A method and apparatus for target intent recognition based on hyperparameter search using a data optimizer.

Abstract

This invention discloses a method and apparatus for target intent recognition based on hyperparameter search using a data optimizer. The method includes: training and optimizing corresponding intent recognition models using out-of-domain data combinations and hyperparameters from multiple different training parameter combinations; forming an intent recognition system with each intent recognition model and intent recognition rules, and obtaining a score for each intent recognition system; inputting each score into a data optimizer to select a target training parameter combination; and forming a target intent recognition system with the intent recognition model corresponding to the target training parameter combination and intent recognition rules to identify the target intent. This invention embeds the intent recognition model into a complex online intent recognition system to obtain a score, and optimizes the performance of intent recognition models with different training parameter combinations in actual online production using a data optimizer. This improves the model training effect while avoiding uncertainties caused by interaction with intent recognition rules, thereby improving the accuracy of intent recognition in actual production.

Description

Technical Field

[0001] This invention relates to the field of natural speech recognition technology, and more specifically, to a method and apparatus for target intent recognition based on hyperparameter search using a data optimizer. Background Technology

[0002] Intent recognition is an important branch of natural language understanding and plays a crucial role in robotics. Generally, an intent recognition system consists of intent recognition rules and an intent recognition model. The intent recognition model is typically a neural network model trained on data.

[0003] like Figure 1 Existing intent recognition models primarily utilize training data and fixed hyperparameters during training. A predetermined accuracy metric is used as the objective function, and the model's parameters are updated via gradient backpropagation to obtain a trained intent recognition model. The inventors, through... Figure 1 After analyzing the model training method, it was found that it has at least the following technical problems:

[0004] 1. Figure 1 Although the model parameters were updated in the training method, hyperparameters, data selection, and other important factors affecting the model training effect were not optimized.

[0005] 2. Figure 1 The training method only trains the intent recognition model separately, without considering the complex interactions between the intent recognition model and intent recognition rules in actual production. In reality, a high-precision intent recognition model trained independently often fails to achieve satisfactory intent recognition results after interacting with intent recognition rules. Summary of the Invention

[0006] In view of this, the main objective of the present invention is to provide a method, apparatus, electronic device and computer-readable medium for target intent recognition based on hyperparameter search using a data optimizer, in order to at least partially solve at least one of the above-mentioned technical problems.

[0007] To address the aforementioned technical problems, the first aspect of this invention proposes a target intent recognition method based on hyperparameter search using a data optimizer, the method comprising:

[0008] The corresponding intent recognition model is trained and optimized using out-of-domain data combinations and hyperparameters from multiple different training parameter combinations.

[0009] Each intent recognition model and intent recognition rule together constitute an intent recognition system, and a score is obtained for each intent recognition system.

[0010] Select the target training parameter combination from the input data optimizer for each test score;

[0011] The target training parameters are combined to form an intent recognition model and intent recognition rules to create a target intent recognition system that identifies the target intent.

[0012] According to a preferred embodiment of the present invention, training and optimizing the corresponding intent recognition model using a combination of out-of-domain data and hyperparameters in a single training parameter combination includes:

[0013] The out-of-domain data combinations in the training parameter combinations are merged with the training data to form a training set;

[0014] The intent recognition model is trained using the training set and the hyperparameters in the training parameter combination;

[0015] The hyperparameters are updated using the gradient inversion algorithm to optimize the objective function.

[0016] According to a preferred embodiment of the present invention, the hyperparameters in the training parameter combination are selected from a range of selectable hyperparameters, and the out-of-domain data combination is composed of at least two out-of-domain data combinations selected from a set of selectable out-of-domain data.

[0017] According to a preferred embodiment of the present invention, the intent recognition model and intent recognition rules form an intent recognition system through a preset interaction method; the preset interaction method includes: preferentially using intent recognition rules to recognize intent, or simultaneously using intent recognition rules and intent recognition model to recognize intent.

[0018] According to a preferred embodiment of the present invention, each intent category is divided into multiple intent subcategories, and the simultaneous use of intent recognition rules and intent recognition models to identify intent includes:

[0019] Determine whether the recognition result of the intent recognition model is the same as the recognition result of the intent recognition rule. If they are the same, output the same recognition result as the intent recognition result.

[0020] If they are different, determine whether the intent sub-category corresponding to the recognition result of the intent recognition model and the intent sub-category corresponding to the recognition result of the intent recognition rule belong to the same intent category. If they belong, output the intent category to which it belongs.

[0021] If it does not belong to the category, the intent is re-identified using both the intent recognition rules and the intent recognition model until the intent category is output.

[0022] According to a preferred embodiment of the present invention, obtaining the scores of each intent recognition system includes:

[0023] The sample data is input into each intent recognition system to obtain intent recognition results;

[0024] The accuracy rate of the intent recognition results is statistically analyzed to obtain a score.

[0025] To address the aforementioned technical problems, a second aspect of the present invention provides a target intent recognition device based on hyperparameter search using a data optimizer, the device comprising:

[0026] The training module is used to train and optimize the corresponding intent recognition model using out-of-domain data combinations and hyperparameters from multiple different training parameter combinations.

[0027] An evaluation module has been launched to enable the various intent recognition models and intent recognition rules to work together to form an intent recognition system and to obtain evaluation scores for each intent recognition system.

[0028] The selection module is used to combine the target training parameters selected from the input data optimizer for each test score.

[0029] The recognition module is used to combine the target training parameters with the corresponding intent recognition model and intent recognition rules to form a target intent recognition system to recognize the target intent.

[0030] According to a preferred embodiment of the present invention, the training module includes:

[0031] The merging module is used to merge the out-of-domain data combinations in the training parameter combinations with the training data to form a training set;

[0032] The sub-training module is used to train the intent recognition model using hyperparameters from the training set and the training parameter combination;

[0033] The update module is used to update the hyperparameters using the gradient inversion algorithm to optimize the objective function.

[0034] According to a preferred embodiment of the present invention, the hyperparameters in the training parameter combination are selected from a range of selectable hyperparameters, and the out-of-domain data combination is composed of at least two out-of-domain data combinations selected from a set of selectable out-of-domain data.

[0035] According to a preferred embodiment of the present invention, the online evaluation module combines the intent recognition model and intent recognition rules into an intent recognition system through a preset interaction method; the preset interaction method includes: preferentially using intent recognition rules to identify intent, or simultaneously using intent recognition rules and intent recognition model to identify intent.

[0036] According to a preferred embodiment of the present invention, each intent category is divided into multiple intent subcategories, and the online evaluation module includes:

[0037] The first judgment module is used to determine whether the recognition result of the intent recognition model is the same as the recognition result of the intent recognition rule. If they are the same, the same recognition result is output as the intent recognition result.

[0038] The second judgment module is used to determine whether the intention sub-category corresponding to the recognition result of the intention recognition model and the intention sub-category corresponding to the recognition result of the intention recognition rule belong to the same intention category if they are different. If they do, the intention category to which they belong is output.

[0039] The loop module is used to re-identify the intent using both the intent recognition rules and the intent recognition model if the intent is not found, until the intent category is output.

[0040] According to a preferred embodiment of the present invention, the online evaluation module further includes:

[0041] The input module is used to input sample data into various intent recognition systems to obtain intent recognition results.

[0042] The statistics module is used to calculate the accuracy of the intent recognition results and obtain a score.

[0043] This invention employs multiple different combinations of training parameters, including out-of-domain data combinations and hyperparameters, to train and optimize corresponding intent recognition models. By training and optimizing models based on different combinations of out-of-domain data and hyperparameters, various intent recognition models are obtained. Each intent recognition model, together with intent recognition rules, constitutes an intent recognition system, and a score is obtained for each system. A mapping between training parameter combinations and online score is derived, and this score is used as an indicator for a data optimizer to obtain the optimal training parameter combination. This embeds the intent recognition model into a complex online intent recognition system to obtain a score. The data optimizer optimizes the performance of intent recognition models with different training parameter combinations in actual online production, improving model training effectiveness while avoiding uncertainties caused by interaction with intent recognition rules in the online intent recognition system, thereby improving the accuracy of intent recognition in actual production. Compared with existing technologies, this invention has at least the following advantages:

[0044] Beneficial effects:

[0045] 1. By using a data optimizer, the evaluation score of the intent recognition system is used as an indicator to guide the selection and updating of training parameter combinations, thereby increasing the model optimization space and improving model performance.

[0046] 2. Considering the interaction between the actual online intent recognition model and intent recognition rules, the intent recognition model is trained on real intent recognition rules to avoid the uncertainty of model performance caused by training the intent recognition model separately and interacting with intent recognition rules, thereby improving the accuracy of intent recognition in actual production. Attached Figure Description

[0047] To make the technical problems solved by this invention, the technical means adopted, and the technical effects achieved clearer, specific embodiments of this invention will be described in detail below with reference to the accompanying drawings. However, it should be noted that the drawings described below are merely drawings of exemplary embodiments of this invention. Those skilled in the art can obtain drawings of other embodiments based on these drawings without any creative effort.

[0048] Figure 1 This is a schematic diagram of the training process of an existing intent recognition model;

[0049] Figure 2 This is a flowchart illustrating a target intent recognition method based on hyperparameter search using a data optimizer, according to an embodiment of the present invention.

[0050] Figure 3 This is a flowchart illustrating another target intent recognition method based on hyperparameter search using a data optimizer according to an embodiment of the present invention.

[0051] Figure 4 This is a schematic diagram of the structural framework of a target intent recognition device based on hyperparameter search using a data optimizer, according to an embodiment of the present invention. Detailed Implementation

[0052] Subject to the inventive concept, the structures, performance, effects or other features described in a particular embodiment may be combined in any suitable manner with one or more other embodiments.

[0053] In the description of specific embodiments, detailed descriptions of structures, performance, effects, or other features are provided to enable those skilled in the art to fully understand the embodiments. However, it is not excluded that those skilled in the art can implement the present invention with technical solutions that do not contain the above-described structures, performance, effects, or other features under specific circumstances. The flowcharts in the accompanying drawings are merely exemplary flowchart demonstrations and do not imply that the solution of the present invention must include all the contents, operations, and steps in the flowcharts, nor do they imply that the execution must be carried out in the order shown in the figures.

[0054] Please see Figure 2 , Figure 2 This invention provides a target intent recognition method based on hyperparameter search using a data optimizer, such as... Figure 2 As shown, the method includes:

[0055] S1. Train and optimize the corresponding intent recognition model using out-of-domain data combinations and hyperparameters from multiple different training parameter combinations;

[0056] In this embodiment, the training parameter combination includes at least one factor that affects the model training effect. For example, factors affecting the model training effect may include hyperparameters, data selection, etc. Therefore, the training parameter combination may include hyperparameters, data combinations, or hyperparameters and data combinations. Hyperparameters correspond to adjustable parameters in a model training process, such as the model training learning rate, model structure, etc. Data combinations refer to data composed of important variables obtained through data selection.

[0057] Typically, intent recognition models trained on data lack the ability to identify data outside the training domain. When encountering such data, they may misclassify it as belonging to a specific category (even though the data doesn't actually belong to any category). "Outside the training domain" refers to data that doesn't belong to the same broad category as the training data. For example, if the training data is related to "finance," then data related to "news" would be considered outside the training domain. Considering the diversity of data categories in actual production, this embodiment introduces outside the training domain into the model, enabling the model to identify it. However, there are many categories of outside the training domain, and some may have the opposite effect. For instance, data categorized as "financial news" may have some relevance to "finance," but it doesn't belong to the "finance" category, making it unsuitable as outside the training domain. Therefore, this embodiment filters both the training data and outside the training domain. The training parameter combinations can further include: combinations of outside the training domain data, combinations of hyperparameters and outside the training domain data, combinations of data and outside the training domain data, or combinations of hyperparameters, data, and outside the training domain data. The hyperparameters can be selected from the optional hyperparameter range, for example, the optional range of hyperparameter A is [1, 3]. The out-of-domain data combination is formed by selecting at least two out-of-domain data combinations from the optional out-of-domain data set. For example, if the out-of-domain data set contains out-of-domain data a, out-of-domain data b, out-of-domain data c, and out-of-domain data d, then at least two out-of-domain data combinations can be selected from it to obtain the out-of-domain data combination.

[0058] In a preferred example, the training parameter combination includes a hyperparameter and out-of-domain data combination. Each training parameter combination includes a set of hyperparameters and a set of out-of-domain data. The hyperparameters or out-of-domain data combinations differ in each training parameter combination. For example: in the first training parameter combination, hyperparameter A is 0.13, and the out-of-domain data combination includes out-of-domain data a and out-of-domain data b; in the second training parameter combination, hyperparameter A is 0.13, and the out-of-domain data combination includes out-of-domain data a; in the third training parameter combination, hyperparameter A is 0.15, and the out-of-domain data combination includes out-of-domain data a. Figure 3 For each combination of training parameters, training and optimizing the corresponding intent recognition model using the out-of-domain data combination and hyperparameters from a single training parameter combination includes:

[0059] S11. Merge the out-of-domain data combination in the training parameter combination with the training data to form a training set;

[0060] S12. Train the intent recognition model using the hyperparameters in the training set and training parameter combination;

[0061] For example, the hyperparameters in the training parameter combination can be used as hyperparameters of the intent recognition model, and the training set can be input into the intent recognition model to obtain the intent prediction result.

[0062] S13. Update the hyperparameters using the gradient inversion algorithm to optimize the objective function.

[0063] In this embodiment, the accuracy index is used as the objective function. For example, if the accuracy of the intent recognition result is greater than 95%, the accuracy of the intent recognition result is calculated based on the intent prediction result and the actual intent. When the accuracy of the intent recognition result is greater than 95%, the training ends and the trained intent recognition model is obtained.

[0064] S2. Each intent recognition model and intent recognition rule constitutes an intent recognition system, and the evaluation score of each intent recognition system is obtained.

[0065] In actual production processes, online intent recognition systems predict intent through the interaction of intent recognition models and intent recognition rules. Intent recognition rules are predefined intent recognition methods without trainable parameters, requiring no extensive data training. Each rule may include pre-defined keywords and corresponding intents; if text or audio matches the keyword, the corresponding intent is output. For example, in one instance of an intent rule, if the text contains the word "hello," it is recognized as a "greeting" intent. The intent recognition model is a neural network model trained on data. It can predict intent based on text, audio, or other information. Generally, training the intent recognition model and defining intent recognition rules are done independently. However, because intent recognition rules and models combine to identify intent, the rules actually affect the model's performance. Therefore, this invention embeds various intent recognition models into a complex online intent recognition system, interacting with intent recognition rules to obtain evaluation scores, guiding the model's training process. This directly optimizes the model's performance in actual online production, avoiding the uncertainty caused by interaction with intent recognition rules.

[0066] For example, the intent recognition model and intent recognition rules form an intent recognition system through a preset interaction method; the preset interaction method includes: prioritizing the use of intent recognition rules to identify intent, or simultaneously using intent recognition rules and the intent recognition model to identify intent. Specifically: prioritizing the use of intent recognition rules means first identifying intent through intent recognition rules; if the test text or audio matches a certain intent recognition rule, the intent corresponding to that rule is output; if not all intent recognition rules are matched, then the intent recognition model is used for identification. Simultaneously using intent recognition rules and the intent recognition model to identify intent means simultaneously using the intent recognition model and intent recognition rules to identify intent, and providing a comprehensive intent recognition result based on the recognition results of the intent recognition model and the matching results of the intent recognition rules.

[0067] For example, when using both intent recognition rules and intent recognition models to identify intent, the weights of the intent recognition rules and models can be pre-set, and the recognition result with the larger weight can be used as the intent recognition result. For instance, if the weight of the intent recognition model is 0.7 and the weight of the intent recognition rule is 0.3, then the recognition result of the intent recognition model can be used as the intent recognition result. Furthermore, the intent category can be further divided into multiple intent subcategories. First, it is determined whether the recognition result of the intent recognition model is the same as the recognition result of the intent recognition rule. If they are the same, the same recognition result is output as the intent recognition result. If they are different, it is determined whether the intent subcategory corresponding to the recognition result of the intent recognition model and the intent subcategory corresponding to the recognition result of the intent recognition rule belong to the same intent category. If they belong, the corresponding intent category is output. If they do not belong, both the intent recognition rules and the intent recognition model are used to re-identify the intent until the intent category is output.

[0068] It should be noted that the above is merely an example of the preset interaction method between the intent recognition model and the intent recognition rules. In actual production, the interaction method between the intent recognition rules and the intent recognition model may be very complex, and the interaction method may also be adjusted and changed according to the accuracy of the recognition results.

[0069] In this embodiment, obtaining the scores of each intent recognition system may include:

[0070] S21. Input the sample data into each intent recognition system to obtain the intent recognition results;

[0071] Wherein: the sample data includes: input data and the true intent corresponding to the input data; the input data may be: text data and / or audio data. For example, the sample data may be text data and / or audio data recognized within a predetermined time period (e.g., N days) after various intent recognition models and intent recognition rules are combined to form a corresponding intent recognition system.

[0072] S22. Calculate the accuracy of the intent recognition results to obtain the score.

[0073] The score can be the accuracy of the intent recognition result. For example, the recognition results of each intent recognition system can be compared with their corresponding true intents to calculate the accuracy of the intent recognition result and obtain the score.

[0074] Through the above steps S1 and S2, a mapping between different training parameter combinations and online test scores is obtained. It should be noted that: in steps S1 and S2, multiple different training parameter combinations can be pre-configured, and steps S1 and S2 can be executed separately for each different training parameter combination to obtain the corresponding test score; alternatively, an initial training parameter combination can be configured first, steps S1 and S2 can be executed to obtain the test score corresponding to the initial training parameter combination, and then the initial training parameter combination can be updated according to a preset rule before executing steps S1 and S2 again to obtain the test score corresponding to each training parameter combination.

[0075] S3. Input each test score into the data optimizer and select the target training parameter combination;

[0076] The data optimizer is an optimization method for hyperparameter search. Its core idea is that each experiment corresponds to a scalar metric, and this scalar metric is treated as a Gaussian process. The properties of the Gaussian process are used to predict and optimize the experimental parameters. Existing data optimizers only select and optimize hyperparameters. This invention modifies conventional data optimizers by using the performance score of the intent recognition model trained with different parameter sets, interacting with intent recognition rules to form an online intent recognition system, as an indicator to guide the selection and updating of training parameter sets to achieve optimal online performance. Preferably, the data optimizer can be a Bayesian optimizer; the data optimization method can be a Bayesian optimization method.

[0077] In a preferred example, such as Figure 3 The training parameter set includes a set of hyperparameters and a set of out-of-domain data combinations. The Bayesian optimizer uses the evaluation scores of the intent recognition models trained with different parameter sets on online intent recognition systems as indicators to guide the selection and updating of hyperparameters and the selection of out-of-domain data. While general Bayesian optimizers select hyperparameters, this paper uses Bayesian optimization methods to select out-of-domain data. This avoids the situation where the model accuracy is negatively affected by the similarity between out-of-domain data and data in the training set, thereby further improving the ability of the intent recognition model to recognize out-of-domain data and preventing unsuitable out-of-domain data from participating in model training.

[0078] S4. Combine the target training parameters with the corresponding intent recognition model and intent recognition rules to form a target intent recognition system to recognize the target intent.

[0079] Specifically, the intent recognition model corresponding to the combination of target training parameters is deployed online, and together with intent recognition rules, they form a target intent recognition system to identify the target intent. Preferably, the target intent can be the user intent.

[0080] Figure 4 This invention relates to a target intent recognition device based on hyperparameter search using a data optimizer, such as... Figure 4 As shown, the device includes:

[0081] Training module 41 is used to train and optimize the corresponding intent recognition model using out-of-domain data combinations and hyperparameters from multiple different training parameter combinations;

[0082] The online evaluation module 42 is used to combine various intent recognition models and intent recognition rules to form an intent recognition system and obtain the evaluation scores of each intent recognition system.

[0083] Module 43 is selected to combine the target training parameters selected from the input data optimizer for each test score.

[0084] The recognition module 44 is used to combine the target training parameters with the corresponding intent recognition model and intent recognition rules to form a target intent recognition system to recognize the target intent.

[0085] In one embodiment, the training module 41 includes:

[0086] The merging module is used to merge the out-of-domain data combinations in the training parameter combinations with the training data to form a training set;

[0087] The sub-training module is used to train the intent recognition model using hyperparameters from the training set and the training parameter combination;

[0088] The update module is used to update the hyperparameters using the gradient inversion algorithm to optimize the objective function.

[0089] Wherein: the hyperparameters in the training parameter combination are selected from the range of optional hyperparameters, and the out-of-domain data combination is composed of at least two out-of-domain data combinations selected from the optional out-of-domain data set.

[0090] The online evaluation module 42 combines the intent recognition model and intent recognition rules into an intent recognition system through a preset interaction method; the preset interaction method includes: prioritizing the use of intent recognition rules to recognize intent, or simultaneously using intent recognition rules and intent recognition model to recognize intent.

[0091] Optionally, each intent category is divided into multiple intent subcategories, and the online evaluation module 42 includes:

[0092] The first judgment module is used to determine whether the recognition result of the intent recognition model is the same as the recognition result of the intent recognition rule. If they are the same, the same recognition result is output as the intent recognition result.

[0093] The second judgment module is used to determine whether the intention sub-category corresponding to the recognition result of the intention recognition model and the intention sub-category corresponding to the recognition result of the intention recognition rule belong to the same intention category if they are different. If they do, the intention category to which they belong is output.

[0094] The loop module is used to re-identify the intent using both the intent recognition rules and the intent recognition model if the intent is not found, until the intent category is output.

[0095] Furthermore, the online evaluation module 42 also includes:

[0096] The input module is used to input sample data into various intent recognition systems to obtain intent recognition results.

[0097] The statistics module is used to calculate the accuracy of the intent recognition results and obtain a score.

[0098] Those skilled in the art will understand that the modules in the above-described device embodiments can be distributed throughout the device as described, or they can be modified accordingly and distributed in one or more devices different from the above embodiments. The modules in the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules.

[0099] In summary, the present invention can be implemented by methods, apparatus, electronic devices, or computer-readable media that execute computer programs. In practice, some or all of the functions of the present invention can be implemented using general-purpose data processing devices such as microprocessors or digital signal processors (DSPs).

[0100] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the present invention is not inherently related to any specific computer, virtual device, or electronic device, and various general-purpose devices can also implement the present invention. The above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A user intent recognition method based on hyperparameter search using a data optimizer, characterized in that, For natural language intent recognition, the method includes: The intention recognition model for a corresponding neural network type is trained and optimized using out-of-domain data combinations and hyperparameters from multiple different training parameter combinations. Specifically, training and optimizing the intention recognition model for a corresponding neural network type using out-of-domain data combinations and hyperparameters from a single training parameter combination includes: merging the out-of-domain data combinations from the training parameter combination with the training data to form a training set; training the intention recognition model using the training set and the hyperparameters from the training parameter combination; and updating the hyperparameters using a gradient inversion algorithm to optimize the objective function. Each intent recognition model and a predefined intent recognition rule without training parameters are combined through a preset interaction method to form an online intent recognition system. Input data and the corresponding real user intent are input into each online intent recognition system to obtain user intent recognition results. The accuracy of the user intent recognition results is statistically analyzed to obtain a score for each online intent recognition system. The input data is text data and / or audio data. The preset interaction method includes: prioritizing the use of intent recognition rules to identify user intent; or, simultaneously using intent recognition rules and intent recognition models to identify user intent. Specifically, each intent category is divided into multiple intent subcategories. The system determines whether the recognition result of the intent recognition model is the same as the recognition result of the intent recognition rule. If they are the same, the same recognition result is output as the user intent recognition result. If they are different, the system determines whether the intent subcategory corresponding to the recognition result of the intent recognition model and the intent subcategory corresponding to the recognition result of the intent recognition rule belong to the same intent category. If they belong, the corresponding intent category is output. If they do not belong, the system re-identifies the intent using both intent recognition rules and intent recognition models until the intent category is output. Each test score is input into a data optimizer of the Bayesian optimizer type to select the target training parameter combination; The target training parameters are combined to form the corresponding intent recognition model and intent recognition rules to form a target intent recognition system, in order to recognize user intent.

2. The method according to claim 1, characterized in that, The hyperparameters in the training parameter combination are selected from the range of optional hyperparameters, and the out-of-domain data combination is composed of at least two out-of-domain data combinations selected from the optional out-of-domain data set.

3. A user intent recognition device based on hyperparameter search using a data optimizer, characterized in that, For natural language intent recognition, the device includes: The training module is used to train and optimize an intent recognition model of a corresponding neural network type using out-of-domain data combinations and hyperparameters from multiple different training parameter combinations; wherein: training and optimizing an intent recognition model of a corresponding neural network type using out-of-domain data combinations and hyperparameters from a single training parameter combination includes: merging the out-of-domain data combinations from the training parameter combination with the training data to form a training set; training the intent recognition model using the training set and the hyperparameters from the training parameter combination; updating the hyperparameters using a gradient inversion algorithm to optimize the objective function; An online evaluation module is used to jointly form an online intent recognition system with various intent recognition models and predefined intent recognition rules without training parameters through preset interaction methods. Input data and the corresponding real user intents are input into each online intent recognition system to obtain user intent recognition results. The accuracy of the user intent recognition results is calculated to obtain a score for each online intent recognition system. The input data is text data and / or audio data. The preset interaction methods include: prioritizing the use of intent recognition rules to identify user intents; or simultaneously using intent recognition rules and intent recognition models to identify user intents. Specifically, each intent category is divided into multiple intent subcategories. The module determines whether the recognition results of the intent recognition model and the intent recognition rules are the same. If they are the same, the same recognition result is output as the user intent recognition result. If they are different, the module determines whether the intent subcategory corresponding to the recognition result of the intent recognition model and the intent subcategory corresponding to the recognition result of the intent recognition rules belong to the same intent category. If they belong, the corresponding intent category is output. If they do not belong, the intent is re-identified using both intent recognition rules and intent recognition models until the intent category is output. The selection module is used to input various test scores into a data optimizer of the Bayesian optimizer type to select the target training parameter combination; The recognition module is used to combine the target training parameters with the corresponding intent recognition model and intent recognition rules to form a target intent recognition system in order to recognize user intent.

4. The apparatus according to claim 3, characterized in that, The hyperparameters in the training parameter combination are selected from the range of optional hyperparameters, and the out-of-domain data combination is composed of at least two out-of-domain data combinations selected from the optional out-of-domain data set.

Images