifies the user command and its corresponding judgment command to obtain a new classification result. Then, the classification result that matches the new classification result in either the first or second classification result is taken as the accurate classification result. Secondly, a new judgment command is generated based on the user command, the first classification result, the judgment command corresponding to the first classification result, the second classification result, and the judgment command corresponding to the second classification result. This new judgment command is then input into the new major model or the first major model for classification to obtain a third classification result. For example, suppose the user command is "navigate to the nearest charging station." The set of judgment commands corresponding to this user command includes: a first judgment command (determining whether the user command belongs to the navigation category and the definition of the corresponding navigation category); a second judgment command (determining whether the user command belongs to the entertainment category and the definition of the corresponding entertainment category); and a third judgment command (determining whether the user command belongs to the vehicle control category and the definition of the corresponding vehicle control category). Suppose the first classification result output by the first major model is "the above user command belongs to the navigation category"; and the second classification result output by the target multi-classification model is "the above user command belongs to the entertainment category." Based on the first method described above, it can be directly analyzed that the judgment result of the target multi-classification model is incorrect. Therefore, the first classification result output by the first large model is directly used as the third classification result. Based on the second method described above, the user instruction "navigate to the nearest charging station", the first classification result "the above user instruction belongs to the navigation category", the definition content of the navigation category, the second classification result "the above user instruction belongs to the entertainment category", and the definition content of the entertainment category are input into the first large model for classification to obtain the third classification result, which indicates that "the above user instruction belongs to the navigation category".
[0068] When the first classification result includes multiple classification results, and the second classification result includes one classification result, the terminal can first select an accurate classification result from the first classification result including multiple classification results as the fourth classification result, and compare this fourth classification result with the second classification result. If the fourth classification result matches the second classification result, then either the fourth classification result or the second classification result is used as the third classification result; if the fourth classification result does not match the second classification result, then the third classification result can be generated based on the user instruction, the fourth classification result, and the second classification result. The method for generating the third classification result based on the user instruction, the fourth classification result, and the second classification result is the same as the method described above for generating the third classification result based on the user instruction, the first classification result, and the second classification result, and will not be repeated here. Additionally, the method for selecting an accurate classification result from the first classification result including multiple classification results is illustrated by example. See [link to documentation]. Figure 3As shown, the Prompt set (Prompt_1, Prompt_2, ..., Prompt_n, Prompt_ALL) is input into the teacher model A for classification, resulting in the classification results for Prompt_1 (Answer_1), Prompt_2 (Answer_2), ..., Prompt_n (Answer_n), and Prompt_ALL (Answer_ALL). Here, Prompt_1, Prompt_2, ..., Prompt_n are binary classification results, and Answer_ALL is a multi-class classification result. For multiple binary classification results (Prompt_1, Prompt_2, ..., Prompt_n), if any one of the binary classification results is "yes," then that classification result is the accurate classification result. If multiple binary classification results show "No", it indicates overlap between categories or uncertainty in the classification model, requiring further evaluation. In this case, the accurate classification result can be determined by combining the multi-class classification results (Answer_ALL). Specifically, if the multi-class classification results (Answer_ALL) indicate "Query belongs to class A", and the classification result (Answer_2) corresponding to Prompt_2 among the multiple classification results (Prompt_1, Prompt_2, ..., Prompt_n) also indicates "Query belongs to class A", then that classification result (Answer_2) is the accurate classification result. The above steps can be followed by... Figure 3 The above method, by judging the consistency between the first and second classification results, selects accurate classification results to add to the preset database. This can, to a certain extent, ensure that the preset database stores accurate classification results, thereby improving the training effect of the classification model based on the target training sample data.
[0069] S302, construct target training sample data based on user instructions, third classification results and training sample data.
[0070] The training sample data includes training data collected by the terminal during historical time periods (including user commands collected during historical time periods, classification results corresponding to user commands, etc.).
[0071] For example, after obtaining the corresponding third classification result based on the user command in the current round, the user command and the corresponding third classification result of the current round can be stored in a preset database to accumulate and expand the training sample data. When the accumulated data in the preset database reaches the standard data volume, the training data stored in the preset database for historical time periods (including user commands collected in historical time periods, classification results corresponding to user commands, etc.) can be extracted and combined with the user command and the third classification result of the current round to form the target training sample data, which is convenient for training the target multi-classification model based on the target training sample data in the later stage.
[0072] Optionally, when storing the training data of the current round (including the user commands and corresponding third-class classification results of the current round) into the preset database, the version number of the target multi-class classification model can also be added to the preset database accordingly, that is, the version number of the target multi-class classification model is bound to the corresponding training data. Referring to Table 1 below, the terminal can bind the user commands, corresponding third-class classification results, and corresponding version number of the target multi-class classification model for each round, and store them in the preset database as shown in Table 1 below to form training sample data. The above steps can be found in [reference needed]. Figure 3 The process of S7 in China.
[0073] Table 1
[0074] User commands Classification results Model version Instruction a Classification result a V1 Instruction b Classification result b V1 Instruction c Classification result c V2 …… …… ……
[0075] The version number of the target multi-class classification model serves as the source tag corresponding to the third classification result. This source tag is used to identify the classification model that obtained the third classification result, facilitating subsequent lookup of the required training data based on the version number, thereby improving the efficiency of constructing training sample data later. The version number of this target multi-class classification model can also be used for version rollback and data repair when the model's self-training performance is unsatisfactory. For details on version rollback and data repair, please refer to the subsequent embodiments; they will not be elaborated here.
[0076] It should be noted that before storing the training data into the preset database, it is possible to first determine whether there is duplicate data in the preset database. If it is determined that there is no duplicate data, the training data of the current round (the user instructions and the third classification results of the current round) can be stored into the preset database, so that the training sample data can be directly extracted from the preset database for model training later.
[0077] The training process of the above classification model does not require manual data annotation or manual construction of training datasets. That is, the above training method realizes the autonomous construction and expansion of training sample data, reduces the annotation cost of training samples, and thus reduces the training cost of classification model.
[0078] In one exemplary embodiment, a method is provided for generating a third classification result based on user instructions, a first classification result, and a second classification result, such as... Figure 5 As shown, the method includes:
[0079] S401, generate a new judgment instruction based on the user instruction, the judgment instruction corresponding to the first classification result, the second classification result, and the judgment instruction corresponding to the second classification result.
[0080] For example, based on the foregoing analysis, when a third classification result needs to be generated, it indicates that one of the first or second classification results is inaccurate. Therefore, a new judgment instruction can be reassembled to facilitate subsequent reclassification of the user instruction based on this new instruction, resulting in an accurate classification result. During this reassembly of the new judgment instruction, since the first and second classification results contain conflicting information, and the judgment instruction carries definition content, concatenating the user instruction, the first classification result, the judgment instruction corresponding to the first classification result, the second classification result, and the judgment instruction corresponding to the second classification result as a new judgment instruction allows this new instruction to contain more effective information. Therefore, when the second model is used for classification based on this new judgment instruction, an accurate classification result can be obtained.
[0081] S402, input the new judgment instruction into the second large model for classification, and obtain the third classification result.
[0082] The second major model can be a "teacher model" that performs single-class and multi-class classification based on the input judgment instructions. This second major model can be the same as the first major model. For example, after obtaining a new judgment instruction based on the aforementioned steps, the new instruction can be input into the second classification model for classification, resulting in a third classification result. The above steps can be found in [reference needed]. Figure 3 The process of S6 in China.
[0083] The above training method describes how to generate a third classification result when the first and second classification results are inconsistent. That is, by re-concatenating a new judgment instruction containing more effective information, the classification is performed again. Based on the new judgment instruction, an accurate third classification result can be obtained. This ensures that accurate classification results are stored in the subsequent preset database, which can avoid training failures caused by erroneous data to a certain extent. It can also improve the efficiency and quality of training the classification model based on the training sample data in the preset database to a certain extent.
[0084] In one exemplary embodiment, such as Figure 6 As shown, the aforementioned Figures 2-5 The method described in the embodiments further includes the following steps:
[0085] S501, compare the performance of the optimized multi-classification model with that of the target multi-classification model to determine whether the performance of the optimized multi-classification model is higher than that of the target multi-classification model. If yes, proceed to step S502; otherwise, proceed to step S503.
[0086] S502, replace the target multi-class classification model with the optimized multi-class classification model;
[0087] S503 performs data cleaning on the target training sample data.
[0088] Among them, the optimized multi-classification model can be a new version of the classification model, while the target multi-classification model can be an old version of the classification model.
[0089] For example, after training a target multi-class classification model using target training sample data to obtain an optimized multi-class classification model, the performance of the new version of the optimized multi-class classification model can be evaluated, and the performance of the new version of the optimized multi-class classification model and the old version of the target multi-class classification model can be compared. For example, the accuracy and learning rate of the optimized multi-class classification model and the target multi-class classification model can be detected. If the performance of the new version of the optimized multi-class classification model is higher than that of the target multi-class classification model, it means that the performance of the optimized multi-class classification model trained using the target training sample data is better (for example, the accuracy has increased from 60% to 85%). At this time, the new version of the optimized multi-class classification model can be directly used to replace the old version of the target multi-class classification model to complete the model update deployed on the system. If the performance of the new version of the optimized multi-class classification model is lower than that of the old version, it indicates that the performance of the optimized multi-class classification model trained using the target training sample data is worse (e.g., the accuracy improved from 60% to 50%). This also suggests that there may be errors in the new data. In this case, the data self-cleaning function can be activated, which extracts the problematic data from the preset database and uses a large model to re-classify it to correct the errors. This completes the data cleaning of the preset database. Then, the target multi-class classification model is retrained based on the cleaned training sample data until a better-performing model is trained. The above step S501 can be described as follows: Figure 3 The "model comparison" in process S10; the above step S502 can correspond to Figure 3 The "model update" in process S11; the above step S503 can correspond to Figure 3 The "Start Dataset Self-Cleaning" step in process S12.
[0090] It should be noted that when the performance of the new version of the optimized multi-class classification model is determined to be lower than that of the old version of the target multi-class classification model, the system can quickly identify which batch of training samples is problematic from a pre-set database using the version number of the classification model, without needing to inefficiently detect problematic data from historical data in the database. The problematic training samples are then repaired later, completing the cleaning and repair of the dataset containing the target training samples.
[0091] The above method, after improving the performance of the optimized multi-class classification model after training, can automatically update the corresponding target multi-class classification model, achieving supervised self-improvement of multi-class classification model performance, i.e., autonomous training of the classification model. Compared with traditional offline classification model training methods, the above method enables the system where the target multi-class classification model is deployed to achieve real-time online training, ensuring that the target multi-class classification model on the system can maintain excellent performance in real time to a certain extent. In addition, the above training method also introduces a rollback mechanism, that is, after the model's performance declines, the current training sample data can be screened, re-labeled, and updated to update the training sample data, and retraining autonomously until the optimal classification model with the best performance is obtained. This method can improve the quality of the trained classification model.
[0092] In summary, based on all the above embodiments, a method for training a classification model is also provided, such as... Figure 7 As shown, the method includes:
[0093] S601 generates a set of judgment instructions based on the input user instructions.
[0094] S602, input the judgment instruction set into the first large model for classification, and obtain the first classification result.
[0095] S603: Input the user command into the target multi-classification model for classification and obtain the second classification result.
[0096] S604. Compare the first classification result and the second classification result to determine whether the first classification result and the second classification result are consistent. If they are, proceed to step S605; otherwise, proceed to step S606.
[0097] S605, take the first classification result or the second classification result as the third classification result;
[0098] S606: Based on the user instruction, the judgment instruction corresponding to the first classification result, the second classification result, and the judgment instruction corresponding to the second classification result, generate a new judgment instruction, and input the new judgment instruction into the second large model for classification to obtain the third classification result;
[0099] S607 stores user instructions and third classification results in a preset database, and when the accumulated data in the preset database reaches the standard data volume, constructs the target training sample dataset based on the standard data volume.
[0100] S608, Train the target multi-classification model based on the target training sample dataset to obtain the trained optimized multi-classification model.
[0101] S609, compare the performance of the optimized multi-classification model with that of the target multi-classification model to determine whether the performance of the optimized multi-classification model is higher than that of the target multi-classification model. If yes, proceed to step S610; otherwise, proceed to step S611.
[0102] S610 replaces the target multi-class classification model with the optimized multi-class classification model;
[0103] S611 performs data cleaning on the target training sample dataset.
[0104] Each of the above steps has been described in the foregoing embodiments. For details, please refer to the foregoing descriptions. They will not be repeated here.
[0105] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.
[0106] Based on the same inventive concept, this application also provides a training apparatus for a classification model to implement the training method for the classification model described above. The solution provided by this apparatus is similar to the implementation described in the above method; therefore, the specific limitations of one or more training apparatus embodiments for classification models provided below can be found in the limitations of the classification model training method described above, and will not be repeated here.
[0107] In one exemplary embodiment, such as Figure 8 As shown, a training device for a classification model is provided, comprising:
[0108] The first classification module 11 is used to generate a judgment instruction set based on the input user instructions, and input the judgment instruction set into the first large model for classification to obtain the first classification result; the judgment instruction set includes multiple judgment instructions and the definition content corresponding to each judgment instruction; the definition content is used to explain the specific explanation and functional scope of the classification of the corresponding judgment instruction.
[0109] The second classification module 12 is used to input the user question into the target multi-classification model for classification and obtain the second classification result.
[0110] The training module 13 is used to construct target training sample data based on the user instruction, the first classification result, the second classification result and the training sample data, and to train the target multi-classification model based on the target training sample data to obtain an optimized multi-classification model.
[0111] In an exemplary embodiment, the training module 13 described above includes:
[0112] A comparison unit is used to compare the first classification result and the second classification result to obtain a third classification result;
[0113] The construction unit is used to construct target training sample data based on the user instructions, the third classification result, and the training sample data.
[0114] In an exemplary embodiment, the comparison unit includes:
[0115] The first classification subunit is used to take either the first classification result or the second classification result as the third classification result when the first classification result and the second classification result are consistent.
[0116] The second classification subunit is used to generate the third classification result based on the user instruction, the first classification result, and the second classification result when the first classification result and the second classification result are inconsistent.
[0117] In an exemplary embodiment, the second classification subunit is specifically used to generate a new judgment instruction based on the user instruction, the judgment instruction corresponding to the first classification result, the judgment instruction corresponding to the second classification result, and the judgment instruction corresponding to the second classification result, and input the new judgment instruction into the second large model for classification to obtain the third classification result.
[0118] In an exemplary embodiment, the training apparatus for the above-described classification model further includes:
[0119] The first update module is used to add the version number of the target multi-classification model to a preset database.
[0120] In an exemplary embodiment, the training apparatus for the above-described classification model further includes:
[0121] The comparison module is used to compare the performance of the optimized multi-classification model with that of the target multi-classification model.
[0122] The second update module is used to replace the target multi-class classification model with the optimized multi-class classification model when the performance of the optimized multi-class classification model is higher than the performance of the target multi-class classification model.
[0123] The cleaning module is used to clean the target training sample data when the performance of the optimized multi-classification model is lower than the performance of the target multi-classification model.
[0124] The modules in the training device of the above classification model can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.
[0125] In one exemplary embodiment, a vehicle-side control device is provided. This vehicle-side control device can be a terminal, and its internal structure diagram can be as follows: Figure 9 As shown, the vehicle-mounted control device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When the computer program is executed by the processor, it implements a training method for a classification model. The display unit of the vehicle-mounted control device is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the vehicle-side control equipment can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the housing of the vehicle-side control equipment, or external keyboards, touchpads, or mice, etc.
[0126] Those skilled in the art will understand that Figure 9 The structure shown is a block diagram of a portion of the structure related to the solution of this application, and does not constitute a limitation on the vehicle-side control device to which the solution of this application is applied. The specific vehicle-side control device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.
[0127] In one exemplary embodiment, a vehicle-mounted control device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.
[0128] In one exemplary embodiment, a vehicle is provided, including a memory and a processor, the memory storing a computer program that, when executed by the processor, implements the steps in the above-described method embodiments.
[0129] In one exemplary embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described method embodiments.
[0130] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.
[0131] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.
[0132] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program mentioned can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.
[0133] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.
[0134] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.
Claims
1. A method for training a classification model, characterized in that, The method includes: A judgment instruction set is generated based on the input user command, and the judgment instruction set is input into the first large model for classification to obtain the first classification result; the judgment instruction set includes multiple judgment instructions and the definition content corresponding to each judgment instruction; the definition content is used to explain the specific explanation and functional scope of the corresponding judgment instruction classification; The user command is input into the target multi-classification model for classification, and a second classification result is obtained. Target training sample data is constructed based on the user instructions, the first classification result, the second classification result, and the training sample data. The target multi-classification model is then trained based on the target training sample data to obtain an optimized multi-classification model.
2. The method according to claim 1, characterized in that, The step of constructing target training sample data based on the user instruction, the first classification result, the second classification result, and the training sample data includes: The first classification result and the second classification result are compared to obtain the third classification result; Target training sample data is constructed based on the user instructions, the third classification result, and the training sample data.
3. The method according to claim 2, characterized in that, The step of comparing the first classification result and the second classification result to obtain the third classification result includes: If the first classification result and the second classification result are consistent, then the first classification result or the second classification result shall be used as the third classification result; If the first classification result and the second classification result are inconsistent, the third classification result is generated according to the user instruction, the first classification result and the second classification result.
4. The method according to claim 3, characterized in that, The step of generating the third classification result based on the user instruction, the first classification result, and the second classification result includes: A new judgment instruction is generated based on the user instruction, the judgment instruction corresponding to the first classification result, the second classification result, and the judgment instruction corresponding to the second classification result. The new judgment instruction is input into the second major model for classification to obtain the third classification result.
5. The method according to claim 1, characterized in that, After training the target multi-classification model based on the target training sample data, the method further includes: Add the version number of the target multi-classification model to the preset database.
6. The method according to any one of claims 1-5, characterized in that, The method further includes: The performance of the optimized multi-class classification model is compared with that of the target multi-class classification model; If the performance of the optimized multi-class classification model is higher than that of the target multi-class classification model, then the optimized multi-class classification model replaces the target multi-class classification model. If the performance of the optimized multi-class classification model is lower than that of the target multi-class classification model, then the target training sample data is cleaned.
7. A training device for a classification model, characterized in that, The device includes: The first classification module is used to generate a set of judgment instructions based on the input user instructions, and input the set of judgment instructions into the first large model for classification to obtain the first classification result; the set of judgment instructions includes multiple judgment instructions and the definition content corresponding to each judgment instruction; the definition content is used to explain the specific explanation and functional scope of the classification of the corresponding judgment instruction; The second classification module is used to input the user command into the target multi-classification model for classification and obtain the second classification result. The training module is used to construct target training sample data based on the user instruction, the first classification result, the second classification result, and the training sample data, and to train the target multi-classification model based on the target training sample data to obtain an optimized multi-classification model.
8. A vehicle-end control device, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.
9. A vehicle comprising a memory and a processor, said memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.
11. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.