Research And Application Of Spoken Language Understanding Based On ERNIE

The spoken language understanding task mainly includes three sub-tasks: domain recognition,intent recognition and semantic slot filling.Therefore,the research on spoken language understanding can be divided into rough spoken language understanding that only considers a single subtask and refined spoken language understanding that combines multiple subtasks.Therefore,the research is carried out from the two perspectives of spoken language understanding and the perspective of application,and the main work is as follows:(1)Aiming at the problem that short dialogues can easily lead to sparse text semantics in spoken language understanding,and the randomness of dialogues can lead to ambiguous intentions.An intention recognition model that integrates entity information and time series features is proposed.The problem of semantic sparseness is solved by identifying the entity information in the dialogue and adding it to the training of dynamic word vectors.The bidirectional long short-term memory network is used to extract the relationship of the intention in the contextual dialogue to solve the problem of ambiguity of intention.Finally,the internal autocorrelation mechanism of g MLP is used to adaptively fuse entity information and time series features to further improve the correlation between different features.Experiments on related datasets CCKS2018 and SMP2018 show that the proposed model can effectively improve the performance of intent recognition.(2)Aiming at the limited improvement of single-task modeling performance in the current spoken language understanding task.The problem that the concatenation of multiple sub-tasks will cause errors to propagate between tasks,a bidirectional association multi-round information sharing spoken language understanding model(BA-MIS)is proposed.Firstly,a pre-trained model is used to obtain the dynamic representation of the text;considering the influence of contextual information on intention recognition and semantic slot filling.The bidirectional gated recurrent unit is used to obtain the contextual information,and the Attention is used to extract the two tasks to obtain the information required for the task.Then build a multi-round information sharing network(IN-SN),in the first round,the correlation between intent and shallow semantic slot features is calculated in the IN subnet to obtain the intent enhancement vector.and then the intent enhancement vector is sent to the SN subnet for sharing to obtain the semantic slot enhancement vector.In the second round of sharing,the semantic slot enhancement vector replaces the shallow semantic slot feature for sharing to improve the degree and accuracy of information sharing between tasks.Finally,the classification network is used to perform intent enhancement.Recognition and semantic slot filling use Label Attention Network which can consider global information to label semantic slots.Experiments are carried out on the standard datasets of spoken language understanding ATIS and SNIPS,the results show that the model can effectively improve the performance of spoken language understanding.(3)Based on the BA-MIS model,taking ATIS as the basic data set and investigating the actual demands of the airline booking field,this thesis designs and implements an intelligent spoken language understanding system for airline booking.By uploading the offline trained BA-MIS model to the server in advance,and writing the corresponding back-end code to call the model and the corresponding page code to visualize the results of spoken language understanding.