Research On Joint Modeling Of Intent Detection And Semantic Slot Filling In Spoken Language Understanding

Spoken language understanding is an important function module of the dialogue system.The performance directly affects subsequent conversation management.Intent detection and semantic slot filling are two key sub-tasks of spoken language understanding.This thesis mainly focuses on the spoken language understanding part in the dialogue system.Traditional method,intent detection and semantic slot filling are solved according the independent model which does not consider the correlation of two tasks,so the present stage most researchers solve intent detection and semantic slot jointly using the same model.Characteristics of the model can be shared by two tasks,intent and semantic slot filling can be recognized using a model output,which is helpful for dialogue management in subsequent talks.It can reduce error accumulation.This thesis improves on the basis of joint modeling.The work consists of independent modeling research and joint modeling research integrating multiple methods.The specific works of this thesis are as follows.(1)The first is independent modeling research of the two tasks.Due to the better effect of Support Vector Machine(SVM)in intent detection task,this thesis adopts SVM to conduct independent modeling research on intent detection.At the same time,it adopts Convolutional Neural Network(CNN),Bidirectional Gated Recurrent Unit(BiGRU)and combination of the two models to study intent detection.Semantic slot filling task is similar to the named entity recognition task,which usually adopts the method of sequence annotation,the effect of Conditional Random Fields(CRF)is remarkable in this task.The thesis adopts the CRF researching semantic slot filling.The results of independentmodeling are compared with the joint recognition model which integrates various methods in this study.(2)This thesis proposes a joint recognition model that integrates multiple methods to model the intent and semantic slot filling,so as to optimize the semantic framework.Firstly,the deep learning model Bidirectional Long Term Memory(BiLSTM)is used to acquire contextual semantic grammar features.Secondly,attention mechanism is added to the two tasks to overcome the defect which BiLSTM can't focus on the input sequences,so as to achieve the focused learning of all input sequences at different times and better obtain the input characteristic information.Thirdly,because the result of intent detection has a positive effect on the semantic slot filling task,the result of intent detection is applied to the semantic slot filling task by using the slot-gated mechanism.Fourthly,considering the dependency relationship before and after the semantic slot filling task label,CRF is used as the decoding model in the semantic slot filling task after the feature extraction from the deep learning model.In this thesis,experiments are carried out on two data sets.It compares the proposed joint recognition model with the independent model of two tasks and other joint recognition models.Experimental results show that the experimental model in this study is superior to other models,which proves the positive effect of intent detection results on semantic slot filling.Combined with the statistical model,it can take into account the inter-dependence relationship between before and after the tag sequence,which is of great significance for the research of the follow-up dialog system.