Adjustment method of vehicle question and answer model, vehicle response method and system and electric vehicle
By acquiring and analyzing the question-and-answer data between users and the vehicle-mounted question-and-answer model, and using the discriminant model to perform context judgment and adjustment, the problem of insufficient semantic understanding ability of large language models is solved, and accurate question-and-answer functions are achieved in complex environments.
Patent Information
- Application Number
- CN202510725291.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-03
- Publication Date
- 2025-09-19
AI Technical Summary
The update frequency of existing large language models is low and the update process is not flexible enough, resulting in poor semantic understanding ability, making it difficult to adapt to complex and changing driving environments, and affecting the accuracy of question-answering functions.
By obtaining the jth round of question-and-answer data and the n rounds of question-and-answer data before and after it, the discriminant model is used to make context judgments, and the vehicle-computer question-and-answer model is adjusted based on the execution accuracy to improve the semantic understanding ability.
In complex and changeable driving environments, the overall confidence of the vehicle-computer question-answering model and the accuracy of the question-answering function are improved.
Smart Images
Figure CN120671685A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of new energy vehicle technology, and in particular to a method for adjusting a vehicle-computer question-and-answer model, a vehicle-computer response method, a system, and an electric vehicle. Background Art
[0002] With the rise of big language models, various AI products have incorporated these capabilities to optimize their performance. For example, given that the intelligence of smart cockpits is becoming an increasingly important factor in evaluating the in-vehicle experience, big language models are being implemented in smart cockpits to support question-and-answer functions in new energy vehicles.
[0003] Semantic understanding is a crucial component of the question-and-answer functionality provided by smart cockpits, and its level of semantic understanding determines the overall experience. However, existing large language models suffer from infrequent and inflexible updates, making them difficult to adapt to complex and changing driving environments. This often results in unsatisfactory question-and-answer functionality.
[0004] Therefore, how to improve the semantic understanding ability of large language models and thus enhance the overall confidence of the model is an urgent problem that needs to be solved. Summary of the Invention
[0005] In order to solve or partially solve the technical problem that the language large model has poor semantic understanding ability and low overall confidence due to update problems, the present invention provides an adjustment method, a vehicle-computer question and answer model, a vehicle-computer response method, a system and an electric vehicle. The discriminant model in the language large model is used to perform context judgment on the j-th round of question and answer data and the n rounds of question and answer data before and after it, and the correctness of the execution result of the j-th round of question and answer data is judged in combination with more semantic information. The vehicle-computer question and answer model in the language large model is then adjusted using the correctly judged j-round question and answer data combined with the execution correctness to improve the semantic understanding ability of the language large model and the overall confidence of the vehicle-computer question and answer model, so that it can provide accurate question and answer functions even in complex and changeable driving environments.
[0006] To solve the above technical problems, the first aspect of the present invention discloses a method for adjusting a vehicle-based question-answering model, the method comprising:
[0007] Obtain the j-th round of question-and-answer data and the n-round question-and-answer data before and after it; where j represents any round of question-and-answer data, and n ≥ 1 is a positive integer. The question-and-answer data is generated between the user and the vehicle-computer question-and-answer model in the language model;
[0008] Input the j-th round of question-and-answer data and the n-round question-and-answer data before and after it into the discriminant model of the language large model to obtain an execution discriminant label and execution accuracy of the j-th round of question-and-answer data; wherein the execution discriminant label is used to determine whether the vehicle-computer question-and-answer model correctly responds to the j-th round of question-and-answer data, and the execution accuracy is used to indicate the degree to which the vehicle-computer question-and-answer model correctly responds to the j-round of question-and-answer data;
[0009] If the execution judgment label is a correct label, determine the sample weight of the j-th round of question-and-answer data with reference to the execution correctness;
[0010] The vehicle-computer question-answering model is adjusted using the j-th round question-answering data and the sample weight thereof to obtain the adjusted vehicle-computer question-answering model.
[0011] In the above embodiment, before obtaining the j-th round of question-and-answer data and the n rounds of question-and-answer data before and after it, the method further includes:
[0012] Get user question data;
[0013] Analyzing the user question data using the vehicle-computer question-answering model alone or in combination with historical question-answering data to obtain vehicle-computer execution data;
[0014] The user question data, the historical question and answer data, and the vehicle computer execution data are stored in a cache as a round of question and answer data.
[0015] In the above embodiment, obtaining the j-th round of question-and-answer data and the n rounds of question-and-answer data before and after it specifically includes:
[0016] During a preset off-peak period of vehicle usage, the j-th round of question-and-answer data and the n rounds of question-and-answer data before and after it are obtained from the cache.
[0017] In the above embodiment, the j-th round of question and answer data includes user question data, historical question and answer data and vehicle computer execution data;
[0018] After obtaining the execution discrimination label and execution accuracy of the j-th round of question-and-answer data, the method further includes:
[0019] If the execution determination tag is an error tag, obtaining prompt data for instructing the vehicle computer to correctly respond to the user question data in the j-th round of question-and-answer data;
[0020] The prompt data, the j-th round of question and answer data, and the n-round question and answer data before and after it are input into the discriminant model again to obtain vehicle-computer adjustment data for correctly responding to the user question data in the j-round of question and answer data; wherein the vehicle-computer adjustment data has the correct label and the correct execution degree.
[0021] In the above embodiment, after obtaining the vehicle computer adjustment data for correctly responding to the user question data in the j-th round of question-and-answer data, the method further includes:
[0022] The vehicle computer adjustment data is used to replace the vehicle computer execution data in the j-th round question and answer data in the cache.
[0023] In the above embodiment, after determining the sample weight of the j-th round of question-and-answer data with reference to the execution correctness, the method further includes:
[0024] Obtaining a parameter weight of the j-th round of question-and-answer data; wherein the parameter weight is used to represent the degree of influence of the j-th round of question-and-answer data on the update of model parameters of the vehicle-computer question-and-answer model;
[0025] The model adjustment weight is determined using the parameter weight and the sample weight.
[0026] The adjusting the vehicle-computer question-answering model using the j-th round question-answering data and the sample weight thereof to obtain the adjusted vehicle-computer question-answering model specifically includes:
[0027] The vehicle-computer question-answering model is adjusted using the j-th round question-answering data and the model adjustment weight thereof to obtain the adjusted vehicle-computer question-answering model.
[0028] A second aspect of the present invention discloses a vehicle-computer response method, the method comprising:
[0029] Collect user inquiry data;
[0030] Inputting the user query data into the vehicle-computer question-answering model adjusted according to the steps of the method of the first aspect to obtain vehicle-computer execution data;
[0031] The electric vehicle is controlled to respond to the user according to the vehicle computer execution data.
[0032] A third aspect of the present invention discloses a system for adjusting a vehicle-based question-answering model, the system comprising:
[0033] An acquisition module is used to obtain the j-th round of question-and-answer data and the n-round question-and-answer data before and after it; where j represents any round of question-and-answer data, and n ≥ 1 is a positive integer, and the question-and-answer data is the data generated between the user and the vehicle-computer question-and-answer model in the language model;
[0034] a discrimination module, configured to input the j-th round of question-and-answer data and the n rounds of question-and-answer data preceding and following it into the discrimination model of the language macro model, and obtain an execution discrimination label and execution accuracy of the j-th round of question-and-answer data; wherein the execution discrimination label is used to determine whether the vehicle-computer question-and-answer model correctly responds to the j-th round of question-and-answer data, and the execution accuracy is used to indicate the degree to which the vehicle-computer question-and-answer model correctly responds to the j-th round of question-and-answer data;
[0035] a determination module, configured to determine, if the execution discrimination label is a correct label, a sample weight of the j-th round of question-and-answer data with reference to the execution correctness;
[0036] An adjustment module is used to adjust the vehicle-computer question-answering model using the j-th round of question-answering data and its sample weights to obtain the adjusted vehicle-computer question-answering model.
[0037] In the above embodiment, the system further includes:
[0038] The acquisition module is used to obtain user question data;
[0039] An analysis module, configured to analyze the user question data using the vehicle-computer question-answering model alone or in combination with historical question-answering data to obtain vehicle-computer execution data;
[0040] The storage module is used to store the user question data, the historical question and answer data, and the vehicle computer execution data as a round of question and answer data into a cache.
[0041] In the above embodiment, the acquisition module is specifically configured to acquire the j-th round of question-and-answer data and the n rounds of question-and-answer data before and after it from the cache during a preset low-peak period of vehicle use.
[0042] In the above embodiment, the j-th round of question and answer data includes user question data, historical question and answer data and vehicle computer execution data;
[0043] The system also includes: a write-back module, which is used to obtain prompt data for instructing the vehicle computer to correctly respond to the user question data in the j-th round of question and answer data if the execution judgment label is an incorrect label; input the prompt data, the j-th round of question and answer data and the n rounds of question and answer data before and after it into the judgment model again to obtain vehicle computer adjustment data for correctly responding to the user question data in the j-th round of question and answer data; wherein the vehicle computer adjustment data has the correct label and the correct execution degree.
[0044] In the above embodiment,
[0045] The acquisition module is further configured to acquire a parameter weight of the j-th round of question-and-answer data; wherein the parameter weight is used to represent the degree of influence of the j-th round of question-and-answer data on the update of the model parameters of the vehicle-computer question-and-answer model;
[0046] The determination module is further configured to determine a model adjustment weight using the parameter weight and the sample weight.
[0047] In the above embodiment, the adjustment module is specifically used to adjust the vehicle-computer question-answering model using the j-th round question-answering data and the model adjustment weight thereof to obtain the adjusted vehicle-computer question-answering model.
[0048] A fourth aspect of the present invention discloses a vehicle-mounted response system, the system comprising:
[0049] A collection unit, used to collect user query data;
[0050] An input unit, configured to input the user query data into the vehicle-computer question-answering model adjusted according to the steps of the method of the first aspect, to obtain vehicle-computer execution data;
[0051] A control unit is used to control the electric vehicle to respond to the user according to the vehicle computer execution data.
[0052] In a fifth aspect of the present invention, an electric vehicle is disclosed, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the program.
[0053] Through one or more technical solutions of the present invention, the present invention has the following beneficial effects or advantages:
[0054] The present invention provides an adjustment method for a vehicle-computer question-and-answer model, a vehicle-computer response method, a system, and an electric vehicle, which obtain the j-th round of question-and-answer data and the n rounds of question-and-answer data before and after it, and then use a discriminant model in a large language model to perform context judgment on the j-th round of question-and-answer data and the n rounds of question-and-answer data before and after it, and combine more semantic information to judge the correctness of the execution result of the j-round question-and-answer data. Then, the vehicle-computer question-and-answer model in the large language model is adjusted using the correctly judged j-round question-and-answer data and the execution correctness to improve the semantic understanding ability of the large language model, improve the overall confidence of the vehicle-computer question-and-answer model, and enable it to provide accurate question-and-answer functions even in complex and changeable driving environments.
[0055] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are specifically listed below. BRIEF DESCRIPTION OF THE DRAWINGS
[0056] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present invention. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:
[0057] Figure 1 A schematic flow chart showing a method for adjusting a vehicle-based question-answering model according to an embodiment of the present invention is shown;
[0058] Figure 2 A schematic flow chart of a vehicle computer response method according to an embodiment of the present invention is shown;
[0059] Figure 3 A schematic structural diagram of an adjustment system for a vehicle-computer question-answering model according to an embodiment of the present invention is shown;
[0060] Figure 4 A schematic structural diagram of a vehicle-mounted response system according to an embodiment of the present invention is shown. DETAILED DESCRIPTION
[0061] Exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.
[0062] Exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.
[0063] In the first aspect, the adjustment method of the vehicle-computer question-answering model provided by the embodiment of the present invention is mainly applied to smart cars with intelligent cockpits, such as electric cars among new energy vehicles. Figure 1 As shown, the embodiment of the present invention includes at least the following steps:
[0064] S101, obtain the j-th round of question-answering data and the n rounds of question-answering data before and after it.
[0065] Wherein, j represents any round of question-answering data, n≥1 and is a positive integer. The question-answering data is the data generated between the user and the vehicle-computer question-answering model in the language model.
[0066] Before obtaining the jth round of question-and-answer data and the n rounds of question-and-answer data before and after it, it is necessary to collect the real conversation content between the user and the electric vehicle to ensure the authenticity and validity of the conversation content, so as to subsequently improve the accuracy of the user's semantic understanding in the vehicle-computer dialogue.
[0067] Specifically, user question data is obtained, and the vehicle-computer question and answer model in the language model is used to analyze the user question data alone or in combination with historical question and answer data to obtain vehicle-computer execution data; the user question data, the historical question and answer data, and the vehicle-computer execution data are stored in the cache as a round of question and answer data.
[0068] User question data contains user intent. Historical question and answer data is generated before the user question data. Vehicle-mounted execution data is used to respond to user intent. See Table 1.
[0069] Table 1
[0070]
[0071] By analyzing the in-vehicle question-answering model alone or in combination with historical conversation data and user question data, it can be determined that the user's intention is to navigate to the Yellow Crane Tower, so the in-vehicle execution data is generated: navigate to the Yellow Crane Tower.
[0072] Furthermore, the user question data, the historical question and answer data, and the vehicle computer execution data are stored in a cache as a round of question and answer data. The cache records all input information and corresponding output results of each round of question and answer data.
[0073] In an optional implementation, after obtaining the user question data, similar historical conversation data having similar intentions to the user question data is collected based on the user question data.
[0074] Table 2
[0075]
[0076] In Table 2, the keyword "Yellow Crane Tower" is extracted from the user question data, and similar historical conversation data related to "Yellow Crane Tower" generated before the user question data is collected based on the keyword "Yellow Crane Tower". It is worth noting that there may be other irrelevant conversation data between the similar historical conversation data and the user question data, and the two may not be generated on the same day. Other irrelevant conversation data can be excluded by keyword matching, and the most similar historical conversation data can be retained. The vehicle-computer question and answer model is used to combine similar historical conversation data and user question data for analysis to generate vehicle-computer execution data. Furthermore, the user question data, similar historical question and answer data, and the vehicle-computer execution data are stored in the cache as a round of question and answer data.
[0077] The technical solution of the present invention directly uses the most realistic human-computer dialogue data to train the model in real time. Since the human-computer dialogue data is obtained from complex and changeable driving scenarios, the trained vehicle-computer question-and-answer model can calmly respond to various needs raised by users in complex and changeable driving scenarios.
[0078] Furthermore, in the process of obtaining the j-th round of question-and-answer data and the n rounds of question-and-answer data before and after it, the j-th round of question-and-answer data and the n rounds of question-and-answer data before and after it are obtained from the cache during a preset low-peak period of vehicle use.
[0079] Among them, each round of question and answer data is stored in the cache in sequence according to its own question and answer time point.
[0080] When extracting data, we first determine the preset low-use period, during which the mileage traveled is below the set mileage threshold. For example, if users typically use their car between 6:00 AM and 12:00 PM every day, then extracting data from the low-use period between 12:00 PM and 6:00 AM can reduce server resource usage and improve server resource utilization.
[0081] Furthermore, a threshold for the number of extractions is set to manage the training of the vehicle-mounted question-and-answer model. For example, if the extraction threshold is set to 10,000 rounds of question-and-answer data, then 10,000 rounds of data are extracted from the cache to train the model. Furthermore, the j-th round of question-and-answer data and the n rounds of question-and-answer data before and after it are extracted from the cache in sequence according to the question-and-answer time point. For example, the five rounds of question-and-answer data before and after. The j-th round of data is any round of data in the cache. Furthermore, after the j-th round of question-and-answer data is processed, the j-th round of question-and-answer data is marked as processed data in the cache to prevent duplicate processing.
[0082] S102: Input the j-th round of question-and-answer data and the n rounds of question-and-answer data before and after it into the discriminant model of the language model to obtain the execution discrimination label and execution accuracy of the j-th round of question-and-answer data.
[0083] The jth round of Q&A data includes user question data, historical Q&A data, and vehicle-based execution data. Similarly, the previous n rounds of Q&A data also include their respective user question data, historical Q&A data, and vehicle-based execution data. It's worth noting that the historical Q&A data from the previous n rounds and the jth round of Q&A data may partially overlap.
[0084] In order to improve the accuracy of semantic understanding of the vehicle-computer question-answering model, before training the vehicle-computer question-answering model, the j-th round of question-answering data and the n rounds of question-answering data before and after it are input into the discriminant model of the language model. The discriminant model is combined with the n rounds of historical question data before and after the j-th round of question-answering data (that is, the context data of the j-th round of question-answering data) to judge the correctness of the execution result corresponding to the vehicle-computer execution data in the j-round question-answering data, thereby obtaining the execution discrimination label and execution correctness of the j-round question-answering data.
[0085] The execution discrimination label is used to determine whether the vehicle-mounted question-and-answer model correctly responds to the j-th round of question-and-answer data. Specifically, the execution discrimination label includes a correct label and an incorrect label. The correct label indicates that the vehicle-mounted execution data in the j-th round of question-and-answer data output by the vehicle-mounted question-and-answer model can correctly respond to the user question data in the j-th round of question-and-answer data. The incorrect label indicates that the vehicle-mounted execution data in the j-th round of question-and-answer data output by the vehicle-mounted question-and-answer model does not correctly respond to the user question data in the j-th round of question-and-answer data.
[0086] Execution accuracy is used to indicate the degree to which the vehicle-based question-and-answer model correctly responds to the j-th round of question-and-answer data. Execution accuracy ranges from 0 to 1, with larger values indicating a higher accuracy of the execution results of the vehicle-based execution data in the j-th round of question-and-answer data. If the execution discrimination label of the vehicle-based execution data in the j-th round of question-and-answer data is an error label, the execution accuracy of the vehicle-based execution data in the j-th round of question-and-answer data is 0.
[0087] If the execution discrimination label is an error label, prompt data is obtained to instruct the discrimination model to correctly respond to the user question data in the j-th round of question and answer data. The prompt data is used to instruct the discrimination model to correctly respond to the j-th round of question and answer data. The prompt data, the j-th round of question and answer data, and the n-round question and answer data before and after it are re-input into the discrimination model. By instructing the discrimination model to operate, vehicle-computer adjustment data for correctly responding to the user question data in the j-th round of question and answer data is obtained. Furthermore, after the vehicle-computer adjustment data is output, the correct label and correct execution degree of the vehicle-computer adjustment data are also output synchronously.
[0088] Considering that the cache originally stores vehicle-computer execution data, when the execution discrimination label for the j-th round of question-and-answer data is an error label, it indicates that the vehicle-computer execution data failed to correctly respond to the user question data in the j-th round of question-and-answer data. Therefore, after obtaining the vehicle-computer adjustment data, the vehicle-computer adjustment data is used to replace the vehicle-computer execution data in the j-th round of question-and-answer data in the cache, so that the cache stores vehicle-computer adjustment data that correctly responds to the user question data in the j-th round of question-and-answer data. Furthermore, after replacing the vehicle-computer execution data in the j-th round of question-and-answer data in the cache with the vehicle-computer adjustment data, the vehicle-computer adjustment data is marked as processed data to prevent duplicate processing.
[0089] In this technical solution, by using the discriminant model to use contextual information to discriminate the j-th round of data in advance, the overall confidence of the discriminant model can be improved, and the problem of decreased effect caused by incomplete information in one-way discrimination can be avoided, thereby improving the discrimination accuracy and laying a good foundation for subsequent training of the vehicle-computer question-answering model.
[0090] S103: If the execution judgment label is a correct label, determine the sample weight of the j-th round question-and-answer data with reference to the execution correctness.
[0091] The sample weight of the j-th round of Q&A data represents its importance relative to all training data during training. A higher sample weight indicates greater importance for the j-th round of Q&A data. For example, if the j-th round of Q&A data is more representative or important than other rounds of Q&A data, its sample weight will be higher.
[0092] Furthermore, the execution accuracy is used as the sample weight of the j-th round of question-answering data. For example, if the execution accuracy is 0.1, the sample weight is 0.1, but this is not a limitation. For example, the execution accuracy can also be multiplied to obtain the sample weight.
[0093] In an optional embodiment, after determining the sample weight of the j-th round of question and answer data with reference to the execution correctness, the parameter weight of the j-th round of question and answer data is obtained. The parameter weight is used to characterize the degree of influence of the j-th round of question and answer data on the update of the model parameters of the vehicle-computer question and answer model. The higher the parameter weight, the higher the degree of influence of the j-th round of question and answer data on the update of the model parameters of the vehicle-computer question and answer model. The model adjustment weight is determined using the parameter weight and the sample weight. Exemplarily, the model adjustment weight can be obtained by adding the parameter weight and the sample weight, but this does not form a limitation. For example, the model adjustment weight can also be obtained by multiplying the parameter weight and the sample weight. In order to prevent the output of the wrong label due to the inaccuracy of the judgment model itself, which affects the training of the vehicle-computer question and answer model, the parameter weight is introduced to adjust the influence of the sample weight.
[0094] S104 , adjusting the vehicle-computer question-answering model using the j-th round question-answering data and the sample weights thereof to obtain the adjusted vehicle-computer question-answering model.
[0095] During the adjustment process, the vehicle-computer question-answering model is adjusted using the j-th round question-answering data and the model adjustment weight thereof to obtain the adjusted vehicle-computer question-answering model.
[0096] The above is the update process of the vehicle-computer question-and-answer model. By using real data to extract question-and-answer data in real time during each preset low-peak period of vehicle usage and updating it, and by using real interactive question-and-answer data to train the model, the overall confidence of the model can be improved. In addition, the discriminant model is used as a guiding model to implement the knowledge distillation process, thereby improving the accuracy of the vehicle-computer question-and-answer model's responses to vertical domain questions, enabling it to provide accurate question-and-answer functions even in complex and changeable driving environments.
[0097] In the second aspect, based on the same inventive concept as the method for adjusting the vehicle-computer question-answering model provided in the embodiment of the first aspect, the embodiment of the present invention further provides a vehicle-computer response method, see Figure 2 , comprising the following steps:
[0098] S201, collecting user query data.
[0099] In actual application, user query data is generated in the actual driving environment, and it is expected that the vehicle-mounted question-and-answer system can provide the user's expected response data or perform the user's expected operations.
[0100] S202 : Input the user query data into the vehicle-computer question-answering model adjusted according to the steps of the method provided in the first aspect to obtain vehicle-computer execution data.
[0101] During the specific implementation process, in order to accurately understand the user's intention, the adjusted vehicle-computer question-and-answer model is used to analyze the user query data alone or in combination with historical question-and-answer data before the user query data, and the corresponding vehicle-computer execution data is output.
[0102] S203, controlling the electric vehicle to respond to the user according to the vehicle computer execution data.
[0103] Specifically, the vehicle computer executes data replies or controls the electric vehicle to perform related operations to meet user expectations.
[0104] It is worth noting that after the corresponding vehicle-computer execution data is output, the user query data, the historical question-and-answer data before the user query data, and the vehicle-computer execution data will be stored in the cache together for subsequent model training, thereby completing the training closed loop of the vehicle-computer question-and-answer model, enabling it to provide accurate question-and-answer functions even in complex and changeable driving environments.
[0105] In the third aspect, based on the same inventive concept as the method for adjusting the vehicle-computer question-answering model provided in the embodiment of the first aspect, the embodiment of the present invention further provides an adjustment system for the vehicle-computer question-answering model, see Figure 3 , the system comprising:
[0106] Acquisition module 301 is used to obtain the j-th round of question-and-answer data and the n-round question-and-answer data before and after it; where j represents any round of question-and-answer data, and n ≥ 1 is a positive integer, and the question-and-answer data is data generated between the user and the vehicle-computer question-and-answer model in the language model;
[0107] The discriminant module 302 is configured to input the j-th round of question-and-answer data and the n rounds of question-and-answer data preceding and following it into the discriminant model of the language macro model to obtain an execution discriminant label and execution accuracy of the j-th round of question-and-answer data; wherein the execution discriminant label is used to determine whether the vehicle-computer question-and-answer model correctly responded to the j-th round of question-and-answer data, and the execution accuracy is used to indicate the degree to which the vehicle-computer question-and-answer model correctly responded to the j-th round of question-and-answer data;
[0108] A determination module 303 is configured to determine the sample weight of the j-th round of question-and-answer data by referring to the execution correctness if the execution discrimination label is a correct label;
[0109] The adjustment module 304 is configured to adjust the vehicle-computer question-answering model using the j-th round of question-answering data and the sample weights thereof to obtain the adjusted vehicle-computer question-answering model.
[0110] It should be noted that the adjustment system of the vehicle-computer question-and-answer model provided in the embodiment of the present invention, wherein the specific manner in which each module performs operations has been described in detail in the method embodiment provided in the above-mentioned first aspect. The specific implementation process can refer to the method embodiment provided in the above-mentioned first aspect, and will not be elaborated here.
[0111] In the fourth aspect, based on the same inventive concept as the method for adjusting the vehicle-computer question-answering model provided in the embodiment of the first aspect, the embodiment of the present invention further discloses a vehicle-computer response system, see Figure 4 , the system comprising:
[0112] Collection unit 401, used to collect user query data;
[0113] An input unit 402 is configured to input the user query data into the vehicle-computer question-answering model adjusted according to the steps of the method of the first aspect to obtain vehicle-computer execution data;
[0114] The control unit 403 is used to control the electric vehicle to respond to the user according to the vehicle computer execution data.
[0115] In the fifth aspect, based on the same inventive concept as the method for adjusting the vehicle-computer question-and-answer model provided in the embodiment of the first aspect mentioned above, an embodiment of the present invention also discloses an electric vehicle, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any of the aforementioned methods when executing the program.
[0116] Through one or more embodiments of the present invention, the present invention has the following beneficial effects or advantages:
[0117] The present invention provides an adjustment method for a vehicle-computer question-and-answer model, a vehicle-computer response method, a system, and an electric vehicle, which obtain the j-th round of question-and-answer data and the n rounds of question-and-answer data before and after it, and then use a discriminant model in a large language model to perform context judgment on the j-th round of question-and-answer data and the n rounds of question-and-answer data before and after it, and combine more semantic information to judge the correctness of the execution result of the j-round question-and-answer data. Then, the vehicle-computer question-and-answer model in the large language model is adjusted using the correctly judged j-round question-and-answer data and the execution correctness to improve the semantic understanding ability of the large language model, improve the overall confidence of the vehicle-computer question-and-answer model, and enable it to provide accurate question-and-answer functions even in complex and changeable driving environments.
[0118] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.
[0119] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.
Claims
1. A method for adjusting a vehicle-based question-answering model, characterized in that: The method comprises: Obtain the j-th round of question-and-answer data and the n-round question-and-answer data before and after it; where j represents any round of question-and-answer data, and n ≥ 1 is a positive integer. The question-and-answer data is generated between the user and the vehicle-computer question-and-answer model in the language model; Input the j-th round of question-and-answer data and the n-round question-and-answer data before and after it into the discriminant model of the large language model to obtain an execution discriminant label and execution accuracy of the j-th round of question-and-answer data; wherein the execution discriminant label is used to determine whether the vehicle-computer question-and-answer model correctly responds to the j-th round of question-and-answer data, and the execution accuracy is used to indicate the degree to which the vehicle-computer question-and-answer model correctly responds to the j-round of question-and-answer data; If the execution judgment label is a correct label, determine the sample weight of the j-th round of question-and-answer data with reference to the execution correctness; The vehicle-computer question-answering model is adjusted using the j-th round question-answering data and the sample weight thereof to obtain the adjusted vehicle-computer question-answering model.
2. The method according to claim 1, wherein Before obtaining the j-th round of question-and-answer data and the n rounds of question-and-answer data before and after it, the method further includes: Get user question data; Analyzing the user question data using the vehicle-computer question-answering model alone or in combination with historical question-answering data to obtain vehicle-computer execution data; The user question data, the historical question and answer data, and the vehicle computer execution data are stored in a cache as a round of question and answer data.
3. The method according to claim 2, wherein The step of obtaining the j-th round of question-and-answer data and the n-round question-and-answer data before and after it specifically includes: During a preset off-peak period of vehicle usage, the j-th round of question-and-answer data and the n rounds of question-and-answer data before and after it are obtained from the cache.
4. The method according to claim 1, wherein The j-th round of question and answer data includes user question data, historical question and answer data and vehicle computer execution data; After obtaining the execution discrimination label and execution accuracy of the j-th round of question-and-answer data, the method further includes: If the execution determination tag is an error tag, obtaining prompt data for instructing the vehicle computer to correctly respond to the user question data in the j-th round of question-and-answer data; The prompt data, the j-th round of question and answer data, and the n-round question and answer data before and after it are input into the discriminant model again to obtain vehicle-computer adjustment data for correctly responding to the user question data in the j-round of question and answer data; wherein the vehicle-computer adjustment data has the correct label and the correct execution degree.
5. The method according to claim 1, wherein After determining the sample weight of the j-th round of question-and-answer data with reference to the execution correctness, the method further includes: Obtaining a parameter weight of the j-th round of question-and-answer data; wherein the parameter weight is used to represent the degree of influence of the j-th round of question-and-answer data on the update of model parameters of the vehicle-computer question-and-answer model; The model adjustment weight is determined using the parameter weight and the sample weight.
6. The method according to claim 5, wherein The adjusting the vehicle-computer question-answering model using the j-th round question-answering data and the sample weight thereof to obtain the adjusted vehicle-computer question-answering model specifically includes: The vehicle-computer question-answering model is adjusted using the j-th round question-answering data and the model adjustment weight thereof to obtain the adjusted vehicle-computer question-answering model.
7. A vehicle computer response method, characterized in that: The method comprises: Collect user inquiry data; Inputting the user query data into the vehicle-computer question-answering model obtained by adjusting the steps of the method according to any one of claims 1 to 6 to obtain vehicle-computer execution data; The electric vehicle is controlled to respond to the user according to the vehicle computer execution data.
8. A vehicle-computer question-answering model adjustment system, characterized in that: The system comprises: An acquisition module is used to obtain the j-th round of question-and-answer data and the n-round question-and-answer data before and after it; where j represents any round of question-and-answer data, and n ≥ 1 is a positive integer, and the question-and-answer data is the data generated between the user and the vehicle-computer question-and-answer model in the language model; a discrimination module, configured to input the j-th round of question-and-answer data and the n rounds of question-and-answer data preceding and following it into the discrimination model of the language macro model, and obtain an execution discrimination label and execution accuracy of the j-th round of question-and-answer data; wherein the execution discrimination label is used to determine whether the vehicle-computer question-and-answer model correctly responds to the j-th round of question-and-answer data, and the execution accuracy is used to indicate the degree to which the vehicle-computer question-and-answer model correctly responds to the j-th round of question-and-answer data; a determination module, configured to determine, if the execution discrimination label is a correct label, a sample weight of the j-th round of question-and-answer data with reference to the execution correctness; An adjustment module is used to adjust the vehicle-computer question-answering model using the j-th round of question-answering data and its sample weights to obtain the adjusted vehicle-computer question-answering model.
9. A vehicle-computer response system, characterized in that: The system comprises: A collection unit, used to collect user query data; An input unit, configured to input the user query data into the vehicle-computer question-answering model adjusted according to the steps of the method according to any one of claims 1 to 6, to obtain vehicle-computer execution data; A control unit is used to control the electric vehicle to respond to the user according to the vehicle computer execution data.
10. An electric vehicle comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the method according to any one of claims 1 to 7 are implemented.