Intention recognition method and device, computer readable storage medium and electronic device

Abstract

The invention relates to the technical field of artificial intelligence, and particularly discloses an intention recognition method and device, a computer readable storage medium and an electronic device.The intention recognition method comprises the steps that a rejection training set and an intention recognition training set are obtained; training a preset model according to the rejection training set and the intention recognition training set, so that the preset model has rejection processing capability and intention recognition processing capability at the same time; when the service is called, synchronously performing rejection and intention recognition processing on the input content through a preset model; and if the result of the rejection and intention recognition processing is rejection, returning to the preset underwriting copywriting and terminating the process, and if the result of the rejection and intention recognition processing is non-rejection, falling the intention recognition result to the corresponding agent for processing. Therefore, rejection and intention recognition tasks can be synchronously processed, so that the response speed and the resource utilization rate are improved, the accuracy and robustness are enhanced, the system architecture is simplified, and the maintenance cost and complexity are reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to an intent recognition method and device, a computer readable storage medium, and an electronic device. BACKGROUND

[0002] In recent years, with the rapid development of large model technology and its wide application in various industries, the output quality of large model products has received unprecedented attention. Output quality not only involves the accuracy of content, but also includes response performance and resource cost. Traditional large model design usually adopts a serial flow processing procedure, that is, "input-rejection-intent recognition-domain-single component intelligent agent output-arbitration-output".

[0003] However, the related technical solutions have some significant shortcomings and deficiencies: the traditional serial flow processing method has a long processing procedure due to the dependence between the front and back, resulting in poor response performance. The serial execution of rejection service and intent recognition service significantly increases the total processing time, affecting user experience. In the special scenario of large model deployment, expensive computing resources are not fully utilized, resource utilization is low, and cost is wasted seriously. SUMMARY

[0004] The present application aims to at least solve one of the technical problems in the related art. To this end, one object of the present application is to propose an intent recognition method that can simultaneously process rejection and intent recognition tasks, thereby improving response speed and resource utilization, enhancing accuracy and robustness, simplifying system architecture, and reducing maintenance cost and complexity.

[0005] A second object of the present application is to propose a computer readable storage medium.

[0006] A third object of the present application is to propose an intent recognition device.

[0007] A fourth object of the present application is to propose an electronic device.

[0008] To achieve the above objects, the first aspect of the present application proposes an intent recognition method, wherein the method comprises: obtaining a rejection training set and an intent recognition training set; training a preset model according to the rejection training set and the intent recognition training set, so that the preset model simultaneously has rejection processing capability and intent recognition processing capability; when a service is called, simultaneously performing rejection and intent recognition processing on input content through the preset model; if the results of the rejection and intent recognition processing are rejection, returning a preset bottom-up script and terminating the procedure, and if the results of the rejection and intent recognition processing are not rejection, domain-mapping the intent recognition result to a corresponding intelligent agent processing.

[0009] According to the intent recognition method provided in the embodiments of the present application, the preset model is trained according to the rejection training set and the intent recognition training set, so that the preset model has both rejection processing capability and intent recognition processing capability. When the service is called, the input content is synchronously subjected to rejection and intent recognition processing by the preset model. If the result of the rejection and intent recognition processing is rejection, the preset bottom text is returned and the process is terminated. If the result of the rejection and intent recognition processing is not rejection, the intent recognition result is domained to the corresponding intelligent agent for processing. Thus, the rejection and intent recognition tasks can be synchronously processed, so that the response speed and resource utilization are improved, the accuracy and robustness are enhanced, the system architecture is simplified, and the maintenance cost and complexity are reduced.

[0010] In addition, the intent recognition method provided in the embodiments of the present application can further include the following additional technical features:

[0011] According to an embodiment of the present application, the rejection training set includes multiple rejection tasks, and the intent recognition training set includes multiple intent recognition tasks. The rejection tasks are defined as special categories of the intent recognition tasks, so that the rejection tasks and the intent recognition tasks are completed in a unified classification framework, and the repeated features of the rejection tasks and the intent recognition tasks are eliminated by a shared feature extraction module.

[0012] According to an embodiment of the present application, the preset model includes a rejection sub-model and an intent recognition sub-model. The rejection sub-model includes multiple first model parameters, and the intent recognition sub-model includes multiple second model parameters. The multiple first model parameters and the multiple second model parameters have partially shared model parameters.

[0013] According to an embodiment of the present application, the preset model is trained according to the rejection training set and the intent recognition training set, including: the boundary of the intent recognition sub-model is enhanced trained according to the rejection tasks, and the semantic information of the rejection sub-model is enhanced trained according to the intent recognition tasks.

[0014] According to an embodiment of the present application, the input content is synchronously subjected to rejection and intent recognition processing by the preset model, including: the intent recognition confidence corresponding to the input content is obtained by the preset model; when the intent recognition confidence is lower than a confidence threshold, it is determined that the result of the rejection and intent recognition processing is rejection; when the intent recognition confidence is higher than or equal to the confidence threshold, it is determined that the result of the rejection and intent recognition processing is not rejection.

[0015] According to an embodiment of the present application, the confidence threshold and the application scenario of the preset model have a corresponding relationship.

[0016] According to an embodiment of the present application, when the input content is synchronously recognized and intent-recognized by the preset model, the method further comprises: obtaining context information of the input content; and adjusting the strategy of the recognition and intent-recognition processing according to the context information.

[0017] To achieve the above object, the second aspect of the present application provides a computer readable storage medium, which stores an intent-recognition program, and the intent-recognition program is executed by a processor to implement the intent-recognition method of the foregoing embodiments of the present application.

[0018] The computer readable storage medium according to the embodiments of the present application can synchronously process the recognition and intent-recognition tasks by executing the intent-recognition program by the processor, thereby improving the response speed and resource utilization, enhancing the accuracy and robustness, simplifying the system architecture, and reducing the maintenance cost and complexity.

[0019] To achieve the above object, the third aspect of the present application provides an intent-recognition device, which comprises: an obtaining module configured to obtain a recognition training set and an intent-recognition training set; a training module configured to train a preset model according to the recognition training set and the intent-recognition training set, so that the preset model has both recognition processing capability and intent-recognition processing capability; and a processing module configured to synchronously recognize and intent-recognize input content by the preset model when a service is called. The processing module is further configured to return a preset bottom-up script and terminate the process if the result of the recognition and intent-recognition processing is recognition, and to fall the intent-recognition result into corresponding agent processing if the result of the recognition and intent-recognition processing is non-recognition.

[0020] The intent-recognition device according to the embodiments of the present application can obtain the recognition training set and the intent-recognition training set by the obtaining module, and train the preset model according to the recognition training set and the intent-recognition training set by the training module, so that the preset model has both recognition processing capability and intent-recognition processing capability. The processing module is configured to synchronously recognize and intent-recognize input content by the preset model when a service is called. If the result of the recognition and intent-recognition processing is recognition, the processing module returns a preset bottom-up script and terminates the process. If the result of the recognition and intent-recognition processing is non-recognition, the processing module falls the intent-recognition result into corresponding agent processing. In this way, the recognition and intent-recognition tasks can be synchronously processed, thereby improving the response speed and resource utilization, enhancing the accuracy and robustness, simplifying the system architecture, and reducing the maintenance cost and complexity.

[0021] To achieve the above object, the fourth aspect of the present application provides an electronic device comprising the intent-recognition device of the foregoing embodiments of the present application.

[0022] According to the electronic device of the embodiment of the present application, the rejection and the intention recognition task can be processed synchronously by using the intention recognition device of the above-mentioned embodiment of the present application, and the response speed and the resource utilization rate are improved, the accuracy and the robustness are enhanced, the system architecture is simplified, and the maintenance cost and complexity are reduced.

[0023] Additional aspects and advantages of the present application will be described in the following description, will become apparent from the following description, or will be learned by practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 is a flowchart of an intention recognition method according to an embodiment of the present application;

[0025] Figure 2 is a block diagram of an intention recognition device according to an embodiment of the present application;

[0026] Figure 3 is a block diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0027] Embodiments of the present application are described in detail below with reference to the accompanying drawings. The same or similar components are denoted by the same or similar reference numerals throughout the drawings, and repeated description thereof will be omitted. The embodiments described below are examples for explaining the present application, and should not be construed as limiting the present application.

[0028] An intention recognition method, a computer readable storage medium, an intention recognition device, and an electronic device according to embodiments of the present application are described below with reference to the accompanying drawings.

[0029] Figure 1 is a flowchart of an intention recognition method according to an embodiment of the present application.

[0030] Specifically, in some embodiments of the present application, as shown in Figure 1 the intention recognition method includes:

[0031] S101, obtaining a rejection training set and an intention recognition training set.

[0032] Specifically, in this embodiment, the rejection training set and the intent recognition training set can be obtained from public data sets and actual business scenarios. The rejection training set includes multiple rejection tasks, covering invalid inputs, sensitive information, ambiguous expressions, etc. in different fields, such as rejecting inputs related to political sensitive topics, rejecting processing of obviously meaningless random character combinations, etc. In addition, various boundary conditions such as too short or too long input are also included. The intent recognition training set includes multiple intent recognition tasks, such as weather query, news information acquisition, online shopping, customer service consultation, etc. Here, the content of the rejection task and the intent recognition task can not be specifically limited.

[0033] S102, training the preset model according to the rejection training set and the intent recognition training set, so that the preset model has rejection processing capability and intent recognition processing capability at the same time.

[0034] Specifically, in this embodiment, the preset model is constructed using a deep learning architecture. The model is based on a multi-task learning framework and can handle rejection and intent recognition tasks at the same time. The bottom layer of the preset model is a shared feature extraction module, which uses an advanced Transformer architecture containing multiple self-attention mechanism layers that can capture long-distance dependency relationships and semantic information in the input text, providing deep feature representation for the text. Through joint training on the rejection training set and the intent recognition training set, the preset model can learn the common features of rejection tasks and intent recognition tasks and effectively fuse the features in the shared layer. This can improve system performance and save GPU resources.

[0035] S103, when the service is called, the input content is simultaneously rejected and intent-recognized by the preset model.

[0036] Specifically, in this embodiment, when the service receives the input content of the user, a series of preprocessing operations will be performed on the input content, including text cleaning, word segmentation, noise removal, etc. to ensure that the quality and format of the input data meet the requirements of the model. Then, the preprocessed input content is input into the preset model.

[0037] After receiving the input, the preset model uses its shared feature extraction module to extract features from the input content and generate a unified feature representation. This feature representation contains both semantic information of the input text and features related to rejection and intent recognition tasks. Next, the rejection branch and the intent recognition branch of the model process the feature representation at the same time. The rejection branch judges whether the input content belongs to the category that should be rejected, while the intent recognition branch predicts the intent category of the input content.

[0038] During the processing, the two task branches of the preset model can learn from and complement each other through the shared feature representation, improving the processing efficiency and accuracy. For example, the intent recognition task can provide richer semantic information for the rejection task, helping it better determine whether the input content belongs to the boundary case; and the judgment result of the rejection task can also provide a reference for the intent recognition task, making it more accurately identify the intent in the non-rejection case. Through this synchronous processing method, the preset model can complete the rejection and intent recognition processing at the same time in a short time, greatly reducing the decision delay and improving the response speed of the system.

[0039] S104, if the results of the rejection and intent recognition processing are rejection, return the preset bottom case and terminate the process, if the results of the rejection and intent recognition processing are not rejected, the intent recognition result is domain to the corresponding agent processing.

[0040] Specifically, in this embodiment, if the results of the rejection and intent recognition processing are rejection, it means that the input content belongs to the category that should be rejected, and the system will return the pre-set bottom case to the user and terminate the subsequent processing process. Among them, the bottom case can be designed according to different rejection scenarios, which can not only convey the reason for rejection to the user, but also maintain good user experience and avoid user confusion or dissatisfaction.

[0041] If the results of the rejection and intent recognition processing are not rejected, it means that the input content is within the capability range of the model, and then the system will domain the intent recognition result to the corresponding agent for further processing. Specifically, the system will assign the request to the business agent that matches the predicted intent category in the intent recognition result. For example, if the intent recognition result is "query weather", the request will be assigned to the weather query agent; if the intent recognition result is "buy goods", the request will be assigned to the e-commerce transaction agent.

[0042] After the intent recognition result is domain to the corresponding agent, the agent will process the user's request according to its own business logic and functional modules. For example, the agent will further call related service interfaces, database queries or other external resources to obtain information that meets the user's needs or perform corresponding operations. During the processing, the agent can have multiple rounds of dialogue with the user to clarify the user's intent, obtain more context information or guide the user to complete a specific operation process. Finally, the agent will generate a corresponding answer or execution result according to the processing result and return it to the user, completing the entire service call process. Thus, efficient and accurate processing of user input content is achieved, which improves the response performance and user experience while ensuring the security of the output content, and reduces the complexity and resource cost of the system.

[0043] Further, in some embodiments of the present application, the rejection training set includes multiple rejection tasks, and the intent recognition training set includes multiple intent recognition tasks, wherein the rejection task is defined as a special category of the intent recognition task, so that the rejection task and the intent recognition task complete classification under a unified classification framework, and the repeated features of the rejection task and the intent recognition task are eliminated through a shared feature extraction module.

[0044] Specifically, in this embodiment, the rejection training set includes multiple rejection tasks, including but not limited to identifying inputs containing sensitive information, invalid content, ambiguous expressions, inputs beyond the model's ability range, and various boundary cases, etc. These rejection task samples are carefully screened and labeled to ensure that each sample has a clear rejection label to help the model accurately identify the content that should be rejected. For example, rejecting inputs that answer political sensitive topics, rejecting processing obviously meaningless random character combinations, rejecting responding to queries unrelated to business, etc. The intent recognition training set includes multiple intent recognition tasks, such as weather query, news information acquisition, online shopping, customer service consultation, technical support, etc.

[0045] The rejection task can be defined as a special category of the intent recognition task, such as being regarded as a "unknown intent" or "sensitive topic" category. In this way, the rejection task and the intent recognition task can be completed under the same classification framework. Secondly, through the shared feature extraction module, the rejection task and the intent recognition task can share the underlying feature representation, thereby avoiding repeated calculation and improving resource utilization efficiency. The shared feature extraction module is usually based on a deep learning architecture, such as a Transformer architecture, which can capture long-distance dependency relationships and semantic information in the input text and perform deep feature representation on the text. Through the shared feature extraction module, the model can learn the rejection task while learning the intent recognition task, achieving collaborative optimization of the two tasks.

[0046] Further, in some embodiments of the present application, the preset model includes a rejection sub-model and an intent recognition sub-model, the rejection sub-model includes a plurality of first model parameters, the intent recognition sub-model includes a plurality of second model parameters, and the plurality of first model parameters and the plurality of second model parameters have partially shared model parameters.

[0047] Specifically, in this embodiment, the preset model adopts a multi-task learning architecture, including a rejection sub-model and an intent recognition sub-model. The rejection sub-model is responsible for judging whether the input belongs to the category that should be rejected, and the intent recognition sub-model focuses on identifying the specific intent of the input. The rejection sub-model includes a plurality of first model parameters for rejection task judgment, and the intent recognition sub-model includes a plurality of second model parameters for intent recognition task prediction.

[0048] Among them, there are partially shared model parameters between the first model parameters and the second model parameters. These shared parameters are mainly concentrated in the bottom feature extraction part of the model, which is usually based on the encoding layer of the Transformer architecture. The Transformer architecture contains multiple self-attention mechanism layers, which can capture long-distance dependency and semantic information in the input text, and perform deep feature representation on the text. By sharing these bottom feature extraction modules, the two sub-models can fully utilize the same text features, reduce repeated calculation, and improve resource utilization efficiency.

[0049] During the training process of multi-task learning, the rejection sub-model and the intent recognition sub-model can learn together and promote each other. For example, in terms of training data, the data of the rejection task can help the intent recognition sub-model better understand the boundary conditions and improve its generalization ability; and the data of the intent recognition task can provide richer semantic information for the rejection sub-model to improve its judgment accuracy in complex situations. In this way, the preset model can have rejection processing capability and intent recognition processing capability under the multi-task learning framework, improve the accuracy and efficiency of intent recognition under the premise of ensuring the safety of output content, and realize efficient multi-task processing.

[0050] Further, in some embodiments of the present application, training the preset model according to the rejection training set and the intent recognition training set comprises: enhancing the boundary of the intent recognition sub-model according to the rejection task, and enhancing the semantic information of the rejection sub-model according to the intent recognition task.

[0051] Specifically, in this embodiment, on the one hand, the boundary of the intent recognition sub-model can be enhanced according to the rejection task. In the rejection training set, there are often a large number of boundary condition samples, i.e. input instances that are easily confused with normal intent and are at the edge of judgment. By inputting these boundary condition samples into the intent recognition sub-model, the intent recognition sub-model is trained to learn the features of these difficult cases, so that the intent recognition sub-model can more accurately identify and demarcate its own ability boundary when facing ambiguous input, thereby improving its generalization ability and processing precision for edge cases. For example, for some semantically ambiguous user input that may involve multiple intent categories, the intent recognition sub-model after enhanced training can more accurately judge the true intent and avoid misclassification.

[0052] On the other hand, the semantic information of the rejection sub-model can be enhanced training according to the intent recognition task. The samples in the intent recognition training set are rich in diversified semantic information, which is a valuable knowledge source for the rejection sub-model. Using the samples in the intent recognition training set to train the rejection sub-model can enable the rejection sub-model to deeply understand the semantic connotation of the text, so as to more accurately judge whether the input belongs to the category that should be rejected. For example, for an input containing complex semantics or implicit expression, the rejection sub-model after enhanced training can penetrate the surface words and grasp the deep meaning, and then make more accurate rejection decisions.

[0053] Further, the preset model can achieve significant performance improvement in rejection tasks and intent recognition tasks. The rejection sub-model can accurately identify various types of content that should be rejected while better understanding the semantics of the text and reducing false positives. The intent recognition sub-model enhances the processing capability of boundary conditions based on accurate identification of user intent, improving the overall robustness and generalization ability. Thus, the preset model can more efficiently and intelligently process user inputs in practical applications, providing safer and more accurate services to users.

[0054] Further, in some embodiments of the present application, the rejection and intent recognition processing of the input content by the preset model includes: obtaining an intent recognition confidence corresponding to the input content by the preset model; when the intent recognition confidence is lower than a confidence threshold, determining that the result of the rejection and intent recognition processing is rejection; and when the intent recognition confidence is higher than or equal to the confidence threshold, determining that the result of the rejection and intent recognition processing is non-rejection.

[0055] Specifically, in this embodiment, first, when the service receives the input content of the user, a series of preprocessing operations will be performed on the input content, including but not limited to text cleaning, word segmentation, noise removal, etc., to ensure that the quality and format of the input data meet the requirements of the model. Then, the preprocessed input content is input into the preset model. Inside the preset model, the shared feature extraction module will first extract features from the input content to obtain a feature vector that can represent the input content. The feature vector contains semantic information, context information, and other features related to rejection tasks and intent recognition tasks of the input text.

[0056] Next, the rejection sub-model and the intent recognition sub-model will process the feature vector simultaneously. The rejection sub-model uses its unique first model parameters and part of the model parameters shared with the intent recognition sub-model to judge whether the input content belongs to the category that should be rejected, and outputs a rejection confidence. The intent recognition sub-model uses its unique second model parameters and part of the shared model parameters to predict the intent category of the input content and outputs an intent recognition confidence.

[0057] After obtaining the confidence, a decision stage is entered. A confidence threshold can be set, which is set based on a large amount of experimental data and the requirements of actual application scenarios, and is not specifically limited here. When the rejection confidence is higher than or equal to the confidence threshold, it indicates that the model has a high certainty in judging that the input content belongs to the category that should be rejected, or the intent recognition confidence is lower than the confidence threshold, which indicates that the preset model has a high certainty in intent recognition of the input content, and the result of the rejection and intent recognition processing is determined to be rejection. Correspondingly, if the rejection confidence is lower than the confidence threshold, it indicates that the model has a low certainty in judging that the input content belongs to the category that should be rejected, or the intent recognition confidence is higher than or equal to the confidence threshold, which indicates that the model has a high certainty in intent recognition of the input content, and the result of the rejection and intent recognition processing is determined to be non-rejection, and the intent recognition result is domained to the corresponding agent for further processing.

[0058] In addition, when the result of the rejection and intent recognition processing is determined to be rejection, a preset bottom case can be returned to the user. These bottom cases are carefully designed according to different rejection scenarios, which can convey the reason for rejection to the user and maintain good user experience, avoiding user confusion or dissatisfaction. When the result of the rejection and intent recognition processing is determined to be non-rejection, the intent recognition result can be domained to the corresponding agent for further processing. Further, the agent can call the corresponding service interface, database query or other external resources according to the intent recognition result, execute specific business logic, generate a corresponding answer or execution result, and return it to the user.

[0059] Further, in some embodiments of the present application, there is a corresponding relationship between the confidence threshold and the application scenario of the preset model.

[0060] Specifically, in this embodiment, there is a corresponding relationship between the confidence threshold and the application scenario of the preset model. In actual application, different business scenarios have different requirements for security and accuracy. For example, in the fields of finance, medical treatment and the like, which have very high requirements for security, the auditing standard of the input content is more strict. At this time, the confidence threshold should be appropriately increased. On the contrary, in some scenarios that have higher requirements for user experience and relatively lower requirements for content security, such as entertainment, life service and the like, the confidence threshold can be appropriately reduced.

[0061] Further, in some embodiments of the present application, when the input content is simultaneously subjected to rejection and intent recognition processing by the preset model, the method further comprises: obtaining context information of the input content; and adjusting the strategy of the rejection and intent recognition processing according to the context information.

[0062] Specifically, in this embodiment, when processing the input content, the system first acquires context information related to the input. The context information includes dialogue history, user background information, and details such as time and location. The acquisition of these information aims to provide the model with more comprehensive judgment basis, enabling it to more accurately grasp the semantics and background of the input content. Subsequently, the preset model can dynamically adjust the processing strategy of rejection and intent recognition based on the acquired context information. This adjustment mechanism allows the model to flexibly adapt to different scenarios. For example, the model can dynamically adjust the confidence threshold according to the context, or specifically optimize feature extraction and processing logic to adapt to specific dialogue environments or user preferences. By integrating context information and adjusting the processing strategy accordingly, the model not only improves the accuracy of rejection and intent recognition, but also maintains high performance in variable application scenarios. This processing method enhances the model's adaptability to complex dialogue environments, providing users with a more personalized and accurate service experience.

[0063] In summary, according to the intent recognition method of the embodiment of the present application, by acquiring the rejection training set and the intent recognition training set, and training the preset model according to the rejection training set and the intent recognition training set, the preset model simultaneously has rejection processing capability and intent recognition processing capability, and when the service is called, the input content is simultaneously subjected to rejection and intent recognition processing by the preset model, if the results of rejection and intent recognition processing are rejection, the preset bottom script is returned and the process is terminated, if the results of rejection and intent recognition processing are not rejection, the intent recognition result is domained to the corresponding intelligent agent processing. Thus, the rejection and intent recognition tasks can be simultaneously processed, thereby improving the response speed and resource utilization, enhancing the accuracy and robustness, simplifying the system architecture, and reducing the maintenance cost and complexity.

[0064] Based on the intent recognition method proposed in the foregoing embodiment of the present application, the embodiment of the present application further proposes a computer readable storage medium having an intent recognition program stored thereon, the intent recognition program being executed by a processor to implement the intent recognition method of the foregoing embodiment of the present application.

[0065] The computer readable storage medium according to the embodiment of the present application, by executing the intent recognition program by the processor, can simultaneously process the rejection and intent recognition tasks, thereby improving the response speed and resource utilization, enhancing the accuracy and robustness, simplifying the system architecture, and reducing the maintenance cost and complexity.

[0066] Figure 2 is a block schematic diagram of an intent recognition device according to the embodiment of the present application.

[0067] Specifically, as shown in Figure 2 the intent recognition device 100 includes an acquisition module 10, a training module 20, and a processing module 30.

[0068] The obtaining module 10 is configured to obtain a rejection training set and an intent recognition training set; the training module 20 is configured to train a preset model according to the rejection training set and the intent recognition training set, so that the preset model has both rejection processing capability and intent recognition processing capability; and the processing module 30 is configured to, when a service is invoked, perform rejection and intent recognition processing on input content simultaneously by using the preset model, and if the result of the rejection and intent recognition processing is rejection, return a preset bottom text and terminate the process, and if the result of the rejection and intent recognition processing is non-rejection, domain the intent recognition result to corresponding agent processing.

[0069] In some embodiments of the present application, the rejection training set includes multiple rejection tasks, and the intent recognition training set includes multiple intent recognition tasks, wherein the rejection task is defined as a special category of the intent recognition task, so that the rejection task and the intent recognition task are completed in a unified classification framework, and the repeated features of the rejection task and the intent recognition task are eliminated through a shared feature extraction module.

[0070] In some embodiments of the present application, the preset model includes a rejection sub-model and an intent recognition sub-model, the rejection sub-model includes multiple first model parameters, the intent recognition sub-model includes multiple second model parameters, and the multiple first model parameters and the multiple second model parameters have part of shared model parameters.

[0071] In some embodiments of the present application, the training module 20 is specifically configured to perform enhanced training on the boundary of the intent recognition sub-model according to the rejection task, and perform enhanced training on the semantic information of the rejection sub-model according to the intent recognition task.

[0072] In some embodiments of the present application, the processing module 30 is specifically configured to obtain an intent recognition confidence corresponding to the input content through the preset model; when the intent recognition confidence is lower than a confidence threshold, determine that the result of the rejection and intent recognition processing is rejection; and when the intent recognition confidence is higher than or equal to the confidence threshold, determine that the result of the rejection and intent recognition processing is non-rejection.

[0073] In some embodiments of the present application, the confidence threshold has a corresponding relationship with the application scenario of the preset model.

[0074] In some embodiments of the present application, the processing module 30 is specifically configured to obtain context information of the input content; and adjust the strategy of the rejection and intent recognition processing according to the context information.

[0075] It should be noted that other specific embodiments of the intent recognition device proposed in the embodiments of the present application can refer to the specific embodiments of the intent recognition method of the foregoing embodiments of the present application, and to reduce redundancy, will not be repeated here.

[0076] In summary, according to the intention recognition device in the embodiment of the present application, the rejection training set and the intention recognition training set are obtained by the obtaining module, and the preset model is trained according to the rejection training set and the intention recognition training set by the training module, so that the preset model has rejection processing capability and intention recognition processing capability at the same time. When the service is called, the input content is simultaneously subjected to rejection and intention recognition processing by the preset model through the processing module. If the result of the rejection and intention recognition processing is rejection, the preset bottom text is returned and the process is terminated. If the result of the rejection and intention recognition processing is not rejection, the intention recognition result is domained to the corresponding intelligent agent processing. Thus, the rejection and intention recognition tasks can be simultaneously processed, thereby improving the response speed and resource utilization, enhancing the accuracy and robustness, simplifying the system architecture, and reducing the maintenance cost and complexity.

[0077] Figure 3 is a block schematic diagram of an electronic device according to an embodiment of the present application.

[0078] As shown in Figure 3 , the electronic device 1000 includes the intention recognition device 100 of the above embodiment of the present application.

[0079] According to the electronic device in the embodiment of the present application, by adopting the intention recognition device in the above embodiment of the present application, the rejection and intention recognition tasks can be simultaneously processed, thereby improving the response speed and resource utilization, enhancing the accuracy and robustness, simplifying the system architecture, and reducing the maintenance cost and complexity.

[0080] It is to be appreciated that the above description and the examples that follow are intended to be illustrative only and that changes can be made to the description, as represented by the above listed elements, by the steps recited in the flow charts, and by the examples that follow, without departing from the spirit of the application. Accordingly, the scope of the present application is intended to be defined only by the appended claims.

[0081] It should be understood that aspects of the application can be implemented in hardware, software, firmware or a combination thereof. In the above embodiments, various steps or methods can be implemented in software or firmware that is stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any of the following techniques, which are well known in the art of hardware implementation, can be used: a hybrid of the above techniques, a mixture of two or more of the above techniques, or a combination of the above techniques with other techniques not listed above.

[0082] In the description of the present application, the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" are intended to mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the present application. The illustrative appearances of the above-mentioned terms in various places in the specification are not intended to exclude that the terms in one place can refer to the same feature, structure, material, or characteristic as in another place, even if the above-mentioned terms are not explicitly used in one place. Furthermore, it is intended that the specific features, structures, materials, or characteristics described can be combined in any and all suitable ways.

[0083] In the description of the application, it should be understood that the orientation or positional relationship indicated by the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", "axial", "radial", "circumferential" and the like is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the application and simplifying the description, and does not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the application.

[0084] In addition, the terms "first", "second" are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the application, the meaning of "a plurality of" is at least two, such as two, three, etc., unless otherwise explicitly specified and limited.

[0085] In the present application, unless otherwise explicitly specified and limited, the terms "mounting", "connecting", "connecting", "fixing" and the like should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or it can be integrated; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be the internal communication of two elements or the interaction relationship between two elements, unless otherwise explicitly limited. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0086] In the present application, unless otherwise explicitly specified and limited, the first feature is "on" or "under" the second feature, which can be direct contact between the first and second features, or indirect contact between the first and second features through an intermediate medium. Moreover, the first feature "above", "above" and "above" the second feature can be directly above or obliquely above the first feature, or only indicate that the horizontal height of the first feature is higher than that of the second feature. The first feature "below", "below" and "below" the second feature can be directly below or obliquely below the first feature, or only indicate that the horizontal height of the first feature is less than that of the second feature.

[0087] Although the embodiments of the present application have been shown and described above, it should be understood that the above embodiments are exemplary and cannot be understood as a limitation on the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

Claims

1. A method for identifying intention, characterized in that: The method comprises: Obtain rejection training sets and intent recognition training sets; Training a preset model based on the rejection training set and the intent recognition training set, so that the preset model has both rejection processing capabilities and intent recognition processing capabilities; When the service is called, the input content is simultaneously rejected and the intent is recognized through the preset model; If the result of the rejection and intention recognition processing is rejection, the preset fallback text is returned and the process is terminated. If the result of the rejection and intention recognition processing is not rejection, the intention recognition result is domained to the corresponding intelligent agent for processing.

2. The intention recognition method according to claim 1, characterized in that The rejection training set includes multiple rejection tasks, and the intention recognition training set includes multiple intention recognition tasks, wherein the rejection task is defined as a special category of the intention recognition task, so that the rejection task and the intention recognition task are classified under a unified classification framework, and the repeated features of the rejection task and the intention recognition task are eliminated through a shared feature extraction module.

3. The intention recognition method according to claim 2, characterized in that The preset model includes a rejection sub-model and an intention recognition sub-model, the rejection sub-model includes multiple first model parameters, the intention recognition sub-model includes multiple second model parameters, and the multiple first model parameters and the multiple second model parameters have some shared model parameters.

4. The intention recognition method according to claim 3, characterized in that: Training a preset model according to the rejection training set and the intent recognition training set includes: The boundary of the intention recognition sub-model is enhanced and trained according to the rejection task, and the semantic information of the rejection sub-model is enhanced and trained according to the intention recognition task.

5. The intention recognition method according to claim 1, characterized in that: The synchronously performing rejection and intent recognition processing on the input content by using the preset model includes: Obtaining the intent recognition confidence corresponding to the input content through the preset model; When the intention recognition confidence is lower than a confidence threshold, determining that a result of the rejection and intention recognition processing is rejection; When the intention recognition confidence is higher than or equal to a confidence threshold, the result of the rejection and intention recognition processing is determined to be non-rejection.

6. The intention recognition method according to claim 5, characterized in that: There is a corresponding relationship between the confidence threshold and the application scenario of the preset model.

7. The intention recognition method according to claim 1, characterized in that: When simultaneously performing rejection and intent recognition processing on the input content through the preset model, the method further includes: Obtaining context information of the input content; The strategies for the rejection and intent recognition processing are adjusted according to the context information.

8. A computer-readable storage medium, characterized in that An intention recognition program is stored thereon, and when the intention recognition program is executed by a processor, the intention recognition method according to any one of claims 1 to 7 is implemented.

9. An intention recognition device, characterized in that: The device comprises: An acquisition module is used to obtain a rejection training set and an intent recognition training set; A training module, configured to train a preset model based on the rejection training set and the intent recognition training set, so that the preset model has both rejection processing capabilities and intent recognition processing capabilities; A processing module, configured to simultaneously perform rejection and intent recognition processing on the input content through the preset model when the service is called; The processing module is also used to return the preset fallback text and terminate the process if the result of the rejection and intention recognition processing is rejection; if the result of the rejection and intention recognition processing is not rejection, the intention recognition result is domained to the corresponding intelligent agent for processing.

10. An electronic device, characterized in that: Includes the intention recognition device as described in claim 9.

Images