Function control method, electronic equipment and vehicle

Abstract

The invention relates to the technical field of data retrieval, and provides a function control method, electronic equipment and a vehicle. The method comprises the following steps: receiving an input control statement, and determining at least two types of candidate parameters matched with the control statement; inputting the control statement and the at least two types of candidate parameters into a preset base language model; based on a base language model, determining a target parameter from the candidate parameters, determining a text parameter value matched with the target parameter, combining the target parameter and the text parameter value to obtain a target text instruction, and outputting the target text instruction; and executing a function corresponding to the target text instruction. According to the technical scheme provided by the embodiment of the invention, the problem of low efficiency caused by direct large language analysis of the whole sentence is solved, and the efficiency and accuracy of sentence recognition can be effectively improved; and the advantages of large language analysis of the base language model can be further exerted, and the efficiency of instruction generation is improved.

Description

Technical Field

[0001] The present application relates to the technical field of data retrieval, and in particular to a function control method, electronic equipment and vehicle. Background Art

[0002] With the continuous development of urbanization and industrialization, the convenience of transportation has brought people a freer and faster way of traveling. Therefore, more and more users choose to buy private cars to solve travel problems. With the growing demand for convenient car use by users, manufacturers have also further improved their requirements for the level of vehicle intelligence.

[0003] Currently, more and more vehicles are equipped with human-machine interaction interfaces, through which users can input the functions or information they want to find, or directly input the functions or information they want to find through voice input. Therefore, how to improve the efficiency of vehicle recognition of user input content to better provide services to users has become one of the research focuses of technicians in related fields. Summary of the invention

[0004] In view of this, the purpose of the present application is to propose a function control method, an electronic device and a vehicle to solve the problem of low efficiency of a large language model used in a vehicle in generating instructions for sentences input by a user.

[0005] Based on the above objectives, the present application provides a function control method, including:

[0006] receiving an input control statement and determining at least two types of candidate parameters matching the control statement;

[0007] Inputting the control statement and at least two types of candidate parameters into a preset base language model;

[0008] Based on the base language model, a target parameter is determined from each candidate parameter, and a text parameter value matching the target parameter is determined, and the target parameter and the text parameter value are combined to obtain a target text instruction and output;

[0009] Execute the function corresponding to the target text instruction.

[0010] Based on the same purpose of the above method, the present application also provides a function control device, including:

[0011] A statement receiving module, used for receiving an input control statement and determining at least two types of candidate parameters matching the control statement;

[0012] A model input module, used for inputting a control statement and at least two types of candidate parameters into a preset base language model;

[0013] An instruction output module is used to determine a target parameter from each candidate parameter based on a base language model, and to determine a text parameter value matching the target parameter, and to combine the target parameter and the text parameter value to obtain a target text instruction and output it;

[0014] The function execution module is used to execute the function corresponding to the target text instruction.

[0015] Based on the same purpose, the present application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method provided in any embodiment of the present application when executing the program.

[0016] Based on the same purpose, the present application also provides a vehicle, the vehicle comprising:

[0017] A memory for storing executable program codes;

[0018] A processor is used to call and run the executable program code from the memory so that the vehicle executes the method provided in any embodiment of the present application.

[0019] From the above, it can be seen that the technical solution of the embodiment of the present application determines the target parameter from the candidate parameters through the preset base language model, identifies the control statement input by the user according to different parameter dimensions, and provides inspiration for the vehicle to identify the user's intention. The query statement is identified in different dimensions, and the parameters based on different dimensions are processed in a targeted manner instead of directly identifying the entire sentence, which solves the problem of low efficiency caused by directly performing large-scale language analysis on the entire sentence, and can effectively improve the efficiency and accuracy of sentence recognition; then determine the text parameter value matching the target parameter, and then combine the target parameter and the text parameter to generate a target text instruction to control the vehicle to perform the corresponding function. Based on the base language model, the text parameter value is searched and determined, and combined with the parameter. Under the guidance of the inspiration of the parameter, the key information in the query statement is processed in a targeted manner, which can further give play to the advantages of the large language analysis of the base language model, improve the efficiency of instruction generation, and reduce the consumption of computing resources. Therefore, the technical solution of the embodiment of the present application can effectively balance the accuracy of instruction retrieval matching and the efficiency of instruction generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the present application or related technologies, the drawings required for use in the embodiments or related technical descriptions are briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0021] Figure 1A A flowchart of a function control method provided in an embodiment of the present application;

[0022] Figure 1B A schematic diagram of an improved retrieval enhancement generation model provided in an embodiment of the present application;

[0023] Figure 2 A schematic diagram of the structure of a function control device provided in an embodiment of the present application;

[0024] Figure 3 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0025] In order to make the objectives, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below in combination with specific embodiments and with reference to the accompanying drawings.

[0026] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the embodiments of the present application should be the usual meanings understood by people with ordinary skills in the field to which the present application belongs. The "first", "second" and similar words used in the embodiments of the present application do not represent any order, quantity or importance, but are only used to distinguish different components. "Including" or "comprising" and similar words mean that the elements or objects appearing in front of the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0027] Before explaining the embodiments and implementation methods of the present application, the background of the present application and the problems to be solved are briefly explained. In the field of intelligent control of vehicles, in the human-computer interaction scenarios between users and vehicles, most vehicle computers can only understand some simple or clear intentions. For example, the control instructions input by users through voice are very simple and clear, and it is relatively difficult to understand complex intentions. Therefore, it is necessary to rely on the intervention of a large language model. The larger the parameters, the stronger the model's understanding ability is, but the more resources it consumes, and the longer the inference time is. For enterprises, the cost is high, and for users, the longer the waiting time, the worse the experience. In addition, the training knowledge base of the large model itself lacks understanding of the relevant knowledge of vehicle control and vehicle settings, and the large language model will also have hallucination problems. Therefore, the embodiment of the present application proposes a function control method, which is as follows.

[0028] Figure 1A A function control method provided in an embodiment of the present application can be applied to the case where a control sentence input by a user is analyzed and the user's intention is determined by using an improved large language model. The function control method provided in an embodiment of the present application can be executed by a function control device, which can be integrated into an electronic device by means of hardware and / or software, such as integrated into a vehicle controller. Figure 1A , the method provided in the embodiment of the present application comprises the following steps:

[0029] S110: Receive an input control statement, and determine at least two types of candidate parameters matching the control statement.

[0030] Among them, the control statement can be a command or instruction corresponding to the function that the user expects to achieve. In the intelligent control of the vehicle, the user can issue the control statement to the vehicle computer through the human-computer interaction interface or voice input. It should be noted that the control statement is not necessarily a precise control command statement, but can also be just vague content. For example, "adjust the air conditioner to 15℃" is a kind of evil control command, and "the weather is too hot, turn on the air conditioner to cool down" is a fuzzy control statement. In view of the fact that the user in the car may not be very familiar with the functions in the car and the specific adjustment degree of the function, the vehicle needs to be able to analyze the user's true intention based on the fuzzy control statement proposed by the user, so as to provide the user with the service of the internal functions of the vehicle.

[0031] Candidate parameters can be feature quantities that can characterize verbs, nouns and / or logical relationships in control sentences, for example, they can include candidate conditions and candidate actions. Conditions and actions represent the conditions for the execution of the action to be expressed in the control sentence and the function of the action itself, respectively. Continuing with the previous example, in "It's too hot, turn on the air conditioner to cool down", "It's too hot" can be a condition, and the vehicle can be understood as the condition for turning on the air conditioner when the temperature reaches a certain level; correspondingly, turning on the air conditioner, turning the air conditioner to a certain temperature, and how strong the wind is can be actions.

[0032] The candidate conditions may be multiple conditions of the conditional meanings that the vehicle can recognize in the proximity control sentence and are intended to be expressed, and these conditions are used as candidates. Similarly, the candidate actions may be multiple actions of the action meanings that the vehicle can recognize in the proximity control sentence and are intended to be expressed.

[0033] Specifically, the user can input a control statement into the vehicle computer, and the vehicle analyzes the control statement to obtain many candidate conditions and candidate actions that are close to the meaning of the control statement itself. For example, a machine learning model can be pre-trained, and the model inputs a control statement and outputs multiple candidate conditions and candidate actions.

[0034] S120, inputting the control statement and at least two types of candidate parameters into a preset base language model.

[0035] The base language model may be a large language model with certain language parsing and understanding capabilities, and the base language model usually contains a large amount of data for retrieval and reference. The base language model may be pre-trained by relevant technical personnel, and the base language model may adopt a basic model generated based on retrieval enhancement, such as a 7B model. The control statement and each candidate parameter are input into the base language model.

[0036] S130. Based on the base language model, determine the target parameter from each candidate parameter, and determine the text parameter value that matches the target parameter, combine the target parameter and the text parameter value to obtain the target text instruction and output it.

[0037] The text parameter value can be the specific data that needs to be limited in the target condition, or the specific data that needs to be limited in the action, for example: "When the temperature inside the car is higher than 30°C, turn on the air conditioner and adjust it to 15°C", where 30 and 15 are text parameter values. The target text instruction can be the vehicle's recognition result of the control statement input by the user, and the target text instruction is used to control specific vehicle functional components.

[0038] Based on the base language model, each candidate parameter is screened to obtain the target parameter. Each candidate parameter can be screened according to the semantic similarity with the original control sentence of the user recognized by the base language model. The basis of screening is not limited here. And the numerical value corresponding to the target parameter, that is, the text parameter value, can be retrieved and generated from the model. Then the target parameter and the target parameter value are combined to obtain the target text instruction generated by the vehicle computer recognizing the user's intention.

[0039] S140, executing the function corresponding to the target text instruction.

[0040] By executing the target text instruction determined in the above steps, the vehicle realizes the function corresponding to the target text instruction.

[0041] The technical solution of the embodiment of the present application determines the target parameter from the candidate parameters through a preset base language model, identifies the control statement input by the user according to different parameter dimensions, provides inspiration for the vehicle to identify the user's intention, identifies the query statement in different dimensions, and processes the parameters based on different dimensions in a targeted manner instead of directly identifying the entire sentence, thereby solving the problem of low efficiency caused by directly performing large-scale language analysis on the entire sentence, and can effectively improve the efficiency and accuracy of sentence recognition; then determine the text parameter value matching the target parameter, and then combine the target parameter and the text parameter to generate a target text instruction to control the vehicle to perform the corresponding function, use the base language model as a basis, search and determine the text parameter value, and combine it with conditions and actions, and under the inspirational guidance of the parameters, process the key information in the query statement in a targeted manner, which can further give play to the advantages of the large-scale language analysis of the base language model, improve the efficiency of instruction generation, and reduce the consumption of computing resources. Therefore, the technical solution of the embodiment of the present application can effectively balance the accuracy of instruction retrieval matching and the efficiency of instruction generation.

[0042] Based on the above embodiment, to determine the candidate parameters, it is necessary to first vectorize the control statement and perform a database search operation. In an optional implementation, determining at least two types of candidate parameters matching the control statement in S110 may include:

[0043] S111. Vectorize the control statement to obtain a query vector.

[0044] After receiving the control statement input by the user, the vehicle first vectorizes the control statement and converts the control statement into a query vector to facilitate vectorized operations in the subsequent input model. The vectorization method can adopt any vectorization method in the relevant technology, and the embodiment of the present application does not limit this.

[0045] S112: Input the query vector into at least two types of parameter retrieval databases, and respectively determine at least two types of parameters to be screened corresponding to the control statement.

[0046] The parameter search database may be a pre-set database for searching parameters corresponding to control statements, and different parameters may correspond to different search databases. In each category of parameter search database, the corresponding category of parameters to be screened is determined according to the control statement.

[0047] S113. For each of the at least two categories, execute as a target class respectively: sort the parameters to be screened of the target class to obtain a sorting result of the parameters to be screened, and screen multiple candidate parameters of the target class from the sorting result of the parameters to be screened.

[0048] In each category of parameters to be screened, the retrieved parameters to be screened are sorted to obtain parameter sorting results. The sorting basis may be the sorting according to the semantic relevance between each parameter to be screened and the query vector (or control statement), and then multiple candidate parameters are screened from the parameters to be screened. The principle of screening may be to take the first k parameters to be screened as candidate parameters. Since the parameters may include conditions, S112 may include: P1. Input the query vector into the conditional search database to determine the conditions to be screened corresponding to the control statement.

[0049] The condition retrieval database may be a pre-set database for retrieving conditions corresponding to control statements. Of course, the condition retrieval database may be in the form of a vector library, which is convenient for performing retrieval and calculation operations with the query vector. The query vector is input into the condition retrieval database, and several condition vectors that are close to the conditional meanings to be expressed in the control statement are retrieved from the condition retrieval database as the conditions to be screened.

[0050] The corresponding S113 may include: P2, sorting the conditions to be screened to obtain a sorting result of the conditions to be screened, and screening a plurality of candidate conditions from the sorting result of the conditions to be screened.

[0051] The several conditions to be screened retrieved in the above steps are sorted to obtain a condition sorting result. The basis for sorting can be sorting according to the semantic relevance between each condition to be screened and the query vector (or the control statement), and then multiple candidate conditions are obtained from the several conditions to be screened. The principle of screening can be to take the top k conditions to be screened as candidate conditions.

[0052] In addition, the parameter may also include an action, and S112 may include: Q1. Input the query vector into an action search database to determine the action to be screened corresponding to the control statement.

[0053] Similar to the conditional search database, the action search database can be a pre-set database for searching actions corresponding to control statements. Of course, the action search database can be in the form of a vector library, which is convenient for searching and calculating with the query vector. The query vector is input into the action search database, and several action vectors that are close to the action meaning to be expressed in the control statement are retrieved from the action search database as the actions to be screened.

[0054] The corresponding S113 may include: Q2, sorting the actions to be screened to obtain a sorting result of the actions to be screened, and screening a plurality of candidate actions from the sorting result of the actions to be screened.

[0055] Similarly, the several actions to be screened retrieved in the above steps are sorted to obtain a conditional sorting result. The basis for sorting can be the sorting according to the semantic relevance between each action to be screened and the query vector (or the control statement), and then multiple candidate actions are obtained from the several actions to be screened. The principle of screening can be to take the top k actions to be screened as candidate actions.

[0056] In the above implementation, the control statement input by the user is vectorized to obtain a query vector, and the query vector is respectively input into the condition retrieval database and the action retrieval database to retrieve the conditions to be screened and the actions to be screened, and then a plurality of candidate conditions and a plurality of candidate actions are screened by sorting the conditions to be screened and the actions to be screened. The vectorized query and sorting screening can be used to perform two filterings to obtain candidate conditions and candidate actions that are closer to the conditions and actions that the control statement wants to express. These candidate conditions and candidate actions also provide options for the subsequent determination of target conditions and target actions. The split retrieval and screening of conditions and actions can also be closer to the original intention of the user more accurately, which helps to better understand the control statement input by the user and execute the corresponding function.

[0057] Based on the above-mentioned embodiments and implementation methods, the process of determining the parameters to be screened needs to be matched by searching the database. In an optional implementation method, inputting the query vector into at least two types of parameter search data in S112 to respectively determine at least two types of parameters to be screened corresponding to the control statement may include:

[0058] S1121. For each of at least two types of parameter retrieval databases, input the query vector into the parameter retrieval database, and calculate the parameter similarity between each parameter vector in the parameter retrieval database and the query vector.

[0059] The parameter similarity may be the vector similarity between the parameter vector and the query vector, and the vector similarity between the query vector and each parameter vector in the parameter retrieval database is calculated. The vector similarity may be calculated by calculating cosine similarity, Euclidean distance, Manhattan distance, and Pearson correlation coefficient, etc., which is not limited in the embodiments of the present application.

[0060] S1122. Sort the parameter vectors according to the parameter similarity to obtain a parameter similarity sorting result.

[0061] All parameter vectors in the parameter retrieval database are sorted according to the parameter similarity calculated with the query vector (for example, from large to small or from small to high), and the similarity sorting result of each parameter vector in the parameter retrieval database is obtained.

[0062] S1123. Select the parameter vectors with the largest first preset number of parameter similarities from the parameter similarity sorting results as the parameters to be screened.

[0063] In the parameter similarity ranking results obtained by sorting in the above steps, the first n (a first preset number) parameter vectors with the largest similarity are selected as parameters to be screened.

[0064] Since the parameter may include a condition, the step of inputting the query vector into at least two types of parameter search data in S112 to respectively determine at least two types of parameters to be screened corresponding to the control statement may include:

[0065] A1. Input the query vector into the conditional search database, and calculate the conditional similarity between each conditional vector in the conditional search database and the query vector.

[0066] The conditional similarity may be the vector similarity between the conditional vector and the query vector, and the vector similarity between the query vector and each conditional vector in the conditional retrieval database is calculated. The vector similarity may be calculated by calculating cosine similarity, Euclidean distance, Manhattan distance, and Pearson correlation coefficient, etc., which is not limited in the present embodiment of the application.

[0067] A2. Sort each condition vector according to the condition similarity to obtain the condition similarity sorting result.

[0068] All condition vectors in the condition retrieval database are sorted according to the condition similarity calculated with the query vector (for example, sorted from large to small or from small to high), and the similarity sorting result of each condition vector in the condition retrieval database is obtained.

[0069] A3. Select a first preset number of condition vectors with the largest condition similarities from the condition similarity sorting results as the conditions to be screened.

[0070] In the condition similarity sorting results obtained by sorting in the above steps, the first n (a first preset number) condition vectors with the largest similarity are selected as the conditions to be screened.

[0071] Accordingly, since the parameter may include an action, the step of inputting the query vector into at least two types of parameter search data in S112 to respectively determine at least two types of parameters to be screened corresponding to the control statement may include:

[0072] B1. Input the query vector into the action retrieval database, and calculate the action similarity between each action vector in the action retrieval database and the query vector.

[0073] Similarly, the action similarity can be the vector similarity between the action vector and the query vector, and the vector similarity between the query vector and each action vector in the action retrieval database is calculated. The vector similarity can be calculated by calculating cosine similarity, Euclidean distance, Manhattan distance, and Pearson correlation coefficient, etc., which is not limited in the embodiments of the present application.

[0074] B2. Sort each action vector according to action similarity to obtain an action similarity sorting result.

[0075] All action vectors in the action retrieval database are sorted according to the action similarity calculated with the query vector (for example, sorted from large to small or from small to high), and the similarity sorting result of each action vector in the action retrieval database is obtained.

[0076] B3. Selecting a first preset number of action vectors with the largest action similarities from the action similarity sorting results as the actions to be screened.

[0077] In the action similarity ranking results obtained by sorting in the above steps, the first n (a first preset number) action vectors with the largest similarity are selected as the actions to be screened.

[0078] In the above implementation, the query vectors are input into the condition retrieval database and the action retrieval database respectively, and are sorted by calculating the vector similarity to obtain the condition similarity sorting results and the action similarity sorting results. The first n vectors with the largest similarity in each sorting result are selected as the conditions to be screened or the actions to be screened, and the vector similarity is used as the basis for sorting and screening. Vectorized query, calculation and screening can better utilize the parallel computing capabilities of modern processors, optimize the utilization of hardware resources, reduce computing delays, and respond to query and calculation requests more quickly, thereby improving processing efficiency, thereby helping to improve the vehicle's recognition speed for control statements input by users and improve processing efficiency.

[0079] On the basis of the aforementioned implementation, the embodiment of the present application provides a method for filtering out candidate parameters from the parameters to be filtered. In an optional implementation, the steps of sorting the parameters to be filtered of the target class to obtain a sorting result of the parameters to be filtered in S113, and filtering out multiple candidate parameters of the target class from the sorting result of the parameters to be filtered may include:

[0080] S1131. Determine the parameter semantic relevance between each of the parameters to be screened and the control statement, and sort each of the parameters to be screened according to the parameter semantic relevance to obtain a sorting result of the parameters to be screened.

[0081] Parameter semantic relevance can be an indicator of the semantic relevance between the parameter to be screened and the control statement. In the aforementioned embodiment, the parameter to be screened is obtained by selecting the parameter to be screened from the parameter vector of the parameter retrieval database through vector similarity. At this time, the semantic relevance is calculated by the parameter to be screened and the control statement. The calculation of semantic relevance can use context-related word vectors (such as pre-trained models such as ELMo and BERT) to capture the semantic changes of words in different contexts, or calculate the semantic relevance between texts based on a semantic matching model (at this time, the control statement is directly processed, and there is no need to process the vectorized query vector), or use the knowledge graph method. The knowledge graph provides the semantic relationship between words. By traversing the relationship path in the knowledge graph, the semantic relevance between two words can be calculated (at this time, the conditional words of the condition to be screened are also used to directly process the control statement, without the need to use a vector form). Of course, the embodiment of the present application does not limit the calculation and evaluation method of semantic relevance.

[0082] After calculating the semantic relevance indexes between different parameters to be screened and control statements (or query vectors), the parameters to be screened are sorted according to the magnitude of the semantic relevance to obtain a sorting result of the parameters to be screened.

[0083] S1132. Select a second preset number of parameters to be screened with the greatest semantic relevance from the sorting results of the parameters to be screened as the multiple candidate parameters.

[0084] Based on the sorting results of the parameters to be screened obtained in the above steps, k (a second preset number) parameters to be screened with the greatest conditional semantic relevance are selected as the screening results, i.e., multiple candidate parameters. It can be understood that since the candidate parameters are screened again based on the parameters to be screened, k < n (i.e., the second preset number is less than the first preset number).

[0085] In the expanded implementation, since the parameters include conditions, P2 sorts the conditions to be screened to obtain the sorting results of the conditions to be screened, and screens multiple candidate conditions from the sorting results of the conditions to be screened, which can include: C1, determining the conditional semantic relevance of each condition to be screened and the control statement, and sorting each condition to be screened according to the conditional semantic relevance to obtain the sorting results of the conditions to be screened.

[0086] Among them, the conditional semantic relevance can be an indicator of the semantic relevance between the condition to be screened and the control statement. In the aforementioned implementation, the condition to be screened is obtained by selecting the condition to be screened from the condition vector of the conditional retrieval database through vector similarity. At this time, the semantic relevance is calculated by the condition to be screened and the control statement. The calculation of semantic relevance can use context-related word vectors (such as pre-trained models such as ELMo and BERT) to capture the semantic changes of words in different contexts, or calculate the semantic relevance between texts based on a semantic matching model (at this time, the control statement is directly processed, and there is no need to process the vectorized query vector), or use the knowledge graph method. The knowledge graph provides the semantic relationship between words. By traversing the relationship path in the knowledge graph, the semantic relevance between two words can be calculated (at this time, the conditional words of the condition to be screened and the control statement are also used to directly process, without the need to use a vector form). Of course, the embodiment of the present application does not limit the calculation and evaluation method of semantic relevance.

[0087] After calculating the semantic relevance indexes between different to-be-screened conditions and control statements (or query vectors), the to-be-screened conditions are sorted according to the magnitude of the semantic relevance to obtain a sorting result of the to-be-screened conditions.

[0088] C2. Selecting a second preset number of conditions to be screened with the greatest semantic relevance from the sorted results of the conditions to be screened as multiple candidate conditions.

[0089] Based on the sorting results of the conditions to be screened obtained in the above steps, k (a second preset number) conditions to be screened with the greatest semantic relevance are selected as the results after screening, that is, multiple candidate conditions. It can be understood that since the candidate conditions are screened again based on the conditions to be screened, k < n (that is, the second preset number is less than the first preset number).

[0090] Correspondingly, the parameter also includes actions. In Q2, the actions to be screened are sorted to obtain a sorting result of the actions to be screened, and multiple candidate actions are screened from the sorting result of the actions to be screened, which may include:

[0091] D1. Determine the action semantic relevance between each action to be screened and the control statement, and sort each action to be screened according to the action semantic relevance to obtain a sorting result of the actions to be screened.

[0092] Similarly, the semantic relevance of actions can be an indicator of the semantic relevance between the action to be screened and the control statement. In the aforementioned implementation, the action to be screened is obtained by selecting the vector similarity from the condition vector of the action retrieval database. At this time, the semantic relevance is calculated by the action to be screened and the control statement. The calculation of semantic relevance can use context-related word vectors to capture the semantic changes of words in different contexts, or calculate the semantic relevance between texts based on a semantic matching model (at this time, the control statement is directly processed, and there is no need to process the vectorized query vector), or use the knowledge graph method. The knowledge graph provides the semantic relationship between words. By traversing the relationship path in the knowledge graph, the semantic relevance between two words can be calculated (at this time, the action word of the action to be screened is also used to directly process the control statement, without the need to use a vector form). Of course, the embodiment of the present application does not limit the calculation and evaluation method of semantic relevance.

[0093] After calculating the semantic relevance indexes between different actions to be screened and control statements (or query vectors), the actions to be screened are sorted according to the magnitude of the semantic relevance to obtain a sorting result of the actions to be screened.

[0094] D2. Selecting a second preset number of actions to be screened with the greatest action semantic relevance from the sorted results of the actions to be screened as multiple candidate actions.

[0095] Based on the sorting results of the actions to be screened obtained in the above steps, k (a second preset number) actions to be screened with the greatest semantic relevance are selected as the results after screening, i.e., multiple candidate actions. It can be understood that since the candidate actions are screened again based on the actions to be screened, k < n (i.e., the second preset number is less than the first preset number).

[0096] In the above implementation, multiple candidate conditions and multiple candidate actions are further obtained by semantic relevance screening in each condition to be screened and each action to be screened. The judgment and screening of semantic relevance are introduced here, so that the candidate conditions and candidate actions obtained after this screening can be more semantically appropriate to the control statement input by the user. The candidate conditions and candidate actions obtained by screening provide options for the subsequent confirmation of target conditions and target actions, and the screening of semantic relevance also helps these options improve the information quality, thereby helping to improve the accuracy of identifying the user's true intention.

[0097] In various embodiments and implementations of the present application, the base language model is a large language model for processing query vectors, which is improved on the basis of the prior art. In an optional implementation, the base language model is determined by the following method:

[0098] E1. Obtain a low-rank matrix trained based on a preset amount of vehicle domain data.

[0099] Among them, a low-rank matrix may refer to a matrix whose rank is much smaller than the dimension of the matrix, that is, the matrix contains a large amount of redundant information, and the low-rank matrix proposed in the embodiment of the present application is obtained by training with vehicle field data. In addition, the vehicle field data needs to reach a certain data scale, for example, the amount of data can be hundreds of millions, which will cause the trained low-rank matrix to contain a large amount of relevant content in the vehicle field. Of course, the training method for the low-rank matrix can adopt any method in the relevant field, and the embodiment of the present application does not limit this.

[0100] E2. Multiply the low-rank matrix with the intrinsic parameter matrix in the initial base model to obtain the optimized parameter matrix.

[0101] The initial base model can be the original state of the base language model before the low-rank matrix is ​​adjusted. The initial base model is also a large language model. The initial base model can adopt a large language model with ultra-large data scale, such as the base 7B model, which has 700 million data volumes. Of course, these 700 million data include content from various industries and fields, and are not specialized in a certain field.

[0102] The intrinsic parameter matrix can be a parameter matrix of the initial base model itself, which is used to calculate with the query vector to retrieve content related to the query vector from the large language model. The parameters in the intrinsic parameter matrix are huge. If the ability of the large language model needs to be adjusted, it is necessary to make major changes to the intrinsic parameter matrix, which is more laborious.

[0103] Here, a low-rank matrix is ​​introduced. By multiplying the low-rank matrix with the intrinsic parameter matrix, a new optimization parameter matrix is ​​obtained. The optimization parameter matrix introduces a low-rank matrix relative to the original intrinsic parameter matrix. At the same time, because the low-rank matrix is ​​obtained through training with a large amount of vehicle field data, the introduction of the low-rank matrix brings model parameters related to the vehicle professional field to the initial base model. In addition, due to the low-rank characteristics of the low-rank matrix, when adjusting the model parameters, there is no need to make large-scale adjustments to the intrinsic parameter matrix. Only the parameters in the low-rank matrix need to be adjusted to adjust the multiplied optimization parameter matrix. This can bring an inspiring retrieval effect in the vehicle field to the initial base model, and bring retrieval parameter adjustments in the vehicle field to the initial base model with a huge amount of data, so that the base language model can process vehicle-related query requests more efficiently.

[0104] E3. Reload the optimized parameter matrix into the initial base model to obtain the base language model.

[0105] The optimized parameter matrix improved by the low-rank matrix is ​​reloaded back into the initial base model to obtain the base language model. At this time, the base language model has efficient language processing capabilities in the vehicle field.

[0106] In the above implementation, by introducing a low-rank matrix to improve the initial base model, the problem of cumbersome parameters and complex adjustment of the initial base matrix can be solved. Only by adjusting the low-rank matrix, the overall parameters of the base language model can be adjusted, which can save a lot of computing resources and improve the model processing efficiency. At the same time, since the low-rank matrix is ​​trained with a large amount of vehicle field data, the addition of the low-rank matrix provides the base language model with the ability of detailed retrieval in the vehicle age field, which can greatly reduce the problem of retrieval hallucinations that the original initial base model is prone to, and to a certain extent, ensure the language processing ability of the base language model for vehicle-related sentences.

[0107] The base language model provided in the present application may be improved by adding a low-rank matrix. In a further optional implementation, the low-rank matrix includes a parameter selection matrix, and the optimization parameter matrix in the base language model is determined by the parameter selection matrix.

[0108] Determining a target parameter from each of the candidate parameters based on the base language model in S130 includes:

[0109] S131. Based on the base language model, determine the parameter matching degree between each of the candidate parameters and the control statement.

[0110] Among them, the parameter selection matrix can be a low-rank matrix that plays a role in screening candidate parameters. After the optimized parameter matrix under the influence of the parameter selection matrix is ​​reloaded into the model, the ability of the base language model to screen parameters is improved.

[0111] For example, parameter selection LoRA (Low-Rank Adaptation) can be used. In the RAG (Retrieval Augmented Generation) framework, LoRA can be used as a fine-tuning technique to optimize the performance of the model on specific tasks. LoRA reduces the number of parameters that need to be trained through low-rank decomposition, which makes fine-tuning in the RAG framework more efficient and economical. In RAG, due to the need to process a large amount of knowledge and information, the use of LoRA can significantly reduce the consumption of computing resources and storage space. In the embodiment of the present application, the addition of parameter selection LoRA optimizes the ability of the base language model to analyze parameters in control sentences.

[0112] The parameter matching degree can be a matching degree index between the candidate parameter and the control statement. The matching degree index can be selected in the form of normalized cumulative loss gain, recall rate, average accuracy and natural language evaluation score, which is not limited in the embodiment of the present application. The matching degree index between each candidate parameter and the control statement is calculated through the base language model.

[0113] S132. Taking the candidate parameter with the highest degree of matching the condition as the target parameter.

[0114] The candidate parameter with the highest parameter matching degree calculated in the above steps is used as the target parameter, and the target parameter can be a part of the query result (or target text instruction).

[0115] In the expanded implementation, since the parameter includes conditions, the low-rank matrix includes a condition selection matrix, and the optimization parameter matrix in the base language model is determined by the condition selection matrix;

[0116] S131 may be extended to: determining the target condition from each candidate condition based on the base language model, and corresponding S131 may include:

[0117] F1. Based on the base language model, determine the matching degree between each candidate condition and the control statement.

[0118] Among them, the conditional selection matrix can be a low-rank matrix that plays a role in screening candidate conditions. After the optimization parameter matrix under the influence of the conditional selection matrix is ​​reloaded into the model, the ability of the base language model to screen candidate conditions is improved. For example, the conditional selection LoRA (Low-Rank Adaptation) can be used. In the embodiment of the present application, the addition of the conditional selection LoRA optimizes the ability of the base language model to analyze the conditions in the control statement.

[0119] The condition matching degree can be a matching degree index between the candidate condition and the control statement. The matching degree index can be selected in the form of normalized discounted cumulative gain, recall rate, average accuracy rate and natural language evaluation score, which is not limited in the embodiment of the present application. The matching degree index between each candidate condition and the control statement is calculated through the base language model.

[0120] F2. Take the candidate condition with the highest condition matching degree as the target condition.

[0121] The candidate condition with the highest condition matching degree calculated in the above steps is used as the target condition, and the target condition can be a part of the query result (or target text instruction).

[0122] Similarly, since the parameters include actions, the low-rank matrix also includes an action selection matrix, and the optimization parameter matrix in the base language model is determined by the action selection matrix;

[0123] S131 can be extended to determine the target action from each candidate action based on the base language model. The corresponding S131 may include:

[0124] G1. Based on the base language model, determine the action matching degree between each candidate action and the control statement.

[0125] Among them, the action selection matrix can be a low-rank matrix that plays a role in screening candidate actions. After the optimization parameter matrix under the influence of the action selection matrix is ​​reloaded into the model, the ability of the base language model to screen candidate actions is improved. For example, action selection LoRA can be used. In the embodiment of the present application, the addition of action selection LoRA optimizes the ability of the base language model to analyze actions in control sentences.

[0126] The action matching degree can be a matching degree index between the candidate action and the control statement. The matching degree index can be selected in the form of normalized discounted cumulative gain, recall rate, average accuracy rate and natural language evaluation score, etc., which is not limited in the embodiment of the present application. The matching degree index between each candidate action and the control statement is calculated through the base language model.

[0127] G2. The candidate action with the highest action matching degree is selected as the target action.

[0128] The candidate action with the highest action matching degree calculated in the above steps is taken as the target action, and the target action can be a part of the query result (or target text instruction).

[0129] In the above implementation, by introducing the conditional low-rank matrix and the action low-rank matrix, the matching index between the multiple candidate conditions and the multiple candidate actions screened in the above implementation is calculated with the control statement, and finally the target conditions and target actions with the highest matching degree are determined; by adding the conditional low-rank matrix and the action low-rank matrix, the base language model has an inspiring recognition ability, which improves the ability of the base language model to specifically identify the conditions and actions in the sentence, thereby improving the accuracy and efficiency of the base language model retrieval and recognition, and indirectly improving the efficiency and accuracy of the target text instruction generation.

[0130] In addition to the above parameter selection matrix, the improvement of the base language model by the low-rank matrix may also include other aspects. In another optional implementation, the low-rank matrix includes a parameter filling matrix, and the optimized parameter matrix in the base language model is determined by the parameter filling matrix;

[0131] Determining the text parameter value matching the target parameter based on the base language model in S130 may include:

[0132] Based on the base language model, text information that matches the control statement is searched from the base language model; and parameters in the text information that match the target parameter are used as text parameter values.

[0133] Among them, the parameter filling matrix can be used to help the base language model match the corresponding numerical value for the target parameter, and the parameter filling matrix can also use parameter filling LoRA. The parameters corresponding to the target parameters retrieved by the base language model optimized by the parameter filling matrix are used as text parameter values ​​for the subsequent grouping and generation of target text instructions.

[0134] In the expanded implementation, since the parameter includes a condition and an action, the content in S130 can be extended to: based on the base language model, determining the text parameter value matching the target condition and the target action can include:

[0135] H1. Based on the base language model, search for text information that matches the control statement from the base language model.

[0136] Among them, the parameter filling matrix can be used to help the base language model match the corresponding values ​​for the target conditions and target actions, and the parameter filling matrix can also use parameter filling LoRA. Exemplarily, the user speaks to the car computer: "The weather is too hot, turn on the air conditioner to cool down", and after screening and identifying the base language model of the aforementioned embodiments, it can be determined that the target condition can be "the temperature in the car" and the target action can be "adjust the air conditioner". So, how many degrees Celsius does the temperature in the car rise to meet the user's "the weather is too hot", and how should the air conditioner be adjusted to meet the "cool" requirement. At this time, the base language model optimized by the parameter filling matrix retrieves the text information of the corresponding parameters in the huge amount of data for filling, so as to quantitatively express the control statement input by the user to the greatest extent. Continuing with the previous example, the text information that needs to be filled in corresponding to the target condition can be "greater than 30℃", and the text information that needs to be filled in corresponding to the target action can be "lowered to 15℃".

[0137] H2. The parameters in the text information that match the target condition and the target action are used as text parameter values.

[0138] The parameters corresponding to the target conditions and target actions retrieved by the base language model after optimization of the parameter filling matrix are used as text parameter values ​​for the grouping and generation of subsequent target text instructions.

[0139] In the above implementation, by introducing a parameter filling matrix, the base language model is helped to search for parameter values ​​corresponding to target conditions and target actions in massive data, thereby determining the text parameter values. The addition of the parameter filling matrix provides inspiration and guidance for the retrieval of text parameter values, improves the low efficiency caused by blind retrieval in massive data, and can effectively improve the parameter retrieval and matching capabilities of the base language model itself. Similar to the principles of other low-rank matrices, the entire base language model can be adjusted by simply adjusting the parameters of the parameter filling matrix, without directly adjusting the complex parameters of the base language model, thereby improving the flexibility of the base language model.

[0140] In addition to the above-mentioned parameter selection matrix and parameter filling matrix, the improvement of the base language model by the low-rank matrix may also include other aspects. In another optional implementation, the low-rank matrix also includes a logical relationship matrix, and the optimization parameter matrix in the base language model is determined by the logical relationship matrix;

[0141] In S130, generating and outputting a target text instruction from the target parameter and the text parameter value based on the base language model may include:

[0142] Based on the base language model, the logical correspondence between the target parameter and the text parameter value is adjusted; according to the logical correspondence, the target text instruction is generated and output.

[0143] The logical relationship matrix can be a low-rank matrix used to correct the logical pairing relationship between the target parameter and the text parameter value, for example, the logical relationship and reflection LoRA can be used. It can be understood that the initial base model itself, as a large language model, has a certain logical recognition ability, and can answer the answer sentence with a certain logic according to the query sentence, but there is also a high possibility of model hallucination, that is, the output answer sentence is correct in language structure but completely wrong in content meaning.

[0144] The base language model optimized by the logic relationship matrix can more accurately generate target text instructions with the same meaning as the control statements input by the user when the target parameters and their corresponding text parameter values ​​are combined.

[0145] Continuing with the previous example, since the parameters include conditions and actions, the relevant content in S130 can be extended to: based on the base language model, generating and outputting target text instructions from target conditions, target actions and text parameter values, which may include:

[0146] I1. Based on the base language model, adjust the logical correspondence between target conditions, target actions and text parameter values.

[0147] The logical relationship matrix can be a low-rank matrix used to correct the logical pairing relationship between the target condition, the target action and the text parameter value. For example, the logical relationship and the reflective LoRA can be used. Exemplarily, continuing the previous example, if the text information corresponding to the target condition (the temperature inside the car) needs to be filled in is "greater than 30°C", and the text information corresponding to the target action (adjusting the air conditioner) needs to be filled in is "lowered to 15°C", but due to the illusion of the model, the initial base model outputs the answer: "If the temperature inside the car drops to 15°C, adjust the air conditioner to be greater than 30°C". It seems that there is nothing wrong with the simple understanding, but it runs counter to the meaning of the user input "The weather is too hot, turn on the air conditioner to cool down".

[0148] Therefore, a logical relationship matrix is ​​introduced, which is trained from a large number of logical sentence data sets. It plays a role in adjusting and optimizing the logical correspondence between target conditions, target actions and text parameter values ​​in the base language model.

[0149] I2. Generate and output target text instructions according to the logical correspondence.

[0150] The base language model optimized by the logic relationship matrix can more accurately generate target text instructions with the same meaning as the control statements input by the user when the target conditions, target actions and their corresponding text parameter values ​​are combined.

[0151] In the above implementation, by introducing the logic relationship matrix, the logic relationship of the generated sentences in the initial base model is improved, the accuracy and rationality of sentence generation are improved, and thus it helps to improve the accuracy and efficiency of target text instruction generation.

[0152] It should be noted that the method of the embodiment of the present application can be performed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario and completed by multiple devices cooperating with each other. In the case of such a distributed scenario, one of the multiple devices can only perform one or more steps in the method of the embodiment of the present application, and the multiple devices will interact with each other to complete the described method.

[0153] Based on the above embodiments and implementation methods, the present application also provides a specific implementation method as follows: Figure 1B As shown, the details are as follows:

[0154] Before the large language model of the embodiment of the present application is used, the condition knowledge base and the action knowledge base are vectorized, and the condition retrieval database and the action retrieval database are two vector libraries. In addition, offline fine-tuning training action selection LoRA, condition selection LoRA, condition and action parameter filling LoRA, and logical relationship and reflection LoRA are performed.

[0155] When the large language model involved in the embodiment of the present application is used, first, a control sentence issued by a user is received, and the control sentence is vectorized using a vector model to obtain a query vector. The top n conditions to be screened with the highest vector similarity are searched in the condition search database, and the top n actions to be screened with the highest vector similarity are searched in the action search database.

[0156] The rerank model is used to rerank the conditions and actions to be screened respectively, and the basis for the reranking is the semantic relevance between the conditions and actions to be screened and the control statement, thereby screening the top k candidate conditions and top k candidate actions.

[0157] Then, based on the base Qianwen 7B model, the condition selection LoRA, action selection LoRA, parameter filling LoRA of conditions and actions, and logical relationship and reflection LoRA are loaded to adjust the capabilities of the base model.

[0158] Based on the control statement input by the user and the top k candidate conditions, the target conditions matching the control statement are screened out; based on the control statement and the top k candidate actions, the target actions matching the control statement are screened out.

[0159] Based on the control statement, target conditions and target actions, the corresponding parameter information is retrieved through the base model, and the parameters required in the user's control statement are attached to the target conditions and target actions.

[0160] Based on the control statements, target conditions and condition parameters, target actions and action parameters, corresponding target text instructions are generated for the target conditions and target actions according to the logical requirements in the control statements input by the user, and the target text instructions are transmitted back to the terminal to execute the corresponding functions.

[0161] In the technical solution of the above-mentioned specific implementation method, the layered architecture design of the database, rearrangement model, and large language model adopted is more flexible than the end-to-end large model, and the strategy methods at each level are adjusted step by step to gradually achieve the optimal results. The retrieval based on the vector library and the filtering of the rearrangement model greatly reduces the illusion of the large model and greatly improves the accuracy. Switching from the original 72B large model to the 7B large model, the amount of data has been greatly reduced, the training cost and the reasoning cost have been greatly reduced, and the efficiency has also been greatly improved. The fine-tuning of LoRA can make the format of the model output very stable, and the accuracy of the output content is also greatly improved.

[0162] It should be noted that the above describes some embodiments of the present application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims can be performed in a different order than in the above embodiments and still achieve the desired results. In addition, the processes depicted in the drawings are not necessarily

[0163] The particular order shown, or sequential order, is not required to achieve desirable results.Multitasking and parallel processing are also possible or may be advantageous in certain implementations.

[0164] Based on the same inventive concept, corresponding to any of the above-mentioned embodiment methods, the present application also provides a function control device.

[0165] Figure 2 A schematic diagram of a functional control device provided in an embodiment of the present application, referring to Figure 2 , the function control device 200 may include a statement receiving module 210, a model input module 220, an instruction output module 230 and a function execution module 240, wherein:

[0166] A statement receiving module 210, configured to receive an input control statement and determine at least two types of candidate parameters matching the control statement;

[0167] A model input module 220, for inputting a control statement and at least two types of candidate parameters into a preset base language model;

[0168] The instruction output module 230 is used to determine the target parameter from each candidate parameter based on the base language model, and determine the text parameter value matching the target parameter, combine the target parameter and the text parameter value to obtain the target text instruction and output it;

[0169] The function execution module 240 is used to execute the function corresponding to the target text instruction.

[0170] The technical solution of the embodiment of the present application determines the target parameter from the candidate parameters through a preset base language model, identifies the control statement input by the user according to different parameter dimensions, provides inspiration for the vehicle to identify the user's intention, identifies the query statement in different dimensions, and processes the parameters based on different dimensions in a targeted manner instead of directly identifying the entire sentence, thereby solving the problem of low efficiency caused by directly performing large-scale language analysis on the entire sentence, and can effectively improve the efficiency and accuracy of sentence recognition; then determine the text parameter value matching the target parameter, and then combine the target parameter and the text parameter to generate a target text instruction to control the vehicle to perform the corresponding function, use the base language model as a basis, search and determine the text parameter value, and combine it with conditions and actions, and under the inspirational guidance of the parameters, process the key information in the query statement in a targeted manner, which can further give play to the advantages of the large-scale language analysis of the base language model, improve the efficiency of instruction generation, and reduce the consumption of computing resources. Therefore, the technical solution of the embodiment of the present application can effectively balance the accuracy of instruction retrieval matching and the efficiency of instruction generation.

[0171] In an optional implementation, the statement receiving module 210 may include:

[0172] A vectorization unit, used to vectorize the control statement to obtain a query vector;

[0173] A parameter determination unit to be screened, used for inputting a query vector into at least two types of parameter retrieval databases, and respectively determining at least two types of parameters to be screened corresponding to the control statement;

[0174] The candidate condition screening unit is used to execute, for each of at least two categories as a target class: sorting the parameters to be screened of the target class to obtain a sorting result of the parameters to be screened, and screening multiple candidate parameters from the sorting result of the parameters to be screened.

[0175] In an optional implementation, the to-be-screened parameter determination unit may include:

[0176] A parameter similarity determination subunit is used to input the query vector into the parameter retrieval database and calculate the parameter similarity between each parameter vector in the parameter retrieval database and the query vector;

[0177] The parameter sorting subunit is used to sort each parameter vector according to the parameter similarity to obtain the parameter similarity sorting result;

[0178] The parameter screening subunit is used to select a first preset number of parameter vectors with the largest parameter similarity from the parameter similarity sorting results as parameters to be screened.

[0179] In an optional implementation, the candidate parameter screening unit may include:

[0180] The parameter correlation sorting subunit is used to determine the parameter semantic correlation of each parameter to be screened and the control statement, and sort each parameter to be screened according to the parameter semantic correlation to obtain the sorting result of the parameter to be screened;

[0181] The candidate parameter determination subunit is used to select a second preset number of parameters to be screened with the greatest semantic relevance from the sorting results of the parameters to be screened as multiple candidate parameters.

[0182] In an optional implementation, the base language model is determined by:

[0183] Obtain a low-rank matrix obtained by training based on a preset amount of vehicle domain data;

[0184] Multiply the low-rank matrix with the intrinsic parameter matrix in the initial base model to obtain the optimized parameter matrix;

[0185] The optimized parameter matrix is ​​reloaded into the initial base model to obtain the base language model.

[0186] In a further optional embodiment, the low-rank matrix includes a parameter selection matrix, and the optimization parameter matrix in the base language model is determined by the parameter selection matrix;

[0187] The instruction output module 230 may include:

[0188] A parameter matching degree determination unit, used to determine the parameter matching degree between each candidate parameter and the control sentence based on the base language model;

[0189] The target parameter determination unit is used to take the candidate parameter with the highest parameter matching degree as the target parameter.

[0190] In another optional implementation, the low-rank matrix includes a parameter filling matrix, and the optimized parameter matrix in the base language model is determined by the parameter filling matrix;

[0191] The instruction output module 230 may include:

[0192] A text information search unit, used for searching text information that matches the control sentence from the base language model based on the base language model;

[0193] The text parameter value determination unit is used to take the parameter in the text information that matches the target parameter as the text parameter value.

[0194] In yet another optional implementation, the low-rank matrix includes a logical relationship matrix, and the optimization parameter matrix in the base language model is determined by the logical relationship matrix;

[0195] The instruction output module 230 may further include:

[0196] A logical correspondence determination unit, used for adjusting the logical correspondence between the target parameter and the text parameter value based on the base language model;

[0197] The target text instruction generating unit is used to generate and output the target text instruction according to the logical corresponding relationship.

[0198] For the convenience of description, the above device is described in terms of functions divided into various modules. Of course, when implementing the present application, the functions of each module can be implemented in the same or multiple software and / or hardware.

[0199] The device of the above embodiment is used to implement the corresponding function control method in any of the above embodiments, and has the beneficial effects of the corresponding method embodiment, which will not be described in detail here.

[0200] Based on the same inventive concept, corresponding to any of the above-mentioned embodiments and methods, the present application also provides an electronic device, including 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 function control method described in any of the above embodiments is implemented.

[0201] Figure 3 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application, Figure 3 A specific electronic device hardware structure is shown, and the device may include: a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040 are connected to each other through the bus 1050 in the device.

[0202] The processor 1010 can be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0203] The memory 1020 may be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 may store an operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program codes are stored in the memory 1020 and are called and executed by the processor 1010.

[0204] The input / output interface 1030 is used to connect the input / output module to realize information input and output. The input / output module can be configured in the device as a component (not shown in the figure), or it can be externally connected to the device to provide corresponding functions. The input device may include a keyboard, a mouse, a touch screen, a microphone, various sensors, etc., and the output device may include a display, a speaker, a vibrator, an indicator light, etc.

[0205] The communication interface 1040 is used to connect a communication module (not shown) to realize communication interaction between the device and other devices. The communication module can realize communication through a wired mode (such as USB, network cable, etc.) or a wireless mode (such as mobile network, WIFI, Bluetooth, etc.).

[0206] The bus 1050 includes a path that transmits information between the various components of the device (eg, the processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040).

[0207] It should be noted that, although the above device only shows the processor 1010, the memory 1020, the input / output interface 1030, the communication interface 1040 and the bus 1050, in the specific implementation process, the device may also include other components necessary for normal operation. In addition, it can be understood by those skilled in the art that the above device may also only include the components necessary for implementing the embodiments of the present specification, and does not necessarily include all the components shown in the figure.

[0208] The electronic device of the above embodiment is used to implement the corresponding function control method in any of the above embodiments, and has the beneficial effects of the corresponding method embodiment, which will not be described in detail here.

[0209] Based on the same inventive concept, corresponding to any of the above-mentioned embodiments and methods, the present application further provides a vehicle, the vehicle comprising:

[0210] A memory for storing executable program codes;

[0211] A processor is used to call and run the executable program code from the memory so that the vehicle executes the method provided in any embodiment of the present application.

[0212] Based on the same inventive concept, corresponding to any of the above-mentioned embodiment methods, the present application also provides a non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to execute the function control method described in any of the above embodiments.

[0213] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, read-only compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device.

[0214] The computer instructions stored in the storage medium of the above embodiments are used to enable the computer to execute the function control method described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0215] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present application (including the claims) is limited to these examples. In line with the concept of the present application, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of the different aspects of the embodiments of the present application as described above, which are not provided in detail for the sake of simplicity.

[0216] In addition, to simplify the description and discussion, and in order not to make the embodiments of the present application difficult to understand, the known power supply / ground connection with the integrated circuit (IC) chip and other components may or may not be shown in the provided drawings. In addition, the device can be shown in the form of a block diagram to avoid making the embodiments of the present application difficult to understand, and this also takes into account the fact that the details of the implementation of these block diagram devices are highly dependent on the platform to be implemented in the embodiments of the present application (that is, these details should be fully within the scope of understanding of those skilled in the art). In the case of elaborating specific details (e.g., circuits) to describe exemplary embodiments of the present application, it is obvious to those skilled in the art that the embodiments of the present application can be implemented without these specific details or when these specific details are changed. Therefore, these descriptions should be considered to be illustrative rather than restrictive.

[0217] Although the present application has been described in conjunction with specific embodiments of the present application, many replacements, modifications and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may use the embodiments discussed.

[0218] The embodiments of the present application are intended to cover all such substitutions, modifications and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present application should be included in the scope of protection of the present application.

Claims

1. A function control method, characterized in that: The method comprises: receiving an input control statement, and determining at least two types of candidate parameters matching the control statement; Inputting the control sentence and at least two types of candidate parameters into a preset base language model; Based on the base language model, determine a target parameter from each of the candidate parameters, and determine a text parameter value matching the target parameter, combine the target parameter and the text parameter value to obtain a target text instruction and output it; Execute the function corresponding to the target text instruction.

2. The method according to claim 1, characterized in that The determining of at least two types of candidate parameters matching the control statement includes: Vectorizing the control statement to obtain a query vector; Inputting the query vector into at least two types of parameter retrieval databases to respectively determine at least two types of parameters to be screened corresponding to the control statement; For each of the at least two categories, it is respectively executed as a target class: the parameters to be screened of the target class are sorted to obtain a sorting result of the parameters to be screened, and a plurality of candidate parameters of the target class are screened from the sorting result of the parameters to be screened.

3. The method according to claim 2, characterized in that The step of inputting the query vector into at least two types of parameter retrieval data to respectively determine at least two types of parameters to be screened corresponding to the control statement includes: For each of at least two types of parameter retrieval databases, input the query vector into the parameter retrieval database, and calculate the parameter similarity between each parameter vector in the parameter retrieval database and the query vector; Sorting each of the parameter vectors according to the parameter similarity to obtain a parameter similarity sorting result; The parameter vectors having the largest first preset number of parameter similarities are selected from the parameter similarity sorting results as the parameters to be screened.

4. The method according to claim 2, characterized in that: The step of sorting the parameters to be screened of the target class to obtain a sorting result of the parameters to be screened, and screening a plurality of candidate parameters of the target class from the sorting result of the parameters to be screened includes: Determine the parameter semantic relevance between each of the parameters to be screened and the control statement, and sort each of the parameters to be screened according to the parameter semantic relevance to obtain a sorting result of the parameters to be screened; A second preset number of parameters to be screened with the greatest semantic relevance are selected from the sorting results of the parameters to be screened as the multiple candidate parameters.

5. The method according to claim 1, characterized in that The base language model is determined by: Obtain a low-rank matrix obtained by training based on a preset amount of vehicle domain data; Multiplying the low-rank matrix with the intrinsic parameter matrix in the initial base model to obtain an optimized parameter matrix; The optimized parameter matrix is ​​reloaded into the initial base model to obtain the base language model.

6. The method according to claim 5, characterized in that The low-rank matrix includes a parameter selection matrix, and the optimization parameter matrix in the base language model is determined by the parameter selection matrix; The step of determining a target parameter from each of the candidate parameters based on the base language model includes: Based on the base language model, determining the parameter matching degree between each of the candidate parameters and the control sentence; The candidate parameter with the highest matching degree to the condition is used as the target parameter.

7. The method according to claim 5, characterized in that The low-rank matrix includes a parameter filling matrix, and the optimization parameter matrix in the base language model is determined by the parameter filling matrix; The step of determining a text parameter value matching the target parameter based on the base language model includes: Based on the base language model, searching the base language model for text information that matches the control sentence; The parameter in the text information that matches the target parameter is used as the text parameter value.

8. The method according to claim 5, characterized in that The low-rank matrix includes a logical relationship matrix, and the optimization parameter matrix in the base language model is determined by the logical relationship matrix; The step of generating and outputting a target text instruction based on the target parameter and the text parameter value comprises: Based on the base language model, adjusting the logical correspondence between the target parameter and the text parameter value; According to the logical correspondence, the target text instruction is generated and output.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 8 is implemented.

10. A vehicle, characterized in that: The vehicle comprises: A memory for storing executable program codes; A processor, configured to call and run the executable program code from the memory so that the vehicle executes the method according to any one of claims 1 to 8.

Images