Intention recognition method and device

Abstract

The invention discloses an intention recognition method and device, and the method comprises the steps: inputting first question information inputted by a user into a first model for information enhancement processing, obtaining a first processing result, inputting the first processing result into a second model for intention recognition processing, and obtaining a second processing result; the second processing result comprises at least one candidate intention corresponding to the first question information; determining a target intention corresponding to the first problem information based on the second processing result; wherein the first model is a large model, and the second model is a deep learning model which aims at the same problem information and has higher intention recognition speed than the large model.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence applications, and in particular to an intent recognition method and device. Background Art

[0002] In today's field of artificial intelligence (AI) technology, intent recognition, as a key natural language processing task, is widely used in scenarios such as intelligent customer service, voice assistants, search engines, and personalized recommendation systems. It aims to understand the true intentions behind human language and provide more personalized and efficient services.

[0003] In practical applications, it is usually necessary to fine-tune pre-trained models such as Bert using labeled data to identify the intent of the input user query. However, this requires a large amount of labeled data, which not only increases the cost of data collection, but also affects the generalization and scalability of the model. In addition, when the user query is rich and too colloquial, or even expresses multiple intents in one sentence, it is difficult for the model to accurately identify the true intent of the user query, which reduces the user interaction experience. Summary of the invention

[0004] In view of the above problems, this application provides the following solutions:

[0005] The first aspect of the present application provides an intention recognition method, the method comprising:

[0006] Obtain the first question information input by the user;

[0007] Inputting the first question information into a first model for information enhancement processing to obtain a first processing result;

[0008] Inputting the first processing result into a second model for intent recognition processing to obtain a second processing result; the second processing result includes at least one candidate intent corresponding to the first question information;

[0009] Based on the second processing result, determining a target intention corresponding to the first question information;

[0010] Among them, the first model is a large model, and the second model is a deep learning model that recognizes the intent of the same question information faster than the large model.

[0011] In a possible implementation, inputting the first question information into a first model for information enhancement processing to obtain a first processing result, and inputting the first processing result into a second model for intent recognition processing to obtain a second processing result includes:

[0012] Inputting the first question information into the first model to decompose the complex question, and obtaining at least one second question information contained in the first question information; the second question information has a single intention or a simple intention, and the first question information has multiple intentions or a complex intention;

[0013] Input at least one of the second question information into the second model for intention recognition processing to obtain a second processing result corresponding to at least one of the second question information; the second processing result includes at least one predicted intention corresponding to the second question information and the prediction confidence corresponding to each of the predicted intentions.

[0014] In a possible implementation, determining the target intention corresponding to the first question information based on the second processing result includes:

[0015] Based on the prediction confidence, select k candidate intentions from the predicted intentions corresponding to each of the second question information;

[0016] Obtaining intent description information of each of the candidate intents;

[0017] At least the k candidate intents and the intent description information of each candidate intent are input into the first model, so that the first model processes the k candidate intents at least based on the intent description information to determine the target intent corresponding to the first question information.

[0018] In a possible implementation, determining the target intention corresponding to the first question information based on the second processing result includes at least one of the following:

[0019] If it is determined that the prediction confidence of at least one of the predicted intentions corresponding to each of the second question information is greater than or equal to the confidence threshold, the at least one predicted intention is determined as the target intention corresponding to the first question information;

[0020] If it is determined that the prediction confidences of the predicted intentions corresponding to the second question information are all less than the confidence threshold, at least each predicted intention is input into the first model, and the first model processes each predicted intention based on at least the intention description information of each predicted intention to determine the target intention corresponding to the first question information;

[0021] If it is determined that the prediction confidences of the predicted intentions corresponding to the respective second question information are all less than the confidence threshold, determine candidate intentions among the predicted intentions, at least input the candidate intentions into the first model, and the first model processes the candidate intentions based on at least the intention description information of the candidate intentions to determine the target intention corresponding to the first question information;

[0022] If the prediction confidence of the predicted intention corresponding to at least one of the multiple second question information is less than the confidence threshold, determine the candidate intention among the predicted intentions corresponding to the at least one second question information, at least input the candidate intention into the first model, and the first model processes the candidate intention at least based on the intention description information of the candidate intention to obtain the target intention corresponding to the second question information; wherein, the target intention corresponding to the second question information and the predicted intentions corresponding to other second question information whose prediction confidence is greater than or equal to the confidence threshold constitute the target intention corresponding to the first question information.

[0023] In a possible implementation, determining the target intention corresponding to the first question information based on the second processing result further includes:

[0024] Based on the candidate intent or the predicted intent, obtaining fine-tuning data;

[0025] The fine-tuning data is used to adjust the parameters of the first model to obtain an adjusted first model that can understand the candidate intent or the predicted intent, so as to process the candidate intent or the predicted intent through the adjusted first model to obtain the target intent.

[0026] In a possible implementation, inputting the first question information into the first model to decompose the complex question and obtaining at least one second question information included in the first question information includes at least one of the following:

[0027] Input the first question information into a first model, perform semantic analysis on the first question information and historical question information input adjacent to the first question information, and obtain at least one second question information included in the first question information;

[0028] The first question information is input into a first model for intent recognition, and the first question information is decomposed into multiple second question information according to multiple intents or complex intents of the first question information.

[0029] In a possible implementation, the method further includes:

[0030] Performing a complexity evaluation on the first question information to determine that the first question information belongs to a target category question;

[0031] If the target category problem is a first category problem, executing the step of inputting the first problem information into the first model for information enhancement processing;

[0032] If the target category question is a second category question, input the first question information into the second model, convert the first question information into a first question vector through the embedding network in the second model, perform intent recognition on the first question vector through the converter network of the second model, and determine the target intent corresponding to the first question information;

[0033] The complexity of the first category of problems is higher than the complexity of the second category of problems.

[0034] In a possible implementation, performing complexity evaluation on the first question information to determine whether the first question information belongs to a target category question includes:

[0035] Inputting the first question information into a question classification model for complexity evaluation, and determining that the first question information belongs to a target category question that matches a complexity evaluation result;

[0036] The question classification model is obtained by training a classifier using training questions with category labels, and the category labels are determined based on the intention recognition results of the first model and the second model for the corresponding training questions.

[0037] In a possible implementation, the method further includes at least one of the following:

[0038] If the target intention includes a first intention to control the electronic device, in response to a first working state of the electronic device not matching the first intention, outputting a device control page matching the first intention, and in response to a control input operation on the device control page, controlling the electronic device to switch from the first working state to a second working state matching the first intention;

[0039] If the target intention includes a first intention to control the electronic device, in response to the second working state of the electronic device matching the first intention, outputting corresponding matching prompt information;

[0040] If the target intent includes a second intent in the question-answering task, the first model outputs target content for answering the first question information based on the second intent;

[0041] If the target intent includes a third intent to call a cloud service, in response to the third intent, output a cloud service calling interface or a cloud service interaction interface of the called cloud service; the cloud service calling interface displays at least one calling method of the intended calling cloud service;

[0042] If the target intent includes a fourth intent to launch an application, in response to the fourth intent, an application interaction interface intended to launch the application is output.

[0043] A second aspect of the present application provides an intention recognition device, the device comprising:

[0044] A question information obtaining module, used to obtain first question information input by a user;

[0045] A first processing module, used for inputting the first question information into a first model for information enhancement processing to obtain a first processing result;

[0046] A second processing module is used to input the first processing result into a second model for intent recognition processing to obtain a second processing result; the second processing result includes at least one candidate intent corresponding to the first question information;

[0047] a target intention determination module, configured to determine a target intention corresponding to the first question information based on the second processing result;

[0048] Among them, the first model is a large model, and the second model is a deep learning model that recognizes the intent of the same question information faster than the large model. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. Throughout the accompanying drawings, the same or similar reference numerals represent the same or similar elements. It should be understood that the drawings are schematic and the originals and elements are not necessarily drawn to scale.

[0050] Figure 1 A flowchart of an intention recognition method proposed in Example 1 of the present application;

[0051] Figure 2 A flowchart of an intention recognition method proposed in Embodiment 2 of the present application;

[0052] Figure 3 A flowchart of an intention recognition method proposed in Embodiment 3 of the present application;

[0053] Figure 4 A flowchart of an intention recognition method proposed in Embodiment 4 of the present application;

[0054] Figure 5 A flowchart of an intention recognition method proposed in Embodiment 5 of the present application;

[0055] Figure 6 A schematic diagram of the structure of an intention recognition device provided in an embodiment of the present application;

[0056] Figure 7The present invention is a schematic diagram of the hardware structure of an electronic device suitable for the intention recognition method proposed in this application. DETAILED DESCRIPTION

[0057] The embodiments of the present application are described below in conjunction with the drawings in the embodiments of the present application. The terms used in the implementation mode of the present application are only used to explain the specific embodiments of the present application and are not intended to limit the present application. The embodiments of the present application are described below in conjunction with the drawings. It is known to those of ordinary skill in the art that with the development of technology and the emergence of new scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0058] The term "first", "second" etc. that the full text of the application relates to is used to distinguish similar objects, and need not be used to describe a specific order or sequential order. It should be understood that the terms used in this way can be interchangeable in appropriate circumstances, and this is only to describe the distinction mode adopted by the objects of the same attributes when describing in the embodiments of the application. In addition, the terms "including" and "having" and any of their deformations are intended to cover non-exclusive inclusions, so that the process, method, system, product or equipment comprising a series of units need not be limited to those units, but may include other units that are not clearly listed or inherent to these processes, methods, products or equipment.

[0059] As described in the background technology section, with the rapid development of artificial intelligence (AI), it is proposed to call LLM (Large Language Model) or large models (LM) to identify the true intention of complex user queries, such as using prompt engineering to input user queries and intention systems into large models / LLMs, and use the model's language understanding ability to accurately identify the true intentions of user queries. Among them, when there are many intent categories, it is often limited by the length limit of the LLM / large model context. The candidate intent list is often roughly recalled first, and then the LLM / large model is used to judge it to determine the target intent of the user query. This requires the LLM / large model to be started for a long time, which increases the intent recognition response time, and the long-term occupation of a large amount of computing and memory resources of electronic devices will lead to excessive power consumption and affect other tasks performed by electronic devices.

[0060] In order to solve the above technical problems, the embodiment of the present application provides an intent recognition method and device. For the query input by the user, the information enhancement processing can be performed using a large model first, and then a deep learning model with a higher intent recognition response speed (whose model parameter amount is much less than the model parameter amount of the large model) can be input for intent recognition processing. Based on the obtained processing result containing at least one candidate intent corresponding to the user query, the target intent corresponding to the user query is quickly and accurately determined. Compared with the implementation method of only using a large model to identify the true intent of the user query, the present application uses the deep learning model to more quickly and accurately recall the candidate intent corresponding to the user query, improve the accuracy and efficiency of the target intent recognition corresponding to the user query, and reduce the startup time of the large model, thereby reducing the occupation of the computing and memory resources of the electronic device, reducing the power consumption of the electronic device and the adverse effects on other tasks performed by the electronic device. The intent recognition method of the embodiment of the present application is described in detail below in conjunction with the accompanying drawings.

[0061] Reference Figure 1 , is a flowchart of an intent recognition method proposed in the first embodiment of the present application. The intent recognition method can be applied to electronic devices, which can be terminal devices or servers, such as applications with question-answering processing functions running on the terminal device or cloud services provided by cloud servers, etc., or servers can cooperate with terminal devices to implement the intent recognition method proposed in the present application. Figure 1 As shown, the intention recognition method proposed in this embodiment may include:

[0062] Step S11, obtaining first question information input by the user;

[0063] In an embodiment of the present application, the first question information may be a question Query input by a user using an input component of an electronic device (i.e., a user terminal), and may be one or more combinations of text information, image information, video information, or audio information. For different types of first question information, corresponding types of input components may be used to complete the input operation. The present application does not impose any restrictions on the input implementation method of the first question information.

[0064] Exemplarily, the user can input first question information such as "Check whether the computer's flight power mode has customized settings recommended by the manufacturer", "Query pictures with the theme of AI transformation", "Turn on Bluetooth and eye protection mode", and "Increase screen brightness, increase it a little more" through voice; or input / specify a document while inputting the first question information in voice / text format such as "Can you help me review the content of this PPT? Please adjust the incoherent language and improve the unclear parts"; the user can use one or more appropriate input methods and language types to input the corresponding first question information to the electronic device according to the actual task processing requirements. One question information can correspond to one or more intentions. The present application does not limit the content of the first question information input each time.

[0065] It should be understood that for the first question information or the object to be processed contained therein (such as the content itself or its access address, etc. of a specified document, video or image, etc.), other input components such as fingers, stylus, mouse keys or joysticks can also be used to perform input operations on the touch screen or its accessories (such as selecting and inputting various objects displayed on the interactive interface), and the electronic device responds to the input operation to obtain the first question information input by the user. It can be seen that the present application can flexibly select one or more combined input methods to complete the input operation of the first question information according to the actual task processing requirements, including but not limited to the implementation methods listed in the present application.

[0066] Step S12, inputting the first question information into the first model for information enhancement processing to obtain a first processing result;

[0067] In an embodiment of the present application, in order to improve the response speed of intent recognition, relative to inputting the first question information into a large model for intent recognition, determining the implementation method of the real intention corresponding to the first question information input by the user, as described above in the relevant description of the technical solution of the present application, the present application proposes to perform information enhancement processing on the first question information in advance, so that the first processing result obtained can express each intention more clearly and explicitly relative to the first question information, such as the various intentions involved in the first question information input by the user to request the electronic device to perform the expected task, which helps to improve the processing efficiency, reliability and accuracy of subsequent intent recognition processing steps.

[0068] In practical applications, the first model can use appropriate information enhancement (i.e., data enhancement) methods based on the semantic recognition results of the first question information and the information type to process the first question information input by the user into a model that matches its meaning. Figure 1 The present application does not limit the information enhancement processing method of the first model on the input first question information.

[0069] For example, the original first question information may be processed by text transformation, expansion or improvement, or by using image data enhancement methods based on deep learning / machine learning (such as one or more combinations of adversarial generation, random flipping / scaling / attribute adjustment, and meta-learning data enhancement, etc.) to enhance the image data contained in the first question information; for the audio data contained in the first question information or the voice signal input by the user into the first question information, one or more methods such as changing the volume and adjusting the speech speed may be used to enhance the audio data of the first question information. According to actual needs, one or more combinations of the text data enhancement methods, image data enhancement methods, and audio data enhancement methods listed in this application may be flexibly selected but not limited to implementation, and the embodiments of this application will not be described in detail here.

[0070] Step S13, inputting the first processing result into the second model for intent recognition processing to obtain a second processing result; wherein the first model is a large model, and the second model is a deep learning model that recognizes intent for the same question information faster than the large model; the second processing result includes at least one candidate intent corresponding to the first question information;

[0071] Following the above analysis, after using the large model to obtain the enhanced first processing result of the input first question information, the present application selects a second model with fewer model parameters than the large model (even at least one parameter level smaller) to continue to perform intent recognition processing on the first processing result. The second model is usually a deep learning model with thousands or millions of model parameters, which is equivalent to using a large model with millions, billions or hundreds of billions of model parameters to perform intent recognition processing on the first processing result. The present application greatly reduces the amount of model calculations in the intent recognition processing process and shortens the model response time, thereby reducing the resource occupation time and resource consumption of the electronic device during the model startup period, reducing the power consumption of the electronic device, and better adapting to the end-side device with limited resources (i.e., user terminal).

[0072] It should be noted that the present application does not restrict the network structure of the second model and its method of processing the intent recognition of the input first processing result. In order to avoid the second model being heavily dependent on labeled training data, resulting in excessively high costs for labeled training data and affecting the generalization and scalability of the model, the second model is not obtained by fine-tuning pre-trained models such as Bert, and can reuse deep learning models with intent recognition capabilities that electronic devices have or can call. The present application does not restrict the method of using the second model.

[0073] It can be seen that the present application uses the second model to perform intent recognition processing on the first processing result, and quickly and accurately obtains all possible true intentions corresponding to the first question information (the true intention can be recorded as the target intention). The possible true intention can be recorded as a candidate intention, which is equivalent to roughly retrieving the candidate intention directly from the intention library. The present application improves the recall rate of true intentions. The number of candidate intentions contained in the second processing result obtained by the second model can be determined based on actual conditions and the pre-configured recall rate (which can be configured based on experiments or experience). The present application does not limit the content contained in the first processing result.

[0074] Step S14: determine the target intention corresponding to the first question information based on the second processing result.

[0075] In combination with the above description of at least one candidate intent corresponding to the first question information in the second processing result, the present application can directly analyze the at least one candidate intent based on the intent recognition accuracy / confidence of each candidate intent by the second model, and determine the target intent corresponding to the first question information, such as selecting the candidate intent with the highest intent recognition accuracy of the second model or reaching a specified preset as the target intent; it can also use the language understanding ability of the large model to correctly and fully understand each candidate intent, determine the target intent corresponding to the first question information, etc. The present application does not limit the implementation method of step S14.

[0076] In summary, in the process of realizing the intention recognition of the first question information input by the user, the present application will first use the large model to perform flexible and reasonable information enhancement processing on the first question information to ensure that the first processing result obtained is consistent with the intention of the first question information. Figure 1 On the basis of the above, it can express each intention more clearly and accurately, which helps to improve the efficiency and accuracy of subsequent intention recognition processing. In the subsequent intention recognition processing, this application will use a deep learning model that has a faster intention recognition speed than the large model, that is, a second model, to quickly and accurately obtain a second processing result containing at least one candidate intention corresponding to the first question, and thus efficiently and accurately determine the target intention corresponding to the first question information.

[0077] It can be seen that during the entire intent recognition process, the large model is not always started to occupy a large amount of computing and memory resources of the electronic device, thereby reducing the power consumption of the electronic device; and the present application uses the large model to first enhance the first question information and then input it into the second model for intent recognition processing. Compared with directly inputting the first question information into the second model for intent recognition processing, especially when the first question information is complex or the user's query expresses ambiguous intent in spoken language, the present application improves the accuracy and reliability of intent recognition and ensures the user interaction experience.

[0078] Reference Figure 2, is a flow chart of an intention recognition method proposed in Embodiment 2 of the present application. This embodiment can describe an optional refined implementation method of how to use the first model and the second model in the intention recognition method proposed above to obtain a second processing result of the first question information, in which the first model is a large model, the second model is a deep learning model that recognizes the intention of the same question information faster than the large model, and the model parameter amount of the second model is much smaller than the model parameter amount of the first model, such as Figure 2 As shown, the intention recognition method proposed in this embodiment may include:

[0079] Step S21, obtaining first question information input by the user;

[0080] Regarding the input method and content of the first question information, reference may be made to but not limited to the description of the corresponding part of the above embodiment, such as the description of step S11, and this embodiment will not be described in detail here.

[0081] Step S22, inputting the first question information into the first model to decompose the complex question, and obtaining at least one second question information contained in the first question information; the second question information has a single intention or a simple intention, and the first question information has multiple intentions or a complex intention;

[0082] Based on the above analysis of the technical solution proposed in this application, in order to improve the accuracy of intent recognition of the second model, for the complex query (question information) input by the user, that is, the first question information with multiple intentions or complex intentions, a large model will be used to decompose the complex question, and the first question information will be processed into at least one second question information. Each second question information has a single intent or a simple intent, so as to avoid inputting complex question information into the second model, thereby ensuring the accuracy and reliability of intent recognition of the second model from the input source.

[0083] Among them, the intent of the question information Query usually refers to the actual needs or purposes expressed or potentially expressed by the user through the input question information. The intent can be divided into two categories: simple intent and complex intent based on factors such as the clarity and diversity of user needs and the complexity of the operations required to execute the query. Therefore, the simple intent of this application can refer to a type of intent with the characteristics of a single purpose (that is, the user's needs are very clear, usually involving only one action or information point), direct and clear (that is, the user's questions or requests can be understood without additional contextual information), easy to understand (that is, the model can easily identify the user's needs and directly provide answers to the questions or perform corresponding operations), no need for complex reasoning (that is, no need for in-depth analysis or reasoning of the user's intentions), and few parameters or conditions (that is, no need for multiple parameters or conditions to limit the answer to the question or the operation to be performed).

[0084] Relatively speaking, the complex intent of the present application may refer to a type of intent with multiple purposes (i.e., the user's needs involve multiple actions or information points, and multiple conditions may need to be met at the same time), implicit or ambiguous (i.e., the user's query may contain implicit intent, and the true intent / actual needs or purpose need to be clarified through context or additional information), requiring complex parsing (i.e., the model needs to perform complex semantic analysis and reasoning to understand the user's needs), involving multiple steps (i.e., it may be necessary to perform multiple steps or operations or candidate solutions to meet the user's query, such as filtering first and then sorting, etc.), with many parameters or conditions (i.e., the user's query may involve multiple parameters or conditions, and multiple factors need to be considered comprehensively) or involving tasks outside the intention system (i.e., the task may not be a task within the intention system, but requires a combination of multiple intentions in the intention system to achieve the task).

[0085] It can be seen that the demand for a simple intent Query is clear and direct, and it usually belongs to a single intent, but the single intent of a Query does not necessarily belong to a simple intent. A single intent with at least one characteristic as described in the complex intent above may also be a complex intent. Usually, a Query with multiple intents has a complex intent. Since the second model has limited understanding ability of question information with complex intent, it is difficult to fully and accurately identify the various intents or complex intents of the question information. This application proposes calling a large model to decompose the first question information with complex intent or multiple intents, to ensure that the intent of the second question information subsequently input to the second model is simple and single, so that the second model can accurately identify the intent corresponding to the input question information based on its own limited understanding ability.

[0086] In some embodiments, such as when a user continuously inputs multiple question messages to express multiple rounds of intent, the expression of question messages that are not input for the first time may be relatively concise, such as omitting or referring to objects / content that already exist in historical question messages that are input adjacently. The user's intent cannot be identified only through the question information input this time. For example, the user inputs "increase screen brightness" in the first round, and "higher" in the second round. The first question message of "higher" does not clearly express the user's actual needs or purposes, and the model cannot directly identify the intent from it. In response to this, an embodiment of the present application proposes to rewrite the first question message currently input by the user in combination with the context of the question message to obtain a simple and clear expression of intent. For example, "higher" is rewritten as "higher screen brightness" and the like.

[0087] Based on this, in the implementation process of step S22, after the first question information is input into the first model, the first model can obtain the historical question information input adjacent to the first question information, perform semantic analysis on the historical question information and the first question information, and obtain at least one second question information contained in the first question information. As in the above example, the first model can determine the intent of the first question information or the sentence contained in it that is unclear in intent based on the semantic analysis result, and accordingly change the first question information or the sentence contained in it into the second question information that is simple and clear in intent expression.

[0088] In some other embodiments, combined with the above description of the complex intent or multiple intents of the first question information, the large model can adopt a complex question splitting method to split a first question information into multiple second question information, ensuring that each second question information contains a simple intent or a single intent. That is, the first question information is input into the first model for intent recognition, and according to the multiple intents or complex intents of the first question information, the first question information is decomposed into multiple second question information.

[0089] In practical applications, the first question information with complex intent or multiple intents belongs to a complex query, which usually contains multiple parameters or conditions, connections, sub-questions, grouping or sorting and other intent expressions. The first question information with different intent expressions has different corresponding splitting methods, such as conditional decomposition method based on logical symbols, sub-question decomposition method or index-based decomposition method. For these splitting methods, the present application can pre-configure corresponding complex question decomposition prompt words. Based on this, in some embodiments, the first question information can also be subjected to semantic or structural analysis, and after determining the matching complex question decomposition prompt words, the first question information and the complex question decomposition prompt words are input into the large model, and according to the complex question decomposition prompt words, the large model is guided to decompose the first question information into multiple second question information.

[0090] Exemplarily, the first question information input by the user is: "Turn on Bluetooth and eye protection mode", which can be input into the large model for intent recognition. It is determined that the first question information contains two intentions, and it can be decomposed into two second question information expressing a single intention, such as "Turn on Bluetooth" and "Turn on eye protection mode". Of course, the first question information is analyzed to contain the logical symbol "and", and the conditional prompt word based on the logical symbol decomposition is input into the large model with the first question information, and the two clauses connected by "and" in the first question information are split into independent sentences with clear intentions according to the splitting method represented by the prompt word, which is to split each clause connected by the logical symbol in the question information into independent sentences with clear intentions. However, it is not limited to the various complex problem decomposition methods described in this application.

[0091] In some other embodiments, in combination with the above description of complex intents, when the intent of the first question information input by the user is vague or implies multiple intents / complex intents (i.e., contains multiple / complex potential intents), the big model can use its own knowledge to semantically understand the first question information during the process of complex problem decomposition of such first question information, analyze the various potential / implicit intentions for implementing the task corresponding to the first question information, such as the intent corresponding to at least one execution plan / strategy for implementing the task, and then obtain the second question information for each intent based on this.

[0092] For example, the first problem information input by the user is "eye discomfort". In order to solve the task of user's eye discomfort, the execution plan / strategy generated or selected by the big model may include lowering the device screen brightness, reducing blue light, adjusting the font size of the displayed content, adjusting the brightness of the ambient light, turning off the screen after a preset period of time and then restoring, or calling a specified service (such as an eye examination appointment service, etc.) and other task description information. For each execution plan / strategy, there is at least one intention. The big model decomposes the complex problem of "eye discomfort" and obtains the second problem information containing a single intent / simple intent based on at least one intention corresponding to at least one execution plan / strategy that matches or is selected, such as the task description information corresponding to the execution plan / strategy, or splits the task description information containing multiple intentions / complex intentions into the second problem information containing a single intent / simple intent, etc.

[0093] Step S23, inputting at least one second question information into the second model for intent recognition, and obtaining a second processing result corresponding to the at least one second question information; the second processing result includes at least one predicted intent corresponding to the corresponding second question information and a prediction confidence corresponding to each predicted intent;

[0094] Combined with the above description of the second model, after the large model processes the first question information into at least one second question information through a suitable information enhancement method, since each second question information has a single intention or a simple intention, it is input into the second model for intention prediction, and the predicted intentions corresponding to the second question information and their corresponding prediction confidences can be accurately obtained. The prediction confidence can be the probability that the predicted intention corresponding to the second question information is the true intention of the second question information. The larger the prediction confidence, the more likely the corresponding predicted intention is to belong to the true intention of the corresponding second question information. This application does not elaborate on the intention prediction steps performed by the second model on each second question information.

[0095] Among them, in the above intention prediction process, if the first question information contains multiple second question information, the multiple second question information can be input into the second model in sequence, and the second model performs intention prediction on each input second question information to obtain at least one predicted intention of the second question information and its corresponding prediction confidence. Alternatively, all the multiple second question information can be input into the second model, and the intention prediction proposed in this application can be performed to obtain the second processing result of each second question information.

[0096] Step S24, selecting k candidate intentions from the predicted intentions corresponding to each second question information based on the prediction confidence;

[0097] According to but not limited to the method described above, after obtaining the second processing results corresponding to each second question information, based on each prediction confidence contained in the second processing result, the corresponding predicted intentions contained in the second processing result can be selected to be more likely to be the expression of the first question information or the potential true intention, and each selected predicted intention can be recorded as a candidate intention. Therefore, the present application can select several predicted intentions with higher prediction confidence as candidate intentions according to the ranking of the prediction confidence of each predicted intention obtained by the second model, and the specific implementation method is not limited in the present application.

[0098] In a possible implementation of determining at least one candidate intention of the first question information contained in the second processing result, for each second question information having a second processing result, based on the prediction confidence of at least one predicted intention corresponding to the second question information, a candidate intention of the second question information can be selected from the at least one predicted intention corresponding to the second question information in the manner described above; if the first question information contains one second question information, the candidate intention selected for the second question information is all the candidate intentions corresponding to the first question information; if the first question information contains multiple second question information, the candidate intention selected for the second question information is part of all the candidate intentions corresponding to the first question information.

[0099] Therefore, in the case that the first question information contains multiple second question information, after selecting the candidate intent corresponding to each second question information according to the candidate intent determination method described above, the candidate intents corresponding to these multiple second question information constitute all the candidate intents corresponding to the first question information (that is, the k candidate intents mentioned above). The number k of candidate intents corresponding to the first question information determined by this method is an integer greater than 2, which includes at least one predicted intent from each second question information.

[0100] In another possible implementation, the present application may also sort the prediction confidences corresponding to all the predicted intentions corresponding to all the second question information, and then select k (here an integer greater than or equal to 1) predicted intentions with higher prediction confidences or prediction confidences reaching a certain threshold as candidate intentions. In this case, the at least one candidate intention corresponding to the first question information determined does not necessarily include the predicted intentions from each second question information. For each predicted intention corresponding to the second question information that is not included, it is usually impossible to be the true intention expressed or potentially expressed by the first question information, and can be directly eliminated to reduce the subsequent calculation amount.

[0101] Step S25, obtaining the intention description information of each candidate intention;

[0102] Step S26, input at least k candidate intentions and the intention description information of each candidate intention into the first model, process the k candidate intentions at least based on the intention description information, and determine the target intention corresponding to the first question information.

[0103] In an embodiment of the present application, according to but not limited to the method described above, after screening out at least one candidate intent corresponding to the first question information from the predicted intent corresponding to at least one second question information contained in the first question information predicted by the second model, in order to improve the accuracy of intent recognition, the large model can continue to be called to process all the screened candidate intents to determine the target intent corresponding to the first question information.

[0104] In a possible implementation, in order to enable the large model to better understand the definition description of intent, i.e., intent description information, the intent description information corresponding to different categories of intent can be configured by experts in the corresponding field or industry knowledge. For example, the intent description information of the intent "Get factory information of non-local devices" can be: Query factory information of non-local electronic devices; the intent description information of the intent "Get device brand" can be: What brand is my computer; the intent description information of the intent "Project to my PC" can be: Project to My PC' allows wireless projection of content from mobile devices to Windows PCs for easy sharing and display on larger screens (Project to My PC allows users to wirelessly project content on mobile devices to Windows PCs for easy sharing and display on larger screens).

[0105] In this way, the present application can correctly understand the k candidate intents through large model learning or based on the intent description information of each of the k selected candidate intents, and adopt a reasonable and correct processing method to process the k candidate intents to ensure that the target intent corresponding to the first question information is the true intent. For example, the first question information input by the user is "What model is the Dongfanghong 3?" According to the method described above, the corresponding candidate intent is determined to include the intent of "obtaining the brand of the device". Through its intent description information, it is determined whether the intent is the true intent of the first question information, or whether the match with the first question information is large enough to indicate that the intent is the target intent, etc.

[0106] It should be understood that the definition description of one or more of the screened candidate intents may be different in different fields / industries or enterprises, or it may be a custom intent within a specific field / enterprise, such as the meaning of "upper screen" in the intent of "adjusting the brightness of the upper screen" is unclear, resulting in the large model being unable to accurately determine whether this intent is a true intent. Therefore, in order to ensure that the large model correctly understands the intent of the candidate intent within the specific field / enterprise to which the first question information belongs / is involved, before the k candidate intents are input into the large model, the intent description information of each of the k candidate intents is obtained first. This application does not limit the source of the intent description information of each candidate intent (i.e., the definition description of the intent category to which each candidate intent belongs), which can be obtained from the intent system database, or queried from the knowledge base of a specific field / enterprise, etc.

[0107] Afterwards, the present application can at least combine k candidate intentions and the intention description information of each candidate intention to form an intention processing prompt information prompt, such as splicing at least k candidate intentions and the intention description information of each candidate intention to obtain a prompt, and then input the prompt into the big model, so that the big model can sort or select the k candidate intentions according to the intention description information in the prompt, and determine the target intention corresponding to the first question information. The target intention refers to the true intention of the first question information determined by the big model, which may be one or more candidate intentions that are most likely to be the true intention among the k candidate intentions, or an intention generated by these one or more candidate intentions (i.e., an extended intention), etc. The present application does not limit how the big model processes the k candidate intentions and determines the target intention corresponding to the first question information, and it is determined based on actual conditions.

[0108] In some other embodiments, the present application may also pre-fine-tune the large model so that the adjusted first model can understand each candidate intent, so that k candidate intents can be input into the adjusted first model, and the k candidate intents can be processed by the adjusted first model to determine the target intent corresponding to the first question information. Based on this, the present application can obtain fine-tuning data based on the k candidate intents, and use the fine-tuning data to fine-tune the parameters of the first model to obtain the adjusted first model that can understand the k candidate intents.

[0109] In one possible implementation, after obtaining the intent description information of each candidate intent, the present application can determine the intent description information as fine-tuning data, or use the intent description information to constitute fine-tuning data to achieve parameter fine-tuning of the large model, so that the adjusted first model can know the definition and boundaries of each candidate intent in advance. In this case, the k screened candidate intents can be input into the first model for processing to determine the target intent corresponding to the first question information.

[0110] In some other embodiments, in order to improve the reliability and accuracy of the first model in determining the target intent corresponding to the first question information, in addition to obtaining the intent description information of each candidate intent, historical interaction information (such as historical conversation information) that matches the first question information can also be obtained, so as to input k candidate intents, the intent description information of each candidate intent, and the historical interaction information into the first model, and the first model processes the k candidate intents based on the intent description information and the historical interaction information to determine the target intent corresponding to the first question information. Compared with the implementation method of step S26, this embodiment, in combination with the historical interaction information, can more accurately understand the real intent expressed or potentially expressed by the first question information, and after correctly understanding the k candidate intents based on the intent description information, can more reasonably process (such as sorting, selecting or expanding one or more processing methods) the k candidate intents, thereby improving the accuracy and reliability of the determined target intent.

[0111] In summary, the embodiment of the present application inputs the first question information into the large model for complex problem decomposition, ensuring that at least one second question information obtained is a question information containing a single intention or a simple intention, and then inputs it into the second model with a faster intention recognition speed than the large model for intention recognition. Through the limited language understanding ability of the second model, it can also accurately determine at least one predicted intention corresponding to each second question information and its corresponding prediction confidence, and then based on the prediction probability that the corresponding predicted intention represented by the prediction confidence is the true intention, select the k candidate intentions corresponding to the first question information from all predicted intentions. This process does not require starting the large model to occupy a large amount of computing and memory resources of the electronic device. Since the model parameter amount of the second model is much less than that of the large model, the resource consumption of the electronic device is greatly reduced, and the intention recognition processing speed is improved. Afterwards, at least k candidate intentions and their corresponding intention description information are input into the large model, so that the large model correctly understands the corresponding candidate intentions at least based on the intention description information, realizes the reasonable processing of the k candidate intentions, and ensures the accuracy and reliability of the target intention corresponding to the first question information determined thereby.

[0112] It can be seen that the intent recognition method proposed in this application greatly reduces the startup time of the large model and the power consumption of electronic devices compared to the implementation method of only calling the large model to identify the first question information to determine the target intent, making it better suitable for end-side devices with limited resources such as computing and memory. Compared with the implementation method of determining the target intent by fine-tuning the pre-trained model such as BERT, this application can accurately identify complex or ambiguous target intents and improve the accuracy of intent recognition. The length of the input question information processed by the large model is not limited, which improves the user interaction experience. It does not need to rely too much on labeled training data, reduces the cost of labeled training data, and improves the generalization and extensibility of the model.

[0113] Reference Figure 3 , is a flowchart of an intent recognition method proposed in Embodiment 3 of the present application. This embodiment can describe another possible detailed implementation method of how to determine the target intent corresponding to the first question information based on the second processing result in the intent recognition method proposed above. Based on this, Figure 3 As shown, the intention recognition method proposed in this embodiment may include:

[0114] Step S31, obtaining first question information input by the user;

[0115] Step S32, inputting the first question information into the first model to decompose the complex question, and obtaining at least one second question information contained in the first question information; the second question information has a single intention or a simple intention, and the first question information has multiple intentions or a complex intention;

[0116] Step S33, inputting at least one second question information into the second model for intent recognition, and obtaining a second processing result corresponding to the at least one second question information; the second processing result includes at least one predicted intent corresponding to the corresponding second question information and a prediction confidence corresponding to each predicted intent;

[0117] Regarding the implementation process of step S31-step S33, reference may be made to the description of the corresponding part of the context implementation example, such as the description of the corresponding part of step S21-step S23, which will not be elaborated in this embodiment.

[0118] Step S34, comparing the prediction confidence of the prediction intention corresponding to each second question information with the confidence threshold;

[0119] In the embodiment of the present application, since the second model predicts the intention of each second question information with a single intention or a simple intention, it is possible to obtain at least one predicted intention corresponding to the second question information and its corresponding prediction confidence. Since the prediction confidence represents the predicted probability that the corresponding predicted intention is the true intention (i.e., the target intention) of the second question information or the first question information, when the prediction confidence is high enough, the corresponding predicted intention can be considered to be the true intention. The intention prediction result of the second model is directly used, and there is no need to call the large model to determine the target intention corresponding to the first question information, thereby further reducing the startup time of the large model and reducing the power consumption of the electronic device.

[0120] Therefore, the present application can pre-configure the minimum confidence that the prediction confidence is high enough and the corresponding predicted intention is the true intention, which is recorded as the confidence threshold, such as 85% or 90%, etc. The present application does not limit the value of the confidence threshold, and its representation is consistent with the prediction confidence representation, including but not limited to the prediction probability or prediction score. Based on this, it can be determined whether the prediction confidence of each predicted intention corresponding to each second question information obtained by the second model is greater than or equal to the confidence threshold, thereby determining whether there is a true intention among all the predicted intentions corresponding to the second question information.

[0121] Step S35: if it is determined that the prediction confidence of at least one predicted intention corresponding to each second question information is greater than or equal to the confidence threshold, the at least one predicted intention is determined as the target intention corresponding to the first question information;

[0122] Following the above analysis, after comparing the prediction confidence of each predicted intention with the confidence threshold, if it is determined that the prediction confidence of at least one predicted intention corresponding to each second question information is greater than or equal to the confidence threshold, it means that the second model has identified at least one true intention for each second question information. Since the first question information is composed of the second question information, this means that the intention prediction result of the second model contains the true intention of the first question information. By comparing the prediction confidence with the confidence threshold, the target intention corresponding to the first question information, that is, the true intention, can be quickly and reliably determined. There is no need to start the large model to process the predicted intention or the k candidate intentions selected therefrom, which greatly reduces the startup time of the large model, reduces the time occupied by the computing and memory resources of the electronic device due to starting the large model, and reduces the power consumption of the electronic device.

[0123] Step S36: if it is determined that the prediction confidence of each predicted intention corresponding to each second question information is less than the confidence threshold, at least each predicted intention is input into the first model, and the first model processes each predicted intention based on at least the intention description information of each predicted intention to determine the target intention corresponding to the first question information;

[0124] After the comparison process in step S34, it is determined that the intention prediction of the second model is inaccurate, that is, the prediction confidence of each predicted intention corresponding to all the second question information is less than the confidence threshold, and the predicted intention cannot be directly used to determine the target intention corresponding to the first question information. At this time, in order to determine the true intention, it is still necessary to call the first model to determine the target intention corresponding to the first question information. In a possible implementation, the predicted intentions obtained by the second model from the input second question information can be directly input into the first model for processing, and the first model's own knowledge can be used to process the input predicted intentions to accurately determine the target intention corresponding to the first question information.

[0125] In another possible implementation, the present application can also obtain intention description information of each predicted intention, and at least use each predicted intention and its corresponding intention description information to form intention processing prompt information, so as to input the intention processing prompt information into the first model, and the first model processes each predicted intention obtained by the second model based on at least the intention description information of each predicted intention, and determines the target intention corresponding to the first question information.

[0126] Optionally, according to actual needs, combined with the description of the corresponding part of Example 2 above, the present application can also combine historical interaction information to form intention processing prompt information input into the first model, and the first model processes each predicted intention based on the intention description information of each predicted intention and the historical interaction information to determine the target intention. During this processing, the first model can also process each predicted intention in combination with the first question information that has been obtained. The present application does not limit how the first model processes the predicted intention and determines the implementation method of the target intention corresponding to the first question information. The content of the intention processing prompt information in different implementation methods can be adaptively adjusted, including but not limited to the content listed above.

[0127] It should be understood that after using the intention description information of each predicted intention to fine-tune the parameters of the large model in advance, when using the adjusted large model to implement the processing of each predicted intention, it is no longer necessary to input the intention description information into the first model, and the target intention corresponding to the first question information can be accurately determined. For the implementation process of fine-tuning the parameters of the large model, please refer to the corresponding part of the above embodiment, that is, the relevant description of the implementation steps of fine-tuning the parameters of the first model using the fine-tuning data obtained from the candidate intentions. This embodiment will not be described in detail here.

[0128] In some other embodiments proposed in the present application, combined with the method for determining the target intent corresponding to the first question information described in Example 2, after step S34, it is determined that the predicted intent in the second processing result of each second question information obtained by using the second model is not accurate, that is, when it is determined that the prediction confidence of at least one predicted intent corresponding to each second question information is greater than or equal to the confidence threshold, k candidate intents can still be selected from the obtained predicted intents based on the prediction confidence, and then at least the candidate intents are input into the first model (or the adjusted first model obtained according to the parameter fine-tuning method described above) for processing, and the first model processes the k candidate intents based on at least the intent description information of each of the k candidate intents (which can be input into the first model as a component of the intent processing prompt information, or can be pre-set in the first model, or the first model can pre-learn the intent description information of each type of intent so that each candidate intent can be correctly understood, etc., and the present application does not limit the source of the intent description information) to accurately determine the target intent corresponding to the first question information.

[0129] Optionally, in the process of constructing the intent processing prompt information, in addition to the candidate intents and their corresponding intent description information described above, historical interaction information matching the first question information can also be obtained as needed. Of course, the historical interaction information can also be retrieved by the large model after receiving the input information. Afterwards, based on the intent description information and historical interaction information of each candidate intent, the target intent corresponding to the first question information is determined.

[0130] Step S37, if it is determined that the prediction confidence of the predicted intention corresponding to at least one second question information among multiple second question information is less than the confidence threshold, at least the candidate intention among the predicted intentions corresponding to the at least one second question information is input into the first model, and the first model processes the candidate intention at least based on the intention description of the candidate intention, and determines the target intention corresponding to the second question information and the predicted intentions corresponding to other second question information whose prediction confidence is greater than or equal to the confidence threshold as the target intention corresponding to the first question information.

[0131] In an embodiment of the present application, after comparison in step S34, it is determined that the predicted confidence of the predicted intention corresponding to at least one of the multiple second question information is less than the confidence threshold, and the predicted confidence of at least one predicted intention corresponding to other second question information (i.e., at least one of the other second question information among the multiple second question information) is greater than or equal to the confidence threshold, indicating that the second model accurately predicts the intention of a part of the second question information constituting the first question information, and inaccurately predicts the intention of another part of the second question information. In order to more accurately determine the target intention corresponding to the first question information, the first model can still be called to process the predicted intention corresponding to this part of the second question information with inaccurate intention prediction, and accurately determine the target intention corresponding to this part of the second question information, and together with the predicted intention of another part of the second question information accurately predicted by the second model (whose prediction confidence is greater than or equal to the confidence threshold) constitute the target intention of the first question information.

[0132] It can be seen that the second question information contains multiple clauses, and the intention complexity of different clauses is different. The first model decomposes the complex question and obtains multiple second question information with different intention complexity and expression clarity, which leads to different intention prediction accuracy of the second model for different second question information. In this case, the present application can adopt the method described in step S37 to call the large model to accurately determine the target intention corresponding to the second question information whose intention cannot be accurately predicted by the second model, that is, the real intention corresponding to the second question information. Among them, in the process of calling the large model to determine the target intention corresponding to the second question information, the candidate intention in the predicted intention corresponding to this part of the second question information can be determined as in step S37, and then at least the candidate intention is input into the first model for processing; or at least the candidate intention and its corresponding intention description information, and even historical interaction information can be input into the first model for processing. The implementation process can refer to the description of the corresponding part of the above embodiment.

[0133] In another possible implementation, for step S37, in the process of calling the large model to determine the target intent corresponding to the second question information, the predicted intent corresponding to this part of the second question information can also be directly input into the first model for processing, or the predicted intent, as well as the intent description information and / or historical interaction information corresponding to the predicted intent can be input into the first model (these information can also be spliced ​​or other methods to form a fixed format of intent processing prompt information and then input) The first model processes the predicted intent corresponding to this part of the second question information based on the received intent description information and / or historical interaction information, and determines the target intent corresponding to the second question information. The implementation process is not described in detail in this application.

[0134] To summarize, a large model is used to decompose the complex problem of the first question information to obtain at least one second question information with a single intention or a simple intention. In order to shorten the response time of intent recognition, the second model will be used to perform intent recognition processing on each second question information, and at least one predicted intention corresponding to each second question information and its corresponding prediction confidence will be predicted. Afterwards, by comparing each prediction confidence with the confidence threshold, it is determined whether the intention prediction of the second model is accurate. If each second question information has an accurate predicted intention, it is directly determined as the target intention without calling the large model again; if the predicted intentions are not accurate or the intention predictions of some second question information are inaccurate, the large model is called to determine the target intention corresponding to the first question information. Compared with calling the large model to implement the entire intention recognition method, the startup time of the large model is shortened, the system delay is reduced, and the power consumption of the electronic device is reduced due to the reduced resource occupation time of the electronic device, thereby ensuring that the intention recognition method of the present application can be reliably applied to end-side devices with limited resources (such as user terminals).

[0135] In some other embodiments proposed in the present application, after the comparison in step 34, if it is determined that the prediction confidence of at least one predicted intention of at least one second question information is greater than or equal to the confidence threshold, the predicted intention can be determined as the target intention corresponding to the first question information, and there is no need to call the large model to process other predicted intentions or candidate intentions determined therefrom.

[0136] It should be noted that in the above embodiments, the big model processes the candidate intents / predicted intents to determine the target intent corresponding to the first question information. The processing method performed by the big model may include but is not limited to intent selection or sorting, etc., which can be determined according to actual conditions or according to pre-configured intent processing prompts. For example, the intent processing prompts for "greeting" type intents may include selecting this intent as the target intent when the user's Query (that is, the first question information input by the user) is in the form of a greeting or a happy wish; the intent processing prompts for "information request" type intents may include selecting this intent as the target intent when the user requests information or detailed information on a specific topic; the intent processing prompts for "problem solving" type intents may include selecting this intent as the target intent when users express the problems or challenges they face and seek help, etc. In this way, when the first question information is "Good morning! I am looking for best practices information for sustainable living", the large model performs the processing task of selecting the most appropriate intent from given intents (such as the predicted intent or candidate intent mentioned above) according to the user query according to the method described above. It can determine which of the three types of intents above is most suitable for the first question information input by the user, that is, determine the target intent corresponding to the first question information, and can also output the reason for selecting the target intent, etc.

[0137] Reference Figure 4 , is a flow chart of an intention recognition method proposed in the fourth embodiment of the present application. In the intention recognition method proposed above, this embodiment can also determine whether the question information input by the user is a complex question or a simple question. For simple questions, the second model is directly used to process them without starting the first model, thereby further improving the response speed of intention recognition. Based on this, Figure 4 As shown, the intention recognition method proposed in this embodiment may include:

[0138] Step S41, obtaining first question information input by the user;

[0139] Step S42, performing complexity evaluation on the first question information to determine whether the first question information belongs to a target category question;

[0140] In the embodiment of the present application, the complexity evaluation method such as semantic analysis and / or sentence structure analysis can be used to determine whether the first question information belongs to the first category of questions (i.e., complex questions with multiple intentions or complex intentions) or the second category of questions (i.e., simple questions with a single intention or simple intentions). It can be seen that the complexity of the first category of questions is higher than that of the second category of questions, and the present application does not limit the method for evaluating the complexity of the question.

[0141] In some embodiments, a question classification model can be pre-trained to determine the question category of the first question information, and a training question with a category label, i.e., a training data set, can be obtained and input into a classifier for training to obtain a question classification model that meets the training termination conditions (such as the problem classification accuracy or the prediction loss between the predicted problem category and the category label converges, or the number of training times reaches a specified number, etc.). The present application does not impose any restrictions on the training implementation method of the question classification model.

[0142] Among them, the category label for the training problem can be determined based on the intent recognition results of the corresponding training problem by the first model and the second model respectively. For example, the first model and the second model are used to identify the intent of the same training problem respectively. After the intent of the training problem is obtained, if the intent obtained by the first model and the second model is the same and both are the true intent of the training problem, it can be considered that the training problem belongs to a simple problem, and the category label of the second type of problem is configured for the training problem; if the intent obtained by the first model and the second model are different, and the intent obtained by the first model is the true intent of the training problem, but the intent obtained by the second model is not the true intent of the training problem, it is considered that the training problem belongs to a complex problem, and the category label of the first type of problem is configured for the training problem

[0143] It should be noted that the present application includes but is not limited to the above-mentioned category label determination method. The second model can also be used directly to identify the intent of the training data. Based on whether the obtained intent is accurate or whether the degree of match between the intent and the true intent reaches a matching threshold (the intent is considered to be the minimum match degree for the true intent, etc., and the present application does not impose any restrictions on its numerical value), it is determined whether the training problem is a complex problem or a simple intent, and the corresponding category label is obtained.

[0144] Based on this, the first question information input by the user can be input into the question classification model for complexity evaluation, and it is determined that the first question information belongs to the target category problem that matches the complexity evaluation result, such as a complex question or a simple question, etc. This application does not elaborate on the complexity evaluation implementation process of the question classification model for the input first question information. The complexity evaluation result can be a category label, which represents the target category problem; it can also be the predicted probability that the first question information belongs to a complex question, and the predicted probability that it belongs to a simple question, and the category problem with a larger predicted probability is determined as the target category problem, etc.

[0145] Step S43, if the target category problem is a first category problem, input the first problem information into the first model for information enhancement processing to obtain a first processing result; the first model is a large model;

[0146] Step S44, inputting the first processing result into the second model for intent recognition processing to obtain a second processing result; the second model is a deep learning model that recognizes intent for the same question information faster than the large model;

[0147] Step S45, determining the target intention corresponding to the first question information based on the second processing result;

[0148] It can be seen that the embodiment of the present application determines that the first question information input by the user belongs to a complex question, that is, a first category question, and directly assigns the first question information to the large model for processing. In this processing process, in order to solve the problem of long response time when only using the large model to identify the intent of the first question information, and long-term startup of the large model occupies the resources of the electronic device, resulting in excessive power consumption of the electronic device, the present application uses the large model to perform information enhancement processing on the first question information, and then transmits it to the second model for intent recognition, and then based on the second processing result obtained, quickly and accurately determines the target intention corresponding to the first question information, while ensuring the accuracy of intent recognition, ensuring the low latency and low power consumption of the intent recognition system. Regarding the implementation process of the scheme described in steps S43-step S45, please refer to the description of the corresponding parts of the above embodiments, and this embodiment will not be repeated here.

[0149] Step S46, if the target category question is a second category question, the first question information is input into the second model, the first question information is converted into a first question vector through the embedding network in the second model, the first question vector is subjected to intent recognition through the converter network of the second model, and the target intent corresponding to the first question information is determined.

[0150] Following the above analysis, when it is determined that the first question information input by the user belongs to a simple question, that is, a question of the second category, there is no need to call the large model. The second model is directly used to identify the intent of the first question information and determine the target intent corresponding to the first question information. For example, among the various predicted intents corresponding to the first question information predicted by the second model, the one with the highest prediction confidence or several higher predicted intents are selected to form the target intent.

[0151] Among them, in order to reduce the occupation of system resources, the second model can reuse the embedding network of the large model (that is, one or more embedding models in the large model are combined or weighted together), combined with at least one layer of the transformer network of the transformer architecture, to realize the intention recognition of the first question information. Since the embedding network and the transformer network of the second model can be decoupled, when there is an embedding model in the system, such as the embedding model of the large model, the embedding model can be directly reused to form the second model, reducing the overall resource occupation of the system. It should be noted that the network structure of the second model of the above-mentioned function of the present application, including but not limited to the content described in this embodiment, can be flexibly determined according to actual conditions.

[0152] In summary, if Figure 5 As shown in the flow chart, the present application comprehensively considers the balance between accuracy of intent recognition, response time and power consumption, first determines whether the first question information input by the user is a complex question or a simple question, assigns complex questions to the large model for processing, and assigns simple questions to the second model for processing. During the processing of complex questions, the second model is used to provide k candidate intentions for the first question information, ensuring that the candidate intentions contain all possible real intentions and improving the intent recall rate, which not only improves the accuracy of intent recognition for complex questions, but also ensures low latency and low power consumption of the system.

[0153] For the intent recognition method described in the above embodiments, after determining the target intent corresponding to the first question information input by the user, the corresponding task operation can be performed according to the intent category of at least one intent contained in the target intent. Therefore, the intent recognition method proposed in this application may also include but is not limited to the task operation implementation method described below:

[0154] When the target intent corresponding to the first question information includes the first intention to control the electronic device, that is, the intention of the device control type, in response to the first working state of the electronic device not matching the first intention, a device control page matching the first intention can be output, and in response to the control input operation on the device control page, the electronic device can be controlled to switch from the first working state to the second working state matching the first intention.

[0155] Among them, the state content of the first working state and the second working state can be determined according to the first intention content. Exemplarily, such as the first intention of turning on the eye protection mode of the electronic device, it can be determined whether the electronic device is in the eye protection mode. If it is currently in the non-eye protection mode, that is, the eye protection mode is not turned on, that is, the first working state of the electronic device does not match the first intention, and the eye protection mode setting interface can be output for the user to turn on the eye protection mode on the eye protection mode setting interface. In response to the control input operation, the electronic device is controlled to turn on the eye protection mode, that is, the electronic device is controlled to enter the second working state matching the first intention.

[0156] In another possible implementation, in response to the first working state of the electronic device not matching the first intention, a control button for the eye protection mode may be output, and in response to the triggering operation of the control button, the eye protection mode of the electronic device may be turned on; or, the electronic device may be directly controlled to switch from the first working state to the second working state matching the first intention, that is, the eye protection mode of the electronic device may be automatically turned on without user operation.

[0157] In response to the second working state of the electronic device matching the first intention, corresponding matching prompt information may also be output, such as the eye protection mode is turned on, etc. In this case, no processing may be performed to ensure that the electronic device is in the second working state matching the first intention.

[0158] It should be understood that for the first intention of other device control categories, the implementation method of the electronic device responding to the first intention is similar. The corresponding API can be called to achieve control of the electronic device itself or its interconnected devices (such as external imaging devices, speakers or other electronic devices such as displays). The control process is similar, and this application does not give detailed examples one by one.

[0159] If the target intent corresponding to the first question information includes the second intent in the question-answering task, such as content generation / search intent, or the intent implied by the task corresponding to the first question information, which is not directly expressed in the first question information, in response to the second intent, the first model can output the target content for answering the first question information based on the second intent.

[0160] In the above implementation process, a pre-configured task prompt word corresponding to the second intention can be obtained, and at least the task prompt word and the second intention constitute the question and answer prompt information input into the first model, and the first model is guided by the task prompt word based on the second intention to output the target content for answering the first question information. For example, to generate a summary of document A, the task prompt word can be a pre-configured rule on how to generate a summary. The generation implementation process corresponding to other categories of content is similar, and this application will not elaborate. Of course, the task prompt word and the second intention can also be directly input into the first model to output the target content. This application will not elaborate on how the large model performs the question and answer task based on the prompt project.

[0161] It should be understood that in the process of responding to the above-mentioned second intention, the user can also interact with the content output by the big model, perform question supplement operations on the first question information based on the prompt information of the question to be supplemented output by the big model, and input the question supplement information into the big model to obtain the target content that meets the user's expectations. The implementation process will not be described in detail.

[0162] If the target intent corresponding to the first question information includes a third intent to call a cloud service, a cloud service calling interface or a cloud service interaction interface of the called cloud service can be output in response to the third intent; the cloud service calling interface displays at least one calling method of the intended calling cloud service. For example, according to the intent recognition method described above, it is determined that the user inputs the first question information and needs to call the customer service of a certain product. The electronic device can output a customer service calling interface containing a calling method such as a customer service phone number or a customer service system access address link, or directly request to communicate with the product's manual customer service, or a customer interaction interface (such as a chat window, etc.).

[0163] If the target intent corresponding to the first question information includes the fourth intent of launching an application APP, in response to the fourth intent, the application interaction interface of the intent to launch the application is output. That is, directly opening the APP, such as opening a browser or a word document, etc. The fourth intent can be the intent directly expressed by the first question information, or it can be determined according to the above-mentioned intent recognition method that the execution of the task needs to launch the application APP to implement the case, the application APP needs to be launched, and the corresponding application interaction interface can be output. The implementation process is not described in detail in this application.

[0164] In the actual application of the present application, after identifying the intent of the first question information and determining that its target intent includes one or more intentions listed above, the electronic device can respond to each intention in accordance with but not limited to the processing method of the corresponding intention to meet the task processing requirements of the user inputting the first question information.

[0165] The above introduces an intention recognition method provided by an embodiment of the present application, and the following will introduce a device for executing the above intention recognition method.

[0166] Reference Figure 6 , is a schematic diagram of the structure of an intention recognition device provided in an embodiment of the present application. Figure 6 As shown, the intention recognition device may include:

[0167] The question information obtaining module 61 is used to obtain the first question information input by the user;

[0168] A first processing module 62, configured to input the first question information into a first model for information enhancement processing to obtain a first processing result;

[0169] A second processing module 63 is used to input the first processing result into a second model for intent recognition processing to obtain a second processing result; the second processing result includes at least one candidate intent corresponding to the first question information;

[0170] A target intention determination module 64, configured to determine a target intention corresponding to the first question information based on the second processing result;

[0171] Among them, the first model is a large model, and the second model is a deep learning model that recognizes the intent of the same question information faster than the large model.

[0172] In some embodiments, the first processing module 62 may include:

[0173] A complex problem decomposition unit, used for inputting the first problem information into a first model to decompose the complex problem, and obtaining at least one second problem information contained in the first problem information; the second problem information has a single intention or a simple intention, and the first problem information has multiple intentions or a complex intention;

[0174] Based on this, the second processing module 63 may include

[0175] An intention recognition processing unit is used to input at least one of the second question information into a second model for intention recognition processing to obtain a second processing result corresponding to at least one of the second question information; the second processing result includes at least one predicted intention corresponding to the corresponding second question information and the prediction confidence corresponding to each of the predicted intentions.

[0176] In one possible implementation, the target intention determination module 64 may include:

[0177] a candidate intention selection unit, configured to select k candidate intentions from the predicted intentions corresponding to each of the second question information based on the prediction confidence;

[0178] An intention description information obtaining unit, used to obtain the intention description information of each candidate intention;

[0179] The first input unit is used to input at least the k candidate intents and the intent description information of each candidate intent into the first model, so that the first model processes the k candidate intents at least based on the intent description information to determine the target intent corresponding to the first question information.

[0180] In a possible implementation, the target intention determination module 64 may include at least one of the following determination units:

[0181] A first determination unit, configured to determine at least one predicted intention as a target intention corresponding to the first question information when it is determined that the prediction confidence of at least one predicted intention corresponding to each of the second question information is greater than or equal to a confidence threshold;

[0182] a second determination unit, configured to, when it is determined that the prediction confidences of the prediction intentions corresponding to the respective prediction intentions of the respective second question information are all less than the confidence threshold, at least input each prediction intention into the first model, wherein the first model processes each prediction intention based on at least the intention description information of each prediction intention, and determines the target intention corresponding to the first question information;

[0183] A third determination unit is used to determine candidate intentions among the predicted intentions when it is determined that the prediction confidences of the predicted intentions corresponding to the respective second question information are all less than the confidence threshold, at least input the candidate intentions into the first model, and the first model processes the candidate intentions based on at least the intention description information of the candidate intentions to determine the target intention corresponding to the first question information;

[0184] A fourth determination unit is used to determine a candidate intention among the predicted intentions corresponding to at least one of the multiple second question information, when the predicted confidence of the predicted intention corresponding to at least one of the second question information is less than the confidence threshold, and at least input the candidate intention into the first model, and the first model processes the candidate intention based on at least the intention description information of the candidate intention to obtain the target intention corresponding to the second question information; wherein the target intention corresponding to the second question information and the predicted intentions corresponding to other second question information whose prediction confidence is greater than or equal to the confidence threshold constitute the target intention corresponding to the first question information.

[0185] In some further embodiments, the target intention determination module 64 may further include:

[0186] A fine-tuning data obtaining unit, configured to obtain fine-tuning data based on the candidate intent or the predicted intent;

[0187] A parameter fine-tuning unit is used to use the fine-tuning data to adjust the parameters of the first model to obtain an adjusted first model that can understand the candidate intent or the predicted intent, so as to process the candidate intent or the predicted intent through the adjusted first model to obtain the target intent.

[0188] In some embodiments, the complex problem decomposition unit may include at least one of the following units:

[0189] a semantic analysis unit, configured to input the first question information into a first model, perform semantic analysis on the first question information and historical question information input adjacent to the first question information, and obtain at least one second question information included in the first question information;

[0190] A decomposition unit is used to input the first question information into a first model for intent recognition, and decompose the first question information into multiple second question information according to multiple intents or complex intents of the first question information.

[0191] In summary, the intention recognition device may further include:

[0192] A complexity evaluation module is used to perform a complexity evaluation on the first question information to determine whether the first question information belongs to a target category question; if the target category question is a first category question, trigger the first processing module 62 to input the first question information into a first model for information enhancement processing;

[0193] an intention recognition module, configured to, if the target category question is a second category question, input the first question information into the second model, convert the first question information into a first question vector through an embedding network in the second model, perform intention recognition on the first question vector through a converter network of the second model, and determine a target intention corresponding to the first question information;

[0194] Among them, the complexity of the first category of questions is higher than that of the second category of questions. The first category of questions can be called complex questions, which have multiple intentions or complex questions; the second category of questions can be called simple questions, which have a single or simple answer intention.

[0195] In a possible implementation, the complexity evaluation module may include:

[0196] A complexity evaluation unit, configured to input the first question information into a question classification model for complexity evaluation, and determine whether the first question information belongs to a target category question that matches a complexity evaluation result;

[0197] The question classification model is obtained by training a classifier using training questions with category labels, and the category labels are determined based on the intention recognition results of the first model and the second model for the corresponding training questions.

[0198] In summary, the intention recognition device may further include at least one of the following response modules:

[0199] a first response module, configured to, if the target intention includes a first intention to control the electronic device, output a device control page matching the first intention in response to a first working state of the electronic device not matching the first intention, and control the electronic device to switch from the first working state to a second working state matching the first intention in response to a control input operation on the device control page;

[0200] a second response module, configured to output corresponding matching prompt information in response to a second working state of the electronic device matching the first intention if the target intention includes a first intention to control the electronic device;

[0201] a third response module, configured to, if the target intent includes a second intent in the question-answering task, cause the first model to output target content for answering the first question information based on the second intent;

[0202] A fourth response module is used for outputting a cloud service calling interface or a cloud service interaction interface of the called cloud service in response to the third intent if the target intent includes a third intent to call a cloud service; the cloud service calling interface displays at least one calling method of the intended calling cloud service;

[0203] The fifth response module is used to output an application interaction interface for the intention to launch the application in response to the fourth intent if the target intent includes the fourth intent for launching the application.

[0204] An embodiment of the present application also provides a computer program product, including computer-readable instructions. When the computer-readable instructions are executed on an electronic device, the electronic device implements any one of the intent recognition methods provided in the embodiments of the present application.

[0205] Wherein, when loading and executing computer-readable instructions on a computer, all or part of the process or function described in the embodiment of the present application is generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer-readable instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer-readable instructions may be transmitted from a website site, a computer, a training device, or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, training device, or data center, to meet the data / instruction transmission requirements in the execution process of the three-dimensional scene rendering processing method proposed in the embodiment of the present application.

[0206] A computer-readable storage medium is also provided in an embodiment of the present application. The storage medium carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any one of the intent recognition methods provided in the embodiment of the present application.

[0207] The computer-readable storage medium may be any available medium that can be stored by a computer or a data storage device such as a training device or a data center that includes one or more available media. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid state drive (SSD)).

[0208] Reference Figure 7 , is a hardware structure diagram of an electronic device applicable to the intent recognition method proposed in this application, the electronic device may include but is not limited to mobile phones, tablet computers, wearable devices, vehicle-mounted devices, augmented reality (AR) / virtual reality (VR) devices, laptops, netbooks, robots or business terminals, etc., and a server may also be configured to implement the intent recognition method proposed in this application. Figure 7 As shown, the electronic device may include but is not limited to: at least one communication element 71, at least one memory 72 and at least one processor 73, wherein:

[0209] At least one communication element 71, at least one memory 72 and at least one processor 73 can communicate with each other via a bus. The bus can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 7 Only one bidirectional line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0210] The communication element 71 can be used to obtain the first question information input by the user, realize communication between internal components of the electronic device, and communication between the electronic device and the server, so as to transmit the data or instructions required in the execution process of the intention recognition method, such as when calling the big model to execute the corresponding steps, sending the corresponding instructions and data (such as the first question information, candidate intentions / predicted intentions, and as needed, it can also include pre-configured task processing prompt words that match the determined target intention, etc.) to the server where the big model is deployed, and receiving the server's response based on the big model feedback, such as receiving the target content sent by the server for replying to the first question information output by the big model in response to the target intention, etc.

[0211] It can be seen that the communication element 71 may include a communication element corresponding to wireless communication methods such as wifi, Bluetooth, 5G / 6G and other mobile communications, so that the electronic device can realize data transmission with other devices through the communication element, and may also include one or more interfaces supporting wired communication methods, such as general-purpose input / output (GPIO) interface, USB interface, universal asynchronous receiver / transmitter (UART) interface, etc., to realize data transmission between various components inside the electronic device. This application does not limit the composition structure of the communication element 71 to realize this function and its corresponding communication transmission mechanism, which can be determined according to the circumstances.

[0212] The memory 72 can be used to store computer program instructions for implementing the intention recognition method proposed in the embodiment of the present application; the processor 73 can load and execute the computer program instructions stored in the memory 72 to implement the various steps of the intention recognition method proposed in the embodiment of the present application. The implementation process can refer to the description of the corresponding part of the method embodiment above.

[0213] In the embodiment of the present application, the memory 72 may include the storage media listed above. The processor 71 may include any one or more of a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor (MP), a digital signal processor (DSP), an application specific integrated circuit (ASIC), or a field-programmable gate array (FPGA).

[0214] It should be understood that Figure 7 The structure of the electronic device shown does not constitute a limitation on the electronic device in the embodiment of the present application. In practical applications, the electronic device may include Figure 7 More or fewer components as shown, or a combination of certain components, such as a microphone (which can collect voice signals input by the user), a speaker, a display, various sensors, an antenna, a power module, a radio frequency component, an external port, and other input / output components, etc., can be determined according to the processing function requirements, and this application does not give detailed examples one by one.

[0215] It should also be noted that the device embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed over multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, in the drawings of the device embodiments provided by the present application, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines.

[0216] Through the description of the above implementation mode, the technicians in the field can clearly understand that the present application can be implemented by means of software plus necessary general hardware, and of course, it can also be implemented by special hardware including special integrated circuits, special CPUs, special memories, special components, etc. In general, all functions completed by computer programs can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be various, such as analog circuits, digital circuits or special circuits. However, for the present application, software program implementation is a better implementation mode in more cases. Based on such an understanding, the technical solution of the present application is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a readable storage medium, such as a computer floppy disk, a U disk, a mobile hard disk, a ROM, a RAM, a disk or an optical disk, etc., including a number of instructions to enable a computer device (which can be a personal computer, a training device, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0217] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments can be implemented in the form of a computer program product. In this specification, each embodiment is described in a progressive or parallel manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the embodiments can be referred to each other. For the device and electronic device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

Claims

1. A method for identifying intent, the method comprising: Obtain the first question information input by the user; Inputting the first question information into a first model for information enhancement processing to obtain a first processing result; Inputting the first processing result into a second model for intention recognition processing to obtain a second processing result; The second processing result includes at least one candidate intent corresponding to the first question information; Based on the second processing result, determining a target intention corresponding to the first question information; Among them, the first model is a large model, and the second model is a deep learning model that recognizes the intent of the same question information faster than the large model.

2. According to the method of claim 1, the step of inputting the first question information into a first model for information enhancement processing to obtain a first processing result, and inputting the first processing result into a second model for intent recognition processing to obtain a second processing result comprises: Inputting the first question information into a first model to decompose the complex question, and obtaining at least one second question information contained in the first question information; The second question information has a single intention or a simple intention, and the first question information has multiple intentions or a complex intention; Inputting at least one of the second question information into a second model for intention recognition processing, and obtaining a second processing result corresponding to at least one of the second question information; The second processing result includes at least one predicted intention corresponding to the second question information and the prediction confidence corresponding to each predicted intention.

3. The method according to claim 2, wherein determining the target intention corresponding to the first question information based on the second processing result comprises: Based on the prediction confidence, select k candidate intentions from the predicted intentions corresponding to each of the second question information; Obtaining intent description information of each of the candidate intents; At least the k candidate intents and the intent description information of each candidate intent are input into the first model, so that the first model processes the k candidate intents at least based on the intent description information to determine the target intent corresponding to the first question information.

4. The method according to claim 2, wherein determining the target intention corresponding to the first question information based on the second processing result comprises at least one of the following: If it is determined that the prediction confidence of at least one of the predicted intentions corresponding to each of the second question information is greater than or equal to the confidence threshold, the at least one predicted intention is determined as the target intention corresponding to the first question information; If it is determined that the prediction confidences of the predicted intentions corresponding to the second question information are all less than the confidence threshold, at least each predicted intention is input into the first model, and the first model processes each predicted intention based on at least the intention description information of each predicted intention to determine the target intention corresponding to the first question information; If it is determined that the prediction confidences of the predicted intentions corresponding to the respective second question information are all less than the confidence threshold, determine candidate intentions among the predicted intentions, at least input the candidate intentions into the first model, and the first model processes the candidate intentions based on at least the intention description information of the candidate intentions to determine the target intention corresponding to the first question information; If the prediction confidence of the predicted intention corresponding to at least one of the multiple second question information is less than the confidence threshold, determine the candidate intention among the predicted intentions corresponding to the at least one second question information, at least input the candidate intention into the first model, and the first model processes the candidate intention at least based on the intention description information of the candidate intention to obtain the target intention corresponding to the second question information; wherein, the target intention corresponding to the second question information and the predicted intentions corresponding to other second question information whose prediction confidence is greater than or equal to the confidence threshold constitute the target intention corresponding to the first question information.

5. The method according to claim 3 or 4, wherein determining the target intention corresponding to the first question information based on the second processing result further comprises: Based on the candidate intent or the predicted intent, obtaining fine-tuning data; The fine-tuning data is used to adjust the parameters of the first model to obtain an adjusted first model that can understand the candidate intent or the predicted intent, so as to process the candidate intent or the predicted intent through the adjusted first model to obtain the target intent.

6. The method according to claim 2, wherein the step of inputting the first problem information into the first model to decompose the complex problem and obtaining at least one second problem information contained in the first problem information comprises at least one of the following: Input the first question information into a first model, perform semantic analysis on the first question information and historical question information input adjacent to the first question information, and obtain at least one second question information included in the first question information; The first question information is input into a first model for intent recognition, and the first question information is decomposed into multiple second question information according to multiple intents or complex intents of the first question information.

7. The method according to any one of claims 1 to 4, further comprising: Performing a complexity evaluation on the first question information to determine that the first question information belongs to a target category question; If the target category problem is a first category problem, executing the step of inputting the first problem information into the first model for information enhancement processing; If the target category question is a second category question, input the first question information into the second model, convert the first question information into a first question vector through the embedding network in the second model, perform intent recognition on the first question vector through the converter network of the second model, and determine the target intent corresponding to the first question information; The complexity of the first category of problems is higher than the complexity of the second category of problems.

8. According to the method of claim 7, the step of performing complexity evaluation on the first question information to determine whether the first question information belongs to a target category question comprises: Inputting the first question information into a question classification model for complexity evaluation, and determining that the first question information belongs to a target category question that matches a complexity evaluation result; The question classification model is obtained by training a classifier using training questions with category labels, and the category labels are determined based on the intention recognition results of the first model and the second model for the corresponding training questions.

9. The method according to claim 1, further comprising at least one of the following: If the target intention includes a first intention to control the electronic device, in response to a first working state of the electronic device not matching the first intention, outputting a device control page matching the first intention, and in response to a control input operation on the device control page, controlling the electronic device to switch from the first working state to a second working state matching the first intention; If the target intention includes a first intention to control the electronic device, in response to the second working state of the electronic device matching the first intention, outputting corresponding matching prompt information; If the target intent includes a second intent in the question-answering task, the first model outputs target content for answering the first question information based on the second intent; If the target intent includes a third intent to call a cloud service, in response to the third intent, output a cloud service calling interface or a cloud service interaction interface of the called cloud service; The cloud service calling interface displays at least one calling method of intending to call a cloud service; If the target intent includes a fourth intent to launch an application, in response to the fourth intent, an application interaction interface intended to launch the application is output.

10. An intention recognition device, the device comprising: A question information obtaining module, used to obtain first question information input by a user; A first processing module, used for inputting the first question information into a first model for information enhancement processing to obtain a first processing result; A second processing module, used for inputting the first processing result into a second model for intention recognition processing to obtain a second processing result; The second processing result includes at least one candidate intent corresponding to the first question information; a target intention determination module, configured to determine a target intention corresponding to the first question information based on the second processing result; Among them, the first model is a large model, and the second model is a deep learning model that recognizes the intent of the same question information faster than the large model.

Images