Method and device for constructing ASR/NLU engine cross optimization data closed loop

Abstract

The invention provides a method for constructing an ASR / NLU engine cross optimization data closed loop, and the method comprises the steps: extracting an abnormal log related to user voice interactionbased on a pre-stored log database, restoring a user interaction process related to the abnormal log, and forming a cross verification log; sending a to-be-identified voice log in the cross validation log to the ASR engine, obtaining a to-be-identified text log, sending the to-be-identified text log to the NLU engine, and obtaining a validation intention log; classifying and comparing the obtained to-be-identified text log with a to-be-compared text log in the cross validation log, and the validation intention log with a to-be-compared intention log in the cross validation log, and obtainingan engine problem classification result; and pushing the abnormal user interaction category to a related equipment terminal for display according to an engine problem classification result, and constructing a corresponding abnormal test set according to the abnormal user interaction category. A data closed loop optimized by the ASR / NLU engine is conveniently formed, and the initiative of discovering the abnormal problem of the engine is improved.

Description

technical field [0001] The present invention relates to the technical field of data closed-loop, in particular to a method and device for constructing ASR / NLU engine cross-optimized data closed-loop. Background technique [0002] In the voice interaction scenario, it is not easy to perceive whether the user experience is good or not. Generally, the following two methods are used to determine whether the user experience is good: one is client feedback, including: ASR engine speech recognition is inaccurate, service response is slow or service is unavailable, NLU engine semantic understanding is inaccurate, service response is slow, service is unavailable etc.; Second, general testers will regularly submit fixed test sets for testing, trying to find engine problems, but in the process of using the above two methods to determine, because the proactive performance of engine abnormal problem discovery is poor, it needs to be started after user feedback Troubleshooting, resulting...

AI Technical Summary

Problems solved by technology

Generally, the following two methods are used to determine whether the user experience is good: one is client feedback, including: ASR engine speech recognition is inaccurate, service response is slow or service is unavailable, NLU engine semantic understanding is inaccurate, service response is slow, service is unavailable etc.; Second, general testers will regularly submit fixed test sets for testing, trying to find engine problems, but in the process of using the above two methods to determine, because the proactive performance of engine abnormal problem discovery is poor, it needs to be started after user feedback Troubleshoot issues that lead to poor user experience

Images