User task processing method and device based on multi-expert model, equipment and medium

Abstract

The invention discloses a user task processing method and device based on a multi-expert model, equipment and a medium. The method comprises the steps of obtaining a total task of a target user according to multi-modal data; decomposing the general task to obtain a plurality of sub-tasks; constructing a multi-expert model; and for each sub-task, executing steps S141 to S144: step S141, controlling the gating network unit to determine the confidence corresponding to each vertical expert network processing sub-task; step S142, determining a plurality of target vertical expert networks; step S143, adjusting the target vertical expert network; step S144, obtaining a multi-dimensional processing result of the subtask based on a plurality of target vertical expert networks; and obtaining a processing result of the total task of the target user according to the plurality of multi-dimensional processing results. The invention belongs to the field of task processing. According to the invention, the result hit accuracy can be improved.

Description

Technical Field

[0001] The present invention relates to the field of task processing, and in particular to a method, device, equipment and medium for user task processing based on a multi-expert model. Background Art

[0002] When dealing with complex tasks, large models usually use deep learning architectures (such as Transformer) for hierarchical feature extraction, and adapt to specific task requirements through pre-training and fine-tuning. At the same time, data enhancement, regularization techniques (such as Dropout), and optimization algorithms (such as Adam) are used to improve the generalization ability of the model, and distributed training and hardware acceleration (such as GPU, TPU) are used to improve training efficiency.

[0003] However, a single model may not achieve optimal performance when processing diverse tasks. Therefore, how to improve the accuracy of processing results is an urgent problem to be solved. Summary of the invention

[0004] The present invention solves the technical problem in the prior art that a single model may not achieve optimal performance when processing diversified tasks by providing a user task processing method, device, equipment and medium based on a multi-expert model, thereby achieving the technical effect of improving the accuracy of processing result hits.

[0005] In a first aspect, the present invention provides a method for processing user tasks based on a multi-expert model, the method comprising: Obtain multimodal data of the target user, and obtain the total task of the target user based on the multimodal data; Decompose the total task into several subtasks; Constructing a multi-expert model, wherein the multi-expert model includes a gated network unit and a vertical expert network cluster, wherein the vertical expert network cluster includes a plurality of vertical expert networks, each vertical expert network corresponds to a preset field and all the vertical expert networks are open networks; For each subtask, steps S141-S144 are executed, including: Step S141, controlling the gated network unit to determine the confidence corresponding to each vertical expert network processing subtask; Step S142, determining a number of target vertical expert networks according to the confidence of each vertical expert network; Step S143, adjusting the target vertical expert network according to the dynamic processing information of the target vertical expert network or the correlation between the target vertical expert networks, wherein the dynamic information includes the running state and the processing state; Step S144, obtaining a multi-dimensional processing result of the subtask based on a plurality of target vertical expert networks; Based on several multi-dimensional processing results, the processing result of the total task of the target user is obtained.

[0006] Furthermore, the control gate network unit determines the confidence corresponding to each vertical expert network processing subtask, including: Perform feature preprocessing and dimension expansion on the initial features of the subtask to obtain expanded fusion features; The expanded fusion features are mapped to the same dimensional space as the number of vertical expert networks through the dimensionality reduction layer, and the confidence distribution is generated based on the Softmax function. The confidence distribution includes the confidence of each vertical expert network. The confidence distribution is adjusted based on the historical processing capability of the vertical expert network, where the historical processing capability includes the historical accuracy rate and the processing time limit compliance rate.

[0007] Furthermore, based on the historical processing capability of the vertical expert network, the confidence distribution is adjusted, including:

[0008] in, After adjustment The confidence of a vertical expert network, Before adjustment The confidence of a vertical expert network, is the historical accuracy, It is the processing time limit achievement rate.

[0009] Furthermore, according to the confidence of each vertical expert network, several target vertical expert networks are determined, including: According to the confidence distribution, the confidence density is determined; Set a reference value for the number of expert networks, and determine the number G of target vertical expert networks based on the confidence density of the vertical expert network cluster and the reference value for the number of expert networks; The top G vertical expert networks are selected from the confidence distribution as the target vertical expert networks.

[0010] Further, including:

[0011] in, is the confidence density, is the number of vertical expert networks; Also includes:

[0012] in, is the number of target vertical expert networks and G is rounded up, is the reference value of the number of expert networks.

[0013] Furthermore, according to the dynamic processing information of the target vertical expert network or the correlation between the target vertical expert networks, the target vertical expert network is adjusted, including: According to the operation status of the target vertical expert network, determine whether the target vertical expert network is in a fault state or an occupied state; If so, the target vertical expert network will be eliminated and filled with the vertical expert networks that have not been selected in order; if neither of them is in the same situation, no elimination will be performed; Among several target vertical expert networks, determining a first expert network; Based on the preset knowledge graph, the degree of correlation between the first expert network and the remaining target vertical expert networks is determined, and the remaining target vertical expert networks are sorted according to the degree of correlation.

[0014] Furthermore, based on several target vertical expert networks, the multi-dimensional processing results of this subtask are obtained, including: Based on a number of target vertical expert networks, the subtask is processed to obtain a first single-dimensional processing result of each target vertical expert network; According to the ranking of the target vertical expert networks, the result weights corresponding to each target vertical expert network are determined; According to the result weights corresponding to each target vertical expert network and the corresponding first result of the single-dimensional processing, a second result of the single-dimensional processing is obtained; According to the second results of the single-dimensional processing, a multi-dimensional processing result of the subtask is obtained.

[0015] In a second aspect, the present invention provides a user task processing device based on a multi-expert model, the device comprising: An acquisition module is used to acquire multimodal data of a target user and obtain the total task of the target user based on the multimodal data; Decomposition module, used to decompose the overall task into several subtasks; A construction module is used to construct a multi-expert model, wherein the multi-expert model includes a gated network unit and a vertical expert network cluster, the vertical expert network cluster includes a plurality of vertical expert networks, each vertical expert network corresponds to a preset field and the vertical expert networks are all open networks; The multi-dimensional processing module is used to execute steps S141-S144 for each subtask, including: step S141, controlling the gated network unit to determine the confidence corresponding to each vertical expert network processing subtask; step S142, determining a number of target vertical expert networks according to the confidence of each vertical expert network; step S143, adjusting the target vertical expert network according to the dynamic processing information of the target vertical expert network or the correlation between the target vertical expert networks, wherein the dynamic information includes the running state and the processing state; step S144, obtaining the multi-dimensional processing result of the subtask based on the several target vertical expert networks; The result module is used to obtain the processing result of the total task of the target user based on a number of multi-dimensional processing results.

[0016] In a third aspect, the present invention provides an electronic device, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to execute to implement the user task processing method based on the multi-expert model as provided in the first aspect.

[0017] In a fourth aspect, the present invention provides a non-temporary computer-readable storage medium. When the instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to implement the user task processing method based on a multi-expert model as provided in the first aspect.

[0018] One or more technical solutions provided in the present invention have at least the following technical effects or advantages: The present invention combines the historical performance (accuracy and timeliness) of experts with the relevance to the current task by comprehensively considering the historical accuracy rate and the processing time limit compliance rate, avoids the deviation of a single indicator, improves the reliability of decision-making, and achieves precise matching of task requirements and expert capabilities through time limit and accuracy indicators. Ultimately, it takes into account both response speed and resource utilization while ensuring high accuracy.

[0019] The present invention can optimize the configuration quantity of the target vertical expert network through the confidence density, improve the accuracy of result generation, and avoid the waste of computing resources.

[0020] The present invention first pre-sorts several target vertical expert networks by presetting the knowledge graph and confidence distribution, and assigns weights to each result according to the sorting results, so that the processing results corresponding to the subtasks are more in line with the actual situation and the answer hit rate is improved.

[0021] The present invention integrates the problems through the total (total task)-sub (sub-task)-total (total result) structure, and splits them for processing, and combines the results of the split processing as the total result, which can improve the hit rate of the results. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0023] Figure 1 A flowchart of a method for processing user tasks based on a multi-expert model provided by the present invention; Figure 2 This is a schematic diagram of the structure of a user task processing device based on a multi-expert model provided by the present invention. DETAILED DESCRIPTION

[0024] The embodiment of the present invention solves the technical problem in the prior art that a single model may not achieve optimal performance when processing diverse tasks by providing a user task processing method based on a multi-expert model.

[0025] The technical solution of the present invention is to solve the above technical problems, and the overall idea is as follows: A user task processing method based on a multi-expert model, the method comprising: obtaining multimodal data of a target user, and obtaining a total task of the target user according to the multimodal data; decomposing the total task to obtain a plurality of subtasks; constructing a multi-expert model, wherein the multi-expert model comprises a gated network unit and a vertical expert network cluster, wherein the vertical expert network cluster comprises a plurality of vertical expert networks, each vertical expert network corresponds to a preset field and the vertical expert networks are all open networks; for each subtask, executing steps S141-S144, comprising: step S141, controlling the gated network unit to determine the confidence corresponding to each vertical expert network processing subtask; step S142, determining a plurality of target vertical expert networks according to the confidence of each vertical expert network; step S143, adjusting the target vertical expert network according to the dynamic processing information of the target vertical expert network or the correlation between the target vertical expert networks, wherein the dynamic information comprises the running state and the processing state; step S144, obtaining a multidimensional processing result of the subtask based on the plurality of target vertical expert networks; and obtaining a processing result of the total task of the target user according to the plurality of multidimensional processing results.

[0026] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0027] First of all, the term "and / or" in this article is only a description of the association relationship of the associated objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0028] The present invention provides Figure 1 The user task processing method based on the multi-expert model shown includes steps S11-S15: Step S11, acquiring multimodal data of the target user, and obtaining the total task of the target user according to the multimodal data.

[0029] The target user refers to the user with the processing result to be obtained, and the multimodal data refers to the data coming from different sources or in different forms, which may include: text data, image data, audio data, video data, behavior data or sensor data, etc.

[0030] After obtaining the multimodal data, preprocessing can be performed on the multimodal data, including cleaning, alignment, labeling, and normalization. These are conventional data processing methods and will not be described in detail.

[0031] Then, in the feature extraction stage, multimodal data is merged and feature interaction is performed to obtain fused features (such as concatenating text word vectors with image CNN features and inputting them into LSTM), and a pre-trained model (BERT / XLNet) is used to perform intent classification on the fused features (for example, input fused features → fully connected layer → Softmax classification), identify task-critical information (such as date, location, amount), and obtain the task description based on the task-critical information, which is the overall task referred to in the present invention (for example, order a gluten-free Margherita pizza on August 24, 2024 and deliver it to the designated address A).

[0032] Step S12, decomposing the overall task into several subtasks.

[0033] After the overall task is obtained in step S11 , the description of the overall task may be analyzed based on natural language processing (NLP) technology to extract key information.

[0034] Take the above example: "Order a gluten-free Margherita pizza on August 24, 2024 and deliver it to the specified address A": Key information, including: Time: 2024.8.24 Pizza type: Margherita.

[0035] Special needs: Gluten free.

[0036] Delivery address: A.

[0037] Payment: Payment is required to complete.

[0038] Based on the extracted key information, the overall task is decomposed into several subtasks. Each subtask can correspond to one or more fields. For example, if image recognition is required, it is divided into the image recognition field, if text reasoning is required, it is divided into the text reasoning recognition field, and if speech processing is required, it is divided into the speech recognition field. The preset field can be set according to the actual situation and there is no restriction here.

[0039] Step S13, constructing a multi-expert model, wherein the multi-expert model includes a gated network unit and a vertical expert network cluster, the vertical expert network cluster includes a number of vertical expert networks, each vertical expert network corresponds to a preset field and the vertical expert networks are all open networks.

[0040] In practical applications, vertical expert network clusters are usually feed-forward networks, and the gating network determines which input data is sent to which vertical expert network.

[0041] The vertical expert network cluster referred to in the present invention includes several vertical expert networks, which may be domain-specific experts, data modality-specific experts, algorithm-specific experts, problem complexity-specific experts, user group-specific experts, geographic region-specific experts, etc.; Data modality-specific experts can include experts in specific fields. For example, a corresponding vertical expert network can be built for text, image, and audio respectively.

[0042] Each expert network is an independent neural network. The method in the embodiment of this specification can be applied to a server equipped with multiple expert models, or a business processing platform.

[0043] In practical applications, standardized interaction interfaces can be designed to facilitate interaction with vertical expert networks. For example, using RESTful API interfaces, a unified request and response format can be defined to facilitate integration between different systems.

[0044] Step S14, for each subtask, execute steps S141-S144, including: Step S141, controlling the gated network unit to determine the confidence level corresponding to each vertical expert network processing subtask.

[0045] Specifically include: The initial features of the subtasks are preprocessed and dimensionally expanded to obtain expanded fusion features. The expanded fusion features are mapped to the same dimensional space as the number of vertical expert networks through the dimensionality reduction layer, and a confidence distribution is generated based on the Softmax function, which includes the confidence of each vertical expert network. The confidence distribution is adjusted based on the historical processing capabilities of the vertical expert networks, where the historical processing capabilities include the historical accuracy rate and the processing time limit compliance rate.

[0046] In the multi-expert network system, the role of the gating network is to assign a confidence score to each vertical expert network to decide which vertical expert networks should be called when processing subtasks and how much weight to give them.

[0047] Feature preprocessing may include normalization, etc.

[0048] In order to better capture the feature information of the subtask, the preprocessed features need to be dimensionally expanded. This can be achieved through some nonlinear or linear transformations, such as using a multi-layer perceptron (MLP) or a convolutional neural network (CNN) (see also patent document CN118520903A). Dimensional expansion can increase the expressive power of features, allowing the model to learn more complex feature relationships.

[0049] The expanded fusion features are mapped to a dimensional space with the same number of vertical expert networks through a dimensionality reduction layer, and the confidence distribution is generated based on the Softmax function.

[0050] The dimension of the expanded fusion feature may be high, which is not conducive to subsequent calculation and analysis. Therefore, it is necessary to map it to the same dimensional space as the number of vertical expert networks through a dimensionality reduction layer. The dimensionality reduction layer can be implemented using a fully connected layer, which converts the high-dimensional feature vector into a vector equal to the number of vertical expert networks.

[0051] Based on the historical processing capabilities of the vertical expert network, the confidence distribution is adjusted, including:

[0052] in, After adjustment The confidence of a vertical expert network, Before adjustment The confidence of a vertical expert network, is the historical accuracy, It is the processing time limit achievement rate.

[0053] The historical accuracy rate refers to the accuracy rate of the vertical expert network in processing subtasks in this field in the past. For example, if 90 out of 100 tasks are accurate, the accuracy rate is 90%. The processing time limit compliance rate refers to the timeliness rate of completing tasks within the processing time limit when the vertical expert network processed subtasks in this field in the past. For example, if 80 out of 100 tasks are completed within the time limit, the accuracy rate is 80%.

[0054] The present invention combines the historical performance (accuracy and timeliness) of experts with the relevance to the current task by comprehensively considering the historical accuracy rate and the processing time limit compliance rate, avoids the deviation of a single indicator, improves the reliability of decision-making, and achieves precise matching of task requirements and expert capabilities through time limit and accuracy indicators. Ultimately, it takes into account both response speed and resource utilization while ensuring high accuracy.

[0055] Step S142: determining a number of target vertical expert networks according to the confidence of each vertical expert network.

[0056] Specifically, the method includes: determining the confidence density according to the confidence distribution; setting a reference value for the number of expert networks, and determining the number G of target vertical expert networks according to the confidence density of the vertical expert network cluster and the reference value for the number of expert networks; and selecting the first G vertical expert networks from the confidence distribution as the target vertical expert networks.

[0057] include:

[0058] in, is the confidence density, is the number of vertical expert networks; Also includes:

[0059] in, is the number of target vertical expert networks and G is rounded up, It is a reference value for the number of expert networks. It can be understood that multiple vertical expert networks may be required to cooperate in processing a certain subtask, and the confidence of each vertical expert network in processing the subtask is different. The reference value K of the number of expert networks can be determined according to actual conditions, and it can also be the number of all vertical expert networks in the present invention.

[0060] For example, when processing a certain subtask, the confidence distribution of each vertical expert network is 98%, 90%, 50%, and 0%, then the confidence differences are 8%, 40%, and 50%, respectively, from which the confidence density can be obtained.

[0061] If the confidence density is larger, it means that the difference between adjacent confidences is large, the confidence distribution is steep, and a few experts are significantly better than other experts. In this case, the number of target vertical expert networks can be reduced to reduce the waste of computing resources and avoid generating results with low relevance. If the confidence density is larger, it means that the difference between adjacent confidences is small, the confidence distribution is average, and the results generated by each expert are highly correlated and relatively average. In this case, more experts can be used as much as possible to improve the accuracy of result generation.

[0062] The present invention can optimize the configuration quantity of the target vertical expert network through the confidence density, improve the accuracy of result generation, and avoid the waste of computing resources.

[0063] Step S143: adjusting the target vertical expert network according to the dynamic processing information of the target vertical expert network or the correlation between the target vertical expert networks, wherein the dynamic information includes the running state and the processing state.

[0064] Specifically, the method includes: judging whether the target vertical expert network is in a fault state or an occupied state according to the operating state of the target vertical expert network; if it is in a fault state or an occupied state, eliminating the target vertical expert network and filling it in order with the vertical expert networks that have not been selected; if it is not in a fault state or an occupied state, not eliminating the target vertical expert network; determining a first expert network among several target vertical expert networks; determining the degree of correlation between the first expert network and the remaining target vertical expert networks based on a preset knowledge graph, and sorting the remaining target vertical expert networks according to the degree of correlation.

[0065] If the target vertical expert network is in a faulty state or occupied state, the target vertical expert network is removed and filled in according to the order in the confidence distribution. For example, the confidence distribution of each vertical expert network is 98%, 90%, 50%, and 0%. If the vertical expert network corresponding to 90% is in a faulty state or occupied state, it is removed and filled in with the vertical expert network corresponding to 50%, and the vertical expert network with the highest confidence is used as the first expert network.

[0066] The preset knowledge graph may include knowledge in several fields. In the preset knowledge graph, the association between the field in which the first expert network is located and the fields in which the other target vertical expert networks are located can be found. Specifically, the more associations there are between the field in which the first expert network is located and the other target vertical expert networks, the higher the correlation.

[0067] For example, the first expert network is in the field of image processing, and a target vertical expert network is speech recognition. Then "speech" and "image" can be used as keywords, and the two words appearing in the same relationship are recorded as an association. According to the number of associations, the target vertical expert networks are sorted. It should be noted that the target vertical expert networks that participate in the sorting must participate in the processing of the subtasks.

[0068] Step S144, based on a plurality of target vertical expert networks, a multi-dimensional processing result of the subtask is obtained.

[0069] Based on several target vertical expert networks, a multidimensional processing result of the subtask is obtained, including: based on several target vertical expert networks, the subtask is processed to obtain a first single-dimensional processing result of each target vertical expert network; according to the order of the target vertical expert networks, the result weight corresponding to each target vertical expert network is determined; according to the result weight corresponding to each target vertical expert network and the corresponding first single-dimensional processing result, a second single-dimensional processing result is obtained; according to several second single-dimensional processing results, a multidimensional processing result of the subtask is obtained.

[0070] The subtasks are processed with the target vertical expert network to obtain the corresponding first result of the single-dimensional processing. The result weights can be assigned according to the ranking. For example, the present invention provides fixed assignments of 70%, 20%, and 10%. That is to say, only a fixed number of assignments are taken, and the first three corresponding first results of the single-dimensional processing are taken. Of course, the assignments can also be performed according to the remaining rules, as long as the result weights are positively correlated with the ranking.

[0071] The first result of the single-dimensional processing is converted into structured data (in vector form) and weighted according to the corresponding result weight. Then, for different data types (such as text, image), feature-level or decision-level fusion can be used to obtain the multi-dimensional processing result of the subtask.

[0072] The present invention first pre-sorts several target vertical expert networks by presetting the knowledge graph and confidence distribution, and assigns weights to each result according to the sorting results, so that the processing results corresponding to the subtasks are more in line with the actual situation and the answer hit rate is improved.

[0073] Step S15, obtaining the processing result of the total task of the target user according to the plurality of multi-dimensional processing results.

[0074] The multi-dimensional processing results corresponding to each subtask are converted into vector form, and feature-level or decision-level fusion is adopted to obtain the processing result of the overall task.

[0075] In summary, the present invention provides a user task processing method based on a multi-expert model, the method comprising: obtaining multimodal data of a target user, and obtaining a total task of the target user according to the multimodal data; decomposing the total task to obtain a plurality of subtasks; constructing a multi-expert model, wherein the multi-expert model comprises a gated network unit and a vertical expert network cluster, the vertical expert network cluster comprises a plurality of vertical expert networks, each vertical expert network corresponds to a preset field and the vertical expert networks are all open networks; for each subtask, executing steps S141-S144, comprising: step S141 , control the gated network unit to determine the confidence corresponding to each vertical expert network processing subtask; step S142, determine several target vertical expert networks according to the confidence of each vertical expert network; step S143, adjust the target vertical expert network according to the dynamic processing information of the target vertical expert network or the correlation between the target vertical expert networks, wherein the dynamic information includes the running state and the processing state; step S144, obtain the multidimensional processing result of the subtask based on several target vertical expert networks; obtain the processing result of the total task of the target user according to several multidimensional processing results. The present invention combines the historical accuracy rate and the processing time limit compliance rate, combines the historical performance (accuracy and timeliness) of the expert with the current task relevance, avoids the deviation of a single indicator, improves the reliability of decision-making, realizes the precise matching of task requirements and expert capabilities through time limit and accuracy indicators, and finally takes into account the response speed and resource utilization while ensuring high accuracy. The present invention can optimize the configuration quantity of the target vertical expert network through the confidence density, improve the accuracy of result generation, and avoid the waste of computing resources. The present invention first pre-sorts several target vertical expert networks by presetting knowledge graphs and confidence distribution, and assigns weights to each result according to the sorting results, so that the processing results corresponding to the subtasks are more in line with the actual situation and the answer hit rate is improved. The present invention integrates the questions through the total (total task)-sub (subtask)-total (total result) structure, splits the problems, and merges the results of the split processing as the total result, which can improve the hit rate of the results.

[0076] Based on the same inventive concept, the present invention provides Figure 2 The user task processing device based on the multi-expert model shown includes: An acquisition module 21 is used to acquire multimodal data of a target user and obtain a total task of the target user based on the multimodal data; A decomposition module 22 is used to decompose the overall task into a plurality of subtasks; A construction module 23 is used to construct a multi-expert model, wherein the multi-expert model includes a gated network unit and a vertical expert network cluster, the vertical expert network cluster includes a plurality of vertical expert networks, each vertical expert network corresponds to a preset field and the vertical expert networks are all open networks; The multi-dimensional processing module 24 is used to execute steps S141-S144 for each subtask, including: step S141, controlling the gated network unit to determine the confidence corresponding to each vertical expert network processing subtask; step S142, determining a number of target vertical expert networks according to the confidence of each vertical expert network; step S143, adjusting the target vertical expert network according to the dynamic processing information of the target vertical expert network or the correlation between the target vertical expert networks, wherein the dynamic information includes the running state and the processing state; step S144, obtaining the multi-dimensional processing result of the subtask based on the several target vertical expert networks; The result module 25 is used to obtain the processing result of the overall task of the target user according to the multiple multi-dimensional processing results.

[0077] Based on the same inventive concept, the present invention also provides an electronic device as shown, including: processor; a memory for storing processor-executable instructions; The processor is configured to execute to implement the user task processing method based on the multi-expert model as provided above.

[0078] Based on the same inventive concept, the present invention also provides a non-temporary computer-readable storage medium. When the instructions in the storage medium are executed by a processor of an electronic device, the electronic device can execute the user task processing method based on the multi-expert model as provided above.

[0079] Since the electronic device introduced in this embodiment is an electronic device used to implement the information processing method in the embodiment of the present invention, based on the information processing method introduced in the embodiment of the present invention, a person skilled in the art can understand the specific implementation of the electronic device of this embodiment and its various variations, so how the electronic device implements the method in the embodiment of the present invention is not described in detail here. As long as the electronic device used by a person skilled in the art to implement the information processing method in the embodiment of the present invention, it belongs to the scope of protection of the present invention.

[0080] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0081] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0082] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0083] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0084] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0085] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A user task processing method based on a multi-expert model, characterized by: The method comprises: Acquire multimodal data of a target user, and obtain a total task of the target user based on the multimodal data; Decomposing the overall task into several subtasks; Construct a multi-expert model, where the multi-expert model includes a gated network unit and a vertical expert network cluster. The vertical expert network cluster includes several vertical expert networks, each of which corresponds to a preset field and is an open network. For each subtask, execute steps S141-S144, including: Step S141, controlling the gated network unit to determine the confidence level corresponding to each vertical expert network processing subtask; Step S142, determining a number of target vertical expert networks based on the confidence of each vertical expert network; Step S143, adjusting the target vertical expert network according to the dynamic processing information of the target vertical expert network or the correlation between the target vertical expert networks, wherein the dynamic information includes the running state and the processing state; Step S144, obtaining a multi-dimensional processing result of the subtask based on a plurality of target vertical expert networks; According to the multiple multi-dimensional processing results, a processing result of the total task of the target user is obtained.

2. The user task processing method based on the multi-expert model according to claim 1, characterized in that: Controlling the gated network unit to determine the confidence corresponding to each vertical expert network processing subtask includes: Perform feature preprocessing and dimension expansion on the initial features of the subtask to obtain expanded fusion features; The expanded fusion features are mapped to a dimensional space with the same number of vertical expert networks through a dimensionality reduction layer, and a confidence distribution is generated based on the Softmax function. The confidence distribution includes the confidence of each vertical expert network. The confidence distribution is adjusted based on the historical processing capabilities of the vertical expert network, where the historical processing capabilities include historical accuracy and processing time limit compliance rate.

3. The user task processing method based on the multi-expert model according to claim 2, characterized in that: Based on the historical processing capabilities of the vertical expert network, the confidence distribution is adjusted, including: in, After adjustment The confidence of a vertical expert network, Before adjustment The confidence of a vertical expert network, is the historical accuracy, The processing time limit compliance rate.

4. The user task processing method based on the multi-expert model according to claim 1, characterized in that: According to the confidence of each vertical expert network, several target vertical expert networks are determined, including: According to the confidence distribution, the confidence density is determined; Setting a reference value for the number of expert networks, and determining the number G of target vertical expert networks based on the confidence density of the vertical expert network cluster and the reference value for the number of expert networks; The top G vertical expert networks are selected from the confidence distribution as the target vertical expert networks.

5. The user task processing method based on the multi-expert model according to claim 4, characterized in that: include: in, is the confidence density, is the number of vertical expert networks; Also includes: in, is the number of target vertical expert networks and G is rounded up, is the reference value of the number of expert networks.

6. The user task processing method based on a multi-expert model according to claim 1, characterized in that: According to the dynamic processing information of the target vertical expert network or the correlation between the target vertical expert networks, the target vertical expert network is adjusted, including: According to the operation status of the target vertical expert network, it is determined whether the target vertical expert network is in a fault state or an occupied state; If so, the target vertical expert network will be eliminated and filled with the vertical expert networks that have not been selected in order; if neither of them is in, no elimination will be performed; Determine a first expert network among several target vertical expert networks; Based on the preset knowledge graph, the degree of correlation between the first expert network and the remaining target vertical expert networks is determined, and the remaining target vertical expert networks are sorted according to the degree of correlation.

7. The user task processing method based on the multi-expert model according to claim 6, characterized in that: Based on several target vertical expert networks, we obtain the multi-dimensional processing results of this subtask, including: Based on several target vertical expert networks, the subtask is processed to obtain a first single-dimensional processing result of each target vertical expert network; According to the ranking of the target vertical expert networks, determine the result weight corresponding to each target vertical expert network; According to the result weights corresponding to each target vertical expert network and the corresponding first result of the single-dimensional processing, the second result of the single-dimensional processing is obtained; A multi-dimensional processing result of the subtask is obtained according to the second results of the single-dimensional processing.

8. A user task processing device based on a multi-expert model, characterized in that: The device comprises: An acquisition module, configured to acquire multimodal data of a target user and obtain a total task of the target user based on the multimodal data; A decomposition module, used to decompose the overall task into several subtasks; A construction module is used to construct a multi-expert model, wherein the multi-expert model includes a gated network unit and a vertical expert network cluster. The vertical expert network cluster includes several vertical expert networks, each of which corresponds to a preset field and is an open network. The multidimensional processing module is used to execute steps S141-S144 for each subtask, including: step S141, controlling the gated network unit to determine the confidence corresponding to each vertical expert network processing subtask; step S142, determining a plurality of target vertical expert networks based on the confidence of each vertical expert network; step S143, adjusting the target vertical expert network based on dynamic processing information of the target vertical expert network or the correlation between the target vertical expert networks, wherein the dynamic information includes the running state and the processing state; step S144, obtaining the multidimensional processing result of the subtask based on the plurality of target vertical expert networks; The result module is used to obtain the processing result of the total task of the target user according to a plurality of multi-dimensional processing results.

9. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute to implement the user task processing method based on a multi-expert model as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, characterized in that When the instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to implement the user task processing method based on a multi-expert model according to any one of claims 1 to 7.

Images