Study On End-to-End Dialogue State Tracking Based On Deep Learning

Dialogue system is a hot topic in the field of artificial intelligence.At present,most of the latest research on dialogue system is done by deep learning technology.Deep learning uses big data to learn feature representations and strategies,requiring only a small amount of manual feature extraction.Dialogue state tracking task has become a hot topic in the field of dialogue in recent years,especially the end-to-end dialogue state tracking research based on deep learning in the past three years.Dialogue state tracking is the core component of task-oriented dialogue system,which aims to predict the target of the current round of users according to the context information.On the one hand,traditional dialogue state tracking method takes the output of the natural language understanding module as the input,which increases the error transmission to some extent.On the other hand,its ability to interact with the historical information and the current information is not strong.With the increase of the dialogue round,the recognition effect of new slot value is not good.At present,the research corpus of conversation state tracking is mainly English dataset.Therefore,in order to explore the effect of dialogue state tracking model on Chinese corpus cross-domain data set,a dialogue state tracking method based on hierarchical attention network is designed,and it is compared with Slot-sentence state tracking network for comparative analysis.The end-to-end dialogue state tracking algorithm based on deep learning designed in this paper can obtain the dialogue state directly from user utterance and system feedback and can update the dialogue state in real time.Regarding the problems faced by traditional dialogue state tracking,this article quotes and constructs a BERT-based slot-sentence belief tracking network,the model uses BERT as encoder,respectively for candidate trough,the current round of trough value,the current round of users,the current round of feedback system for coding,the coding results using long attention mechanism to interact,then a dialogue on results of the interaction and dialogue in the input LSTM model state tracking.The model has achieved good results on English WOZ2.0 single-domain dataset and Chinese CrossWOZ cross-domain dataset.In order to solve the problem that traditional dialogue state tracking methods cannot handle long rounds of dialogue and the slower speed of groove-sentence belief tracking networks,this paper makes simple modifications on the basis of previous models and proposes a hierarchical attention network based on ALBERT.The model inputs all historical dialogues on the basis of the slot-sentence belief tracking network,uses the multi-head attention mechanism to interact with candidate slots and all dialogues in multiple rounds,and obtains information between candidate slots and all rounds of dialogue sentences(global information),the interactive information between the candidate slot and the current round of dialogue.Subsequently,the LSTM structure is used to model the global information and the current round information to obtain the dialogue state of the current round.Experiments show that compared to the groove-sentence belief tracking network,the model has the advantages of faster prediction speed,higher accuracy,and less memory in both Chinese and English corpus.