An intent recognition method, system and related apparatus

Abstract

The application discloses an intention recognition method, system and related device, the method comprises the following steps: obtaining the to-be-recognized information of a target object; obtaining a plurality of reference question templates; obtaining a candidate question template from all the reference question templates based on the content correlation between the to-be-recognized information and the reference question templates; obtaining a plurality of reference intention information; and obtaining candidate intention information from all the reference intention information based on the semantic correlation between the plurality of intention sub-information corresponding to the reference intention information and the to-be-recognized information; wherein the plurality of intention sub-information of the reference intention information comprises intention description, example text and intention scene description; and obtaining the target intention corresponding to the to-be-recognized information based on the candidate question template and the candidate intention information. Through the above manner, the application can improve the accuracy of intention recognition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing, in particular to an intent recognition method, system and related device. BACKGROUND

[0002] Intent recognition is one of the core tasks in the field of natural language processing, aiming to accurately capture the deep purpose or behavior orientation behind the input (such as text or voice) of a target object. The current traditional intent recognition method relies on matching the input of the target object with the intent information in a single intent library, and recalling the intent information with the highest matching degree. However, due to the fixed form and limited expression ability of the intent information in the single intent library, there is a certain limitation in recognition accuracy.

[0003] Therefore, how to propose an intent recognition method with high accuracy has become a problem to be solved. SUMMARY

[0004] The technical problem solved by the present application is to provide an intent recognition method, system and related device, which can improve the accuracy of intent recognition.

[0005] To solve the above technical problem, the first aspect of the present application provides an intent recognition method, comprising: obtaining to-be-recognized information of a target object; obtaining a plurality of reference question templates, and obtaining a candidate question template from all the reference question templates based on the content correlation between the to-be-recognized information and the reference question templates; and obtaining a plurality of reference intent information, and obtaining a candidate intent information from all the reference intent information based on the semantic correlation between the plurality of intent sub-information corresponding to the reference intent information and the to-be-recognized information; wherein the plurality of intent sub-information of the reference intent information includes intent description, example text and intent scene description; and obtaining a target intent corresponding to the to-be-recognized information based on the candidate question template and the candidate intent information.

[0006] To solve the above technical problem, the second aspect of the present application provides an intent recognition system, comprising: an acquisition module configured to obtain to-be-recognized information of a target object; a first processing module configured to obtain a plurality of reference question templates, and obtain a candidate question template from all the reference question templates based on the content correlation between the to-be-recognized information and the reference question templates; a second processing module configured to obtain a plurality of reference intent information, and obtain a candidate intent information from all the reference intent information based on the semantic correlation between the plurality of intent sub-information corresponding to the reference intent information and the to-be-recognized information; wherein the plurality of intent sub-information of the reference intent information includes intent description, example text and intent scene description; and an identification module configured to obtain a target intent corresponding to the to-be-recognized information based on the candidate question template and the candidate intent information.

[0007] To solve the above technical problems, the third aspect of the present application provides an electronic device, comprising a memory and a processor coupled with each other, the memory stores program instructions, and the processor is configured to execute the program instructions to implement the method mentioned in the above technical solution.

[0008] To solve the above technical problems, the fourth aspect of the present application provides a computer readable storage medium, which stores program instructions, and the program instructions are executed by a processor to implement the method mentioned in the above technical solution.

[0009] The beneficial effects of the present application are: different from the prior art, the intention recognition method proposed in the present application, after obtaining the to-be-recognized information of the target object, determines the candidate question template from the multiple reference question templates according to the relevance of the content, and determines the candidate intention information from the multiple reference intention information according to the relevance of the semantics between each intention sub-information corresponding to each reference intention information and the to-be-recognized information. In combination with the above candidate question template and candidate intention information, the target intention corresponding to the to-be-recognized information is obtained, that is, the target intention is obtained according to different types of information obtained in different ways, which helps to improve the accuracy of intention recognition. BRIEF DESCRIPTION OF DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor. Among them:

[0011] Figure 1 is a flowchart of an embodiment of the intention recognition method of the present application;

[0012] Figure 2 is Figure 1 corresponding to another embodiment of step S101 in

[0013] Figure 3 is Figure 1 corresponding to another embodiment of step S102 in

[0014] Figure 4 is Figure 3 corresponding to another embodiment of step S302 in

[0015] Figure 5 is Figure 1 corresponding to another embodiment of step S103 in

[0016] Figure 6 is Figure 5 corresponding to another embodiment of the method of identifying intention in step S502 of the flowchart;

[0017] Figure 7 is Figure 1 corresponding to another embodiment of the method of identifying intention in step S104 of the flowchart;

[0018] Figure 8 is Figure 7 corresponding to another embodiment of the method of identifying intention in step S703 of the flowchart;

[0019] Figure 9 is a structural diagram of an embodiment of the system for identifying intention of the present application;

[0020] Figure 10 is a structural diagram of an embodiment of the electronic device of the present application;

[0021] Figure 11 is a structural diagram of an embodiment of the computer readable storage medium of the present application. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application, and the adaptive combination between different embodiments can be made. Based on the embodiments in the present application, all the other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0023] The terms "system" and "network" are often used interchangeably in this document. The term "and / or" in this document is only used to describe the associated relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which means that there are three cases of A alone, A and B together, and B alone. In addition, the character " / " in this document generally represents an "or" relationship between the front and rear associated objects. In addition, "multiple" in this document means two or more than two.

[0024] The intention recognition method proposed in the present application relies on an application on a smart terminal or a smart terminal integrated with at least an intention recognition function, and the corresponding execution subject is a processing unit capable of data processing. The smart terminal can be a smart office book, a mobile phone, a tablet computer, a personal computer or a wearable smart device, etc.

[0025] Please refer to Figure 1 , Figure 1 is a flowchart of an embodiment of the method of identifying intention of the present application. Specifically, the implementation process of the method includes:

[0026] S101: Obtain to-be-recognized information of a target object.

[0027] In an embodiment, to-be-recognized information of a target object for performing intent recognition is obtained.

[0028] In some implementation scenarios, text input by a target object on a smart terminal is obtained, and the text is taken as to-be-recognized information, so that the text input by the target object is subjected to intent recognition through subsequent steps.

[0029] In some implementation scenarios, audio input by a target object on a smart terminal is obtained, and the input audio is subjected to character recognition to generate corresponding transcribed text. The obtained transcribed text is taken as to-be-recognized information.

[0030] In an embodiment, in response to text input by a target object on a smart terminal, the input text is processed to remove redundant elements in the text, and the processed text is taken as to-be-recognized information. Alternatively, in response to audio input by a target object on a smart terminal, the input audio is subjected to noise reduction processing, the audio subjected to noise reduction processing is subjected to character recognition, and the transcribed text recognized is taken as to-be-recognized information.

[0031] S102: Obtain a plurality of reference question templates, and obtain a candidate question template from all reference question templates based on content correlation between to-be-recognized information and the reference question templates.

[0032] In an embodiment, natural language questions of different users are taken as samples, and reference question templates are obtained by processing the above samples to remove ambiguities. It is easy to understand that each reference question template is a clear, accurate, structured, and easy-to-understand natural language for expressing an intent, for example, “How do I download my monthly service invoice?”.

[0033] Further, content correlation between to-be-recognized information and each reference question template is obtained according to content corresponding to the to-be-recognized information and content corresponding to each reference question template. The candidate question template is obtained from all reference question templates according to the above content correlation.

[0034] In some implementation scenarios, an FAQ database including a plurality of reference question templates is constructed in advance. Content correlation between to-be-recognized information and each reference question template in the FAQ database is determined, and a reference question template corresponding to maximum content correlation is taken as a candidate question template.

[0035] In an embodiment, a target scenario to which an intent recognition method is applied is determined, and an FAQ database is constructed according to domain knowledge of the target scenario, so that the FAQ database includes a plurality of reference question templates related to the target scenario.

[0036] S103: Obtain a plurality of reference intent information, and obtain candidate intent information from all reference intent information based on semantic association between a plurality of intent sub-information corresponding to the reference intent information and the to-be-identified information; wherein the plurality of intent sub-information of the reference intent information includes intent description, example text and intent scene description.

[0037] In an embodiment, an intent information library including a plurality of reference intent information is obtained, and each reference intent information includes a plurality of intent sub-information. For each reference intent information, the semantics of each intent sub-information corresponding thereto and the semantics of the to-be-identified information are obtained, and the semantic association between each intent sub-information and the to-be-identified information is determined. According to all semantic associations corresponding to the above-mentioned reference intent information, candidate intent information is obtained from all reference intent information.

[0038] Specifically, a plurality of intents are determined in advance, and a plurality of description texts matching each intent are obtained, the description texts are taken as intent sub-information, and a plurality of intent sub-information matching the same intent are constructed as reference intent information. According to the semantics of each intent sub-information corresponding to the to-be-identified information and the reference intent information, the similarity between the two is obtained, and the reference intent information corresponding to the maximum similarity is taken as the candidate intent information. Wherein, the plurality of intent sub-information matching the above-mentioned reference intent information includes example text and intent scene description in addition to intent description.

[0039] In some implementation scenarios, the above-mentioned description text is obtained by a relevant technical person according to the predetermined intent; for example, for the same intent, a plurality of different description texts obtained by different technical persons are obtained. Alternatively, in order to improve the efficiency of obtaining the reference intent information, the above-mentioned intent sub-information can also be generated by using a large language model to describe the predetermined intent. Wherein, the large language model can include but is not limited to Deep Neural Networks (DNNs), Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) and generative pre-training Transformer model, etc., and the specific structure and specific deployment of the large language model are not limited here.

[0040] In an embodiment, a target scenario to which the intent recognition method is applied is determined, an intent related to the target scenario is determined, and an intent information library that matches the target scenario and includes multiple reference intent information is constructed according to the determined intent, so that the intent information library has a higher relevance to the to-be-recognized information. The determination of each reference intent information in the intent information library can refer to the corresponding embodiment described above.

[0041] Further, candidate intent information is obtained from the intent information library according to the semantics of the to-be-recognized information and the semantics of each intent information corresponding to the reference intent information.

[0042] It should be noted that the implementation order of the step S102 and the step S103 can also be other, for example, the step S102 and the step S103 can be executed simultaneously; or, the step S103 can be executed first, and then the step S102 is executed.

[0043] S104: obtaining a target intent corresponding to the to-be-recognized information based on the candidate question template and the candidate intent information.

[0044] In an embodiment, the to-be-recognized information is subjected to final intent decision in combination with the obtained candidate question template and the candidate intent information, and a target intent corresponding to the to-be-recognized information is determined.

[0045] In some implementation scenarios, a decision large model is obtained, the to-be-recognized information, the candidate question template and the candidate intent information are input into the decision large model, the decision large model is prompted to analyze the input content and perform intent recognition, and finally a target intent generated by the decision large model is obtained.

[0046] Specifically, a target prompt template generated in advance is obtained, for example, “according to the [MASK] input by a target object, the [candidate question template] and the [candidate intent information], the intent of the target object is recognized, and a target intent is generated”. The “[MASK]” in the target prompt template is replaced by the to-be-recognized information, and the candidate question template and the candidate intent information are filled into the corresponding positions in the target prompt template to obtain target prompt information. The target prompt information is input into the decision large model, and a target intent generated by the decision large model is obtained.

[0047] In a specific application scenario, the large language model can include, but is not limited to, deep neural networks (DNNs), convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory networks (LSTM), and generative pre-training Transformer models, and the like. The specific construction and specific deployment of the large language model are not limited here.

[0048] The intent recognition method provided in the present application determines a candidate question template from a plurality of reference question templates according to the relevance of the content after obtaining the to-be-recognized information of the target object, and determines candidate intent information from a plurality of reference intent information according to the relevance of the semantics between each intent sub-information corresponding to each reference intent information and the to-be-recognized information. The target intent corresponding to the to-be-recognized information is obtained in combination with the candidate question template and the candidate intent information, that is, the target intent is obtained according to different types of information obtained in different ways, which helps to improve the accuracy of intent recognition.

[0049] Please refer to Figure 2 , Figure 2 is Figure 1 The flowchart of another embodiment corresponding to step S101 is shown in FIG. 1B. Specifically, the implementation process of step S101 includes:

[0050] S201: Obtain initial dialogue information input by a target object.

[0051] In an embodiment, the initial dialogue information input by the target object on the intelligent terminal is obtained. The initial dialogue information can be text input directly by the target object on the intelligent terminal, or the initial dialogue information can also be text obtained by audio recognition on audio input by the target object on the intelligent terminal.

[0052] S202: Input the initial dialogue information to the reasoning large model to obtain intent reasoning information generated by the reasoning large model and matched with the initial dialogue information.

[0053] In an embodiment, the obtained initial dialogue information is input to the reasoning large model, and the reasoning large model is prompted to perform intent reasoning on the initial dialogue information until the intent reasoning information generated by the reasoning large model and matched with the initial dialogue information is obtained.

[0054] In some implementation scenarios, the reasoning large model is a large language model with good data analysis capability. The corresponding reasoning prompt information is generated according to the initial dialogue information, and the reasoning prompt information is input into the reasoning large model, so that the reasoning large model generates matched intent reasoning information after detailed interpretation of the reasoning prompt information.

[0055] Specifically, a pre-generated reasoning prompt template is obtained, for example, "According to the [MASK] input by the target object, infer the intent of the target object and generate intent reasoning information". After obtaining the initial dialogue information input by the target object, the [MASK] in the reasoning prompt template is replaced by the initial dialogue information to obtain the reasoning prompt information.

[0056] In a specific application scenario, the large language model can include but is not limited to Deep Neural Networks (DNNs), Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM), and generative pre-training Transformer model, etc. The specific construction and specific deployment of the large language model are not limited here.

[0057] S203: Based on the initial dialogue information and the intent reasoning information, obtaining the to-be-recognized information of the target object.

[0058] In an embodiment, the initial dialogue information input by the target object and the intent reasoning information generated by the reasoning large model are spliced to obtain the to-be-recognized information.

[0059] The above scheme, after obtaining the initial dialogue information of the target object, generates corresponding intent reasoning information by using the reasoning large model, and obtains the to-be-recognized information by combining the initial dialogue information and the intent reasoning information, so as to improve the representation ability of the to-be-recognized information, thereby helping to improve the accuracy of subsequent intent recognition.

[0060] Please refer to Figure 3 , Figure 3 is Figure 1 The step S102 in the above embodiment corresponds to the flowchart of another embodiment. Specifically, the implementation process of step S102 includes:

[0061] S301: Obtain a plurality of reference question templates.

[0062] In an embodiment, a plurality of reference question templates are obtained, and the specific process can refer to the above corresponding embodiment, which will not be described in detail here.

[0063] S302: Obtain a first score between the to-be-identified information and the reference question templates based on the respective content corresponding to the to-be-identified information and the reference question templates.

[0064] In an implementation, in response to the to-be-identified information and the reference question templates being texts, the first score between the to-be-identified information and each reference question template is determined according to the respective content corresponding to the to-be-identified information and the reference question templates, and the first score is used to represent the matching degree between the to-be-identified information and the corresponding reference question template, that is, the higher the first score, the higher the matching degree between the to-be-identified information and the corresponding reference question template.

[0065] In some implementation scenarios, the respective content corresponding to the to-be-identified information and the reference question templates includes character elements. The number of same character elements in the to-be-identified information and the reference question templates is obtained. The first score between the to-be-identified information and the corresponding reference question template is determined according to the number of same character elements, that is, the more the number of same character elements, the higher the first score.

[0066] S303: Obtain a plurality of candidate question templates from all reference question templates based on the first score corresponding to each reference question template.

[0067] In an implementation, after obtaining the first score corresponding to each reference question template, all reference question templates are sorted according to the order of the first score from large to small to obtain a reference question template sequence. The reference question templates located in the front first number of the reference question template sequence are used as the candidate question templates.

[0068] The above scheme determines the candidate question templates from the plurality of reference question templates to provide a reference basis for subsequently determining the target intent corresponding to the to-be-identified information, which helps to improve the accuracy of intent recognition.

[0069] Please refer to Figure 4 , Figure 4 is Figure 3 the flowchart of another implementation corresponding to step S302. Specifically, the implementation process of step S302 includes:

[0070] S401: Obtain a content similarity between the to-be-identified information and the reference question templates based on the respective content corresponding to the to-be-identified information and the reference question templates.

[0071] In an implementation, the BM25 (Best Matching 25, information retrieval) algorithm is used to calculate the relevance of the to-be-identified information and the reference question templates to obtain the corresponding content similarity.

[0072] Specifically, the BM25 algorithm queries the frequency of occurrence of each character of the to-be-identified information and the reference question template, so as to obtain the content similarity according to the frequency of occurrence. The specific application process of the BM25 algorithm can refer to the prior art, and will not be described in detail here.

[0073] S402: Based on the semantics corresponding to the to-be-identified information and the reference question template respectively, the semantic similarity between the to-be-identified information and the reference question template is obtained.

[0074] In an embodiment, the to-be-identified information and the reference question template are respectively subjected to semantic extraction, so as to obtain the semantics corresponding to the to-be-identified information and the reference question template respectively. According to the extracted semantics, the semantic similarity between the to-be-identified information and the reference question template is obtained.

[0075] Specifically, the to-be-identified information and the reference question template are respectively subjected to semantic extraction by using the trained semantic extraction model, so as to obtain the corresponding semantics. In response to the extracted semantics being a feature vector, the semantic similarity between the to-be-identified information and the corresponding reference question template is obtained by calculating the vector cosine distance between the semantics corresponding to the to-be-identified information and the semantics corresponding to the reference question template.

[0076] In some implementation scenarios, the specific structure of the above-mentioned semantic extraction model can refer to the existing neural network model structure, and will not be described in detail here.

[0077] It should be noted that in actual application, the order of obtaining the content similarity and the semantic similarity can also be other, for example, the content similarity and the semantic similarity can be obtained at the same time; or, the semantic similarity can be obtained first, and then the content similarity can be obtained.

[0078] S403: For the content similarity and the semantic similarity corresponding to each reference question template, a first score corresponding to the reference question template is obtained.

[0079] In an embodiment, after obtaining the content similarity and the semantic similarity corresponding to the reference question template through the above-mentioned steps, for each reference question template, the corresponding content similarity and semantic similarity are weighted and summed to obtain the first score.

[0080] Specifically, the specific calculation formula of the first score corresponding to each reference question template is as follows:

[0081]

[0082] wherein, the first score corresponding to the reference question template is represented by f, the content similarity corresponding to the same reference question template is represented by s, characterize the semantic similarity corresponding to the same reference question template, characterize the weight corresponding to the content similarity, characterize the weight corresponding to the semantic similarity. The weight and the weight The specific values of the weight and the weight can be obtained by estimation, or can be obtained by relevant technical personnel through multiple experiments.

[0083] The above scheme, for the to-be-recognized information and the reference question template, combines the corresponding content and semantics to determine the first score, so that the first score can accurately characterize the matching degree between the to-be-recognized information and the corresponding reference question template.

[0084] Please refer to Figure 5 , Figure 5 is Figure 1 The flowchart of step S103 in the embodiment is shown in FIG. 3. Specifically, the implementation process of step S103 includes the following steps.

[0085] S501: Obtain a plurality of reference intent information; wherein at least part of the reference intent information matches the associated information.

[0086] In an embodiment, an intent information library is obtained, the intent information library includes reference intent information matched with a plurality of intents, and at least part of the reference intent information matches the associated information. Moreover, each reference intent information includes a plurality of intent sub-information.

[0087] In some implementation scenarios, for each reference intent information, the plurality of intent sub-information included therein includes at least part of intent description, example text, and intent scene description. The intent description includes intent name and description text matched with the corresponding intent, for example, “intent name: query balance, description text: the user indicates the need to handle the bank card to query the balance in the card”. The example text includes examples matched with the corresponding intent, for example, “I want to see how much money is in my card, please help me check the balance”. The intent scene description includes the limitation and / or description of the application scene of the corresponding intent, for example, “this intent only includes the balance query in the bank card scene, and other scenes such as membership cards do not belong to this intent”. Each reference intent information includes a plurality of different types of intent sub-information, so that the matching degree between the to-be-recognized information and different intents is determined according to different types of intent sub-information, which helps to improve the accuracy of intent recognition.

[0088] In some implementation scenarios, the association information between different reference intent information corresponds to a task topology structure, to represent the development sequence between different reference intent information associated with each other. For example, in a bank service scenario, the "query balance" intent is often associated with the "withdrawal" or "transfer" intent, and the priority is "query balance", then "deposit" or "withdrawal". By determining the association information between at least part of the reference intent information, a reference basis is provided for subsequent determination of candidate intent information, and the accuracy of determining candidate intent information is improved.

[0089] In an embodiment, the implementation process of step S501 can further include: obtaining an intent information library including a plurality of reference intent information. Obtain attribute sub-information matched with each reference intent information. Based on the attribute sub-information respectively matched with different reference intent information, it is determined that there is association information between at least part of the reference intent information.

[0090] Specifically, information extraction is performed on each reference intent information to obtain entity content and behavior content in the reference intent information, and the extracted content is used as matched attribute sub-information. The attribute sub-information respectively matched with any two reference intent information is compared, and the association information between the corresponding reference intent information is determined according to the comparison result.

[0091] In some implementation scenarios, in response to the existence of common elements between the attribute sub-information respectively corresponding to different reference intent information, it is considered that the different reference intent information are associated, and the common elements are used as the corresponding association information. For example, for a first reference intent information and a second reference intent information, the first attribute sub-information corresponding to the first reference intent information includes "bank, balance, query", and the second attribute sub-information corresponding to the second reference intent information includes "bank, withdrawal", and the first reference intent information and the second reference intent information have a common element "bank", and it is determined that they are associated.

[0092] Of course, in other implementation scenarios, the number of common elements between the attribute sub-information respectively corresponding to different reference intent information can also be used to determine whether the different reference intent information are associated. For example, when the number of common elements corresponding to different reference intent information exceeds a preset number threshold, it is considered that there is association between the different reference intent information.

[0093] Alternatively, after the attribute sub-information matched with each reference intent information is extracted, a key element is determined from the attribute sub-information. In response to different reference intent information corresponding to the same key element, it is considered that there is an association between the different reference intent information, and the common key element is taken as the association information. For example, the first attribute sub-information corresponding to the first reference intent information includes "bank, balance, query", and the key element "bank" is determined; the second attribute sub-information corresponding to the second reference intent information includes "bank, withdrawal", and the key element "bank" is also determined. In response to the key elements corresponding to the first reference intent information and the second reference intent information being the same, it is determined that the first reference intent information and the second reference intent information are associated with each other.

[0094] In a specific application scenario, the above process can be performed by processing the reference intent information through natural language processing technology to extract the attribute sub-information corresponding to the reference intent information. In addition, for the extracted attribute sub-information, the key element in the attribute sub-information is determined by using natural language processing technology.

[0095] S502: Based on the semantics corresponding to each intent sub-information and the semantics corresponding to the to-be-identified information, the reference similarity between the to-be-identified information and each intent sub-information is obtained.

[0096] In an embodiment, for each intent sub-information in the reference intent information, the reference similarity between the to-be-identified information and the intent sub-information is obtained according to the semantics corresponding to the intent sub-information and the semantics corresponding to the to-be-identified information. The process of extracting semantics from the intent sub-information and the to-be-identified information can refer to the above corresponding embodiments.

[0097] In some implementation scenarios, the semantic features of the to-be-identified information and each intent sub-information are extracted respectively, and the corresponding reference similarity is calculated according to the extracted semantic features.

[0098] It should be noted that, in order to improve the efficiency of intent recognition and reduce the consumption of computing resources, after the to-be-identified information is obtained, the to-be-identified information is extracted once to make the extracted semantics directly called in the subsequent corresponding embodiments to calculate the relevant similarity.

[0099] In some implementation scenarios, in order to improve the accuracy of calculating the reference similarity, for each type of intent sub-information, a corresponding semantic discrimination model is trained. The to-be-identified information and the intent sub-information are input into the corresponding semantic discrimination model, and the corresponding reference similarity is output by using the semantic discrimination model. The higher the reference similarity is, the higher the matching degree between the to-be-identified information and the corresponding intent sub-information is, that is, the higher the matching degree between the to-be-identified information and the corresponding reference intent information is. The specific structure of the semantic discrimination model can refer to the existing neural network model structure, and will not be described in detail here.

[0100] S503: Obtain a second score between the to-be-recognized information and the corresponding reference intent information based on the reference similarity corresponding to each intent sub-information.

[0101] In an embodiment, for each reference intent information, the reference similarities corresponding to each intent sub-information are weighted and summed to obtain the second score between the to-be-recognized information and the reference intent information.

[0102] In an embodiment, after obtaining the second score corresponding to each reference intent information, the method further includes updating the second score corresponding to at least part of the reference intent information by using the association information.

[0103] In some implementation scenarios, in response to the second score corresponding to the reference intent information being greater than a preset score threshold, an update weight is obtained, and the second score of the associated other reference intent information is updated by using the update weight to increase the second score of the associated other reference intent information. The value of the update weight is greater than 1. This way updates the second score corresponding to at least part of the reference intent information by combining the association information, improving the flexibility of obtaining the second score.

[0104] In a specific application scenario, the reference intent information corresponding to "query balance" is associated with the reference intent information corresponding to "withdrawal". In response to the second score of the reference intent information corresponding to "query balance" being 0.9 and the second score being greater than the preset score threshold 0.8, the second score 0.6 of the reference intent information corresponding to "withdrawal" is updated. Specifically, the second score 0.6 is multiplied by the update weight 1.2 to obtain the updated second score 0.72.

[0105] S504: Obtain a plurality of candidate intent information from all reference intent information based on the second score corresponding to each reference intent information and the association information.

[0106] In an embodiment, after obtaining the second score corresponding to each reference intent information, all reference intent information is sorted according to the order of the second score from large to small to obtain a reference intent information sequence. The first second number of reference intent information in the reference intent information sequence is taken as the candidate intent information.

[0107] In an embodiment, the implementation process of the above step S504 can further include sorting all reference intent information based on the order of the second score from large to small to obtain a reference intent information sequence, and obtaining a plurality of screening intent information from the reference intent information sequence. The second score corresponding to the screening intent information is greater than the second score corresponding to the remaining reference intent information in the reference intent information sequence.

[0108] Further, based on the associated information corresponding to the screening intent information, at least part of other reference intent information associated with the screening intent information is obtained. Based on the screening intent information and the at least part of other reference intent information associated therewith, a plurality of candidate intent information is obtained.

[0109] In some implementation scenarios, all the reference intent information is sorted in descending order of the second scores, to obtain a sequence of reference intent information. The plurality of reference intent information corresponding to the maximum second score in the sequence of reference intent information is taken as the screening intent information, that is, the second score corresponding to any screening intent information is greater than the second score corresponding to the rest of the reference intent information in the sequence of reference intent information. For the obtained screening intent information, other reference intent information associated with the screening intent information is determined from all the reference intent information other than the screening intent information according to the associated information corresponding to each screening intent information. The screening intent information and the other reference intent information associated therewith are taken as the candidate intent information. This manner determines the plurality of candidate intent information related to the to-be-recognized information by combining the second score and the associated information between different reference intent information, and improves the accuracy and comprehensiveness of obtaining the candidate intent information.

[0110] In some implementation scenarios, in response to the screening intent information being associated with a plurality of other reference intent information, all the screening intent information is taken as the candidate intent information; and at least one is selected from all the other reference intent information corresponding to the screening intent information as the candidate intent information according to the corresponding second score.

[0111] In a specific application scenario, all the reference intent information is sorted in descending order of the second score to obtain a sequence of reference intent information, and the first third quantity of reference intent information is extracted from the sequence of reference intent information as the screening intent information. For each screening intent information, all the other reference intent information associated therewith is obtained from the sequence of reference intent information. The part of other reference intent information with the highest second score is selected according to the second score corresponding to all the other reference intent information associated with the screening intent information, and the selected part of other reference intent information and the screening intent information are taken as the candidate intent information.

[0112] The above scheme determines the candidate intent information from the plurality of reference intent information, to provide a reference basis for subsequently determining the target intent corresponding to the to-be-recognized information, which helps to improve the accuracy of intent recognition.

[0113] Please refer to Figure 6 , Figure 6 is Figure 5 The flowchart of another embodiment corresponding to step S502 is shown in FIG. 5B. Specifically, the plurality of intent sub-information of the reference intent information includes intent description, example text and intent scene description, based on which the implementation process of step S502 includes:

[0114] S601: Obtain a description similarity between the to-be-identified information and the corresponding reference intent information based on semantics corresponding to the intent description and the to-be-identified information respectively; wherein the intent description comprises an intent name.

[0115] In an embodiment, semantics of the intent description and the to-be-identified information are extracted respectively to obtain semantics corresponding to the to-be-identified information and the intent description respectively. Based on the extracted semantics, a description similarity between the to-be-identified information and the intent description is obtained.

[0116] Specifically, in response to the extracted semantics being a feature vector, a vector cosine distance between a semantic feature vector corresponding to the to-be-identified information and a semantic feature vector corresponding to the intent description is calculated to obtain a description similarity between the to-be-identified information and the corresponding reference intent information.

[0117] S602: Obtain an example similarity between the to-be-identified information and the corresponding reference intent information based on semantics corresponding to the example text and the to-be-identified information respectively.

[0118] In an embodiment, semantics of the example text and the to-be-identified information are extracted respectively to obtain semantics corresponding to the to-be-identified information and the example text respectively. Based on the extracted semantics, an example similarity between the to-be-identified information and the example text is obtained. The specific implementation process can be referred to the above corresponding embodiment, which will not be described in detail here.

[0119] S603: Obtain a scene similarity between the to-be-identified information and the corresponding reference intent information based on semantics corresponding to the intent scene description and the to-be-identified information respectively.

[0120] In an embodiment, semantics of the intent scene description and the to-be-identified information are extracted respectively to obtain semantics corresponding to the to-be-identified information and the example text respectively. Based on the extracted semantics, an example similarity between the to-be-identified information and the example text is obtained.

[0121] The above scheme, the plurality of intent sub-information of the reference intent information comprises the intent description, the example text and the intent scene description, and the reference similarity comprises an intent similarity between the to-be-identified information and the intent description, an example similarity between the to-be-identified information and the example text, and a scene similarity between the to-be-identified information and the intent scene description. By obtaining the reference similarity according to different types of intent sub-information, it is helpful to improve the accuracy of subsequent calculation of the second score.

[0122] In an embodiment, in response to the reference similarity obtained by the above embodiment comprising the description similarity, the example similarity and the scene similarity, the obtaining step of the second score comprises: obtaining the second score based on the description similarity, the example similarity and the scene similarity corresponding to the reference intent information.

[0123] Specifically, for each reference intention information, the corresponding description similarity, example similarity and scene similarity are weighted and summed to obtain a second score for representing the correlation degree between the corresponding reference intention information and the to-be-identified information. The specific calculation formula of the second score is as follows:

[0124]

[0125] wherein, representing the second score corresponding to the reference intention information, representing the description similarity corresponding to the reference intention information, representing the example similarity corresponding to the same reference intention information, representing the scene similarity corresponding to the same reference intention information, representing the weight corresponding to the description similarity, representing the weight corresponding to the example similarity, representing the weight corresponding to the scene similarity. The specific values of the weights , the weight and the weight may be obtained by estimation or may be obtained by relevant technical personnel through multiple experiments and backstepping.

[0126] In addition, it should be noted that in other embodiments, the multiple intention sub-information in the reference intention information can also include part of the intention description, example text and intention scene description. For example, the multiple intention sub-information includes the intention description and the intention scene description.

[0127] Please refer to Figure 7 , Figure 7 is Figure 1 the flowchart of another embodiment corresponding to step S104. Specifically, in response to obtaining the multiple candidate question templates and the multiple candidate intention information, the implementation process of step S104 includes:

[0128] S701: obtaining a first decision result based on the intention label matched by the multiple candidate question templates.

[0129] In an embodiment, each candidate question template is matched with an intention label, and the to-be-identified information is preliminarily decided according to the intention label matched by each candidate question template to obtain a first decision result.

[0130] Specifically, when constructing the AFQ database, each reference question template is matched with a corresponding intent label, and the intent label is used to directly represent the intent of the corresponding reference question template. Alternatively, the intent label also includes an intent solution for efficiently solving the intent corresponding to the reference question template. Each candidate question template obtained through the above-mentioned corresponding embodiments is matched with an intent label, so that the target intent is generated according to the intent label subsequently, and the accuracy of intent recognition is improved.

[0131] In some implementation scenarios, a decision large model is obtained. In response to obtaining a plurality of candidate question templates, the to-be-recognized information, the plurality of candidate question templates, and the intent labels matched therewith are input into the decision large model, the input information is analyzed by using the decision large model, and a first decision result for representing the intent of the to-be-recognized information is generated. The above-mentioned decision large model is a large language model with relatively optimal data analysis capability. According to the to-be-recognized information, the candidate question templates, and the intent labels matched therewith, corresponding first decision prompt information is generated, and the first decision prompt information is input into the decision large model, so that the decision large model generates the first decision result after detailed interpretation of the first decision prompt information.

[0132] Specifically, a pre-generated decision prompt template is obtained, for example, “according to the [MASK] input by the target object, [candidate question template 1] and the intent label [intent label 1] matched therewith, and [candidate question template 2] and the intent label [intent label 2] matched therewith, the intent of the target object is inferred, and the first decision information is generated”. The “[MASK]” in the decision prompt template is replaced by using the to-be-recognized information, and each candidate question template and the intent label matched therewith is filled into the corresponding position in the decision prompt template to obtain the first decision prompt information. The first decision prompt information is input into the decision large model to obtain the first decision result generated by the decision large model. The specific construction of the above-mentioned large language model can refer to the corresponding embodiments described above.

[0133] In an embodiment, to further improve the decision accuracy of the decision large model, the to-be-recognized information, the plurality of candidate question templates, the intent labels matched therewith, and the first scores are input into the decision large model, and the decision large model generates a first decision result for representing the intent of the to-be-recognized information by combining the first scores between the to-be-recognized information and each candidate question template.

[0134] S702: Based on the plurality of candidate intent information, a second decision result is obtained.

[0135] In an embodiment, the to-be-recognized information is preliminarily decided according to each candidate intent information, and a second decision result is obtained.

[0136] In some implementation scenarios, the to-be-recognized information and the plurality of candidate intent information are input into the decision large model, the input information is analyzed by using the decision large model, and a second decision result for representing an intent of the to-be-recognized information is generated. The specific structure of the decision large model and the specific obtaining process of the second decision result can refer to the corresponding implementation manners described above, and will not be described in detail herein.

[0137] Alternatively, the to-be-recognized information, the plurality of candidate intent information, and the second scores matched therewith can also be input into the decision large model, and the decision large model is used to generate a second decision result for representing an intent of the to-be-recognized information.

[0138] S703: The first decision result and the second decision result are input into the decision large model, and a target intent generated by the decision large model based on the first decision result and the second decision result is obtained.

[0139] In an implementation manner, after the first decision result and the second decision result are obtained by different manners, the first decision result and the second decision result are input into the decision large model, and the decision large model is used to generate a target intent with the highest fitting degree with the to-be-recognized information by combining the first decision result and the second decision result.

[0140] In some implementation scenarios, the first decision result and the second decision result are input into the decision large model, and the decision large model is prompted to perform intent recognition on the to-be-recognized information by referring to the first decision result and the second decision result, and a target intent with higher accuracy generated by the decision large model is obtained.

[0141] In some implementation scenarios, the target intent generated by the decision large model includes an intent solution, so that subsequent demand processing is performed on the target object according to the intent solution.

[0142] Please refer to Figure 8 , Figure 8 is Figure 7 The step S703 in the method 700 corresponds to a flowchart of another implementation manner. Specifically, the implementation process of the step S703 includes:

[0143] S801: Obtain historical dialogue information of a target object, and obtain target preference information matched with the target object based on the historical dialogue information.

[0144] In an implementation manner, historical dialogue information input by the target object on the intelligent terminal is obtained, the historical dialogue information is subjected to semantic analysis, and target preference information matched with the target object is extracted from the historical dialogue information.

[0145] In some implementation scenarios, in response to the to-be-identified information being obtained based on initial dialogue information input by the target object and the intelligent terminal, the timestamp of the historical dialogue information matching is earlier than the timestamp of the initial dialogue information matching. The historical dialogue information is input to the decision large model, prompting the decision large model to extract preference information from the historical dialogue information, and obtaining target preference information generated by the decision large model.

[0146] In some implementation scenarios, the preference information extraction large model is obtained, and the historical dialogue information is input to the preference information extraction large model to obtain target preference information generated by the preference information extraction large model. The preference information extraction large model is obtained by fine-tuning a large language model using a plurality of training data, so that the obtained preference information extraction large model has better preference information extraction capability and helps to improve the accuracy of extracting target preference information. The specific structure of the large language model can refer to the corresponding implementation mode described above, and the process of fine-tuning the large language model can refer to the existing technology, which will not be described in detail here.

[0147] S802: Based on the target preference information, the first decision result and the second decision result, obtain a target intent generated by a decision large model.

[0148] In an implementation mode, the target preference information, the first decision result and the second decision result are input to the decision large model, and the decision large model is prompted to refer to the target preference information, the first decision result and the second decision result to identify the to-be-identified information, and obtain a target intent generated by the decision large model with higher accuracy. The target intent is used to represent an intent solution matched with the target preference information.

[0149] In a specific application scenario, when the initial dialogue information input by the target object and the intelligent terminal is "my bank card is lost", and the target intent determined by any of the above implementation modes includes the intent solution "loss suspension". In response to the target preference information representing that the target object pays attention to processing efficiency, in order to improve the efficiency, the target object is no longer asked whether to need to suspend the loss, and the process of suspending the bank card loss is directly executed.

[0150] Please refer to Figure 9 , Figure 9 is a structural schematic diagram of an implementation mode of an intent recognition system of the present application. The intent recognition system includes an acquisition module 10, a first processing module 20, a second processing module 30 and an identification module 40 which are coupled to each other. Specifically:

[0151] The acquisition module 10 is used to acquire to-be-identified information of a target object.

[0152] The first processing module 20 is configured to obtain a plurality of reference question templates, and obtain a candidate question template from all the reference question templates based on content correlation between the to-be-identified information and the reference question templates.

[0153] The second processing module 30 is configured to obtain a plurality of reference intention information, and obtain a candidate intention information from all the reference intention information based on semantic correlation between a plurality of intention sub-information corresponding to the reference intention information and the to-be-identified information.

[0154] The recognition module 40 is configured to obtain a target intention corresponding to the to-be-identified information based on the candidate question template and the candidate intention information.

[0155] In an embodiment, the obtaining module 10 obtains the to-be-identified information of the target object, including: obtaining initial dialogue information input by the target object; inputting the initial dialogue information to the large inference model to obtain intention inference information generated by the large inference model and matched with the initial dialogue information; and obtaining the to-be-identified information of the target object based on the initial dialogue information and the intention inference information.

[0156] In an embodiment, the first processing module 20 obtains a plurality of reference question templates, and obtains a candidate question template from all the reference question templates based on content correlation between the to-be-identified information and the reference question templates, including: obtaining a plurality of reference question templates; obtaining a first score between the to-be-identified information and the reference question templates based on content corresponding to the to-be-identified information and the reference question templates respectively; and obtaining a plurality of candidate question templates from all the reference question templates based on the first score corresponding to each reference question template.

[0157] In an embodiment, the first processing module 20 obtains a first score between the to-be-identified information and the reference question templates based on content corresponding to the to-be-identified information and the reference question templates respectively, including: obtaining content similarity between the to-be-identified information and the reference question templates based on content corresponding to the to-be-identified information and the reference question templates respectively; obtaining semantic similarity between the to-be-identified information and the reference question templates based on semantics corresponding to the to-be-identified information and the reference question templates respectively; and obtaining the first score corresponding to each reference question template based on the content similarity and the semantic similarity corresponding to the reference question template.

[0158] In an embodiment, the second processing module 30 obtains a plurality of reference intention information, obtains candidate intention information from all the reference intention information based on semantic relevance between each intention sub-information corresponding to the reference intention information and the to-be-identified information, comprising: obtaining a plurality of reference intention information; wherein at least part of the reference intention information matches the associated information; obtaining a reference similarity between the to-be-identified information and each intention sub-information based on the semantics of the to-be-identified information and each intention sub-information; obtaining a second score between the to-be-identified information and the corresponding reference intention information based on the reference similarity corresponding to each intention sub-information; obtaining a plurality of candidate intention information from all the reference intention information based on the second score corresponding to each reference intention information and the associated information.

[0159] In an embodiment, the plurality of intention sub-information of the reference intention information comprises intention description, example text and intention scene description, the second processing module 30 obtains a reference similarity between the to-be-identified information and each intention sub-information based on the semantics of the to-be-identified information and each intention sub-information, comprising: obtaining a description similarity between the to-be-identified information and the corresponding reference intention information based on the semantics of the intention description and the to-be-identified information; wherein the intention description comprises an intention name; and obtaining an example similarity between the to-be-identified information and the corresponding reference intention information based on the semantics of the example text and the to-be-identified information; and obtaining a scene similarity between the to-be-identified information and the corresponding reference intention information based on the semantics of the intention scene description and the to-be-identified information.

[0160] In an embodiment, the reference similarity comprises the description similarity, the example similarity and the scene similarity, the second processing module 30 obtains a second score between the to-be-identified information and the corresponding reference intention information based on the reference similarity corresponding to each intention sub-information, comprising: obtaining the second score based on the description similarity, the example similarity and the scene similarity corresponding to the reference intention information.

[0161] In an embodiment, the second processing module 30 obtains a plurality of candidate intention information from all the reference intention information based on the second score corresponding to each reference intention information and the associated information, comprising: sorting all the reference intention information based on the order of the second score from large to small to obtain a reference intention information sequence, and obtaining a plurality of screening intention information from the reference intention information sequence; wherein the second score corresponding to the screening intention information is greater than the second score corresponding to the remaining reference intention information in the reference intention information sequence; obtaining at least part of the reference intention information associated with the screening intention information based on the associated information corresponding to the screening intention information; obtaining a plurality of candidate intention information based on the screening intention information and the at least part of the reference intention information associated therewith.

[0162] In an embodiment, in response to obtaining the plurality of candidate question templates and the plurality of candidate intent information, the candidate question templates are matched with the intent labels, the identification module 40 obtains the target intent corresponding to the to-be-identified information based on the candidate question templates and the candidate intent information, including: obtaining a first decision result based on the intent labels matched by the plurality of candidate question templates; and obtaining a second decision result based on the plurality of candidate intent information; inputting the first decision result and the second decision result into the decision large model to obtain the target intent generated by the decision large model based on the first decision result and the second decision result.

[0163] In an embodiment, the identification module 40 inputs the first decision result and the second decision result into the decision large model to obtain the target intent generated by the decision large model based on the first decision result and the second decision result, including: obtaining historical dialogue information of the target object, obtaining target preference information matched with the target object based on the historical dialogue information; obtaining the target intent generated by the decision large model based on the target preference information, the first decision result and the second decision result.

[0164] Please refer to Figure 10 , Figure 10 is a structural schematic diagram of an embodiment of an electronic device of the present application. The electronic device includes a memory 50 and a processor 60 coupled with each other. The memory 50 stores program instructions, and the processor 60 is configured to execute the program instructions to implement the method mentioned in any of the above embodiments. Specifically, the electronic device includes but is not limited to a desktop computer, a notebook computer, a tablet computer, a server, etc., which are not limited herein. In addition, the processor 60 can also be referred to as a CPU (Center Processing Unit). The processor 60 can be an integrated circuit chip with signal processing capability. The processor 60 can also be a general processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The general processor can be a microprocessor or the processor can also be any conventional processor. In addition, the processor 60 can be implemented by an integrated circuit chip together.

[0165] Please refer to Figure 11 , Figure 11is a structural schematic diagram of an embodiment of a computer readable storage medium of the present application, and the computer readable storage medium 70 stores program instructions 80 capable of being executed by a processor, and the program instructions 80 are executed by the processor to implement the method mentioned in any of the above embodiments.

[0166] In several embodiments provided in the present application, it should be understood that the disclosed methods and devices can be implemented in other manners. For example, the above-described device embodiments are merely schematic, and the division of the modules or units is merely a logical function division, and there can be another division manner in actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections can be indirect couplings or communication connections through some interfaces, devices or units, and can be electrical, mechanical or in other forms.

[0167] The units described as separated components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purposes of the embodiments.

[0168] In addition, each functional unit in the embodiments of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The above integrated unit can be implemented in the form of hardware, or in the form of software functional units.

[0169] If the integrated unit is implemented in the form of software functional units and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on such an understanding, the technical solutions of the present application essentially or the part that makes contributions to the prior art, or all or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor (processor) to execute all or part of the steps of the methods in the embodiments of the present application. The foregoing storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk, and various media that can store program codes.

[0170] The above merely describes the embodiments of the present application, and does not limit the protection scope of the present application, and any equivalent structure or equivalent process transformation made by using the content of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the protection scope of the present application.

Claims

1. An intent recognition method, characterized in that, include: Obtain the identification information of the target object; Multiple reference question templates are obtained, and candidate question templates are obtained from all the reference question templates based on the content correlation between the information to be identified and the reference question templates. as well as, Multiple reference intent information is obtained, and candidate intent information is obtained from all the reference intent information based on the semantic correlation between the multiple intent sub-information corresponding to the reference intent information and the information to be identified; wherein, the multiple intent sub-information of the reference intent information includes intent description, example text and intent scenario description; Based on the candidate question template and the candidate intent information, the target intent corresponding to the information to be identified is obtained; The step of acquiring multiple reference intent information, and acquiring candidate intent information from all the reference intent information based on the semantic correlation between the multiple intent sub-information corresponding to the reference intent information and the information to be identified, includes: acquiring multiple reference intent information; wherein at least some of the reference intent information are matched with correlation information; wherein the correlation information represents the development order between different related reference intent information; acquiring a reference similarity between the information to be identified and each of the intent sub-information based on the semantics corresponding to each of the intent sub-information; acquiring a second score between the information to be identified and the corresponding reference intent information based on the reference similarity corresponding to each of the intent sub-information; and acquiring multiple candidate intent information from all the reference intent information based on the second score corresponding to each of the reference intent information and the correlation information.

2. The intent recognition method according to claim 1, characterized in that, The process of obtaining the target object's identification information includes: Obtain the initial dialogue information input by the target object; The initial dialogue information is input into the reasoning model to obtain the intent reasoning information generated by the reasoning model that matches the initial dialogue information; Based on the initial dialogue information and the intent reasoning information, the information to be identified of the target object is obtained.

3. The intent recognition method according to claim 1, characterized in that, The step of obtaining multiple reference question templates involves obtaining candidate question templates from all the reference question templates based on the content correlation between the information to be identified and the reference question templates, including: Obtain multiple reference question templates; Based on the content corresponding to the information to be identified and the reference question template, a first score is obtained between the information to be identified and the reference question template. Based on the first score corresponding to each of the reference question templates, a plurality of candidate question templates are obtained from all the reference question templates.

4. The intent recognition method according to claim 3, characterized in that, The step of obtaining a first score between the information to be identified and the reference question template based on their respective contents includes: Based on the content corresponding to the information to be identified and the reference question template, the content similarity between the information to be identified and the reference question template is obtained; and... Based on the semantics corresponding to the information to be identified and the reference question template, the semantic similarity between the information to be identified and the reference question template is obtained. For each reference question template, the first score corresponding to the reference question template is obtained based on the content similarity and semantic similarity.

5. The intent recognition method according to claim 1, characterized in that, The step of obtaining a reference similarity between the information to be identified and each of the intent sub-informations based on their respective semantics includes: Based on the semantics corresponding to the intent description and the information to be identified, a description similarity is obtained between the information to be identified and the corresponding reference intent information; wherein, the intent description includes an intent name; and, Based on the semantics corresponding to the example text and the information to be identified, the example similarity between the information to be identified and the corresponding reference intent information is obtained; and... Based on the intent scene description and the semantics corresponding to the information to be identified, the scene similarity between the information to be identified and the corresponding reference intent information is obtained.

6. The intent recognition method according to claim 5, characterized in that, The reference similarity includes the description similarity, the example similarity, and the scene similarity. The step of obtaining a second score between the information to be identified and the corresponding reference intent information based on the reference similarity corresponding to each of the intent sub-information includes: The second score is obtained based on the description similarity, example similarity, and scene similarity corresponding to the reference intent information.

7. The intent recognition method according to claim 1, characterized in that, The step of obtaining multiple candidate intent information from all the reference intent information based on the second score corresponding to each of the reference intent information and the association information includes: All the reference intent information is sorted in descending order based on the second score to obtain a reference intent information sequence, and multiple filtering intent information is obtained from the reference intent information sequence; wherein, the second score corresponding to the filtering intent information is greater than the second score corresponding to the other reference intent information in the reference intent information sequence. Based on the association information corresponding to the filtering intent information, at least some other reference intent information associated with the filtering intent information is obtained; Based on the filtering intent information and at least some of the other reference intent information associated with it, a plurality of candidate intent information is obtained.

8. The intent recognition method according to claim 1, characterized in that, In response to obtaining multiple candidate question templates and multiple candidate intent information, wherein the candidate question templates are matched with intent tags, the step of obtaining the target intent corresponding to the information to be identified based on the candidate question templates and the candidate intent information includes: Based on the intent tags matched with multiple candidate question templates, a first decision result is obtained; and, Based on the multiple candidate intent information, a second decision result is obtained; The first decision result and the second decision result are input into the decision model to obtain the target intent generated by the decision model based on the first decision result and the second decision result.

9. The intent recognition method according to claim 8, characterized in that, The step of inputting the first decision result and the second decision result into the decision-making big model to obtain the target intent generated by the decision-making big model based on the first decision result and the second decision result includes: Obtain the historical dialogue information of the target object, and based on the historical dialogue information, obtain target preference information that matches the target object; Based on the target preference information, the first decision result, and the second decision result, the target intent generated by the decision-making big model is obtained.

10. An intent recognition system, characterized in that, include: The acquisition module is used to acquire the identification information of the target object; The first processing module is used to obtain multiple reference question templates and, based on the content correlation between the information to be identified and the reference question templates, obtain candidate question templates from all the reference question templates. The second processing module is used to acquire multiple reference intent information, and based on the semantic correlation between the multiple intent sub-information corresponding to the reference intent information and the information to be identified, acquire candidate intent information from all the reference intent information; wherein, the multiple intent sub-information of the reference intent information includes intent description, example text and intent scenario description; The identification module is used to obtain the target intent corresponding to the information to be identified based on the candidate question template and the candidate intent information; The step of acquiring multiple reference intent information, and acquiring candidate intent information from all the reference intent information based on the semantic correlation between the multiple intent sub-information corresponding to the reference intent information and the information to be identified, includes: acquiring multiple reference intent information; wherein at least some of the reference intent information are matched with correlation information; wherein the correlation information represents the development order between different related reference intent information; acquiring a reference similarity between the information to be identified and each of the intent sub-information based on the semantics corresponding to each of the intent sub-information; acquiring a second score between the information to be identified and the corresponding reference intent information based on the reference similarity corresponding to each of the intent sub-information; and acquiring multiple candidate intent information from all the reference intent information based on the second score corresponding to each of the reference intent information and the correlation information.

11. An electronic device, characterized in that, include: A memory and a processor are coupled to each other, the memory storing program instructions, and the processor executing the program instructions to implement the method as described in any one of claims 1-9.

12. A computer-readable storage medium having program instructions stored thereon, characterized in that, When the program instructions are executed by the processor, they implement the method as described in any one of claims 1-9.

Images