Research On Improving Dialogue State Tracking Accuracy Based On Task-oriented Dialogue System

With the rapid development of the information era,human-computer interaction can make it easier and faster for humans to obtain information.Dialogue state tracking is an important part of the human-computer dialogue system,which is the basis for generating dialogue policy.With the process of dialogue interaction,the accuracy of updating dialogue state can directly affect the performance of the dialogue system.Therefore,the research of dialogue state tracking has a more important significance.The traditional dialogue state tracking task is based on the results of the upstream model which called natural language understanding module.Due to the accumulation of errors,the update belief state accuracy will decline.Therefore,in recent years,the research on dialogue state tracking tasks is based on joint modeling of natural language understanding and dialogue state tracking.At present,there are also some problems.Aiming at two of them,including multi-granularity feature extraction in dialogue information context and accurate recognition of dialog states composed by rare slot value pairs,the thesis conducts a series of researches.And the main contributions are as follows.Firstly,aiming at the semantic feature extraction is relatively simple which in the context modeling of dialogue history information.The thesis considers multi-granularity feature fusion for semantic modeling,which is based on Memory Network and add self-attention mechanism,at the same time adding structured semantic information as one of input.Using multi-dimensional semantic features to enhance model learning.Secondly,aiming at the capture of rare slot value pairs in the dialogue corpus,the thesis adds slot value pairs as one of input of the model for learning and adds a scoring module to compute the confidence score for evaluating slot value pairs.The model learns the feature of slot value pairs and two classifications's method in the model can effectively alleviate the problem of identifying rare slot value pairs.Thirdly,combining the first two parts,under the dialogue state tracking task,the thesis designs two models to solve the two proposed problems,MBSAP(Mem_BiLstm_Self_Attention_Parallel)model and MBSAS(Mem_BiLstm_Self_Attention_Serial)model.Experimental results on the WOZ dataset are 85.7%and 84.5%,both exceeding the comparative model,which verify the validity of the model designed in this thesis.Finally,based on the DST model designed in this thesis,a dialogue state tracking display platform for inquiry restaurant is implemented.