Research And Implementation Of LSTM-based Dialog State Tracking Model

With the rapid development of information technology,flexible and quick way of human-computer interaction has gradually come into people's lives.The dialog state tracking is an important part of human-machine dialog system,which is the basis of system decision-making.Therefore,The dialog state tracking has important research significance.The dialog state tracking is based on the users' speech information in multi-turn of dialog,which updates belief state and provides support for the dialog decision-making.The current problems are as follows:The dialog state tracking is usually carried out on the results of spoken language understanding,which has the problems of the accumulation of errors.And the accuracy of the dialog state tracking will decrease with the increase in new-value problem.This thesis is the research on the above problems.And the main contributions are as follows.In this thesis,a dialog state tracking model based on cascaded neural network model is proposed,which includes hierarchical model and similarity calculation.First,hierarchical model combines spoken language understanding with dialog state tracking,which avoids the error transfer caused by spoken language understanding.The model consists of two layers of neural networks,the lower layer is encoding users' semantic information,and the upper layer uses the LSTM model to get multi-turn of dialog information.In the feature extraction of the lower layer,we present a variety of vector representations of speech recognition results,which abstracts the deeper semantic features and reduces time cost.Besides,a similarity calculation method is proposed to calculate the similarity between users' semantic information and slot values using the attention model or edit distance.The cascade neural network model combines hierarchical model with similarity computing model to effectively alleviate the new-value problem in the s dialog state tracking.Finally,the experimental results on the public dataset DSTC2 are 72.4%,and the experimental results on DSTC3 which contains the new-value problem are 52.2%.The cascaded neural network modehas better performance than the compared ones,which verify the validity of the model.And this thesis implements the display platform based on proposed model,which is used to show the results of the dialog state tracking.