Dynamic training method, device, equipment and storage medium for intent recognition model
By receiving the input information to be identified and automatically labeling and adding training corpus, dynamic training of the intent recognition model is solved, and the problem that the intent recognition system in the prior art relies on a large number of labeled corpus is improved, the model accuracy is improved and the dynamic training cost is reduced.
Patent Information
- Application Number
- CN201810658950.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2018-06-25
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2038-06-25
AI Technical Summary
The existing intention recognition system relies on a model that has been dynamically trained in advance, and the more corpus the training, the better, but it increases the cost and practicality of the model. How to dynamically train and update the model without increasing the length of the labeled corpus to Improve recognition accuracy.
By receiving the input information to be identified, automatically labeling and adding training corpus. When the corpus length exceeds the threshold, dynamically train the intent recognition model to generate a new model, and use diversified inputs to update the model.
Without increasing the length of the labeled corpus, the accuracy of the intention recognition model is improved, the cost of dynamic training is reduced, and the shortcomings of existing models needing a large number of labeled corpus.
Smart Images

Figure CN110633724B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence, and in particular to a dynamic training method, device, equipment and storage medium for an intention recognition model. Background Art
[0002] Intent recognition is an important research direction in the field of natural language processing, which is mainly used to identify the behavioral intentions of users. Existing intent recognition systems all perform recognition on intention recognition models that have been dynamically trained in advance, that is, the recognition effect of the intent recognition system depends on the intention recognition model that has been dynamically trained in advance, and the generation of the intention recognition model depends on the training corpus that has been annotated in advance. The more and more complete the training corpus that has been annotated in advance, the better the recognition effect of the intent recognition system. However, on the other hand, too much training corpus will also increase the dynamic training cost of the intent recognition model, and the practicality of this model building method is not strong.
[0003] Therefore, how to effectively train and update the pre-trained intent recognition model according to the user's diverse input without increasing the length of the annotated corpus, improve the accuracy of the intent recognition model, and then improve the recognition accuracy of the intent recognition system is an urgent problem to be solved. Summary of the invention
[0004] In view of this, the purpose of the embodiments of the present invention is to provide a method, device and storage medium for dynamic training of an intent recognition model, so as to solve the problem of effectively dynamically training and updating the pre-dynamically trained intent recognition model according to the diverse inputs of users without increasing the length of the annotated corpus, thereby improving the accuracy of the intent recognition model.
[0005] The technical solution adopted by the embodiment of the present invention to solve the above technical problems is as follows:
[0006] According to one aspect of the present invention, a dynamic training method for an intent recognition model is provided, comprising:
[0007] After receiving the input information to be recognized, the input information is recognized according to the existing intention recognition model;
[0008] When recognition is successful, the input information and intent recognition results are automatically annotated, and the annotated results are used as additional training corpus;
[0009] When the length of the additional corpus exceeds a preset threshold, the existing intent recognition model is dynamically trained according to the training corpus to generate a new intent recognition model.
[0010] According to another aspect of the present invention, a dynamic training device for an intention recognition model is provided, the device comprising:
[0011] An input module, used for receiving input information to be identified;
[0012] The recognition module is used to perform intent recognition on the input information according to the existing intent recognition model after receiving the input information to be recognized;
[0013] The automatic annotation module is used to automatically annotate the input information and intent recognition results after successful recognition, and use them as additional training corpus.
[0014] The training module is used to dynamically train the existing intent recognition model again according to the training corpus to generate a new intent recognition model when the length of the additional corpus exceeds a preset threshold.
[0015] According to another aspect of the present invention, a dynamic training device for an intent recognition model is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the steps of the dynamic training method for the above-mentioned intent recognition model are implemented.
[0016] According to another aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the dynamic training method of the above-mentioned intent recognition model are implemented.
[0017] The dynamic training method, apparatus, device and computer-readable storage medium of the intent recognition model provided in the embodiments of the present invention can effectively utilize diversified inputs to dynamically train and update the existing intent recognition model without increasing the length of the annotated corpus, thereby making up for the deficiency that the existing intent recognition model requires a large amount of annotated corpus for dynamic training, improving the accuracy of the intent recognition model and minimizing the dynamic training cost of the intent recognition model. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Figure 1 A flowchart of a dynamic training method for an intent recognition model provided in Embodiment 1 of the present invention;
[0019] Figure 2 A flowchart of another dynamic training method for an intent recognition model provided in Embodiment 2 of the present invention;
[0020] Figure 3 A flow chart of an automatic labeling method based on input information and intention recognition results provided in Embodiment 3 of the present invention;
[0021] Figure 4 A flow chart of an automatic assisted manual labeling method based on input information provided in Embodiment 4 of the present invention;
[0022] Figure 5A flowchart of a dynamic training method for an intention recognition model in the football field provided in Embodiment 5 of the present invention;
[0023] Figure 6 This is a module structure diagram of the dynamic training device for the intent recognition model provided in Example 6 of the present invention.
[0024] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION
[0025] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention clearer and more understandable, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.
[0026] Embodiment 1
[0027] like Figure 1 As shown, an embodiment of the present invention provides a dynamic training method for an intent recognition model, the method comprising:
[0028] S101: After receiving input information to be identified, perform intent identification on the input information according to an existing intent identification model.
[0029] S102: When the recognition is successful, the input information and the intention recognition result are automatically annotated, and the annotated result is used as additional training corpus.
[0030] Specifically, if the intent recognition is successful, the user's input information has been segmented, entity recognized, and attribute relationship extracted, so it can be automatically labeled based on the user's input information and the recognition result, and the labeled result can be used as additional training corpus. If the intent recognition fails, the user needs to manually label the input information that failed to be recognized, and then the labeled result can be used as additional training corpus.
[0031] S103: When the length of the additional corpus exceeds a preset threshold, the existing intent recognition model is dynamically trained according to the training corpus to generate a new intent recognition model.
[0032] Specifically, when the length of the additional corpus exceeds a preset threshold, the existing intent recognition model is dynamically trained again based on the training corpus to generate a new intent recognition model with higher accuracy.
[0033] It should be noted that the intent recognition system loads the intent recognition model in a hot loading manner, that is, the dynamic training of the intent recognition model will not affect the recognition of the intent recognition system. In addition, the specific method of dynamically training the intent recognition model can be based on existing statistical learning methods or existing machine learning methods.
[0034] In the embodiments of the present invention, without increasing the length of the annotated corpus, the existing intent recognition model can be effectively trained and updated using diversified inputs, thereby making up for the deficiency that the existing intent recognition model requires a large amount of annotated corpus for dynamic training, improving the accuracy of the intent recognition model, and minimizing the dynamic training cost of the intent recognition model.
[0035] Embodiment 2
[0036] like Figure 2 As shown, an embodiment of the present invention provides another dynamic training method for an intent recognition model, the method comprising:
[0037] S201: After receiving input information to be identified, perform intent identification on the input information according to an existing intent identification model.
[0038] S202: When the recognition is successful, the input information and the intention recognition result are automatically marked.
[0039] S203: When recognition fails, the input information is automatically assisted in manual annotation, and the annotation results are used as additional training corpus.
[0040] S204: When the length of the additional corpus exceeds a preset threshold, the existing intent recognition model is dynamically trained again according to the training corpus to generate a new intent recognition model.
[0041] Based on the above example 1, the embodiment of the present invention assists in manual annotation of the input information that fails to be recognized and also uses it as additional corpus to dynamically train the intent recognition model.
[0042] Embodiment 3
[0043] like Figure 3 As shown, an embodiment of the present invention provides an automatic labeling method based on input information and recognition results, the method comprising:
[0044] S301, obtaining semantic analysis results and intent recognition results of input information obtained by intent recognition.
[0045] Among them, the semantic analysis results include word segmentation results, entity recognition results, attribute extraction results and / or relationship extraction results.
[0046] Specifically, since intent recognition has been successful, it means that semantic analysis such as word segmentation, entity recognition, and attribute relationship extraction has been successfully performed on the user's input, so the semantic analysis results and recognition results of the user's input information in the recognition process are obtained, and the user input and recognition results are automatically labeled.
[0047] S302: Automatically label entity labels, attribute labels and / or relationship labels according to the semantic analysis results.
[0048] Specifically, based on the word segmentation results and entity recognition results, the words corresponding to the identified entities are marked as entity labels; based on the word segmentation results and attribute extraction results, the words corresponding to the identified attribute names are marked as attribute labels; based on the word segmentation results and relationship extraction results, the words corresponding to the identified relationships are marked as relationship labels.
[0049] S303: Mark the intent recognition result as an intent label.
[0050] In the embodiment of the present invention, by utilizing the semantic analysis results and intent recognition results of the user input information corresponding to the successful recognition, the input information is automatically annotated, thereby eliminating the need for manual annotation and improving the efficiency and accuracy of training.
[0051] Embodiment 4
[0052] like Figure 4 As shown, an embodiment of the present invention provides an automatic assisted manual labeling method based on input information, the method comprising:
[0053] S401 、 Perform word segmentation on the input information to obtain the word segmentation result.
[0054] S402 、 Dependency syntactic analysis is performed based on the word segmentation results to obtain a set of dependency relations.
[0055] Specifically, dependency syntactic analysis uses the dependency relationship between words in a sentence to represent the syntactic structure information of words. For example, the structural relationship between subject and predicate, verb and object, attributive, etc., and uses a tree structure to represent the structure of the whole sentence (such as subject, predicate, object, attributive, adverbial, complement, etc.). This set mainly includes the word part of speech and the relationship between words (such as: master-slave relationship, master-slave-passive relationship, verb-object relationship, parallel relationship, correlation relationship, etc.); dependency syntactic analysis can be directly analyzed using the existing Harbin Institute of Technology LTP.
[0056] S403 、 The nouns or personal pronouns marked as subjects in the dependency set are output as candidate subject sets to the user for confirmation, and the subjects confirmed by the user are marked as entity labels.
[0057] Specifically, first search whether there are words with a master-slave relationship in the dependency set. If so, directly obtain the subject of the user's input information from the words with a master-slave relationship; if not, take all nouns or personal pronouns marked as subjects in the dependency set as a candidate subject set, and output the candidate subject set to the user to manually determine the subject, and mark the subject determined by the user as an entity label.
[0058] S404: Output the verbs or nouns marked as objects in the dependency set as candidate object sets to the user for confirmation, and mark the objects confirmed by the user as intent labels / entity labels.
[0059] Specifically, the object is searched from the dependency set to determine the intent label or entity label. When the object cannot be found, the dependency set is searched to see if there are words with a subject-object relationship. If so, the object of the user input information is directly obtained from the words with a subject-object relationship. If not, all verbs or nouns marked as objects in the dependency set are used as a candidate object set, and the candidate object set is output to the user to manually determine the object. If the determined object is a verb, it is marked as an intent label; if the determined object is a noun, it is marked as an entity label.
[0060] S405: Output all words with dependency relationships in the dependency relationship set as a candidate relationship set to the user for confirmation, and mark the relationship confirmed by the user as a relationship label.
[0061] S406: Exclude the subject and object determined by the user from the dependency set, output all nouns from the excluded dependency set as candidate attribute sets to the user for confirmation, and mark the attributes confirmed by the user as attribute tags.
[0062] In an embodiment of the present invention, a candidate subject set, a candidate object set, a candidate relationship set, and / or a candidate attribute set are generated through a dependency set, and output to the user to determine the entity label, intent label, attribute label, and / or relationship label, thereby reducing the difficulty of manual labeling and improving the efficiency of manual labeling.
[0063] Embodiment 5
[0064] like Figure 5 As shown, the embodiment of the present invention is illustrated by taking the dynamic training of the intention recognition model in the football field as an example. For the sake of simplicity, the dynamic training of the intention recognition model in the football field is specified. It should be noted that the present invention is also applicable to the dynamic training of the intention recognition model in other fields.
[0065] S501. Obtain all information about leagues and teams related to football, and construct a standard question-and-answer database based on the information.
[0066] Specifically, all information about football-related leagues and teams can be captured from the Internet through crawlers or other technologies, and a standard question-answer database can be constructed based on the information.
[0067] S502: Select typical leagues and teams from all leagues and teams, and manually annotate their information to generate original training corpus.
[0068] Specifically, for example, if we select the Barcelona team in the La Liga, the specific implementation method is as follows:
[0069] First, all attributes of the players in the team set are determined, such as the team attributes include team name, team city, team home, team establishment time, team head coach, etc. The attributes of the players include player name, player number, player height, player weight, player value, player nationality, etc. In addition, there is an ownership relationship between the player and the team. Then, according to the corpus in step S501, the team name, the player name, the determined team and player attributes, and the ownership relationship between the player and the team, the entity label, attribute label and relationship label are respectively marked.
[0070] S503: Perform dynamic training based on the generated training corpus and generate an initial intent recognition model.
[0071] S504: After receiving the user's inquiry information, identify the user's intention based on the initial intention recognition model and feature template.
[0072] Among them, the initial intention recognition model corresponds to an initial feature template for identifying the user's question. For example, the user asks "What is the value of a player from the Barcelona team?" If the user's intention can be successfully identified based on the initial intention recognition model and the feature template, then go to step S506; otherwise, go to step S507.
[0073] S505, determine whether the recognition is successful, if successful, proceed to step S506; otherwise, proceed to step S507.
[0074] S506. Automatically annotate the user's input information and recognition results and use them as additional training corpus. When the length of the additional corpus exceeds a preset threshold, dynamically train the initialized intent recognition model again to generate a new intent recognition model.
[0075] Specifically, the automatic dynamic training submodule will automatically annotate the user's input information and recognition results and use them as additional training corpus. It will also determine whether to dynamically train the existing intent recognition model again based on the preset length threshold of the additional corpus, and dynamically update the feature template used to recognize user questions; it will dynamically train the initialized intent recognition model again to generate a new intent recognition model with higher accuracy.
[0076] S507. Annotate the user's input information through automatic assisted manual annotation, and use the annotation results as additional training corpus. When the length of the additional corpus exceeds a preset threshold, train the initialized intent recognition model to obtain a new intent recognition model, and synchronously update the feature template.
[0077] In the embodiment of the present invention, the dynamic training of the intention recognition model in the football field is specified only for the sake of simplicity. It should be noted that the present invention is also applicable to the dynamic training of the intention recognition model in other fields.
[0078] Embodiment 6
[0079] like Figure 6 As shown, an embodiment of the present invention provides a dynamic training device for an intent recognition model, which includes an input module 10, a recognition module 20, an automatic labeling module 30, an auxiliary manual labeling module 40, and a training module 50.
[0080] An input module 10, for receiving input information to be identified;
[0081] The recognition module 20 is used to perform intent recognition on the input information according to the existing intent recognition model after receiving the input information to be recognized;
[0082] The automatic annotation module 30 is used to automatically annotate the input information and the intention recognition result after the recognition is successful, and use them as additional training corpus.
[0083] The auxiliary manual annotation module 40 is used to automatically assist in manual annotation of the input information when recognition fails, and use the annotation results as additional training corpus.
[0084] The training module 50 is used to dynamically train the existing intent recognition model again according to the training corpus to generate a new intent recognition model when the length of the additional corpus exceeds a preset threshold.
[0085] Specifically, the input module 10 receives the user's input request and sends the user's input information to the recognition module 20. The intention recognition module 20 recognizes the user's behavioral intention based on the existing intention recognition model. When the recognition is successful, the corresponding user input and recognition result are automatically marked and output to the training module 50 as additional training corpus.
[0086] When recognition fails, the auxiliary manual annotation module 40 automatically assists in manual annotation of the input information, performs word segmentation on the user input information that fails to be recognized, performs dependency syntactic analysis on the word segmentation results, uses the dependency relationship between words in the sentence to represent the syntactic structure information of the words (such as subject-predicate, verb-object, attributive, etc.), and uses a tree structure to represent the structure of the entire sentence (such as subject-predicate-object, attributive-adverbial-complement, etc.), and generates a dependency relationship set, which mainly includes the word part of speech and the relationship between words (such as: master-slave relationship, master-slave-passive relationship, verb-object relationship, parallel relationship, correlation relationship, etc.); the dependency syntactic analysis can be directly analyzed using the existing Harbin Institute of Technology LTP. The user input information corresponding to the recognition failure and the annotation results are output to the training module 50 as additional training corpus.
[0087] The training module 50 is used to dynamically train the existing intent recognition model again based on the training corpus to generate a new intent recognition model when the length of the additional corpus exceeds a preset threshold. This includes retraining the existing intent recognition model based on automatically annotated corpus and manually annotated corpus. It should be noted that the intent recognition system loads the intent recognition model by hot loading, that is, the dynamic training of the intent recognition model will not affect the recognition of the intent recognition system. In addition, the specific method of dynamically training the intent recognition model can be based on existing statistical learning methods or existing machine learning methods.
[0088] It should be noted that the dynamic training device for the above-mentioned intention recognition model and the dynamic training method embodiments for the intention recognition model in Examples 1 to 5 belong to the same concept, and the specific implementation process is detailed in the method embodiments, and the technical features in the method embodiments are correspondingly applicable in the device embodiments, which will not be repeated here.
[0089] Embodiment 7
[0090] An embodiment of the present invention also provides a dynamic training device for an intent recognition model, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the steps of the dynamic training method for the intent recognition model of embodiments one to five are implemented.
[0091] In addition, an embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the dynamic training method of the intent recognition model of Embodiments 1 to 5 are performed.
[0092] It should be noted that the dynamic training device and computer-readable storage medium of the above-mentioned intent recognition model belong to the same concept as the embodiment of the dynamic training method of the intent recognition model. The specific implementation process is detailed in the method embodiment, and the technical features in the method embodiment are correspondingly applicable in the device embodiment, which will not be repeated here.
[0093] The dynamic training method, apparatus, device and storage medium of the intent recognition model provided in the embodiments of the present invention can effectively utilize diversified inputs to dynamically train and update the existing intent recognition model without increasing the length of the annotated corpus, thereby making up for the deficiency that the existing intent recognition model requires a large amount of annotated corpus for dynamic training, improving the accuracy of the intent recognition model and minimizing the dynamic training cost of the intent recognition model.
[0094] Those skilled in the art will appreciate that all or some of the steps in the methods disclosed above, and the functional modules / units in the systems and devices may be implemented as software, firmware, hardware, or a suitable combination thereof.
[0095] In hardware implementation, the division between the functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed by several physical components in cooperation. Some physical components or all physical components may be implemented as software executed by a processor, such as a central processing unit, a digital signal processor or a microprocessor, or implemented as hardware, or implemented as an integrated circuit, such as an application-specific integrated circuit. Such software may be distributed on a computer-readable medium, which may include a computer storage medium (or non-transitory medium) and a communication medium (or temporary medium). As known to those of ordinary skill in the art, the term computer storage medium includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disk (DVD) or other optical disk storage, magnetic cassettes, magnetic tapes, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, it is well known to those skilled in the art that communication media typically embodies computer readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.
[0096] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but the scope of the present invention is not limited thereby. Any modification, equivalent substitution and improvement made by those skilled in the art without departing from the scope and essence of the present invention shall be within the scope of the present invention.
Claims
1. A dynamic training method for an intent recognition model, the method comprising: After receiving the input information to be identified, performing intent identification on the input information according to the existing intent identification model; When the recognition is successful, the semantic analysis result of the input information obtained by the intent recognition and the intent recognition result are obtained, wherein the semantic analysis result includes a word segmentation result, an entity recognition result, an attribute extraction result and / or a relationship extraction result; Automatically label entity labels, attribute labels and / or relationship labels according to the semantic analysis results; Label the intent recognition results as intent labels; Use the annotation results as additional training corpus; In response to the length of the additional corpus exceeding a preset threshold, the existing intent recognition model is dynamically trained according to the training corpus to generate a new intent recognition model.
2. The dynamic training method of the intention recognition model according to claim 1 is characterized in that: The automatically labeling entity labels, attribute labels and / or relationship labels according to the semantic analysis results includes: According to the word segmentation results and entity recognition results, the words corresponding to the recognized entities are marked as entity tags; According to the word segmentation results and attribute extraction results, the words corresponding to the identified attribute names are marked as attribute labels; According to the word segmentation results and relationship extraction results, the words corresponding to the identified relationships are marked as relationship labels.
3. The dynamic training method of the intention recognition model according to claim 1 is characterized in that: After performing intent recognition on the input information according to the existing intent recognition model, the method further includes: When recognition fails, the input information is automatically assisted in manual annotation, and the annotation results are used as additional training corpus.
4. The dynamic training method of the intention recognition model according to claim 2 is characterized in that: The automatic auxiliary manual annotation of the input information specifically includes: Performing word segmentation processing on the input information to obtain a word segmentation result; Performing dependency syntactic analysis according to the word segmentation results to obtain a dependency relationship set; A candidate subject set, a candidate object set, a candidate relationship set, and / or a candidate attribute set are generated according to the dependency set, and output to the user to determine the entity label, intent label, attribute label, and / or relationship label.
5. The dynamic training method of the intention recognition model according to claim 4 is characterized in that: The generating of a candidate subject set, a candidate object set, a candidate attribute set and / or a candidate relationship set according to the dependency relationship set specifically includes: The nouns or personal pronouns marked as subjects in the dependency set are taken as candidate subject sets; The verbs or nouns marked as objects in the dependency set are taken as candidate object sets; All words with dependency relationships in the dependency relationship set are taken as candidate relationship sets; The subject and object determined by the user are excluded from the dependency set, and all nouns in the excluded dependency set are used as candidate attribute sets.
6. The dynamic training method of the intention recognition model according to claim 4, characterized in that: The output is provided to the user to determine entity tags, intent tags, attribute tags, and / or relationship tags, specifically including: Mark the subject determined by the user as an entity tag; When the object determined by the user is a verb, the object is marked as an intention label, and when the object determined by the user is a noun, the object is marked as an entity label; Labeling the relationship determined by the user as a relationship label; Labels user-defined attributes as attribute labels.
7. A dynamic training device for an intent recognition model, the device comprising: An input module, used for receiving input information to be identified; The recognition module is used to perform intent recognition on the input information according to the existing intent recognition model after receiving the input information to be recognized; An automatic annotation module is used to obtain semantic analysis results and intent recognition results of the input information obtained by intent recognition after successful recognition, wherein the semantic analysis results include word segmentation results, entity recognition results, attribute extraction results and / or relationship extraction results; The automatic labeling module is further used to automatically label entity labels, attribute labels and / or relationship labels according to the semantic analysis results; The automatic labeling module is also used to label the intent recognition results as intent labels; The automatic annotation module is also used to use the annotation results as additional training corpus. The training module is used to dynamically train the existing intent recognition model according to the training corpus in response to the length of the additional corpus exceeding a preset threshold, so as to generate a new intent recognition model.
8. A dynamic training device for an intent recognition model, characterized in that: The device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the steps of the dynamic training method of the intent recognition model as described in any one of claims 1 to 6 are implemented.
9. A storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by the processor, the steps of the dynamic training method of the intent recognition model as described in any one of claims 1 to 6 are implemented.
