Optimization method, voice interaction method, server and computer readable storage medium

Abstract

The application discloses a method for optimizing a model for vehicle voice interaction, a voice interaction method, a server and a computer readable storage medium. The method comprises: first, performing a rejection modification process on an original training data set of the model to obtain a first training data set and a second training data set. Then, a third training data set is obtained according to the original training data set and an original log data set. Finally, the model is trained according to the original training data set, the first training data set, the second training data set and the third training data set to improve the rejection capability of the model. In this way, by performing a rejection modification process on the original training data set and performing a specific process according to the original training data set and the original log data set, the number of rejection training data sets is increased and the scope of the rejection training data sets is widened, and the rejection capability of the model is improved through the rejection training data sets, so that the model can be prevented from executing a request that may cause danger incorrectly.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of voice interaction, and in particular to a model optimization method for vehicle voice interaction, a voice interaction method, a server and a computer readable storage medium. BACKGROUND

[0002] The in-vehicle intelligent system often trains its rejection ability by inputting rejection training data, so as to improve the ability of the system to reject inappropriate or invalid inputs in actual application. However, due to the insufficient number of rejection training data and the low breadth of rejection training data, the rejection ability of the in-vehicle intelligent system is poor. SUMMARY

[0003] The present application provides a model optimization method for vehicle voice interaction, a voice interaction method, a server and a computer readable storage medium.

[0004] The present application provides a model optimization method for vehicle voice interaction, the method comprising:

[0005] performing rejection modification processing on the original training data set of the model to obtain a first training data set and a second training data set;

[0006] obtaining a third training data set according to the original training data set and an original log data set;

[0007] training the model according to the original training data set, the first training data set, the second training data set and the third training data set to improve the rejection ability of the model.

[0008] In this way, the server performs rejection modification processing on the original training data set of the model to obtain a first training data set and a second training data set. Then, the server obtains a third training data set according to the original training data set and an original log data set. Finally, the server trains the model according to the original training data set, the first training data set, the second training data set and the third training data set to improve the rejection ability of the model. In this way, by performing rejection modification processing on the original training data set and performing specific processing according to the original training data set and the original log data set, the number of rejection training data sets is increased and the breadth of the rejection training data sets is widened, and the rejection ability of the model is improved through these rejection training data, so that the model can be prevented from executing a request that may cause danger incorrectly.

[0009] In some embodiments, the rejection modification processing on the original training data set of the model to obtain a first training data set and a second training data set comprises:

[0010] The first training data set is obtained based on a preset test method according to the preset rejection description, an original rejection data subset and an original valid data subset in the original training data set.

[0011] In this way, the server obtains the first training data set based on a preset test method according to the preset rejection description, an original rejection data subset and an original valid data subset in the original training data set. In this way, the server uses the preset test method and the preset rejection description, and combines the original rejection data subset and the original valid data subset in the original training data set to obtain the first training data set, thereby helping the model to better understand under what circumstances the input should be rejected for processing, and improving the safety and reliability of the model in actual application.

[0012] In some embodiments, the original valid data subset includes a plurality of original valid data, and the first training data set is obtained based on a preset test method according to the original training data set and a preset rejection description, including:

[0013] Each of the original valid data in the original valid data subset is segmented based on a preset granularity to obtain a plurality of target rejection data, wherein each of the target rejection data corresponds to one of the original valid data;

[0014] A plurality of temporary rejection data sets are obtained according to the preset rejection description and the plurality of target rejection data;

[0015] The temporary rejection data sets are tested according to the original rejection data subset to confirm that at least one of the temporary rejection data sets is added to the first training data set.

[0016] In this way, the server segments each of the original valid data in the original valid data subset based on a preset granularity to obtain a plurality of target rejection data, wherein each of the target rejection data corresponds to one of the original valid data. Then, the server obtains a plurality of temporary rejection data sets according to the preset rejection description and the plurality of target rejection data. Finally, the server tests the temporary rejection data sets according to the original rejection data subset to confirm that at least one of the temporary rejection data sets is added to the first training data set. In this way, through detailed data processing and testing, it is ensured that the generated first training data set can effectively train the model, so that the model can run more safely and reliably in actual application.

[0017] In some embodiments, the target rejection data includes at least one sentence, and the plurality of temporary rejection data sets are obtained according to the preset rejection description and the target rejection data, including:

[0018] The preset rejection description is respectively placed at a segmentation position in each of the target rejection data, to obtain a plurality of temporary rejection data sets, and the segmentation position is a connection position of each segment in the target rejection data.

[0019] In this way, the server places the preset rejection description at the segmentation position in each of the target rejection data, to obtain a plurality of temporary rejection data sets, and the segmentation position is the connection position of each segment in the target rejection data. In this way, the server places the preset rejection description at the segmentation position of each first rejection data, creates a plurality of new data sets, i.e., temporary rejection data sets, and generates a first training data set according to the temporary rejection data set, to train the model and improve the rejection ability of the model in actual application.

[0020] In some embodiments, the rejection modification processing of the original training data set of the model obtains a first training data set and a second training data set, including:

[0021] The original effective data subset is modified to obtain the second training data set.

[0022] In this way, the server modifies the original effective data subset to obtain the second training data set. In this way, the second training data set is obtained, which can help the model better understand under what circumstances the input should be rejected for processing, thereby improving the security and reliability of the model in actual application.

[0023] In some embodiments, the modification processing of the original effective data subset obtains the second training data set, including:

[0024] The original effective data in the original effective data subset is parsed to obtain target information to be modified;

[0025] The target information to be modified is modified to obtain target error information;

[0026] The second training data set is obtained according to the target error information.

[0027] In this way, the server parses the original effective data in the original effective data subset to obtain target information to be modified. Then, the server modifies the target information to be modified to obtain target error information. Finally, the server obtains the second training data set according to the target error information. In this way, the second training data set is generated by replacing the target information to be modified with the generated target error information, which can train the model to better understand under what circumstances the input should be rejected for processing, thereby improving the security and reliability of the model in actual application.

[0028] In some embodiments, the obtaining, according to the original training data set and the original log data set, a third training data set comprises:

[0029] obtaining, according to the log data in the original log data set and the original effective data subset in the original training data set, a similarity of each of the log data based on a preset similarity algorithm;

[0030] determining the third training data set according to the preset similarity threshold and the similarity.

[0031] In this way, the server obtains, according to the log data in the original log data set and the original effective data subset in the original training data set, a similarity of each of the log data based on a preset similarity algorithm. Then, the server determines the third training data set according to the preset similarity threshold and the similarity. In this way, the log data that is not similar or abnormal to the original effective data subset is obtained, which can be used to train the model to improve the ability of the model to reject inappropriate or abnormal input in actual application.

[0032] In some embodiments, the original effective data subset comprises a plurality of original effective data, and the obtaining, according to the log data in the original log data set and the original effective data subset in the original training data set, a similarity of each of the log data based on a preset similarity algorithm comprises:

[0033] determining a log data vector of the log data according to the preset similarity algorithm and the log data;

[0034] determining an original effective data vector of the original effective data according to the preset similarity algorithm and the original effective data;

[0035] obtaining the similarity according to the log data vector and the original effective data vector.

[0036] In this way, the server determines a log data vector of the log data according to the preset similarity algorithm and the log data. Then, the server determines an original effective data vector of the original effective data according to the preset similarity algorithm and the original effective data. Finally, the server obtains the similarity according to the log data vector and the original effective data vector. In this way, the server can evaluate the similarity between the log data and the original effective data, so as to better understand the relationship between the data and provide a basis for subsequent data processing and analysis.

[0037] In some embodiments, the determining, according to the preset similarity algorithm and the log data, a log data vector of the log data comprises:

[0038] determining a word vector in the log data according to the log data;

[0039] determine a target word according to the log data;

[0040] obtain a TF-IDF weight of the target word according to the target word and the original effective data subset;

[0041] obtain a log data vector of the log data according to the word vector and the TF-IDF weight.

[0042] In this way, the server determines a word vector in the log data according to the log data. Then, the server determines a target word according to the log data. Next, the server obtains a TF-IDF weight of the target word according to the target word and the original effective data subset. Finally, the server obtains a log data vector of the log data according to the word vector and the TF-IDF weight. In this way, a log data vector composed of TF-IDF weights of the target word is obtained, which can effectively represent the characteristics and content of the log data, so that the content and theme of the data are better understood for subsequent similarity calculation.

[0043] In some embodiments, the training of the model according to the original training data set, the first training data set, the second training data set and the third training data set to improve the recognition ability of the model comprises:

[0044] obtaining a recognition training data set according to the second training data set and the third training data set;

[0045] training the model according to the original training data set, the first training data set and the recognition training data set based on a preset proportion to improve the recognition ability of the model.

[0046] In this way, the server obtains a recognition training data set according to the second training data set and the third training data set. Then, the server trains the model according to the original training data set, the first training data set and the recognition training data set based on a preset proportion to improve the recognition ability of the model. In this way, the server determines the usage amount of data in the original training data set, the first training data set and the recognition training data set based on a preset proportion to achieve good training effect, thereby improving the ability of the model to reject inappropriate or abnormal input in actual application.

[0047] The embodiments of the present application provide a voice interaction method based on the model optimized by the above optimization method, and the method comprises:

[0048] receiving a voice request forwarded by a vehicle;

[0049] The voice request is recognized according to the model to complete the voice interaction.

[0050] In this way, the server receives the voice request forwarded by the vehicle. Then, the server recognizes the voice request according to the model to complete the voice interaction. In this way, based on the model trained by the above optimization method, the server can ensure that the voice interaction between the vehicle and the user is both safe and reliable.

[0051] The embodiment of the present application provides a server, which comprises a processor and a memory, and the memory stores a computer program, and the computer program is executed by the processor to realize the optimization method or the voice interaction method.

[0052] The embodiment of the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by the processor to realize the steps of the optimization method or the voice interaction method.

[0053] Additional aspects and advantages of the embodiments of the present application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0054] The above and / or additional aspects and advantages of the present application will become apparent and be readily appreciated from the following description, including the references to the figures, in which:

[0055] Figure 1 is one of the flowcharts of the optimization method of the model of some embodiments of the present application;

[0056] Figure 2 is another of the flowcharts of the optimization method of the model of some embodiments of the present application;

[0057] Figure 3 is a third of the flowcharts of the optimization method of the model of some embodiments of the present application;

[0058] Figure 4 is a fourth of the flowcharts of the optimization method of the model of some embodiments of the present application;

[0059] Figure 5 is a fifth of the flowcharts of the optimization method of the model of some embodiments of the present application;

[0060] Figure 6 is a sixth of the flowcharts of the optimization method of the model of some embodiments of the present application;

[0061] Figure 7 is a seventh of the flowcharts of the optimization method of the model of some embodiments of the present application;

[0062] Figure 8 Figure 8 is a flowchart of an optimization method of a model according to some embodiments of the present application;

[0063] Figure 9 Figure 9 is a flowchart of an optimization method of a model according to some embodiments of the present application;

[0064] Figure 10 Figure 10 is a flowchart of an optimization method of a model according to some embodiments of the present application;

[0065] Figure 11 Figure 11 is a flowchart of a model training data processing method according to some embodiments of the present application;

[0066] Figure 12 Figure 12 is a flowchart of a voice interaction method according to some embodiments of the present application. DETAILED DESCRIPTION

[0067] Embodiments of the present application are described in detail below with reference to the accompanying drawings, in which the same or similar components have the same or similar reference numbers throughout the various figures. The embodiments described below are examples of the present application, and are not intended to limit the present application.

[0068] The on-board intelligent system usually trains its rejection ability through input rejection training data to improve the system's ability to reject inappropriate or invalid inputs in actual application. However, due to the lack of rejection training data and the low breadth of rejection training data, the rejection ability of the on-board intelligent system is poor, which may lead to a decrease in the safety of the on-board intelligent system. In this way, the on-board intelligent system may have the following problems in actual application:

[0069] First, due to the lack of rejection training data, the on-board intelligent system may not be able to accurately identify inappropriate or invalid inputs, resulting in misidentification. For example, the on-board intelligent system may mistakenly identify the driver's request as a valid request and perform the wrong operation.

[0070] Second, the low breadth of rejection training data may result in the on-board intelligent system failing to cover all possible inappropriate or invalid input situations. Therefore, the system may miss some inappropriate or invalid inputs, resulting in the inability to reject processing these inputs.

[0071] Based on the above problems, please refer to Figure 1 The embodiments of the present application provide an optimization method of a model for vehicle voice interaction, the method comprising:

[0072] 01: The original training data set of the model is subjected to a rejection modification process to obtain a first training data set and a second training data set;

[0073] 02: According to the original training data set and the original log data set, a third training data set is obtained;

[0074] 03: The original training data set, the first training data set, the second training data set and the third training data set are used to train the model to improve the rejection ability of the model.

[0075] The voice interaction method of the embodiment of the application can be implemented by the server of the embodiment of the application. Specifically, the memory stores a computer program, and the processor is configured to perform rejection modification on the original training data set of the model to obtain a first training data set and a second training data set. According to the original training data set and the original log data set, a third training data set is obtained. And according to the original training data set, the first training data set, the second training data set and the third training data set, the model is trained to improve the rejection ability of the model.

[0076] The embodiment of the application also provides a model optimization device. The voice interaction method of the embodiment of the application can be implemented by the model optimization device of the embodiment of the application. Specifically, the model optimization device includes a rejection modification module, a processing module and a training module. The rejection modification module is configured to perform rejection modification on the original training data set of the model to obtain a first training data set and a second training data set. The processing module is configured to obtain a third training data set according to the original training data set and the original log data set. The training module is configured to train the model according to the original training data set, the first training data set, the second training data set and the third training data set to improve the rejection ability of the model.

[0077] Specifically, the original training data set refers to a set of data used to train a machine learning model, which usually includes input samples and corresponding output labels, and is used to let the model learn how to extract features from input data and make correct predictions or decisions. The original training data set usually includes original valid data that can output correct responses and original rejection data that the system needs to reject, most of which are original valid data. In the context of a vehicle-mounted intelligent system, the original training data set may include various driving scenarios, driver requests, vehicle states and other information, which is used to train the model to identify and respond to different driving situations.

[0078] The rejection reforming process refers to the process of processing the original training dataset to generate the first training dataset and the second training dataset. Through this rejection reforming process, the original valid data in the original training dataset is converted into rejection data for training the machine learning model to improve its rejection ability. The rejection data usually includes specific requests and corresponding rejection labels, as well as various inappropriate or invalid inputs, and instructions on how the model should reject the processing when encountering these inputs. The methods of rejection reforming process may include data augmentation, etc., to improve the model's ability to reject inappropriate or invalid inputs in actual application.

[0079] The first training dataset refers to the training dataset obtained by performing rejection reforming on the original training dataset according to the preset rejection description.

[0080] The second training dataset refers to the training dataset obtained by directly reforming the original training dataset, including different input samples and corresponding output labels from the first training dataset, for further improving the performance and generalization ability of the model. By using the second training dataset, the model can learn more features and patterns, thereby improving its performance in actual application.

[0081] The third training dataset includes log data selected from the original log dataset, which has a low similarity to the data in the original training dataset, including different input samples and corresponding output labels from the first two training datasets, for further improving the performance and generalization ability of the model.

[0082] The rejection ability of the model refers to the ability of the model to recognize and reject processing of unfamiliar, abnormal or inappropriate inputs, thereby ensuring the safety and reliability of the model in actual application.

[0083] The server performs rejection reforming on the original training dataset of the model to obtain the first training dataset and the second training dataset.

[0084] Then, the server obtains the third training dataset according to the original training dataset and the original log dataset. That is, the original log dataset may include samples input by users in actual application, which may contain erroneous or abnormal inputs. By analyzing these log data, problems that the model may encounter in actual application can be identified, and rejection samples can be generated accordingly.

[0085] Finally, the server trains the model according to the original training dataset, the first training dataset, the second training dataset and the third training dataset to improve the rejection ability of the model, so that the model learns to correctly reject processing when encountering inputs that are unfamiliar or not suitable for the current task, instead of trying to make predictions or decisions that may lead to errors.

[0086] In summary, the method for optimizing a model for voice interaction of a vehicle and the server provided by the embodiments of the present application are configured to perform recognition modification processing on an original training data set of the model to obtain a first training data set and a second training data set. Then, the server obtains a third training data set according to the original training data set and an original log data set. Finally, the server trains the model according to the original training data set, the first training data set, the second training data set, and the third training data set to improve the recognition ability of the model. In this way, by performing recognition modification processing on the original training data set and performing specific processing according to the original training data set and the original log data set, the number of recognition training data sets is increased and the scope of the recognition training data sets is widened, and the recognition ability of the model is improved through these recognition training data sets, thereby preventing the model from incorrectly executing a request that may cause danger.

[0087] Please refer to Figure 2 In some embodiments, step 01 (performing recognition modification processing on the original training data set of the model to obtain the first training data set and the second training data set) includes:

[0088] 011: obtaining the first training data set according to the preset recognition description, the original recognition data subset, and the original effective data subset in the original training data set based on a preset testing method.

[0089] In some embodiments, the recognition modification module is further configured to obtain the first training data set according to the preset recognition description, the original recognition data subset, and the original effective data subset in the original training data set based on a preset testing method.

[0090] In some embodiments, the processor is further configured to obtain the first training data set according to the preset recognition description, the original recognition data subset, and the original effective data subset in the original training data set based on a preset testing method.

[0091] Specifically, the preset testing method (Predefined Testing Methods) refers to a designed test scheme that can be used to generate the first training data set, process and convert the original recognition data subset and the original in the original training data set to generate the first training data set.

[0092] The preset rejection description refers to a set of predefined rules or conditions used to identify and classify which input data should be rejected by the model. In the context of a vehicle-mounted system, these descriptions are usually based on an understanding of the task or problem, as well as prior knowledge of the data set. The preset rejection description can ensure that the model can correctly reject inappropriate or invalid inputs when faced with them. For example, "if the request does not belong to the above functions, terminate the process and directly output the rejection label", "if the location name contained in the user's voice request does not exist in the system's location database, terminate the process and directly output the rejection label", and "if the user's voice request contains a request that the system does not support, terminate the process and directly output the rejection label" are all preset rejection descriptions.

[0093] The original rejection data subset refers to a part of the original training data set that is separated out and used to test and evaluate the model's ability to reject, including input data that the model should reject, such as ambiguous requests, format errors, requests outside the system's function range, etc. The original rejection data subset can verify whether the model can correctly identify and reject these inappropriate or invalid inputs, thereby verifying whether the model can correctly identify and reject these inappropriate or invalid inputs. For example, in a vehicle-mounted system, the original rejection data subset may include non-existent location names, unsupported requests, ambiguous voice inputs, etc.

[0094] The original valid data subset refers to a part of the original training data set that is separated out and can be correctly understood and processed, including input data that the model should normally process, such as clear requests, format correct requests, requests within the system's function range, etc. The original valid data subset is used to train the model, so that it can learn how to correctly respond and process these valid inputs, so that the system can establish the ability to identify and process valid inputs, thereby effectively responding to user requests and requests in actual application. For example, in a vehicle-mounted system, the original valid data subset may include navigation to a specific location, playing music, adjusting the temperature in the car, etc.

[0095] The server uses the preset test method and the preset rejection description, combined with the original rejection data subset and the original valid data subset in the original training data set, to generate a first rejection data set. For example, the original valid data is: "Please identify and execute the corresponding operation according to the following driver's voice request: 'Please open the window.'", and after adding the preset rejection description, it may be: "Please identify and execute the corresponding operation according to the following driver's voice request: 'Please open the window.' If the requested device is not the window, terminate the process and directly output the rejection label".

[0096] Thus, the server uses the preset test method and the preset rejection description, and combines the original rejection data subset and the original valid data subset in the original training data set to obtain the first training data set, thereby helping the model to better understand under what circumstances the input should be rejected for processing, and improving the safety and reliability of the model in actual application.

[0097] Referring to Figure 3 In some embodiments, the original valid data subset includes a plurality of original valid data, and step 011 (obtaining the first training data set based on the preset test method and the original training data set and the preset rejection description) includes:

[0098] 0111: performing segmentation processing on each original valid data in the original valid data subset based on a preset granularity to obtain a plurality of target rejection data;

[0099] 0112: obtaining a plurality of temporary rejection data sets according to the preset rejection description and the plurality of target rejection data;

[0100] 0113: performing test processing on the temporary rejection data sets according to the original rejection data subset, and confirming to add at least one of the temporary rejection data sets to the first training data set.

[0101] In some embodiments, the rejection modification module is further configured to perform segmentation processing on each original valid data in the original valid data subset based on a preset granularity to obtain a plurality of target rejection data, obtain a plurality of temporary rejection data sets according to the preset rejection description and the plurality of target rejection data, and perform test processing on the temporary rejection data sets according to the original rejection data subset, and confirm to add at least one of the temporary rejection data sets to the first training data set.

[0102] In some embodiments, the processor is further configured to perform segmentation processing on each original valid data in the original valid data subset based on a preset granularity to obtain a plurality of target rejection data, obtain a plurality of temporary rejection data sets according to the preset rejection description and the plurality of target rejection data, and perform test processing on the temporary rejection data sets according to the original rejection data subset, and confirm to add at least one of the temporary rejection data sets to the first training data set.

[0103] Specifically, the preset granularity refers to a minimum unit of data segmentation or processing defined by the server in advance when processing the original valid data subset, which can be a paragraph, a sentence, a word, etc. In the embodiments of the present application, a sentence is taken as the preset granularity for illustration.

[0104] The target rejection data refers to data obtained by performing segmentation processing on the original valid data based on the preset granularity.

[0105] A temporary rejection dataset refers to a collection of temporary rejection data generated based on target rejection data and a preset rejection description. Each target rejection data corresponds to a temporary rejection dataset. In addition, the temporary rejection data in a temporary rejection dataset is not necessarily effective.

[0106] Based on a preset granularity, the server segments each raw valid data in the raw valid data subset to generate multiple target rejection data, where each target rejection data corresponds to a single raw valid data. Specifically, the raw valid data prompt is segmented at the sentence granularity level (P_prompt = P_1 + P_2 + ... + P_k, where P_1 is the first sentence of the prompt), where P_1 + P_2 + ... + P_k is a single target rejection data. For example, if the raw valid data is: "Please recognize and perform the corresponding action based on the following driver's voice request: 'Please open the window.'" , the corresponding target rejection data generated might be: "[]Please recognize and perform the corresponding action based on the following driver's voice request: []'Please open the window.'[]".

[0107] Next, the server generates multiple temporary rejection datasets based on the preset rejection descriptions and multiple target rejection data. That is, the server adds corresponding preset rejection descriptions between sentences in the target rejection data (for P_i, P_rejection is generated, P_prompt rejectioni = P_1+P_2+...+P_i+P_rejection+...+P_k). Continuing with the above example, the generated temporary rejection dataset may include: "[If the request does not support the following functions, terminate the process and directly output the rejection label] Please recognize and perform the corresponding operation based on the following driver's voice request: 'Please open the window.'" It should be noted that corresponding preset rejection descriptions can be added between every two sentences, at the beginning, and at the end of each sentence in the target rejection data, which will not be repeated here.

[0108] The temporary rejection dataset may also include: "Please recognize and perform the corresponding action based on the following driver's voice request: 'Please open the window.' [If the request does not support the above functions, terminate the process and directly output the rejection label]."

[0109] Finally, the server tests the temporary rejected data sets based on the original rejected data subsets and confirms that at least one of the temporary rejected data sets should be added to the first training data set. Specifically, each of the temporary rejected data sets obtained above is tested using data from the original rejected data subsets to calculate the rejection accuracy rate. Then, based on specific requirements, one or more temporary rejected data sets with the highest rejection accuracy rates are selected from each temporary rejected data set and added to the first training data set.

[0110] Need to explain, usually in the original valid data beginning of the description task part and in the original valid data example few shot part add preset refusal description, refusal ability improvement effect is more obvious.

[0111] In this way, through careful data processing and testing, it is ensured that the generated first training data set can effectively train the model, so that the model can run more safely and reliably in actual application.

[0112] Please refer to Figure 4 In some embodiments, the target refusal data includes at least one sentence, and step 0112 (obtaining a plurality of temporary refusal data sets according to the preset refusal description and the target refusal data) includes:

[0113] 01121: Place the preset refusal description at the segmentation position in each target refusal data to obtain a plurality of temporary refusal data sets.

[0114] In some embodiments, the refusal modification module is further configured to place the preset refusal description at the segmentation position in each target refusal data to obtain a plurality of temporary refusal data sets.

[0115] In some embodiments, the processor is further configured to place the preset refusal description at the segmentation position in each target refusal data to obtain a plurality of temporary refusal data sets.

[0116] Specifically, the segmentation position refers to the position at which the sentences in the target refusal data are connected when processing the data. These positions are usually located after punctuation marks, spaces or other separators in the data, and are used to identify different parts or units in the data. For example, the original valid data is: "Please identify and execute the corresponding operation according to the following driver's voice request: 'Please open the window.'". Then the corresponding segmentation position is the position of the following "[]" box: "[] Please identify and execute the corresponding operation according to the following driver's voice request: [] 'Please open the window.' []".

[0117] The server places the preset refusal description at the segmentation position in each target refusal data to obtain a plurality of temporary refusal data sets, and the segmentation position is the connection position of each sentence in the target refusal data. That is, the preset refusal description is added at different segmentation positions of a target refusal data to obtain a temporary refusal data set, and a plurality of target refusal data sets obtain a plurality of temporary data sets.

[0118] In this way, the server places the preset refusal description at the segmentation position of each first refusal data to create a plurality of new data sets, i.e. temporary refusal data sets, so as to generate a first training data set according to the temporary refusal data set to train the model and improve the refusal ability of the model in actual application.

[0119] Referring to Figure 5 In some embodiments, step 01 (refusal modification processing on the original training data set of the model to obtain a first training data set and a second training data set) comprises:

[0120] 012: Modification processing on the original valid data subset to obtain a second training data set.

[0121] In some embodiments, the refusal modification module is further used for modification processing on the original valid data subset to obtain a second training data set.

[0122] In some embodiments, the processor is further used for modification processing on the original valid data subset to obtain a second training data set.

[0123] Specifically, modification processing refers to adding noise, modifying data features, creating outliers, etc., with the purpose of generating samples that the model should refuse to process.

[0124] Adding noise refers to introducing random or artificial interference in the speech data to simulate noisy environments in the real world. For example, background noise such as wind noise, engine sound or other people's conversation can be added to the driver's voice command to test the robustness of the model in different noise environments.

[0125] Modifying data features refers to intentionally changing some properties or features of the speech data to create new samples. For example, the range of capabilities of the vehicle-mounted voice assistant can be modified, or the tone, speed or volume of the voice can be changed, or different accents and pronunciation methods can be simulated to generate new training samples.

[0126] Creating outliers refers to changing part of the original valid data in the original valid data subset to generate some samples that are significantly different from most of the speech data in the original training data set.

[0127] The server modification processes the original valid data subset to obtain a second training data set.

[0128] In this way, the second training data set is obtained, which can help the model better understand under what circumstances it should refuse to process the input, thereby improving the safety and reliability of the model in actual application.

[0129] Referring to Figure 6 In some embodiments, step 012 (modification processing on the original valid data subset to obtain a second training data set) comprises:

[0130] 0121: parse the original valid data in the original valid data subset to obtain target information to be transformed;

[0131] 0122: transform the target information to be transformed to obtain target error information;

[0132] 0123: obtain a second training data set according to the target error information.

[0133] In some embodiments, the recognition transformation processing module is configured to parse the original valid data in the original valid data subset to obtain target information to be transformed, transform the target information to be transformed to obtain target error information, and obtain a second training data set according to the target error information.

[0134] In some embodiments, the processor is further configured to parse the original valid data in the original valid data subset to obtain target information to be transformed, transform the target information to be transformed to obtain target error information, and obtain a second training data set according to the target error information.

[0135] Specifically, the target information to be transformed refers to data information extracted from the original valid data subset, which will have a large change in meaning after modification. For example, the original valid data is “as an intelligent voice assistant, you need to play a navigator to provide the best play route for the user”, and the target information to be transformed can be “navigator” or “best play route”. If the original valid data is “please refer to the steps and examples of the above task to output the content in json format”, the target information to be transformed can be “json format”. If the original valid data is “move the chair of the co-pilot back”, the target information to be transformed can be “co-pilot” or “chair”.

[0136] The target error information refers to information obtained after transformation processing, which is an error sample that the model should refuse to process. For example, when the target information to be transformed is “navigator”, the corresponding target error information can be “psychological counselor”, “salesperson”, and “educator”, etc. When the target information to be transformed is “best play route”, the corresponding target error information can be “worst play route” and “best shopping route”, etc. When the target information to be transformed is “json format”, the corresponding target error information can be “JavaScript format” and “list format”, etc. When the target information to be transformed is “co-pilot”, the corresponding target error information can be “window” and “air conditioner”, etc. When the target information to be transformed is “chair”, the corresponding target error information can be “mobile phone” and “handbag”, etc.

[0137] The server parses the original valid data in the original valid data subset to obtain target information to be transformed. Then, the server transforms the target information to be transformed to obtain target error information.

[0138] Finally, the server obtains a second training data set according to the target error information. That is, the server replaces the target information to be transformed with the target error information to generate new training data. For example, replace the "best play route" with the "worst play route", and the original valid data becomes "as an intelligent voice assistant, you need to play a navigator and provide the worst play route for the user".

[0139] In this way, by using the generated target error information to replace the target information to be transformed to generate the second training data set, the model can be trained to better understand in which case it should refuse to process the input, thereby improving the safety and reliability of the model in actual application.

[0140] Please refer to Figure 7 In some embodiments, step 02 (obtaining a third training data set according to the original training data set and the original log data set) comprises:

[0141] 021: Based on a preset similarity algorithm, the similarity of each log data is obtained according to the log data in the original log data set and the original valid data subset in the original training data set.

[0142] 022: Determine the third training data set according to the preset similarity threshold and the similarity.

[0143] In some embodiments, the processing module is configured to obtain the similarity of each log data based on a preset similarity algorithm according to the log data in the original log data set and the original valid data subset in the original training data set. And determine the third training data set according to the preset similarity threshold and the similarity.

[0144] In some embodiments, the processor is further configured to obtain the similarity of each log data based on a preset similarity algorithm according to the log data in the original log data set and the original valid data subset in the original training data set. And determine the third training data set according to the preset similarity threshold and the similarity.

[0145] Specifically, the preset similarity algorithm refers to an algorithm used to calculate the similarity between the log data and the original effective data in the original effective data subset, including TF-IDF weighting (Term Frequency-Inverse Document Frequency), cosine similarity (Cosine Similarity), Jaccard similarity (Jaccard Similarity), and the like. For the convenience of illustration, the embodiments of the present application take TF-IDF weighting as an example.

[0146] The original log data set refers to the user voice request records recorded by the vehicle-mounted system, including various inputs encountered by the model in actual application.

[0147] The preset similarity threshold refers to a similarity threshold value set in advance when calculating the similarity. When the similarity between two texts exceeds this threshold, the two texts can be considered similar. If the similarity is lower than the threshold, they are considered dissimilar. In the embodiments of the present application, 0.5 is used as the preset similarity threshold, and other numerical values can also be used as the preset similarity threshold in some embodiments, which are not limited herein.

[0148] The server obtains the similarity of each log data based on the preset similarity algorithm according to the log data in the original log data set and the original effective data subset in the original training data set. In this way, each log data will obtain a similarity value, indicating its similarity degree with the known effective samples in the original training data.

[0149] Then, the server determines the third training data set according to the preset similarity threshold and the similarity. It is used to judge whether the log data is similar enough to the data in the original effective data subset. If the similarity of the log data is lower than the threshold, it can be considered as dissimilar or abnormal input, which is assembled with a predetermined prompt word, and then stored in the third training data set.

[0150] For example, suppose we have an original log data set containing user voice requests, and an original training data set containing known effective voice request samples. The server uses the similarity algorithm to calculate the similarity between each log data and the original effective data subset. Suppose we set a similarity threshold of 0.5. For each log data, if the calculated similarity value is lower than 0.5, for example, 0.3, the server will consider it as dissimilar or abnormal input. Then, the server assembles this log data with a predetermined prompt word (for example, “this is not a valid request”). Finally, the assembled data is stored in the third training data set for training the model’s rejection ability.

[0151] Thus, log data that is dissimilar or abnormal to the original valid data subset is obtained, which can be used to train the model to improve the ability of the model to reject inappropriate or abnormal inputs in actual application.

[0152] Referring to Figure 8 In some embodiments, the original valid data subset includes a plurality of original valid data, and step 021 (obtaining a similarity of each log data based on a preset similarity algorithm according to the log data in the original log data set and the original valid data subset in the original training data set) includes:

[0153] 0211: determining a log data vector of the log data according to the preset similarity algorithm and the log data;

[0154] 0212: determining an original valid data vector of the original valid data according to the preset similarity algorithm and the original valid data;

[0155] 0213: obtaining the similarity according to the log data vector and the original valid data vector.

[0156] In some embodiments, the processing module is configured to determine a log data vector of the log data according to the preset similarity algorithm and the log data, determine an original valid data vector of the original valid data according to the preset similarity algorithm and the original valid data, and obtain the similarity according to the log data vector and the original valid data vector.

[0157] In some embodiments, the processor is further configured to determine a log data vector of the log data according to the preset similarity algorithm and the log data, determine an original valid data vector of the original valid data according to the preset similarity algorithm and the original valid data, and obtain the similarity according to the log data vector and the original valid data vector.

[0158] Specifically, the server determines a log data vector of the log data according to the preset similarity algorithm and the log data. Then, the server determines an original valid data vector of the original valid data according to the preset similarity algorithm and the original valid data.

[0159] Finally, the server obtains the similarity according to the log data vector and the original valid data vector. This similarity value can be used for subsequent data processing, such as identifying inputs that are dissimilar or abnormal to the training data, to train the rejection ability of the model. In some embodiments, the similarity calculation formula is: wherein V 日志 is the log data vector of the log data, V1 is the original valid data vector of the original valid data, and V 日志 ·V is the dot product of the two vectors, which measures their similarity in direction. ||V 日志|| and ||V1|| are the Euclidean norms (i.e. lengths) of the log data vector and the original valid data vector, respectively. Max represents the maximum value of the similarity between the log data vector V 日志 The similarity is calculated by finding the most similar original valid data vector in the original valid data.

[0160] For example, assume we have a log data vector Vlog representing a user's voice request, we need to calculate its similarity with the original valid data vectors V1.

[0161] First, we calculate the dot product of Vlog and each V1, as well as their Euclidean norms. Then, we calculate the similarity value according to the similarity formula. Finally, we select the V1 with the highest similarity as the most similar original valid data vector. If the calculated similarity value is very low, for example, 0.2, we can consider this log data vector Vlog as abnormal or dissimilar to the training data. We can use it to train the model's rejection ability to improve the model's ability to handle abnormal inputs.

[0162] In this way, the server can evaluate the similarity between the log data and the original valid data, better understand the relationship between the data, and provide a basis for subsequent data processing and analysis.

[0163] Please refer to Figure 9 In some embodiments, step 0211 (determining a log data vector of the log data according to a preset similarity algorithm and the log data) comprises:

[0164] 02111: determining a word vector in the log data according to the log data;

[0165] 02112: determining a target word according to the log data;

[0166] 02113: obtaining a term frequency-inverse document frequency weight of the target word according to the target word and the subset of original valid data;

[0167] 02114: obtaining a log data vector of the log data according to the word vector and the term frequency-inverse document frequency weight.

[0168] In some embodiments, the processing module is configured to determine a word vector in the log data according to the log data, and determine a target word according to the log data. The processing module is configured to obtain a term frequency-inverse document frequency weight of the target word according to the target word and the subset of original valid data, and obtain a log data vector of the log data according to the word vector and the term frequency-inverse document frequency weight.

[0169] In some embodiments, the processor is further configured to determine a word vector in the log data according to the log data, and determine a target word according to the log data. The processor is further configured to obtain a term frequency-inverse document frequency weight of the target word according to the target word and the original effective data subset, and obtain a log data vector of the log data according to the word vector and the term frequency-inverse document frequency weight.

[0170] Specifically, the target word refers to a word in the log data that needs to be processed. For example, the log data is "turn on the air conditioner", and the target word can be one of "turn", "on", "air" and "conditioner".

[0171] The term frequency-inverse document frequency weight TF-IDF refers to a statistical method for evaluating the importance of a word to a document or a file in a corpus. Term frequency (TF) refers to the frequency of a word (or word) in a document, and inverse document frequency (IDF) is a measure of the rarity of a word (word) in the entire corpus. The TF-IDF weight is the product of the term frequency and the inverse document frequency, which can highlight those words that frequently appear in a specific document but are less common in the entire corpus. The term frequency-inverse document frequency weight usually includes two parts: first, calculate the term frequency (TF) calculation formula is TF=i / N, where i is the number of occurrences of the word in the text, and N is the total number of words in the text. Second, calculate the inverse document frequency, the calculation formula is IDF=log2(n / k), where n is the total number of sentences or articles in the text library, and k is the number of sentences or articles in which the word appears. Third, calculate the term frequency-inverse document frequency weight, which is the product of TF and IDF, the calculation formula is TF-IDF=TF*IDF, which can represent the importance of each word in the sentence.

[0172] First, use a pre-trained word embedding model such as BERT to convert each word in the log data into a vector. BERT (Bidirectional Encoder Representations from Transformers) is a pre-trained language representation model that can understand the contextual meaning of words in a sentence. For each word in a sentence, BERT generates a fixed-dimensional vector that contains the semantic information of the word.

[0173] Then, according to the log data obtained from the log data set, the target word is determined. For example, the obtained log data is "open the door of the car", and the target word "car" is determined. Subsequently, the original effective data subset is taken as a text library, that is, the total number of words in the original effective data subset is N, the total number of "car" in the original effective data subset is i, the total number of rain gear in the original effective data subset is n, and the number of sentences in which "car" appears in the original effective data subset is k. The term frequency-inverse document frequency weight TF-IDF of the target word "car" is calculated. Then, the term frequency-inverse document frequency weight TF-IDF of the word "open" in the target data "open the door of the car" is calculated, the term frequency-inverse document frequency weight TF-IDF of the word "door" in the target data "open the door of the car" is calculated, and the term frequency-inverse document frequency weight TF-IDF of the word "door" in the target data "open the door of the car" is calculated.

[0174] Finally, the term frequency-inverse document frequency weight TF-IDF of each word in the log data and the word vector of each word are weighted and summed to obtain a log data vector of the log data. It should be noted that the original effective data vector is obtained based on similar steps, which will not be described here.

[0175] In this way, a log data vector composed of the TF-IDF weight of the target word is obtained, which can effectively represent the characteristics and content of the log data, so as to better understand the content and theme of the data for subsequent similarity calculation.

[0176] Please refer to Figure 10 In some embodiments, step 03 (training the model according to the original training data set, the first training data set, the second training data set and the third training data set to improve the recognition ability of the model) comprises:

[0177] 031: obtaining a recognition training data set according to the second training data set and the third training data set;

[0178] 032: training the model according to the original training data set, the first training data set and the recognition training data set based on a preset proportion to improve the recognition ability of the model.

[0179] In some embodiments, the training module is configured to obtain a recognition training data set according to the second training data set and the third training data set, and train the model according to the original training data set, the first training data set and the recognition training data set based on a preset proportion to improve the recognition ability of the model.

[0180] In some embodiments, the processor is configured to obtain a rejection training data set according to the second training data set and the third training data set, and train the model according to the preset proportion, the original training data set, the first training data set and the rejection training data set, so as to improve the rejection capability of the model.

[0181] Specifically, the preset proportion refers to the proportional relationship between the original training data set, the first training data set and the rejection training data set when training the model.

[0182] Referring to Figure 11 , Figure 11 is a schematic diagram of a model training data processing flow. First, the original effective data in the original training data is modified by adding a specific rejection task prompt (i.e., a preset rejection description) to adapt to the needs of the rejection model, thereby obtaining the first training data set. Second, some task information in the original training data is updated, i.e., some information (target to be modified information) in the original effective data that can normally give a reply is modified. The updated data is further modified to better adapt to the needs of the rejection model, thereby obtaining the second training data set. Third, log data is obtained, which may include problems and challenges encountered in actual application, and the log data is filtered according to the original effective data to select sentences with low similarity to the original training data to increase the generalization capability of the model. Subsequently, the filtered data is further modified to better adapt to the needs of the rejection model, thereby obtaining the third training data set. Fourth, the original training data set, the first training data set, the second training data set and the third training data set are used together to enhance the rejection capability of the model. Figure 1 The process of training the rejection model according to the obtained training data is shown, starting from the original training data, through increasing the rejection task, updating the data, filtering and modifying, etc.

[0183] In this way, the server determines the usage amount of data in the original training data set, the first training data set and the rejection training data set based on the preset proportion to achieve good training effect, thereby improving the ability of the model to reject inappropriate or abnormal input in actual application.

[0184] Referring to Figure 12 , the embodiment of the present application provides a voice interaction method based on the model optimized by the above-mentioned optimization method, the method comprising:

[0185] 04: receiving a voice request forwarded by a vehicle;

[0186] 05: rejecting the voice request according to the model to complete voice interaction.

[0187] The embodiment of the application further provides a server comprising a memory and a processor. The voice interaction method of the embodiment of the application can be implemented by the server of the embodiment of the application. Specifically, the memory stores a computer program, and the processor is configured to receive a voice request forwarded by a vehicle, and perform recognition processing on the voice request according to a model to complete voice interaction.

[0188] The embodiment of the application further provides a voice interaction device. The voice interaction method of the embodiment of the application can be implemented by the voice interaction device of the embodiment of the application. Specifically, the voice interaction device comprises a receiving module and a recognition module. The receiving module is configured to receive a voice request forwarded by a vehicle, and the recognition module is configured to perform recognition processing on the voice request according to a model to complete voice interaction.

[0189] Specifically, the server receives a voice request forwarded by a vehicle. Then, the server analyzes the received voice request according to the obtained model recognition model to complete voice interaction. The task of the recognition model is to determine whether the voice request is suitable or can be processed by the system. If the request is not related to the function of the system or cannot be understood, the model will identify and reject this voice request.

[0190] For example, suppose a user says in the vehicle: "Please help me find the nearest gas station." The voice request is forwarded by the vehicle to the server. The server analyzes this request using the recognition model. The model identifies that this is a valid request because it is related to the navigation function of the vehicle system. The server then processes this request and returns the navigation information of the nearest gas station to the vehicle.

[0191] If the user says: "Please tell me a joke.", the recognition model may identify that this is a request that is not suitable for the vehicle system to process, because it is not related to the functions of the vehicle, such as navigation, music playing, etc. Therefore, the model will reject this request and may prompt the user to say: "I'm sorry, I cannot execute this request."

[0192] In summary, in the voice interaction method and the server provided by the embodiment of the application, the server receives a voice request forwarded by a vehicle. Then, the server performs recognition processing on the voice request according to a model to complete voice interaction. In this way, based on the model obtained by training the above-mentioned optimization method, the server can ensure that the voice interaction between the vehicle and the user is both safe and reliable.

[0193] The application further provides a computer readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the optimization method or the voice interaction method described above are implemented.

[0194] It can be understood that the computer program includes computer program code. The computer program code can be in the form of source code, object code, executable files, or some intermediate forms, etc. The computer readable storage medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), and software distribution medium, etc.

[0195] In the description of the present specification, the description referring to the terms "specifically", "further", "particularly", "it can be understood that", etc. means that the specific features, structures, materials or characteristics described in connection with the embodiments or examples are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not intend to refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the present specification and the features of the different embodiments or examples without contradiction.

[0196] Any process or method descriptions in flow charts or described elsewhere herein can be understood as representing one or more modules, segments, or portions of code that include one or more executable instructions for implementing specific logic functions or steps in the process. The scope of preferred embodiments of the present application includes alternative implementations including processes that are performed out of the order described in the figures or discussed herein, including functions that are performed substantially concurrently or in reverse order, as will be understood by those skilled in the art.

[0197] Although the embodiments of the present application have been shown and described above, it can be understood that the above-described embodiments are exemplary and cannot be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above-described embodiments within the scope of the present application.

Claims

1. A method for optimizing a model for vehicle voice interaction, the method comprising: The method comprises: The original training data set of the model is subjected to a recognition modification process to obtain a first training data set and a second training data set, wherein the first training data set is obtained by subjecting the original training data set to a recognition modification process according to a preset recognition description, and the preset recognition description is a rule defined in advance for guiding the model to perform a recognition process; and the second training data set is obtained by directly modifying the original training data set; According to the original training data set, an original log data set is subjected to a screening process to obtain a third training data set; The model is trained according to the original training data set, the first training data set, the second training data set and the third training data set to improve the recognition ability of the model.

2. The optimization method of claim 1, wherein, The original training data set of the model is subjected to a recognition modification process to obtain a first training data set and a second training data set, comprising: According to the original recognition data subset and the original effective data subset in the original training data set, the first training data set is obtained based on a preset test method and a preset recognition description.

3. The optimization method of claim 2, wherein, The original effective data subset comprises a plurality of original effective data, and the first training data set is obtained based on a preset test method and a preset recognition description. According to the preset granularity, each original effective data in the original effective data subset is subjected to a segmentation process to obtain a plurality of target recognition data, wherein each target recognition data corresponds to one original effective data. According to the preset recognition description and the plurality of target recognition data, a plurality of temporary recognition data sets are obtained. According to the original recognition data subset, the temporary recognition data set is tested to confirm that at least one of the temporary recognition data set is added to the first training data set.

4. The optimization method of claim 3, wherein, The target recognition data comprises at least one sentence, and the plurality of temporary recognition data sets are obtained according to the preset recognition description and the target recognition data, comprising: The preset recognition description is placed at a segmentation position in each target recognition data to obtain a plurality of temporary recognition data sets, and the segmentation position is a connection position of each sentence in the target recognition data.

5. The optimization method of claim 2, wherein, The original training data set of the model is subjected to a recognition modification process to obtain a first training data set and a second training data set, comprising: The original effective data subset is subjected to a modification process to obtain the second training data set.

6. The optimization method of claim 5, wherein, The original effective data subset is subjected to a modification process to obtain the second training data set, comprising: The original effective data in the original effective data subset is subjected to an analysis process to obtain target information to be modified; The target information to be modified is subjected to a modification process to obtain target error information; The second training data set is obtained according to the target error information.

7. The optimization method of claim 1, wherein, The original log data set is subjected to a screening process to obtain a third training data set according to the original training data set, comprising: The similarity of each log data is obtained based on a preset similarity algorithm and the log data in the original log data set and the original effective data subset in the original training data set. The third training data set is determined according to a preset similarity threshold and the similarity.

8. The optimization method of claim 7, wherein, The original effective data subset includes a plurality of original effective data, and the similarity of each log data is obtained based on a preset similarity algorithm and the log data in the original log data set and the original effective data subset in the original training data set, including: A log data vector of the log data is determined according to the preset similarity algorithm and the log data. An original effective data vector of the original effective data is determined according to the preset similarity algorithm and the original effective data. The similarity is obtained according to the log data vector and the original effective data vector.

9. The optimization method of claim 8, wherein, The log data vector of the log data is determined according to the preset similarity algorithm and the log data, including: A word vector in the log data is determined according to the log data. A target word is determined according to the log data. A term frequency-inverse document frequency weight of the target word is obtained according to the target word and the original effective data subset. The log data vector of the log data is obtained according to the word vector and the term frequency-inverse document frequency weight.

10. The optimization method of claim 1, wherein, The model is trained according to the original training data set, the first training data set, the second training data set and the third training data set to improve the recognition ability of the model, including: A rejection training data set is obtained according to the second training data set and the third training data set. The model is trained according to the original training data set, the first training data set and the rejection training data set based on a preset proportion to improve the recognition ability of the model.

11. A voice interaction method, characterized by, The voice interaction method is based on the model optimized by the optimization method of any one of claims 1-10, and the method includes: Receiving a voice request forwarded by a vehicle; Performing rejection processing on the voice request according to the model to complete the voice interaction.

12. A server, characterized by The server includes a processor and a memory, and the memory stores a computer program, which, when executed by the processor, implements the optimization method of any one of claims 1-11.

13. A computer readable storage medium having stored thereon a computer program, characterized in that The program is executed by the processor to implement the steps of the method of any one of claims 1-11. The program is executed by the processor to implement the steps of the method of any one of claims 1-11.

Images