Data analysis method and electronic equipment based on multiple analysis tasks
By prioritizing and predicting the results in parallel data analysis of multi-analysis tasks, we can determine whether to wait for the second analysis task to generate real results, which solves the problem of time-consuming and inefficient in the existing technology, and realizes more efficient data analysis and more Fast output result.
Patent Information
- Application Number
- CN202011461545.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2020-03-17
- Filing Date
- 2020-12-08
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2040-12-08
AI Technical Summary
In the parallel data analysis of multi-analysis tasks, the existing processing models take a long time and are less efficient due to differences in analysis tasks, and cannot effectively improve the accuracy and timeliness of the output results.
By triggering multiple parallel analysis tasks after obtaining the pending data, and waiting for all tasks to generate results, priority sorting is done to determine whether you need to wait for the second analysis task to generate the real result. If there is no waiting value, the highest priority will be The result is used as an output result, saving waiting time.
It improves the efficiency of data analysis, shortens the time-consuming task waiting, and improves the timeliness and accuracy of analysis results, especially in scenarios with high real-time requirements, which significantly improves processing efficiency.
Smart Images

Figure CN113408567B_ABST
Abstract
Description
[0001] This application claims the priority of the Chinese patent application filed with the State Intellectual Property Office on March 17, 2020, with application number 202010185620.2 and application name “Data analysis method and electronic equipment based on multiple analysis tasks”, the entire contents of which are incorporated by reference in this application. Technical Field
[0002] The present application belongs to the field of artificial intelligence technology, and in particular, relates to a data analysis method and electronic equipment based on multiple analysis tasks. Background Art
[0003] When performing data analysis and outputting results, existing processing models often start multiple parallel analysis tasks in order to improve the accuracy of the output results. Each analysis task independently analyzes the data to be processed and obtains the corresponding analysis results. The processing model then conducts a comprehensive analysis of all the analysis results and selects the best one as the final output result.
[0004] Due to the differences in the analysis principles, dependent resources, data transmission conditions, and other factors of each analysis task, the time consumption of each analysis task will be greatly different. Therefore, although parallel data analysis of multiple analysis tasks can improve the accuracy of the final output results, it will also make the data analysis of the final processing model take longer and less efficient. Summary of the invention
[0005] In view of this, the embodiments of the present application provide a data analysis method and electronic device based on multiple analysis tasks, which can solve the problem of long time consumption and low efficiency in multi-analysis task data analysis in the field of artificial intelligence.
[0006] A first aspect of an embodiment of the present application provides a data analysis method based on multiple analysis tasks, comprising:
[0007] After obtaining the data to be processed, multiple parallel first analysis tasks and multiple second analysis tasks are triggered to process the data to be processed. And wait for all first analysis tasks to generate first results, and the second results obtained by predicting the results of the second analysis tasks. After obtaining all first results and second results, these results are prioritized, and it is determined whether it is necessary to wait for the second analysis task to generate the real analysis result based on the sorting results. Finally, if it is determined that it is not necessary to wait for the second analysis task to generate the analysis result, the result with the highest priority in the sorting results is used as the output result.
[0008] In the embodiment of the present application, the existing real analysis results and predicted analysis results are prioritized, and the ungenerated real results are judged whether they have waiting value according to the ranking results. When it is judged that the ungenerated results have no waiting value, the result with the highest priority among the existing analysis results is used as the output result. Therefore, a lot of waiting time for the real results can be saved, thereby improving the analysis time.
[0009] In a first possible implementation manner of the first aspect, the operation of determining whether waiting is required includes:
[0010] If the result with the highest priority in the sorted results is the first result, it is determined that there is no need to wait for the second analysis task to generate an analysis result.
[0011] Since the first result is a true result, when the first result has the highest priority, it means that no matter what the analysis result generated by the second analysis task is, it will not affect the actual priority order. This further means that the waiting value of the ungenerated true result is small and it is not necessary to wait.
[0012] A second possible implementation of the first aspect includes:
[0013] If the result with the highest priority in the sorting results is the second result, the second analysis task corresponding to the result with the highest priority is used as the target analysis task, and it is determined that it is necessary to wait for the target analysis task to generate an analysis result.
[0014] Correspondingly, the data analysis method based on multiple analysis tasks also includes:
[0015] If it is determined according to the sorting result that it is necessary to wait for the target analysis task to generate an analysis result, the analysis result generated by the target analysis task is obtained and used as the output result.
[0016] When the result with the highest priority in the sorting result is the second result, it means that the second result is the most suitable output result. However, since the second result is the predicted result of the second analysis task, it is not the real result generated by the second analysis task. Therefore, the second result cannot be directly used as the output result, but it is necessary to wait until the real result corresponding to the second result is obtained before using the corresponding real result as the output result.
[0017] Based on the first or second possible implementation manner of the first aspect, in a third possible implementation manner of the first aspect, obtaining a sorting result obtained by sorting the first result and the second result by priority includes:
[0018] If a third result generated by one or more second analysis tasks is obtained, a sorting result obtained by sorting the first result, the second result, and the third result by priority is obtained.
[0019] Correspondingly, judging whether it is necessary to wait for the second analysis task to generate an analysis result according to the sorting result includes:
[0020] If the result with the highest priority in the sorted results is the first result or the third result, it is determined that there is no need to wait for the second analysis task to generate an analysis result.
[0021] Considering the actual situation, some of the real results generated by the second analysis tasks may be obtained before sorting. At this time, the embodiment of the present application will prioritize these real results together. Correspondingly, since the third result is also a real result. Therefore, when the first result or the third result is the result with the highest priority, it means that the waiting value of the remaining ungenerated real results is small. That is, there is no need to wait for these second analysis tasks to generate real results.
[0022] Based on the first or second possible implementation manner of the first aspect, in a fourth possible implementation manner of the first aspect, obtaining a sorting result obtained by sorting the first result and the second result by priority includes:
[0023] If a third result generated by one or more second analysis tasks is obtained, the result corresponding to the third result is eliminated from the obtained second results, and the first result, the third result and the second result remaining after the elimination operation are prioritized to obtain a sorted result.
[0024] Correspondingly, judging whether it is necessary to wait for the second analysis task to generate an analysis result according to the sorting result includes:
[0025] If the result with the highest priority in the sorted results is the first result or the third result, it is determined that there is no need to wait for the second analysis task to generate an analysis result.
[0026] Considering the actual situation, some of the real results generated by the second analysis tasks may be obtained before sorting. At this time, the embodiment of the present application will remove the predicted results corresponding to these real results, and then prioritize them together. Correspondingly, since the third result is also a real result. Therefore, when the first result or the third result is the result with the highest priority, it means that the waiting value of the remaining ungenerated real results is small. That is, there is no need to wait for these second analysis tasks to generate real results.
[0027] Based on the first or second or third possible implementation manner of the first aspect, in a fifth possible implementation manner of the first aspect, the operation of obtaining a second result obtained by predicting a result of the second analysis task includes:
[0028] A pre-trained data analysis model corresponding to the second analysis task is obtained, and the data to be processed is input into the data analysis model for processing, wherein the data analysis model is used to predict the result of the corresponding second analysis task. A second result generated by the data analysis model is obtained.
[0029] In the embodiment of the present application, a model for predicting the analysis result is pre-trained, and the model is used to predict the result of the second analysis task, so as to ensure the effectiveness of the prediction.
[0030] Based on the fifth possible implementation manner of the first aspect, in a sixth possible implementation manner of the first aspect, the training process of the single data analysis model further includes:
[0031] Obtain sample data and update parameters of the data analysis model based on the sample data.
[0032] After each parameter update operation is completed, the data analysis model analyzes the accuracy data of the sample data based on the preset classifier analysis after the parameter update operation is completed.
[0033] Based on the accuracy data, the data analysis model is trained by iterative parameter update until the data analysis model meets the preset convergence conditions and the training is completed.
[0034] Since the analysis task may have a strong response in actual situations, in order to improve the situation of strong response of the data analysis model, a classifier that can analyze the accuracy of the analysis result will be pre-trained in an embodiment of the present application. When the data analysis model is trained, each time the data analysis model generates an analysis result, the classifier will be used to determine whether the analysis result is a strong response result. And the judgment result is used as the feature data for the next iterative update of the data analysis model parameters. Until the data analysis model converges, the model training is completed. The trained data analysis model is less likely to have a strong response than the original second type of analysis task. Thereby, the accuracy of the final output result is improved.
[0035] Based on the fifth or sixth possible implementation of the first aspect, in the seventh possible implementation of the first aspect, the construction and training of the data analysis model can be achieved based on a teacher-student model of knowledge distillation.
[0036] Since the teacher-student model can realize the learning from the teacher model to the student model, the student model can imitate the behavior of the teacher model and achieve considerable accuracy. Therefore, the data analysis model obtained in the embodiment of the present application can realize the effective simulation of each second-class analysis task, and ensure the effective prediction and priority sorting analysis of the classification results of each second-class analysis task. At the same time, knowledge distillation can make the student model more lightweight, thereby making the volume of the data analysis model smaller, and the model efficiency is higher for the corresponding second-class analysis tasks. Thereby, the data analysis model in the embodiment of the present application can generate analysis results faster while ensuring accuracy.
[0037] Based on the fifth or sixth possible implementation manner of the first aspect, an eighth possible implementation manner of the first aspect includes:
[0038] The third result generated by each second analysis task is obtained, and based on the data to be processed and the third result, parameters of each data analysis model are updated.
[0039] In the embodiment of the present application, by continuously learning and updating the data analysis model online, the simulation effect of the data analysis model on the second type of analysis task can be guaranteed, and the dynamic change requirements of the matching relationship between the data to be processed and the analysis results can be adapted. Thus, the reliability of the analysis and processing of the data to be processed in the embodiment of the present application can be guaranteed.
[0040] Based on the eighth possible implementation manner of the first aspect, in a ninth possible implementation manner of the first aspect, in the operation of updating parameters of each data analysis model based on the data to be processed and the third result, the operation of updating parameters of the data analysis model corresponding to the single second analysis task includes:
[0041] A matching analysis is performed on the data to be processed and the third result generated by the second analysis task.
[0042] If the data to be processed matches the third result generated by the second analysis task, parameters of the data analysis model corresponding to the second analysis task are updated based on the data to be processed and the third result generated by the second analysis task.
[0043] The embodiment of the present application will analyze the third result and the corresponding data to be processed each time it is obtained. If the third result does not match the data to be processed, the corresponding data analysis model will not be updated online to ensure the effectiveness of the online update of the data analysis model.
[0044] A second aspect of an embodiment of the present application provides a data analysis method based on multiple analysis tasks, including:
[0045] After obtaining the data to be processed, multiple parallel first analysis tasks and multiple second analysis tasks are triggered to process the data to be processed. And wait for all first analysis tasks to generate first results, and the second results obtained by predicting the results of the second analysis tasks. After obtaining all first results, second results and the confidence levels corresponding to each second result, these results are prioritized. And based on the sorting results and confidence levels, it is determined whether it is necessary to wait for the second analysis task to generate the real analysis results. Finally, if it is determined that it is not necessary to wait for the second analysis task to generate the analysis results, the result with the highest priority in the sorting results is used as the output result.
[0046] In the embodiment of the present application, the existing real analysis results and predicted analysis results are prioritized, and the ungenerated real results are judged whether they have waiting value based on the ranking results and confidence. When it is judged that the ungenerated results have no waiting value, the result with the highest priority among the existing analysis results is used as the output result. Therefore, a lot of waiting time for the real results can be saved, thereby improving the analysis time.
[0047] In a first possible implementation manner of the second aspect, the operation of determining whether waiting is required includes:
[0048] If the target result is not included in the top n results with the highest priority in the first sorting result, and the result with the highest priority in the first sorting result is the first result, it is determined that there is no need to wait for the second analysis task to generate an analysis result, where the target result is the second result whose confidence level is higher than a preset first threshold, and n is an integer greater than 1.
[0049] Because the confidence level can reflect the matching degree between the analysis result and the data to be processed. For the second result that is a predicted result, when the second result confidence level is high, it means that it is more likely to be the analysis result that matches the data to be processed. Therefore, after obtaining the first sorting result, the embodiment of the present application will first identify whether the top n analysis results with the highest priority contain a second result with high confidence. When the high confidence result is not included and the highest priority is the true result, it means that no matter what the analysis result generated by the second analysis task is, it will not affect the actual priority order. This further illustrates that the waiting value of the ungenerated true result is small and there is no need to wait.
[0050] In a second possible implementation manner of the second aspect, the operation of determining whether waiting is required includes:
[0051] If the target result is not included in the top n results with the highest priority in the first sorting result, and the result with the highest priority in the first sorting result is the second result, the second analysis task corresponding to the result with the highest priority is used as the target analysis task, and it is determined that it is necessary to wait for the target analysis task to generate an analysis result, where the target result is the second result whose confidence level is higher than a preset first threshold, and n is an integer greater than 1.
[0052] Correspondingly, the data analysis method based on multiple analysis tasks also includes:
[0053] If it is determined according to the first sorting result and the confidence level that it is necessary to wait for the target analysis task to generate an analysis result, the analysis result generated by the target analysis task is obtained and used as the output result.
[0054] When the top n analysis results with the highest priority do not contain a second result with high confidence, and the result with the highest priority is the second result, it means that the second result is the most suitable output result. However, since the second result is a predicted result of the second analysis task, it is not the real result generated by the second analysis task. Therefore, the second result cannot be directly used as the output result, but it is necessary to wait until the real result corresponding to the second result is obtained before using the corresponding real result as the output result.
[0055] Based on the first or second possible implementation manner of the second aspect, in a third possible implementation manner of the first aspect, the operation of determining whether waiting is required further includes:
[0056] If the target result is included in the top n results with the highest priority in the first sorting result, and the result with the highest priority in the first sorting result is the target result, the second analysis task corresponding to the result with the highest priority is used as the target analysis task, and it is determined that it is necessary to wait for the target analysis task to generate an analysis result, wherein the target result is the second result whose confidence level is higher than a preset first threshold, and n is an integer greater than 1.
[0057] Correspondingly, the data analysis method based on multiple analysis tasks also includes:
[0058] If it is determined according to the first sorting result and the confidence level that it is necessary to wait for the target analysis task to generate an analysis result, the analysis result generated by the target analysis task is obtained and used as the output result.
[0059] When the analysis result with the highest priority is the target result, it indicates that the target result has a high matching degree with the data to be processed. However, since the target result is a predicted result rather than a real result, the embodiment of the present application will obtain the third result corresponding to the target result as the output result.
[0060] Based on the first, second or third possible implementation manner of the second aspect, in a fourth possible implementation manner of the first aspect, the operation of determining whether waiting is required further includes:
[0061] If the target result is included in the top n results with the highest priority in the first sorting result, and the result with the highest priority in the first sorting result is not the target result, it is determined that it is necessary to wait for the second analysis task to generate an analysis result, where the target result is the second result whose confidence level is higher than a preset first threshold, and n is an integer greater than 1.
[0062] The current n results contain the target result and the one with the highest priority is not the target result, indicating that the target result and the analysis result with the highest priority do not have much difference in matching degree with the data to be processed. That is, the target results of the first n results may have a certain impact on the final output result, and these target results still have waiting value. Therefore, the embodiment of the present application will determine that it is necessary to wait for the analysis result of the second analysis task to ensure the accuracy of the final output result.
[0063] Based on the fourth possible implementation manner of the second aspect, in a fifth possible implementation manner of the first aspect, further comprising:
[0064] If it is necessary to wait for the second analysis task to generate an analysis result based on the first sorting result and confidence level, obtain the third result corresponding to each target result in the first n results, wherein the third result is the analysis result generated by the second analysis task corresponding to the target result.
[0065] The first result among the first n results, the second result whose confidence is lower than or equal to the first threshold, and the obtained third result are prioritized to obtain a corresponding second sorting result.
[0066] If the result with the highest priority in the second sorting results is the first result or the third result, the result with the highest priority in the second sorting results is used as the output result.
[0067] The embodiment of the present application will obtain the third result corresponding to all target results. After obtaining the third result corresponding to all target results, the first result in the first n analysis results, the second result with a confidence level lower than the first threshold, and the obtained third result are prioritized again, and the corresponding second sorting result is obtained. Since the first result and the third result are both true results. Therefore, it is explained that the waiting value of the second analysis task that has not generated an analysis result is small and no waiting is required. The true result with the highest priority can be used as the output result.
[0068] Based on the fifth possible implementation manner of the second aspect, in a sixth possible implementation manner of the first aspect, further comprising:
[0069] If the result with the highest priority in the second sorting result is the second result, obtain the third result corresponding to the second result, and use the obtained third result as the output result, wherein the third result is the analysis result generated by the second analysis task corresponding to the result with the highest priority in the second sorting result.
[0070] If the second result has the highest priority in the second sorting result, it means that the second result has a high degree of matching with the data to be processed. However, since the second result is a predicted result rather than a real result, the embodiment of the present application will obtain the third result corresponding to the second result as the output result.
[0071] Based on the fifth possible implementation manner of the second aspect, in a seventh possible implementation manner of the first aspect, further including:
[0072] If it is necessary to wait for the second analysis task to generate an analysis result based on the first sorting result and confidence level, obtain the third result corresponding to each second result in the first n results, wherein the third result is the analysis result generated by the second analysis task corresponding to the second result.
[0073] Prioritize the first result in the first n results and the obtained third result to obtain a corresponding second sorted result. The result with the highest priority in the second sorted result is used as the output result.
[0074] In the embodiment of the present application, the real results corresponding to all the predicted results in the first n results will be waited for, and then the secondary sorting will be performed. At this time, since the content of the second priority sorting is the first result and the third result, these results are all real results. Therefore, there is no need to wait any longer, and the result with the highest priority can be used as the output result. The embodiment of the present application can prevent the mistaken abandonment of the high confidence prediction result and improve the accuracy of the final output result.
[0075] On the basis of any one of the first to fifth possible implementations of the second aspect, in an eighth possible implementation of the first aspect, the operation of obtaining a second result obtained by predicting a result of the second analysis task includes:
[0076] A pre-trained data analysis model corresponding to the second analysis task is obtained, and the data to be processed is input into the data analysis model for processing, wherein the data analysis model is used to predict the result of the corresponding second analysis task.
[0077] A second result generated by the data analysis model is obtained.
[0078] In the embodiment of the present application, a model for predicting the analysis result is pre-trained, and the model is used to predict the result of the second analysis task, so as to ensure the effectiveness of the prediction.
[0079] Based on the eighth implementation manner of the second aspect, in a ninth possible implementation manner of the first aspect, the training process of the single data analysis model further includes:
[0080] Obtain sample data and update parameters of the data analysis model based on the sample data.
[0081] After each parameter update operation is completed, the data analysis model analyzes the accuracy data of the sample data based on the preset classifier analysis after the parameter update operation is completed.
[0082] Based on the accuracy data, the data analysis model is trained by iterative parameter update until the data analysis model meets the preset convergence conditions and the training is completed.
[0083] Since the analysis task may have a strong response in actual situations, in order to improve the situation of strong response of the data analysis model, a classifier that can analyze the accuracy of the analysis result will be pre-trained in an embodiment of the present application. When the data analysis model is trained, each time the data analysis model generates an analysis result, the classifier will be used to determine whether the analysis result is a strong response result. And the judgment result is used as the feature data for the next iterative update of the data analysis model parameters. Until the data analysis model converges, the model training is completed. The trained data analysis model is less likely to have a strong response than the original second type of analysis task. Thereby, the accuracy of the final output result is improved.
[0084] Based on the eighth or ninth possible implementation of the second aspect, in the tenth possible implementation of the second aspect, the construction and training of the data analysis model can be achieved based on a teacher-student model of knowledge distillation.
[0085] Since the teacher-student model can realize the learning from the teacher model to the student model, the student model can imitate the behavior of the teacher model and achieve considerable accuracy. Therefore, the data analysis model obtained in the embodiment of the present application can realize the effective simulation of each second-class analysis task, and ensure the effective prediction and priority sorting analysis of the classification results of each second-class analysis task. At the same time, knowledge distillation can make the student model more lightweight, thereby making the volume of the data analysis model smaller, and the model efficiency is higher for the corresponding second-class analysis tasks. Thereby, the data analysis model in the embodiment of the present application can generate analysis results faster while ensuring accuracy.
[0086] Based on the eighth or ninth possible implementation manner of the second aspect, in an eleventh possible implementation manner of the second aspect, further comprising:
[0087] The third result generated by each second analysis task is obtained, and based on the data to be processed and the third result, parameters of each data analysis model are updated.
[0088] In the embodiment of the present application, by continuously learning and updating the data analysis model online, the simulation effect of the data analysis model on the second type of analysis task can be guaranteed, and the dynamic change requirements of the matching relationship between the data to be processed and the analysis results can be adapted. Thus, the reliability of the analysis and processing of the data to be processed in the embodiment of the present application can be guaranteed.
[0089] Based on the eleventh possible implementation manner of the second aspect, in a twelfth possible implementation manner of the second aspect, the present invention further includes:
[0090] A matching analysis is performed on the data to be processed and the third result generated by the second analysis task.
[0091] If the data to be processed matches the third result generated by the second analysis task, parameters of the data analysis model corresponding to the second analysis task are updated based on the data to be processed and the third result generated by the second analysis task.
[0092] The embodiment of the present application will analyze the third result and the corresponding data to be processed each time it is obtained. If the third result does not match the data to be processed, the corresponding data analysis model will not be updated online to ensure the effectiveness of the online update of the data analysis model.
[0093] A third aspect of an embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, the electronic device implements the steps of the data analysis method based on multiple analysis tasks as described in any one of the first aspects above, or implements the steps of the data analysis method based on multiple analysis tasks as described in any one of the second aspects above.
[0094] A fourth aspect of an embodiment of the present application provides a computer-readable storage medium, comprising: storing a computer program, characterized in that when the computer program is executed by a processor, the electronic device implements the steps of the data analysis method based on multiple analysis tasks as described in any one of the first aspects above, or implements the steps of the data analysis method based on multiple analysis tasks as described in any one of the second aspects above.
[0095] A fifth aspect of an embodiment of the present application provides a computer program product. When the computer program product runs on an electronic device, the electronic device executes the data analysis method based on multiple analysis tasks described in any one of the first aspects above, or implements the steps of the data analysis method based on multiple analysis tasks described in any one of the second aspects above.
[0096] It can be understood that the beneficial effects of the third to fifth aspects mentioned above can be found in the relevant descriptions of the first and second aspects mentioned above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS
[0097] Figure 1A A schematic diagram of a scenario provided for an embodiment of the present application;
[0098] Figure 1B A flowchart of a data analysis method based on multiple analysis tasks provided in an embodiment of the present application;
[0099] Figure 1C A flowchart of a data analysis method based on multiple analysis tasks provided in an embodiment of the present application;
[0100] Figure 1D A flowchart of a data analysis method based on multiple analysis tasks provided in an embodiment of the present application;
[0101] Figure 2A A flowchart of a data analysis method based on multiple analysis tasks provided in an embodiment of the present application;
[0102] Figure 2B A flowchart of a data analysis method based on multiple analysis tasks provided in an embodiment of the present application;
[0103] Figure 2C A schematic diagram of a scenario provided for an embodiment of the present application;
[0104] Figure 2D A schematic diagram of a flow chart of model updating in a data analysis method based on multiple analysis tasks provided in an embodiment of the present application;
[0105] Figure 3 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION
[0106] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.
[0107] In order to facilitate the understanding of the present application, the embodiments of the present application are briefly described here:
[0108] When performing multi-task parallel data analysis, the existing processing models first wait for all parallel analysis tasks to generate corresponding analysis results, and then prioritize all analysis results. Finally, the analysis result with the highest priority is used as the final output result. For example, in the processing model for human-computer dialogue in the field of natural language processing (NLP) such as voice assistants. When the user dialogue content is obtained, the dialogue content will be used as the data to be processed to synchronously start multiple dialogue analysis tasks. After obtaining the analysis results of all dialogue analysis tasks, the analysis results will be prioritized. Finally, the analysis result with the highest priority is used as the output result for subsequent analysis of the user dialogue response.
[0109] In order to reduce the total waiting time for parallel tasks and improve the analysis efficiency when multiple analysis tasks perform data analysis. One optional approach is to divide the analysis tasks into platform self-developed tasks and tasks provided by third-party service providers, and set a maximum waiting time. Among them, platform self-developed tasks refer to analysis tasks provided by the service platform corresponding to the processing model. After triggering multiple analysis tasks, the timing starts. On the one hand, it will wait for the platform's self-developed tasks to generate results, and on the other hand, it will collect the results returned by all tasks provided by third-party service providers within the maximum waiting time. After exceeding the maximum waiting time, it will no longer wait for tasks provided by third-party service providers, and sort and select all obtained results. Since the platform's self-developed tasks are generally shorter, by setting a limit on the maximum waiting time, the waiting time for third-party service providers can be effectively controlled. Avoid the situation of waiting for tasks from third-party service providers for a long time.
[0110] In conjunction with examples, examples are given to illustrate. In an embodiment of the present application, the application scenario is human-computer dialogue, that is, the electronic device needs to conduct human-computer dialogue with the user, and implement human-computer dialogue management (Dialog Manager, DM). Among them, the analysis tasks that are responsible for the voice processing platform (i.e., service platform, referred to as platform) corresponding to the analysis model in the electronic device include telephone and alarm clock, etc. The analysis tasks provided by a third-party service provider (also referred to as a non-platform) include music, weather, search, encyclopedia and chat, etc. The maximum waiting time is set to 200 milliseconds. Among them, music and weather are analysis tasks provided by third-party service provider 1, and search, encyclopedia and chat are analysis tasks provided by third-party service provider 2. There are also third-party service providers 3 that can also provide some analysis tasks (not shown in the figure), such as navigation and other analysis tasks.
[0111] Assume that during the actual operation, the analysis tasks of the platform have all generated results, and the analysis tasks of third-party service providers 1 and 2 have all returned results within 200 milliseconds. However, the analysis task of third-party service provider 3 has not returned results within 200 milliseconds. Figure 1A . At this time, the non-platform operation results such as music, weather, search, encyclopedia and chat, as well as platform operation results such as phone and alarm clock returned within 200 milliseconds will be comprehensively sorted. Finally, the final result is determined based on the sorting results. Practice has proved that through the method of the embodiment of the present application, the DM time consumption can be greatly shortened (practice has proved that the shortened time can generally reach more than 100 milliseconds), and the efficiency of multi-task analysis can be improved.
[0112] Although parallel data analysis of multiple analysis tasks can improve the accuracy of the final output results. However, it is found in actual applications that, on the one hand, there are still certain differences in the help of each analysis task in improving the accuracy of the final output results. In different scenarios, this difference may be reflected in the importance of the analysis task itself, or in the degree of influence of the analysis results generated by the analysis task on the final output results. For analysis tasks that are weak or even unhelpful in improving the accuracy of the final output results. If these analysis tasks are continuously waited for, the actual significance is small and the task waiting time will be greatly increased. Therefore, the waiting value of these analysis tasks is small. On the other hand, for many scenarios, the accuracy and timeliness of the final output results of the processing model (that is, the time consumption of data analysis) are two very important performance indicators. For example, if the voice assistant cannot respond to the user's conversation in time, even if the final response result is very accurate, it loses its meaning at this time. Therefore, in these scenarios with high real-time requirements, in order to meet the needs of shortening the waiting time and timeliness. Some time-consuming analysis tasks and some tasks that are weak in helping to improve the accuracy of the final output results are tasks with low waiting value.
[0113] Based on the above principles, on the one hand, the embodiment of the present application will pre-divide the analysis tasks into two categories according to the task time consumption: the first type of analysis tasks with shorter time consumption and the second type of analysis tasks with longer time consumption. Practice has found that the second type of analysis tasks with longer time consumption often lead to longer data analysis time consumption and lower efficiency for the final processing model. Among them, the embodiment of the present application does not impose too many restrictions on the division criteria for the length of time consumption, as well as the content and quantity of tasks specifically included in the first type of analysis tasks and the second type of analysis tasks. It can be set by technical personnel according to the actual time consumption of each analysis task in actual application, and the demand for task timeliness in the application scenario.
[0114] On the other hand, the embodiment of the present application will pre-train the corresponding data analysis model for each second-class analysis task, and these data analysis models are used to simulate the corresponding second-class analysis task for data processing. For example, in the voice assistant, assuming that the task of weather query belongs to the second-class analysis task, a data analysis model for weather query can be pre-trained. When performing data analysis of the data to be processed, while triggering each first-class analysis task and the second-class analysis task, these data analysis models will also be used to simulate each second-class analysis task for data analysis. After triggering each analysis task and data analysis model, the embodiment of the present application will wait for all first-class analysis tasks and data analysis models to complete data analysis, and then prioritize the obtained analysis results. Finally, according to the sorting result, it is judged whether the final suitable output result is the analysis result corresponding to the first-class analysis task or the analysis result corresponding to the data analysis model. If it is the analysis result corresponding to the first-class analysis task, the result can be used as the output result. If it is the analysis result corresponding to the data analysis model, it will wait for the analysis result actually returned by the corresponding second-class analysis task to be received, and then the analysis result actually returned will be used as the output result. Because the data analysis model can be designed and stored according to the actual scene requirements. Relative to the second type of analysis tasks, the resources that the data analysis model relies on can be pre-set, and data transmission is more efficient and reliable. The data analysis model analysis is time-consuming and more efficient. Therefore, the prediction results output by the data analysis model are comprehensively sorted with the real results of the first type of analysis tasks, and the waiting value of each second type of analysis task can be accurately known. On this basis, the final output result is evaluated according to the sorting result to see if it is the analysis result corresponding to the second type of task. If not, the real result with the highest priority can be used as the output result, so that the embodiment of the present application does not need to wait for the second type of analysis task to return the result. Then the total duration of task waiting is reduced and efficiency is improved. When the evaluation result is that the output result is the analysis result corresponding to the second type of task, then wait for the corresponding second type of task to return the analysis result. At this time, the waiting for some analysis tasks can be selectively abandoned, thereby reducing the total duration of waiting for parallel tasks and improving the analysis efficiency when multiple analysis tasks perform data analysis.
[0115] At the same time, the execution subject of the data analysis method based on multiple analysis tasks in the embodiment of the present application is: an electronic device that can realize operations such as triggering multiple parallel analysis tasks and outputting analysis results in the embodiment of the present application. Among them, the specific type of electronic device is not limited here and needs to be determined according to the actual application scenario. For example, it can be a terminal device such as a mobile phone, a computer, and a wearable device, or it can be a server and other devices.
[0116] In order to illustrate the technical solution described in this application, a specific embodiment is provided below for illustration.
[0117] Figure 1B The following is a flowchart of the implementation of the data analysis method based on multiple analysis tasks provided in the first embodiment of the present application, which is described in detail as follows:
[0118] S101, obtaining data to be processed, triggering at least one first analysis task and at least one second analysis task in parallel, wherein the first analysis task and the second analysis task are both used to analyze the data to be processed.
[0119] Among them, the data content and data type of the data to be processed need to be determined according to the actual application scenario and are not limited here. For example, when the embodiment of the present application is applied to the scenario of human-computer dialogue, such as applied to a voice assistant to achieve a response to the content of the user's dialogue. At this time, the data to be processed can be the content of the user's dialogue, such as the user's question "What's the weather like today?" For another example, the embodiment of the present application is applied to user product push to achieve personalized push of the user's favorite products. At this time, the data to be processed can be the user's personal information and some historical consumption records, such as the user's gender, age and historical purchase records in the past six months.
[0120] At the same time, under different application scenarios, the types and quantities of analysis tasks actually required to be used may also have certain differences, which can be selected or set by the technician according to actual needs, and are not limited here. For example, in the scenario of human-computer dialogue, in theory, there is a positive correlation between the number of analysis tasks and the accuracy of the final analysis result. Therefore, the technician can set different numbers of parallel analysis tasks according to different accuracy requirements. In order to improve the comprehensiveness of the analysis results and thus improve the accuracy of the final output results, the technician can also set the type of analysis tasks actually triggered according to the actual scenario requirements. For example, in order to enrich the functions of the voice assistant, the analysis tasks corresponding to the functions such as phone, alarm clock, music, weather, encyclopedia search and chat can be used as the analysis tasks required to be triggered in the embodiment of the present application. On the other hand, for the actual time consumption of each analysis task in the actual application scenario, and the application scenario's demand for the timeliness of the output results. The embodiment of the present application will pre-classify the analysis tasks that may be used in the actual application scenario into the above-mentioned first type of analysis tasks and second type of analysis tasks. So that after determining the analysis tasks required to be used, the embodiment of the present application can simultaneously determine the types of each analysis task. In the embodiment of the present application, all the first-category analysis tasks in the triggered analysis tasks are named first analysis tasks, and all the second-category analysis tasks are named second analysis tasks.
[0121] As an optional way to classify analysis tasks in this application. In actual applications, it is found that the analysis tasks performed by the service platform corresponding to the processing model itself, its analysis principles and dependent resources are all controllable. For example, more different voice analysis models can be integrated in the Bot platform (Bot Platform) corresponding to some voice assistants to perform different voice analysis tasks. That is, the analysis time corresponding to these analysis tasks in actual applications is often controllable, and technicians can adjust according to the actual scene requirements to meet the timeliness requirements of the application scenario. Therefore, in actual applications, the analysis tasks performed by the service platform corresponding to the processing model itself can be divided into the first category of analysis tasks. On the other hand, for the analysis tasks performed by third-party service providers. For example, some voice services such as music, weather, takeaway and taxi of voice assistants are often provided by some third-party service providers. When the corresponding data processing is required, the data to be processed is sent to the corresponding third-party service provider, and the third-party service provider returns the corresponding analysis results after completing the data processing. Because the situation of the third-party service provider is uncontrollable. As a result, in actual applications, there will be large differences in the analysis principles, dependent resources and data transmission conditions of each analysis task provided by the third-party service provider. As a result, these analysis tasks often take a long time. Therefore, in practical applications, the analysis tasks performed by third-party service providers can be classified as the second type of analysis tasks.
[0122] In addition, under different application scenarios, there may be certain differences in the executors of the analysis tasks, so that there may also be certain differences in the corresponding analysis task triggering methods and the acquisition operations of the analysis results. For example, for the analysis tasks that the electronic device itself is responsible for, its executor is the electronic device itself. Correspondingly, triggering the analysis task at this time means that the electronic device itself starts to execute the corresponding analysis task, and the acquisition operation refers to obtaining the analysis results generated by the electronic device itself. For some analysis tasks that are responsible for third-party devices, the executor is the third-party device that is specifically responsible for the analysis task. For example, for the analysis tasks that a third-party service provider is responsible for, the executor is the device corresponding to the third-party service provider. Correspondingly, triggering the analysis task at this time means sending the data to be processed to the third-party device to inform the third-party device to analyze the received data to be processed. The acquisition operation refers to waiting for the third-party device to return the analysis results. Therefore, the embodiment of the present application does not limit the specific analysis task triggering method, which needs to be determined according to the actual analysis task situation.
[0123] Based on the above description, when the data to be processed is obtained, the embodiment of the present application triggers corresponding parallel one or more first analysis tasks and one or more second analysis tasks, so that each analysis task independently performs data analysis on the data to be processed.
[0124] S102, obtaining a data analysis model corresponding to each second analysis task, and inputting the data to be processed into the data analysis model for processing, wherein the data analysis model is used to predict the result of the corresponding second analysis task.
[0125] In actual situations, multi-task data analysis will only select an analysis result with the highest priority as the final output result. Although each analysis task may generate many analysis results, only one determined analysis result will be generated in the end to participate in the priority sorting of the final output results. On this basis, when the electronic device has obtained one or more real results, for a single analysis task that has not generated an analysis result, there may be two situations:
[0126] Case 1: The priority of the analysis result finally generated by the analysis task is higher than all existing real results. This means that the analysis task may affect the final output result.
[0127] Case 2: The priority of the analysis result finally generated by the analysis task is not higher than all existing true results, that is, there are true results with higher priorities. In this case, the analysis task will not affect the final output result.
[0128] When the analysis task belongs to situation 2, it means that on the basis of the existing real results, whether or not to wait for the analysis result of the analysis task, it will not affect the final output result. It is based on this principle, in order to reduce the total task waiting time. The embodiment of the present application will simultaneously predict the analysis results of each second analysis task when the second analysis task is triggered, as a reference data for subsequent judgment on whether to wait for the analysis task. Among them, the embodiment of the present application does not limit the method of result prediction, and can be selected or set by technical personnel according to actual needs.
[0129] As an optional embodiment for predicting the analysis results of the present application. The present application embodiment will pre-train the corresponding data analysis model for each second analysis task to simulate the corresponding second analysis task for data analysis, thereby realizing the prediction of the result of the second analysis task. For example, assuming that the takeaway, map and music skills corresponding to the voice assistant are all second analysis tasks provided by a third-party service provider, and are all second analysis tasks required to be triggered by the present application embodiment. At this time, the present application embodiment will pre-train the corresponding takeaway processing model, map processing model and music processing model for takeaway, map and music. These data analysis models will simulate the analysis principle of the corresponding second analysis task to realize the processing of the data to be processed. Among them, the present application embodiment does not impose too many restrictions on the model type and training method of the data analysis model, which can be selected or set by the technician according to actual needs. For example, in some optional embodiments, the data analysis model can be constructed and trained based on knowledge distillation and transfer learning. So that the trained data analysis model is close to the output distribution of the corresponding second analysis task, thereby realizing the simulation of the second analysis task.
[0130] As another optional embodiment of the analysis result prediction of the present application. In order to ensure the credibility of the prediction result, two optional result prediction methods are included in the embodiment of the present application:
[0131] a. Obtain all possible analysis results of the second analysis task on the data to be processed, sort them by priority, and then use the analysis result with the highest priority as the prediction result.
[0132] At this time, for the data analysis model, it is not necessary to simulate the second analysis task to perform data analysis. Instead, it is necessary to pre-set the second analysis task corresponding to the data analysis model, as well as the analysis results that the second analysis task may generate for various data to be processed. At the same time, set the corresponding priority sorting rules. On this basis, when the data to be processed is obtained, the data analysis model executes the operation of prediction method a. The priority sorting rules are not limited here and can be selected or set by technical personnel according to actual scenario requirements.
[0133] b. Obtain all possible analysis results of the second analysis task on the data to be processed, and use the most likely analysis result as the prediction result.
[0134] At this time, for the data analysis model, there is no need to simulate the second analysis task to perform data analysis. Instead, the data analysis model needs to set up the corresponding second analysis task in advance, as well as the analysis results that the second analysis task may generate for various data to be processed. At the same time, the analysis method that is most likely to generate the analysis results is set. On this basis, when the data to be processed is obtained, the data analysis model executes the operation of prediction method b. The specific analysis method that is most likely to generate the results is not limited here and can be selected or set by the technician. For example, in some embodiments, analysis and prediction can be performed based on the record of historically generated analysis results of the second analysis task. In other embodiments, analysis and prediction can also be performed based on the historical records of analysis results generated by other tasks similar to the second analysis task.
[0135] In the embodiments of the present application, any one of the above result prediction methods can be selected as the result prediction method actually used.
[0136] At the same time, the embodiment of the present application does not limit the storage method of the data analysis model. It can be stored in an electronic device that serves as the execution subject of the embodiment of the present application. At this time, it is only necessary to call the data analysis model stored in the electronic device to perform data analysis. It can also be stored in other devices, such as some servers. At this time, if the data analysis model needs to be used, the electronic device needs to send the data to be processed to the corresponding device to achieve the result prediction of the second analysis task. Corresponding to this situation, the operation of obtaining the data analysis model corresponding to each second analysis task in S102 refers to the electronic device finding out the device corresponding to each data analysis model required to be used.
[0137] The embodiment of the present application will search for the data analysis model corresponding to each triggered second analysis task, and when each second analysis task is triggered, the data to be processed is synchronously input into each data analysis model. These data analysis models are used to predict the analysis results of each second analysis task.
[0138] S1031, obtaining a first result generated by a first analysis task and a second result generated by a data analysis model. If all results generated by the first analysis task and the data analysis model are obtained, the obtained first results and second results are prioritized to obtain corresponding sorting results.
[0139] In the embodiment of the present application, the real result generated by the first analysis task is uniformly named as the first result. The predicted result generated by the data analysis model is uniformly named as the second result. At the same time, the real result generated by the second analysis task is uniformly named as the third result. From the above analysis, it can be seen that the second result in the embodiment of the present application is actually the predicted result of the third result.
[0140] After triggering each analysis task and inputting the data to be processed into the data analysis model, the embodiment of the present application will wait for each analysis task and the processing data model to generate their own analysis results. However, the difference from the prior art is that the embodiment of the present application does not prioritize the analysis results after all analysis tasks have generated analysis results. Instead, after all the first analysis tasks and all the data analysis models have generated their own analysis results, the generated results are prioritized. Since the data analysis model is a prediction of the analysis results of the second analysis task, the priority of the second result in the sorting result can reflect the priority of the corresponding third result. And then the simulation sorting of each third result can be achieved.
[0141] Among them, the embodiments of the present application do not limit the specific sorting method, which can be selected or set by the technician according to the actual scene requirements. For example, in some embodiments, some fixed sorting rules can be preset by the technician according to the needs of the actual scene. On this basis, the priority sorting in the embodiments of the present application can be implemented according to these preset sorting rules. In another embodiment, some models with sorting functions can also be used to implement the priority sorting in the embodiments of the present application. For example, various machine learning sorting, such as lamdda MART machine learning sorting and Learning to Rank machine learning sorting, can be used to implement the analysis result sorting in the embodiments of the present application. At the same time, according to the different priority sorting methods actually used, the priority of the analysis result in the embodiments of the present application can also have a variety of possible forms. For example, in some optional embodiments, it is assumed that a sorting model is used for priority sorting. At the same time, it is assumed that the results output by the sorting model are the priority levels of each analysis result, such as the first priority, the second priority and the third priority. At this time, the priority in the embodiments of the present application refers to these priority levels. In other optional embodiments, it is assumed that a scoring model is used to implement the priority sorting in the embodiments of the present application. At the same time, it is assumed that the scoring model will score each input analysis result, and after all analysis results are scored, the scores corresponding to the analysis results are output in order from high to low. At this time, the priority is not the priority level in the previous embodiment, but refers to the order of the score corresponding to a certain analysis result among all the analysis result scores. For example, assuming that the score of analysis result A is 100 points, and the scores of the other analysis results are all lower than 100 points, then the priority of analysis result A is the first.
[0142] Considering that the priority sorting of analysis results in actual applications may be very complicated. In particular, in some scenarios with strong real-time performance, the priority level itself changes in real time. For example, in the scenario of product recommendation to users, the user's preference for products will change, so the priority level corresponding to each product changes in real time. For example, in the human-computer dialogue scenario, if a continuous dialogue is conducted with the user, the meaning of the same noun represented in different contexts may be different. At this time, the priority level of the analysis result corresponding to the same noun will also change to a certain extent. For example, some nouns are place names and may also be song nouns. At this time, it is necessary to determine whether the user's needs are to query place names or song nouns based on the content of the previous dialogue, and then determine the corresponding priority level based on the actual user needs. In order to cope with the priority sorting requirements in these scenarios, as an optional embodiment of the present application, an independent sorting model or functional module can be set to perform the priority sorting operation of the analysis results. Therefore, the present application embodiment does not limit the sorting method of the analysis results, including but not limited to: sorting by the processing model when performing data analysis. Alternatively, the processing model collaborates with an independent sorting model or functional module, and the processing model outputs the first result to the sorting model or functional module for priority sorting, and returns it to the processing model.
[0143] It should be noted that in actual applications, not all analysis tasks and data analysis models can normally analyze and respond to the data to be processed. For example, for an analysis task or data analysis model that provides voice weather query, assume that the data to be processed is "Remind me to bring an umbrella, the weather is bad today" said by the user. At this time, what the user actually needs is a reminder of matters, so in theory, the analysis task or data analysis model cannot respond normally. In order to deal with this situation, there are generally two optional processing methods: 1. Strong response to the data to be processed, that is, matching a result with the highest correlation for output. As in the above example, the weather data of the day can be generated as the corresponding analysis result. 2. Output a default analysis result to declare that it cannot respond to the data to be processed. Among them, the specific data content of the default analysis result is not limited here, and can be set by the technician according to actual needs. For example, the default analysis result can be set to "unable to respond" or "unknown". In combination with the above example, it is assumed that the default analysis result is set to "unable to respond". At this time, the analysis task or data analysis model will generate an "unable to respond" analysis result to declare that it cannot respond to the data to be processed. Therefore, in the embodiment of the present application, if method 2 is selected to generate the result when the response cannot be performed normally. At this time, the result generated by the first analysis task and the data analysis model may also be the above-mentioned default analysis result. In this case, when the embodiment of the present application sorts the first result and the second result, these default analysis results will be eliminated first to improve the efficiency of the sorting.
[0144] As an optional embodiment of the present application, considering that in actual applications, before all first analysis tasks and data analysis models generate analysis results, some second analysis tasks may return analysis results. That is, before sorting, some third results generated by the second analysis task may also be obtained. If the third result is obtained, it is necessary to determine whether to add the third result to the sorting when sorting. In the embodiment of the present application, any one of the following methods can be selected for sorting:
[0145] 1. Only the first and second results are sorted.
[0146] 2. Comprehensively sort all the first results, second results and third results obtained.
[0147] 3. Eliminate the second result corresponding to the acquired third result, and comprehensively sort the first result, second result, and third result remaining after the elimination operation.
[0148] For sorting mode 1, the operation step is S1031. At this time, since the sorting content includes the predicted result corresponding to the third result, the reliable sorting of the third result can be guaranteed.
[0149] For sorting method 2, the sorting results will include both the obtained third results and all the predicted second results. Figure 1C , the above S1031 can be replaced by:
[0150] S1032, obtaining a first result generated by the first analysis task, a second result generated by the data analysis model, and a third result generated by the second analysis task. If all the results generated by the first analysis task and the data analysis model are obtained, the obtained first result, second result, and third result are prioritized to obtain a corresponding sorting result.
[0151] For sorting method 3, the corresponding data analysis model will be determined based on the second analysis task to which the third result belongs, and then the second results corresponding to these models will be found. Then the second result corresponding to the third result will be determined. By eliminating the second result corresponding to the third result that has been obtained, the workload of sorting can be reduced. Corresponding to sorting method 2, refer to Figure 1D , the above S1031 can be replaced by:
[0152] S1033, obtaining the first result generated by the first analysis task, the second result generated by the data analysis model, and the third result generated by the second analysis task. If all the results generated by the first analysis task and the data analysis model are obtained, the second result corresponding to the obtained third result is eliminated, and the first result, the second result, and the third result remaining after the elimination operation are prioritized to obtain the corresponding sorting result.
[0153] In addition to the above three sorting methods, technicians can also select or set other sorting methods according to actual needs to process the obtained third result. The present application embodiment does not impose too many restrictions on the sorting methods.
[0154] S1041: If the result with the highest priority in the sorting results is the first result, the first result with the highest priority is used as the output result.
[0155] When the result with the highest priority in the sorting result is the first result, it means that for the second analysis task that has not generated an analysis result, the priority of the analysis result finally generated must be lower than the real result with the highest priority, that is, it will not affect the final output result. Therefore, in this case, the embodiment of the present application will determine that there is no need to wait for the second analysis task to generate an analysis result, and use the first result with the highest priority as the final output result.
[0156] For example, assuming that five analysis results are obtained in S1031, the priorities are as follows: first result A, second result B, first result C, second result D, and second result E. At this time, the embodiment of the present application will use first result A as the final output result.
[0157] For the cases of S1032 and S1033, refer to Figure 1C and Figure 1D , then S1041 can be replaced by:
[0158] S1042: If the result with the highest priority in the sorted results is the first result or the third result, use the first result or the third result with the highest priority as the output result.
[0159] Since the third result is the same as the first result, both are true results. Therefore, when the analysis result with the highest priority is the first result or the third result, it means that the second analysis task that has not generated an analysis result currently does not need to wait. Therefore, after S1032 and S1033, the operation of S1042 can be executed.
[0160] S105: If the result with the highest priority in the sorting result is the second result, a third result corresponding to the second result with the highest priority is obtained, and the third result is used as the output result. The third result is an analysis result generated by the target analysis task, the target analysis task is the second analysis task corresponding to the target data analysis model, and the target data analysis model is the data analysis model to which the second result with the highest priority belongs.
[0161] When the result with the highest priority in the sorting results is the second result, it means that the second result is the most suitable output result. However, since the second result is a predicted result for the second analysis task, it is not the actual result generated by the second analysis task. Therefore, the second result cannot be directly used as the output result, but it is necessary to wait for the third result corresponding to the second result to be obtained before the corresponding third result is used as the output result. Among them, the determination operation of the third result corresponding to the second result may include: first determining the data analysis model to which the second result belongs, then finding the second analysis task corresponding to the data analysis model, and waiting for the second analysis task to generate the corresponding third result.
[0162] It should also be noted that the output object of the output result in the embodiment of the present application needs to be determined according to the actual application situation to meet the actual application needs, including but not limited to: 1. Other functional modules or processing layers of the processing model itself. 2. Other models that collaborate with the processing model. 3. Other devices other than the device as the execution subject. 4. Output directly from the device to the user, such as voice broadcast.
[0163] In the embodiment of the present application, the pre-trained data analysis model is first used to predict the results of the second type of analysis tasks. After all the first type of analysis tasks and the data analysis model have generated corresponding analysis results, the generated sub-results are prioritized. Finally, it is determined whether it is necessary to continue waiting for the second analysis task to generate the analysis results according to the situation of the sorting results. Since the operations such as the result prediction and sorting of the second type of analysis tasks are performed in the process of waiting for the second type of analysis tasks to generate the analysis results, it is a reuse of the waiting time without adding additional waiting time. When the result with the highest priority is the true result, the true result can be used as the output result. At this time, for the unfinished second type of analysis tasks, there is no need to wait for them to generate the analysis results. Therefore, compared with the traditional waiting for all analysis tasks to generate the analysis results, a large amount of waiting time can be reduced, and the processing efficiency can be improved. When the result with the highest priority is the predicted result, it is only necessary to wait for the generation of the true result corresponding to the predicted result. Compared with the traditional waiting for all analysis tasks to generate the analysis results, the waiting for most of the second analysis tasks is abandoned. The total waiting time of the embodiment of the present application is greatly reduced, and the processing efficiency is improved.
[0164] Figure 2A The following is a flowchart of the implementation of the data analysis method based on multiple analysis tasks provided in the second embodiment of the present application, which is described in detail as follows:
[0165] S201, obtaining data to be processed, triggering at least one first analysis task and at least one second analysis task in parallel, wherein the first analysis task and the second analysis task are both used to analyze the data to be processed.
[0166] Among them, the specific operating details and principles of S201 are Figure 1B The step S101 in the illustrated embodiment is the same, so reference may be made to the description of S101 , which will not be described in detail here.
[0167] S202, obtaining a data analysis model corresponding to each second analysis task, and inputting the data to be processed into the data analysis model for processing, wherein the data analysis model is used to predict the result of the corresponding second analysis task.
[0168] Among them, the specific operating details and principles of S202 are Figure 1B The step S102 in the illustrated embodiment is the same, so reference may be made to the description of S102 and no further description is given here.
[0169] S2031, obtaining the first result generated by the first analysis task and the second result generated by the data analysis model. Obtain the confidence level corresponding to each second result. If all the results generated by the first analysis task and the data analysis model are obtained, prioritize the obtained first results and second results to obtain the corresponding first sorting result.
[0170] Among them, the specific operating details and principles of S2031 are Figure 1B In the illustrated embodiment, step S1031 is basically the same, so the relevant description of S1031 can be referred to, and will not be repeated here. Here, only the differences between S2031 and S1031 are described as follows:
[0171] In actual applications, it is found that the ability of each analysis task to respond to data analysis may vary. For the processing of a piece of data to be processed, the analysis results output by a single analysis task may have the following situations:
[0172] 1. The analysis result has a high degree of match with the data to be processed, and can accurately respond to the data to be processed. For example, in the voice assistant, the user's voice content is "today's weather". For the analysis task responsible for weather query, if the analysis result is the weather data of the day, it means that the analysis result has a high degree of match with the data to be processed.
[0173] 2. The analysis result has a low match with the data to be processed, so it cannot accurately respond to the data to be processed. For example, suppose in the voice assistant, the user's voice content is "Recommend a delicious hot pot nearby." For the analysis task responsible for food query, if the analysis result is a bakery with a high score near the user, it means that the analysis result has a low match with the data to be processed.
[0174] 3. The analysis result does not match the data to be processed. This is because the analysis task obtains the result by making a strong response to the data to be processed. At this time, the data to be processed cannot be responded to normally. For example, suppose that in the voice assistant, the user's voice content is "Recommend a delicious hot pot nearby." For the analysis task responsible for weather query, it is theoretically impossible to obtain effective analysis results. However, in actual applications, third-party service providers may set a strong response strategy for these irrelevant data to be processed in order to increase the user's usage of their own services and increase their own traffic. For example, the analysis result of the weather on that day may be generated. Since the weather on that day is not related to "Recommend a delicious hot pot nearby", the analysis result does not match the data to be processed at this time.
[0175] As can be seen from the above description, for different analysis tasks, there may be certain differences in the matching degree between the generated analysis results and the data to be processed. These differences will affect the final response to the data to be processed. In order to improve the matching degree between the final output result and the data to be processed, when performing multi-analysis task data processing in the embodiment of the present application, the matching degree between the analysis results generated by each analysis task and the data to be processed will be considered simultaneously. Specifically, these matching degree situations will be quantified in the form of confidence. Among them, the embodiment of the present application does not limit the data form of the confidence, which can be set by the technician. For example, the confidence of the second result can be set to a specific numerical value, or multiple confidence levels can be set first, and the confidence of the second result can be set to a specific confidence level.
[0176] Since the data analysis model is a simulated prediction of the second analysis task. Its purpose is to get as close to the input and output of the second analysis task as possible, rather than to improve the accuracy and efficiency of the analysis results. That is, the confidence of the analysis results of different data analysis models may be different. Therefore, when obtaining the second result, the embodiment of the present application will also obtain the confidence of each second result to provide data for subsequent analysis result sorting and screening. Among them, the embodiment of the present application does not impose too many restrictions on the acquisition method of the confidence of the second result, which can be selected or set by the technician. In some optional embodiments, the confidence of each second result can be generated by the data analysis model itself. At this time, it is necessary to detect the confidence of the actual output analysis result of the analysis model according to the trained sample data when training the data analysis model. In other optional embodiments, the historical data to be processed and the analysis results of each second analysis task can also be matched. Analysis is performed to determine the confidence of the analysis result corresponding to each second analysis task. Then the confidence is set to the confidence of the corresponding second result.
[0177] As an optional embodiment of the present application, S2031 may also be replaced by S2032 or S2033, wherein S2032 and S2033 are as follows:
[0178] S2032, obtaining a first result generated by the first analysis task, a second result generated by the data analysis model, and a third result generated by the second analysis task. Obtain the confidence level corresponding to each second result. If all the results generated by the first analysis task and the data analysis model are obtained, prioritize the obtained first results, second results, and third results to obtain a corresponding first sorting result.
[0179] S2033, obtaining the first result generated by the first analysis task, the second result generated by the data analysis model, and the third result generated by the second analysis task. Obtain the confidence level corresponding to each second result. If all the results generated by the first analysis task and the data analysis model are obtained, the second result corresponding to the obtained third result is eliminated, and the first results, second results, and third results remaining after the elimination operation are prioritized to obtain the corresponding first sorted result.
[0180] The specific operation details and principles of S2032 and S2033 are basically the same as those of S1032 and S1033, so reference may be made to the relevant description of S1032 and S1033, which will not be described in detail here. The difference from S1032 and S1033 is that S2032 and S2033 will synchronously obtain the confidence of each second result. For the relevant description of confidence, reference may be made to the description in S2031, which will not be described in detail here.
[0181] As an optional embodiment of the present application, considering that the confidence level is too low, it means that the analysis result has a low match with the data to be processed, and may even be the result of a strong response. Therefore, in order to improve the effectiveness of the sorting of the analysis results and improve the efficiency of the analysis of the data to be processed. In an embodiment of the present application, S2031, S2032, and S2033 can first eliminate the second result whose confidence level is lower than the preset second threshold before prioritizing. Then sort the remaining analysis results after elimination. The specific value of the second threshold can be set by the technician according to actual needs, and is not limited here.
[0182] S204, identifying whether the target result is included in the first n analysis results with the highest priority in the first sorting result, wherein the target result is the second result with a confidence level higher than the first threshold. n is an integer greater than 1. If the target result is not included in the first n analysis results, execute step S2051 or S206. If the target result is included in the first n analysis results, execute step S207 or S2081.
[0183] Because the confidence level can reflect the matching degree between the analysis result and the data to be processed. For the second results that are all prediction results, when the confidence level of the second result is high, it means that it is more likely to be the analysis result that matches the data to be processed. For the second result whose confidence level is not very high, it means that it must be the analysis result that matches the data to be processed. In the embodiment of the present application, the second results with higher confidence levels are uniformly named as target results. At the same time, in order to distinguish between high and low confidence levels, the embodiment of the present application will pre-set a first threshold. And the second result with a confidence level higher than the first threshold is judged as the second result with a higher confidence level, and the second result with a confidence level not higher than the first threshold is judged as the second result with a low confidence level. The specific value of the first threshold is not limited here, and can be set by the technician according to the data form of the confidence level and the actual needs of the application scenario. For example, when the confidence levels are all numerical values, the first threshold is also a numerical value. When the confidence level is a level, the first threshold is a specific level.
[0184] It should be noted that when combined with the above optional embodiment: "S2031, S2032 and S2033 can first eliminate the second results with confidence lower than the preset second threshold before priority sorting. Then sort the remaining analysis results after elimination." When applied. The second threshold should be lower than the first threshold. At this time, according to the first threshold and the second threshold, the second result can be divided into three states: high confidence, medium confidence and low confidence. High confidence means that the confidence of the second result is higher than the first threshold. Medium confidence means that the confidence of the second result is higher than the second threshold and lower than the first threshold. Low confidence means that the confidence of the second result is lower than the second threshold.
[0185] exist Figure 1BIn the illustrated embodiment, when the analysis results in the first sorting result are selected for output, only the priority situation is considered for all the analysis results. However, it is found in practice that for some second results with high confidence, the third result corresponding to them is more likely to be the analysis result actually matched by the data to be processed. However, since the data analysis model realizes the result prediction of the second analysis task by simulation or result evaluation, it is difficult to ensure that the predicted second result is exactly the same as the third result actually generated by the second analysis task every time. This leads to the priority situation of the second result, which may sometimes be different from the priority situation of the corresponding third result. Therefore, for the second result with higher confidence, if it is not selected as the output result in the end because of the difference in priority situation between it and the corresponding third result, it is possible that the second result is mistakenly discarded, which leads to a decrease in the accuracy of the final output result. But on the other hand, it is also found in actual application that the difference between the priority situation of the second result and the corresponding third result is generally not very large. Therefore, for some second results with lower priority and high confidence, the corresponding third result often has a lower priority.
[0186] Based on the above principle analysis, after obtaining the first sorting result, the embodiment of the present application will first identify whether the top n analysis results with the highest priority contain a second result with high confidence, and perform differentiated processing according to the identified results. At this time, there is no need to analyze the confidence level of the analysis results with lower priority. As a result, while ensuring the accuracy of the output results, the embodiment of the present application does not need to perform too much analysis and processing of the confidence level of the analysis results, so that the efficiency of the output result analysis can also be improved. Among them, the specific value of n can be set by the technician according to the actual scene requirements or experience, and is not limited here. For example, in some optional embodiments, it can be set to any value from 2 to 5.
[0187] S2051: If the result with the highest priority in the first sorting result is the first result, use the result with the highest priority as the output result.
[0188] When the current n analysis results do not include the second result with high confidence, and the analysis result with the highest priority is the first result, it means that the analysis result with the highest priority is the result that best matches the data to be processed. Therefore, at this time, the embodiment of the present application will use the analysis result with the highest priority as the final output result.
[0189] Corresponding to the situation of S2032 and S2033, S2051 can be replaced by:
[0190] S2052: If the result with the highest priority in the first sorting result is the first result or the third result, use the first result or the third result with the highest priority as the output result.
[0191] Since the third result is a true result like the first result, when the first sorting result includes the third result, and the analysis result with the highest priority is the first result or the third result, it means that the second analysis task that has not generated the analysis result does not need to wait. Therefore, after S2032 and S2033, the operation of S2052 can be executed.
[0192] S206: If the result with the highest priority in the first sorting result is the second result, a third result corresponding to the second result with the highest priority is obtained, and the third result is used as the output result. The third result is an analysis result generated by the target analysis task, the target analysis task is the second analysis task corresponding to the target data analysis model, and the target data analysis model is the data analysis model to which the second result with the highest priority belongs.
[0193] When the current n analysis results do not contain a second result with a high confidence level, and the analysis result with the highest priority is the second result. This indicates that the second result with the highest priority level currently has a high degree of match with the data to be processed. However, since the second result is a predicted result rather than a true result, the embodiment of the present application will obtain a third result corresponding to the second result as an output result. Among them, the specific determination operation of the third result is basically the same as S105, and the relevant description in S105 can be referred to, which will not be repeated here.
[0194] S207: If the result with the highest priority in the first sorting result is the target result, obtain the third result corresponding to the target result with the highest priority, and use the third result as the output result.
[0195] When the analysis result with the highest priority is the target result, it indicates that the target result has a high degree of match with the data to be processed. However, since the target result is a predicted result rather than a real result, the embodiment of the present application obtains the third result corresponding to the target result as the output result. Among them, the specific determination operation of the third result is basically the same as S105, and the relevant description in S105 can be referred to, which will not be repeated here.
[0196] S2081: If the result with the highest priority in the first sorting result is not the target result, obtain the third result corresponding to all the target results in the first n analysis results, and sort the first result in the first n analysis results, the second result with a confidence lower than the first threshold, and the obtained third result by priority to obtain the corresponding second sorting result.
[0197] Although the result with the highest priority in the first sorting result is not the target result, the target result is still in the top n analysis results with the highest priority in the first sorting result, which means that the target result and the analysis result with the highest priority do not have much difference in matching degree with the data to be processed. At the same time, it is considered that there may be certain differences in the priority ranking of the real result and the corresponding predicted result. That is to say, the third result corresponding to the target result still has a certain possibility of being a result that matches the data to be processed. At this time, if the target result is ignored and the result with the highest priority in the first sorting result is used as the output result, it is very likely that the target result will be mistakenly abandoned. This will lead to a reduction in the accuracy of the selection of the final output result.
[0198] In order to prevent the target result from being mistakenly discarded, the target result is included in the first n analysis results with the highest priority in the first sorting result, and the result with the highest priority is not the target result. The embodiment of the present application will obtain the third result corresponding to all the target results. After obtaining the third results corresponding to all the target results, the first result in the first n analysis results, the second result with a confidence level lower than the first threshold, and the obtained third result are prioritized again, and the corresponding second sorting result is obtained.
[0199] It should be noted that the embodiment of the present application will only wait for the third result corresponding to the target result, and will not wait for the third result corresponding to the second result of medium confidence. However, it is found in actual application that before the third results corresponding to all target results are obtained, the second result of medium confidence may also generate a third result. In order to cope with this situation, as an optional embodiment of the present application, before the third results corresponding to all target results are obtained, the second result of medium confidence also generates a third result. At this time, these third results can replace the corresponding second results of medium confidence and participate in the priority sorting in S2081. Alternatively, the corresponding second results of medium confidence may not be replaced, but these third results may be directly included in the sorting object. Specifically, the technician may select one of the methods for processing, or may set the actual processing method according to actual needs, which is not limited here.
[0200] S209: If the result with the highest priority in the second sorting result is the first result or the third result, use the first result or the third result with the highest priority as the output result.
[0201] After obtaining the second sorting result, if the result with the highest priority is the first result or the third result, since the first result and the third result are both true results, it means that the second analysis task that has not generated the analysis result does not need to wait, and the true result with the highest priority can be used as the output result.
[0202] S210: If the result with the highest priority in the second sorting result is the second result, obtain the third result corresponding to the second result with the highest priority, and use the third result as the output result.
[0203] If the second result has the highest priority in the second sorting result, it means that the second result has a high degree of match with the data to be processed. However, since the second result is a predicted result rather than a real result, the embodiment of the present application will obtain the third result corresponding to the second result as the output result. Among them, the specific determination operation of the third result is basically the same as S105, and the relevant description in S105 can be referred to, which will not be repeated here.
[0204] As an optional embodiment of the present application, when the result with the highest priority in the first sorting result obtained in S207 is not the target result, it is also possible to wait until the third result corresponding to all the second results in the first n analysis results are obtained, and then perform a second priority sorting. At this time, the sorted contents are all real results, and there is no need to consider the analysis result with the highest priority in the second sorting result. Correspondingly, refer to Figure 2B , S2081 can be replaced by S2082:
[0205] S2082: If the result with the highest priority in the first sorting result is not the target result, obtain the third results corresponding to all the second results in the first n analysis results, and sort the first results in the first n analysis results and the obtained third results by priority to obtain the corresponding second sorting result.
[0206] The difference between S2082 and S2081 is that it will wait for the third results corresponding to all the second results in the first n analysis results, rather than just waiting for the third result corresponding to the target result. Therefore, there is no need to perform the judgment operations of S209 and S210 after S2082. Correspondingly, after S2082, the operation of S211 is executed:
[0207] S211: Take the result with the highest priority in the second sorting result as the output result.
[0208] Since the second priority sorting content is the first result and the third result, these results are all true results. Therefore, there is no need to wait any longer, and the result with the highest priority can be used as the output result.
[0209] Considering that there may be differences between the predicted results and the actual results in actual situations, there may also be differences in the priority between the second result and the third result. On this basis, when the second result with high confidence has a higher priority but is not the highest priority, if the result with the highest priority is directly used as the output result. The second result with high confidence may be mistakenly abandoned. This leads to a decrease in the accuracy of the output result. In order to prevent this from happening, after obtaining the first sorting result of the first result and the second result, the embodiment of the present application will determine whether the first n results with the highest priority contain the second result with high confidence. And when the second result with high confidence is included and the second result with high confidence is not the result with the highest priority, wait for the third result corresponding to the second result with high confidence to be generated. Finally, the first result in the first n results and the third result obtained are sorted twice, and the result with the highest priority is used as the output result. Thus, the embodiment of the present application can realize the reordering of the third result corresponding to the second result with high confidence, and prevent the mistaken abandonment of the second result with high confidence. This improves the accuracy of the final output result.
[0210] At the same time, since the operations of the above-mentioned embodiments of the present application are all performed in the process of waiting for the second type of analysis task to generate the analysis result, the waiting time is reused without adding an additional waiting time. And according to the actual priority sorting situation, the embodiments of the present application can abandon the waiting for most of the second analysis tasks (and may even abandon the waiting for all the second analysis tasks), so that the total waiting time of the embodiments of the present application is greatly reduced, and the processing efficiency is improved.
[0211] Take an example Figures 1B to 2B The embodiments shown are used as examples for illustration. Figure 2C . In an embodiment of the present application, it is assumed that there are a total of four analysis tasks, namely, analysis task 1, analysis task 2, analysis task 3 and analysis task 4 (more or fewer analysis tasks may be included in actual applications). Among them, analysis task 4 is a second type of analysis task provided by a third-party service provider. At the same time, n=2 is assumed. On this basis, the electronic device provides a unified data access interface to obtain the data to be processed, and will parse the data to be processed when the data to be processed is obtained, and determine the corresponding related threads to be started. Then execute the corresponding threads. After the thread is executed, multiple independent analysis tasks corresponding to the data to be processed are triggered. At the same time, a data analysis model for simulating analysis task 4 will be determined, and the data analysis model will be used to simulate analysis task 4 to analyze the data to be processed. After triggering multiple analysis tasks, on the one hand, analysis tasks 1 to 3 will generate corresponding first results. On the other hand, the data analysis model will obtain the predicted second result. On the other hand, analysis task 4 may also generate a real third result.
[0212] When analysis tasks 1 to 3 and the data analysis model generate analysis results, if analysis task 4 has not yet generated a result. At this time, the embodiment of the present application will sort the results generated by analysis tasks 1 to 3 and the data analysis model, and select the results. The selection rules are:
[0213] When the two results with the highest priority do not include the result generated by the data analysis model, the result with the highest priority is used as the final output result.
[0214] When the two results with the highest priority include the results generated by the data analysis model, wait for the result of analysis task 4 to return the result, and then re-sort and select the selected results.
[0215] When analysis tasks 1 to 3 and the data analysis model all generate analysis results, if analysis task 4 has generated a third result, the confidence level of the third result will be determined based on the first threshold and the second threshold.
[0216] When the confidence of the third result is high or medium, the third result will be retained and sorted with the results generated by analysis tasks 1 to 3. When the confidence of the third result is low, the third result will be discarded and the results of analysis tasks 1 to 3 will be sorted. After sorting, the result with the highest priority will be used as the output result.
[0217] After the output result is determined, the result is executed (i.e., the output result is outputted). After the execution result, the processing of the next data to be processed is continued.
[0218] Among them, a master control module is provided in the electronic device for controlling the execution of the above-mentioned operations such as triggering of multiple analysis tasks, sorting of results, selection of results and execution of results.
[0219] right Figure 1B The illustrated embodiments and Figure 2A Some relevant descriptions of the illustrated embodiment are as follows:
[0220] 1. When the second result is determined to be a result with a higher matching degree in S105, S206, S207 and S210, the operations of obtaining a third result corresponding to the second result and using the obtained third result as the output result can all be replaced by:
[0221] Based on the data to be processed, a cache result corresponding to the second result is obtained, and the obtained cache result is used as the output result.
[0222] Considering that there are many uncertain factors in the second type of analysis task, it may sometimes take a long time to get the analysis result of the second analysis task. Therefore, when waiting for the real result corresponding to the predicted result, it may also be necessary to wait for a long time. In order to improve this situation, after obtaining the analysis result of the second type of analysis task, the embodiment of the present application will associate the analysis result with the data to be processed at that time for cache. When it is necessary to obtain the analysis result of the second type of analysis task, the embodiment of the present application does not need to wait for the second type of analysis task to generate the analysis result, but reads the corresponding cache result. Among them, in order to ensure the accuracy of the response to the data to be processed, some corresponding matching rules can be set when reading the cache result. For example, in some embodiments, it can be set to require that the data to be processed at that time and the cached data to be processed are the same, and the analysis result corresponding to the cached data to be processed can be used as the cached result corresponding to the second result. In other embodiments, in order to ensure the timeliness of the cached result, a cache time limit can also be added on the basis of the same data to be processed. That is, when the data to be processed is the same and the cache time is close, the analysis result corresponding to the cached data to be processed is used as the cached result corresponding to the second result.
[0223] As an optional embodiment of the present application, the above-mentioned waiting for the real result and obtaining the cached result can also be applied in combination. That is, a waiting time threshold is pre-set. When S105, S206, S207 and S210 determine that the second result is a result with a higher matching degree, first wait for the third result corresponding to the second result. If the waiting time is longer than the waiting time threshold, the waiting is ended and the cached result corresponding to the second result is read. At this time, the real result can be used as the output result as much as possible, and the problem of the second analysis task taking too long causing the entire data analysis process to take too long can be prevented, thereby ensuring the efficiency of the processing model.
[0224] 2. About Figure 1B The illustrated embodiments and Figure 2A Several points about the data analysis model in the illustrated embodiment are as follows:
[0225] 1. The teacher-student model based on knowledge distillation can be used to build and train data analysis models.
[0226] Knowledge distillation is a kind of same-task transfer learning, in which transfer learning is a branch of machine learning, and transfer learning is to transfer the performance of one model to another model. The teacher-student model in the embodiment of the present application is a kind of transfer learning. In the embodiment of the present application, the teacher model in the teacher-student model is the second type of analysis task, and the student model is the data analysis model in the embodiment of the present application.
[0227] Since the teacher-student model can realize the learning from the teacher model to the student model, the student model can imitate the behavior of the teacher model and achieve considerable accuracy. Therefore, the data analysis model obtained in the embodiment of the present application can realize the effective simulation of each second-class analysis task, and ensure the effective prediction and priority sorting analysis of the classification results of each second-class analysis task. At the same time, knowledge distillation can make the student model more lightweight, thereby making the volume of the data analysis model smaller, and the model efficiency is higher for the corresponding second-class analysis tasks. Thereby, the data analysis model in the embodiment of the present application can generate analysis results faster while ensuring accuracy.
[0228] 2. During the training of the data analysis model, a classifier can be added to identify whether the output result is accurate. Based on the classifier, it is judged whether the results generated during the model training process are accurate, and the feedback is fed back to the data analysis model for parameter update training to reduce the situation where the data analysis model has a strong response.
[0229] Depend on Figure 1B The illustrated embodiments and Figure 2A It can be seen from the description in the illustrated embodiment that a strong response may occur in an analysis task. That is, in order to respond to the data to be processed, the analysis task will generate an analysis result that is unrelated to the data to be processed. An analysis result with a strong response may lead to certain errors in the final prioritization of the analysis results and the determination of the output results. For example, when an analysis result with a strong response is determined as the output result, the output of the processing model will be inaccurate. As a simulation model for the second type of analysis task, if only traditional transfer learning is used to train the data analysis model, the obtained data analysis model may also have a strong response.
[0230] In order to improve the situation of strong response of the data analysis model, a classifier that can analyze the accuracy of the analysis results will be pre-trained in the embodiment of the present application. Among them, the construction and training method of the classifier are not limited here, and can be selected or set by the technician according to actual needs. For example, in some optional embodiments, the pending data and the corresponding analysis results of the historical processing of the second type of analysis task can be recorded, and whether it is a strong response can be marked. On this basis, it can be determined by querying the recorded data to determine whether the analysis result generated by the data analysis model each time is a strong response result. In other optional embodiments of the present application, the pending data and the corresponding analysis results of the historical processing of the second type of analysis task can also be used as sample data. After marking whether the sample data is a strong response, the model training of the classifier is performed to obtain a classifier model that can analyze the strong response of the analysis results to the processed data. Among them, the embodiment of the present application does not limit the model type of the classifier, and can be selected or set by the technician according to actual needs. For example, the Deep Gradient Boosting Machine (DeepGBM) model can be selected as the classifier model.
[0231] On the basis of training the classifier, when the embodiment of the present application performs model training on the data analysis model, each time the data analysis model generates an analysis result, the classifier is used to determine whether the analysis result is a strong response result. The judgment result is used as the feature data for the next iteration and update of the data analysis model parameters. Until the data analysis model converges, the model training is completed.
[0232] As another optional implementation of the present application, a more accurate classifier can also be constructed. That is, the classifier can not only identify whether the analysis result is a strong response, but also analyze the accuracy of the analysis result response to the data to be processed. Among them, the principle of construction can refer to the aforementioned description of the construction of the classifier for identifying whether it is a strong response. At this time, when marking the analysis result as sample data, it is necessary not only to mark whether it is a strong response but also to mark its corresponding accuracy.
[0233] In the embodiment of the present application, a classifier that can identify whether the analysis result is a strong response is added during the data analysis model training process, and feedback on the update of the data analysis model parameters is realized according to the judgment result of the classifier. The trained data analysis model is less likely to have a strong response than the original second type of analysis task. This improves the accuracy of the final output result.
[0234] 3. The data analysis model can be updated online, so that the data analysis model can better adapt to the dynamic changes in the matching relationship between the data to be processed and the analysis results, and realize the effective simulation of the second type of analysis tasks.
[0235] Because in actual situations, the matching relationship between the data to be processed and the analysis results may have certain dynamic changes. As a result, the best response analysis results corresponding to the same data to be processed at different times may have certain differences. For example, in the human-computer interaction scenario, for the analysis task of the takeaway function. The same conversation content "recommendations of delicious takeaways around" generally has certain differences in the corresponding recommended content at different times. For example, for some analysis tasks of personalized push functions based on user preferences, such as personalized music push for users. For the same user information, the content pushed at different times may also be different. Therefore, for a single second-category analysis task, its own response to the data to be processed may undergo certain dynamic changes. In order to ensure the simulation effect of the data analysis model on the second-category analysis task, based on the existing data analysis model, the embodiment of the present application will also update and train the data analysis model online, as described in detail as follows:
[0236] exist Figure 1B The illustrated embodiments and Figure 2A While the illustrated embodiment is being executed, the embodiment of the present application will wait for each second-type analysis task to generate a corresponding third result. When the third result generated at the time is obtained, the third result and the data to be processed at the time will be used as sample data to perform parameter update training on the corresponding data analysis model. The updated data analysis model can be used to process the next data to be processed to simulate a real-time second-type analysis task.
[0237] Take an example to illustrate, for example, assuming that data analysis model A is a model that simulates the second analysis task B. Figure 1B The illustrated embodiments and Figure 2A While the embodiment shown is being executed, regardless of Figure 1B The illustrated embodiments and Figure 2A The embodiment shown actually determines whether the result needs to wait for the second analysis task B to generate an analysis result. The embodiments of the present application will wait for the second analysis task B to process the current data to be processed C and generate a third result D. After obtaining the third result D, the embodiment of the present application will use the data to be processed C and the third result D as sample data to perform parameter update training on the data analysis model A. The updated data analysis model A can be used to process the next data to be processed E to simulate the processing of the data to be processed E by the second analysis task B.
[0238] In the embodiment of the present application, by continuously learning and updating the data analysis model online, the simulation effect of the data analysis model on the second type of analysis task can be guaranteed, and the dynamic change requirements of the matching relationship between the data to be processed and the analysis results can be adapted. Thus, the reliability of the analysis and processing of the data to be processed in the embodiment of the present application can be guaranteed.
[0239] 4. Based on point 3, the third result and the corresponding data to be processed can be analyzed each time. The third result is judged based on the classifier whether it is accurate, and the judgment result is used as input data for online updating of the data analysis model to update the data analysis model online, so as to reduce the situation where the data analysis model has a strong response.
[0240] It can be seen from the description of the above points that the analysis result generated by the analysis task may be a result of a strong response. The analysis result generated at this time does not match the data to be processed. In the process of online updating of the data analysis model, if the third result of the time is the result of a strong response. At this time, the online update of the data analysis model based on the third result of the strong response may cause a strong response to occur in the subsequent data analysis process of the data analysis model, which will lead to certain errors in the final priority sorting of the analysis results and the determination of the output results. In order to avoid this situation, the embodiment of the present application will first detect the third result and the data to be processed obtained at this time before each online update of the data analysis model to determine whether the third result is accurate, and provide at least the following three optional processing methods:
[0241] Method 1: After obtaining the third result, detect whether the third result matches the data to be processed (i.e., whether it is a strong response result; if not, it is determined to be a strong response result; if it matches, it is not a strong response result).
[0242] If it is a strong response result, the embodiment of the present application will not use the third result and the data to be processed as sample data to update the data analysis model. This is to ensure the effectiveness of the online update of the data analysis model. If it is not a strong response result, the third result and the data to be processed are used as sample data to update the data analysis model.
[0243] Among them, the embodiment of the present application does not limit the detection method of whether the specific third result and the data to be processed match, and can be selected or set by the technical staff according to actual needs. For example, in some embodiments, the data to be processed and the corresponding analysis results of the historical processing of the second type of analysis task can be recorded, and whether it is a strong response can be marked. In the application, the third result of the time and the data to be processed are queried according to the recorded data to see whether they are marked strong responses. If so, it is determined that the third result of the time does not match the data to be processed and is a strong response. Alternatively, a classifier can be constructed to identify whether the output result is accurate, and the classifier can be used to implement matching detection.
[0244] Method 2: After obtaining the third result, detecting whether the third result matches the data to be processed.
[0245] If the result is a strong response, negative feedback data is generated as feature data for the next iteration of the data analysis model parameter update until the data analysis model converges and the model training is completed.
[0246] If it is not a strong response result, positive feedback data is generated as feature data for the next iteration of the data analysis model parameter update until the data analysis model converges and the model training is completed.
[0247] Method 3: After obtaining the third result, detecting whether the third result matches the data to be processed.
[0248] If the result is a strong response, negative feedback data is generated as feature data for the next iteration of the data analysis model parameter update until the data analysis model converges and the model training is completed.
[0249] If it is not a strong response result, the third result is evaluated for accuracy, and positive feedback data is generated based on the evaluation result, which is used as feature data for the next iteration of the data analysis model parameter update, until the data analysis model converges and the model training is completed.
[0250] In addition, methods 2 and 3 can also be combined. Figure 2D In the embodiment of the present application, after obtaining the third result, it is detected whether the third result is a strong response result.
[0251] If the result is a strong response, negative feedback data is generated as feature data for the next iteration of the data analysis model parameter update until the data analysis model converges and the model training is completed.
[0252] If it is not a strong response result, you can choose whether to evaluate the accuracy of the third result. If not, directly generate positive feedback data as the feature data for the next data analysis model parameter iteration update. If evaluated, generate positive feedback data based on the evaluation results as the feature data for the next data analysis model parameter iteration update. Until the data analysis model converges, the model training is completed.
[0253] The embodiment of the present application does not make too many restrictions on the specific data forms of positive feedback data and negative feedback data, and can be set by technical personnel according to actual needs. For example, the feedback data can include the third result of the current processing, the data to be processed, and a weight coefficient that characterizes the matching degree of the third result and the data to be processed. For positive feedback data, the weight coefficient can be set to a larger value. For negative feedback data, the weight coefficient is set to a smaller value. The weight coefficient can be used as a reference for iterative updates of the data analysis model. The numerical values corresponding to the larger and smaller values of the specific weight coefficient can be set by the technicians themselves, and no excessive restrictions are made here.
[0254] The embodiment of the present application does not impose too many restrictions on the accuracy evaluation method of the third result, which can be set by technical personnel according to actual conditions. As an optional embodiment of the present application, considering that in actual applications, the processing frequency of the processing model for the data to be processed may be relatively high. For example, in a large human-computer dialogue processing platform, tens of thousands of different dialogue contents may need to be processed in one second. Therefore, in order to ensure that the data analysis model can efficiently and accurately analyze the data to be processed, when the data analysis model is updated online, the timeliness of the online update is often required to be high. Correspondingly, the timeliness requirements for identifying whether the third result obtained each time and the data to be processed are strong response situations are also relatively high. At the same time, the timeliness of the accuracy evaluation of the third result is also required to be relatively high.
[0255] In order to meet the rapid identification of whether the third result and the data to be processed are strong responses. In an embodiment of the present application, the data to be processed and the corresponding analysis results of the historical processing of the second type of analysis task will first be recorded, and whether it is a strong response will be marked. On this basis, a model with a faster recognition speed is selected as the recognition model in the embodiment of the present application, and the recorded data is used as sample data to train the recognition model. Thereby, a model that can quickly identify whether the third result and the data to be processed are strong responses is obtained. Similarly, a model with a faster calculation speed for the accuracy of the third result will also be selected as the accuracy assessment model in the embodiment of the present application. The specific types of recognition models and accuracy assessment models are not limited here, and can be selected or set by technical personnel according to actual needs. For example, in some optional embodiments, the DeepGBM model can be selected as the recognition model in the embodiment of the present application, and the PointwiseHSIC model can be selected as the accuracy assessment model in the embodiment of the present application.
[0256] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of the present application. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.
[0257] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.
[0258] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or combinations thereof.
[0259] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.
[0260] As used in the specification and appended claims of this application, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [described condition or event] is detected" or "in response to detecting [described condition or event]", depending on the context.
[0261] In addition, in the description of the present specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish descriptions, and cannot be understood as indicating or suggesting relative importance. It should also be understood that although the terms "first", "second", etc. are used to describe various elements in some embodiments of the present application in the text, these elements should not be limited by these terms. These terms are only used to distinguish one element from another element. For example, the first table can be named as the second table, and similarly, the second table can be named as the first table without departing from the scope of the various described embodiments. The first table and the second table are both tables, but they are not the same table.
[0262] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.
[0263] The data analysis method based on multiple analysis tasks provided in the embodiments of the present application can be applied to electronic devices such as mobile phones, tablet computers, wearable devices, vehicle-mounted devices, augmented reality (AR) / virtual reality (VR) devices, laptop computers, ultra-mobile personal computers (UMPC), netbooks, personal digital assistants (PDA), etc. The embodiments of the present application do not impose any restrictions on the specific types of electronic devices.
[0264] For example, the electronic device may be a station (STAION, ST) in a WLAN, a cellular phone, a cordless phone, a Session Initiation Protocol (SIP) phone, a Wireless Local Loop (WLL) station, a Personal Digital Assistant (PDA) device, a handheld device with wireless communication function, a computing device or other processing device connected to a wireless modem, a vehicle-mounted device, a vehicle networking terminal, a computer, a laptop computer, a handheld communication device, a handheld computing device, a satellite wireless device, a wireless modem card, a TV set top box (STB), a customer premises equipment (CPE) and / or other devices for communicating on a wireless system and a next generation communication system, for example, an electronic device in a 5G network or an electronic device in a future evolved Public Land Mobile Network (PLMN) network, etc.
[0265] As an example but not limitation, when the electronic device is a wearable device, the wearable device can also be a general term for wearable devices that are intelligently designed and developed using wearable technology for daily wear, such as glasses, gloves, watches, clothing and shoes. A wearable device is a portable device that is worn directly on the body or integrated into the user's clothes or accessories. Wearable devices are not just hardware devices, but also achieve powerful functions through software support, data interaction, and cloud interaction. Broadly speaking, wearable smart devices include those that are fully functional, large in size, and can achieve complete or partial functions without relying on smartphones, such as smart watches or smart glasses, as well as those that only focus on a certain type of application function and need to be used in conjunction with other devices such as smartphones, such as various types of smart bracelets and smart jewelry for vital sign monitoring.
[0266] Figure 3 Schematic diagram of the structure of an electronic device provided by an embodiment of the present application. Figure 3 As shown, the electronic device 3 of this embodiment includes: at least one processor 30 ( Figure 3 Only one is shown in the figure), a memory 31, in which a computer program 32 that can be run on the processor 30 is stored. When the processor 30 executes the computer program 32, the steps in the above-mentioned data analysis method embodiment based on multiple analysis tasks are implemented, for example Figure 1B Steps 101 to 106 are shown. Alternatively, Figure 1C The processor 30 can also implement the functions of each module / unit in the above-mentioned device embodiments when executing the computer program 32.
[0267] The electronic device 3 may be a computing device such as a desktop computer, a notebook, a PDA, or a cloud server. The electronic device may include, but is not limited to, a processor 30 and a memory 31. Those skilled in the art will appreciate that Figure 3 It is only an example of the electronic device 3 and does not constitute a limitation of the electronic device 3. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the electronic device may also include an input sending device, a network access device, a bus, etc.
[0268] The processor 30 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.
[0269] In some embodiments, the memory 31 may be an internal storage unit of the electronic device 3, such as a hard disk or memory of the electronic device 3. The memory 31 may also be an external storage device of the electronic device 3, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the electronic device 3. Further, the memory 31 may also include both an internal storage unit of the electronic device 3 and an external storage device. The memory 31 is used to store an operating system, an application program, a boot loader (BootLoader), data, and other programs, such as the program code of the computer program, etc. The memory 31 may also be used to temporarily store data that has been sent or is to be sent.
[0270] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.
[0271] An embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in the above-mentioned method embodiments can be implemented.
[0272] An embodiment of the present application provides a computer program product. When the computer program product runs on an electronic device, the electronic device can implement the steps in the above-mentioned method embodiments when executing the computer program product.
[0273] If the integrated module / unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc.
[0274] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.
[0275] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.
[0276] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0277] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.
[0278] Finally, it should be noted that the above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.
Claims
1. A data analysis method based on multiple analysis tasks, It is characterized in that include: Acquire data to be processed, and trigger at least one first analysis task and at least one second analysis task in parallel, wherein both the first analysis task and the second analysis task are used to analyze the data to be processed; Obtaining a first result generated by each of the first analysis tasks, and a second result obtained by predicting a result of each of the second analysis tasks; Obtaining a sorting result obtained by prioritizing the first result and the second result, and determining whether it is necessary to wait for the second analysis task to generate an analysis result according to the sorting result; If it is determined according to the sorting result that there is no need to wait for the second analysis task to generate an analysis result, the result with the highest priority in the sorting result is used as the output result.
2. The data analysis method based on multiple analysis tasks according to claim 1, It is characterized in that The determining, according to the sorting result, whether it is necessary to wait for the second analysis task to generate an analysis result includes: If the result with the highest priority in the sorted results is the first result, it is determined that there is no need to wait for the second analysis task to generate an analysis result.
3. The data analysis method based on multiple analysis tasks according to claim 1, It is characterized in that The determining, according to the sorting result, whether it is necessary to wait for the second analysis task to generate an analysis result includes: If the result with the highest priority in the sorting results is the second result, taking the second analysis task corresponding to the result with the highest priority as the target analysis task, and determining that it is necessary to wait for the target analysis task to generate an analysis result; Accordingly, the data analysis method based on multiple analysis tasks also includes: If it is determined according to the sorting result that it is necessary to wait for the target analysis task to generate an analysis result, the analysis result generated by the target analysis task is obtained and used as an output result.
4. The data analysis method based on multiple analysis tasks according to any one of claims 1 to 3, It is characterized in that The obtaining a sorting result obtained by sorting the first result and the second result by priority includes: If a third result generated by one or more of the second analysis tasks is obtained, obtaining a sorting result obtained by sorting the first result, the second result, and the third result by priority; Correspondingly, judging whether it is necessary to wait for the second analysis task to generate an analysis result according to the sorting result includes: If the result with the highest priority in the sorted results is the first result or the third result, it is determined that there is no need to wait for the second analysis task to generate an analysis result.
5. The data analysis method based on multiple analysis tasks according to any one of claims 1 to 3, It is characterized in that The obtaining a sorting result obtained by sorting the first result and the second result by priority includes: If a third result generated by one or more of the second analysis tasks is obtained, the result corresponding to the third result is removed from the obtained second results, and the first result, the third result, and the second result remaining after the removal operation are prioritized to obtain the sorted result; Correspondingly, judging whether it is necessary to wait for the second analysis task to generate an analysis result according to the sorting result includes: If the result with the highest priority in the sorted results is the first result or the third result, it is determined that there is no need to wait for the second analysis task to generate an analysis result.
6. The data analysis method based on multiple analysis tasks according to any one of claims 1 to 3, It is characterized in that The operation of obtaining a second result obtained by predicting the result of the second analysis task includes: Obtaining a pre-trained data analysis model corresponding to the second analysis task in one-to-one correspondence, and inputting the data to be processed into the data analysis model for processing, wherein the data analysis model is used to predict the result of the corresponding second analysis task; The second result generated by the data analysis model is obtained.
7. The data analysis method based on multiple analysis tasks according to claim 6, It is characterized in that The training process of a single data analysis model further includes: Acquire sample data, and update parameters of the data analysis model based on the sample data; After each parameter update operation is completed, analyzing the accuracy data of the sample data analyzed by the data analysis model based on a preset classifier after the parameter update operation is completed; Based on the accuracy data, the data analysis model is trained by iterative parameter update until the data analysis model meets a preset convergence condition and the training is completed.
8. The data analysis method based on multiple analysis tasks according to claim 7, It is characterized in that Also includes: The third result generated by each of the second analysis tasks is obtained, and based on the data to be processed and the third result, parameters of each of the data analysis models are updated.
9. The data analysis method based on multiple analysis tasks according to claim 8, It is characterized in that In the operation of updating parameters of each of the data analysis models based on the data to be processed and the third result, the operation of updating parameters of the data analysis model corresponding to a single second analysis task includes: Performing a matching analysis on the data to be processed and the third result generated by the second analysis task; If the data to be processed matches the third result generated by the second analysis task, parameters of the data analysis model corresponding to the second analysis task are updated based on the data to be processed and the third result generated by the second analysis task.
10. A data analysis method based on multiple analysis tasks, It is characterized in that include: Acquire data to be processed, and trigger at least one first analysis task and at least one second analysis task in parallel, wherein both the first analysis task and the second analysis task are used to analyze the data to be processed; Obtaining first results generated by each of the first analysis tasks, obtaining second results obtained by predicting results of each of the second analysis tasks, and confidence levels corresponding to each of the second results; Obtaining a first sorting result obtained by prioritizing the first result and the second result, and determining whether it is necessary to wait for the second analysis task to generate an analysis result according to the first sorting result and the confidence level; If it is determined according to the first sorting result and the confidence level that there is no need to wait for the second analysis task to generate an analysis result, the result with the highest priority in the first sorting result is used as the output result.
11. The data analysis method based on multiple analysis tasks according to claim 10, It is characterized in that The determining, according to the first sorting result and the confidence level, whether it is necessary to wait for the second analysis task to generate an analysis result includes: If the target result is not included in the top n results with the highest priority in the first sorting result, and the result with the highest priority in the first sorting result is the first result, it is determined that there is no need to wait for the second analysis task to generate an analysis result, wherein the target result is the second result whose confidence level is higher than a preset first threshold, and n is an integer greater than 1.
12. The data analysis method based on multiple analysis tasks according to claim 10 or 11, It is characterized in that The determining, according to the first sorting result and the confidence level, whether it is necessary to wait for the second analysis task to generate an analysis result includes: If the target result is not included in the top n results with the highest priority in the first sorting result, and the result with the highest priority in the first sorting result is the second result, the second analysis task corresponding to the result with the highest priority is taken as the target analysis task, and it is determined that it is necessary to wait for the target analysis task to generate an analysis result, wherein the target result is the second result whose confidence level is higher than a preset first threshold, and n is an integer greater than 1; Accordingly, the data analysis method based on multiple analysis tasks also includes: If it is determined according to the first sorting result and the confidence level that it is necessary to wait for the target analysis task to generate an analysis result, the analysis result generated by the target analysis task is obtained and used as an output result.
13. The data analysis method based on multiple analysis tasks according to claim 10 or 11, It is characterized in that The determining, according to the first sorting result and the confidence level, whether it is necessary to wait for the second analysis task to generate an analysis result includes: If the first n results with the highest priority in the first sorting result include a target result, and the result with the highest priority in the first sorting result is the target result, the second analysis task corresponding to the result with the highest priority is used as the target analysis task, and it is determined that it is necessary to wait for the target analysis task to generate an analysis result, wherein the target result is the second result whose confidence level is higher than a preset first threshold, and n is an integer greater than 1; Accordingly, the data analysis method based on multiple analysis tasks also includes: If it is determined according to the first sorting result and the confidence level that it is necessary to wait for the target analysis task to generate an analysis result, the analysis result generated by the target analysis task is obtained and used as an output result.
14. The data analysis method based on multiple analysis tasks according to claim 10 or 11, It is characterized in that The determining, according to the first sorting result and the confidence level, whether it is necessary to wait for the second analysis task to generate an analysis result includes: If the target result is included in the top n results with the highest priority in the first sorting result, and the result with the highest priority in the first sorting result is not the target result, it is determined that it is necessary to wait for the second analysis task to generate an analysis result, wherein the target result is the second result whose confidence level is higher than a preset first threshold, and n is an integer greater than 1.
15. The data analysis method based on multiple analysis tasks according to claim 14, It is characterized in that Also includes: If it is necessary to wait for the second analysis task to generate an analysis result according to the first sorting result and the confidence level, obtaining a third result corresponding to each of the target results in the first n results, wherein the third result is an analysis result generated by the second analysis task corresponding to the target result; Prioritize the first result among the first n results, the second result whose confidence is lower than or equal to the first threshold, and the acquired third result to obtain a corresponding second sorting result; If the result with the highest priority in the second sorting results is the first result or the third result, the result with the highest priority in the second sorting results is used as the output result.
16. The data analysis method based on multiple analysis tasks according to claim 15, It is characterized in that Also includes: If the result with the highest priority in the second sorting results is the second result, obtain the third result corresponding to the second result, and use the obtained third result as the output result, wherein the third result is the analysis result generated by the second analysis task corresponding to the result with the highest priority in the second sorting results.
17. The data analysis method based on multiple analysis tasks according to claim 14, It is characterized in that Also includes: If it is necessary to wait for the second analysis task to generate an analysis result according to the first sorting result and the confidence level, obtaining a third result corresponding to each of the second results in the first n results, wherein the third result is an analysis result generated by the second analysis task corresponding to the second result; Prioritize the first result in the first n results and the acquired third result to obtain a corresponding second ranking result; The result with the highest priority among the second sorting results is used as the output result.
18. The data analysis method based on multiple analysis tasks according to any one of claims 10, 11, 15 to 17, It is characterized in that The operation of obtaining a second result obtained by predicting the result of the second analysis task includes: Obtaining a pre-trained data analysis model corresponding to the second analysis task in one-to-one correspondence, and inputting the data to be processed into the data analysis model for processing, wherein the data analysis model is used to predict the result of the corresponding second analysis task; The second result generated by the data analysis model is obtained.
19. The data analysis method based on multiple analysis tasks according to claim 18, It is characterized in that The training process of a single data analysis model further includes: Acquire sample data, and update parameters of the data analysis model based on the sample data; After each parameter update operation is completed, analyzing the accuracy data of the sample data analyzed by the data analysis model based on a preset classifier after the parameter update operation is completed; Based on the accuracy data, the data analysis model is trained by iterative parameter update until the data analysis model meets a preset convergence condition and the training is completed.
20. The data analysis method based on multiple analysis tasks according to claim 19, It is characterized in that Also includes: The third result generated by each of the second analysis tasks is obtained, and based on the data to be processed and the third result, parameters of each of the data analysis models are updated.
21. The data analysis method based on multiple analysis tasks according to claim 20, It is characterized in that In the operation of updating parameters of each of the data analysis models based on the data to be processed and the third result, the operation of updating parameters of the data analysis model corresponding to a single second analysis task includes: Performing a matching analysis on the data to be processed and the third result generated by the second analysis task; If the data to be processed matches the third result generated by the second analysis task, parameters of the data analysis model corresponding to the second analysis task are updated based on the data to be processed and the third result generated by the second analysis task.
22. An electronic device, It is characterized in that The electronic device includes a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, it implements the method according to any one of claims 1 to 9, or implements the method according to any one of claims 10 to 21.
23. A computer-readable storage medium storing a computer program. It is characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 9 is implemented, or the method according to any one of claims 10 to 21 is implemented.
