A method and system for voice processing with flexible reuse of nodes using a multi-task model

Abstract

There is provided a method and system for artificial intelligence voice processing with flexible reuse of nodes using a multi-task model. The method comprises: inputting all corpus training data and performing clustering analysis operations based on clustering algorithm to obtain corpus training data that meets preset category rules in each category, and use it as training set; obtaining first processing flow structure based on first voice requirement; obtaining second processing flow structure based on second voice requirement, characterized in that each non-terminal node of the first processing flow structure and the second processing flow structure is used as a sub-task for multi- task training; initializing intent matching model template, which comprises task-specific layers corresponding to each sub-task in the processing flow structure;training the initialized intent matching model template using corpus from the first processing flow structure in the training set to obtain first intent matching model;updating the task- specific layers in the initialized first intent matching model that are different from the sub-tasks of the first processing flow structure by training with corpus from the second processing flow structure in the training set to obtain second intent matching model;performing artificial intelligence voice processing based on the second intent matching model.

Description

[0001] A METHOD AND SYSTEM FOR VOICE PROCESSING WITH FLEXIBLE REUSE OF NODES USING A MULTI-TASK MODEL

[0002] FIELD OF INVENTION

[0003] The present invention relates to a method and system for voice processing and specifically relates to a voice processing with flexible reuse of nodes using a multi-task model.

[0004] BACKGROUND

[0005] With the continuous increase in operating costs, telephone customer service centers are also following the general trend of transitioning from "cost centers to profit centers." When a business needs to communicate or notify many clients, manually contacting each one individually is not only time-consuming and laborious but also inefficient.

[0006] In recent years, significant advancements in Al technology has led to widespread application across various industries. For example, in the field of outbound telephone calls, Al has enabled efficient automated operations by integrating computer technology and voice technology. Using an outbound telephone call system, automated notifications can be sent. After automatically detecting that the callee has picked up the phone, the system can be configured to send pre-defined voice or text notifications. Upon detecting that the callee has hung up, the system can automatically proceed to contact the next person with no down time, making the process very convenient and efficient. Al voice processing systems effectively address the issues of high service-side costs and low efficiency. Corporations can pre-define the required outbound voice call processes based on business needs, and the system can implement requisite processes such as user identity verification, business consultation, and business processing.

[0007] Currently, outbound telephone call systems mainly rely on rule-based intent recognition methods, which require manual configuration of corresponding rules for each dialogue node. The disadvantage of such methods is lack of flexibility and accuracy. Specifically, rule-based intent recognition methods typically have a cumbersome configuration process and also often suffer from low recognition accuracy. Whenever there is a slight variation in user dialogue, the system may fail to accurately recognize it, leading to misjudgment or communication difficulties.

[0008] Furthermore, it is currently not possible to solve the problem of different meanings of the same text at different nodes within the same dialogue process, which greatly limits the intelligence and practicality of the system. At the same time, rule-based intent recognition methods also suffer from difficulties in reuse and inability to update in realtime.

[0009] SUMMARY

[0010] In a first aspect, there is provided a method for artificial intelligence voice processing with flexible reuse of nodes using a multi-task model, comprising: inputting all corpus training data and performing clustering analysis operations based on clustering algorithm to obtain corpus training data that meets preset category rules in each category, and use it as training set; obtaining first processing flow structure based on first voice requirement; obtaining second processing flow structure based on second voice requirement, where each non-terminal node of the first processing flow structure and the second processing flow structure is used as a sub-task for multi-task training; initializing intent matching model template, which comprises task-specific layers corresponding to each sub-task in the processing flow structure; training the initialized intent matching model template using corpus from the first processing flow structure in the training set to obtain first intent matching model; updating the task-specific layers in the initialized first intent matching model that are different from the sub-tasks of the first processing flow structure by training with corpus from the second processing flow structure in the training set to obtain second intent matching model; performing artificial intelligence voice processing based on the second intent matching model. In a second aspect, there is provided a system for artificial intelligence voice processing with flexible reuse of nodes using a multi-task model, characterized in that the system comprises: first obtaining module for obtaining corpus training data that meets preset category rules in each category and using it as training set; second obtaining module for obtaining first processing flow structure based on first voice requirement; third obtaining module for obtaining second processing flow structure based on second voice requirement; first initialization module for initializing intent matching model template, which comprises task-specific layers corresponding to each sub-task in the processing flow structure; fourth obtaining module for training the initialized intent matching model template using corpus from the first processing flow structure in the training set to obtain first intent matching model; fifth obtaining module for updating the task-specific layers in the initialized first intent matching model that are different from the sub-tasks of the first processing flow structure by training with corpus from the second processing flow structure in the training set to obtain second intent matching model; and first outbound module for performing artificial intelligence voice processing based on the second intent matching model.

[0011] In a final aspect, there is provided a system for artificial intelligence voice processing with flexible reuse of nodes using a multi-task model, comprising a memory, a processor, and a computer program stored on the memory and runnable on the processor, where the processor executes the program to implement the steps of any one of the aforementioned methods.

[0012] It will be appreciated that the broad forms of the invention and their respective features can be used in conjunction, interchangeably and / or independently, and reference to separate broad forms is not intended to be limiting.

[0013] DESCRIPTION OF FIGURES By reading the detailed description of exemplary embodiments in the following text, those skilled in the art will understand the advantages and benefits described herein, as well as other advantages and benefits. The accompanying drawings are only for the purpose of demonstrating exemplary embodiments and are not considered a limitation on the application. And throughout all drawings, the same components are represented by the same numbers. In the attached drawings:

[0014] FIG 1 illustrates a flowchart for an embodiment of the application illustrating the method for artificial intelligence voice processing with flexible reuse of nodes using a multi-task model;

[0015] FIG 2 illustrates a flowchart for an embodiment of the application illustrating optimizing the intent matching model template by minimizing the difference between the maximum intra-class distance and the minimum inter-class distance to obtain first intent matching model;

[0016] FIG 3 illustrates a flowchart for an embodiment of the application illustrating training and updating the task-specific layer corresponding to the first intent matching model to obtain second intent matching model;

[0017] FIG 4 illustrates a flowchart for an embodiment of the application illustrating performing Al voice processing based on the first prediction result;

[0018] FIG 5 illustrates a flowchart for an embodiment of the application illustrating allocating training data and corresponding intent labels for each sub-task;

[0019] FIG 6 illustrates a schematic diagram illustrating a data modeling system according to an embodiment of the application; and

[0020] FIG 7 illustrates a schematic diagram illustrating the structure of an exemplary electronic device according to an embodiment of the application.

[0021] DETAILED DESCRIPTION

[0022] Exemplary embodiments of the present application will be described in more detail below with reference to all accompanying drawings. Although the exemplary embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments described here. On the contrary, these embodiments are provided to enable a more thorough understanding of the present application and to fully convey the scope of the present application to those skilled in the art.

[0023] In the description of the embodiments of the application, it should be understood that terms such as "including" or "having" are intended to indicate the presence of disclosed features, numbers, steps, actions, components, parts, or combinations thereof in this specification, and do not exclude the possibility of the existence of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0024] Unless otherwise specified, " / " represents the meaning of "or". For example, A / B can represent A or B; "and / or" in this document is just a way to describe the relationship between associated objects, indicating that there can be three types of relationships. For example, A and / or B can represent three situations: the existence of A alone, the simultaneous existence of A and B, and the existence of B alone.

[0025] Terms such as "first", "second", and the like are used to distinguish between identical or similar technical features for descriptive convenience only and should not be interpreted as indicating or implying the relative importance or quantity of these technical features. Thus, features defined by "first", "second", etc., can explicitly or implicitly include one or more of these features. In the description of the embodiments of the application, unless otherwise specified, the term "multiple" means two or more.

[0026] It should also be noted that, without conflict, the embodiments and features in the embodiments in the application can be combined with each other. The following will refer to the accompanying drawings and combine embodiments to illustrate the application in detail.

[0027] In existing rule-based intent recognition technologies, cumbersome rules need to be manually set for each dialogue node. However, this approach is rigid and inefficient, resulting in a significantly reduced recognition accuracy. Once there is a slight variation in the customer's dialogue, the system may typically misjudge and fail to handle situations where the same text has different meanings under different dialogue nodes. Furthermore, this method has poor reusability and cannot achieve real-time updates, severely restricting the efficiency and flexibility of intelligent marketing. An embodiment of the application introduces a multi-task intent classification model to achieve an innovative solution. This model is configured to allow free configuration of the required corpus and labels at different nodes in the processing flow, ensuring that the intent labels at each node do not interfere with each other. By directly updating the prototype corpus under each node, the model can be configured to be easily and instantly updated and configured, thereby significantly enhancing the reusability of the model.

[0028] The technical solution in an exemplary embodiment of the application will be described clearly and in detail with reference to the accompanying drawings. Obviously, the described embodiment is only some of the embodiment of the application, rather than all of them. It should be understood that the application is not limited by the exemplary embodiment described. All other embodiments obtained by those skilled in the art without creative input based on the embodiment of the application falls within the scope of the application.

[0029] Embodiment 1

[0030] As shown in FIG 1 , an embodiment of the application provides a method 98 for artificial intelligence voice processing with flexible reuse of nodes using a multi-task model, where the method 98 is applied to a relevant system. The method 98 comprises:

[0031] At step S100, inputting all corpus training data and performing clustering analysis operations based on clustering algorithm to obtain corpus training data that meets preset category rules in each category, and using it as training set.

[0032] The clustering analysis operation, also known as cluster analysis, is a statistical analysis method based on the principle of similarity, widely used in research, sample classification, and indicator analysis. In the scenario of the application, clustering analysis is used to process corpus training data to optimize the training effect of the model. Before obtaining all the corpus training data, the domain direction to which the system will be applied is first determined. After determination, all the corpus training data is obtained, and then the obtained corpus is subjected to word segmentation processing to divide long texts into several words. After the text is segmented into words, it needs to be further converted into vectors. All vocabulary contained in the corpus training data is constructed into a term list without repeated terms, and then a vector is constructed for each text. The dimension of the vector is the same as the dimension of the term list, and the value of the vector is the number of times each term in the term list appears in the text. Then, it is assigned to the most similar cluster based on its similarity to the reference center. After the assignment is completed, the corpus training data that meets the preset category rules in each category is obtained and used as training set. The obtained training set is the basic data for model training and setting up dialogue nodes and intent recognition. The obtaining of the training set is the foundation for building the system for artificial intelligence voice processing with flexible reuse of nodes using a multi-task model.

[0033] At step S200, obtaining first processing flow structure based on first voice requirement.

[0034] At step S300, obtaining second processing flow structure based on second voice requirement, where each non-terminal node of the first processing flow structure and the second processing flow structure is used as a sub-task for multi-task training.

[0035] Specifically, the first voice requirement includes selecting any scenario in the application domain of the system for artificial intelligence voice processing with flexible reuse of nodes using a multi-task model; the second voice requirement is similar to the first voice requirement, selecting any scenario different from the first voice requirement in the application domain of the system for artificial intelligence voice processing with flexible reuse of nodes using a multi-task model. Obtaining the first processing flow structure based on the first voice requirement comprises formulating a corresponding processing flow structure according to the specific business process of the first voice requirement; obtaining the second processing flow structure based on the second voice requirement comprises formulating a corresponding processing flow structure according to the specific business process of the second voice requirement. The processing flow structure comprises terminal nodes and nonterminal nodes. The terminal nodes in the processing flow structure are the endpoints of the processing flow structure, corresponding to the end of the business process in the corresponding business scenario; non-term inal nodes are the sub-business processes in the corresponding scenario and are also the sub-tasks for multi-task training in the multi-task model to be constructed. The first processing flow structure and the second processing flow structure are obtained to make the system for artificial intelligence voice processing with flexible reuse of nodes using a multi-task model more standardized in application; multiple sub-tasks corresponding to non-terminal nodes are for handling various business modes that may occur in the corresponding business scenario, making the business process more flexible.

[0036] At step S400, initializing intent matching model template. The intent matching model template comprises task-specific layers corresponding to each sub-task in the processing flow structure.

[0037] The intent matching model template is a model framework designed for multiple business scenarios with similarities. Although the tasks in these business scenarios have some similarities, the degree of similarity is not high. The core goal of the multitask matching model is to effectively enhance the execution performance and accuracy of each business scenario by mining and leveraging the shared knowledge and characteristics among these different business scenarios to achieve collaborative assistance among various scenarios. This templated matching model not only improves the versatility and flexibility of the model but also enables the model to better adapt to diverse business needs and scenario changes. Initializing the intent matching model template refers to formulating a corresponding intent matching model template based on the first voice requirement and the second voice requirement before training the intent matching model template. The intent matching model template comprises a number of sub-tasks, which is the number of task-specific layers. The intent matching model template is set based on specific scenarios and will become more intelligent after training with the training set, able to perform corresponding operations based on the business content it processes.

[0038] At step S500, training the initialized intent matching model template using the corpus from the first processing flow structure in the training set to obtain first intent matching model. Training the initialized intent matching model template using the corpus from the first processing flow structure in the training set comprises randomly selecting any number of objects from the already processed training set as initial reference centers and calculating the remaining unselected objects. Based on their similarity to the selected objects, i.e. , the distance to the reference center, they are assigned to the most similar cluster. That is, the corpus of the obtained training set is classified based on similarity. If there is a more similar reference center compared to the currently assigned distance class center after the assignment, reclassification is performed based on the similarity principle until the classification result is relatively optimal. According to the obtained classification result of the training set, feature extraction is performed on the corpus training data to obtain corresponding feature vectors. Different feature vectors correspond to different task-specific layers, enabling intent matching for different texts, and the first intent matching model is obtained. The first intent matching model can derive the corresponding expression intent based on the input corpus.

[0039] At step S600, updating the task-specific layers in the initialized first intent matching model that are different from the sub-tasks of the first processing flow structure by training with corpus from the corpus of the second processing flow structure in the training set to obtain second intent matching model.

[0040] Using the corpus of the second processing flow structure in the training set, training and updating the task-specific layers in the initialized first intent matching model that are different from the sub-task of the first processing flow structure. Randomly selecting a non-term inal node in the second processing flow structure as the sub-task for this training, then using the data corresponding to this sub-task as the input data for this training task, and distinguishing multiple categories of intents in the training set and select multiple texts as the data required for training. Preprocessing the input training texts to obtain the feature vectors, then calculating the distances between each feature vector, and obtaining the maximum and minimum distances of each feature vector in its respective class, i.e., the maximum intra-class distance and the minimum inter-class distance. Optimizing the model by calculating the difference between the maximum intra-class distance and the minimum inter-class distance, reselect new sub-task nodes and corresponding training data, and repeat until an ideal difference is obtained. End the optimization process to obtain the second intent matching model. The first multi-task intent model and the second multi-task intent model can extract common features among tasks in various scenarios by training multiple related tasks, and then transmit these common features into different subtask scenarios for calculation, refining the process handling and text processing results for each scenario.

[0041] At step S700, performing artificial intelligence voice processing based on the second intent matching model.

[0042] The second intent matching model is implemented based on the first intent matching model. The second intent matching model can configure the required corpus for different business scenarios, i.e., different task nodes, to complete the business processes in the corresponding scenarios. Each execution of a business also updates the model, eliminating the need for manual periodic updates. Therefore, the second intent matching model has higher reusability and can be used for artificial intelligence voice processing.

[0043] Furthermore, as shown in FIG 2, step S500 also comprises:

[0044] At step S501 , obtaining the first sub-task in the first processing flow structure;

[0045] At step S502, randomly selecting intent categories from the first sub-task, and selecting texts as the first training set for each intent category;

[0046] At step S503, inputting all training texts of the first training set into the corresponding task-specific layer in the intent matching model template to obtain the first feature vector matrix;

[0047] At step S504, calculating the Euclidean distance between each pair of feature vectors in the first feature vector matrix;

[0048] At step S505, for each category, calculating the maximum intra-class distance and the minimum inter-class distance; At step S506, obtaining the difference between the maximum intra-class distance and the minimum inter-class distance for each category;

[0049] At step S507, optimizing the intent matching model template by minimizing the difference between the maximum intra-class distance and the minimum inter-class distance to obtain the first intent matching model.

[0050] Obtaining the first sub-task in the first processing flow structure means randomly selecting a non-terminal node from the first processing flow structure as the first task. Randomly selecting intent categories from the first sub-task comprises selecting specific data from the included data as the intent classification for the first sub-task, and selecting intent labels as the first training set for each intent category, which means selecting intent labels as training data for each selected intent category, i.e., selecting texts as the first training set. Inputting the first training set into a shared encoder results in vectors of length d. Inputting all training texts of the first training set into the corresponding task-specific layer in the intent matching model template comprises inputting all training texts of the obtained first training set into the taskspecific layer corresponding to the first sub-task, obtaining feature vectors of length d processed by the task-specific layer, and obtaining the first feature vector matrix. Then, calculate the Euclidean distance between each feature vector based on the obtained first feature vector matrix, i.e., calculate the distance between each feature vector and other vectors under the same intent category. For the calculated Euclidean distances between vectors in each category, determine the maximum intra-class distance and the minimum inter-class distance within each category. Then, obtain the difference between the maximum intra-class distance and the minimum inter-class distance for each category. Optimize the intent matching model template by minimizing the difference between the maximum intra-class distance and the minimum inter-class distance to obtain the first intent matching model. The first intent matching model is trained based on any business scenario and is suitable for business processing in that scenario.

[0051] Furthermore, step S507 also comprises: an optimization function of the first intent matching model being:

[0052] Wherein, ' / represents the optimization target difference of the intent matching model template; a represents the selected text vector number in the intent category; p represents the positive sample text vector number corresponding to text vector a; n represents the negative sample text vector number corresponding to text vector; i represents the category of the intent category;

[0053] D represents the distance between vectors;

[0054] Nc represents the number of intent categories;

[0055] Ns represents the number of samples contained in each intent category.

[0056] The optimization function of the first intent matching model comprises obtaining the maximum intra-class distance and the minimum inter-class distance in the category according to the first sub-task, and is constructed with the goal of shortening the maximum intra-class distance and the minimum inter-class distance. The optimization function of the first intent matching model can reduce the distance between samples belonging to the same category and increase the distance between samples belonging to different categories.

[0057] Furthermore, as shown in FIG 3, S600 also comprises:

[0058] At step S601 , obtaining the non-terminal nodes of the first processing flow structure. At step S602, obtaining the non-terminal nodes of the second processing flow structure.

[0059] At step S603, using the shared encoder of the first intent matching model as the shared encoder of the second intent matching model and updating the task-specific layers by training to obtain second intent matching model.

[0060] Some of the non-terminal nodes of the first processing flow structure and the second processing flow structure are the same, and the first intent matching model has been trained using the training set through the first processing flow structure. That is, the first intent matching model comprises a feature matrix and four task-specific layers. The second intent matching model uses the shared encoder of the first intent matching model for encoding, obtains the task-specific layers corresponding to the non-terminal nodes of the second processing flow structure, and trains and updates these taskspecific layers to obtain the second intent matching model. That is, the task-specific layers of the second intent matching model include six layers, and the model covers a more comprehensive scope and involves more detailed business content.

[0061] Furthermore, as shown in FIG 4, step S700 also comprises:

[0062] At step S701 , for each non-terminal node in the first processing flow structure and / or the second processing flow structure.

[0063] At step S702, inputting unrecognized data and the intent matching model template to obtain first prediction result.

[0064] At step S703, performing artificial intelligence voice processing based on the first prediction result.

[0065] Based on the obtained non-terminal nodes in the first processing flow structure and / or the second processing flow structure, the obtained non-terminal nodes may belong to a common sub-task in both the first processing flow structure and the second processing flow structure, or they may belong to a sub-task in either the first processing flow structure or the second processing flow structure. Based on the obtained non-term inal nodes and the unrecognized data, input them into the intent matching model template for intent matching. After calculation, obtain the intent of the unrecognized data, i.e., obtain the first prediction result. Perform outbound calls and follow up on business processes based on the obtained first prediction result.

[0066] Furthermore, as shown in FIG 5, S701 also comprises:

[0067] At step S7011 , inputting all text under the non-terminal node into the second intent matching model to obtain first feature vector.

[0068] At step S7012, inputting the unrecognized data into the second intent matching model to obtain second feature vector.

[0069] At step S7013, randomly selecting several pieces of data as prototype data for each category under the non-terminal node. The prototype data comprises text and corresponding intent labels.

[0070] At step S7014, calculating the Euclidean distance between the first feature vector and the second feature vector, and assign the intent label corresponding to the feature vector with the shortest Euclidean distance as the intent label of the unrecognized data, and return the intent label.

[0071] The first feature vector comprises feature vectors of all input text under this task node. Then, input the unrecognized data into the second intent matching model to obtain second feature vector. At the same time, select all data under the non-terminal node as prototype data for the second intent matching model. The prototype data comprises text and corresponding intent labels. Calculate the Euclidean distance between the first feature vector and the second feature vector based on the obtained text of the prototype data and the intent labels corresponding to the prototype data. Then, assign the intent label with the shortest calculated Euclidean distance as the intent label of the input unrecognized data, and return the obtained label of the unrecognized data to the second intent matching model for further calculation. By returning the intent label of the unrecognized data to the second intent matching model for further calculation, the intent label of the unrecognized data can be obtained, and the second intent matching model can perform business processing based on the intent label.

[0072] Compared with the prior art, the present invention has at least the following beneficial effects:

[0073] All input corpus training data is subjected to clustering analysis operations based on clustering algorithm to obtain corpus training data in each category that conforms to preset category rules, which is used as training set; first processing flow structure is obtained based on first voice requirement; second processing flow structure is obtained based on second voice requirement, where each non-terminal node of the first processing flow structure and the second processing flow structure is respectively used as a sub-task for multitask training; intention matching model template is initialized, and the intention matching model template comprises task-specific layers that respectively correspond to each sub-task in the processing flow structure; the initialized intention matching model template is trained using corpus from the first processing flow structure in the training set to obtain first intention matching model; the task-specific layers in the initialized first intention matching model that are different from the sub-tasks of the first processing flow structure are trained and updated using corpus of the second processing flow structure in the training set to obtain second intention matching model; artificial intelligence voice processing is performed based on the second intention matching model, achieving the technical effect of using a multitask intention classification model to realize free configuration of corpus and labels required for different nodes in the processing flow structure, such that intention labels under different nodes do not affect each other, and the model can be updated and configured by directly updating prototype corpus contained under each node, improving model reusability.

[0074] Embodiment 2

[0075] The application provides a system 6000 for artificial intelligence voice processing with flexible reuse of nodes using a multi-task model, as shown in FIG 6. The system 6000 comprises: first obtaining module 11 , which is configured to obtain corpus training data in each category that conforms to preset category rules and use it as training set; second obtaining module 12, which is configured to obtain first processing flow structure based on first voice requirement; third obtaining module 13, which is configured to obtain second processing flow structure based on second voice requirement; first initialization module 14, which is configured to initialize intention matching model template, the intention matching model template comprises task-specific layers that respectively correspond to each sub-task in the processing flow structure; fourth obtaining module 15, which is configured to train the initialized intention matching model template using corpus from the first processing flow structure in the training set to obtain first intention matching model; fifth obtaining module 16, which is configured to update the task-specific layers in the initialized first intention matching model that are different from the sub-tasks of the first processing flow structure by training with corpus from the second processing flow structure in the training set to obtain second intention matching model; first outbound module 17, which is configured to perform artificial intelligence voice processing based on the second intention matching model.

[0076] Furthermore, the system 6000 also optionally comprises: sixth obtaining module, which is configured to obtain first sub-task in the first processing flow structure; first selection module, which is configured to randomly select intention categories from the first sub-task, and select text as first training set from each intention category; seventh obtaining module, which is configured to input all training text in the first training set into the task-specific layer corresponding to the intention matching model template to obtain first feature vector matrix; first calculation module, which is configured to calculate the Euclidean distance between each pair of feature vectors in the first feature vector matrix; second calculation module, which is configured to calculate the maximum intra-class distance and the minimum inter-class distance for each category; eighth obtaining module, which is configured to obtain the difference between the maximum intra-class distance and the minimum inter-class distance for each category; first optimization module, which is configured to optimize the intention matching model template by minimizing the difference between the maximum intra-class distance and the minimum inter-class distance to obtain first intention matching model.

[0077] Furthermore, the system 6000 also optionally comprises: first function module, which is configured for the optimization function of the first intent matching model as:

[0078] Wherein, / represents the optimization target difference of the intent matching model template; a represents the selected text vector number in the intent category; p represents the positive sample text vector number corresponding to text vector a; n represents the negative sample text vector number corresponding to text vector; i represents the category of the intent category;

[0079] D represents the distance between vectors;

[0080] Nc represents the number of intent categories;

[0081] Ns represents the number of samples contained in each intent category; second category module, which is used for the intention category that the feature vectors belong to.

[0082] Furthermore, the system 6000 also optionally comprises: ninth obtaining module, which is configured to acquire non-terminal nodes of the first processing flow structure; tenth obtaining module, which is configured to acquire non-terminal nodes of the second processing flow structure; eleventh obtaining module, which is configured to train and update the task-specific layer corresponding to the first intent matching model by using the shared encoder of the first intent matching model as the shared encoder of the second intent matching model, and obtaining the second intent matching model.

[0083] Furthermore, the system 6000 also optionally comprises: second selection module, which is configured for each non-terminal node in the first processing flow structure and / or the second processing flow structure; twelfth obtaining module, which is configured to input unrecognized voice data and the input intent matching model template to obtain first prediction result; first prediction module, which is configured to perform artificial intelligence voice processing based on the first prediction result.

[0084] Furthermore, the system 6000 also optionally comprises: first input module, which is configured to input all texts a specific context into the second intent matching model to obtain first feature vector; second input module, which is configured to input the unrecognized voice data into the second intent matching model to obtain second feature vector; third selection module, which is configured to randomly select several pieces of voice data as prototype voice data for each category under the non-terminal node. The prototype voice data comprises text, as well as corresponding intent labels; third calculation module, which is configured to calculate the Euclidean distance between the first feature vector and the second feature vector, assign the intent label corresponding to the feature vector with the shortest Euclidean distance as the intent label of the unrecognized voice data, and return the intent label.

[0085] The variations and specific examples of the method for artificial intelligence voice processing with flexible reuse of nodes using a multi-task model in embodiment 1 , as described earlier in FIG 1 , are also applicable to the system for artificial intelligence voice processing with flexible reuse of nodes using a multi-task model in the embodiment 2. Through the detailed description of the method provided earlier, those skilled in the art can clearly understand the system in this embodiment. Therefore, further details will not be repeated here.

[0086] Embodiment 3

[0087] FIG 7 illustrates a structural diagram of an electronic device 50 according to an embodiment of the application. Referring to FIG 7, the electronic device 50 in this embodiment of the application can be a mobile device itself or a standalone device separate from it, which stores a computer program. When the program is executed by the electronic device 50, it implements any of the steps of the methods described earlier.

[0088] As shown in FIG 7, the electronic device 50 comprises one or more processors 51 and a memory 52.

[0089] The processor 51 can be a Central Processing Unit (CPU) or another form of processing module configured with data processing capabilities and / or instruction execution capabilities, and can be configured to control other components in the electronic device 50 to perform desired functions.

[0090] The memory 52 can include one or more computer program products, which can include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, Random Access Memory (RAM) and / or cache memory. The non-volatile memory may include, for example, Read-Only Memory (ROM), hard disks, flash memory, etc. One or more computer program instructions can be stored on the computer-readable storage media, and the processor 51 can execute these program instructions to implement the methods of various embodiment of the application described above and / or other desired functions.

[0091] In one example, the electronic device 50 may also include: an input device 53 and an output device 54, which are interconnected through a bus system and / or other forms of connection mechanisms (not shown).

[0092] The embodiment of the application provides a node flexible reuse artificial intelligence voice processing method using a multi-task model, characterized in that the method, comprises: inputting all corpus training data to perform clustering analysis operations based on clustering algorithm, obtaining the corpus training data that meets preset category rules in each category, and using it as training set; obtaining first processing flow structure based on first voice requirement; obtaining second processing flow structure based on second voice requirement, where each non-terminal node of the first processing flow structure and the second processing flow structure is used as a sub-task for multi-task training, respectively; initializing intent matching model template, where the intent matching model template comprises task-specific layers, each corresponding to each sub-task in the processing flow structure; training the initialized intent matching model template using corpus from the first processing flow structure in the training set to obtain first intent matching model; updating and training the task-specific layers in the initialized first intent matching model that are different from the sub-tasks of the first processing flow structure by training with corpus from corpus of the second processing flow structure in the training set to obtain second intent matching model; performing artificial intelligence voice processing based on the second intent matching model, achieving the technical effect of implementing free configuration of required corpus and labels for different nodes in the processing flow structure through a multi-task intent classification model, so that intent labels under different nodes do not affect each other, and the prototype corpus contained under each node can be directly updated to complete the model update and configuration, thereby improving the reusability of the model.

[0093] Through the description of the above implementation modes, technical personnel in the field can clearly understand that the application can be implemented by software plus necessary general hardware, and of course, it can also be implemented through special hardware, comprising application-specific integrated circuits, special CPUs, special memories, special components, etc. In general, any function completed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special circuits. However, for the application, software program implementation is a better implementation mode in more cases. Based on this understanding, the technical solution of the application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a readable storage medium, such as a computer's floppy disk, U disk, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and comprises several instructions for enabling a computer device to execute the method described in the embodiment of the application. In the above embodiment, all or part of the implementation can be achieved through software, hardware, firmware, or any combination of them. When implemented using software, it can be fully or partially implemented in the form of computer program products.

[0094] The computer program product comprises one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiment of the application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, or a magnetic tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a Solid State Disk (SSD)), etc.

[0095] It should be understood that the "an embodiment" or "an embodiment" mentioned throughout the specification means that specific features, structures, or characteristics related to the embodiment are included in at least one embodiment of the application. Therefore, the appearance of "in an embodiment" or "in an embodiment" in various places throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in one or more embodiment in any suitable manner. It should be understood that in the various embodiment of the application, the size of the sequence numbers of the above processes does not necessarily imply the order of execution, and the execution order of the processes should be determined by their functions and internal logic rather than limiting the implementation process of the embodiment of the application.

[0096] In addition, the terms "system" and "network" are often used interchangeably in this article. The term "and / or" in this article is merely used to describe the association relationship between associated objects and indicates that there can be three relationships, for example, A and / or B can indicate: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character in this article generally indicates an "or" relationship between the associated objects before and after it.

[0097] Ordinary technical personnel in this field can realize that the modules and algorithm steps of the examples described in conjunction with the embodiment disclosed in this article can be implemented through electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of their functions in the above description. Whether these functions are performed through hardware or software depends on specific applications and design constraints of the technical solutions. Professional technicians can use different methods for each specific application to implement the described functions, but such implementation should not be considered as the scope of the application.

[0098] In summary, what has been described above is only a preferred embodiment of the technical solution of the application, and is not used to limit the protection scope of the application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the application should be included within the protection scope of the application.

[0099] Throughout this specification and claims which follow, unless the context requires otherwise, the word “comprise”, and variations such as “comprises” or “comprising”, will be understood to imply the inclusion of a stated integer or group of integers or steps but not the exclusion of any other integer or group of integers.

[0100] Persons skilled in the art will appreciate that numerous variations and modifications will become apparent. All such variations and modifications which become apparent to persons skilled in the art, should be considered to fall within the spirit and scope that the invention broadly appearing before described.

Claims

CLAIMS1 . A method for artificial intelligence voice processing with flexible reuse of nodes using a multi-task model, comprises: inputting all corpus training data and performing clustering analysis operations based on clustering algorithm to obtain corpus training data that meets preset category rules in each category, and use it as training set; obtaining first processing flow structure based on first voice requirement; obtaining second processing flow structure based on second voice requirement, where each non-terminal node of the first processing flow structure and the second processing flow structure is used as a sub-task for multi-task training; initializing intent matching model template, which comprises task-specific layers corresponding to each sub-task in the processing flow structure; training the initialized intent matching model template using corpus from the first processing flow structure in the training set to obtain first intent matching model; updating the task-specific layers in the initialized first intent matching model that are different from the sub-tasks of the first processing flow structure by training with corpus from the second processing flow structure in the training set to obtain second intent matching model; performing artificial intelligence voice processing based on the second intent matching model.

2. The method according to claim 1 , characterized in that training the initialized intent matching model template using corpus from the first processing flow structure in the training set to obtain first intent matching model, comprises: obtaining first sub-task from the first processing flow structure; randomly selecting intent categories from the first sub-task and selecting text as first training set from each intent category; inputting all training texts in the first training set into the corresponding task-specific layer in the intent matching model template to obtain first feature vector matrix; calculating the Euclidean distance between each pair of feature vectors in the first feature vector matrix;for each category, calculating the maximum intra-class distance and the minimum inter-class distance; obtaining the difference between the maximum intra-class distance and the minimum inter-class distance for each category; optimizing the intent matching model template by minimizing the difference between the maximum intra-class distance and the minimum inter-class distance to obtain first intent matching model.

3. The method according to claim 2, characterized in that optimizing the intent matching model template by minimizing the difference between the maximum intra- class distance and the minimum inter-class distance to obtain first intent matching model, comprises: the optimization function of the first intent matching model is:wherein, L represents the optimization target difference of the intent matching model template; a represents the selected text vector number in the intent category; p represents the positive sample text vector number corresponding to text vector a; n represents the negative sample text vector number corresponding to text vector; i represents the category of the intent category;D represents the distance between vectors;Nc represents the number of intent categories;Ns represents the number of samples contained in each intent category.

4. The method according to claim 1 , characterized in that updating the task-specific layers in the initialized first intent matching model that are different from the sub-tasks of the first processing flow structure by training with corpus from the second processing flow structure in the training set to obtain second intent matching model, comprises:obtaining non-terminal nodes of the first processing flow structure; obtaining non-terminal nodes of the second processing flow structure; using the shared encoder of the first intent matching model as the shared encoder of the second intent matching model, and updating the task-specific layers by training to obtain second intent matching model.

5. The method according to claim 1 , characterized in that the method further, comprises: for each non-terminal node in the first processing flow structure and / or the second processing flow structure, inputting unrecognized corpus into the intent matching model template to obtain first prediction result; performing artificial intelligence voice processing based on the first prediction result.

6. The method according to claim 5, characterized in that the method further, comprises: inputting all texts into the second intent matching model to obtain first feature vector; inputting the unrecognized corpus into the second intent matching model to obtain second feature vector; randomly selecting several pieces of corpus as prototype corpus for each category under the non-terminal node, where the prototype corpus comprises text and corresponding intent labels; calculating the Euclidean distance between the first feature vector and the second feature vector, and using the intent label corresponding to the feature vector with the shortest Euclidean distance as the intent label of the unrecognized corpus, and returning the intent label; the number of non-terminal nodes in the processing flow structure is the same as the number of sub-tasks, allocating training data that will be used for each sub-task, where the training data comprises text and corresponding intent labels,.

7. A system for artificial intelligence voice processing with flexible reuse of nodes using a multi-task model, characterized in that the system comprises: first obtaining module for obtaining corpus training data that meets preset category rules in each category and using it as training set;second obtaining module for obtaining first processing flow structure based on first voice requirement; third obtaining module for obtaining second processing flow structure based on second voice requirement; first initialization module for initializing intent matching model template, which comprises task-specific layers corresponding to each sub-task in the processing flow structure; fourth obtaining module for training the initialized intent matching model template using corpus from the first processing flow structure in the training set to obtain first intent matching model; fifth obtaining module for updating the task-specific layers in the initialized first intent matching model that are different from the sub-tasks of the first processing flow structure by training with corpus from the second processing flow structure in the training set to obtain second intent matching model; and first outbound module for performing artificial intelligence voice processing based on the second intent matching model.

8. A system for artificial intelligence voice processing with flexible reuse of nodes using a multi-task model, comprising a memory, a processor, and a computer program stored on the memory and runnable on the processor, where the processor executes the program to implement the steps of any one of the methods according to claims 1-6.

Images